SigmaXL Version 8 Workbook

SigmaXL Version 8.
0 ®
Workbook
Contact Information:
Technical Support: 1-866-475-2124 (Toll Free in North America) or
1-519-579-5877
Sales: 1-888-SigmaXL (888-744-6295)
E-mail: [email protected] Web: www.SigmaXL.com
Copyright © 2004-2017, SigmaXL Inc.

Published January, 2017
Table of Contents
SigmaXL® What’s New in Version 8.0, System Requirements and Getting Help .......................1
What’s New in Version 8.0............................................................................................................. 2
Activating SigmaXL ................................................................................................................... 6
Error Messages ........................................................................................................................ 8
SigmaXL® Defaults and Menu Options ........................................................................................... 9
Clear Saved Defaults ............................................................................................................... 9
Data Selection Default ............................................................................................................. 9
Menu Options (Classical or DMAIC) ....................................................................................... 10
SigmaXL® System Requirements .................................................................................................. 11
Getting Help and Product Registration ........................................................................................ 12
Introduction to SigmaXL® Data Format and Tools Summary .................................................13
Introduction ........................................................................................................................... 15
The Y=f(X) Model ................................................................................................................... 15
Data Types: Continuous Versus Discrete ............................................................................... 16
Stacked Data Column Format versus Unstacked Multiple Column Format .......................... 17
Summary of Graphical Tools ........................................................................................................ 18
Summary of Statistical Tools ........................................................................................................ 19
SigmaXL: Measure Phase Tools .............................................................................................21
Part A - Basic Data Manipulation ................................................................................................. 23
Introduction to Basic Data Manipulation .............................................................................. 23
Category Subset ..................................................................................................................... 23
Random Subset ...................................................................................................................... 24
Numerical Subset ................................................................................................................... 25
Date Subset ............................................................................................................................ 25
Transpose Data ...................................................................................................................... 26
Stack Subgroups Across Rows ................................................................................................ 26
Stack Columns ........................................................................................................................ 28
Random Data ......................................................................................................................... 29
Box-Cox Transformation ........................................................................................................ 29
Standardize Data .................................................................................................................... 29
iii
SigmaXL: Table of Contents
Data Preparation – Remove Blank Rows and Columns ......................................................... 30

Data Preparation – Change Text Data Format to Numeric.................................................... 32
Recall SigmaXL Dialog ............................................................................................................ 33
Activate Last Worksheet ........................................................................................................ 33
Worksheet Manager .............................................................................................................. 34
Part B – Templates & Calculators ................................................................................................ 35
Introduction to Templates & Calculators .............................................................................. 35
SigmaXL Templates & Calculators.......................................................................................... 36
Team/Project Charter ............................................................................................................ 38
SIPOC Diagram ....................................................................................................................... 39
Data Measurement Plan ........................................................................................................ 39
Cause & Effect (Fishbone) Diagram ....................................................................................... 40
Cause & Effect (Fishbone) Template...................................................................................... 41
Cause & Effect (XY) Matrix Example ...................................................................................... 42
Failure Mode & Effects Analysis (FMEA) Example ................................................................. 43
Quality Function Deployment (QFD) Example ....................................................................... 46
Pugh Concept Selection Matrix Example ............................................................................... 47
Control Plan............................................................................................................................ 48
Lean – Takt Time Calculator Example .................................................................................... 49
Lean – Value Analysis/Process Load Balance......................................................................... 50
Lean – Value Stream Mapping ............................................................................................... 51
Basic Graphical Templates – Pareto Example ........................................................................ 53
Basic Graphical Templates – Histogram Example .................................................................. 54
Basic Graphical Templates – Run Chart Example .................................................................. 56
Basic Statistical Templates – Sample Size – Discrete ............................................................. 58
Basic Statistical Templates – Sample Size – Continuous ....................................................... 60
Basic Statistical Templates – Minimum Sample Size for Robust t-Tests and ANOVA ........... 61
Basic Statistical Templates – 1 Sample Z-Test and Confidence Interval for Mean................ 65
Basic Statistical Templates – 1 Sample t-Test and Confidence Interval for Mean ................ 66
Basic Statistical Templates – 2 Sample t-Test and Confidence Interval (Compare 2 Means)67
Basic Statistical Templates – 1 Sample Equivalence Test for Mean ...................................... 68
Basic Statistical Templates – 2 Sample Equivalence Test (Compare 2 Means) ..................... 69
iv
Basic Statistical Templates – 1 Sample Chi-Square Test and CI for Standard Deviation ....... 70
Basic Statistical Templates – 2 Sample F-Test and CI (Compare 2 Standard Deviations)...... 71
Basic Statistical Templates – 1 Proportion Test and Confidence Interval ............................. 72
Basic Statistical Templates – 2 Proportions Test and Confidence Interval ........................... 74
Basic Statistical Templates – 2 Proportions Equivalence Test............................................... 76
Basic Statistical Templates – 1 Poisson Rate Test and Confidence Interval .......................... 78
Basic Statistical Templates – 2 Poisson Rates Test and Confidence Interval ........................ 80
Basic Statistical Templates – 2 Poisson Rates Equivalence Test............................................ 82
Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit Test............................. 84
Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit Test - Exact ................. 85
Probability Distribution Calculators – Normal ....................................................................... 88
Basic MSA Templates – Type 1 Gage Study ........................................................................... 89
Basic MSA Templates – Gage Bias and Linearity Study ......................................................... 91
Basic MSA Templates – Gage R&R Study (MSA) .................................................................... 93
Basic MSA Templates - Gage R&R: Multi-Vari & X-bar R Charts............................................ 95
Basic MSA Templates – Attribute Gage R&R (MSA) .............................................................. 96
Basic Process Capability Templates – Process Sigma Level – Discrete Data ......................... 98
Basic Process Capability Templates – Process Sigma Level – Continuous Data .................... 99
Basic Process Capability Templates – Process Capability Indices........................................ 100
Basic Process Capability Templates – Process Capability & Confidence Intervals .............. 101
Basic Process Capability Templates – Tolerance Interval Calculator (Normal Exact).......... 102
Basic Control Chart Templates – Individuals Chart.............................................................. 103
Part C – Descriptive/Summary Statistics ................................................................................... 106
Descriptive Statistics ............................................................................................................ 106
Descriptive Statistics - Options ............................................................................................ 108
Part D – Histograms ................................................................................................................... 115
Basic Histogram Template ................................................................................................... 115
Single (Basic) Histogram....................................................................................................... 115
Multiple Histograms............................................................................................................. 116
Part E – Dotplots ........................................................................................................................ 118
Dotplots................................................................................................................................ 118
Part F – Boxplots ........................................................................................................................ 119
v
Boxplots ............................................................................................................................... 119

Part G – Normal Probability Plots .............................................................................................. 123
Normal Probability Plots ...................................................................................................... 123
Part H– Run Charts ..................................................................................................................... 127
Basic Run Chart Template .................................................................................................... 127
Run Charts ............................................................................................................................ 127
Overlay Run Charts .............................................................................................................. 130
Part I – Measurement Systems Analysis .................................................................................... 131
Type 1 Gage Study Template ............................................................................................... 131
Gage Bias and Linearity Study Template ............................................................................. 131
Gage R&R Study (MSA) Template ........................................................................................ 131
Attribute Gage R&R (MSA) Template .................................................................................. 131
Create Gage R&R (Crossed) Worksheet............................................................................... 132
Analyze Gage R&R (Crossed)................................................................................................ 134
Attribute MSA (Binary)......................................................................................................... 138
Attribute MSA (Ordinal) ....................................................................................................... 149
Attribute MSA (Nominal) ..................................................................................................... 155
Part J – Process Capability ......................................................................................................... 157
Process Capability Templates .............................................................................................. 157
Histograms and Process Capability ...................................................................................... 158
Capability Combination Report (Individuals) ....................................................................... 160
Capability Combination Report (Subgroups) ....................................................................... 162
Capability Combination Report (Individuals Nonnormal) ................................................... 164
Distribution Fitting ............................................................................................................... 172
Box-Cox Transformation ...................................................................................................... 176
Part K – Reliability/Weibull Analysis .......................................................................................... 178
Reliability/Weibull Analysis.................................................................................................. 178
SigmaXL: Analyze Phase Tools ............................................................................................183
Part A – Stratification with Pareto ............................................................................................. 185
Basic Pareto Chart Template ............................................................................................... 186
Basic (Single) Pareto Charts ................................................................................................. 186
Advanced (Multiple) Pareto Charts ..................................................................................... 188
vi
Part B - EZ-Pivot ......................................................................................................................... 191

Example of Three X’s, No Response Y’s ............................................................................... 191
Example of Three X’s and One Y .......................................................................................... 194
Example of 3 X’s and 3 Y’s .................................................................................................... 195
Part C – Confidence Intervals..................................................................................................... 196
One Sample Confidence Interval Templates ....................................................................... 196
Confidence Intervals – Descriptive Statistics ....................................................................... 196
Part D – Hypothesis Testing – One Sample Z and t-Test ........................................................... 199
One Sample Z, t and Equivalence Test Templates ............................................................... 199
One Sample t-Test................................................................................................................ 200
Part E – Power and Sample Size ................................................................................................. 203
Power and Sample Size – One Sample t-Test – Customer Data .......................................... 203
Power and Sample Size – One Sample t-Test – Graphing the Relationships between Power,
Sample Size, and Difference ................................................................................................ 207
Part F – One Sample Nonparametric Tests ................................................................................ 209
Introduction to Nonparametric Tests .................................................................................. 209
Exact Nonparametric Tests .................................................................................................. 209
One Sample Sign Test........................................................................................................... 211
One Sample Wilcoxon Signed Rank Test.............................................................................. 212
One Sample Wilcoxon Signed Rank Test - Exact .................................................................. 213
Part G – Two Sample t-Test ....................................................................................................... 219
Two Sample t and Equivalence Test Templates .................................................................. 219
Two Sample t-Test ............................................................................................................... 220
Paired t-Test ......................................................................................................................... 224
Unpaired 2 Sample t-Test vs. Paired t-Test ......................................................................... 226
Power & Sample Size for 2 Sample T-Test ........................................................................... 227
Part H – Two Sample F and Comparison Test ............................................................................ 229
Two Sample F-Test Template ............................................................................................... 229
Two Sample Comparison Test.............................................................................................. 230
Part I – Two Sample Nonparametric Test: Mann-Whitney....................................................... 232
Two Sample Mann-Whitney Test......................................................................................... 232
Two Sample Mann-Whitney Test - Exact ............................................................................. 233
vii
Part J – One-Way ANOVA & Means Matrix ............................................................................... 238

One-Way ANOVA & Means Matrix ...................................................................................... 238
One-Way ANOVA & Means Matrix - Options ...................................................................... 241
Power & Sample Size for One-Way ANOVA......................................................................... 247
Part K – Two-Way ANOVA .......................................................................................................... 249
Two-Way ANOVA ................................................................................................................. 249
Part L – Tests for Equal Variance & Welch’s ANOVA ................................................................ 253
Bartlett’s Test ....................................................................................................................... 253
Levene’s Test........................................................................................................................ 257
Welch’s ANOVA Test (Assume Unequal Variance) .............................................................. 261
Part M – Nonparametric Multiple Comparison ......................................................................... 268
Kruskal-Wallis ....................................................................................................................... 268
Kruskal-Wallis – Exact .......................................................................................................... 271
Mood’s Median Test ............................................................................................................ 275
Mood’s Median Test - Exact ................................................................................................ 277
Part N – Nonparametric Runs Test ............................................................................................ 283
Nonparametric Runs Test for Randomness ......................................................................... 283
Nonparametric Runs Test for Randomness - Examples....................................................... 284
Nonparametric Runs Test for Randomness - Exact ............................................................. 287
Part O – Attribute/Discrete Data Tests ...................................................................................... 292
1 Proportion Test and Confidence Interval Template ......................................................... 292
2 Proportions Test and Confidence Interval Template ........................................................ 292
2 Proportions Equivalence Test Template ........................................................................... 292
One-Way Chi-Square Goodness-of-Fit Template................................................................. 292
One-Way Chi-Square Goodness-of-Fit Template - Exact ..................................................... 293
1 Poisson Rate Test and Confidence Interval Template ...................................................... 293
2 Poisson Rates Test and Confidence Interval Template..................................................... 293
2 Poisson Rates Equivalence Test Template ........................................................................ 293
2 Proportions Test and Confidence Interval Template Example ......................................... 294
Chi-Square Test – Two-Way Table Data............................................................................... 295
Chi-Square Test – Two-Way Table Data: Advanced Tests and Measures of Association – Nominal
Categories ............................................................................................................................ 300
viii
Computing Odds Ratio and Confidence Interval for 2x2 Table ........................................... 304
Chi-Square Test – Two-Way Table Data: Advanced Tests and Measures of Association – Ordinal
Categories ............................................................................................................................ 305
Chi-Square Test (Stacked Column Format Data) ................................................................. 310
Chi-Square Test – Fisher’s Exact .......................................................................................... 315
Chi-Square Test – Two Way Table Data – Fisher’s Exact ..................................................... 319
References for Fisher’s Exact ............................................................................................... 323
Power & Sample Size for One Proportion Test .................................................................... 324
Power & Sample Size for Two Proportions Test .................................................................. 326
Part P – Analysis of Means (ANOM) Charts ............................................................................... 328
Introduction – What is ANOM ............................................................................................. 328
ANOM Normal One-Way ..................................................................................................... 332
ANOM Normal Two-Way (with Main Effects and Slice Charts) ........................................... 334
ANOM Binomial Proportions One-Way ............................................................................... 342
ANOM Binomial Proportions Two-Way ............................................................................... 344
ANOM Poisson Rates One-Way ........................................................................................... 346
ANOM Poisson Rates Two-Way ........................................................................................... 348
ANOM Nonparametric Transformed Ranks......................................................................... 350
ANOM Variances .................................................................................................................. 354
ANOM Levene Robust Variances ......................................................................................... 356
Part Q – Multi-Vari Charts .......................................................................................................... 358
Multi-Vari Charts .................................................................................................................. 358
Part R – Scatter Plots.................................................................................................................. 362
Scatter Plots ......................................................................................................................... 362
Scatter Plot Matrix ............................................................................................................... 364
Part S – Correlation Matrix ........................................................................................................ 365
Correlation Matrix................................................................................................................ 365
Reference ............................................................................................................................. 366
Part T – Multiple Regression ...................................................................................................... 367
Multiple Regression ............................................................................................................. 367
Multiple Regression with Continuous and Categorical Predictors ...................................... 371
Part U – Logistic Regression ....................................................................................................... 373
ix
Binary Logistic Regression.................................................................................................... 373

Ordinal Logistic Regression .................................................................................................. 379
SigmaXL: Improve Phase Tools: Design of Experiments (DOE) .............................................383
Part A – Overview of Basic Design of Experiments (DOE) Templates ....................................... 385
Part B – Three Factor Full Factorial Example Using DOE Template ........................................... 386
Part C – Design and Analysis of Catapult Full Factorial Experiment .......................................... 395
Part D – Design and Analysis of Response Surface Experiment – Cake Bake ............................ 406
SigmaXL: Control Phase Tools: Statistical Process Control (SPC) Charts ............................... 413
Part A - Individuals Charts .......................................................................................................... 415
Individuals Chart Template .................................................................................................. 415
Individuals Chart .................................................................................................................. 416
Individuals Charts: Advanced Limit Options – Historical Groups ........................................ 424
Individuals & Moving Range Charts ..................................................................................... 426
Individuals Charts for Nonnormal Data (Box-Cox Transformation).................................... 427
Part B - X-Bar & Range/StDev Charts ......................................................................................... 431
X-Bar & R Charts ................................................................................................................... 431
X-Bar & R Charts – Exclude Subgroups ................................................................................ 437
X-Bar & S Charts ................................................................................................................... 439
Part C – Attribute Charts (P, NP, C, U) ....................................................................................... 441
P-Charts ................................................................................................................................ 441
P-Charts: Advanced Limit Options – Historical Groups ....................................................... 442
NP-Charts ............................................................................................................................. 444
C-Charts ................................................................................................................................ 445
U-Charts ............................................................................................................................... 446
Part D – P’ & U’ Charts (Laney) .................................................................................................. 447
P’-Charts ............................................................................................................................... 447
U’-Charts .............................................................................................................................. 449
Part E – Control Chart Selection Tool ........................................................................................ 451
Data Types and Definitions .................................................................................................. 451
Control Chart Selection Tool – Individuals Chart ................................................................. 452
Control Chart Selection Tool – X-Bar & R Chart ................................................................... 453
Control Chart Selection Tool – P-Chart ................................................................................ 454
x
Part F – Advanced Charts: I-MR-R/S .......................................................................................... 455

I-MR-R Charts ....................................................................................................................... 455
SigmaXL Appendix ..............................................................................................................459
SigmaXL Version 8.0 Feature List Summary............................................................................... 460
Descriptive Statistics Options and Tolerance Interval Calculator (Normal Exact) ..................... 464
Shapiro-Wilk (SW) and Kolmogorov-Smirnov-Lilliefors (KSL) Normality Test ..................... 464
Doornik-Hansen (DH) Normality Test .................................................................................. 464
Confidence and Tolerance Intervals .................................................................................... 464
Tolerance Interval Calculator (Normal Exact) ...................................................................... 465
Percentile (Nonparametric) Confidence and Tolerance Intervals ....................................... 466
References for Descriptive Statistics Options and Tolerance Interval Calculator (Normal
Exact) .................................................................................................................................... 467
Statistical Details for Nonnormal Distributions and Transformations ...................................... 468
Maximum Likelihood Estimation (MLE) ............................................................................... 468
Distributions ....................................................................................................................... 469
Transformations ................................................................................................................ 481
Automatic Best Fit ............................................................................................................. 484
Process Capability Indices (Nonnormal) ........................................................................ 485
Control Charts (Nonnormal) ............................................................................................ 486
Exact and Monte Carlo P-Values for Nonparametric and Contingency Tests ........................... 487
Introduction ......................................................................................................................... 487
Full Enumeration .................................................................................................................. 487
Exact ..................................................................................................................................... 488
Monte Carlo ......................................................................................................................... 489
Two Tailed (Sided) Tests ...................................................................................................... 489
Sample Sizes for Exact.......................................................................................................... 490
Power ................................................................................................................................... 490
Validation of SigmaXL Exact P-Values .................................................................................. 491
1 Sample Tests ..................................................................................................................... 491
2 Sample Tests ..................................................................................................................... 491
K Sample Tests ..................................................................................................................... 491
Contingency Tests ................................................................................................................ 492
xi
References for Exact and Monte Carlo P-Values ................................................................. 492

Hypothesis Test Assumptions Report ........................................................................................ 494
Normality ............................................................................................................................. 494
Robustness ........................................................................................................................... 494
Outliers (Boxplot Rules) ....................................................................................................... 497
Randomness (Independence) .............................................................................................. 498
Equal Variance ..................................................................................................................... 499
Attribute Measurement Systems Analysis................................................................................. 501
Percent Confidence Intervals (Exact Versus Wilson Score) ................................................. 501
Kappa ................................................................................................................................... 501
Concordance ........................................................................................................................ 502
Kendall’s Coefficient of Concordance .................................................................................. 502
Kendall’s Correlation Coefficient ......................................................................................... 503
References for Attribute Measurement Systems Analysis .................................................. 505
Chi-Square Tests and (Contingency) Table Associations ........................................................... 506
Chi-Square Tests for Nominal Categories ............................................................................ 506
Measures of Association for Nominal Categories ............................................................... 507
Tests for Ordinal Categories ................................................................................................ 509
Measures of Association for Ordinal Categories ................................................................. 510
References for Chi-Square Tests and (Contingency) Table Associations............................. 512
Multiple Comparison of Means and Variances (a.k.a. Post-Hoc Tests) .................................... 514
One-Way ANOVA ................................................................................................................. 514
Welch’s ANOVA (Unequal Variance).................................................................................... 515
Bartlett’s Test for Equal Variance ........................................................................................ 515
Levene’s Test for Equal Variance ......................................................................................... 515
References for Multiple Comparisons of Means and Variances ......................................... 516
Analysis of Means (ANOM) Charts ............................................................................................. 517
ANOM Normal One-Way ..................................................................................................... 517
ANOM Normal Two-Way ..................................................................................................... 518
ANOM Binomial Proportions One-Way ............................................................................... 519
ANOM Binomial Proportions Two-Way ............................................................................... 520
ANOM Poisson Rates One-Way ........................................................................................... 521
xii
ANOM Poisson Rates Two-Way ........................................................................................... 521

ANOM Nonparametric Transformed Ranks......................................................................... 522
ANOM Variances .................................................................................................................. 522
ANOM Levene Robust Variances ......................................................................................... 524
References for Analysis of Means (ANOM) Charts .............................................................. 524
About SigmaXL, Inc. ............................................................................................................ 527
xiii
SigmaXL® What’s New in Version 8.0, System
Requirements and Getting Help

SigmaXL: What’s New, Installation Notes, Getting Help and Product Registration
What’s New in Version 8.0
See Appendix: SigmaXL Version 8.0 Feature List Summary for a complete feature list summary.
New features in SigmaXL Version 8.0 include:
• Templates and Calculators

o 1 Sample Z-test and Confidence Interval for Mean
o Tolerance Interval Calculator (Normal Exact)
o Equivalence Tests - Two One-Sided Tests (TOST)
 1 Sample Equivalence Test for Mean
 2 Sample Equivalence Test (Compare 2 Means)
 2 Proportions Equivalence Test
 2 Poisson Rates Equivalence Test
o Type 1 Gage Study
o Gage Bias and Linearity Study
• Descriptive Statistics
o New features are available by clicking the dialog Options button
o Percentile Report
 27 values from 0.135 to 99.865
o Percentile Ranges
 75 - 25 (50%, Interquartile Range IQR)
 90 - 10 (80%, Interdecile Range IDR)
 95 - 5 (90%, Span)
 97.5 - 2.5 (95%, +/- 1.96 Sigma Equivalent)
 99 - 1 (98%)
 99.5 - 0.5 (99%)
 99.865 - 0.135 (99.73%, +/- 3 Sigma Equivalent)
o Percentile Confidence and Tolerance Intervals
 Interpolated or Exact
 Minimum sample size reported if unable to compute CI or TI
 Quartile Confidence Intervals (25, 50, 75)
 Percentile Confidence Intervals (27 values from 0.135 to 99.865)
 Percentile Tolerance Intervals (50%, 80%, 90%, 95%, 98%, 99%, 99.73%)
o Additional Descriptive Statistics
 5% Trimmed Mean
 Standard Error of Mean
 Variance
 Coefficient of Variation
 Short Term StDev (MR-bar/d2)
2
o Additional Normality Tests

 Shapiro-Wilk (n <= 5000) and Kolmogorov-Smirnov-Lilliefors (n > 5000)
 Doornik-Hansen
• Univariate omnibus test based on Skewness and Kurtosis
• Best for data with ties (“chunky” data)
o Outlier Tests
 Boxplot: Potential 1.5(IQR), Likely 2.2(IQR), Extreme 3.0(IQR)
 Grubbs
o Randomness Test
 Nonparametric Runs Test (Exact)
o Normality, Outlier and Randomness Tests display the same Green, Yellow, Red
highlight used in Version 7 hypothesis tests.
• Analysis of Means (ANOM) Charts

o All charts support balanced & unbalanced data
o Normal One-Way
 Exact critical values via table lookup are used rather than approximation
formulas
o Normal Two-Way with Main Effects and Wludyka Slice Charts.
 Slice Charts are a modified ANOM chart developed by Dr. Peter Wludyka that
enables one to easily interpret the effects in the presence of an interaction.
 Yellow highlight automatically recommends Main Effects (if interaction is not
significant) or Slice Chart (if interaction is significant).
 Interaction P-Value automatically determined from ANOVA
 Option to specify correction to alpha for multiple chart family-wise error rate
(Bonferroni)
o Binomial Proportions One-Way

 Warning is given if normal approximation to binomial is not applicable
(np, nq <= 5)
3
o Binomial Proportions Two-Way with Main Effects and Wludyka & Noguera Slice
Charts
 In collaboration, Peter Wludyka and John Noguera of SigmaXL extended the
Slice Charts to Binomial and Poisson.
 Interaction P-Value automatically determined from Logistic regression
o Poisson Rates One-Way
 Warning is given if normal approximation to Poisson is not applicable
(nu <= 5)
o Poisson Rates Two-Way with Main Effects and Wludyka & Noguera Slice Charts
 Interaction P-Value automatically determined from Poisson regression
o Nonparametric Transformed Ranks
o Variances
o Levene Robust Variances
• One-Way ANOVA
o Multiple Comparison of Means Probability Methods (a.k.a. Post-Hoc)
 Fisher
 Tukey
 Dunnett with Control
o Display ANOM Normal One-Way Chart
o Confidence Level
o Residual Plots
• Welch ANOVA (Assume Unequal Variance)

o Multiple Comparison of Means Probability Methods
 Welch Pairwise
 Games-Howell
o Confidence Level
o Residual Plots
• Two-Way ANOVA
o Display ANOM Normal Two-Way Chart
4
• Bartlett Equal Variance

o Multiple Comparison of Variances Probability Methods
 F-Test Pairwise
 F-Test with Bonferroni Correction
o Display ANOM Variances Chart
o Confidence Level
• Levene Equal Variance

o Multiple Comparison of Variances Probability Methods
 Levene Pairwise
 Tukey ADM (Absolute Deviations from Median)
o Display ANOM Levene Robust Variances Chart
o Confidence Level
• Chi-Square Tests & Table Associations

o Improved dialog labels for stacked data (Rows, Cols, Frequency)
o Adjusted Residuals
 Red font highlight denotes significant cell residual value
 Bold red highlight denotes significant cell residual value with Bonferroni
adjustment
 Note: red highlight is only active if Chi-Square P-Value is significant
o Cell’s Contribution to Chi-Square
o Confidence Level
o Additional Chi-Square Tests (optional):
 Likelihood Ratio
 McNemar-Bowker Symmetry (Square Table)
o Measures of Association for Nominal Categories (optional)
 Pearson's Phi
 Cramer's V
 Contingency Coefficient
 Cohen's Kappa (Agreement - Square Table)
 Goodman-Kruskal Lambda (Cols & Rows Dependent, Symmetric)
 Goodman-Kruskal Tau (Cols & Rows Dependent)
 Theil's Uncertainty (Cols & Rows Dependent, Symmetric)
o Tests of Association for Ordinal Categories (optional)
 Concordant - Discordant
 Spearman Rank Correlation
o Measures of Association for Ordinal Categories (optional)
 Spearman Rank Correlation
 Kendall's Tau-B (Square Table)
 Kendall-Stuart Tau-C (Rectangular Table)
 Goodman-Kruskal Gamma
 Somers' D (Cols & Rows Dependent, Symmetric)
5
Activating SigmaXL
Please proceed with the following steps if you have a valid serial number and your computer is
connected to the Internet. If you do not have an Internet connection, please email
[email protected] from a different system or call 1-888-SigmaXL (1-888-744-6295) or 1-
519-579-5877. If you purchased SigmaXL as a download, you received the serial number by e-
mail. If you purchased a CD, the serial number is on the label of the CD.
If your trial has timed out and you do not have a serial number but wish to purchase a SigmaXL
license, please click Purchase SigmaXL in the Activation Wizard Box and this will take you to
SigmaXL’s order page http://www.sigmaxl.com/Order%20SigmaXL.shtml. You can also call 1-
888-SigmaXL (1-888-744-6295) or 1-519-579-5877 to place an order.
1. In the Activation Wizard box select Activate SigmaXL (Enter a serial number.)
2. Click Next. Enter your serial number. We recommend that you copy and paste the serial
number into the serial number field to avoid typos.
6
3. Click Next. The activation process continues and is confirmed as shown:
4. Click OK. In Excel 2007 and newer, SigmaXL appears as a new Ribbon:
5. In Excel 2011 for Mac, SigmaXL is added to Excel’s menu:
7
Error Messages
1. Incorrect Serial Number:
This is due to an incorrect serial number entry in the previous registration window – click OK
and please reenter your serial number.
2. Serial Number Used in Previous Activation:
The license is currently activated on the maximum allowable number of computers. You will
need to deactivate SigmaXL from your old computer while connected to the Internet. To
deactivate a SigmaXL license go to SigmaXL > Help > Deactivate SigmaXL.
3. Internet Connection Problem:
If you do not have an Internet connection, please email [email protected] from a different
system or call 1-888-SigmaXL (1-888-744-6295) or 1-519-579-5877.
8
SigmaXL® Defaults and Menu Options
Clear Saved Defaults

Clear Saved Defaults will reset all saved defaults such as Pareto and Multi-Vari Chart settings,
saved control limits, and dialog box settings. All settings are restored to the original installation
defaults.
Click SigmaXL > Help > SigmaXL Defaults > Clear Saved Defaults. A warning message is given prior
to clearing saved defaults.
Data Selection Default

The Data Selection Default setting is: Prompt me to select my data range and/or ‘Use Data
Labels’.
This can be changed to: Always use my pre-selected data range without prompting. ‘Data Labels’
will be used. This setting saves you from having to click Next at the start of every function, but the
user is responsible to ensure that the proper data selection is made prior to starting any menu
item.
Click SigmaXL > Help > SigmaXL Defaults > Data Selection Default to make this change. This will
apply permanently unless you revert back to the Prompt me setting or click Clear Saved Defaults
shown above.
9
Menu Options (Classical or DMAIC)

The default SigmaXL menu system groups tools by category, but this can be changed to the Six
Sigma DMAIC format.
Click SigmaXL > Help > SigmaXL Defaults > Menu Options – Set SigmaXL’s Menu to Classical or
DMAIC. The Set Menu dialog allows you to choose between Classical (default) and DMAIC:
If you select the DMAIC Menu, the DMAIC Menu Ribbon appears as shown (Excel 2007 and newer):
All SigmaXL tools are available with this menu format, but they are categorized using the Six Sigma
DMAIC phase format. Note that some tools will appear in more than one phase.
This workbook uses the classical (default) menu format, but the chapters are organized as
Measure, Analyze, Improve and Control.
10
SigmaXL® System Requirements
Minimum System Requirements:

Computer and processor: 1 gigahertz (GHz) processor or higher.
Memory: 2 GB of RAM or greater.
Hard disk: 200 MB of available hard-disk space.
Drive: CD-ROM or DVD drive.
Display: 1024x768 or higher resolution monitor.
Windows Operating System: Microsoft Windows 7 with current Service Packs, or later operating
system.
Microsoft Windows Excel version: Excel 2007 and newer with latest service packs installed.
Mac Operating System: OS X Snow Leopard (10.6) or later
Mac Excel Version: Excel 2011 or later with latest service packs installed.
Administrative Rights to install software.
11
Getting Help and Product Registration
To access the help system, please click SigmaXL > Help > Help.
Technical support is available by phone at 1-866-475-2124 (toll-free in North America) or 1-519-

579-5877 or by e-mail [email protected].
Please note that users obtain free technical support and upgrades for one year from date of
purchase. Optional maintenance is available for purchase prior to the anniversary date.
To register by web, simply click SigmaXL > Help > Register SigmaXL.
12
Introduction to SigmaXL®
Data Format and Tools Summary

SigmaXL: Introduction to Data Format and Tools Summary
Introduction
SigmaXL is a powerful but easy to use Excel Add-In that will enable you to Measure, Analyze,
Improve and Control your service, transactional, and manufacturing processes. This is the ideal
cost effective tool for Six Sigma Green Belts and Black Belts, Quality and Business Professionals,
Engineers, and Managers.
SigmaXL will help you in your problem solving and process improvement efforts by enabling you to
easily slice and dice your data, quickly separating the “vital few” factors from the “trivial many”.
This tool will also help you to identify and validate root causes and sources of variation, which then
helps to ensure that you develop permanent corrective actions and/or improvements.
The Y=f(X) Model

SigmaXL utilizes the “Y=f(X)” model in its dialog boxes. Y denotes a key process output metric; X
denotes a key process input metric. This process is shown pictorially as:
X Process Y
The mathematical expression Y = f(X) denotes that the variable Y is a function of X. Y can also be
viewed as the effect of interest and X is the cause. For example, Y could be customer satisfaction
as measured on a survey and X could be location or responsiveness to calls (also measured on a
survey). The goal is to figure out which X’s from among many possible are the key X’s and to what
extent do they impact the Y’s of interest. Solutions and improvements then focus on those key X’s.
15
Data Types: Continuous Versus Discrete

X and Y metrics can each be continuous or discrete. A continuous measure will have readings on a
continuous scale where a mid-point has meaning. For example, in a customer satisfaction survey
using a 1 to 5 score, the value 3.5 has meaning. Other examples of continuous measures include
cycle time, thickness, and weight. A discrete measure is categorical in nature. If we have Customer
Types 1, 2, and 3, customer type 1.5 has no meaning. Other examples of discrete measures include
defect counts and number of customer complaints.
It is possible to have various combinations of discrete/continuous X’s and discrete/continuous Y’s.

Some examples are given below:
Examples of Discrete (Category) X and Discrete Y
• X = Customer Type, Y = Number of Complaints
• X = Product Type, Y = Number of Defects
• X = Day Shift vs. Night Shift, Y = Proportion of Defective Units
Examples of Discrete (Category) X and Continuous Y
• X = Customer Type, Y = Customer Satisfaction (1-5)
• X = Before Improvement vs. After Improvement, Y = Customer Satisfaction (1-5)
• X = Location, Y = Order to Delivery Time
Examples of Continuous X and Discrete Y
• X = Responsiveness to Calls (1-5), Y = Number of Complaints
• X = Process Temperature, Y = Number of Defects
Examples of Continuous X and Continuous Y
• X = Responsiveness to Calls (1-5), Y = Customer Satisfaction (1-5)
• X = Amount of Loan ($), Y = Cycle Time (Loan Application to Approval)
Note that in SigmaXL, a discrete X can be text or numeric, but a continuous X must be numeric. Y’s
must be numeric. If Y is discrete, count data will be required. If the data of interest is discrete text,
it should be referenced as X1 and SigmaXL will automatically search through the text data to obtain
a count (applicable for Pareto, Chi-Square and EZ-Pivot tools).
16
Stacked Data Column Format versus Unstacked Multiple Column

Format
SigmaXL can accommodate two data formats: stacked column and unstacked multiple column.
The stacked column format has an X column also referred to as the “Group Category” column and a
Y column that contains the data of interest. The following is an example of data in stacked column
format, with three unique groups of Customer Type:
If the data is in unstacked multiple column format, each unique group of X corresponds to a
different column. The above data is now shown in unstacked format with customer satisfaction
scores for each customer type in separate columns:
17
Tool What Type of Data When to Use Location in SigmaXL
Pareto Chart Plots a bar chart of the response in Y=Discrete (e.g., To separate the vital few from the trivial  SigmaXL > Templates & Calculators >
descending order with a cumulative Defect Count) or many, help specify a problem statement, and Basic Graphical Templates > Pareto Chart
sum line. Continuous (e.g., Cost; prioritize potential root causes. This chart is  SigmaXL > Graphical Tools > Basic
must be additive) based on the Pareto principle, which states Pareto Chart (Single)
X=Discrete that typically 80% of the defects in a process  SigmaXL > Graphical Tools > Advanced
(Category) or product are caused by only 20% of the Pareto Charts (Multiple)
possible causes.
Pivot Chart Plots a stacked (or clustered) bar Y=Discrete or To easily ‘slice and dice’ your data, quickly  SigmaXL > Graphical Tools > EZ-
chart from an Excel Pivot Table. Continuous look at different X factors and their Pivot/Pivot Charts
X = Discrete contribution to the total. It is similar to the
(Category) Pareto Chart, but without the descending bar
order.
Histogram Visual display of one variable Y=Continuous 1. Summarize large amounts of data  SigmaXL > Templates & Calculators >
showing data center, spread, shape X=Discrete 2. To get a ‘feel for the data’ Basic Graphical Templates > Histogram
and outliers. (Category) 3. To compare actual description to  SigmaXL > Graphical Tools > Basic
customer specifications Histogram (Single)
 SigmaXL > Graphical Tools > Histograms
& Descriptive Statistics
 SigmaXL > Graphical Tools > Histograms
& Process Capability
Dotplots Visual display of one variable Y=Continuous 1. Small sample size (n < 30)  SigmaXL > Graphical Tools > Dotplots
showing data center, spread, shape X=Discrete 2. To get a ‘feel for the data’
and outliers. (Category)
Boxplots Visual display of the summary of Y Y=Continuous Summary display to visualize differences in  SigmaXL > Graphical Tools > Boxplots
data grouped by category of X. X=Discrete data center, spread and outliers across
18
(Category) categories.
Normal Plots data in a straight line if the Y=Continuous To check for Normality and Outliers.  SigmaXL > Graphical Tools > Normal
Probability Plot data is normally distributed. X=Discrete Probability Plots
(Category)
Run Charts Plots observations in time sequence Y=Continuous or To view process performance over time for  SigmaXL > Templates & Calculators >
Discrete trends, shifts or cycles. Basic Graphical Templates > Run Chart
 SigmaXL > Graphical Tools > Run Chart
To test for Randomness using the  SigmaXL > Graphical Tools > Overlay
Nonparametric Runs Test Run Chart
Multi-Vari Charts Plots vertical lines with dots to allow Y=Continuous To visually compare subgroups by individual  SigmaXL > Graphical Tools > Multi-Vari
comparison of subgroups on one X=Discrete data points and the mean. Charts
variable. (Category) To identify major sources of variation (e.g.,
Summary of Graphical Tools
within a subgroup, between subgroups, or

over time).
Scatter Plot Plots a response Y versus a Y=Continuous To understand the possible relationships  SigmaXL > Graphical Tools > Scatter Plot
(Diagram) predictor X. X=Continuous between two variables.  SigmaXL > Graphical Tools > Scatter Plot
Matrix
To identify possible root causes which are
related to Y.
Analysis of Plots response Y mean for each Y=Continuous ANOM is a complement to ANOVA showing  SigmaXL > Graphical Tools > Analysis of
Means (ANOM) level of X category with decision X=Discrete which group means are significantly different Means (ANOM)
limits. (Category) than the grand mean.
Control Charts Plots observations in time sequence Y=Continuous or To monitor the process over time for trends,  SigmaXL > Control Charts >
against a mean and control limits. Discrete shifts or cycles in order to control and
improve process performance.
To identify special causes.

Tool What Type of Data When to Use Location in SigmaXL
t-Test Determine if there is a significant Y=Continuous 1. Test if mean = specified value SigmaXL > Templates & Calculators > Basic Statistical Templates >
difference between two group X=Discrete (Category) 2. Test if 2 sample means are equal 1 Sample t-Test and Confidence Interval for Mean
means or if the true mean of the 3. Paired t: to reduce variation when comparing 2 Sample t-Test and Confidence Interval (Compare 2 Means)
data is equal to a standard value. two sample means
4. Multiple pairwise comparisons SigmaXL > Statistical Tools >
1 Sample t-Test & Confidence Intervals / Paired t-Test
2 Sample t-Test / 2 Sample Comparison Tests
One-Way ANOVA & Means Matrix
One-Way ANOVA (Analysis Determine if there is a difference Y=Continuous Determine if there is a statistically significant SigmaXL > Statistical Tools > One-Way ANOVA & Means Matrix (for
of Variance) in mean among many groups. X=Discrete (Category) difference in means among the groups. equal variance)
SigmaXL > Statistical Tools > Equal Variance Tests > Welch’s ANOVA
(for unequal variance)
Two-Way ANOVA (Analysis Determine if there is a difference Y=Continuous Determine if there is a statistically significant SigmaXL > Statistical Tools > Two-Way ANOVA
of Variance) in mean among many groups for X1=Discrete (Category) difference in means among the groups.
two factors plus their interaction. X2=Discrete (Category)
Nonparametric Tests Determine if there is a difference Y=Continuous 1. Test if median = specified value: SigmaXL > Statistical Tools > Nonparametric Tests
between two or more group X=Discrete (Category) 1 Sample Sign Test or Wilcoxon
medians or if the median of the 2. Test if 2 sample medians are equal: SigmaXL > Statistical Tools > Nonparametric Tests – Exact (for small
data is equal to a standard value. 2 Sample Mann-Whitney sample)
3. Test if there is a difference in medians among
the groups: Kruskal-Wallis or Mood’s Median
CI for Standard Deviation/ Determine if there is a difference Y=Continuous 1. Determine the confidence interval for a single SigmaXL > Templates & Calculators > Basic Statistical Templates >
F-test / between two or more group X=Discrete (Category) standard deviation. 1 Sample Chi-Square Test and CI for Standard Deviation
Bartlett’s Test/ Levene’s Test variances or determine the 2. Test if 2 sample variances (standard 2 Sample F-Test and CI (Compare 2 Standard Deviations)
confidence interval of a single deviations) are equal.
standard deviation. 3. Determine if there is a statistically significant SigmaXL > Statistical Tools > Two Sample Comparison Tests
difference for the variances among the
groups. Use Bartlett’s test for normal data. SigmaXL > Statistical Tools > Equal Variance Tests >Bartlett / Levene
19
Use Levene’s test for non-normal data.
Proportions Test Determine if there is a difference Y=Discrete (Proportion) 1. Determine the confidence interval for a single SigmaXL > Templates & Calculators > Basic Statistical Templates >
between two proportions or X=Discrete (Category) proportion. 1 Proportion Test & Confidence Interval
determine the confidence interval 2. Determine if there is a statistically significant 2 Proportions Test & Confidence Interval
of a single proportion. difference for two proportions.
Poisson Rate Test Determine if there is a difference Y=Discrete (Count) 1. Determine the confidence interval for a single SigmaXL > Templates & Calculators > Basic Statistical Templates >
between two rates or determine X=Discrete (Category) rate. 1 Poisson Rate Test and Confidence Interval
the confidence interval of a single 2. Determine if there is a statistically significant 2 Poisson Rates Test and Confidence Interval
rate. difference for two rates.
One-Way Chi-Square Determine if the observed Y=Discrete (Count) Test the distribution of observed frequency counts SigmaXL > Templates & Calculators > Basic Statistical Templates >
Goodness-of-Fit frequencies for one discrete against expected (typically uniform discrete). One-Way Chi-Square Goodness-of-Fit Test
variable are distributed equally One-Way Chi-Square Goodness-of-Fit Test Exact (for small sample)
Χ2 Determine if there is a difference Y=Discrete (Count) Determine if there is a relationship between two SigmaXL > Statistical Tools > Chi-Square Tests >
Chi Square for observed frequencies of two X=Discrete (Category) discrete variables. Chi-Square Test (for raw discrete data in stacked column format)
discrete variables. Chi-Square Test – Two Way Table Data (for pivot or contingency table)
Summary of Statistical Tools
SigmaXL > Statistical Tools > Chi-Square Tests – Exact (for small
sample)
Anderson-Darling Normality Determine if the data is normally Y=Continuous Test if the sample data is normally distributed. SigmaXL > Graphical Tools > Histograms & Descriptive Statistics
Test distributed.
SigmaXL > Statistical Tools > Descriptive Statistics
Correlation Quantify strength of relationships. Y=Continuous Determine if there is evidence of a relationship SigmaXL > Statistical Tools > Correlation Matrix
X=Continuous between Xs and Ys, quantify the relationship,
identify root causes.
Regression (Simple Linear & Summarizes, describes, predicts Y=Continuous 1. Determine if there is evidence of a relationship SigmaXL > Statistical Tools > Regression > Multiple Regression
Multiple) and quantifies relationships. X=Continuous between Xs and Ys.
2. Model data to develop a mathematical
Logistic Regression Summarizes, describes, predicts Y=Discrete (Binary or SigmaXL > Statistical Tools > Regression >
equation to quantify the relationship.
and quantifies relationships. Ordinal) Binary Logistic Regression
3. Identify root causes.
X=Continuous Ordinal Logistic Regression
4. Make predictions using the model.
Design of Experiments Systematic and efficient proactive Y=Continuous or Discrete To establish cause and effect relationship between SigmaXL > Design of Experiments
(DOE) approach to testing relationships. X=Continuous or Discrete Ys and Xs. To identify ‘vital few’ Xs.
SigmaXL: Measure Phase Tools

Part A - Basic Data Manipulation
Introduction to Basic Data Manipulation

Open Customer Data.xlsx (to access, click SigmaXL > Help > Sample Data or Start > Programs >
SigmaXL > Sample Data). This data is in stacked column format. This format is highly
recommended for use with SigmaXL. Note that all pertinent information is provided in each record
(row). Also note that only one row is used for column headings (labels) and there are no blank
rows or columns. Each column contains a consistent format of either numeric, text, or date. This is
also the data format used by other major statistical software packages.
Note that Loyalty, Overall Satisfaction, Responsive to Calls, Ease of Communications, and Staff
Knowledge were obtained from surveys. A Likert scale of 1 to 5 was used, with 1 being very
dissatisfied, and 5 very satisfied. Survey results were averaged to obtain non-integer results.
Category Subset
1. Click SigmaXL > Data Manipulation > Category Subset.
2. If you are working with a portion of a dataset, specify the appropriate range, otherwise check
Use Entire Data Table.
3. Click Next.
23
4. Select Customer Type, 1, >> as shown:
5. Click OK.
A new subset worksheet is created containing only Customer Type 1.

Note: We could have chosen more than one Customer Type and had the option to create a
subset which included or excluded these Customer Types.
Random Subset
1. Click Sheet 1 Tab of Customer Data.xlsx.
2. Click SigmaXL > Data Manipulation > Random Subset.
3. Ensure that the entire data table is selected. If not, check Use Entire Data Table. Click Next.
4. Enter Number of Rows in Random Subset as 30. The default Sort Data selection is Original
Order.
24
5. Click OK. A new worksheet is created that contains a random subset of 30 rows.
This feature is useful for data collection to ensure a random sample, e.g., given a list of
transaction numbers select a random sample of 30 transactions.
Numerical Subset
2. Click SigmaXL > Data Manipulation > Numerical Subset.
3. Ensure that entire data table is selected. If not, check Use Entire Data Table. Click Next.
4. Select Overall Satisfaction, >=, Enter Value as 4.
5. Click OK.
A new subset worksheet is created containing only those rows with Overall Satisfaction >= 4.
Date Subset
2. Click SigmaXL > Data Manipulation > Date Subset.
25
4. Select Order Date, select 1/9/2016, click Start Date, select 1/12/2016, click End Date.
5. Click OK. A new subset worksheet is created containing only those rows with Order Date
between 1/9/2016 to 1/12/2016.
Transpose Data
1. Open Catapult Data Row Format.xlsx.
2. Manually select the entire data table if the data is not already selected.
3. Click SigmaXL > Data Manipulation > Transpose Data.
4. This will transpose rows to columns or columns to rows. It is equivalent to Copy, Paste Special,
Transpose.
Stack Subgroups Across Rows

5. Now we will stack the subgroups across rows for the transposed data. Ensure that the
Transposed Data Sheet of Catapult Data Row Format.xlsx is active.
6. Click SigmaXL > Data Manipulation > Stack Subgroups Across Rows.
7. Check Use Entire Data Table, click Next.
8. Click on Shot 1. Shift Click on Shot 3 to highlight the three columns of interest.
26
9. Click Numeric Data Variables (Y) >> to select Shot 1 to Shot 3. Enter Distance as the Stacked
Data (Y) Column Heading and Shot No as the Category (X) Column Heading. Select Operator,
click Additional Category Columns (X) >>:
Note that any selected column may be removed by highlighting and double-clicking or clicking
the Remove button.
10. Click OK. The resulting stacked data is shown:
27
Stack Columns
1. Open Customer Satisfaction Unstacked.xlsx.
2. Click SigmaXL > Data Manipulation > Stack Columns.
4. Shift Click on Overall Satisfaction_3 to highlight all three column names. Click Select Columns
>>. Enter the Stacked Data (Y) Column Heading (Optional) as Overall Satisfaction. Enter the
Category (X) Column Heading (Optional) as Customer Type.
5. Click OK. Shown is the resulting stacked column format:
28
6. Data that is in stacked column format can be unstacked using Data Manipulation > Unstack
Columns.
Random Data
The normal random data generator is used to produce normal random data. Column headings are
automatically created with Mean and Standard Deviation values (e.g. 1: Mean = 0; StDev = 1). This
utility works with Recall SigmaXL Dialog (F3) to append columns to the current Normal Random
Data worksheet. An example is shown in Measure Phase Tools, Part G – Normal Probability Plots.
Additional random number generators include Uniform (Continuous & Integer), Lognormal,
Exponential, Weibull and Triangular. The column headings show the specified parameter values.
Box-Cox Transformation
This tool is used to convert nonnormal data to normal by applying a power transformation.
Examples of use are given in Measure Phase Tools, Part J – Process Capability for Nonnormal Data
and Control Phase Tools, Part A – Individuals Charts for Nonnormal Data.
Standardize Data
This tool is used to Standardize ((Yi – Mean)/StDev) or Code (Ymax = +1, Ymin = -1) your data. This
is particularly useful when performing Multiple Regression. Standardized Predictors have better
statistical properties. For example, the importance of model coefficients can be determined by the
relative size because units are removed. Another statistical benefit is reduced multicollinearity
when investigating two-factor interactions.
29
2. Click SigmaXL > Data Manipulation > Standardize Data.
4. Select Responsive to Calls and Ease of Communications and click Numeric (Y) Columns to
Standardize >>
5. Click OK. The results are given on the Standardize sheet:
Data Preparation – Remove Blank Rows and Columns

This data preparation utility is provided as a convenient way to prepare data for analysis by
deleting any empty rows and/or columns.
1. Open Customer Data.xlsx. Click Sheet 1 Tab.
2. Insert a new column in B; Click Column B heading, Right-Click > Insert > Columns.
3. Insert a new row in row 2. Click Row 2 label, Right-Click > Insert > Rows as shown:
30
4. This is now an example of a data set that requires deletion of empty rows and columns. Click
SigmaXL >Data Manipulation >Data Preparation >Remove Blank Rows and Columns.
5. Check Delete Empty Rows and Delete Empty Columns.
31
6. Click OK. A warning message is given prior to the deletion step.
7. Click Yes. The empty rows and columns are deleted automatically.
Data Preparation – Change Text Data Format to Numeric

This Data Preparation utility will convert data that represents numeric values but are currently in
text format. This sometimes occurs when importing data into Excel from another application or
text file.
32
Recall SigmaXL Dialog

Recall SigmaXL Dialog is used to activate the last data worksheet and recall the last dialog, making
it very easy to do repetitive analysis. To access, in Excel 2007 and newer, click the Recall SigmaXL
Dialog menu button as shown below:
Alternatively, you can use the Hot Key F3 (not available in Excel Mac 2011). This feature can also
be accessed by clicking SigmaXL > Help > Hot Keys > Recall SigmaXL Dialog.
Note that Recall SigmaXL Dialog may not be available for all functions.
Activate Last Worksheet

Activate Last Worksheet is used to activate the last data worksheet without recalling the dialog. To
access, press hot key F4 (not available in Excel Mac 2011). This feature can also be accessed by
clicking SigmaXL > Help > Hot Keys > Activate Last Sheet.
33
Worksheet Manager
The Worksheet Manager is a powerful utility that allows you to list all open Excel workbooks,
display all worksheets and chart sheets in a selected workbook and quickly select a worksheet or
chart sheet of interest. To access, click SigmaXL > Worksheet Manager.
34
Part B – Templates & Calculators
Introduction to Templates & Calculators

To use SigmaXL templates, select the appropriate template, enter the inputs and the resulting
outputs are produced immediately. If the template does not automatically perform the
calculations, click Tools > Options, select Calculation, Automatic, and click OK (Excel 2007: Office
Button | Excel Options | Formulas, select Automatic, click OK; Excel 2010: File | Options…).
Templates and Calculators are protected worksheets by default, but this may be modified by
clicking SigmaXL > Help > Unprotect Worksheet. Alternatively, you can click Excel Tools >
Protection > Unprotect Sheet (Excel 2007 and newer: Review | Changes | Unprotect Sheet). Gage
R&R is password protected, so it cannot be unprotected.
Click SigmaXL > Templates & Calculators to access the templates and calculators. Basic Graphical
Templates can also be accessed by clicking SigmaXL > Graphical Tools > Basic Graphical Templates.
Basic Statistical Templates can be accessed by clicking SigmaXL > Statistical Tools > Basic Statistical
Templates. Basic MSA Templates can be accessed by clicking SigmaXL > Measurement Systems
Analysis > Basic MSA Templates. Basic Process Capability Templates can be accessed by clicking
SigmaXL > Process Capability > Basic Process Capability Templates. Basic DOE Templates can also
be accessed by clicking SigmaXL > Design of Experiments > Basic DOE Templates. Basic Control
Chart Templates can also be accessed by clicking SigmaXL > Control Charts > Basic Control Chart
Templates.
35
SigmaXL Templates & Calculators

• DMAIC & DFSS Templates:
o Team/Project Charter
o SIPOC Diagram
o Data Measurement Plan
o Cause & Effect (Fishbone) Diagram and Quick Template
o Cause & Effect (XY) Matrix
o Failure Mode & Effects Analysis (FMEA)
o Quality Function Deployment (QFD)
o Pugh Concept Selection Matrix
o Control Plan
• Lean Templates:
o Takt Time Calculator
o Value Analysis/Process Load Balance
o Value Stream Mapping
• Basic Graphical Templates:

o Pareto Chart, Histogram and Run Chart
• Basic Statistical Templates:

o Sample Size – Discrete and Continuous
o Minimum Sample Size for Robust t-Tests and ANOVA
o1 Sample Z-Test and Confidence Interval for Mean
o1 Sample t-Test and Confidence Interval for Mean
o2 Sample t-Test and Confidence Interval (Compare 2 Means)
o1 Sample Equivalence Test for Mean
o2 Sample Equivalence Test (Compare 2 Means)
o1 Sample Chi-Square Test and CI for Standard Deviation
o2 Sample F-Test and CI (Compare 2 Standard Deviations)
36
o1 Proportion Test and Confidence Interval
o2 Proportions Test and Confidence Interval
o2 Proportions Equivalence Test
o1 Poisson Rate Test and Confidence Interval
o2 Poisson Rates Test and Confidence Interval
o2 Poisson Rates Equivalence Test
o One-Way Chi-Square Goodness-of-Fit Test
o One-Way Chi-Square Goodness-of-Fit Test Exact
• Probability Distribution Calculators:

o Normal, Inverse Normal, Lognormal, Exponential, Weibull
o Binomial, Poisson, Hypergeometric
• Basic MSA Templates:

o Type 1 Gage Study
o Gage Bias and Linearity Study
o Gage R&R Study – with Multi-Vari Analysis
o Attribute Gage R&R (Attribute Agreement Analysis)
• Basic Process Capability Templates:

o Process Sigma Level – Discrete and Continuous
o Process Capability & Confidence Intervals
o Tolerance Interval Calculator (Normal Exact)
• Basic DOE Templates:

o2 to 5 Factors
o Main Effects & Interaction Plots
• Basic Control Chart Templates:

o Individuals
o C-Chart
37
Team/Project Charter
Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Team/Project Charter to
access the Team/Project Charter template.
38
SIPOC Diagram
Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > SIPOC Diagram to access
the SIPOC Diagram template.
Data Measurement Plan

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Data Measurement Plan to
access the Data Measurement Plan template.
39
Cause & Effect (Fishbone) Diagram

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Cause & Effect (Fishbone)
Diagram to access the Cause & Effect Diagram form.
Notes:
1. Overwrite the text in the Cause, Sub-Cause and Problem Statement/Effect as appropriate.
2. Use copy and paste to create additional causes or sub-causes.
3. The arrows with text are grouped. To ungroup use Excel's Draw > Ungroup tool.
40
Cause & Effect (Fishbone) Template

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Cause & Effect (Fishbone)
Template to access the Cause & Effect Diagram template.
Simply fill in the template table and click the Fishbone Diagram button to create the diagram.
41
Cause & Effect (XY) Matrix Example

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Cause & Effect (XY) Matrix
to access the Cause & Effect (XY) Matrix template.
The following example is given in SigmaXL > Help > Template Examples > C&E Matrix. This is a
simple Cause and Effect Matrix example for a Call Center. If prompted, please ensure that macros
are enabled in order to allow Pareto Charts to be created.
Notes for use of the Cause and Effect Matrix template, also known as the XY Matrix:
1. Weight the Output Variables (Y's) on a scale of 1 to 10 with 10 indicating most important to the
Customer.
2. For Root Cause Analysis, assign the association/effect multiplier score for each X to Y using a
scale of 0, 1, 3, 9, where 0 = None, 1 = Weak, 3 = Moderate, and 9 = Strong. Initially this
assignment will likely be a team subjective assessment. Data should be collected and the
degree of association should be validated with Graphical and Statistical Tools.
3. For Project Selection or Solution Selection, assign the association multiplier score for each X to
Y using a scale of 1 to 10, with 10 indicating strong association.
Click the Pareto Chart

button to create a Pareto
Chart of Weighted Scores
by Input/Process
Variables:
42
Failure Mode & Effects Analysis (FMEA) Example

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Failure Mode & Effects
Analysis (FMEA) template.
The following example is given in SigmaXL > Help > Template Examples > FMEA. This is a simple
Failure Mode and Effects example for Stocking Inventory. If prompted, please ensure that macros
are enabled in order to allow sort by RPN.
If you hover the mouse cursor over the Severity, Occurrence or Detection heading the
recommended scale will appear as a comment:
43
The recommended scales for Severity, Occurrence, and Detection are shown below:
44
Click Risk Priority Number (RPN) Sort to sort the rows by RPN in descending order.
Click Revised Risk Priority Number Sort to sort the rows by Revised RPN in descending order.
Click Row Number Sort to restore the FMEA worksheet to the original row order (ascending).
45
Quality Function Deployment (QFD) Example

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Quality Function
Deployment (QFD) > 10 by 10 QFD (or 20 by 20, 30 by 30) to access the Quality Function
Deployment (QFD) template. Enter data as shown.
46
Pugh Concept Selection Matrix Example

Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Pugh Concept Selection
Matrix to access the Pugh Concept Selection Matrix template. Enter data as shown.
47
Control Plan
Click SigmaXL > Templates & Calculators > DMAIC & DFSS Templates > Control Plan to access the
Control Plan template.
48
Lean – Takt Time Calculator Example

Click SigmaXL > Templates & Calculators > Lean > Takt Time Calculator to access the Takt Time
calculator. The template gives the following default example:
49
Lean – Value Analysis/Process Load Balance

Click SigmaXL > Templates & Calculators > Lean > Value Analysis/Process Load Balance to access
the Value Analysis/Process Load Balance template. The template gives the following default
example:
50
Lean – Value Stream Mapping

Click SigmaXL > Templates & Calculators > Lean > Value Stream Mapping to access the Value
Stream Mapping template. The example on the following page is given in SigmaXL > Help >
Template Examples > Value Stream Mapping.
51
SigmaXL Inc., Copyright © 2011-2014
Portions courtesy and copyright of Strategos,

Inc. www.strategosinc.com
Notes:
1. Copy and Paste the above symbols into a

blank worksheet as needed to create the value
stream map. Resize if necessary.
2. Recommended resource for Value Stream

Mapping: Quarterman Lee and Brad Snyder,
"The Strategos Guide to Value Stream &
Process Mapping", Enna Products Corp., 2006.
Open the file Template & Calculator Examples.xlsx. Click on the worksheet named VSM Example.
This is a present state value stream map of a manufacturing operation. This example is courtesy
and copyright of Strategos, Inc. www.strategosinc.com.
52
Basic Graphical Templates – Pareto Example

Click SigmaXL > Templates & Calculators > Basic Graphical Templates > Pareto Chart. Enter data as
shown:
Click the Pareto Chart button to produce the Pareto Chart:
Pareto Chart
100 100%
90 90%
80 80%
70 70%
60 60%
Count
50 50%
40 40%
30 30%
20 20%
10 10%
0 0%
ls r ng r e
al de lo bl
-c or -lo co ila
rn t o- t oo g- va
e tu lt - s- r on t- a
R ff icu ke W No
Di r- ta
e
rd
O
Category
Notes:
1. This Pareto Chart template should be used with count data like number of errors. You can also
create Pareto charts with cost data.
2. You can replace the Category and Count column headings with any headings that you wish.
3. Enter the Pareto categories in the Category column. These can be Name, Location, Error Type
or other text information. Pareto categories are required and will appear on the horizontal X-
Axis of the Pareto Chart.
4. Enter your count (or cost) data in the Count column.
5. Click the Pareto Chart button to create a Pareto Chart.
53
Basic Graphical Templates – Histogram Example

Click SigmaXL > Templates & Calculators > Basic Graphical Templates > Histogram. Enter data as
shown:
Click the Histogram button to produce the Histogram:
54
Notes:
1. This Histogram template should be used with continuous data like cycle time.
2. You can replace the Data column heading with any heading that you wish.
3. Enter your data in the Data column.
4. Click the Histogram button to create a Histogram.
55
Basic Graphical Templates – Run Chart Example

Click SigmaXL > Templates & Calculators > Basic Graphical Templates > Run Chart. Enter the data
as shown (copy and paste the histogram data):
Click the Run Chart button to produce the Run Chart:
56
Run Chart
60
50
Mean: 37.95
40
Run Chart: Data
30
20
10
0
re
re
re
re
re
r
te
te
te
te
te
fo
fo
fo
fo
fo
Af
Af
Af
Af
Af
Be
Be
Be
Be
Be
Notes:
1. This Run Chart template can be used with a variety of data types: continuous data like cycle
time, count data like number of errors, or cost data. The data must be in chronological time-
sequence order.
2. You can replace the X-Axis Label and Data column headings with any headings that you wish.
4. Enter labels in X-Axis Label column. Labels can be Date, Time, Name, or other text information.
These labels are optional and will appear on the horizontal X-Axis of the Run Chart.
5. Click the Run Chart button to create a Run Chart.
57
Basic Statistical Templates – Sample Size – Discrete

Click SigmaXL > Templates & Calculators > Basic Statistical Templates > Sample Size – Discrete to
access the Sample Size – Discrete calculator. The template gives the following default example:
Notes for use of the Sample Size – Discrete Calculator:
1. P is estimate of proportion for outcome of interest. Use P = 0.5 if unknown.
2. Delta is desired proportion margin of error. Enter as the half-width, i.e. if the desired margin of
error is +/- 3%, enter 0.03.
3. Enter population size N to adjust for small populations (N < 10000).
4. np should be >= 5. If necessary, reduce delta to adjust.
5. Power (1 – Beta) is not considered in these calculations. Power and Sample Size may be
calculated using SigmaXL > Statistical Tools > Power & Sample Size Calculators.
This sample size calculator is based on a confidence interval approach, where you enter the desired
half-interval value (delta). Let’s say that you desire the proportion margin of error to be +/- 3%. As
noted above, you simply enter a delta (half-interval) value of .03.
The challenge with this sample size calculator is that we need an estimate of the population
proportion. This is a bit of a chicken and egg situation – which comes first? I want to calculate an
appropriate sample size to determine an outcome, but I am being asked to enter an estimate of the
population proportion. First, keep in mind that this tool is a planning tool – the true confidence
intervals will be determined after you collect your data. Estimating the population proportion can
be done if you have good historical data to draw from, for example, in historical customer surveys
the percentage of satisfied customers was 80%. In this case, you could use P=0.8. If you do not have
a priori knowledge, then leave P=0.5 which gives the most conservative value (i.e., largest estimate
58
of sample size). If you enter a value other than 0.5 it will result in a smaller sample size
requirement.
59
Basic Statistical Templates – Sample Size – Continuous

Click SigmaXL > Templates & Calculators > Basic Statistical Templates > Sample Size – Continuous
to access the Sample Size – Continuous calculator. The template gives the following default
example:
Notes for use of the Sample Size – Continuous Calculator:
1. Delta uses the same units as the standard deviation. Enter as the half-width, i.e., if the desired
margin of error is +/- 0.25, enter 0.25.
2. Enter (optional) population size N to adjust for small populations (N < 1000).
3. Power (1 – Beta) is not considered in these calculations. Power and Sample Size may be
calculated using SigmaXL > Statistical Tools > Power & Sample Size Calculators.
Similar to the Sample Size – Discrete calculator, this sample size calculator is based on a confidence
interval approach, where you enter the desired half-interval value (delta). If a survey had responses
that were on a continuous scale of 1 to 5, and you desired a margin of error on the mean to be +/-
.25, then you would use .25 as the delta value. Note that in this continuous case the units are not
percentages but level of satisfaction.
Here we need an a priori estimate of the population standard deviation. If you have no idea what
the standard deviation will be, then you could take a small sample to get a rough estimate of the
standard deviation.
60
Basic Statistical Templates – Minimum Sample Size for Robust t-

Tests and ANOVA
It is well known that the central limit theorem enables the t-Test and ANOVA to be fairly robust to
the assumption of normality. A question that invariably arises is, “How large does the sample size
have to be?” A popular rule of thumb answer for the one sample t-Test is “n = 30.” While this rule
of thumb often does work well, the sample size may be too large or too small depending on the
degree of nonnormality as measured by the Skewness and Kurtosis. Furthermore it is not
applicable to a One Sided t-Test, 2 Sample t-Test or One Way ANOVA.
To address this issue, we have developed a unique template that gives a minimum sample size
needed for a hypothesis test to be robust.
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > Minimum Sample Size for
Robust t-Tests and ANOVA to access this template. It includes minimum sample size for
robustness for the 1 Sample t-Test, 2 Sample t-Test and the One Way ANOVA.
The user may specify the alternative hypothesis as “Less Than” (one sided), “Not Equal To” (two
sided) or “Greater Than” (one sided). Confidence levels of 90% (α = 0.1), 95% (α = .05) or 99% (α =
.01) may also be specified:
To use the template, simply select the appropriate Hypothesis Test, Alternative Hypothesis and
Confidence Level using the drop down selection. Enter Skewness and Kurtosis values as shown in
the yellow highlighted cells.
Note that in the example shown, the rule of thumb for a 1 Sample t-Test “n = 30” is confirmed with
a moderate skew value of 1.
61
Now change the Alternative Hypothesis to “Less Than” and Confidence Level to 99% as shown:
The minimum sample size required for robustness is now 752!
On the other hand, if you want to perform a standard One Way ANOVA, enter the values as shown:
Now the minimum sample size requirement is only 3. This value applies to each sample or group,
so for the 3 Sample ANOVA that would mean each sample has n = 3 for a total number of
observations = 9.
Note that this calculator is strictly addressing the question of alpha robustness to nonnormality.
Power is not considered here.
If the minimum sample size requirements cannot be met, you should use a nonparametric
equivalent to the parametric hypothesis test (i.e. One Sample Sign or Wilcoxon, Two Sample Mann-
Whitney, Kruskal-Wallis or Mood’s Median: SigmaXL > Statistical Tools > Nonparametric Tests).
62
Skewness and Kurtosis

Sample Skewness and Kurtosis values can be obtained from SigmaXL’s descriptive statistics:
SigmaXL > Statistical Tools > Descriptive Statistics.
A slight Skew is +/- 0.5, moderate Skew is +/- 1, severe Skew is +/- 2 and extreme Skew is +/- 5.The
Skewness range used should be -5 to +5. Values beyond this range are extrapolated so may be
inaccurate.
Kurtosis for a normal distribution is 0. Kurtosis must be greater than or equal to:
(Skew ^ 2 - 1.48).
Kurtosis “delta” is Kurt - Skew2 and is used in the regression equation. Kurtosis delta range should
be -1.48 to +1.48. A kurtosis delta less than -1.48 denotes a bimodal distribution so this is a lower
boundary. Values above +1.48 are extrapolated so may be inaccurate.
The calculator assumes that all samples have the same Skewness and Kurtosis.
Monte Carlo Simulation

The data for minimum sample size formulas are derived from extensive Monte Carlo simulations
for n = 2 to 2000. Observed alpha values were determined empirically from the P-Values of
100,000 replicate hypothesis tests for each n. Nonnormal data with Skew = -5 to +5 and Kurt delta
= -1.48 to +1.48 was generated using the Pearson Family function (see SigmaXL DiscoverSim
Workbook: Appendix for details).
Minimum sample size for robustness occurs when the simulated observed alpha is within +/- 20%
of the specified alpha (Bradley 1980 and Rhiel 1996).
Regression Models
A separate regression model was constructed for each hypothesis test, alternative hypothesis, and
confidence level (total of 21 models), using coding as shown.
The model predictor terms are Skewness^2 and Kurtosis Delta. The response is n minimum. The
term coefficient values are stored in the template as shown.
63
Model R-Square values are typically over 99%, with some exceptions (96%) due to small estimated
sample sizes.
The model used is the one which matches the selected hypothesis test, alternative hypothesis, and
confidence level. The selected n is highlighted in red and is also displayed in the template Results
cell. If n > 2000, a text display “> 2000” is shown in the Results cell.
References
Boos, D.D. and Hughes-Oliver, J.M. (2000), “How Large Does n Have to be for Z and t Intervals?”
The American Statistician, 54(2), 121-128.
Bradley, J. V. (1980), “Nonrobustness in Z, t, and F Tests at Large Sample Sizes,” Bulletin of the
Psychonomics Society, 16(5), 333-336.
Rhiel, G. S., and Chaffin, W. W. (1996), “An Investigation of the Large-Sample/Small-Sample

Approach to the One-Sample Test for a Mean (Sigma Unknown),” Journal of Statistics Education, 4,
No. 3 (www.amstat.org/publications/jse).
SigmaXL, Inc., DiscoverSim Version 1.1 Workbook.
64
Basic Statistical Templates – 1 Sample Z-Test and Confidence

Interval for Mean
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Sample Z-Test and
Confidence Interval for Mean to access the 1 Sample Z-Test calculator. The template gives the
following default example.
Notes:
1. Enter summarized Sample Data, Known Population Standard Deviation, Null Hypothesis and
Confidence Level in cells with yellow highlight. Do not modify any other part of this worksheet.
2. Select Alternative Hypothesis using drop-down.
3. This calculator will automatically detect the version of Excel and use the appropriate statistical
functions. If Excel 2010 or higher, then the Excel 2010 statistical functions are used.
65
Basic Statistical Templates – 1 Sample t-Test and Confidence

Interval for Mean
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Sample t-Test and
Confidence Interval for Mean to access the 1 Sample t-Test calculator. The template gives the
Notes:
1. Enter summarized Sample Data, Null Hypothesis and Confidence Level in cells with yellow
highlight. Do not modify any other part of this worksheet.
66
Basic Statistical Templates – 2 Sample t-Test and Confidence

Interval (Compare 2 Means)
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Sample t-Test &
Confidence Interval (Compare 2 Means) to access the 2 Sample t-Test calculator. The template
gives the following default example.
Notes:
2. Select Alternative Hypothesis and "Assume Equal Variances" using drop-down.
67
Basic Statistical Templates – 1 Sample Equivalence Test for Mean

Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Sample Equivalence Test
for Mean to access the 1 Sample Equivalence Test (Two One-Sided t-Tests TOST) calculator. The
template gives the following default example.
Notes:
1. Enter summarized Sample Data, Target, Upper & Lower Equivalence Limits and Confidence
Level in cells with yellow highlight. Do not modify any other part of this worksheet.
2. Null hypothesis for P1: Difference >= UEL; Null hypothesis for P2: Difference <= LEL. Both null
hypothesis must be rejected to conclude that equivalence is true.
3. LEL and UEL establish the zone or region of equivalence and are determined by what size
difference is considered practically significant.
68
Basic Statistical Templates – 2 Sample Equivalence Test (Compare

2 Means)
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Sample Equivalence Test
(Compare 2 Means) to access the 2 Sample Equivalence Test (Two One-Sided t-Tests TOST)
calculator. The template gives the following default example.
Notes:
1. Enter summarized Sample Data, Upper & Lower Equivalence Limits and Confidence Level in
cells with yellow highlight. Do not modify any other part of this worksheet.
2. Select "Assume Equal Variances" using drop-down.
3. Null hypothesis for P1: Mean Difference >= UEL; Null hypothesis for P2: Mean Difference <= LEL.
Both null hypothesis must be rejected to conclude that equivalence is true.
4. LEL and UEL establish the zone or region of equivalence and are determined by what size mean
69
Basic Statistical Templates – 1 Sample Chi-Square Test and CI for

Standard Deviation
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Sample Chi-Square Test
and CI for Standard Deviation to access the 1 Sample Chi-Square-Test for Standard Deviation
Notes:
70
Basic Statistical Templates – 2 Sample F-Test and CI (Compare 2

Standard Deviations)
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Sample F-Test and CI
(Compare 2 Standard Deviations) to access the 2 Sample F-Test calculator. The template gives the
Notes:
4. The F-test for two sample standard deviations/variances assumes that samples are normally
distributed.
71
Basic Statistical Templates – 1 Proportion Test and Confidence

Interval
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Proportion Test and
Confidence Interval to access the 1 Proportion Test calculator. The template gives the following
default example.
Notes:
2. Select Alternative Hypothesis and Methods using drop-down. If x = 0, it is recommended that

the Alternative Hypothesis be set to one-sided "Less Than." If x = n, set the Alternative
Hypothesis to one-sided "Greater Than."
72
3. Hypothesis Test Method:
• Binomial Exact is recommended.
• Use Normal Approximation only when n is large and proportion is close to 0.5 (check that
np >= 5 and nq >= 5, or the more conservative npq, shown above, >= 5). See Rosner (2010).
4. Confidence Interval Method (recommendations based on literature review, see references

below):
• Confidence intervals for binomial proportions have an "oscillation" phenomenon where the
coverage probability varies with n and p.
• Exact (Clopper-Pearson) is strictly conservative and will guarantee the specified confidence
level as a minimum coverage probability, but results in wide intervals. Recommended only
for applications requiring strictly conservative intervals.
• Wilson Score and Jeffreys Beta have mean coverage probability matching the specified
confidence interval. We recommend Wilson for two-sided confidence intervals and Jeffreys
for one-sided.
• Normal Approximation (Wald) is not recommended, but included here to validate hand
calculations. Use only when n is large and proportion is close to 0.5 (check that npq, shown
above, >= 5).
REFERENCES
1. Agresti, A. and Coull, B.A. (1998), Approximate is Better than “Exact” for Interval Estimation of
Binomial Proportions. The American Statistician, 52,119–126.
2. Brown, L. D., Cai, T. T. and DasGupta, A. (2001), Interval estimation for a binomial proportion.
Statistical Science, 16, 101-133. With comments and a rejoinder by the authors.
3. Cai, T. T. (2005), One-sided conﬁdence intervals in discrete distributions. J. Statist. Plann. Inference,
131, 63-88.
4. Clopper, C.J., and Pearson, E.S.(1934), The Use of Confidence or Fiducial Limits Illustrated in the
Case of the Binomial. Biometrika, 26, 404–413.
5. Newcombe, R. (1998a), Two-sided confidence intervals for the single proportion: Comparison of
seven methods, Statistics in Medicine, 17, 857-872.
6. Rosner, B. (2010), Fundamentals of Biostatistics. Seventh Edition. Equation 5.15, page 133.
Duxbury Press.
7. Vollset, S.E. (1993). Confidence intervals for a binomial proportion, Statistics in Medicine, 12, 809–
824.
73
Basic Statistical Templates – 2 Proportions Test and Confidence

Interval
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Proportions Test and
Confidence Interval to access the 2 Proportions Test calculator. The template gives the following
default example.
Notes:
2. Select Alternative Hypothesis and Methods using drop-down.
74
• Fisher's Exact is recommended, but is only valid for H0: P1 - P2 = 0.
• Use Normal Approximation only when minimum expected cell value, shown above, is >= 5.
Unpooled is used if H0: P1 - P2 <> 0. Use pooled if H0: P1 - P2 = 0. (See Peltier).
4. Confidence Interval Method:
• Confidence intervals for difference in binomial proportions have an "oscillation"

phenomenon where the coverage probability varies with n and p.
• Newcombe-Wilson Score is recommended and has a mean coverage probability that is close
to the specified confidence interval. (See Newcombe method 10).
• Newcombe-Wilson Score (CC = Continuity Corrected) is conservative and will typically meet
the specified confidence level as a minimum coverage probability, but results in wider
intervals. (See Newcombe method 11).
• Jeffreys-Perks is similar to Newcombe-Wilson, and included as an option because it is the

preferred interval for some practitioners. (See Radhakrishna; Newcombe method 4).
• Normal Approximation is not recommended, but included here to validate hand

calculations. Use only when minimum expected cell value, shown above, is >= 5.
• Exact methods that are strictly conservative (like Clopper-Pearson for the one proportion
case) are not included in this template because they are computationally intensive and
slow.
5. Due to the complexity of calculations, this template uses vba macros rather than Excel
formulas. SigmaXL must be initialized and appear on the menu in order for this template to
function.
REFERENCES
1. Beal, S. L. (1987), Asymptotic confidence intervals for the difference between two binomial
parameters for use with small samples. Biometrics, 43, 941-950.
2. Newcombe, R. G. (1998b), Interval estimation for the difference between independent proportions:
Comparison of eleven methods. Statistics in Medicine, 17:873–890.
3. Peltier, C., Why Do We Pool for the Two-Proportion z-Test?

http://apcentral.collegeboard.com/apc/members/courses/teachers_corner/49013.html.
4. Radhakrishna S., Murthy B.N., Nair N.G., Jayabal P., Jayasri R. (1992), Confidence intervals in
medical research. Indian J Med Res., Jun;96:199-205.
75
Basic Statistical Templates – 2 Proportions Equivalence Test

Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Proportions Equivalence
Test to access the 2 Proportions Equivalence Test (Two One-Sided Tests TOST) calculator. The
Notes:
2. Select Confidence Interval Method using drop-down.
• Confidence intervals for difference in binomial proportions have an "oscillation" phenomenon

where the coverage probability varies with n and p.
• Newcombe-Wilson Score is recommended and has a mean coverage probability that is close to the
specified confidence interval. (See Newcombe method 10).
• Newcombe-Wilson Score (CC = Continuity Corrected) is conservative and will typically meet the
specified confidence level as a minimum coverage probability, but results in wider intervals. (See
Newcombe method 11).
• Jeffreys-Perks is similar to Newcombe-Wilson, and included as an option because it is the preferred
interval for some practitioners. (See Radhakrishna; Newcombe method 4).
• Normal Approximation will match the P-Value calculations, but should only be used when minimum
expected cell value, shown above, is >= 5.
76
• Exact methods that are strictly conservative (like Clopper-Pearson for the one proportion case) are
not included in this template because they are computationally intensive and slow.
3. Null hypothesis for P1: Proportion Difference >= UEL; Null hypothesis for P2: Proportion
Difference <= LEL. Both null hypothesis must be rejected to conclude that equivalence is true.
The P-Values are based on the normal approximation unpooled method.
4. LEL and UEL establish the zone or region of equivalence and are determined by what size
proportion difference is considered practically significant.
5. Since the confidence interval options use different methods than the hypothesis test (except
for Normal option), it is possible that the conclusion from the Equivalence P-Value will be
different from that of the confidence interval. In this case, we recommend using just the
confidence interval method.
6. Due to the complexity of calculations, this template uses VBA macros rather than Excel
function.
REFERENCES
1. Beal, S. L. (1987), Asymptotic confidence intervals for the difference between two binomial parameters for
use with small samples. Biometrics, 43, 941-950.
2. Newcombe, R. G. (1998b), Interval estimation for the difference between independent proportions:
Comparison of eleven methods. Statistics in Medicine, 17:873–890.
3. Radhakrishna S., Murthy B.N., Nair N.G., Jayabal P., Jayasri R. (1992), Confidence intervals in medical
research. Indian J Med Res., Jun;96:199-205.
77
Basic Statistical Templates – 1 Poisson Rate Test and Confidence

Interval
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 1 Poisson Rate Test and
Confidence Interval to access the 1 Poisson Rate calculator. The template gives the following
default example.
Notes:
2. Sample Size (area of opportunity) can be a count, time, length or other unit that defines the
area of opportunity for an occurrence.
3. Select Alternative Hypothesis and Methods using drop-down. If x = 0, it is recommended that

the Alternative Hypothesis be set to one-sided "Less Than."
78
• Poisson Exact is recommended.
• Use Normal Approximation only when the number of occurrences, x > 10.
5. Confidence Interval Method (recommendations based on literature review, see references

below):
• Confidence intervals for Poisson rates have an "oscillation" phenomenon where the
coverage probability varies with x and n.
• Exact (Garwood Chi-Square) is strictly conservative and will guarantee the specified
confidence level as a minimum coverage probability, but results in wide intervals.
Recommended only for applications requiring strictly conservative intervals.
• Jeffreys has mean coverage probability matching the specified confidence interval and is
recommended.
• Normal Approximation (Wald) is not recommended, but included here to validate hand
calculations. Use only when the number of occurrences, x > 10.
REFERENCES
1. Garwood, F. (1936). Fiducial limits for the Poisson distribution. Biometrika 28:437–442.
2. Swift, M.B. (2009). Comparison of confidence intervals for a Poisson Mean-Further considerations,
Communication in Statistics — Theory and Methods, 38, 748–759.
79
Basic Statistical Templates – 2 Poisson Rates Test and Confidence

Interval
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Poisson Rates Test and
Confidence Interval to access the 2 Poisson Rates calculator. The template gives the following
default example.
Notes:
3. Select Alternative Hypothesis and Methods using drop-down.
80
• Exact (Binomial Method) is recommended, but is only valid for H0: λ1 - λ2 = 0.
• Use Normal Approximation only when the number of occurrences for each sample is
greater than 10. Unpooled is used if H0: λ1 - λ2 <> 0. Use pooled if H0: λ1 - λ2 = 0.
5. Confidence Interval Method:
• Confidence intervals for difference in Poisson rates have an "oscillation" phenomenon

where the coverage probability varies with x and n.
• Jeffreys Hybrid is recommended and has a mean coverage probability that is close to the
specified confidence interval. (See Li, 2011).
• Normal Approximation is not recommended, but included here to validate hand

calculations. Use only when the number of occurrences for each sample is greater than 10.
function.
REFERENCES
1. Li, H.Q., Tang, M.L., Poon, W.Y. and Tang, N.S. (2011), Confidence intervals for difference between
two Poisson rates, Comm. Stat. Simul. Comput. 40, pp. 1478–1491.
2. Przyborowski, J., Wilenski, H., (1940), Homogeneity of results in testing samples from Poisson
series. Biometrika 31, 313–323.
81
Basic Statistical Templates – 2 Poisson Rates Equivalence Test

Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Poisson Rates
Equivalence Test to access the 2 Poisson Rates Equivalence Test (Two One-Sided Tests TOST)
Notes:
3. Select Confidence Interval Method using drop-down.
• Confidence intervals for difference in Poisson rates have an "oscillation" phenomenon where the
coverage probability varies with x and n.
• Jeffreys Hybrid is recommended and has a mean coverage probability that is close to the specified
confidence interval. (See Li, 2011).
• Normal Approximation will match the P-Value calculations, but should only be used when the
number of occurrences for each sample is greater than 10.
82
4. Null hypothesis for P1: Rate Difference >= UEL; Null hypothesis for P2: Rate Difference <= LEL.
Both null hypothesis must be rejected to conclude that equivalence is true. The P-Values are
based on the normal approximation unpooled method.
5. LEL and UEL establish the zone or region of equivalence and are determined by what size rate
6. Since the Jeffreys Hybrid confidence interval option uses a different method than the normal
based hypothesis test, it is possible that the conclusion from the Equivalence P-Value will be
different from that of the confidence interval. In this case, we recommend using just the
confidence interval method.
7. Due to the complexity of calculations, this template uses VBA macros rather than Excel
function.
REFERENCES
1. Li, H.Q., Tang, M.L., Poon, W.Y. and Tang, N.S. (2011), Confidence intervals for difference between two
Poisson rates, Comm. Stat. Simul. Comput. 40, pp. 1478–1491.
2. Przyborowski, J., Wilenski, H., (1940), Homogeneity of results in testing samples from Poisson series.
Biometrika 31, 313–323.
83
Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit

Test
Click SigmaXL > Templates & Calculators > Basic Statistical Templates > One-Way Chi-Square
Goodness-of-Fit Test to access the One Way Chi-Square Goodness-of-Fit Test calculator. The
Notes:
1. Enter Category ID, Observed Count values and (optional) Historical Counts. Do not modify any
other part of this worksheet.
2. If optional Historical Counts are not specified, chi-square is calculated using equal expected
proportions.
3. If optional Historical Counts are specified, a value must be entered for each observed count.
4. Chi-Square statistic requires that no more than 20% of cells have an expected count less than 5.
84
Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit

Test - Exact
Goodness-of-Fit Test - Exact to access the One Way Chi-Square Goodness-of-Fit Test – Exact
Notes:
1. Enter Category ID, Observed Count values and (optional) Historical Counts. Do not modify any
other part of this worksheet.
2. If optional Historical Counts are not specified, chi-square is calculated using equal expected
proportions.
3. If optional Historical Counts are specified, a value must be entered for each observed count.
4. The Chi-Square statistic requires that no more than 20% of cells have an expected count less
than 5 (and none of the cells have an expected count less than 1). If this assumption is not
satisfied, the Chi-Square approximation may be invalid and Exact or Monte Carlo P-Values
should be used. In this example 40% of the cells have an expected count less than 5 so Exact
should be used.
5. This example shows that the Exact P-Value is 0.5013 and the “large sample” Chi-Square P-Value
is 0.4897.
6. Chi-Square Exact solves the permutation problem using enhanced enumeration. For further
details refer to the Appendix Exact and Monte Carlo P-Values for Nonparametric and
Contingency Test.
7. It is important to note that while exact P-Values are “correct,” they do not increase (or
decrease) the power of a small sample test, so they are not a solution to the problem of failure
to reject the null due to inadequate sample size.
8. For data that requires more computation time than specified, Monte Carlo P-Values provide an
approximate (but unbiased) P-Value that typically matches exact to two decimal places using
10,000 replications. One million replications give a P-Value that is typically accurate to three
decimal places. A confidence interval (99% default) is given for the Monte Carlo P-Values.
85
Monte Carlo Example:
1. The Exact P-Value for this example is solved very quickly, so Monte Carlo is not needed, but we
will run it for continuity in the example. Click Monte Carlo P-Value. Select Number of
Replications = 10000 and Confidence Level for P-Value = 99%.
Tip: The Monte Carlo 99% confidence interval for P-Value is not the same as a confidence
interval on the test statistic due to data sampling error. The confidence level for the hypothesis
test statistic is still 95%, so all reported P-Values less than .05 will be highlighted in red to
indicate significance. The 99% Monte Carlo P-Value confidence interval is due to the
uncertainty in Monte Carlo sampling, and it becomes smaller as the number of replications
increases (irrespective of the data sample size). The Exact P-Value will lie within the stated
Monte Carlo confidence interval 99% of the time.
2. Click OK. Results:
The Monte Carlo P-Value here is 0.5012 with a 99% confidence interval of 0.4896 to 0.5128.
This will be slightly different every time it is run (the Monte Carlo seed value is derived from the
system clock). The true Exact P-Value = 0.5013 lies within this confidence interval.
Small Sample Exact Example:
3. Now we will consider a small sample problem. Enter the following values for sample data in the
yellow highlight region. Note that the displayed Monte Carlo (or Exact) P-Values are cleared
when new data is entered in the template:
86
This example is adapted from Mehta, C.R. and Patel, N. R., IBM SPSS Exact Tests, IBM Corp.,
page 44.
ftp://public.dhe.ibm.com/software/analytics/spss/documentation/statistics/21.0/en/client/Ma
nuals/IBM_SPSS_Exact_Tests.pdf
4. Click Exact P-Value. Select Time Limit for Exact Computation = 60 seconds.
Note that the Exact P-Value is 0.0523 which is a “fail-to-reject” of the null hypothesis (H0), but
the “large sample” or “asymptotic” Chi-Square P-Value incorrectly rejected H0 with a P-Value of
0.046. The exact P-Value matches that given in the reference. The error can also go the other
way, where a large sample Chi-Square P-Value is a “fail-to-reject” of the null hypothesis and the
Exact P-Value is a rejection of H0.
In conclusion, always use the Exact (or Monte Carlo) P-Value when the Chi-Square large sample
assumptions are not met.
87
Probability Distribution Calculators – Normal

Click SigmaXL > Templates & Calculators > Probability Distribution Calculators > Normal to access
the Normal Distribution Probability calculator. The template gives the following default example:
88
Basic MSA Templates – Type 1 Gage Study

Click SigmaXL > Templates & Calculators > Basic MSA Templates > Gage Bias and Linearity Study
to access the Gage Bias and Linearity Study template.
The following example is given in SigmaXL > Help > Template Examples > MSA > Type 1 Gage
Study.
89
Notes:
1. This calculator assumes that the data are normally distributed.
2. Enter your data in the Measurement Data column. Minimum recommended sample size is 25,
but 50 is preferred. This template has a maximum sample size of 1000, suitable for automated
test systems.
3. Enter Tolerance (Upper Specification Limit USL - Lower Specification Limit LSL).
4. Enter Part Reference Value. If missing, Gage Capability Cgk is not computed. Bias will assume a
0 value.
5. Enter Gage Resolution. If missing, Resolution % of Tolerance is not computed.
6. The default StDev Multiplier is 6 (99.73% coverage). Change this to 5.15 if AIAG convention is
being used (99% coverage). Dietrich [1] recommends using 4 for approximate 95% coverage.
We recommend 6 for typical and critical measurement systems, and 4 for non-critical.
7. The default Percent of Tolerance for Cg/Cgk is 20. Stacey [2] suggests using 10 for critical
measurement systems.
8. If Cg and Cgk are >= 1.33, the measurement device is capable. If Cg or Cgk are < 1.33, the
measurement device is not capable and needs to be improved.
9. If Resolution % of Tolerance (RE%) is <= 5%, the resolution is acceptable. If RE% is > 5% the
resolution is not acceptable.
10. Click the Type 1 Gage Run Chart button to create a Run Chart with center line = Ref, upper limit
= Ref + 0.1*Tol and lower limit = Ref - 0.1*Tol (assuming 20% of Tolerance).
11. This calculator requires Excel 2010 or higher.
REFERENCES
1. Dietrich, E. (2002), "Measurement System Capability" Reference Manual, http://www.q-

das.com/uploads/tx_sbdownloader/Leitfaden_v21_me.pdf.
2. Stacey, J., Mathematics of Measurement Systems Analysis, Rolls-Royce plc.,

http://www.mei.org.uk/files/pdf/Mathematics_of_Measurement_Systems_Analysis.pdf
90
Basic MSA Templates – Gage Bias and Linearity Study

Click SigmaXL > Templates & Calculators > Basic MSA Templates > Gage Bias and Linearity Study
to access the Gage Bias and Linearity Study template.
The following example is given in SigmaXL > Help > Template Examples > MSA > Gage Bias and
Linearity Study.
91
Notes:
1. Enter Process Variation (SV from Gage R&R or 6 * Historical StDev). If neither are available, use
Process Tolerance. % Average Bias will be blank if not specified. Note that % Linearity does not
require Process Variation.
2. Enter part Reference Values. These should be relative to a traceable standard, but if not
available, see AIAG MSA Reference Manual [1], pp. 92-93 for suggestions on how to obtain
reference values. Reference values should cover the operating range of the gage.
3. Enter part measurement data in the yellow highlight region below the respective
part/reference value. AIAG recommends a minimum of 10 trials per part/reference. It is
important that the parts be selected at random for each trial in order to minimize appraiser
"recall" bias.
4. Click Gage Linearity Report button to generate the Linearity Report.
5. If a P-Value is < .05, it is highlighted in bold red to indicate significance. If the Average Bias P-
Value or Linearity Slope P-Value are < .05, the respective bar in Percent of Process Variation will
also be red to highlight significance.
6. Statistical significance is the best indicator of a problem with bias or linearity. Some
practitioners also use a rule of thumb that a good measurement system will have % Average
Bias < 5% and % Linearity < 5%.
7. The dashed lines on the Linearity scatter plot are the regression fit 95% confidence intervals. If
the zero reference line is not within the intervals, the linearity slope is significant, indicating
that the measurement system linearity is not acceptable.
8. This calculator requires Excel 2010 or higher. Linearity regression uses vba macros rather than
Excel formulas. SigmaXL must be initialized and appear on the menu in order for this template
to function.
9. Rows 34-39 and 42 are hidden to simplify the output report. These may be unhidden to view
how the calculations are performed.
REFERENCE
1. Automotive Industry Action Group AIAG (2010). Measurement Systems Analysis MSA Reference
Manual, 4th Edition
92
Basic MSA Templates – Gage R&R Study (MSA)

Click SigmaXL > Templates & Calculators > Basic MSA Templates > Gage R&R Study (MSA) to
access the Gage R&R Study template.
The following example is given in SigmaXL > Help > Template Examples > MSA > Gage R&R. If
prompted, please ensure that macros are enabled.
Click Create Stacked Column Format

for “Analyze Gage R&R” >> if you wish
to analyze the above data using
SigmaXL’s menu Gage R&R Analysis
t l
Notes for use of the Gage R&R Template:
1. The Automotive Industry Action Group (AIAG) recommended study includes 10 Parts, 3
Operators and 3 Replicates. The template calculations will work with a minimum of 2
Operators, 2 Parts and 2 Replicates. The data should be balanced with each operator
measuring the same number of parts and the same number of replicates. Use SigmaXL >
Measurement Systems Analysis to specify up to 30 Parts, 10 Operators and 10 Replicates.
2. Enter process Upper Specification Limit (USL) and Lower Specification Limit (LSL) in the Process
Tolerance window. This is used to determine the % Tolerance metrics. If the specification is
single-sided, leave both entries blank.
93
3. The default StDev multiplier is 6. Change this to 5.15 if AIAG convention is being used.
4. The cells shaded in light blue highlight the critical metrics Gage R&R % Total Variation (also
known as %R&R) and %Tolerance: < 10% indicates a good measurement system;
> 30% indicates an unacceptable measurement system.
94
Basic MSA Templates - Gage R&R: Multi-Vari & X-bar R Charts

Click SigmaXL > Templates & Calculators > Basic MSA Templates > Gage R&R: Multi-Vari & X-bar R
Charts to access the Multi-Vari & X-bar R Charts template. An example is given in SigmaXL > Help >
Template Examples > MSA > Gage R&R - Multi-Vari and Gage R&R – X-Bar R:
Gage R&R Multi-Vari Gage R&R Multi-Vari Gage R&R Multi-Vari Gage R&R Multi-Vari Gage R&R Multi-Vari
1.5685 1.5685 1.5685 1.5685 1.5685
1.5185 1.5185 1.5185 1.5185 1.5185
1.4685 1.4685 1.4685 1.4685 1.4685
1.4185 1.4185 1.4185 1.4185 1.4185

Reading
1.3685 1.3685 1.3685 1.3685 1.3685
1.3185 1.3185 1.3185 1.3185 1.3185
1.2685 1.2685 1.2685 1.2685 1.2685
1.2185 1.2185 1.2185 1.2185 1.2185
1.1685 1.1685 1.1685 1.1685 1.1685

Operator A Operator B Operator C Operator A Operator B Operator C Operator A Operator B Operator C Operator A Operator B Operator C Operator A Operator B Operator C
Operator - Part 01 Operator - Part 02 Operator - Part 03 Operator - Part: Part 04 Operator - Part: Part 05
Gage R&R - X-Bar by Operator
1.6890
1.5890
X-Bar - Reading1 - Reading 3
1.4890 1.46
1.42
1.3890 1.38
1.2890
1.1890
1.0890
-1
-2
-3
-4
-5
-6
-7
-8
-9
0
0
-1
-2
-3
-4
-5
-6
-7
-8
-9
-1
-2
-3
-4
-5
-6
-7
-8
-9
-1
-1
-1
C
A
C
A
Gage R&R - R-Chart by Operator
0.06970
0.07
0.06
R - Reading1 - Reading3
0.05
0.04
0.03
0.02133
0.02
0.01
0.00000
0
-1
-2
-3
-4
-5
-6
-7
-8
-9
0
0
-1
-2
-3
-4
-5
-6
-7
-8
-9
-1
-2
-3
-4
-5
-6
-7
-8
-9
-1
-1
-1
C
A
C
A
Notes for use of the Gage R&R: Multi-Vari & X-bar R Charts:
1. The Gage R&R Multi-Vari and X-bar & R charts can only be generated if a Gage R&R template
has been completed and is the active worksheet. Select the Gage R&R worksheet and click
SigmaXL > Templates & Calculators > Gage R&R: Multi-Vari & X-Bar R Charts to create the
above charts.
2. The Multi-Vari chart shows each Part as a separate graph. Each Operator’s response readings
are denoted as a vertical line with the top tick corresponding to the Maximum value, bottom
tick is the Minimum, and the middle tick is the Mean. The horizontal line across each graph is
the overall average for each part.
3. When interpreting the X-bar and R chart for a Gage R&R study, it is desirable that the X-bar
chart be out-of-control, and the Range chart be in-control. The control limits are derived from
within Operator repeatability.
95
Basic MSA Templates – Attribute Gage R&R (MSA)

Click SigmaXL > Templates & Calculators > Basic MSA Templates > Attribute MSA to access the
Attribute MSA Study template. An example is given in SigmaXL > Help > Template Examples >
Attribute MSA.
Click Create Stacked Column Format to Analyze with “Attribute

MSA (Binary)” >> if you wish to analyze the above data using
SigmaXL’s menu Attribute MSA Analysis tool.
Notes for use of the Attribute Gage R&R (MSA) Template:
1. Attribute Gage R&R is also known as Attribute Agreement Analysis.
96
2. Recommend for study: 3 Appraisers, 2 to 3 Replicates, Minimum of 10 Good Parts and 10 Bad
Parts. The data should be balanced with each appraiser evaluating the same number of parts
and the same number of replicates.
3. Specify the Good Part or Unit as G or other appropriate text (P, Y, etc.). Specify the Bad Part or
Unit as NG or other appropriate text (F, N, etc.). Be careful to avoid typing or spelling errors
when entering the results. A space accidentally inserted after a character will be treated as a
different value leading to incorrect results.
97
Basic Process Capability Templates – Process Sigma Level –

Discrete Data
Click SigmaXL > Templates & Calculators > Basic Process Capability Templates > Process Sigma
Level – Discrete to access the Process Sigma Level – Discrete calculator. The template gives the
Notes for use of the Process Sigma Calculator for Discrete Data:
1. Total number of defects should include defects made and later fixed.
2. Sample size should be large enough to observe 5 defects.
98
Basic Process Capability Templates – Process Sigma Level –

Continuous Data
Click SigmaXL > Templates & Calculators > Basic Process Capability Templates > Process Sigma
Level – Continuous to access the Process Sigma Level – Continuous calculator. The template gives
the following default example.
Note: This calculator assumes that the Mean and Standard Deviation are computed from data that
are normally distributed.
99
Basic Process Capability Templates – Process Capability Indices

Click SigmaXL > Templates & Calculators > Basic Process Capability Templates > Process Capability
to access the Process Capability Indices calculator. The template gives the following default
example.
Notes for use of the Process Capability Indices Calculator:
1. This calculator assumes that the Mean and Standard Deviation are computed from data that
2. Reports Cp, Cpk if entered S is Within or Short Term (using a control chart).
3. Reports Pp, Ppk if entered S is Overall or Long Term.
100
Basic Process Capability Templates – Process Capability &

Confidence Intervals
Click SigmaXL > Templates & Calculators > Basic Process Capability Templates > Process Capability
& Confidence Intervals to access the Process Capability Indices & Confidence Intervals calculator.
The template gives the following default example.
Notes for use of the Process Capability & Confidence Intervals Calculator:
1. This calculator assumes that the Mean and Standard Deviation are computed from data that
2. Reports Cp, Cpk if entered S is Within or Short Term (using a control chart).
3. Reports Pp, Ppk if entered S is Overall or Long Term.
101
Basic Process Capability Templates – Tolerance Interval

Calculator (Normal Exact)
Click SigmaXL > Templates & Calculators > Basic Process Capability Templates > Tolerance Interval
Calculator (Normal Exact) to access the Tolerance Interval Calculator (Normal Exact). The template
gives the following default example.
Notes:
1. Enter summarized Sample Data, Population Coverage and Confidence Level in cells with yellow
highlight.
2. Select Tolerance Interval Type using drop-down.
3. This calculator assumes that the summary statistics come from data that is normally
distributed. Tolerance intervals are exact, one-sided uses the noncentral t distribution, two-
sided uses the noncentral chi-square distribution. See Appendix Tolerance Interval Calculator
(Normal Exact) for statistical details.
function.
102
Basic Control Chart Templates – Individuals Chart

Click SigmaXL > Templates & Calculators > Basic Control Chart Templates > Individuals Chart.
Enter the data as shown (or copy and paste the “Before” data from the histogram data above):
Click the Individuals Control Chart button to produce the Individuals Chart (note that in actual
application, one should use a minimum of 20 to 30 data points to compute the initial control
limits).
Individuals Control Chart
Data UCL CL (Mean) LCL
70 64.34
60
Mean CL: 47.2

50
Individuals: Data
40
30.06
30
20
10
0
Before Before Before Before Before Before Before Before Before Before
103
Now enter the data as shown (or copy and paste the “After” data from the histogram data used
earlier):
Click the Add Data button to add the “After” data to the Individuals chart.
Individuals Control Chart
Data UCL CL (Mean) LCL
70
64.34
60
50 Mean CL: 47.2

Individuals: Data
40
30.06
30
20
10
0
re
re
re
re
re
r
te
te
te
te
te
fo
fo
fo
fo
fo
Af
Af
Af
Af
Af
Be
Be
Be
Be
Be
104
Notes:
1. This Individuals Control Chart template should be used with continuous data like cycle time.
The data must be in chronological time-sequence order.
2. You can replace the X-Axis Label and Data column headings with any headings that you wish.
4. Enter labels in X-Axis Label column. Labels can be Date, Time, Name, or other text information.
These labels are optional and will appear on the horizontal X-Axis of the Individuals Control
Chart.
5. Click the Individuals Control Chart button to create an Individuals Control Chart.
6. After the control chart has been created and additional new data entered into the Data
column, click the Add Data button to add the data to the existing chart.
105
Part C – Descriptive/Summary Statistics
Descriptive Statistics
1. Open Customer Data.xlsx. (To access, click SigmaXL > Help > Sample Data or Start > Programs
> SigmaXL > Sample Data). Click Sheet 1 Tab.
2. Click SigmaXL > Statistical Tools > Descriptive Statistics.
4. Select Overall Satisfaction, click Numeric Data Variables (Y) >>. Select Customer Type, click
Group Category (X1) >> as shown:
5. Click OK.
6. Descriptive Statistics are given for Customer Satisfaction grouped by Customer Type:
Which Customer Type has the highest mean satisfaction score? Clearly Type 2. However, we have
to be careful concluding that there is a significant difference in satisfaction solely by looking at the
Means. In the Analyze Phase, we will run tests of hypothesis to validate that Type 2 Customers are,
in fact, significantly more satisfied.
106
Tip: Click on Column B, click View | Window | Split, Freeze Panes. This freezes Column A and
allows you to scroll across the Descriptive Statistics for each level of the Group Category. This is
particularly beneficial when there are a large number of columns.
7. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Change the format selected
to Column Format as shown:
8. Click OK. Descriptive Statistics are given for Customer Satisfaction broken out by Customer
Type in Column Format:
Tip: This column format is useful to create subsequent graphs on the summary statistics.
107
Descriptive Statistics - Options

1. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Click Options. Check Select
All and change Percentile Confidence Intervals to Percentile to display all Percentile values in
the report.
Tip: Select only those options that are of interest in order to minimize the size of the report.
Here we are selecting all options for demonstration purposes. Note that when any option is
checked, Row Format is automatically selected, Column Format and Group Category (X2) are
greyed out. These display options are limited due to the amount of information displayed in
the extended report.
2. Click OK. Extended Descriptive Statistics are given for Customer Satisfaction grouped by
Customer Type:
108
3. The Additional Descriptive Statistics are:

• 5% Trimmed Mean. The highest 5% and lowest 5% are excluded and mean calculated with
the rest of the data. This gives a robust alternative to the Median as a measure of central
tendency in the presence of outliers.
• Standard Error of Mean (StDev/√𝑁𝑁)
• Variance (StDev2)
• Coefficient of Variation (100 * StDev/Mean)
• Short Term StDev (MR-bar/d2)
4. The Additional Normality Tests are:

• Shapiro-Wilk (n <= 5000) and Kolmogorov-Smirnov-Lilliefors (KSL, n > 5000)
o This is a popular alternative to Anderson Darling.
• Doornik-Hansen (DH)
o Univariate omnibus test based on Skewness and Kurtosis. (Note, the bivariate DH
test is used in Correlation Matrix to test bivariate normality).
o Best for data with ties, i.e. “chunky” data. Anderson-Darling, Shapiro-Wilk and KSL
are severely affected by ties in the data and will trigger a low P-Value even if the
data are normal.
o See Appendix Doornik-Hansen (DH) Normality Test for further details and
references.
109
5. The Percentile Report gives 27 values from 0.135 to 99.865.
6. The Percentile Ranges are:

• 75 - 25 (50%, Interquartile Range IQR)
• 90 - 10 (80%, Interdecile Range IDR)
• 95 - 5 (90%, Span)
• 97.5 - 2.5 (95%, +/- 1.96 Sigma Equivalent)
• 99 - 1 (98%)
• 99.5 - 0.5 (99%)
• 99.865 - 0.135 (99.73%, +/- 3 Sigma Equivalent)
110
7. The Percentile Confidence Intervals give 27 values from 0.135 to 99.865.
8. The Quartile Confidence Intervals give 3 values: 25, 50 and 75.
9. The Percentile Tolerance Intervals are 50%, 80%, 90%, 95%, 98%, 99%, and 99.73%.
10. Confidence Intervals and Tolerance Intervals can be exact or interpolated. If exact, the actual
exact confidence level will be a value greater than or equal to specified, due to percentile
values being discrete in nature. The actual exact level will also be reported in this case. If
interpolated, the result will be an interpolated estimate of the specified confidence level
(typically 95.0%) and is the recommended setting. See Appendix Percentile (Nonparametric)
Confidence and Tolerance Intervals for further details.
11. If the Confidence Interval or Tolerance Interval cannot be computed due to inadequate sample
size, a minimum sample size is reported.
111
12. The Outlier (Boxplot Rules) Tests are: Potential 1.5(IQR), Likely 2.2(IQR), Extreme 3.0(IQR)
13. Grubbs Outlier Test is more powerful at detecting a single outlier as maximum or minimum but
assumes that the remainder of the data are normally distributed.
14. The Randomness Runs Test is a nonparametric exact runs test.
15. The Outlier and Randomness Tests use the same Green, Yellow, Red highlight given in the
automatic assumptions report that are included in t-tests and ANOVA.
112
16. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Click Options. Uncheck
Select All to clear the selections and check Percentile Confidence Intervals, select Exact and
Percentile. Check Percentile Tolerance Intervals, and select Exact as shown:
113
17. Click OK. The Percentile Confidence Intervals and Tolerance Intervals are displayed:
18. The specified 95% is a guaranteed minimum. The exact confidence level is given with each
reported interval.
114
Part D – Histograms
Basic Histogram Template

Click SigmaXL > Templates and Calculators > Basic Graphical Templates > Histogram or
SigmaXL > Graphical Tools > Basic Graphical Templates > Histogram. See Part B – Templates
and Calculators for a Histogram Template example.
Single (Basic) Histogram

1. Click Sheet 1 Tab of Customer Data.xlsx (or press F4 to activate last worksheet).
2. Click SigmaXL > Graphical Tools > Basic Histogram.
4. Select Overall Satisfaction, click Numeric Data Variable (Y) >> as shown:
5. Click Next. Ensure that Normal Curve is checked. Set Start Point = 1. Change the Bin Width to
0.5, and the Number of Bins to 8. Click Update Chart to view the histogram. (If the survey
satisfaction data was pure integer format we would have checked the Integer Data option).
Note: For SigmaXL Mac version, click SigmaXL > Graphical Tools > Basic Histogram Options to
set options. The Histogram Basic Chart is displayed after you click Finish.
115
6. Click Finish. A histogram of Overall Customer Satisfaction is produced.
7. Note that bin one is 1 to < 1.5, bin 2 is 1.5 to < 2, etc.
Tip: Any graph produced by SigmaXL can be Copied/Pasted into Word. It can also be enlarged
by clicking on the graph and dragging the corner. The number of decimal places displayed can
be modified by clicking on the Axis Label and selecting the Number tab to adjust. The text label
alignment can also be modified by: Select Axis, Right Click, Format Axis.
Multiple Histograms
2. Click SigmaXL > Graphical Tools > Histograms & Descriptive Statistics.
4. Select Overall Satisfaction, click Numeric Data Variables (Y) >>, select Customer Type, click
Group Category (X1) >> as shown:
116
5. Click OK. Multiple Histograms and Descriptive Statistics of Customer Satisfaction By Customer
Type are produced:
Clearly Customer Type 2 shows a higher level of overall satisfaction, with the data skewed left.
Note that Customer Type 1 and 3 have data that is normally distributed, but this is not desirable
when the response is a satisfaction score!
6. Note that bin one is 1.72 to < 1.99, bin 2 is 1.99 to < 2.26, etc. The number of decimals
displayed can be changed by double-clicking on the X axis, click Number tab, and adjust decimal
places.
117
Part E – Dotplots
Dotplots
1. Click Sheet 1 Tab of Customer Data.xlsx (or press F4 to activate last worksheet). Click SigmaXL
> Graphical Tools > Dotplots.
3. Select Overall Satisfaction, click Numeric Data Variable (Y) >>, select Customer Type, click
Group Category (X1) >>:
4. Click OK. Multiple dotplots of Customer Satisfaction By Customer Type are produced:
5. Since all data points are shown, dotplots are a useful alternative to Histograms, particularly
when the group sample sizes are small (n < 30).
118
Part F – Boxplots
Boxplots
2. Click SigmaXL > Graphical Tools > Boxplots.
4. Select Overall Satisfaction, click Numeric Data Variable (Y) >>, select Customer Type, click
Group Category (X1) >>, check Show Mean, check Show Legend:
5. Click OK. A boxplot of Customer Satisfaction By Customer Type is produced:
6. The legend indicates that the solid center line is the median. The dashed red line shows the
sample mean. The top of the box is the 75th percentile (Q3). The bottom of the box is the 25th
percentile (Q1). The height of the box is called the Inter-Quartile Range (IQR) and is a robust
measure of spread or sample variability. The data point highlighted for Customer Type 2 is a
potential outlier (< Q1 – 1.5 * IQR or > Q3 + 1.5 * IQR). Note that extreme outliers are
highlighted with a solid dot (< Q1 – 3 * IQR or > Q3 + 3 * IQR).
119
Tip: If you hover your mouse cursor in the middle of any of the Boxplot lines as shown, you will see
the balloon help indicating what the line is and its numeric Y value.
7. Now we will modify the Y axis scale, showing 1 as minimum and 5 as maximum (given that the
response data comes from a survey with 1-5 scale). To do this right click on the Y axis and
select “Format Axis”, modify the minimum value and maximum value. Change Horizontal axis
crosses > Axis value to 1 as shown:
120
8. The Boxplot axis is modified as shown below:
9. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog.
10. Select Overall Satisfaction, click Numeric Data Variable (Y) >>; select Customer Type, click
Group Category (X1) >>; select Size of Customer, click Group Category (X2) >>; check Show
Mean; uncheck Show Legend:
121
11. Click OK. Boxplots of Customer Satisfaction By Customer Type and Size are produced:
12. In order to adjust the Y-axis scale for both charts, click SigmaXL Chart Tools > Set Chart Y-Axis
Max/Min.
13. Click OK. The Y-axis scale maximum and minimum are now modified for both charts.
122
Part G – Normal Probability Plots
Normal Probability Plots

1. Create 100 random normal values as follows: Click SigmaXL > Data Manipulation > Random
Data > Normal. Specify 1 Column, 100 Rows, Mean of 100 and Standard Deviation of 25 as
shown below:
2. Click OK. Change Column heading to Normal Data.
3. Create a Histogram & Descriptive Statistics for this data. Your data will be slightly different due
to the random number generation:
If the P-Value of the Anderson-Darling Normality test is greater than or equal to .05, the data is
considered to be normal (interpretation of P-Values will be discussed further in Analyze).
4. Create a normal probability plot of this data: Click SigmaXL Random Data (1) Sheet, Click
SigmaXL > Graphical Tools > Normal Probability Plots.
6. Select Normal Data, click Numeric Data Variable (Y) >>. Check Add Title. Enter Example
Normal Prob Plot. Click OK.
123
7. Click OK. A Normal Probability Plot of simulated random data is produced (again, your plot will
be slightly different due to the random number generation):
The data points follow the straight line fairly well, indicating that the data is normally distributed.
Note that the data will not likely fall in a perfectly straight line. The eminent statistician George
Box uses a “Fat Pencil” test where the data, if covered by a fat pencil, can be considered normal!
We can also see that the data is normal since the points fall within the normal probability plot 95%
confidence intervals (confidence intervals will be discussed further in Analyze).
9. Click SigmaXL > Graphical Tools > Normal Probability Plots.
124
11. Select Overall Satisfaction; click Numeric Data Variable (Y) >>. Click OK. A Normal Probability
Plot of Customer Satisfaction data is produced:
Is this data normally distributed? See earlier histogram and descriptive statistics of Customer
Satisfaction data.
12. Now we would like to stratify the customer satisfaction score by customer type and look at the
normal probability plots.
13. Click Sheet 1 of Customer Data.xlsx. Click SigmaXL > Graphical Tools > Normal Probability
Plots. Ensure that Entire Table is selected, click Next. (Alternatively, press F3 or click Recall
SigmaXL Dialog to recall last dialog).
14. Select Overall Satisfaction, click Numeric Data Variable (Y) >>; select Customer Type as Group
Category (X) >>. Click OK. Normal Probability Plots of Overall Satisfaction by Customer Type
are produced:
Reviewing these normal probability plots, along with the previously created histograms and
descriptive statistics, we see that the satisfaction data for customer type 2 is not normal, and
skewed left, which is desirable for satisfaction data! Note that although the customer type 2 data
falls within the 95% confidence intervals, the Anderson Darling test from descriptive statistics
125
shows p < .05 indicating nonnormal data. Smaller sample sizes tend to result in wider confidence
intervals, but we still see that the curvature for customer type 2 is quite strong.
Tip: Use the Normal Probability Plot (NPP) to distinguish reasons for nonnormality. If the data fails
the Anderson-Darling (AD) test (with p < 0.05) and forms a curve on the NPP, it is inherently
nonnormal or skewed. Calculations such as Sigma Level, Pp, Cp, Ppk and Cpk assume normality and
will therefore be affected. Consider transforming the data using LN(Y) or SQRT(Y) or using the Box-
Cox Transformation tool (SigmaXL > Data Manipulation > Box-Cox Transformation) to make the
data normal. Of course, whatever transformation you apply to your data, you must also apply to
your specification limits. See also the Process Capability for nonnormal data tools.
If the data fails the AD normality test, but the bulk of the data forms a straight line and there are
some outliers, the outliers are driving the nonnormality. Do not attempt to transform this data!
Determine the root cause for the outliers and take corrective action on those root causes.
126
Part H– Run Charts
Run charts, also known as trend charts and time series plots, add the dimension of time to the
graphical tools. They allow us to see trends and outliers in the data. Run Charts are a precursor to
control charts, which add calculated control limits. Note that Run Charts should be used only on
unsorted data, in its original chronological sequence.
Basic Run Chart Template

Click SigmaXL > Templates and Calculators > Basic Graphical Templates > Run Chart or
SigmaXL > Graphical Tools > Basic Graphical Templates > Run Chart. See Part B – Templates
and Calculators for a Run Chart Template example.
Run Charts
> Graphical Tools > Run Chart. Ensure that entire data table is selected. If not, check Use Entire
Data Table. Click Next.
2. Select Overall Satisfaction, click Numeric Data Variable (Y) >>. Select Show Mean. Uncheck
Nonparametric Runs Test (to be discussed later in Part N of Analyze Phase).
3. Click OK. A Run Chart of Overall Satisfaction with Mean center line is produced.
127
4. Select the Y axis, Right Click, Format Axis, to activate the Format Axis dialog. Change Minimum
to 1, Maximum to 5, Horizontal axis crosses > Axis value to 1:
5. The chart is automatically updated.
6. Are there any obvious trends? Some possible cycling, but nothing clearly stands out. It may be
interesting to look more closely at a specific data point. Any data point value can be identified
by simply moving the cursor over it:
128
7. A label can be added to a data point by one single-click on the data point, followed by a Right
mouse click, and select Add Data Label. See also SigmaXL Chart Tools > Add Data Label in
Control Phase Tools, Part B - X-Bar & Range Charts.
8. Click OK. Resulting Run Chart with label attached to data point:
9. This label can be changed to a text comment. Single-click three times on the label and type in a
comment as shown:
129
Overlay Run Charts

> Graphical Tools > Overlay Run Chart.
3. Select Overall Satisfaction, Responsive to Calls and Ease of Communications. Click Numeric Data
Variable (Y) >>.
4. Click OK. An Overlay Run Chart of Overall Satisfaction, Responsive to Calls and Ease of
Communications is produced.
Overall Satisfaction
3 Responsive to Calls
Ease of Communications
0
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97
130
Part I – Measurement Systems Analysis
Type 1 Gage Study Template

Click SigmaXL > Templates and Calculators > Basic MSA Templates > Type 1 Gage Study or
SigmaXL > Measurement Systems Analysis > Basic MSA Templates > Type 1 Gage Study.
See Basic MSA Templates – Type 1 Gage Study for an example of the Type 1 Gage Study template.
Gage Bias and Linearity Study Template

Click SigmaXL > Templates and Calculators > Basic MSA Templates > Gage Bias and Linearity
Study or SigmaXL > Measurement Systems Analysis > Basic MSA Templates > Gage Bias and
Linearity Study.
See Basic MSA Templates – Gage Bias and Linearity Study for an example of the Gage Bias and
Linearity Study template.
Gage R&R Study (MSA) Template

Click SigmaXL > Templates and Calculators > Basic MSA Templates > Gage R&R Study (MSA) or
SigmaXL > Measurement Systems Analysis > Basic MSA Templates > Gage R&R Study (MSA).
See Basic MSA Templates – Gage R&R Study (MSA) for an example of the Gage R&R Study (MSA)
template. If you wish to analyze the template data with the Gage R&R Analysis Tool, click Create
Stacked Column Format for “Analyze Gage R&R” >>.
After entering the data into the Gage R&R Study template, you can create Gage Multi-Vari & X-bar
R charts. Click SigmaXL > Templates and Calculators > Basic MSA Templates > Gage R&R: Multi-
Vari & X-bar R Charts or SigmaXL > Measurement Systems Analysis > Basic MSA Templates >
Gage R&R: Multi-Vari & X-bar R Charts. See Basic MSA Templates - Gage R&R: Multi-Vari & X-bar
R Charts for an example.
Attribute Gage R&R (MSA) Template

Click SigmaXL > Templates and Calculators > Basic MSA Templates > Attribute Gage R&R (MSA) or
SigmaXL > Measurement Systems Analysis > Basic MSA Templates > Attribute Gage R&R (MSA).
131
See Basic MSA Templates – Attribute Gage R&R (MSA) for an example of the Attribute Gage R&R
(MSA) template.
Create Gage R&R (Crossed) Worksheet

1. To create a blank Gage R&R worksheet, click SigmaXL > Measurement Systems Analysis >
Create Gage R&R (Crossed) Worksheet. A crossed Gage R&R study means that every
operator/appraiser will measure every part/sample.
2. The default settings are 10 Parts, 3 Operators, 3 Replicates with Parts and Operators
randomized. Clicking the Reset button will restore the dialog to these default settings.
3. Part and Operator names can be edited.
4. Click OK. The following worksheet is produced:
132
5. Your worksheet order will be different due to the randomization. The results of the Gage R&R
study are entered in the yellow highlighted Measurement cells.
133
Analyze Gage R&R (Crossed)

1. Open the file Gage RR – AIAG.xlsx. This is an example from the Automotive Industry Action
Group (AIAG) MSA Reference Manual, 3rd Edition, page 101. Note that parts were measured in
random order, but the worksheet is given in standard order.
2. Click SigmaXL > Measurement Systems Analysis > Analyze Gage R&R (Crossed). The data
worksheet is recognized by SigmaXL and highlighted automatically. Click Next.
3. Select Part, Operator and Measurement as shown. Check Display Multi-Vari & X-Bar R Charts.
Check Tolerance/Historical StDev (Optional) and enter 8 for the Process Tolerance Upper-
Lower Spec.
134
4. Click OK. The following Gage R&R Study Report is produced:
135
5. The first ANOVA table shows that the Operator * Part Interaction is not significant with the P-
Value = 0.9741 (greater than the alpha-to-remove value of 0.1). The second ANOVA table
shows the results excluding the Part * Operator Interaction.
6. The cells shaded in light blue highlight the critical metrics: Gage R&R % Total Variation (also
known as %R&R) and %Tolerance: < 10% indicates a good measurement system;
> 30% indicates an unacceptable measurement system. The Number of Distinct Categories
should be at least 5. Initial analysis shows that this is a marginal measurement system.
(Traditionally NDC is truncated to an integer value, but SigmaXL reports a more informative one
decimal place).
7. SigmaXL also reports the confidence intervals for the Gage R&R metrics. Note that when the
confidence intervals are taken into account, it is possible that this measurement system is
totally inadequate. This strongly suggests that the measurement system needs to be improved
but it also points out the weakness of the traditional Gage R&R study with 10 Parts, 3
Operators, and 3 Replicates. These sample sizes will typically yield wide confidence intervals.
For statistical details of the Gage R&R confidence intervals see Burdick, R. K., Borror, C. M., and
Montgomery, D. C., “Design and Analysis of Gauge R&R Studies: Making Decisions with
Confidence Intervals in Random and Mixed ANOVA Models”, ASA-SIAM Series on Statistics and
Applied Probability, 2005.
8. Click on the Gage R&R – Multi-Vari Sheet tab to view the Multi-Vari Chart:
The Multi-Vari chart shows each Part as a separate graph. Each Operator’s response readings
are denoted as a vertical line with the top tick corresponding to the Maximum value, bottom
tick is the Minimum, and the middle tick is the Mean. The horizontal line across each graph is
the overall average for each part.
136
9. Click on the Gage R&R – X-Bar R Sheet tab to view the X-Bar & R Chart:
When interpreting the X-bar and R chart for a Gage R&R study, it is desirable that the X-bar
chart be out-of-control, and the Range chart be in-control. The control limits are derived from
within Operator repeatability. The Range chart indicates a problem with Operator B being
inconsistent.
137
Attribute MSA (Binary)

Attribute MSA is also known as Attribute Agreement Analysis. Use the Binary option if the assessed
result is text or numeric binary (e.g., 0/1, Pass/Fail, Good/Bad, G/NG, Yes/No).
1. Open the file Attribute MSA – AIAG.xlsx. This is an example from the Automotive Industry
Action Group (AIAG) MSA Reference Manual, 3rd edition, page 127 (4th Edition, page 134).
There are 50 samples, 3 appraisers and 3 trials with a 0/1 response. A “good” sample is denoted
as a 1. A “bad” sample is denoted as a 0. Note that the worksheet data must be in stacked
column format and the known reference values must be consistent for each sample.
2. Click SigmaXL > Measurement Systems Analysis > Attribute MSA (Binary). Ensure that the
entire data table is selected. Click Next.
3. Select Part, Appraiser, Assessed Result and Reference as shown. Check Report Information and
enter AIAG Attribute MSA Binary for Product/Unit Name. Select Percent Confidence Interval
Type – Exact. The default Good Level of “1”will be used as specified in the AIAG manual :
Tip: The Good Level definition is used to determine Type I and Type II error rates. It is
applicable only when a True Standard is selected.
Tip: Percent Confidence Interval Type applies to the Percent Agreement and Percent
Effectiveness Confidence Intervals. These are binomial proportions that have an "oscillation"
phenomenon where the coverage probability varies with the sample size and proportion value.
Exact is strictly conservative and will guarantee the specified confidence level as a minimum
coverage probability, but results in wider intervals. Wilson Score has a mean coverage
138
probability that matches the specified confidence interval. Since the intervals are narrower and
thereby more powerful, Wilson Score is recommended for use in attribute MSA studies due to
the small sample sizes typically used. Exact is selected in this example for continuity with the
results from SigmaXL Version 6.
4. Click OK. The Attribute MSA Binary Analysis Report is produced. The tables and associated
graphs are described separately by section for clarity.
Tip: While this report is quite extensive, a quick assessment of the attribute measurement
system can be made by viewing the Kappa color highlights: Green - very good agreement
(Kappa >= 0.9); Yellow - marginally acceptable, improvement should be considered (Kappa 0.7
to < 0.9); Red - unacceptable (Kappa < 0.7). Further details on Kappa are given below.
139
Within Appraiser Agreement is an assessment of each appraiser’s consistency of ratings across

trials and requires at least two trials. This is analogous to Gage R&R Repeatability. Note that the
reference standard is not considered, so an appraiser may be perfectly consistent but consistently
wrong!
Percent/CI: Within Appraiser Agreement Graph:
Kappa/CI: Within Appraiser Agreement Graph:
Within Appraiser Agreement Table:
Within Appraiser Percent Agreement will decrease as the number of trials increase because a
match occurs only if an appraiser is consistent across all trials. Use the Kappa/CI: Within Appraiser
Agreement Graph to determine adequacy of the Within Appraiser agreement. See below for
additional interpretation guidelines.
Tip: Hover the mouse pointer over the heading cells to view the following report comments.
140
# Inspected: Number of parts or samples.
# Matched: A match occurs only if an appraiser is consistent across all trials.
Percent Agreement = (# Matched / # Inspected) * 100
LC = Percent Lower Confidence Limit. UC = Percent Upper Confidence Limit. Confidence intervals
(CI) for binomial proportions have an "oscillation" phenomenon where the coverage probability
varies with n and p. Exact is strictly conservative and will guarantee the specified confidence level
as a minimum coverage probability, but results in wide intervals. Wilson Score has a mean
coverage probability that matches the specified confidence interval. Since the intervals are
narrower and thereby more powerful, they are recommended for use in attribute MSA studies due
to the small sample sizes typically used. See Appendix Percent Confidence Intervals (Exact Versus
Wilson Score) for references.
Fleiss’ Kappa statistic is a measure of agreement that is analogous to a “correlation coefficient” for
discrete data. Kappa ranges from -1 to +1: A Kappa value of +1 indicates perfect agreement. If
Kappa = 0, then agreement is the same as would be expected by chance. If Kappa = -1, then there
is perfect disagreement. “Rule-of-thumb” interpretation guidelines: >= 0.9 very good agreement
(green); 0.7 to < 0.9 marginally acceptable, improvement should be considered (yellow); < 0.7
unacceptable (red). See Appendix Kappa for further details on the Kappa calculations and “rule-of-
thumb” interpretation guidelines.
Fleiss’ Kappa P-Value: H0: Kappa = 0. If P-Value < alpha (.05 for specified 95% confidence level),
reject H0 and conclude that agreement is not the same as would be expected by chance. Significant
P-Values are highlighted in red.
Fleiss' Kappa LC (Lower Confidence) and Fleiss' Kappa UC (Upper Confidence) limits use a kappa
normal approximation. Interpretation Guidelines: Kappa lower confidence limit >= 0.9: very good
agreement. Kappa upper confidence limit < 0.7: the attribute agreement is unacceptable. Wide
confidence intervals indicate that the sample size is inadequate.
In this example, we have marginal Within Appraiser Agreement for each of the appraisers.
141
Each Appraiser vs. Standard Agreement is an assessment of each appraiser’s ratings across trials
compared to a known reference standard. This is analogous to Gage R&R Accuracy.
Percent/CI: Each Appraiser vs. Standard Agreement Graph:
Each Appraiser vs. Standard Agreement Table:
Tip: The Percent/CI Each Appraiser vs. Standard Agreement Graph can be used to compare
agreement to standard across the appraisers, but should not be used as an absolute measure of
agreement. Each Appraiser vs. Standard Agreement will decrease as the number of trials increase
because a match occurs only if an appraiser agrees with the standard consistently across all trials.
Use Fleiss’ Kappa in the Each Appraiser vs. Standard Agreement Table to determine the adequacy
of Each Appraiser versus Standard agreement.
# Inspected: Number of parts or samples.
# Matched: A match occurs only if an appraiser agrees with the standard consistently across all
trials.
Percent Agreement = (# Matched / # Inspected) * 100
Kappa is interpreted as above: >= 0.9 very good agreement (green); 0.7 to < 0.9 marginally
acceptable, improvement should be considered (yellow); < 0.7 unacceptable (red).
Appraisers A and C have marginal agreement versus the standard values. Appraiser B has very good
agreement to the standard.
142
Each Appraiser vs. Standard Disagreement is a breakdown of each appraiser’s rating

misclassifications (compared to a known reference standard). This table is applicable only to binary
two-level responses (e.g., 0/1, G/NG, Pass/Fail, True/False, Yes/No).
A Type I Error occurs when the appraiser consistently assesses a good part/sample as bad. "Good"
is defined by the user in the Attribute MSA analysis dialog.
Type I Error % = (Type I Error / # Good Parts or Samples) * 100
A Type II Error occurs when a bad part/sample is consistently assessed as good.
Type II Error % = (Type II Error / # Bad Parts or Samples) * 100
A Mixed Error occurs when the assessments across trials are not identical.
Mixed Error % = (Mixed Error / # Parts or Samples) * 100
Between Appraiser Agreement is an assessment of the appraisers’ consistency of ratings across

trials and between each other. At least two appraisers are required. This is analogous to Gage R&R
Reproducibility. Note that the reference standard is not considered, so the appraisers may be
perfectly consistent, but consistently wrong!
All Appraisers vs. Standard Agreement is an assessment of all appraisers’ ratings across trials
compared to a known reference standard. This is analogous to Gage R&R Accuracy.
acceptable (yellow); < 0.7 unacceptable (red).
Since the Between Appraiser Agreement and All Appraisers vs. Standard Agreement are
marginally acceptable, improvements to the attribute measurement should be considered. Look
for unclear or confusing operational definitions, inadequate training, operator distractions or poor
lighting. Consider the use of pictures to clearly define a defect.
143
The Attribute Effectiveness Report is similar to the Attribute Agreement Report, but treats each
trial as an opportunity. Consistency across trials or appraisers is not considered. This has the
benefit of providing a Percent measure that is unaffected by the number of trials or appraisers.
Also, the increased sample size for # Inspected results in a reduction of the width of the Percent
confidence interval. The Misclassification report shows all errors classified as Type I or Type II.
Mixed errors are not relevant here. This report requires a known reference standard and includes:
Each Appraiser vs. Standard Effectiveness, All Appraisers vs. Standard Effectiveness, and
Effectiveness and Misclassification Summary.
Each Appraiser vs. Standard Effectiveness is an assessment of each appraiser’s ratings compared
to a known reference standard. This is analogous to Gage R&R Accuracy. Unlike the Each Appraiser
vs. Standard Agreement table above, consistency across trials is not considered here - each trial is
considered as an opportunity. This has the benefit of providing a Percent measure that is
unaffected by the number of trials. Also, the increased sample size for # Inspected results in a
reduction of the width of the Percent confidence interval.
Percent/CI: Each Appraiser vs. Standard Effectiveness Graph:
Kappa/CI: Each Appraiser vs. Standard Effectiveness Graph:
144
Each Appraiser vs. Standard Effectiveness Table:
# Inspected = # Parts or Samples * # Trials
# Matched: A match occurs if an appraiser agrees with the standard (consistency across trials is not
considered here).
Percent Effectiveness = (# Matched / # Inspected) * 100. Interpretation Guidelines for Percent

Effectiveness: => 95% very good; 85% to <95% marginal, may be acceptable but improvement
should be considered; < 85% unacceptable. These guidelines assume an equal number of known
good and known bad parts/samples.
acceptable (yellow); < 0.7 unacceptable (red).
Tip: The Kappa values in the Effectiveness tables are very similar to those in the Agreement tables
(the slight difference is due to average Kappa for unstacked versus Kappa for stacked data). This is
why the Kappa/CI Each Appraiser vs. Standard Agreement graph is not shown. It would essentially
be a duplicate of the Kappa/CI Each Appraiser vs. Standard Effectiveness graph.
Appraisers A and C have marginal agreement versus the standard values, with less than 95%
Effectiveness and Kappa < 0.9. Appraiser B has very good agreement to the standard.
145
All Appraisers vs. Standard Effectiveness is an assessment of all appraisers’ ratings compared to a
known reference standard. This is analogous to Gage R&R Accuracy. Unlike the All Appraiser vs.
Standard Agreement table above, consistency across trials and appraisers is not considered here -
each trial is considered as an opportunity. This has the benefit of providing a Percent measure that
is unaffected by the number of trials or appraisers. Also, the increased sample size for # Inspected
results in a reduction of the width of the Percent confidence interval.
# Inspected = # Parts or Samples * # Trials * # Appraisers
# Matched: A match occurs if an appraiser agrees with the standard (consistency across trials and
appraisers is not considered here).
Percent Effectiveness = (# Matched / # Inspected) * 100.
The interpretation guidelines for Kappa and Percent Effectiveness are the same as noted above.
This measurement system is marginally acceptable.
146
Each Appraiser vs. Standard Misclassification is a breakdown of each appraiser’s rating

two-level responses (e.g., 0/1, G/NG, Pass/Fail, True/False, Yes/No). Unlike the Each Appraiser vs.
Standard Disagreement table above, consistency across trials is not considered here. All errors are
classified as Type I or Type II. Mixed errors are not relevant.
A Type I error occurs when the appraiser assesses a good part/sample as bad (consistency across
trials is not considered here). "Good" is defined by the user in the Attribute MSA analysis dialog.
See Misclassification Legend for specific definition of Type I and Type II Errors.
# Inspected = # Good Parts or Samples * # Trials
Type I Error % = (Type I Error / # Inspected Good) * 100
A Type II error occurs when a bad part/sample is assessed as good. See Misclassification Legend
for specific definition of Type I and Type II Errors.
# Inspected = # Bad Parts or Samples * # Trials
Type II Error % = (Type II Error / # Inspected Bad) * 100
All Appraisers vs. Standard Misclassification is a breakdown of all appraisers’ rating

two-level responses (e.g., 0/1, G/NG, Pass/Fail, True/False, Yes/No). Unlike the All Appraisers vs.
Standard Disagreement table above, consistency across trials and appraisers is not considered
here. All errors are classified as Type I or Type II. Mixed errors are not relevant.
A Type I error occurs when the appraiser assesses a good part/sample as bad (consistency across
trials is not considered here). "Good" is defined by the user in the Attribute MSA analysis dialog.
See Misclassification Legend for specific definition of Type I and Type II Errors.
# Inspected = # Good Parts or Samples * # Trials * # Appraisers
Type I Error % = (Type I Error / # Inspected Good) * 100
A Type II error occurs when a bad part/sample is assessed as good. See Misclassification Legend
for specific definition of Type I and Type II Errors.
# Inspected = # Bad Parts or Samples * # Trials * # Appraisers
Type II Error % = (Type II Error / # Inspected Bad) * 100
147
Effectiveness and Misclassification Summary is a summary table of all appraisers’ correct rating
counts and misclassification counts compared to the known reference standard values.
Attribute MSA Data is a summary showing the original data in unstacked format. This makes it
easy to visually compare appraiser results by part. If a reference standard is provided, the cells are
color highlighted as follows: agrees with reference (green); Type I error (yellow); Type II error (red):
In conclusion, with the Kappa scores in the “yellow zone” (< 0.9) and Percent Effectiveness less
than 95% this measurement system is marginal and should be improved. Appraiser B is the
exception and does well against the standard. Look for unclear or confusing operational definitions,
inadequate training, operator distractions or poor lighting. Consider the use of pictures to clearly
define a defect. Use Attribute MSA as a way to “put your stake in the ground” and track the
effectiveness of improvements to the measurement system.
148
Attribute MSA (Ordinal)

Attribute MSA is also known as Attribute Agreement Analysis. Use the Ordinal option if the
assessed result is numeric ordinal (e.g., 1, 2, 3, 4, 5). There must be at least 3 response levels in the
assessed result, otherwise it is binary. Examples of ordinal responses used elsewhere in this
workbook include:
• Customer Loyalty –Likely to Recommend score which contains ordinal integer values from 1
to 5, where a 1 indicates that the customer is very unlikely to recommend and a 5 indicates
that the customer is very likely to recommend. This example is used in Ordinal Logistic
Regression.
• Taste Score on a scale of 1-7 where 1 is "awful" and 7 is "delicious." This is used in the cake
bake taste test Design of Experiments: Part B – Three Factor Full Factorial Example Using
DOE Template.
An Ordinal Attribute MSA study should be done prior to formal ordinal data collection for use in
hypothesis testing, regression or design of experiments.
1. Open the file Attribute MSA – Ordinal.xlsx. This is an Ordinal MSA example with 50 samples, 3
appraisers and 3 trials. The response is 1 to 5, grading product quality. One denotes “Very Poor
Quality,” 2 is “Poor,” 3 is “Fair,” 4 is “Good” and a 5 is “Very Good Quality.” The Expert
Reference column is the reference standard from an expert appraisal. Note that the worksheet
data must be in stacked column format and the reference values must be consistent for each
sample.
2. Click SigmaXL > Measurement Systems Analysis > Attribute MSA (Ordinal). Ensure that the
149
3. Select Sample No., Appraiser, Assessed Result and Expert Reference as shown. Check Report
Information and enter Attribute MSA Ordinal for Product/Unit Name. Select Percent
Confidence Interval Type – Wilson Score:
4. Click OK. The Attribute MSA Ordinal Analysis Report is produced.
Tip: While this report is quite extensive, a quick assessment of the attribute measurement
system can be made by viewing the Kendall Concordance and Kendall Correlation color
highlights: Green - very good agreement; Yellow - marginally acceptable, improvement should
be considered; Red - unacceptable. Further details on the Kendall Coefficients are given below.
Tip: Fleiss’ Kappa and Percent Agreement are included in the report for completeness but not
recommended for use with Ordinal response data because they treat each response level as
nominal. Kendall’s Concordance and Correlation take the order of the data into account, so a
deviation of 1 is not as bad as a deviation of 2 or more. See Attribute MSA – Nominal for a
discussion of the Fleiss’ Kappa report.
150
Kendall’s Concordance/CI Within Appraiser Agreement Graph:
Within Appraiser Agreement Table:
Between Appraiser Agreement Table:
Kendall's Coefficient of Concordance (Kendall's W) is a measure of association for discrete ordinal

data, used for assessments that do not include a known reference standard. Kendall’s coefficient
of concordance ranges from 0 to 1: A coefficient value of 1 indicates perfect agreement. If the
coefficient = 0, then the agreement is random, i.e., the same as would be expected by chance.
“Rule-of-thumb” interpretation guidelines: >= 0.9 very good agreement (green); 0.7 to < 0.9
marginally acceptable, improvement should be considered (yellow); < 0.7 unacceptable (red).
See Appendix Kendall’s Coefficient of Concordance for further details on the Kendall Concordance
calculations and “rule-of-thumb” interpretation guidelines.
Kendall's Concordance P-Value: H0: Kendall's Coefficient of Concordance = 0. If P-Value < alpha
(.05 for specified 95% confidence level), reject H0 and conclude that agreement is not the same as
would be expected by chance. Significant P-Values are highlighted in red.
Kendall's Concordance LC (Lower Confidence) limit and Kendall's Concordance UC (Upper

Confidence) limit cannot be solved analytically, so are estimated using bootstrapping. See Appendix
Kendall’s Coefficient of Concordance for further details on the bootstrap confidence intervals.
Interpretation Guidelines: Concordance lower confidence limit >= 0.9: very good agreement.
Concordance upper confidence limit < 0.7: the attribute agreement is unacceptable. Wide
The Within Appraiser Agreement for Joe is marginal, Moe is unacceptable and Sue is very good.
The Between Appraiser Agreement is unacceptable.
151
Kendall’s Correlation/CI Each Appraiser vs. Standard Effectiveness Graph:
Each Appraiser vs. Standard Agreement Table:
All Appraisers vs. Standard Agreement Table:
Each Appraiser vs. Standard Effectiveness Table:
All Appraisers vs. Standard Effectiveness Table:
Kendall's Correlation Coefficient (Kendall's tau-b) is a measure of association for discrete ordinal
data, used for assessments that include a known reference standard. Kendall’s correlation
coefficient ranges from -1 to 1: A coefficient value of 1 indicates perfect agreement. If the
coefficient = 0, then the agreement is random, i.e., the same as would be expected by chance. A
coefficient value of -1 indicates perfect disagreement. “Rule-of-thumb” interpretation guidelines:
>= 0.8 very good agreement (green); 0.6 to < 0.8 marginally acceptable, improvement should be
considered (yellow); < 0.6 unacceptable (red). See Appendix Kendall’s Correlation Coefficient for
further details on the Kendall Correlation calculations and “rule-of-thumb” interpretation
guidelines.
152
Kendall's Correlation P-Value: H0: Kendall's Correlation Coefficient = 0. If P-Value < alpha (.05 for
specified 95% confidence level), reject H0 and conclude that agreement is not the same as would
be expected by chance. Significant P-Values are highlighted in red.
Kendall's Correlation LC (Lower Confidence) and Kendall's Correlation UC (Upper Confidence) limit
use a normal approximation. Interpretation Guidelines: Correlation lower confidence limit >= 0.8:
very good agreement. Correlation upper confidence limit < 0.6: the attribute agreement is
unacceptable. Wide confidence intervals indicate that the sample size is inadequate.
Tip: Kendall’s Correlation values in the Effectiveness tables are very similar to those in the
Agreement tables (the slight difference is due to average Kendall for unstacked versus Kendall for
stacked data). This is why the Kendall’s Correlation/CI Each Appraiser vs. Standard Agreement
graph is not shown. It would essentially be a duplicate of the Kendall’s Correlation/CI Each
Appraiser vs. Standard Effectiveness graph.
Appraiser Joe has marginal agreement versus the standard values. Appraiser Moe has
unacceptable agreement to the standard. Sue has very good agreement to the standard.
Overall, the appraisers have marginal agreement to the standard.
Note that the Percent Agreement results in All Appraisers vs. Standard Agreement Table show
only 2% agreement! This is due to the requirement that all appraisers agree with the standard
across all trials for a 5 level response, which is very unlikely to occur. This highlights the problem
with using Percent Agreement in an Ordinal MSA. Kendall’s coefficients are the key metric to assess
an Ordinal MSA.
153
Effectiveness and Misclassification Summary is a summary table of all appraisers’ correct rating
counts and misclassification counts compared to the known reference standard values.
Attribute MSA Data is a summary showing the original data in unstacked format. This makes it
easy to visually compare appraiser results by part. If a reference standard is provided, the cells are
color highlighted as follows: absolute deviation = 0 (green); absolute deviation = 1 (yellow);
absolute deviation >=2 (red):
In conclusion, this measurement system is marginal and should be improved. Appraiser Moe needs
training and Appraiser Joe needs a refresher. Sue has very good agreement based on Kendall’s
Concordance and Correlation, but would have been considered marginal based on Kappa (< .9) and
Percent Effectiveness (< 95%). As discussed above, Kappa and Percent Effectiveness do not take
the order of the response data into account, so are not as useful as Kendall’s coefficients in an
Ordinal MSA study.
154
Attribute MSA (Nominal)

Attribute MSA is also known as Attribute Agreement Analysis. Use the Nominal option if the
assessed result is numeric or text nominal (e.g., Defect Type 1, Defect Type 2, Defect Type 3).
There must be at least 3 response levels in the assessed result, otherwise it is binary.
1. Open the file Attribute MSA – Nominal.xlsx. This is a Nominal MSA example with 30 samples, 3
appraisers and 2 trials. The response is text “Type_1”, “Type_2” and “Type_3.” The Expert
Reference column is the reference standard from an expert appraisal. Note that the worksheet
data must be in stacked column format and the reference values must be consistent for each
sample.
2. Click SigmaXL > Measurement Systems Analysis > Attribute MSA (Nominal). Ensure that the
3. Select Sample No., Appraiser, Assessed Result and Expert Reference as shown. Check Report
Information and enter Attribute MSA – Nominal for Product/Unit Name. Select Percent
Confidence Interval Type – Wilson Score:
4. Click OK. The Attribute MSA Nominal Analysis Report is produced.
Tip: While this report is quite extensive, a quick assessment of the attribute measurement system
can be made by viewing the Kappa color highlights: Green - very good agreement (Kappa >= 0.9);
Yellow - marginally acceptable, improvement should be considered (Kappa 0.7 to < 0.9); Red -
unacceptable (Kappa < 0.7). See Attribute MSA (Binary) for a detailed discussion of the report
graphs and tables. Here we will just look at Kappa.
155
Attribute Agreement Report:
Attribute Effectiveness Report:
Fleiss’ Kappa statistic is a measure of agreement that is analogous to a “correlation coefficient” for
discrete data. Kappa ranges from -1 to +1: A Kappa value of +1 indicates perfect agreement. If
Kappa = 0, then agreement is the same as would be expected by chance. If Kappa = -1, then there
is perfect disagreement. “Rule-of-thumb” interpretation guidelines: >= 0.9 very good agreement
(green); 0.7 to < 0.9 marginally acceptable, improvement should be considered (yellow); < 0.7
unacceptable (red). See Appendix Kappa for further details on the Kappa calculations and “rule-of-
thumb” interpretation guidelines.
Fleiss’ Kappa P-Value: H0: Kappa = 0. If P-Value < alpha (.05 for specified 95% confidence level),
reject H0 and conclude that agreement is not the same as would be expected by chance. Significant
P-Values are highlighted in red.
Fleiss' Kappa LC (Lower Confidence) and Fleiss' Kappa UC (Upper Confidence) limits use a kappa
normal approximation. Interpretation Guidelines: Kappa lower confidence limit >= 0.9: very good
agreement. Kappa upper confidence limit < 0.7: the attribute agreement is unacceptable. Wide
Fleiss’ Kappa Overall is an overall Kappa for all of the response levels.
Fleiss’ Kappa Individual gives Kappa for each response level (Type_1, Type_2 and Type_3). This is
useful to identify if an appraiser has difficulty assessing a particular defect type.
This is a very good attribute measurement system, but Appraiser C is marginal so a refresher would
be helpful.
156
Part J – Process Capability
Process Capability Templates

Click SigmaXL > Templates and Calculators > Basic Process Capability Templates >
Process Sigma Level – Discrete Data or
Process Sigma Level – Continuous Data or
Process Capability Indices or
Process Capability & Confidence Intervals or
Tolerance Interval Calculator (Normal Exact).
These templates are also located at SigmaXL > Process Capability > Basic Process Capability
Templates.
See Measure Phase Part B – Templates and Calculators for Process Capability Template examples:
Basic Process Capability Templates – Process Sigma Level – Discrete Data
Basic Process Capability Templates – Process Sigma Level – Continuous Data
Basic Process Capability Templates – Process Capability Indices
Basic Process Capability Templates – Process Capability & Confidence Intervals
Basic Process Capability Templates – Tolerance Interval Calculator (Normal Exact)
157
Histograms and Process Capability

1. Open the file Delivery Times.xlsx. This contains continuous data of hotel breakfast delivery
times. Deviation Time is the deviation around targeted delivery time in minutes. The Critical
Customer Requirements (CCR’s) are as follows: USL = 10 minutes late, LSL = -10 minutes (early).
2. Let’s begin with a view of the data using Histograms and Descriptive Statistics. Click SigmaXL >
Graphical Tools > Histograms & Descriptive Statistics.
4. Select Delivery Time Deviation, click Numeric Data Variable (Y) >>. Click OK.
Note that the data has an average of 6, so on average the breakfast delivery is 6 minutes late.
Also note the wide variation in delivery times (StDev = 7.2 minutes). With the AD P-Value > .05,
this data can be assumed to have a normal distribution.
158
5. Click Sheet 1 Tab. Click SigmaXL > Process Capability > Histograms & Process Capability (or
SigmaXL > Graphical Tools > Histograms & Process Capability).
6. Select Delivery Time Deviation, click Numeric Data Variable (Y) >>. Enter USL = 10, Target = 0,
LSL = -10, check Normal Curve as shown below:
7. Click OK. Resulting Histogram & Process Capability report is shown below:
With the Process Performance indices Pp and Ppk < 1, this process is clearly in need of
improvement. Note that the difference between Pp and Ppk is due to the off-center process
mean. (Cp and Cpk indices are provided in the Capability Combination Report or optionally
when creating Control Charts).
159
Capability Combination Report (Individuals)

1. Open the file Delivery Times.xlsx.
2. Click SigmaXL >Process Capability > Capability Combination Report (Individuals).
3. Select Delivery Time Deviation, click Numeric Data Variable (Y) >>. Enter USL = 10, Target = 0,
LSL = -10, as shown below:
4. Click OK. The resulting Process Capability Combination report is shown below:
160
5. To print this report, click SigmaXL Chart Tools > Print Charts. Do not use Excel’s print function.
6. Cp is the “best case” potential capability index. Ppk is the “worst case” actual performance. If
Cp is excellent (say Cp =2) and Ppk is poor (say Ppk < 1), this indicates that the process can be
dramatically improved through centering and stabilizing. If Cp is poor, you will have to
implement fundamental improvements using DMAIC to address common cause variation. If
Ppk = Cp, this indicates that the process is centered and stable, and the short-term (within
subgroup) variation is the same as the long-term (overall) variation.
7. Cpm is a similar statistic but it incorporates a penalty for deviation from target.
8. On a technical note regarding compatibility with other software tools, please be advised that
SigmaXL does not use unbiasing constants when calculating Overall StDev, Pp, Ppk and Cpm.
This is done in order to ensure that the calculated Overall StDev matches the results given in
Descriptive Statistics. Note that Within StDev, Cp and Cpk are based on control chart
techniques (MR-bar/d2 for Individuals Chart, R-bar/d2 for X-Bar & R, S-bar/c4 for X-Bar & S).
161
Capability Combination Report (Subgroups)

1. Open the file Catapult Data – Xbar Control Charts.xlsx. Each operator fires the catapult ball 3
times. The target distance is 100 inches. The Upper Specification Limit (USL) is 108 inches. The
Lower Specification Limit (LSL) is 92 inches.
2. Select B2:F22; here, we will only use the first 20 subgroups to determine the process capability
(subgroups 21 to 25 are studied later in Control Phase Tools: Statistical Process Control (SPC)
Charts: Part B - X-Bar & Range Charts).
3. Select SigmaXL > Process Capability > Capability Combination Report (Subgroups).
4. Do not check Use Entire Data Table!
5. Click Next. Select Subgroups across Rows, select Shot 1, Shot 2, Shot 3, click Numeric Data
Variables (Y) >>. Select Operator, click Optional X-Axis Labels >>. Select X-Bar & R Charts.
Check Tests for Special Causes. Enter USL = 108, Target = 100, LSL = 92, as shown:
162
163
Capability Combination Report (Individuals Nonnormal)

An important assumption for process capability analysis is that the data be normally distributed.
The Capability Combination Report (Individuals Nonnormal) allows you to transform the data to
normality or utilize nonnormal distributions, including:
Box-Cox Transformation (includes an automatic threshold option so that data with negative values
can be transformed)
Johnson Transformation
Distributions supported:
 Half-Normal
 Lognormal (2 & 3 parameter)
 Exponential (1 & 2 parameter)
 Weibull (2 & 3 parameter)
 Beta (2 & 4 parameter)
 Gamma (2 & 3 parameter)
 Logistic
 Loglogistic (2 & 3 parameter)
 Largest Extreme Value
 Smallest Extreme Value
Automatic Best Fit based on AD P-Value
For technical details, see Appendix: Statistical Details for Nonnormal Distributions and
Transformations. See also Andrew Sleeper, Six Sigma Distribution Modeling, for further
information on these methods.
Note that these transformations and distributions are particularly effective for inherently skewed
data but should not be used with bimodal data or where the nonnormality is due to outliers
(typically identified with a Normal Probability Plot). In these cases, you should identify the reason
for the bimodality or outliers and take corrective action. Another common reason for nonnormal
data is poor measurement discrimination leading to “chunky” data. In this case, attempts should
be made to improve the measurement system.
164
SigmaXL’s default setting is to use the Box-Cox transformation which is the most common approach
to dealing with nonnormal data. Box-Cox is used to convert nonnormal data to normal by applying
a power transformation, Y^lambda, where lambda varies from -5 to +5. You may select rounded or
optimal lambda. Rounded is typically preferred since it will result in a more “intuitive”
transformation such as Ln(Y) (lambda=0) or SQRT(Y) (lambda=0.5). If the data includes zero or
negative values, select Lambda & Threshold. SigmaXL will solve for an optimal threshold which is a
shift factor on the data so that all of the values are positive.
1. Open the file Nonnormal Cycle Time2.xlsx. This contains continuous data of process cycle
times. The Critical Customer Requirement is: USL = 1000 minutes.
2. Let’s begin with a view of the data using Histograms and Descriptive Statistics. Click SigmaXL >
Graphical Tools > Histograms & Descriptive Statistics.
4. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. Click OK.
Clearly this is a process in need of improvement. To start, we would like to get a baseline
process capability. The problem with using regular Capability analysis is that the results will be
incorrect due to the nonnormality in the data. The Histogram and AD P-Value < .05 clearly
show that this data is not normal.
5. We will confirm the nonnormality by using a Normal Probability Plot. Click Sheet 1 Tab (or F4).
Click SigmaXL > Graphical Tools > Normal Probability Plots.
165
7. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. Click OK. A Normal Probability
Plot of Cycle Time data is produced:
2.33
1.83
1.33
0.83
NSCORE
0.33
-0.17
-0.67
-1.17
-1.67
-2.17
-4000
-2000
2000
4000
6000
8000
Cycle Time (Minutes)
8. The curvature in this normal probability plot confirms that this data is not normal.
9. For now, let us ignore the nonnormal issue and perform a Process Capability study assuming a
normal distribution. Click Sheet 1 Tab. Click SigmaXL > Process Capability > Capability
Combination Report (Individuals).
10. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. Enter USL = 1000; delete
previous Target and LSL settings.
166
11. Click OK. The resulting Process Capability Report is shown below:
167
Notice the discrepancy between the Expected Overall (Theoretical) Performance and Actual
(Empirical) Performance. This is largely due to the nonnormality in the data, since the expected
performance assumes that the data is normal. So why not just use the actual performance and
disregard the expected? This would not be reliable because the sample size, n = 30, is too small to
estimate performance using pass/fail (discrete) criteria.
Also note that the process appears to be out-of-control on both the individuals and moving range
charts.
12. We will now perform a process capability analysis using the Capability Combination Report for
Nonnormal Individuals. Click Sheet 1 Tab (or F4). Click SigmaXL > Process Capability >
Nonnormal > Capability Combination Report (Individuals Nonnormal). Ensure that the entire
data table is selected. If not, check Use Entire Data Table. Click Next.
13. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. Enter USL = 1000. We will use
the default selection for Transformation/Distribution Options: Box-Cox Transformation with
Rounded Lambda. Check Tests for Special Causes as shown:
168
169
The AD Normality P-Value Transformed Data value of 0.404 confirms that the Box-Cox
transformation to normality was successful. The process capability indices and expected
performance can now be used to establish a baseline performance. Note that there are no out-of-
control signals on the control charts, so the signals observed earlier when normality was assumed
were false alarms.
The Individuals – Original Data chart displays the untransformed data with control limits calculated
as:
UCL = 99.865 percentile
CL = 50th percentile
LCL = 0.135 percentile
The benefit of displaying this chart is that one can observe the original untransformed data. Since
the control limits are based on percentiles, this represents the overall, long term variation rather
than the typical short term variation. The limits will likely be nonsymmetrical.
The Individuals/Moving Range – Normalized Data chart displays the transformed z-values with
control limits calculated using the standard Shewhart formulas for Individuals and Moving Range
charts. The benefit of using this chart is that tests for special causes can be applied and the control
limits are based on short term variation. The disadvantage is that one is observing transformed
data on the chart rather than the original data.
Automatic Best Fit
Now we will redo the capability analysis using the Automatic Best Fit option.
15. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Select Automatic Best Fit as
shown:
170
16. Click OK. The resulting Process Capability Combination report is shown below. Please note that
due to the extensive computations required, this could take up to 1 minute (or longer for large
datasets):
171
The 2 Parameter Loglogistic distribution was selected as the best fit distribution. For details on
how this selection was made, see Appendix: Statistical Details for Nonnormal Distributions
and Transformations.
The Anderson Darling statistic for the Loglogistic distribution is 0.245 which is less than the 0.37
value for the AD Normality test of the Box-Cox transformation indicating a better fit. (Note that
published AD P-Values for this distribution are limited to a maximum value of 0.25. The best fit
selection uses a P-Value estimate that is obtained by transforming the data to normality and
then using a modified Anderson Darling Normality test on the transformed data).
Distribution Fitting
Another helpful tool to evaluate transformations and distributions is Distribution Fitting.
1. Click Sheet 1 Tab (or F4). Click SigmaXL > Process Capability > Nonnormal > Distribution
Fitting. Ensure that the entire data table is selected. If not, check Use Entire Data Table. Click
Next.
2. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. We will use the default
selection for Transformation/Distribution Options: All Transformations & Distributions as
shown:
172
3. Click OK. The resulting Distribution Fitting report is shown below. Please note that due to the
extensive computations required, this could take up to 1 minute (or longer for large datasets):
173
The distributions and transformations are sorted in descending order using the AD Normality P-
Value on the transformed z-score values. Note that the first distribution shown may not be the
selected “best fit”, because the best fit procedure also looks for models that are close but with
fewer parameters.
The reported AD P-Values are those derived from the particular distribution. The AD P-Value is not
available for distributions with a threshold (except Weibull), so the AD Normality P-Value on the
transformed z-score values is used (labeled as Z-Score Est.).
Since the sort order is based on the AD P-Values from Z-Score estimates, it is possible that the
reported distribution based AD P-Values may not be in perfect descending order. However any
discrepancies based on sort order will likely not be statistically or practically significant.
Some data will have distributions and transformations where the parameters cannot be solved
(e.g., 2-parameter Weibull with negative values). These are excluded from the Distribution Fitting
report.
The parameter estimates and percentile report includes a confidence interval as specified in the
Distribution Fitting dialog, with 95% being the default. Note that the wide intervals here are due
to the small sample size, n = 30.
The control limits for the percentile based Individuals chart will be the 0.135% (lower control limit),
50% (center line, median) and 99.865% (upper control limit). Additional percentiles may be entered
in the Distribution Fitting dialog.
After reviewing this report, if you wish to perform a process capability analysis with a particular
transformation or distribution, simply select Specify Distribution from the
Transformation/Distribution Options in the Capability Combination Report (Individuals -
Nonnormal) dialog as shown below (using 2 Parameter Loglogistic):
174
175
This is a standalone tool that allows you to visually see how the Box-Cox transformation selects a
rounded or optimal lambda value.
1. Open the file Nonnormal Cycle Time2.xlsx. Select Sheet 1 Tab.
2. Click SigmaXL > Process Capability > Nonnormal > Box-Cox Transformation (or SigmaXL > Data
Manipulation > Box-Cox Transformation or SigmaXL > Control Charts > Nonnormal > Box-Cox
Transformation). Ensure that the entire data table is selected. If not, check Use Entire Data
Table. Click Next.
3. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. The selected variable must
contain all positive values.
Tip: Note that while this tool is often successful to transform the data to normality, there may
not be a suitable transformation to make the data normal. The output report indicates the
Anderson-Darling P-Value for the transformed data. You may wish to check Do not store if
transformed data is not normal. Another option is Do not store transformed data if Lambda =
1 falls within 95%CI. This latter option prevents you from using transformations that do not
result in a statistically significant improvement towards normality.
176
4. Click OK. The resulting report is shown:
5. The fact that Lambda=1 falls outside of the 95% confidence interval tells us that the
transformation is statistically significant. The Anderson-Darling P-Value of 0.4041 indicates that
we cannot reject the null hypothesis that the transformed data is normal, so the Ln
transformation has successfully converted the data to normality.
177
Part K – Reliability/Weibull Analysis
Reliability/Weibull Analysis
The Weibull distribution is a continuous distribution that was publicized by Waloddi Weibull in
1951. It has become widely used, especially in the reliability field. The Weibull distribution's
popularity resulted from its ability to be used with small sample sizes and its flexibility. See
Dodson, B. “The Weibull Analysis Handbook” Second Edition, ASQ, for further information. For
SigmaXL technical details, see Appendix: Statistical Details for Nonnormal Distributions and
Transformations.
The Weibull distribution can also be used as an alternative to the Box-Cox Transformation to
determine Process Capability Indices or Control Limits for nonnormal data.
Weibull is particularly effective when the data are censored, with exact failure times unknown or
suspension of test. Suspension of test is also referred to as right censored. SigmaXL can analyze
complete or right censored data.
1. Open the file Weibull Dodson.xlsx. This contains right censored time-to-failure data from
Dodson, page 28.
2. Click SigmaXL > Reliability/Weibull Analysis. Ensure that the entire data table is selected. Click
Next.
3. Select Time-to-Fail, click Numeric Response (Y) >>. Note that the selected variable must
contain all positive values. Select Censor Code, click Right Censor Column >>. The default
censor value of 1 will be used. We will use the default method of estimation, Maximum
Likelihood (some reliability practitioners prefer Least Squares Regression (X on Y) for small
sample sizes). Other defaults include Threshold = 0, and Confidence Level = 90%. Enter time
values to estimate survival probabilities as shown:
178
4. The resulting Weibull analysis report is shown:
179
Weibull Probability Plot of Time-to-Fail
0.500
0.000
-0.500
WEIBULL SCORE
-1.000
-1.500
-2.000
-2.500
-3.000
-3.500
3.70
3.90
4.10
4.30
4.50
4.70
4.90
5.10
5.30
5.50
Ln (Time-to-Fail)
180
5. The Hollander-Proschan test is used when there are right censored observations (see Dodson).
If the P-Value is less than 0.05, this indicates significant lack of fit, so Weibull would not be an
appropriate distribution. Given the P-Value of 0.9771, there is no lack-of-fit with this time-to-
fail data. This is confirmed in the Weibull probability plot with the data roughly following a
straight line.
6. If the data is uncensored then the Anderson-Darling unadjusted test is used.
7. The shape parameter, also referred to as Beta, is the slope of the line on the Weibull probability
plot. The Weibull distribution can be used in a wide variety of situations and dependent on the
value of Beta, is equal to or can approximate several other distributions. For example, if Beta =
1, the Weibull distribution is identical to the exponential distribution; if Beta = 2.5, the Weibull
distribution approximates the lognormal distribution; and if Beta = 3.6, the Weibull distribution
approximates the normal distribution. Because of this flexibility, there are few observed failure
rates that cannot be accurately modeled by the Weibull distribution. Furthermore if Beta is < 1,
the system is suffering from infant mortality failures; if Beta is = 1, failures are random – they
are occurring at a constant failure rate and are independent of time; if Beta is > 1, components
are wearing out with age.
8. The scale parameter is the characteristic life. This is the point at which we could expect 63.2%
of the population under study to have failed. The survival probability report shows that the
survival probability at time 200 = 0.386. The cumulative failure probability at 200 is 1 – 0.386 =
0.614.
9. The percentile report tells us that we can expect a 50% failure rate (one-half of the population
will fail) at time = 179.73, i.e. the Median = 179.73.
10. The percentile report can also be used to determine control limits for nonnormal data. The
lower control limit is the 0.135 % value = 22.052; the center line is the 50% value = 179.73; the
upper control limit is the 99.865 % value = 383.55.
181
SigmaXL: Analyze Phase Tools

Part A – Stratification with Pareto
SigmaXL’s Pareto tool allows you to create Basic (Single) or Advanced (Multiple) Pareto Charts.
Advanced Pareto charts are particularly useful in the Analyze Phase because of the ease with which
you can slice and dice (or stratify) your data. Of course, Pareto charts are not limited to the
Analyze Phase – they can also be used to aid project selection and to prioritize in the Measure
Phase.
Consider the following guidelines to help ensure that your Pareto analysis is successful:
• Your Pareto analysis will only be as good as the quality of the data collected. Ensure that
you have the right data and that the data is correct. Use other graphs such as run charts to
apply a sanity check to your data.
• Check process stability using appropriate control charts. If the process is not in control,
your prioritization of defects and root causes could be invalid.
• Avoid collecting data over too short a time period. Your data may not be representative of
the process as a whole. Also keep in mind that since the data is discrete, a minimum sample
size of 500 is recommended with 1000 preferred.
• Conversely, data gathered over too long a time period may include process changes that
could lead to incorrect conclusions. SigmaXL provides a date subsetting feature that allows
you to easily explore different time periods.
• If your initial Pareto analysis does not yield useful results, explore other categories that may
be important. SigmaXL’s Advanced Charts makes it easy for you to 'slice and dice' your data
with different X categories.
• Consider Pareto charting measures such as cost and severity, in addition to defect counts.
SigmaXL enables you to chart multiple Y responses.
185
Basic Pareto Chart Template

Click SigmaXL > Templates and Calculators > Basic Graphical Templates > Pareto Chart or
SigmaXL > Graphical Tools > Basic Graphical Templates > Pareto Chart. See Measure Phase
Part B – Templates and Calculators for a Pareto Template example.
Basic (Single) Pareto Charts

1. Open the file Customer Data.xlsx. Click SigmaXL>Graphical Tools >Basic Pareto Chart.
3. Select Major-Complaint, click Pareto Category (X) >>.
Tip: SigmaXL will automatically count the number of unique items in the Pareto Category. If we
had a separate column with a count (or cost), this count column would be selected as the
Optional Numeric Count (Y).
4. Click Next. Set Basic Chart Options as follows:
Tab “Other” Bar/Cum Sum:

Cum Sum Line – On Top of First Bar
186
Tab Chart Options:

Category (X) Font – Slanted
Data Labels – Check Bars, Cum Sum
Check Save Defaults.
Tip: After you have saved your defaults, you can bypass the above options, by clicking Finish
instead of Next at the original Basic Pareto Chart dialog box. The saved defaults will
automatically be applied.
Note: For SigmaXL Mac version, click SigmaXL > Graphical Tools > Basic Pareto Chart Options
to set options. The Basic Pareto Chart is displayed after you click Finish.
5. Click Finish. The Pareto Chart is produced:
187
Advanced (Multiple) Pareto Charts

1. Click Sheet1 Tab of Customer Data.xlsx (or press F4 to activate last worksheet).
2. Click SigmaXL > Graphical Tools > Advanced Pareto Options.

Note that the Sample Charts have nothing to do with the data set being evaluated. They are
used to dynamically illustrate how your options affect the charts to be produced. (The Sample
Charts are not displayed in Excel for Mac).
3. Set Order of Bars to Same Order on the “Other Bar”/Cum Sum options tab.
This is typically used for comparative purposes. The Descending Order option makes each
Chart a true Pareto Chart, but is less useful for comparison.
4. Click Chart Options tab. Set according to choice – in this case we have selected Data Labels for
the Bars but not for the Cum Sum line.
5. Ensure that Save Defaults is checked. Note that these options will be saved and applied to all
Advanced Pareto Charts. Click Finish
6. SigmaXL automatically takes you to the next step of Chart Generation (This is equivalent to
clicking SigmaXL > Graphical Tools > Advanced Pareto Charts). If necessary, check Use Entire
Data Table.
7. Click Next.
188
8. Select Major Complaint, click Pareto Category (X1) >>; select Customer Type, click Group
Category (X2) >>.
9. Click OK. A Pareto Chart of Major Customer Complaints is produced for each Customer Type.
10. Click Sheet 1 Tab, Click SigmaXL > Graphical Tools > Advanced Pareto Charts.
11. Ensure that entire data table is selected. If not, check Use Entire Data Table. Click Next. (Steps
10 and 11 can be bypassed with the Recall SigmaXL Dialog menu or by pressing F3 to recall last
dialog).
189
12. Select Major Complaint, click Pareto Category (X1) >>; select Size of Customer, click Group
Category (X2) >>; select Product Type, click Group Category (X3) >>.
If a Numeric Count (Y) variable is not specified, SigmaXL automatically determines the counts
from the Pareto Category (X1).
Normally we would use a text column of discrete Xs, but be aware that numeric columns are
also allowed. Be careful here – this could easily generate a very large number of charts.
The total number of charts generated = (# of levels in X2) * (# of levels in X3) * (# of levels in X4)
* (# of Y variables). Please note, Group Category (X3) and Group Category (X4)
13. Click OK. Multiple Paretos are generated:
190
Part B - EZ-Pivot
One of the most powerful features in Excel is the Pivot table. SigmaXL’s EZ-Pivot tool simplifies the
creation of Pivot tables using the familiar X and Y dialog box found in the previous Pareto tools.
Example of Three X’s, No Response Y’s

1. Open Customer Data.xlsx, click Sheet 1 (or press F4 to activate last worksheet). Select SigmaXL
> Graphical Tools > EZ-Pivot/Pivot Charts.
3. Select Major Complaint, click Count Category (X1) >>. Note that if Y is not specified, the Pivot
Table Data is based on a count of X1, hence the name Count Category.
4. Select Customer Type, click Group Category (X2) >>; select Size of Customer, click Group
Category (X3) >> as shown.
5. Click OK. Resulting Pivot Table of Major Complaint by Customer Type is shown:
191
6. This Pivot table shows the counts for each Major Complaint (X1), broken out by Customer Type
(X2), for all Sizes of Customers (X3). (Grand Totals can be added to the Pivot Table by using
Pivot Table Toolbar > Table Options. Check Grand Totals for Columns, Grand Totals for Rows).
7. To display counts for a specific Customer Size, click the arrow adjacent to Size of Customer
(All). Select Large.
8. Click OK. Resulting Pivot Table is:
Note that the Major Complaint “Not-Available” is not shown. Pivot tables only show rows
where there is at least a count of one.
9. The Pivot Chart can be seen by clicking the EZ Pivot Chart (1) tab; reset Size of Customer to All
as shown below:
192
Size of Customer (All)
Count of Major-Complaint
70
60
50
40 Customer Type
3
2
30 1
20
10
0
Difficult-to-order Not-available Order-takes-too-long Return-calls Wrong-color
Major-Complaint
10. Drag Size of Customer from Report Filter to Axis Fields (shown under PivotChart Fields to the
right of the chart) and Excel will automatically split the Pivot Chart showing both Large and
Small Customers.
Count of Major-Complaint
40
35
30
25
Customer Type
3
20
2
1
15
10
0
Difficult-to- Order-takes- Return-calls Wrong-color Difficult-to- Not-available Order-takes- Return-calls Wrong-color
order too-long order too-long
Large Small
Size of Customer Major-Complaint
193
Example of Three X’s and One Y

1. Select Sheet 1 of Customer Data.xlsx; click SigmaXL > Graphical Tools > EZ-Pivot/Pivot Charts;
click Next (alternatively, click Recall SigmaXL Dialog menu or press F3 to recall last dialog).
2. Select Customer Type, click Count Category (X1) >>; select Size of Customer, click Group
Category (X2) >>; select Product Type, click Group Category (X3); select Overall Satisfaction,
click Numeric Responses (Y) >>. Note that the Label for X1 changed from Count Category to
Group Category. The Pivot Table data will now be based on Y data.
3. The Response default uses a Sum of Y. This however can be changed to Average or Standard
Deviation. Select Average. Uncheck Create Pivot Charts (Since we are looking at averages, the
stacked bar Pivot Charts would not be very useful, unless they are changed to clustered column
format using Chart > Chart Type).
4. Click OK. The resulting Pivot Table is:
5. Note that the table now contains Averages of the Customer Satisfaction scores (Y). Again
Product Type (X3) can be varied to show Consumer, Manufacturer, or All. Double clicking on
Average of Overall Satisfaction allows you to switch to Standard Deviation (StDev).
194
Example of 3 X’s and 3 Y’s

1. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog.
2. Select Customer Type, click Group Category (X1) >>; select Size of Customer, click Group
Category (X2) >>; select Product Type, click Group Category (X3) >>. Select Avg Days Order to
Delivery, Loyalty – Likely to Respond, Overall Satisfaction, click Numeric Responses (Y) >>.
Select Average and One Pivot Table (default is separate Pivot Tables for each Y). Uncheck
Create Pivot Charts.
3. Click OK. Resulting Pivot Table:
Again, Product Type (X3) can be varied.
195
Part C – Confidence Intervals
One Sample Confidence Interval Templates

Click SigmaXL > Templates and Calculators > Basic Statistical Templates >
1 Sample Z-Test and Confidence Interval for Mean or
1 Sample t-Test and Confidence Interval for Mean or
1 Sample Chi-Square Test and CI for Standard Deviation or
1 Proportion Test and Confidence Interval or
1 Poisson Rate Test and Confidence Interval.
These confidence interval templates are also located at SigmaXL > Statistical Tools > Basic
Statistical Templates.
See Measure Phase Part B – Templates and Calculators for Confidence Interval template
examples.
Confidence Intervals – Descriptive Statistics

1. Open Customer Data.xlsx. Click Sheet 1 Tab (or press F4 to activate last worksheet).
2. Click SigmaXL > Statistical Tools > Descriptive Statistics.
4. Select Overall Satisfaction, click Numeric Data Variables (Y) >>. Select Customer Type, click
Group Category (X1) >>; Confidence Level default is 95%:
196
5. Click OK. Descriptive Statistics are given for Customer Satisfaction grouped by Customer Type:
We are given the 95% confidence interval for each sample Mean (95% CI Mean) as well as the
95% confidence interval for the Standard Deviation (95% CI Sigma – do not confuse this with
Process Sigma Quality Level).
These confidence intervals are very important in understanding our data and making decisions
from the data. How often are we driven by sample estimates only and fail to consider the
confidence interval or margin of error? For example, newspapers will often fail to take into
account the confidence interval when reporting opinion poll results. (To calculate confidence
intervals for discrete proportion data, use SigmaXL > Templates and Calculators > Basic
Statistical Templates > 1 Proportion Test and Confidence Interval).
Note that a confidence interval of 95% implies that, on average, the true population parameter
(Mean, Median, Standard Deviation, or Proportion) will lie within the interval 19 times out of
20.
A confidence interval or margin of error does not take into account measurement error or
survey bias, so the actual uncertainty may be greater than stated. This should be addressed
with good data collection, reliable measurement systems, and good survey design.
Confidence Intervals for the Mean can be obtained in several ways with SigmaXL: Descriptive
Statistics, Histograms & Descriptive Statistics, 1-Sample t-test & Confidence Intervals, One-Way
ANOVA, and Multi-Vari Charts.
To illustrate confidence intervals for the mean of Overall Satisfaction graphically, we have
generated a Multi-Vari Chart (with 95% CI Mean Options) using the Customer Data.xlsx data.
This chart type will be covered later (Part Q).
197
4.5
4
Overall Satisfaction (Mean Options)
3.5
2.5
1.5
1
#1 #2 #3
Customer Type
The dots correspond to individual data points. The tick marks show the 95% upper confidence
limit, mean, and 95% lower confidence limit. Clearly we can see that Customer Type 2 has a
significantly higher level of mean satisfaction; the lower limit does not overlap with the upper
limit for Types 1 and 3. On the other hand, we see overlap of the CI’s when comparing types 1
and 3. Hypothesis testing will now be used to compare the mean satisfaction scores more
precisely and determine statistical significance for the results.
198
Part D – Hypothesis Testing – One Sample Z and t-Test
One Sample Z, t and Equivalence Test Templates

1 Sample Z-Test and Confidence Interval for Mean or
1 Sample t-Test and Confidence Interval for Mean or
1 Sample Equivalence Test for Mean
These templates are also located at SigmaXL > Statistical Tools > Basic Statistical Templates.
See Measure Phase Part B – Templates and Calculators for One Sample Z, t and Equivalence
Test template examples:
Basic Statistical Templates – 1 Sample Z-Test and Confidence Interval for Mean
Basic Statistical Templates – 1 Sample t-Test and Confidence Interval for Mean
Basic Statistical Templates – 1 Sample Equivalence Test for Mean
199
One Sample t-Test

1. Open Customer Data.xlsx, select Sheet 1 tab (or press F4 to activate last worksheet). Click
SigmaXL > Statistical Tools > 1 Sample t-Test & Confidence Intervals. If necessary, check Use
Entire Data Table, click Next.
2. Ensure that Stacked Column Format is selected. Select Overall Satisfaction, click Numeric Data
Variable (Y) >>, select Customer Type, click Optional Group Category (X) >>.
3. Historically, our average customer satisfaction score has been 3.5. We would like to see if this
has changed, with the results grouped by customer type. Null Hypothesis H0: μ=3.5;
Alternative Hypothesis Ha: μ≠3.5
4. Enter 3.5 for the Null Hypothesis H0: Mean value. Set Ha as Not Equal To, Confidence Level =
95.0%, and check Display Test Assumptions Report.
200
6. Note the P-Values. Customer Type 2 shows a significant change (increase) in Satisfaction Mean
(P-Value < .05), whereas Customer Types 1 and 3 show no change (P-Value ≥ .05). Also note the
confidence intervals around each mean match the results from Descriptive Statistics.
7. In the Measure Phase we determined that Overall Satisfaction for Customer Type 2 has
nonnormal data but this does not imply that the P-Value for the 1 Sample t-test is wrong. The
Central Limit Theorem applies here: the distribution of averages tends to be normal, even if the
individual observations are not-normal. With a sample size of 42, the t-test is fairly robust
against skewed data.
8. 1 Sample t-Test Assumptions Report:
This is a text report with color highlight: Green (OK), Yellow (Warning) and Red (Serious
Violation).
Each sample is tested for Normality using the Anderson Darling test. If not normal, the
minimum sample size for robustness of the 1 sample t-Test is determined utilizing Monte Carlo
201
regression equations (see Basic Statistical Templates – Minimum Sample Size for Robust t-
Tests and ANOVA). If the sample size is inadequate, a warning is given and a suitable
Nonparametric Test is recommended (Wilcoxon if symmetric, Sign Test if not symmetric).
Each sample is tested for Outliers defined as: Potential: Tukey's Boxplot (> Q3 + 1.5*IQR or < Q1
– 1.5*IQR); Likely: Tukey's Boxplot 2.2*IQR; Extreme: Tukey's Boxplot 3*IQR. If outliers are
present, a warning is given and recommendation to review the data with a Boxplot and Normal
Probability Plot. Here we have a potential outlier for Customer Type 2.
Tip: If the removal of outlier(s) result in an Anderson Darling P-Value that is > 0.1, a notice is
given that excluding the outlier(s), the sample data are inherently normal.
Each sample is also tested for Randomness using the Exact Nonparametric Runs Test. If the
sample data is not random, a warning is given and recommendation to review the data with a
Run Chart.
See Appendix Hypothesis Test Assumptions Report for further details.
202
Part E – Power and Sample Size
Power and Sample Size – One Sample t-Test – Customer Data

Using the One Sample t-Test (above, Part D), we determined that Customer Types 1 and 3 resulted
in “Fail to reject H0: μ=3.5”. A failure to reject H0 does not mean that we have proven the null to
be true. The question that we want to consider here is “What was the power of the test?”
Restated, “What was the likelihood that given Ha: μ≠3.5 was true, we would have rejected H0 and
accepted Ha?” To answer this, we will use the Power and Sample Size Calculator.
Tip: Typical sample size rules of thumb address confidence interval size and robustness to
normality (e.g. n=30). Computing power is more difficult because it involves the magnitude of
change in mean to be detected, so one needs to use the power and sample size calculator.
Please use the following guidelines when using the power and sample size calculator:
Power >= .99 (Beta Risk is <= .01) is considered Very High Power
Power >= .95 and < .99 (Beta Risk is <= .05) is High Power
Power >= .8 and < .95 (Beta Risk is <= .2) is Medium Power. Typically a power value of .9 to detect
a difference of 1 standard deviation is considered adequate for most applications. If the data
collection is difficult and/or expensive, then .8 might be used.
Power >= .5 and < .8 (Beta Risk is <= .5) is Low Power – not recommended.
Power < .5 (Beta Risk is > .5) is Very Low Power – do not use!
1. Click SigmaXL > Statistical Tools > Power and Sample Size Calculators > 1 Sample t-Test
Calculator. We will only consider the statistics from Customer Type 3 here. We will treat the
problem as a two sided test with Ha: Not Equal To to be consistent with the original test.
2. Enter 27 in Sample Size (N). The difference to be detected in this case would be the difference
between the sample mean and the hypothesized value i.e. 3.6411 – 3.5 = 0.1411. Enter 0.1411
in Difference. Leave Power value blank, with Solve For Power selected (default). Given any
two values of Power, Sample size, and Difference, SigmaXL will solve for the remaining selected
third value. Enter the sample standard deviation value of 0.6705 in Standard Deviation. Keep
Alpha and Ha at the default values as shown:
203
4. A power value of 0.1836 is very poor. It is the probability of detecting the specified difference.
Alternatively, the associated Beta risk is 1-0.1836 = 0.8164 which is the probability of failure to
detect such a difference. Typically, we would like to see Power > 0.9 or Beta < 0.1. In order to
detect a difference this small we would need to increase the sample size. We could also set the
difference to be detected as a larger value.
5. First we will determine what sample size would be required in order to obtain a Power value of
0.9. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Select the Solve For
Sample Size button as shown. It is not necessary to delete the entered sample size of 27 – it
will be ignored. Enter a Power Value of .9:
204
7. A sample size of 240 would be required to obtain a power value of 0.9. The actual power is
rarely the same as the desired power due to the restriction that the sample size must be an
integer. The actual power will always be greater than or equal to the desired power.
8. Now we will determine what the difference would have to be to obtain a Power value of 0.9,
given the original sample size of 27. Click Recall SigmaXL Dialog menu or press F3 to recall last
dialog. Select the Solve For Difference button as shown:
205
10. A difference of 0.435 would be required to obtain a Power value of 0.9, given the sample size of
27.
206
Power and Sample Size – One Sample t-Test – Graphing the

Relationships between Power, Sample Size, and Difference
In order to provide a graphical view of the relationship between Power, Sample Size, and
Difference, SigmaXL provides a tool called Power and Sample Size with Worksheets. Similar to the
Calculators, Power and Sample Size with Worksheets allows you to solve for Power (1 – Beta),
Sample Size, or Difference (specify two, solve for the third). You must have a worksheet with
Power, Sample Size, or Difference values. Other inputs such as Standard Deviation and Alpha can
be included in the worksheet or manually entered.
1. Open the file Sample Size and Difference Worksheet.xlsx, select the Sample Size & Diff sheet
tab. Click SigmaXL > Statistical Tools > Power & Sample Size with Worksheets > 1 Sample t-
Test. If necessary, check Use Entire Data Table. Click Next.
2. Ensure that Solve For Power (1 – Beta) is selected. Select Sample Size (N) and Difference
columns as shown. Enter the Standard Deviation value of 1. Enter .05 as the Significance Level
value:
207
3. Click OK. The output report is shown below:
4. To create a graph showing the relationship between Power, Sample Size and Difference, click
SigmaXL > Statistical Tools > Power & Sample Size Chart. Check Use Entire Data Table. Click
Next.
5. Select Power (1 – Beta), click Y Axis (Y); select Sample Size (N), click X Axis (X1); select
Difference, click Group Category (X2). Click Add Title. Enter Power & Sample Size: 1 Sample t-
Test:
6. Click OK. Click OK. The resulting Power & Sample Size Chart is displayed:
208
Part F – One Sample Nonparametric Tests
Introduction to Nonparametric Tests

Nonparametric tests make fewer assumptions about the distribution of the data compared to
parametric tests like the t-Test. Nonparametric tests do not rely on the estimation of parameters
such as the mean or the standard deviation. They are sometimes called distribution-free tests.
Nonparametric tests use Medians and Ranks, thus they are robust to outliers in the data. If,
however, the data are normal and free of outliers, nonparametric tests are less powerful than
normal based tests to detect a real difference when one exists.
Nonparametric tests should be used when the data are nonnormal, the data cannot be readily
transformed to normality, and the sample size is too small for robustness due to the Central Limit
Theorem.
Exact Nonparametric Tests

Nonparametric tests do not assume that the sample data are normally distributed, but they do
assume that the test statistic follows a Normal or Chi-Square distribution when computing the
“large sample” or “asymptotic” P-Value. The One-Sample Sign Test, Wilcoxon Signed Rank, Two
Sample Mann-Whitney and Runs Test assume a Normal approximation for the test statistic.
Kruskal-Wallis and Mood’s Median use Chi-Square to compute the P-Value. With very small sample
sizes (see recommendations below), these approximations may be invalid, so exact methods should
be used. SigmaXL computes the exact P-Values utilizing permutations and fast network algorithms.
It is important to note that while exact P-Values are “correct,” they do not increase (or decrease)
the power of a small sample test, so they are not a solution to the problem of failure to detect a
change due to inadequate sample size.
For data that require more computation time than specified, Monte Carlo P-Values provide an
approximate (but unbiased) P-Value that typically matches exact to two decimal places using
10,000 replications. One million replications give a P-Value that is typically accurate to three
decimal places. A confidence interval (99% default) is given for the Monte Carlo P-Values. Note
that the Monte Carlo confidence interval for P-Value is not the same as a confidence interval on
the test statistic due to data sampling error. The confidence level for the hypothesis test statistic is
still 95%, so all reported P-Values less than .05 will be highlighted in red to indicate significance.
The 99% Monte Carlo P-Value confidence interval is due to the uncertainty in Monte Carlo
sampling, and it becomes smaller as the number of replications increases (irrespective of the data
209
sample size). The Exact P-Value will lie within the stated Monte Carlo confidence interval 99% of
the time.
Recommended Sample Sizes for use of Exact Nonparametric Tests
Sign Test: N <= 50
Wilcoxon Signed Rank: N <= 15
Mann-Whitney: Each sample N <= 10
Kruskal-Wallis: Each sample N <= 5
Mood’s Median: Each sample N <= 10
Runs Test (Above/Below) or Runs Test (Up/Down) Test: N <= 50
These are sample size guidelines for when exact nonparametric tests should be used rather than
“large sample” asymptotic based on the Normal or Chi-Square approximation. It is always
acceptable to use an exact test, but computation time can become an issue especially for tests with
two or more samples. In those cases, one can always use a Monte Carlo P-Value with 99%
confidence interval.
Validation of SigmaXL Exact P-Values
SigmaXL Exact P-Values are validated by comparison to textbook examples, published paper
examples and other exact software such as StatXact, SPSS Exact, SAS, and Matlab.
Monte Carlo P-Values are validated using 1e6 replications and compared against exact. Repeated
simulations are used to validate the confidence intervals.
For further details and references refer to the Appendix Exact and Monte Carlo P-Values for
Nonparametric and Contingency Tests.
210
One Sample Sign Test

The sign test is the simplest of the nonparametric tests, and is similar to testing if a two-sided coin
is fair. Count the number of positive values (larger than hypothesized median), the number of
negative values (smaller than the hypothesized median), and test whether there are significantly
more positives (or negatives) than expected. The One Sample Sign Test is a nonparametric
equivalent to the parametric One Sample t-Test.
Historically, our Median customer satisfaction score has been 3.5. We would like to see if this has
changed, with the results grouped by customer type (H0: Median=3.5, Ha: Median≠3.5, α = 0.05).
1. Open Customer Data.xlsx, select Sheet 1 tab. Click SigmaXL > Statistical Tools >
Nonparametric Tests>1 Sample Sign. If necessary, check Use Entire Data Table, click Next.
Variable (Y) >>; select Customer Type, click Optional Group Category (X) >>.
3. Enter 3.5 for the Null Hypothesis H0: Median value. Set Ha as Not Equal To.
211
Note the P-Values. Customer Type 2 shows a significant change (increase) in Satisfaction
Median (P-Value < .05), whereas Customer Types 1 and 3 show no change (P-Value ≥ .05).
While the P-Values are not the same as those given by the 1 sample t-Test, the conclusions do
match.
If Count (N) is less than or equal to 50, the Sign Test computes an exact P-Value using the
binomial distribution. For N > 50, the P-Value is estimated using a normal approximation. Since
this is always done automatically and is very fast, the Sign Test is not included in the separate
Nonparametric Exact menu.
One Sample Wilcoxon Signed Rank Test

The Wilcoxon Signed Rank test is a more powerful nonparametric test than the Sign Test, but it
adds an assumption that the distribution of values is symmetric around the median. An example of
a symmetric distribution is the uniform distribution. Symmetry can be observed with a histogram,
or by checking to see if the Skewness is large (> .5 or < - .5).
The One Sample Wilcoxon Test is a nonparametric equivalent to the parametric One Sample t-Test.
Historically, our Median customer satisfaction score has been 3.5. We would like to see if this has
changed, with the results grouped by customer type (H0: Median=3.5, Ha: Median≠3.5, α = 0.05).
SigmaXL > Statistical Tools > Nonparametric Tests > 1 Sample Wilcoxon. If necessary, check
Use Entire Data Table, click Next.
3. Enter 3.5 for the Null Hypothesis H0: Median value. Keep Ha as Not Equal To.
212
Note the P-Values. Customer Type 2 shows a significant change (increase) in Satisfaction
Median (P-Value < .05), whereas Customer Types 1 and 3 show no change (P-Value ≥ .05).
Although the P-Values are not identical to the sign test and t-Test, the conclusions match.
(Note, in the case of Customer Type 2, the Sign Test is preferred since the data is not
symmetrical but skewed).
One Sample Wilcoxon Signed Rank Test - Exact

We will now redo the above example to compute exact P-Values. Typically this would not be
necessary unless the sample sizes were smaller (N <= 15 for Wilcoxon), but this gives us continuity
on the example. We will consider a small sample problem later.
SigmaXL > Statistical Tools > Nonparametric Tests – Exact > 1 Sample Wilcoxon - Exact. If
necessary, check Use Entire Data Table, click Next.
213
3. Enter 3.5 for the Null Hypothesis H0: Median value. Keep Ha as Not Equal To.
The Wilcoxon Statistics are identical to the above “large sample” or “asymptotic” results. The
Exact P-Values are close but slightly different. This was expected because the sample sizes are
reasonable (N > 15), so the “large sample” P-Values are valid using a normal approximation for
the Wilcoxon Statistic.
Note, if Count (N) is greater than 1000, the Exact P-Value is estimated using a continuity -
corrected normal approximation. Since the Wilcoxon Exact P-Value is computed very quickly
for sample sizes as large as 1000, Monte Carlo P-Values are not required.
5. Now we will consider a small sample problem. Open Nonnormal Task Time Difference – Small
Sample.xlsx. A study was performed to determine the effectiveness of training to reduce the
time required to complete a short but repetitive process task. Fifteen operators were randomly
selected and the difference in task time was recorded in seconds (after training – before
training). A negative value denotes that the operator completed the task in less time after
training than before.
214
6. Click SigmaXL > Graphical Tools > Histograms & Descriptive Statistics. If necessary, check Use
Entire Data Table. Click Next. Select Difference (Seconds), click Numeric Data Variables (Y) >>.
Click OK.
4 Difference (Seconds)
Count = 15
3 Mean = -7.067
Stdev = 12.842
Range = 35.00
2 Minimum = -25
25th Percentile (Q1) = -20
50th Percentile (Median) = -2
1 75th Percentile (Q3) = 6
Maximum = 10
95% CI Mean = -14.18 to 0.05

0 95% CI Sigma = 9.40 to 20.25
10.0
-25.0
-20.0
-15.0
-10.0
0.0
5.0
-5.0
Difference (Seconds) Anderson-Darling Normality Test:

SigmaXLChartSheet A-Squared = 0.841433; P-Value = 0.0227
This small sample data fails the Anderson Darling Normality Test (P-Value = .023). Note that
this is due to the data being uniform or possibly bimodal, not due to a skewed distribution.
Now we will perform a 1 Sample t-Test and review the assumptions.
7. Select Task Time Difference tab (or press F4 to activate last worksheet). Click SigmaXL >
Statistical Tools > 1 Sample t-Test & Confidence Intervals. If necessary, check Use Entire Data
Table, click Next.
8. Ensure that Stacked Column Format is selected. Select Difference (Seconds), click Numeric
Data Variable (Y) >>.
9. This is a one sided test because we have no reason to expect that the task time will increase, so
the Null Hypothesis H0: μ = 0; and the Alternative Hypothesis Ha: μ < 0. Enter 0 for the H0:
Mean value. Select Ha as Less Than, Confidence Level = 95.0%, and check Display Test
Assumptions Report.
215
10. Click OK.
The 1 Sample t-Test Assumptions Report highlights that the data are not normal, but note that
Kurtosis equal to -1.84 is the issue here, not Skewness. This was observed in the Histogram
above with the data being uniform or possibly bimodal.
The sample size is too small for a robust t-Test, so the Exact One Sample Wilcoxon Test is
recommended. The Wilcoxon Test is recommended over the Sign Test because it is a more
powerful test and meets the requirement that the data be symmetrical. The Exact test is
recommended because the sample size is very small (N <= 15).
216
Statistical Tools > Nonparametric Tests – Exact > 1 Sample Wilcoxon - Exact. If necessary,
check Use Entire Data Table, click Next.
Data Variable (Y) >>. Enter 0 for the H0: Median value. Select Ha as Less Than.
With the P-Value = .0497 we reject H0 and conclude that the Median Task Time Difference is
significantly less than 0, so the training is effective.
By way of comparison we will now rerun the analysis using the large sample (asymptotic)
Wilcoxon test.
Statistical Tools > Nonparametric Tests > 1 Sample Wilcoxon. If necessary, check Use Entire
Data Table, click Next.
Data Variable (Y) >>. Enter 0 for the H0: Median value. Select Ha as Less Than.
217
Now with the P-Value = .0513 we incorrectly fail to reject H0.
The difference between exact and large sample P-Value is small but it was enough to lead us to
falsely conclude that the training is ineffective.
In conclusion, whenever you have a small sample size and are performing a nonparametric test,
always use the Exact option.
218
Part G – Two Sample t-Test
Two Sample t and Equivalence Test Templates

2 Sample t-Test and Confidence Interval (Compare 2 Means) or
2 Sample Equivalence Test (Compare 2 Means).
These templates are also located at SigmaXL > Statistical Tools > Basic Statistical Templates.
See Measure Phase Part B – Templates and Calculators for Two Sample t-Test template
examples:
Basic Statistical Templates – 2 Sample t-Test and Confidence Interval (Compare 2 Means)
Basic Statistical Templates – 2 Sample Equivalence Test (Compare 2 Means)
219
Two Sample t-Test

1. Open Customer Data.xlsx, click Sheet 1 tab (or press F4 to activate last worksheet). We will
look at comparing means of Customer Satisfaction by Customer Type (2 vs. 3), using the Two
Sample t-test. H0: μ2=μ3, Ha: μ2≠μ3
2. Click SigmaXL > Statistical Tools > 2 Sample t-test. If necessary, check Use Entire Data Table,
click Next.
3. With stacked column format checked, select Overall Satisfaction, click Numeric Data Variable Y
>>; select Customer Type, click Group Category X >>; H0: Mean Diff = 0; Ha: Not Equal To;
Confidence Level: 95%; ensure that Assume Equal Variances and Display Test Assumption
Report are checked:
220
4. Click OK. Select Customer Type 2 and 3.
5. Click OK. Resulting output:
Given the P-Value of .0007 we reject H0 and conclude that Mean Customer Satisfaction is
significantly different between Customer type 2 and 3. This confirms previous findings.
221
6. 2 Sample t-Test Assumptions Report:
Violation).
Each sample is tested for Normality using the Anderson-Darling test. If not normal, the
minimum sample size for robustness of the 2 sample t-Test is determined utilizing Monte Carlo
Tests and ANOVA). If the sample size is inadequate, a warning is given and the Nonparametric
Mann-Whitney Test is recommended.
Each sample is tested for Randomness using the Exact Nonparametric Runs Test. If the sample
data is not random, a warning is given and recommendation to review the data with a Run
Chart.
222
A test for Equal Variances is also applied. If all sample data are normal, the F-Test is utilized,
otherwise Levene’s Test is used. If the variances are unequal and the test being used is the
equal variance option, then a warning is given and Welch’s test is recommended.
223
Paired t-Test
1. Open the file Dietcola.xlsx. These are the results of a Before and After taste test on sweetness
for diet cola. Ten tasters were used and one month elapsed with the cola in warm storage
between the before and after results. Do a one sample t-test on the column of differences.
2. Click SigmaXL > Statistical Tools > 1-Sample t-test & Confidence Intervals. Ensure that entire
data table is selected. If not, check Use Entire Data Table. Click Next. Select Difference, click
Numeric Data Variable (Y) >>, set H0: Mean = 0, Ha: Less Than (this is a one-sided or one-tail
test – sweetness cannot increase). Uncheck Display Test Assumptions Report for now:
3. Click OK. Result:
Given the P-Value of .012, we reject H0 and conclude that the sweetness has in fact decreased.
4. Now redo the analysis using the paired t-test: Click Sheet 1 Tab; Click SigmaXL > Statistical
Tools > Paired t-Test; Click Next; Select After, click Numeric Data Variable 1 >>, select Before as
224
Numeric Data Variable 2 >>, H0: Mean Diff = 0, Ha: Less Than. Check Display Test
Assumptions Report.
5. Click OK. Results are identical to the One sample t-test analysis of difference column, with the
added assumptions report showing that all assumptions are met:
225
Unpaired 2 Sample t-Test vs. Paired t-Test

1. Open the Dietcola.xlsx file, click the Sheet 1 tab (or press F4 to activate last worksheet).
2. Click SigmaXL > Statistical Tools > 2 Sample t-Test. Ensure that entire data table is selected. If
not, check Use Entire Data Table. Click Next.
3. Check Unstacked Column Format. Select After, Before and click Numeric Data Variables (Y) >>.
H0: Mean Diff = 0, Ha: Less Than; check Assume Equal Variances. Uncheck Display Test
Assumptions Report:
Now the P-Value is .07, indicating a fail to reject H0. What changed? Hint: Compare the SE
Mean of the Paired t-test to the Std Error Difference of the unpaired two-sample t-test. Where
226
does the additional variability come from in the two-sample t-test? The paired t-test is the
appropriate test to use here.
Power & Sample Size for 2 Sample T-Test

To determine Power & Sample Size for a 2 Sample t-Test, you can use the Power & Sample Size
Calculator or Power & Sample Size with Worksheet.
1. Click SigmaXL > Statistical Tools > Power & Sample Size Calculators > 2 Sample t-Test
Calculator.
2. Select Solve For Power (1 – Beta). Enter Sample Size and Difference as shown:
Note that we are calculating the power or likelihood of detection given that Mean1 – Mean2 =
1, with sample size for each group = 30, standard deviation = 1, significance level = .05, and Ha:
Not Equal To (two-sided test).
3. Click OK. The resulting report is displayed:
A power value of 0.97 is good, hence we have the basis for the “minimum sample size n=30”
rule of thumb used for continuous data.
227
4. To determine Power & Sample Size using a Worksheet, click SigmaXL > Statistical Tools >
Power & Sample Size with Worksheet > 2 Sample t-Test.
5. A graph showing the relationship between Power, Sample Size and Difference can then be
created using SigmaXL > Statistical Tools > Power & Sample Size Chart. See Part E for an
example using the 1 Sample t-Test.
228
Part H – Two Sample F and Comparison Test
Two Sample F-Test Template

2 Sample F-Test and CI (Compare 2 Standard Deviations) or SigmaXL > Statistical Tools > Basic
Statistical Templates > 2 Sample F-Test and CI (Compare 2 Standard Deviations) to access the
2 Sample F-Test calculator. The template gives the following default example.
229
Two Sample Comparison Test

We will now do a full comparison test of Customer Satisfaction for Customer Type 1 and 2. This
test checks each sample for normality, equal variance (F-test and Levene’s), 2 sample t-test
(assuming equal and unequal variance), and Mann Whitney test for equal Medians. Depending on
the normality, variance, and sample size results, the appropriate P-Values are highlighted in yellow.
1. Open Customer Data.xlsx, click on Sheet 1 Tab.
2. Click SigmaXL > Statistical Tools > 2 Sample Comparison Tests. Click Next. Check Stacked
Column Format. Select Overall Satisfaction, click Numeric Data Variable (Y) >>, select
Customer Type, click Group Category (X) >>.
230
4. Click OK.
Customer Type 2 has nonnormal data. This makes Levene’s test the appropriate test for
unequal variance. Levene’s test indicates that Customer type 2 has a significantly lower
variance, or standard deviation. The lower standard deviation translates to a consistent level of
satisfaction.
Since Levene’s test indicates unequal variance, the appropriate t-test assumes unequal
variance. The t-test indicates that Customer Type 2 has a significantly higher mean satisfaction.
Clearly the next step would be to determine a root cause or best practices to reduce the
variability in overall satisfaction and increase the mean for all customer types.
231
Part I – Two Sample Nonparametric Test:

Mann-Whitney
Two Sample Mann-Whitney Test

We will look at comparing medians of Customer Satisfaction by Customer Type, using the Two
Sample Mann-Whitney test. (H0: Median2 = Median3, Ha: Median2 ≠ Median3.) The Two
Sample Mann-Whitney Test is the nonparametric equivalent to the parametric Two Sample t-
Test. The test does not assume sample normality but does assume that the samples have equal
shapes. If the shapes are different, the null hypothesis is that the distributions are the same.
1. Open Customer Data.xlsx, click Sheet 1 tab (or press F4 to activate last worksheet).
2. Click SigmaXL > Statistical Tools > Nonparametric Tests > 2 Sample Mann-Whitney. If
3. With Stacked Column Format checked, select Overall Satisfaction, click Numeric Data Variable
(Y) >>; select Customer Type, click Group Category (X) >>; and Ha: Not Equal To:
232
Given the P-Value of .0008 we reject H0 and conclude that Median Customer Satisfaction is
significantly different between Customer types 2 and 3. This confirms previous findings and
matches the results of the 2 Sample t-Test.
Two Sample Mann-Whitney Test - Exact

We will now redo the above example to compute exact P-Values. Typically, this would not be
necessary unless the sample sizes were smaller (each sample N <= 10 for Mann-Whitney), but this
gives us continuity on the example. Due to the large number of permutations, we will use Monte
Carlo Exact for this analysis and will consider a small sample problem later.
2. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > 2 Sample Mann-Whitney –
Exact. If necessary, check Use Entire Data Table, click Next.
(Y) >>; select Customer Type, click Group Category (X) >> and Ha: Not Equal To. Select Monte
Carlo Exact with the default Number of Replications = 10000 and Confidence Level for P-Value
= 99%.
233
Tip 1: If Exact is selected and the computation time limit is exceeded, a dialog will prompt you
to use Monte Carlo or to increase the computation time.
Tip 2: 10,000 replications will result in a Monte Carlo P-Value that is typically correct to two
decimal places. One million (1e6) replications will result in three decimal places of accuracy
and typically require less than 60 seconds to solve for any data set.
Tip 3: The Monte Carlo 99% confidence interval for P-Value is not the same as a confidence
234
4. Click OK. Select Customer Type 2 and 3. Click OK. Results:
Given the Monte Carlo P-Value of .0004 we reject H0 and conclude that Median Customer
Satisfaction is significantly different between Customer types 2 and 3. The Mann Whitney
Statistic is very close to the above “large sample” or “asymptotic” result. This was expected
because the sample size is reasonable (each sample N > 10), so the “large sample” P-Values are
valid using a normal approximation for the Mann-Whitney Statistic.
The Monte Carlo P-Value 99% confidence interval is 0.0000 to 0.0009. Note that the Monte
Carlo P-Value will be slightly different every time it is run (the Monte Carlo seed value is derived
from the system clock). This was demonstrated using 10,000 replications, but with a P-Value
this low, it is recommended that the number of replications be increased to 1e5 or 1e6 to get a
better estimate.
5. Now we will consider a small sample problem. Open Stimulant Test.xlsx. This data is from:
Narayanan, A. and Watts, D. “Exact Methods in the NPAR1WAY Procedure,” SAS Institute Inc.,
Cary, NC. http://support.sas.com/rnd/app/stat/papers/exact.pdf
Researchers conducted an experiment to compare the effects of two stimulants. Thirteen

randomly selected subjects received the first stimulant, and six randomly selected subjects
received the second stimulant. The reaction times are in minutes. We will test the null
hypothesis of no difference between the medians of the two stimulants against the alternative
that stimulant 1 has smaller median reaction time than stimulant 2.
6. Select Reaction Time tab. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > 2
Sample Mann-Whitney - Exact. If necessary, check Use Entire Data Table, click Next.
7. With Stacked Column Format checked, select Reaction Time, click Numeric Data Variable (Y)
>>; select Stimulant, click Group Category (X) >>; and Ha: Less Than. Select Exact with the
default Time Limit for Exact Computation = 60 seconds.
235
8. Click OK. Select Stimulant 1 and 2.
This sets the order for the one sided test, so the alternative hypothesis Ha is Median 1 <
Median 2.
With the P-Value = .0527 we fail to reject H0, so cannot conclude that there is a difference in
median reaction times. This exact P-Value matches that given in the reference paper.
236
By way of comparison we will now rerun the analysis using the “large sample” or “asymptotic”
Mann-Whitney test.
10. Select Reaction Time tab (or press F4 to activate last worksheet). Click SigmaXL > Statistical
Tools > Nonparametric Tests > 2 Sample Mann-Whitney. If necessary, check Use Entire Data
Table, click Next.
11. With Stacked Column Format checked, select Reaction Time, click Numeric Data Variable (Y)
>>; select Stimulant, click Group Category (X) >>; and Ha: Less Than.
12. Click OK. Select Stimulant 1 and 2. Click OK. Results:
Now with the P-Value = .0421 we incorrectly reject H0.
The difference between exact and large sample P-Value is small but it was enough to lead us to
falsely conclude that stimulant 1 resulted in a reduced median reaction time.
237
Part J – One-Way ANOVA & Means Matrix
One-Way ANOVA & Means Matrix

1. One-Way ANOVA and Means Matrix allows you to quickly do multiple pairwise comparisons.
The One-Way ANOVA tests H0: μ1 = μ2 = μ3; Ha: at least one pairwise set of means are not
equal.
2. Open Customer Data.xlsx, click on Sheet 1 tab (or press F4 to activate last worksheet).
3. Click SigmaXL > Statistical Tools > One-Way ANOVA & Means Matrix. Ensure that the entire
data table is selected. If not, check Use Entire Data Table.
4. Click Next. Ensure that Stacked Column Format is checked. Select Overall Satisfaction, click
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>. Uncheck
Display ANOVA Table Details. Check Display Test Assumptions Report.
5. Click OK. The results are shown below:
238
6. The ANOVA P-Value of 0.0000 tells us that at least one pairwise set of means are not equal.
From the Pairwise Mean Difference (Means Matrix), we conclude that Mean Overall
Satisfaction is significantly different between Customer Type 2 and 3, as well as 1 and 2. Note
that the default probabilities are Fisher Pairwise. The P-Values will be slightly different than the
previous 2 Sample t-Test results because the variances from all 3 customer types are “pooled”
here. This also results in slightly different confidence intervals. See below for more details on
the multiple comparison options.
7. A graphical view of the Overall Satisfaction Mean and 95% Confidence Intervals are given to
complement the Means Matrix. The fact that the CI’s for Customer Type 2 do not overlap those
of Type 1 or 3, clearly shows that Customer Type 2 has a significantly higher mean satisfaction
score. The overlap of CI’s for Type 1 and 3 shows that the mean satisfaction scores for 1 and 3
are not significantly different.
8. Note that the Confidence Level shown in the graph can be modified by clicking the Options
button and setting Confidence Level. This also changes the alpha value used to highlight the P-
Values in red:
239
alpha = (100 – Confidence Level)/100
so the default Confidence Level = 95.0%, gives an alpha = 0.05.
9. The R-Square (R-Sq) value of 20.95% indicates that Customer Type “explains” approximately
21% of the variation in Overall Satisfaction. We need to “drill down” to understand the root
causes and best practices associated with Customer Type 2.
10. One-Way ANOVA Assumptions Report:
Violation).
Each sample is tested for Normality using the Anderson Darling test. If not normal, the
minimum sample size for robustness of the ANOVA Test is determined utilizing Monte Carlo
Tests and ANOVA). If the sample size is inadequate, a warning is given and the appropriate
Nonparametric test is recommended (Kruskal-Wallis if there are no extreme outliers, Mood’s
Median if there are extreme outliers).
Chart.
240
A test for Equal Variances is also applied. If all sample data are normal, Bartlett’s Test is
utilized, otherwise Levene’s Test is used. If the variances are unequal, then a warning is given
and Welch’s ANOVA is recommended.
One-Way ANOVA & Means Matrix - Options

Notes on Means Matrix Probability Method (Multiple Comparison of Means) and ANOM Chart:
• Fisher:
o Also known as Fisher’s Least Significant Difference (LSD)
o Pairwise 2 sample t-tests with pooled standard deviation
o Does not correct for family-wise error rate, so should only be used for k = 3 means
and in the restricted case where the ANOVA P-Value is < alpha (this is also known as
Protected Fisher LSD). For k = 3 means, Protected Fisher LSD is more powerful than
Tukey.
• Tukey:
o Similar to LSD, uses pairwise tests with pooled standard deviation, but is a
studentized range statistic that corrects for family-wise error rate. Recommended
for k > 3.
• Dunnett with Control:
o If one of the groups are a control reference group, Dunnett with Control is more
powerful than Tukey because it is doing fewer pairwise comparisons (only considers
those pairwise against the control group).
o Uses pooled standard deviation and a multivariate t distribution that corrects for
family-wise error rate.
• See Appendix Multiple Comparison of Means and Variances

(a.k.a. Post-Hoc Tests) for further details and references.
• Display ANOM Normal One-Way Chart:
o The ANOM alpha = (100 – Confidence Level)/100.
o This chart is also available at SigmaXL > Graphical Tools > Analysis of Means
(ANOM). See Part P – Analysis of Means (ANOM) Charts.
241
1. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Uncheck Display Test Assumptions
Report. Click the Options button. Select Tukey. Check Display Residual Charts and Display
ANOM Normal One-Way Chart; Confidence Level = 95.0 as shown:
Note: The Confidence Level is used to set the level in the Mean/Confidence Interval Plot, the
alpha level (alpha = (100 – CI)/100) used to highlight the P-Values and the alpha level for the
ANOM chart. However, the confidence level used in the Residuals Normal Probability Plot is
always 95%.
2. Click OK. The Pairwise Means Difference (Means Matrix) and Tukey Probability results are:
242
3. The significant Tukey Probability values above have not changed as compared to Fisher, but
note that they are larger than Fisher to compensate for the family-wise error rate:
4. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Click Options. Uncheck all Display
options and select Dunnett with Control Level = 2 (i.e. we are treating Customer Type 2 as the
Control Group):
243
5. Click OK. The Pairwise Means Difference (Means Matrix) and Dunnett Probability results are:
All three mean differences are shown but only two probability values are calculated: Customer
Type 1 – Control Type 2 and Control Type 2 – Type 3. (Typically one would show 1 – 2 and 3 – 2
for Dunnett, but we are displaying the upper triangle for consistency with the other options).
Note that the Dunnett 2 – 3 P-Value is .003 which is smaller than the Tukey .004 but larger than
the Fisher .002. Dunnett is more powerful than Tukey (due to having fewer tests), but less
powerful than Fisher. It does, however, have an advantage over Fisher because it protects the
family-wise error rate for the comparisons being made.
244
6. Click the ANOVA Residuals sheet tab to display the Residual Plots (created at steps 1, 2):
Residuals are the unexplained variation from the ANOVA model (Actual – Predicted/Fitted
values). We expect to see the residuals approximately normally distributed with no obvious
patterns in the above graphs, which is the case here.
245
7. Click ANOM_Normal_One_Way sheet tab to display the ANOM Chart (created at steps 1, 2):
Here we see that Customer Type 1 mean satisfaction score is significantly below the overall
mean and Customer Type 2 is significantly higher. This is consistent with the results that are
observed in the the Pairwise Means Difference (Means Matrix), but is easier to interpret.
The varying decision limits are due to the varying sample sizes for each Customer Type, with
smaller sample size giving wider limits in a manner similar to a control chart. If the data are
balanced the decision limit lines will be constant.
246
Power & Sample Size for One-Way ANOVA

To determine Power & Sample Size for a One-Way ANOVA, you can use the Power & Sample Size
Calculator or Power & Sample Size with Worksheet.
1. Click SigmaXL > Statistical Tools > Power & Sample Size Calculators > One-Way ANOVA
Calculator.
2. Select Solve For Power (1 – Beta). Enter Sample Size and Maximum Difference as shown:
Note that we are calculating the power or likelihood of detection given that the maximum
difference between group means = 1, with sample size for each group = 30, 3 groups, standard
deviation = 1, significance level = .05, and Ha: Not Equal To (two sided test).
A power value of 0.94 is acceptable. Note that this value is less than the power value of 0.97
obtained with the two-sample t-Test.
247
4. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Change the number of groups to 4.
Click OK. Note that the power value is 0.907. If the number of groups (levels) increases, you
will have to increase the sample size in order to maintain statistical power.
Power & Sample Size with Worksheet > One-Way ANOVA.
6. A graph showing the relationship between Power, Sample Size and Maximum Difference can
then be created using SigmaXL > Statistical Tools > Power & Sample Size Chart. See Part E for
an example using the 1 Sample t-Test.
248
Part K – Two-Way ANOVA
Two-Way ANOVA
Two-Way ANOVA tests the following:
H0 (Factor X1): μ1 = μ2 = …= μk
Ha (Factor X1): at least one pairwise set of means are not equal (μi ≠ μj);
H0 (Factor X2): μ1 = μ2 = …= μk
Ha (Factor X2): at least one pairwise set of means are not equal (μi ≠ μj);
H0 (Interaction): There is no interaction between factors X1 and X2

Ha (Interaction): There is an interaction between factors X1 and X2.
A Two-Way ANOVA analysis will typically have balanced data from a designed experiment, with an
equal number of observations for each combination level of X1 and X2. SigmaXL will also
accommodate unbalanced data. The minimum requirement is one observation per combination
level of X1 and X2. An error message will be produced if this minimum requirement is not met.
2. Click SigmaXL > Statistical Tools > Two-Way ANOVA. Ensure that the entire data table is
selected. If not, check Use Entire Data Table.
3. Click Next. Select Avg No of Orders per Mo, click Numeric Data Variable (Y) >>; select Customer
Type, click Group Category Factor (X1) >>; select Size of Customer, click Group Category Factor
(X2) >>. Remove Interaction (Fit Additive Model) should remain unchecked.
Tip: If the Two-Way ANOVA report gives a P-Value for the interaction that is high (> 0.1), you
should then press F3 or click Recall SigmaXL Dialog to recall this dialog, check Remove
Interaction (Fit Additive Model) and rerun the analysis.
Display ANOM Normal Two-Way Chart should be unchecked. We will discuss this tool later in
Part P – Analysis of Means (ANOM) Charts. Note that this option is only available when
Remove Interaction (Fit Additive Model) is unchecked because ANOM Normal Two-Way
always includes the interaction term in the model. The ANOM alpha is (100 – Confidence
Level)/100).
249
250
5. Scroll down to view the Main Effects (with confidence intervals) and Interaction Plots:
Note that the mean values shown are fitted (predicted least squares) means not data means.
This compensates for unbalanced data.
6. From the ANOVA table, we can see that the Size of Customer term is significant with a P-Value
less than .05. Customer Type is not significant by itself, but the Interaction term is significant.
This indicates that the effect of Size of Customer on Average Number of Orders per Month
depends on Customer Type (we could also equivalently say that the effect of Customer Type
depends on Size of Customer).
7. This is also confirmed looking at the Main Effects and Interaction plots. Customer Type by itself
is not significant. Size of Customer is obviously significant. Looking at the Interaction plot, the
different slopes illustrate that the change in Average Number of Orders per Month across
Customer Types depends on Customer Size, albeit this is a relatively small effect.
251
8. Click the Two-Way Residuals sheet to view the residual graphs:

Histogram of Residuals for: Avg No. of orders per mo Normal Probability Plot of Avg No. of orders per mo
25 3.00
2.00
20
1.00
15
NSCORE
Frequency
0.00
10
-1.00
5 -2.00
-3.00
0
00
0
00
00
00
00
0
0
00
00
00
00
00
00
2
.1
.1
.0
.6
.7
.7
.8
.0
5
.0
.0
.0
.0
8.
5.
2.
5.
0.
5.
0.
2.
5.
8.
0.
5.
11
14
17
-9
-6
-3
-0
-5
10
15
20
25
-1
-1
-1
-2
-2
-1
-1
Residuals Residuals
Residuals vs Data Order for: Avg No. of orders per mo Residuals vs Predicted Values for: Avg No. of orders per mo
20.00 20.00
15.00 15.00
10.00 10.00
5.00 5.00
Residuals
Residuals
0.00 0.00
-5.00 -5.00
-10.00 -10.00
-15.00 -15.00
-20.00
-20.00
-25.00
-25.00
0
.0
.0
.0
.0
.0
1
1
11
21
31
41
51
61
71
81
91
10
24
29
34
39
44
Observation Order Fitted Values
Residuals vs Customer Type

Response: Avg No. of orders per mo
16.47
11.47
6.47
Residuals
1.47
-3.54
-8.54
-13.54
-18.54
1
3
2
8
1.
1.
1.
1.
2.
2.
2.
2.
Customer Type
Residuals are the unexplained variation from the ANOVA model (Actual – Predicted/Fitted
values). We expect to see the residuals approximately normally distributed with no obvious
patterns in the above graphs, which is the case here.
The Residuals versus Size of Customer graph is not shown because Size of Customer is text. In
order to display this plot, Size could be coded numerically with 1 = Small and 2 = Large. Simply
create a new column called Size-Coded and use the following Excel formula to create the coded
values for the first record:
=IF(K2="Small",1,IF(K2="Large",2))
Copy and Paste this formula to obtain coded values for all 100 records. Rerun the Two-Way
ANOVA analysis to create the residual graphs
252
Part L – Tests for Equal Variance &

Welch’s ANOVA
Bartlett’s Test
Bartlett’s Test is similar to the 2 sample F-Test (SigmaXL > Statistical Tools > 2 Sample Comparison
Test) but allows for multiple group comparison of variances (or standard deviations). Like the F-
Test, Bartlett’s requires that the data from each group be normally distributed but is more
powerful than Levene’s Test.
Notes on Multiple Comparison of Variances Probability Method and ANOM Chart:
• F-Test Pairwise
o Pairwise 2 sample F-tests
o Does not correct for family-wise error rate, so should only be used for k = 3 groups
and in the restricted case where the Bartlett P-Value is < alpha.
• F-Test with Bonferroni Correction
o Pairwise 2 sample F-tests with Bonferroni correction
o Recommended for k > 3
o Bonferroni P-Value’ = P-Value * m
 m = number of pairwise comparisons k(k-1)/2

• Display ANOM Variances Chart:
253
1. Open Delivery Times.xlsx, click on Sheet 1 tab.
2. Click SigmaXL > Statistical Tools > Equal Variance Tests > Bartlett. Ensure that the entire data
table is selected. If not, check Use Entire Data Table.
3. Click Next. Ensure that Stacked Column Format is checked. Select Delivery Time Deviation,
click Numeric Data Variable (Y) >>; select Floor, click Group Category (X) >>.
5. All 10 Anderson-Darling Test P-Values are > .05 indicating that all group data are normal. Since
the assumption of normality is met, Bartlett’s is the appropriate test to use. If any one of the
groups have a low P-Value for the Normality test, then Levene’s test should be used.
6. With the P-Value = 0.63 we fail to reject H0; we do not have evidence to show that the group
variances are unequal (practically speaking we will assume that the variances are equal).
7. If the Equal Variance test is only being used to test the assumption for use in ANOVA then it is
not necessary to examine the Multiple Comparison of Variances. However, in the context of a
process improvement project, we often do want to know which groups are significantly
different. This can give us important clues to identify opportunities for variance reduction.
8. With a “Fail-to-Reject H0” it is unnecessary to review the Multiple Comparison of Variances or

the ANOM Variances Chart, but we will do so here for demonstration purposes.
254
9. The default Multiple Comparison of Variances is a matrix of F-Test Pairwise Probabilities:
10. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Click Options. Check Display ANOM
Variances Chart.
Note: The Confidence Level determines the alpha level (alpha = (100 – CI)/100) used to
highlight the P-Values and the alpha level for the ANOM chart. However, the alpha level used
to highlight P-Values in the Anderson-Darling Normality Test is always 0.05.
11. We will not run F-Test with Bonferroni Correction in this example, but typically that would be
used when there are more than 3 groups.
255
12. Click OK. Click ANOM_Variances sheet tab to display the ANOM Chart:
The ANOM Variances chart visually shows that none of the group standard deviations are
significantly different from the grand mean of all the standard deviations. It is called an ANOM
Variances Chart but displays Standard Deviations for ease of interpretation (similar to a
Standard Deviation S Control Chart). This does, however, result in non-symmetrical decision
limits. The ANOM Variances chart in SigmaXL > Graphical Tools > Analysis of Means (ANOM) >
ANOM Variances has an option to display Variances.
256
Levene’s Test
Levene’s Test for multiple group comparison of variances is less powerful that Bartlett’s Test, but is
robust to the assumption of normality. (This is a modification of the original Levene’s Test,
sometimes referred to as the Browne-Forsythe Test that uses Absolute Deviations from the Median
or ADM).
Notes on Multiple Comparison of Variances Probability Method and ANOM Chart:
• Levene Pairwise
o Pairwise 2 sample Levene tests
o Does not correct for family-wise error rate, so should only be used for k = 3 groups
and in the restricted case where the Levene P-Value is < alpha.
• Tukey ADM (Absolute Deviations from the Median)
o Application of Tukey on ADM
o Recommended for k > 3
o This post-hoc test is unique to SigmaXL, inspired by the method used in ANOM
Levene Variances.

• Display ANOM Levene Robust Variances Chart:
257
1. Open Customer Data.xlsx, click on Sheet 1 tab.
2. Click SigmaXL > Statistical Tools > Equal Variance Tests > Levene. Ensure that the entire data
table is selected. If not, check Use Entire Data Table.
3. Click Next. Ensure that Stacked Column Format is checked. Select Responsive to Calls, click
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>.
5. The Levene’s Test P-Value of 0.0144 tells us that we reject H0. At least one pairwise set of
variances are not equal. The normality test P-Values indicate that all 3 groups have nonnormal
data (P-Values < .05). Since Levene’s Test is robust to the assumption of normality, it is the
correct test for equal variances (rather than Bartlett’s Test).
6. If the Equal Variances test is only being used to test the assumption for use in ANOVA then it is
not necessary to examine the Multiple Comparison of Variances. However, in the context of a
258
process improvement project, we often do want to know which groups are significantly
different. This can give us important clues to identify opportunities for variance reduction.
7. The default Multiple Comparison of Variances is a matrix of Levene Pairwise Probabilities:
Customer Type 1 versus Customer Type 2 shows a significant difference in variance.
Press F3 or click Recall SigmaXL Dialog to recall last dialog. Click Options. Select Tukey ADM
(Absolute Deviations from Median). Check Display ANOM Levene Robust Variances Chart.
Note: The Confidence Level determines the alpha level (alpha = (100 – CI)/100) used to
highlight the P-Values and the alpha level for the ANOM chart. However, the alpha level used
to highlight P-Values in the Anderson-Darling Normality Test is always 0.05.
8. Click OK. The Multiple Comparison of Variances is a matrix of Tukey ADM Probabilities:
259
The 1 – 2 P-Value is significant but larger than the Levene Pairwise because it adjusts for the
family-wise error rate. Note that it is smaller than the Bonferroni corrected value = .0036 * 3 =
.011, so more powerful than Bonferroni. The difference in power between Tukey and
Bonferroni becomes more prominent with a larger number of groups, so Bonferroni is not
included as an option.
9. Click the ANOM_Levene sheet tab to display the ANOM Chart:
10. The ANOM chart clearly shows Customer Type 1 has significantly higher variance (ADM) than
overall and Customer Type 2 has significantly lower variance.
11. The varying decision limits are due to the varying sample sizes for each Customer Type, with
balanced, the decision limit lines will be constant.
12. Now that we have determined that the variances are not equal, we are presented with a
problem if we want to test for equal group means. Classical ANOVA assumes that the group
variances are equal, so should not be used. A modified ANOVA called Welch’s ANOVA is robust
to the assumption of equal variances and will be demonstrated next.
260
Welch’s ANOVA Test (Assume Unequal Variance)

Welch’s ANOVA is a test for multiple comparison of means. It is a modified One-Way ANOVA that
is robust to the assumption of equal variances. Welch’s ANOVA is an extension of the 2 sample t-
test for means, assuming unequal variance. Nonparametric methods could also be used here but
they are not as powerful as Welch’s ANOVA.
Notes on Means Matrix Probability Method (Multiple Comparison of Means):
• Welch Pairwise:
o Pairwise 2 sample t-tests with unpooled standard deviation and weighted degrees of
freedom (2 sample t-test for unequal variance)
o Does not correct for family-wise error rate, so should only be used for k = 3 means
and in the restricted case where the Welch ANOVA P-Value is < alpha.
• Games-Howell:
o Similar to Welch Pairwise, uses unpooled standard deviation and weighted degrees
of freedom, but is a studentized range statistic that corrects for family-wise error
rate. Recommended for k > 3.
o It is an extension of the Tukey test, but does not assume equal variance.

• ANOM Chart for Welch ANOVA is not included in SigmaXL since it requires two-stage
sampling (see Nelson, Wludyka, Copeland, 2005, Chapter 8, Heteroscedastic Data)
References for Analysis of Means (ANOM) Charts.
261
2. Click SigmaXL > Statistical Tools > Equal Variance Tests > Welch’s ANOVA (Assume Unequal
Variance) or SigmaXL > Statistical Tools > Welch’s ANOVA (Assume Unequal Variance). Ensure
that the entire data table is selected. If not, check Use Entire Data Table.
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>. Check
Display Test Assumptions Report.
262
5. The P-Value for Welch’s ANOVA is 0.0135, therefore we reject H0 and conclude that the group
means for Responsive to Calls are not equal.
6. From the Pairwise Mean Difference (Means Matrix), we conclude that Mean Responsive to Calls
is significantly different between Customer Type 1 and 2. Note that the default probabilities
are Welch Pairwise. See below for more details on the multiple comparison options.
8. A graphical view of the Responsive to Calls Mean and 95% Confidence Intervals are given to
complement the Means Matrix. Note that the standard deviations are unpooled, resulting in
different CI widths for each group. The fact that the CI’s for Customer Type 1 do not overlap
those of Type 2, visually shows that there is a significant difference in mean Responsive to Calls.
The overlap of CI’s for Type 2 and 3 shows that the mean scores for 2 and 3 are not significantly
different.
9. Later, we will explore the relationship between Responsive to Calls and Overall Satisfaction.
263
10. Welch’s ANOVA Assumptions Report:
Violation).
Each sample is tested for Normality using the Anderson-Darling test. If not normal, the
minimum sample size for robustness of the ANOVA Test is determined utilizing Monte Carlo
Tests and ANOVA). If the sample size is inadequate, a warning is given and the appropriate
Nonparametric test is recommended (Kruskal-Wallis if there are no extreme outliers, Mood’s
Median if there are extreme outliers).
Probability Plot
Tip: If the removal of outlier(s) result in an Anderson-Darling P-Value that is > 0.1, a notice is
Chart.
A test for Equal Variances is also applied. If all sample data are normal, Bartlett’s Test is
utilized, otherwise Levene’s Test is used. Since we are using Welch’s ANOVA, it is confirmed as
appropriate.
264
11. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Uncheck Display Test Assumptions
Report. Click the Options button. Select Games-Howell. Check Display Residual Charts as
shown:
Note: The Confidence Level is used to set the level in the Mean/Confidence Interval Plot and
the alpha level (alpha = (100 – CI)/100) used to highlight the P-Values. However, the confidence
level used in the Residuals Normal Probability Plot is always 95%.
12. Click OK. The Pairwise Means Difference (Means Matrix) and Games-Howell Probability results
are:
265
13. The significant Games-Howell Probability values above have not changed as compared to Welch
Pairwise, but note that they are larger to compensate for the family-wise error rate:
Note also that the 1 – 2 Games-Howell probability is smaller than the Bonferroni corrected
value = .0045 * 3 = .0135, so more powerful than Bonferroni. The difference in power between
Games-Howell and Bonferroni becomes more prominent with a larger number of groups, so
Bonferroni is not included as an option.
266
14. Click the Welch Residuals sheet tab to display the Residual Plots:
Residuals are the unexplained variation from the ANOVA model (Actual – Predicted or Fitted
values). Note that the residuals are not normally distributed - as expected from the
assumptions report - but like the regular ANOVA, Welch’s ANOVA is quite robust to the
assumption of normality. Also, as expected from Levene’s test for equal variances, the
variability for Type 2 is less than Type 1, but Welch’s ANOVA is robust to the assumption of
equal variances, so we can trust that the P-Values are valid.
267
Part M – Nonparametric Multiple Comparison
Kruskal-Wallis
The Kruskal-Wallis test is an extension of the Mann-Whitney Rank test, allowing for more than
2 samples. It is a nonparametric equivalent to the parametric One-Way ANOVA. The Null
Hypothesis is: H0: Median1 = Median2 = … = MedianK. Ha: At least two Medians are different.
The test does not assume sample normality but does assume that the samples have equal
shapes. If the shapes are different, the null hypothesis is that the distributions are the same.
2. Click SigmaXL > Statistical Tools > Nonparametric Tests > Kruskal-Wallis. Ensure that the
entire data table is selected. If not, check Use Entire Data Table.
268
Click on cell B16 to view the P-Value with more decimal place precision (or change the cell format
to scientific notation). The P-Value of 0.000023 (2.3 e-5) tells us that we reject H0. At least one
pairwise set of medians are not equal.
5. The Kruskal-Wallis Statistic is based on comparing mean ranks for each group versus the mean
rank for all observations. The Z value for Customer Type 3 is −1.56, the smallest absolute Z-
value. This size indicates that the mean rank for Type 3 differed least from the mean rank for
all observations. The Z value for Customer Type 2 is 4.53, the largest absolute Z-value. This size
indicates that the mean rank differed most from the mean rank for all observations.
269
6. A graphical view of the Overall Satisfaction Median and 95% Confidence Intervals are given to
complement the Z scores. The fact that the CI’s for Customer Type 2 do not overlap those of
Type 1 or 3, clearly shows that Customer Type 2 has a significantly higher median satisfaction
score. The overlap of CI’s for Type 1 and 3 shows that the median satisfaction scores for 1 and
3 are not significantly different.
270
Kruskal-Wallis – Exact
We will now redo the above example to estimate the exact P-Value using Monte Carlo. Typically
this would not be necessary unless the sample sizes were smaller (each sample N <= 5 for Kruskal-
Wallis), but this gives us continuity on the example. We will consider a small sample problem later.
Computing an exact P-Value for Kruskal-Wallis is very computationally intensive. The Network
Model by Mehta and Patel cannot be used for this test (see Appendix Exact and Monte Carlo P-
Values for Nonparametric and Contingency Tests). In this example, the total number of
permutations are:
(31+42+27)! / (31! * 42! * 27!) = 7.42 E44
(i.e., more than the number of stars in the observable universe). So, we will not attempt to
compute the exact, but rather use Monte Carlo!
2. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > Kuskal-Wallis – Exact. If
3. Ensure that Stacked Column Format is checked. Select Overall Satisfaction, click Numeric Data
Variable (Y) >>; select Customer Type, click Group Category (X) >>. Select Monte Carlo Exact
with the Number of Replications = 1e6 and Confidence Level for P-Value = 99%. One million
replications are used because the expected P-Value is very small as estimated from the “large
sample” Kruskal-Wallis above. This will take up to a minute to run, so if you have a slow
computer, use 1e5 replications instead of 1e6.
271
4. Click OK.
Click on cell B16 to view the P-Value with more decimal place precision (or change the cell
format to scientific notation). The Monte Carlo P-Value here is 0.000009 (9 e-6) with a 99%
confidence interval of .000002 (2 e-6) to 0.000016 (1.6 e-5). This will be slightly different every
time it is run (the Monte Carlo seed value is derived from the system clock). So we reject H0: at
least one pairwise set of medians are not equal.
Note that the large sample (asymptotic) P-Value of 2.3 e-5 lies outside of the Monte Carlo exact
confidence interval.
5. Now we will consider a small sample problem. Open Snore Study.xlsx. This data is from:
Gibbons, J.D. and Chakraborti, S. (2010). Nonparametric Statistical Inference (5th Edition). New
York: Chapman & Hall, (Example 10.2.1 data, page 347; Example 10.4.2 analysis, pp. 360 – 362).
An experiment was conducted to determine which device is the most effective in stopping
snoring or at least in reducing it. Fifteen men who are habitual snorers were divided randomly
into three groups to test the devices. Each man’s sleep was monitored for one night by a
machine that measures amount of snoring on a 100-point scale while using a device.
272
6. Select Snore Study Data tab. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact >
Kruskal-Wallis - Exact. If necessary, check Use Entire Data Table, click Next.
7. With Unstacked Column Format checked, select Device A, Device B and Device C, click Numeric
Data Variables (Y) >>. Select Exact with the default Time Limit for Exact Computation = 60
seconds.
With the Exact P-Value = 0.0042 we reject H0, and conclude that there is a significant difference
in median snore study scores. This exact P-Value matches that given in the reference textbook
using SAS and StatXact.
273
By way of comparison we will now rerun the analysis using the “large sample” or “asymptotic”
Kruskal-Wallis test.
9. Select Snore Study Data tab (or press F4 to activate last worksheet). Click SigmaXL > Statistical
Tools > Nonparametric Tests > Kruskal-Wallis. If necessary, check Use Entire Data Table, click
Next.
Data Variables (Y) >>.
With the P-Value = .0118 we reject H0 (using alpha = .05), but note that if we were using alpha
= 0.01, we would have incorrectly failed to reject the null hypothesis. This “large sample” P-
Value matches that given in the reference textbook using Minitab.
274
Mood’s Median Test

Mood’s Median Test is an extension of the One Sample Sign Test, using Chi-Square as the test
statistic. Like the Kruskal-Wallis test, Mood's median test can be used to test the equality of
medians from multiple samples. It provides a nonparametric alternative to the one-way analysis of
variance. The Null Hypothesis is: H0: Median1 = Median2 = … = MedianK. Ha: At least two
Medians are different.
Mood's median test is more robust to outliers than the Kruskal-Wallis test, but is less powerful in
the absence of outliers. You should first look at your data with Boxplots. If there are extreme
outliers, then Mood’s Median should be used rather than Kruskal-Wallis.
2. Click SigmaXL > Statistical Tools > Nonparametric Tests > Mood’s Median Test. Ensure that
the entire data table is selected. If not, check Use Entire Data Table.
275
The P-Value of 0.0012 tells us that we reject H0. At least one pairwise set of medians are not
equal.
5. A graphical view of the Overall Satisfaction Median and 95% Confidence Intervals are given.
This is the same graph provided in the Kruskal-Wallis test report.
276
Mood’s Median Test - Exact

We will now redo the above example to compute exact P-Values. Typically this would not be
necessary unless the sample sizes were smaller (each sample N <= 10 for Mood’s Median), but this
gives us continuity on the example. We will consider a small sample problem later. Unlike Kruskal-
Wallis, computing an exact P-Value is efficient due to the reduction of the data to a Chi-Square
table and the use of a Network Model (see Appendix Exact and Monte Carlo P-Values for
Nonparametric and Contingency Tests)
2. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > Mood’s Median Test – Exact.
If necessary, check Use Entire Data Table, click Next.
(Y) >>; select Customer Type, click Group Category (X) >>. Select Exact with the default Time
Limit for Exact Computation = 60 seconds. Set Values Equal to Overall Median: to Counted as
Below.
Tip: If none of the observations are equal to the overall median, the options for “Values Equal
to Overall Median:” do not affect the exact P-Value. That is the case with this Customer Data
example. In cases where there are observations equal to the overall median, then the exact P-
Value will be different for each option selected. For regular Mood’s Median, SigmaXL uses
Counted as Below (in agreement with Minitab) but will use Counted as Above for small sample
cases if the average cell expected value is higher. StatXact uses Counted as Below for Mood’s
Median. Matlab uses Not Counted. Users should try Counted as Below and Counted as Above
to ensure that the P-Values agree on reject or fail to reject H0. Use of this option will be
demonstrated in the small sample example.
277
Tip: If the exact computation time limit is exceeded a dialog will prompt you to use Monte Carlo
or to increase the computation time. When this occurs, Monte Carlo is recommended.
Click on cell B17 to view the Exact P-Value with more decimal place precision Given the P-Value
of .00124 we reject H0 and conclude that at least one pairwise set of medians are not equal.
The Exact P-Value is very close to the above “large sample” or “asymptotic” result (.00121). This
was expected because the sample size is reasonable (N > 10), so the “large sample” Mood’s
Median P-Values are valid using a chi-square approximation. Note also that the counts in B8:D9
form a two way contingency table and the expected counts are all greater than 5.
The Exact P-Value was computed in seconds, but if the data set was larger, the required
computation time could become excessive, and Monte Carlo would be required.
5. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Select Monte Carlo Exact with the
Number of Replications = 1e6 and Confidence Level for P-Value = 99%. One million
replications are used because the expected P-Value is small as estimated from the “large
sample” Mood’s Median above. This will take up to a minute to run, so if you have a slow
computer, use 1e5 replications instead of 1e6.
278
6. Click OK.
279
Click on cell B17 to view the Monte Carlo P-Value with more decimal place precision. The
Monte Carlo P-Value here is .001249 with a 99% confidence interval of .00117 (B19) to .00133
(B18). This will be slightly different every time it is run (the Monte Carlo seed value is derived
from the system clock). The true Exact P-Value = .00124 lies within this confidence interval.
So we reject H0: at least one pairwise set of medians are not equal.
7. Now we will consider the small sample problem used in Kruskal-Wallis. Open Snore Study.xlsx.
This data is from:
Gibbons, J.D. and Chakraborti, S. (2010). Nonparametric Statistical Inference (5th Edition). New
York: Chapman & Hall, (Example 10.2.1, pp. 347-348).
An experiment was conducted to determine which device is the most effective in stopping
snoring or at least in reducing it. Fifteen men who are habitual snorers were divided randomly
into three groups to test the devices. Each man’s sleep was monitored for one night by a
machine that measures amount of snoring on a 100-point scale while using a device.
8. Select Snore Study Data tab. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact >
Mood’s Median Test - Exact. If necessary, check Use Entire Data Table, click Next.
Data Variables (Y) >>. Select Exact with the default Time Limit for Exact Computation = 60
seconds.
280
With the Exact P-Value = .0676 we fail to reject H0, and cannot conclude that there is a
significant difference in median snore study scores. This exact P-Value matches that given in
the reference textbook using StatXact. Comparing Mood’s Median to Kruskal-Wallis we see
that the Exact P-Value is higher because Mood’s Median is not as powerful as Kruskal-Wallis.
We will now rerun the analysis using the Counted as Above option.
11. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Set Values Equal to Overall
Median: to Counted as Above.
281
With the Exact P-Value = 0.3007 we fail to reject H0, and cannot conclude that there is a
significant difference in median snore study scores. Note that setting Values Equal to Overall
Median to Counted as Above resulted in a different count for Device C, which results in the
dramatic difference in Exact P-Value.
In conclusion, when using Mood’s Median Exact, always try Counted as Below and Counted as
Above to ensure that the P-Values agree with each other.
282
Part N – Nonparametric Runs Test
Nonparametric Runs Test for Randomness

The nonparametric runs test provides a test for randomness or independence. The null hypothesis
is H0: The data is random (or independent). The alternative hypothesis is Ha: The data is not
random (or independent). Note that this test is also provided as an option in Run Charts (SigmaXL
> Graphical Tools > Run Chart). In addition to providing an overall test for randomness, 4 tests are
performed to detect Clustering, Mixtures, Trends, and Oscillations. If any of these patterns are
significant (typically using α = 0.01), we would need to take corrective action before proceeding
with further statistical analysis. (Note that SigmaXL will highlight any P-Values < .05 in red.)
2. Click SigmaXL > Statistical Tools > Nonparametric Tests > Runs Test. Ensure that the entire
data table is selected. If not, check Use Entire Data Table.
3. Click Next. Select Overall Satisfaction, click Numeric Data Variable (Y) >>:
283
5. With all of the P-Values being greater than 0.01, we fail to reject H0, and conclude that the data
is random (or statistically independent). Recall from the run chart of this data that there were
no obvious trends or patterns.
Nonparametric Runs Test for Randomness - Examples

Examples of Clustering, Mixtures, Trends, and Oscillations are given below using the Run Chart to
illustrate. The data for these examples are given in the file Runs Test Example Data.xlsx. (Use
SigmaXL > Graphical Tools > Run Charts to create the run charts).
a. Clustering appears as a group of points in one area of the chart. It may indicate
special cause variation such as sampling or measurement problems.
284
b. Mixtures appear as an absence of data points near the center line. A mixture may
indicate a bimodal distribution due to a regular change of shift, machinery, or raw
materials.
Note that the P-Value for Mixtures = 1 – P-Value for Clustering. They are mutually
exclusive. The P-Value for Lack of Randomness = 2 * minimum of (P-Value
Clustering, P-Value Mixtures).
c. Trends appear as an upward or downward drift in the data and may be due to
special causes such as tool wear.
285
d. Oscillations appear as rapid up/down fluctuations indicating process instability.
Note that the P-Value for Trends = 1 – P-Value for Oscillation. They are mutually
exclusive.
286
Nonparametric Runs Test for Randomness - Exact

The Nonparametric Runs test does not assume that the sample data are normally distributed, but it
does assume that the test statistic follows a Normal distribution when computing the “large
sample” or “asymptotic” P-Value. With a small sample size (N <= 50), this approximation may be
invalid, so exact methods should be used. SigmaXL computes the exact P-Values utilizing
permutations.
trend due to inadequate sample size.
Clustering, Mixtures and Lack of Randomness (Runs Above/Below)
If Count (N) is greater than 1000, the exact P-Value is estimated using a continuity-corrected
normal approximation. Since the Runs Test Exact P-Value is computed very quickly for sample sizes
as large as 1000, Monte Carlo P-Values are not required.
Trends and Oscillation (Runs Up/Down)
Exact P-Values are derived from published tables, given the sample size and the number of
up/down runs. The exact tables apply to N <= 50. If N > 50, a continuity-corrected normal
approximation is used.
Exact P-Values are not available in Run Charts.
For further details and references refer to the Appendix Exact and Monte Carlo P-Values for
Nonparametric and Contingency Tests.
We will now redo the Customer Data Overall Satisfaction example to compute exact P-Values.
Typically this would not be necessary unless the sample sizes were smaller (N <= 50), but this gives
us continuity on the example. We will also consider the above small sample examples later.
1. Open Customer Data.xlsx, click Sheet 1 tab.
2. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > Runs Test – Exact. If
3. Select Overall Satisfaction, click Numeric Data Variables (Y) >>. Set Values Equal to Median: to
Counted as Below.
287
Note on Options for Clustering, Mixtures and Lack of Randomness (Runs Above/Below)
If none of the observations are equal to the sample median, the options for “Values Equal to
Median:” do not affect the exact P-Value. That is the case with this Customer Data example. In
cases where there are observations equal to the sample median, then the exact P-Value will be
different for each option selected. For regular Runs Test, SigmaXL uses Counted as Below (in
agreement with Minitab). StatXact uses Counted as Above for the Runs Test. Matlab uses Not
Counted. Users should try Counted as Below and Counted as Above to ensure that the P-
Values agree on reject or fail to reject H0.
Note on Options for Trends and Oscillation (Runs Up/Down)
The Values Equal to Median options also apply to the Runs Up/Down Test: zeros of the first
order differences (i.e., two consecutive values that are the same). Counted as Below denotes
“counted as negative.” Counted as Above denotes “counted as positive.” Not Counted
denotes that zero first order differences are deleted. This consolidation of options was done to
keep the dialog simple.
Missing values are ignored for all exact runs tests.
Use of these options will be demonstrated in the small sample example.
288
With all of the P-Values being greater than 0.01 (alpha = .01 is preferred for the Runs Test to
minimize false alarms), we fail to reject H0, and conclude that the data is random (or
statistically independent).
The Above/Below and Up/Down Runs counts are identical to the above “large sample” or
“asymptotic” results. The Clustering/Mixtures/Randomness (Above/Below) Exact P-Values are
close but slightly different. This was expected because the sample size is reasonable (N > 50), so
the “large sample” P-Values are valid using a normal approximation.
The P-Values for Trends and Oscillation (Up/Down) use a normal approximation because the
sample size is greater than 50. They are, however, slightly different than the “large sample”
above because a continuity correction is now applied to the normal approximation.
5. Now we will consider the small sample examples. Open Runs Test Example Data.xlsx, click
Runs Test Example Data tab.
6. Click SigmaXL > Statistical Tools > Nonparametric Tests – Exact > Runs Test – Exact. If
7. Select Clustering and shift click to Oscillation, click Numeric Data Variables (Y) >>. Set Values
Equal to Median: to Counted as Below.
289
The Exact P-Values are close to the above “large sample” or “asymptotic” results, but note that
some of the values are now greater than .01 so they would fail-to reject H0. Note that this does
not imply that the large sample runs test is more powerful, but rather we cannot conclude
some of the previously identified patterns at the 99% confidence level.
We will now rerun the analysis using the Counted as Above option.
9. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Set Values Equal to Median: to
Counted as Above.
290
Note the change in Trends and Oscillation Number of Runs and resulting dramatic change to
the Exact P-Values.
In conclusion, when using Runs Test Exact, always try Counted As Below and Counted as Above
to ensure that the P-Values agree with each other. Also, whenever you have a small sample size
and are performing a nonparametric test, always use Exact.
291
Part O – Attribute/Discrete Data Tests
1 Proportion Test and Confidence Interval Template

1 Proportion Test and Confidence Interval or SigmaXL > Statistical Tools > Basic Statistical
Templates > 1 Proportion Test and Confidence Interval.
See Basic Statistical Templates – 1 Proportion Test and Confidence Interval for an example of the
1 Proportion Test and Confidence Interval template.
2 Proportions Test and Confidence Interval Template

2 Proportions Test and Confidence Interval or SigmaXL > Statistical Tools > Basic Statistical
Templates > 2 Proportions Test and Confidence Interval.
See below for an example of the 2 Proportions Test and Confidence Interval template.
2 Proportions Equivalence Test Template

2 Proportions Equivalence Test or SigmaXL > Statistical Tools > Basic Statistical Templates > 2
Proportions Equivalence Test.
See Basic Statistical Templates – 2 Proportions Equivalence Test for an example of the 2
Proportions Equivalence Test template.
One-Way Chi-Square Goodness-of-Fit Template

Goodness-of-Fit Test (or SigmaXL > Statistical Tools > Chi-Square Tests > One Way Chi-Square
Goodness-of-Fit Template) to access the One-Way Chi-Square Goodness-of-Fit Test calculator.
See Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit Test for an example of the
One-Way Chi-Square Goodness-of-Fit Test template.
292
One-Way Chi-Square Goodness-of-Fit Template - Exact

Goodness-of-Fit Test Exact (or SigmaXL > Statistical Tools > Chi-Square Tests – Exact > One Way
Chi-Square Goodness-of-Fit Exact Template) to access the One-Way Chi-Square Goodness-of-Fit
Test – Exact calculator.
See Basic Statistical Templates – One-Way Chi-Square Goodness-of-Fit Test - Exact for an example
of the One-Way Chi-Square Goodness-of-Fit Test - Exact template.
1 Poisson Rate Test and Confidence Interval Template

1 Poisson Rate Test and Confidence Interval or SigmaXL > Statistical Tools > Basic Statistical
Templates > 1 Poisson Rate Test and Confidence Interval.
See Basic Statistical Templates – 1 Poisson Rate Test and Confidence Interval for an example of
the 1 Poisson Rate Test and Confidence Interval template.
2 Poisson Rates Test and Confidence Interval Template

2 Poisson Rates Test and Confidence Interval or SigmaXL > Statistical Tools > Basic Statistical
Templates > 2 Poisson Rates Test and Confidence Interval.
See Basic Statistical Templates – 2 Poisson Rates Test and Confidence Interval for an example of
the 2 Poisson Rates Test and Confidence Interval template.
2 Poisson Rates Equivalence Test Template

2 Poisson Rates Equivalence Test or SigmaXL > Statistical Tools > Basic Statistical Templates > 2
Poisson Rates Equivalence Test.
See Basic Statistical Templates – 2 Poisson Rates Equivalence Test for an example of the 2 Poisson
Equivalence Test template.
293
2 Proportions Test and Confidence Interval Template Example

1. We begin with a scenario where Day Shift is running at 80% Yield and Night Shift has 70% Yield.
This data is based on a random sample of 100 units for each Shift, each unit is either good or
bad. Question: Is Day Shift running differently than Night Shift? Statistically, we call the Null
Hypothesis, H0: Proportion P1 = Proportion P2; the alternative hypothesis, Ha is: P1≠P2. If the
calculated P-Value < .05, then we reject the null hypothesis and conclude that Day Shift and
Night Shift are different.
2. Click SigmaXL > Templates & Calculators > Basic Statistical Templates > 2 Proportions Test and
Confidence Interval (or SigmaXL > Statistical Tools > Basic Statistical Templates > 2
Proportions Test and Confidence Interval).
3. Enter x1 = 80, n1 = 100, x2 = 70, n2 = 100 as shown:
Since the Fisher’s exact P-Value of 0.141 is greater than .05, we fail to reject H0. We do not
have enough evidence to show that there is a significant difference between Day Shift and
Night shift. This does not mean that we have proven that they are the same. In practice
however, we either assume that they are the same or we collect more data.
Note: Fisher’s exact P-Values should be used for any real world problem. The approximate P-
Values based on the normal distribution are provided for instructional purposes, e.g.,
comparing to hand calculations.
4. Now enter x1 = 160, n1 = 200, x2 = 140, n2 = 200. Note that the Fisher’s exact P-Value is now
.028, so we reject H0.
294
Chi-Square Test – Two-Way Table Data

1. Open the file Attribute Data.xlsx, ensure that Example 1 Sheet is active. This data is in Two –
Way Table format, or pivot table format. Note that cells B2:D4 have been pre-selected.
2. Click SigmaXL > Statistical Tools > Chi-Square Tests > Chi-Square Test & Association – Two-
Way Table Data. Note the selection of data includes the Row and Column labels (if we had Row
and Column Totals these would NOT be selected). Do not check Advanced Tests and Measures
of Association.
See Appendix Chi-Square Tests and (Contingency) Table Associations for the Chi-Square Table
Statistics formula details.
3. Click Next. The resulting output is:
295
The P-Value matches that of the 2 Proportions test. Since the P-Value of 0.1 is greater than .05,
we fail to reject H0.
4. Now click Example 2 Sheet tab. The Yields have not changed but we have doubled the sample
size. Repeat the above analysis. The resulting output is:
Since the P-Value is < .05, we now reject the Null Hypothesis, and conclude that Day Shift and
Night Shift are significantly different. The Std. (Standardized) Residuals tell us that Day Shift
failures are less than expected (assuming equal proportions), and Night Shift failures are more
than expected.
Note, by doubling the sample size, we improved the power or sensitivity of the test.
5. Click the Example 3 Sheet tab. In this scenario we have 3 suppliers and an additional marginal
level. A random sample of 100 units per supplier is tested. The null hypothesis here is: No
relationship between Suppliers and Pass/Fail/Marginal rates, but in this case, we can state it as
No difference across suppliers. Redoing the above analysis (for selection B2:E5) yields the
following:
296
The P-Value tells us that we do not have enough evidence to show that there is a difference
across the 3 suppliers.
6. Click the Example 4 Sheet tab. Here we have doubled the sample size to 200 per supplier. Note
that the percentages are identical to example 3. Redoing the above analysis yields the
following:
With the P-Value < .05 we now conclude that there is a significant difference across suppliers.
Examining the Std. (Standardized) Residuals tells us that Supplier A has fewer failures than
297
expected (if there was no difference across suppliers), Supplier B has more marginal parts than
expected and Supplier C has fewer marginal parts than expected.
7. The table row and column cell percentages can be visualized using Excel’s 100% Stacked
Column Chart. Select cells A3:D6 of the Chi-Square sheet. Click Excel’s Insert > Insert Column
or Bar Chart and select 100% Stacked Column as shown.
8. Click to create the 100% stacked column chart. Uncheck the Chart Title as shown.
9. The rows and columns can easily be switched by clicking Design > Switch Row/Column
298
10. These charts make it easy to visualize the cell row and column percentages.
299
Chi-Square Test – Two-Way Table Data: Advanced Tests and

Measures of Association – Nominal Categories
Checking the Nominal Categories option provides additional chi-square statistics and measures of
association, including:
• Adjusted Residuals
o Equivalent to normal z score
o Red font highlight denotes significant cell residual value
o Bold red highlight denotes significant cell residual value with Bonferroni adjustment
o Note: red highlight is only active if Chi-Square P-Value is significant
• Cell’s Contribution to Chi-Square
• Additional Chi-Square Tests
o Likelihood Ratio
o McNemar-Bowker Symmetry (Square Table)
 For a 2x2 table, McNemar’s test is equivalent to a paired two-proportions
test, for example applicable to studying before versus after change in
proportion on the same subject. The returned P-Value is exact, based on the
binomial distribution.
 Bowker extended McNemar’s test for square tables larger than 2x2. The null
hypothesis is that the table is symmetrical (i.e., symmetry of disagreement).
This uses Chi-Square as the test statistic.
• Measures of Association for Nominal Categories
o Pearson's Phi
 Pearson’s Phi is equivalent to Pearson’s correlation coefficient for a 2x2
table. It is the most popular measure of association for 2x2 tables.
 We recommend the following rules-of-thumb, adapted from Cohen (1988):
• < 0.1 = Very Weak
• 0.1 to < 0.3: Weak (“Small” Effect)
• 0.3 to < 0.5: Moderate (“Medium” Effect)
• > 0.5: Strong (“Large” Effect)
 Although Phi is equivalent to Pearson’s correlation for a 2x2 table, we
recommend these rules-of-thumb for use in typical contingency tables,
rather than those commonly used for correlation (i.e., > 0.9 = Strong).
300
o Cramer’s V
 Cramer's V is an extension of Phi for larger tables. It is the most popular
measure of association for tables of any size.
 It varies from 0 to 1, with 0 = no association and 1 = perfect association.
 Use Cohen’s rules-of-thumb given above for Phi.
o Contingency Coefficient
 An alternative to Phi, varies from 0 to < 1.
 Use Cohen’s rules-of-thumb given above for Phi.
o Cohen's Kappa (Agreement - Square Table)
 Kappa is used to measure agreement between two assessors evaluating the
same parts or items.
 For an extended Attribute Measurement Systems Analysis use SigmaXL >
Measurement Systems Analysis > Attribute MSA.
• For Attribute MSA used in Six Sigma quality, the recommendation is

Kappa > 0.9 is strong agreement and < 0.7 is weak agreement, but for
general use, the less stringent guidelines by Fleiss are recommended:
Kappa: >= 0.75 or so signifies excellent agreement, for most purposes,

and <= 0.40 or so signifies poor agreement.
• See Appendix Kappa for further details.
o Goodman-Kruskal Lambda & Tau and Theil's Uncertainty

 Measures of Proportional Reduction in Predictive Error. The basic concept is
a measure that indicates how much knowing the value of the independent
variable improves our ability to estimate the value of the dependent variable.
 They are Directional Measures. If the Y dependent variable is in the Rows
Category, then use the Rows Dependent measure. If the Y dependent
variable is in the Columns Category, then use the Cols Dependent measure. If
there is no clear X-Y dependent-independent relationship, then use the
Symmetric measures (not available for Tau).
 Use Cohen’s rules-of-thumb for these measures.
See Appendix Chi-Square Tests and (Contingency) Table Associations for further formula details
and references. The following are external links with helpful presentations on measures of
association for contingency tables:
• http://course1.winona.edu/bdeppa/STAT%20701%20Online/stat_701%20home.htm
Click 4a) Measures of Association. This is a narrated presentation, but a regular PowerPoint
301
is also available on the web site. It is part of a Biostatistics course by Dr. Brant Deppa,
Department of Mathematics and Statistics, Winona State University.
• http://uregina.ca/~gingrich/ch11a.pdf by Dr. Paul Gingrich, University of Regina.
1. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Check Nominal Categories as
shown:
302
The adjusted residuals are equivalent to normal z values, so for a specified 95% confidence
level, any value greater than 1.96 (or less than -1.96) is highlighted in red. This results in a slight
difference in interpretation from that of the standardized residuals, but the 3 largest magnitude
residuals are consistent.
As noted above, the Chi-Square P-Value tells us that there is a significant difference across
suppliers, in other words, there is association between Supplier and Pass/Fail/Marginal, but it
does not tell us the degree or strength of that association. Cramer’s V is used for tables larger
than 2x2 and from the rules-of-thumb, the 0.1 value is considered weak (or small effect).
303
Computing Odds Ratio and Confidence Interval for 2x2 Table

Odds Ratio and Confidence Intervals are not directly available for 2x2 Tables but can obtained using
Logistic Regression. The table data must be rearranged to stacked column format and the response
changed to “Event/Trial” format. Example 2 of Attribute Data.xlsx:
would be changed to:
Note, coding Shift as continuous numeric is easier to interpret and is valid because it has a range of
1.
Analyze using SigmaXL > Statistical Tools > Regression > Binary Logistic Regression. Select
Response Count (Y) / Sample Size. Pass(Y) is selected as Numeric Response Count (Y), Trials (N) is
Numeric Sample Size (Trials) and Shift (X) is selected as Continuous Predictors (X):
Click OK. The Odds Ratio and Confidence Limits are given as:
The shift change 1 to 2 is 0.583 times as likely to produce passed (good) product with a 95%
confidence interval of 0.368 to 0.924.
304
Chi-Square Test – Two-Way Table Data: Advanced Tests and

Measures of Association – Ordinal Categories
Checking the Ordinal Categories option provides statistics and measures of association appropriate
when both row and column category variables are ordinal:
• Adjusted Residuals and Cell’s Contribution to Chi-Square
• Tests of Association for Ordinal Categories
o Concordant – Discordant
 The P-Value for this hypothesis test is from Kendall’s Tau-B, but is the same
for all of the Concordant – Discordant ordinal measures: Tau-C, Gamma and
Somers’ D. See Agresti (2010). This may differ from other software using an
approximation formula.
o Spearman Rank Correlation
• Measures of Association for Ordinal Categories with Confidence Intervals
o Spearman Rank Correlation
 Equivalent to Pearson’s correlation on ranks
o Kendall's Tau-B (Square Table), Kendall-Stuart Tau-C (Rectangular Table), Goodman-
Kruskal Gamma
 Use Tau-B for square tables (no. rows = no. columns) and Tau-C for
rectangular tables (no. rows <> no. columns).
 Tau-B, Tau-C and Gamma are asymmetric measures, so will give the same
result regardless of variable assignment to Rows and Columns.
 These are all Concordant – Discordant measures.

o Somers' D (Cols & Rows Dependent, Symmetric)
 Also a Concordant – Discordant measure but directional. If the Y dependent

variable is in the Rows Category, then use the Rows Dependent measure. If
the Y dependent variable is in the Columns Category, then use the Cols
Dependent measure. If there is no clear X-Y dependent-independent
relationship, then use the Symmetric measure.
• SigmaXL provides rules-of-thumb for Kendall’s Correlation in Ordinal Attribute MSA (strong
association is > 0.8) and Pearson or Spearman Correlation (strong association is > 0.9),
however these are in the context of measurement systems analysis, design of experiments
or a controlled process study. For typical contingency table applications, we recommend
the rules-of-thumb, adapted from Cohen (1988):
305
o 0.5+: Strong (Large Effect)
o 0.3 to < 0.5: Moderate (Medium Effect)
o 0.1 to < 0.3: Weak (Small Effect)
o < 0.1: Very Weak

See Appendix Chi-Square Tests and (Contingency) Table Associations for further formula details
and references. See also Agresti (2010, Chapter 7 Non-Model-Based Analysis of Ordinal
Association) and Deppa’s presentation on Measures of Association at:
http://course1.winona.edu/bdeppa/STAT%20701%20Online/stat_701%20home.htm.
306
1. Open the file Attribute Data.xlsx, click Example 5 – Salary Sat Sheet tab. This data is in two-
way table format and has ordinal categories: Salary in the Rows and Satisfaction Level in the
Columns. Note that cells A1:E5 have been pre-selected.
2. Click SigmaXL > Statistical Tools > Chi-Square Tests > Chi-Square Test & Association – Two-
Way Table Data. Note the selection of data includes the Row and Column labels (if we had Row
and Column Totals these would NOT be selected). Check Nominal Categories and Ordinal
Categories as shown:
Tip: Even if the categories are ordinal, it is sometimes useful to select nominal categories as
well for comparison purposes.
307
Note that the Chi-Square P-Value is 0.1, indicating that there is no significant association
between Salary and Satisfaction when they are treated as nominal categories (although the
significant result for McNemar-Bowker does show that there is lack of symmetry in the off
diagonals).
Since the Chi-Square P-Value is not significant, the Adjusted Residuals are not highlighted, even
though some values are greater than 1.96 (and less than -1.96). This follows the concept used
in ANOVA called “Fisher Protected” where one considers the significance of post-hoc tests only
when the overall test is significant.
Note: 4 out of 16 cells have expected counts less than 5. If more than 20% of the cells have
expected counts less than 5 (or if any of the cells have an expected count less than 1), the Chi-
Square approximation may be invalid, and Fisher’s Exact should be used. This will be discussed
308
later, but for this example the Fisher’s Monte-Carlo Exact P-Value = 0.095 so does not change
the interpretation of the results for the above Chi-Square analysis.
When Salary and Satisfaction are treated as ordinal categories, the more powerful Concordant
– Discordant and Spearman Rank Correlation P-Values clearly show that there is a significant
association. The Measures of Association for Ordinal Categories show that this is positive, i.e.,
an increase in Salary is associated with an increase in Satisfaction. However, using the rules-of-
thumb given above, we see that the association is weak.
4. The table row and column cell percentages can be visualized using Excel’s 100% Stacked
Column Chart. Select cells A3:E7 of the Chi-Square sheet. Click Excel’s Insert > Insert Column
or Bar Chart and select 100% Stacked Column as shown.
5. Click to create the 100% stacked column chart (uncheck the Chart Title):
6. The rows and columns can easily be switched by clicking Design > Switch Row/Column
309
Chi-Square Test (Stacked Column Format Data)

1. Open Customer Data.xlsx. Click Sheet 1 tab. The discrete data of interest is Complaints and
Customer Type, i.e., does the type of complaint differ across customer type? Formally the Null
Hypothesis is that there is no relationship (or independence) between Customer Type and
Complaints.
2. Click SigmaXL > Statistical Tools > Chi-Square Tests > Chi-Square Test & Association. Ensure
that the entire data table is selected. If not, check Use Entire Data Table. Click Next.
3. Select Major-Complaint, click Rows Category >>; select Customer Type, click Columns Category
>>. We will not be using Options in this example.
310
With the P-Value = 0.142 we fail to reject H0, so we do not have enough evidence to show a
difference in customer complaints across customer types.
Note: 9 out of 15 cells have expected counts less than 5. If more than 20% of the cells have
expected counts less than 5 (or if any of the cells have an expected count less than 1), the Chi-
Square approximation may be invalid. Use Chi-Square Test – Fisher’s Exact).
Tip: Use Advanced Pareto Analysis and Excel’s 100% Stacked Column Chart to complement Chi-
Square Analysis.
Caution: When using stacked column format data with Ordinal Category variables that are text,
SigmaXL will sort alphanumerically which may not result in the correct ascending order for
analysis. We recommend coding text Ordinal variables as numeric (e.g., 1,2,3) or modified text
(e.g., Sat_0, Sat_1).
311
5. Press F3 or click Recall SigmaXL Dialog to recall last dialog. Select Loyalty – Likely to
Recommend, click Rows Category >>; select Sat-Discrete, click Columns Category >>. Click
Options, check Ordinal Categories.
Sat-Discrete is derived from Overall Satisfaction where a score >= 3.5 is considered a 1, and
scores < 3.5 are considered a 0.
312
As expected there is a strong positive association between Loyalty and Discrete Satisfaction.
Note: Kendall-Stuart Tau-C should be used here rather than Tau-B because it is a rectangular
table. Since Satisfaction leads to Loyalty, Loyalty is the Dependent variable, so Rows
Dependent Somers’ D should be used rather than Cols Dependent or Symmetric.
7. In order to visualize the row and column percentages with Excel’s 100% Stacked Column Chart,
we will need to modify the numeric row and column labels as shown, converting to text as
shown:
8. Select cells A3:C8 of the Chi-Square sheet. Click Excel’s Insert > Insert Column or Bar Chart and
select 100% Stacked Column as shown.
313
9. Click to create the 100% stacked column chart (uncheck the Chart Title):
This clearly shows the strong relationship between the Loyalty scores and Discrete Satisfaction.
314
Chi-Square Test – Fisher’s Exact

The Chi-Square Test – Fisher’s Exact utilizes permutations and fast network algorithms to solve
the Exact Fisher P-Value for contingency (two-way row*column) tables. This is an extension of
the Fisher Exact option provided in the Two Proportion Test template. For data that requires
more computation time than specified, Monte Carlo P-Values provide an approximate (but
unbiased) P-Value that typically matches exact to two decimal places using 10,000 replications.
One million replications give a P-Value that is typically accurate to three decimal places. A
confidence interval (99% default) is given for the Monte Carlo P-Values. For further details refer
to the Appendix Exact and Monte Carlo P-Values for Nonparametric and Contingency Tests.
It is important to note that while exact P-Values are “correct,” they do not increase (or
decrease) the power of a small sample test, so they are not a solution to the problem of failure
to detect a change due to inadequate sample size!
1. We will now re-analyze the above Major Complaint by Customer Type data. Open Customer
Data.xlsx. Click Sheet 1 tab. The discrete data of interest is Complaints and Customer Type,
i.e., does the type of complaint differ across customer type? Formally the Null Hypothesis is
that there is no relationship (or independence) between Customer Type and Complaints.
2. Click SigmaXL > Statistical Tools > Chi-Square Tests – Exact > Chi-Square Test – Fisher’s Exact.
Ensure that the entire data table is selected. If not, check Use Entire Data Table. Click Next.
3. Select Major-Complaint, click Rows Category >>; select Customer Type, click Column Category
>>. Select Exact with the default Time Limit for Exact Computation = 60 seconds.
Tip: If the exact computation time limit is exceeded a dialog will prompt you to use Monte Carlo
or to increase the computation time. When this occurs, Monte Carlo is recommended.
315
With Fisher’s Exact P-Value = 0.1469 we fail to reject H0, so we cannot conclude that there is a
difference in customer complaints across customer types. This is close to the approximate Chi-
Square P-Value of 0.142, so either P-Value results in the same conclusion.
However, there are cases where the “large sample” Chi-Square P-Value leads to one conclusion
but Fisher’s Exact P-Value gives another. This will be demonstrated later using Two Way Table
Data – Fisher’s Exact.
The Exact P-Value was computed very quickly, but if the data set was larger, the required
computation time could become excessive, and Monte Carlo would be required. We will rerun
this analysis with Monte Carlo and discuss the output report.
default Number of Replications = 10000 and Confidence Level for P-Value = 99%.
316
Tip: As discussed above 10,000 replications will result in a Monte Carlo P-Value that is correct
to two decimal places. One million (1e6) replications will result in three decimal places of
accuracy and typically require less than 60 seconds to solve for any data set.
317
Fisher’s Monte Carlo P-Value here is 0.1486 with a 99% confidence interval of 0.1403 to 0.1569.
system clock). The true Exact P-Value = 0.1469 lies within this confidence interval. If the
confidence interval is too wide (e.g., upper “fail-to-reject” H0 and lower “reject” H0), simply
rerun the Monte Carlo option using a larger number of replications (use 1e5 or 1e6).
318
Chi-Square Test – Two Way Table Data – Fisher’s Exact

Now we will consider a sparse data set where the Chi-Square approximation fails and Fisher’s Exact
is required to give a correct conclusion for the hypothesis test. This is adapted from a subset of
dental health data (oral lesions) obtained from house to house surveys that were conducted in
three geographic regions of rural India [1, 2]. The Fisher’s Exact P-Value obtained with SigmaXL
may be validated using these references. The data labels have been modified to a generic “A”, “B”,
“C”, etc. for the oral lesions location (rows) and “Region1”, “Region2” and “Region3” for the
geographic regions (columns).
1. Open the file Oral_Lesions.xlsx. This data is in Two –Way Contingency Table (or pivot table)
format. Note that cells A1:D10 have been pre-selected.
2. Click SigmaXL > Statistical Tools > Chi-Square Tests – Exact > Chi-Square Test – Two-Way Table
Data – Fisher’s Exact. Note the selection of data includes the Row and Column labels.
3. Click Next. Results:
319
Note: 25 out of 27 cells have expected counts less than 5. Since more than 20% of the cells
have expected counts less than 5 and several cells have an expected count less than 1, the Chi-
Square P-Value = 0.14 is invalid and leads to an incorrect conclusion: Fail to Reject H0, Site of
Oral Lesion and Geographic Region are independent. However with Fisher’s Exact P-Value =
320
0.0101, we have strong evidence to reject H0, and conclude that there is indeed a relationship
between Site of Oral Lesion and Region.
While this example is originally from a dental health study, it could have been Defect Type
versus Supplier and highlights that the use of Fisher’s Exact for contingency tables with sparse
data can make the difference between a good business decision and a bad business decision!
The Exact P-Value was computed very quickly, but if the data set was larger, the required
computation time could become excessive, and Monte Carlo would be required. We will rerun
this analysis with Monte Carlo and discuss the output report.
default Number of Replications = 10000 and Confidence Level for P-Value = 99%.
Tip: As discussed above 10,000 replications will result in a Monte Carlo P-Value that is correct
to two decimal places. One million (1e6) replications will result in three decimal places of
accuracy and typically require less than 60 seconds to solve for any data set.
5. Click Next. Results:
321
Fisher’s Monte Carlo P-Value here is 0.0094 with a 99% confidence interval of 0.0072 to 0.0116.
322
system clock). The true Exact P-Value = 0.0101 lies within this confidence interval. If the
confidence interval is too wide (e.g., upper is a “fail-to-reject” H0 and lower is a “reject” H0),
simply rerun the Monte Carlo option using a larger number of replications (use 1e5 or 1e6).
References for Fisher’s Exact

[1] Mehta, C. R.; Patel, N. R. (1997) "Exact inference in categorical data," unpublished preprint,
http://www.cytel.com/Papers/sxpaper.pdf. See Table 7 (p. 33) for validation of Fisher’s Exact P-
Value.
[2] Mehta, C.R. ; Patel, N.R. (1998). "Exact Inference for Categorical Data." In P. Armitage and T.
Colton, eds., Encyclopedia of Biostatistics, Chichester: John Wiley, pp. 1411–1422.
323
Power & Sample Size for One Proportion Test

To determine Power & Sample Size for a 1 Proportion Test, you can use the Power & Sample
Size Calculator or Power & Sample Size with Worksheet.
1. Click SigmaXL > Statistical Tools > Power & Sample Size Calculators > 1 Proportion Test
Calculator.
2. Select Solve For Power (1 – Beta). Enter Sample Size and Alternative Proportion as shown:
Note that we are calculating the power or likelihood of detection given that the hypothesized
proportion is 0.5, but the alternative proportion is 0.6, sample size = 300, significance level =
.05, and Ha: Not Equal To (two sided test).
A power value of 0.94 is acceptable, but note that the sample size n = 300, and the difference in
proportion value is 0.1 or 10%! The sample size requirements for discrete data are much higher
than those for continuous data.
324
Power & Sample Size with Worksheets > 1 Proportion Test.
5. A graph showing the relationship between Power, Sample Size and Proportion Value can then
be created using SigmaXL > Statistical Tools > Power & Sample Size Chart. See Part E for an
325
Power & Sample Size for Two Proportions Test

To determine Power & Sample Size for a 2 Proportions Test, you can use the Power & Sample
Size Calculator or Power & Sample Size with Worksheet.
1. Click SigmaXL > Statistical Tools > Power & Sample Size Calculators > 2 Proportions Test
Calculator.
2. Select Solve For Power (1 – Beta). Enter Sample Size and Proportion 1 values as shown:
Note that we are calculating the power or likelihood of detection given that P1 = 0.5 and P2 =
0.6, sample size for each group = 300, significance level = .05, and Ha: Not Equal To (two sided
test).
A power value of 0.69 is unacceptable. Note that this value is much less than the power for the
one proportion test (0.94).
326
4. To compensate, we will double the sample size per group. Press F3 or click Recall SigmaXL
Dialog to recall last dialog. Change the sample size per group from 300 to 600. Note that the
power value is now 0.94.
Power & Sample Size with Worksheets > 2 Proportions Test.
6. A graph showing the relationship between Power, Sample Size and Proportion Values can then
be created using SigmaXL > Statistical Tools > Power & Sample Size Chart. See Part E for an
327
Part P – Analysis of Means (ANOM) Charts
Introduction – What is ANOM

From http://asq.org/glossary/a.html:
A statistical procedure for troubleshooting industrial processes and analyzing the results of
experimental designs with factors at fixed levels. It provides a graphical display of data. Ellis
R. Ott developed the procedure in 1967 because he observed that nonstatisticians had
difficulty understanding analysis of variance. Analysis of means is easier for quality
practitioners to use because it is (like) an extension of the control chart.
From the preface of the book, “The Analysis of Means: A Graphical Method for Comparing Means,
Rates, and Proportions” by Peter R. Nelson, Peter S. Wludyka, and Karen A. F. Copeland:
The goal of statistical data analysis is to use data to gain and communicate knowledge
about processes and phenomena. Comparing means is often part of an analysis, for data
arising in both experimental and observational studies.
The analysis of means (ANOM) is an alternative procedure (to ANOVA) for comparing
means.
ANOM has the advantages of being much more intuitive and providing an easily understood
graphical result, which clearly indicates any means that are different (from the overall
mean) and allows for easy assessment of practical as well as statistical significance.
328
Formula for One-Way Balanced* Normal (from ANOM book 2.4)
where:
ȳ.. = overall mean
h = critical value from multivariate t distribution – SigmaXL uses table exact critical values
(Table B.1)
N = sample size
I = number of levels
√MS𝑒𝑒 SQRT (Mean Square Error) = pooled standard deviation.

* Unbalanced uses critical values from studentized maximum modulus (SMM) distribution. SigmaXL
uses table exact critical values (Table B.2). An adjustment is also made for varying sample size that
results in varying decision limit values. See Appendix Analysis of Means (ANOM) Charts.
Formula for One-Way Balanced Binomial Proportions (from ANOM book 2.6)
This is based on the normal approximation to the Binomial, so requires sample sizes large enough
that np and n(p-1) > 5. SigmaXL automatically checks this and gives a warning message if the
condition is not met.
329
Formula for One-Way Balanced Poisson Rates (from ANOM book 2.9)
This is based on the normal approximation to the Poisson, so requires sample sizes large enough
that nu > 5. SigmaXL automatically checks this and gives a warning message if the condition is not
met.
Formula for Nonparametric Transformed Ranks (from ANOM book 9.3)

The values from the combined sample are ranked from smallest to largest (Rij) and transformed to
a z-score using the inverse normal:
and the above formula for balanced (or unbalanced) Normal is applied to compute decision limits.
Formula for Balanced Variances (from ANOM book 4.4)
where U, L are critical values from Table B.4 (n <= 35). For unbalanced and large sample size see
Appendix Analysis of Means (ANOM) Charts.
Formula for Levene Robust Variances (from ANOM book 4.16)

The Levene Absolute Deviation from Median (ADM) is a simple transformation:
and the above formula for balanced (or unbalanced) Normal is applied to compute decision limits.
330
Two-Way Normal
Main Effects for Two-Way ANOM are similar to One-Way but the mean square error (MSE) is
derived from the ANOVA.
Slice Charts are a modified ANOM chart developed by Dr. Peter Wludyka that enables one to easily
interpret the effects in the presence of an interaction (Wludyka 2013, 2015). The basic idea is to
compare the levels of one factor for each level of the other factor with MSE derived from the Two-
Way ANOVA. A yellow highlight automatically recommends Main Effects (if interaction is not
significant, i.e., P-Value >= 0.1) or Slice Chart (if interaction is significant, P-Value < 0.1). The
interaction P-Value is determined from ANOVA
There is an option to specify correction to alpha for multiple chart family-wise error rate:
Bonferroni alpha’ = alpha/m; m = number of charts
Note that this is strictly to adjust for multiple charts. As seen in the formulas above, the individual
charts also have correction built-in for the number of levels.
A Two-Way ANOM analysis will typically have balanced data from a designed experiment, with an
equal number of observations for each combination level of X1 and X2. SigmaXL will allow
unbalanced data, but we recommend that you use balanced data or slightly unbalanced data.
ANOM charts use actual data means, not fitted (predicted least squares) means as used in Two-
Way ANOVA Main Effects and Interaction Plots.
Two-Way Binomial Proportions and Poisson Rates

In collaboration, Peter Wludyka and John Noguera of SigmaXL extended the Slice Charts to
Binomial and Poisson (Wludyka and Noguera 2016).
As with Normal, the basic idea is to compare the levels of one factor for each level of the other
factor. MSE is derived from the whole model
A yellow highlight automatically recommends Main Effects (if interaction is not significant) or Slice
Chart (if interaction is significant).
The interaction P-Value is automatically determined from Logistic regression for Binomial
Proportions and Poisson regression for Poisson Rates.
See Appendix Analysis of Means (ANOM) Charts for further formula details and references.
331
ANOM Normal One-Way

2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Normal One-Way.
Ensure that the entire data table is selected. If not, check Use Entire Data Table.
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>. Set Alpha
Level = 0.05.
4. Click OK. The ANOM Normal One-Way chart is shown below:
332
5. Here we see that Customer Type 1 mean satisfaction score is significantly below the overall
mean and Customer Type 2 is significantly higher. This is consistent with the results that were
observed in the ANOVA analysis, but is easier to interpret.
333
ANOM Normal Two-Way (with Main Effects and Slice Charts)

1. We will look at two balanced examples from the ANOM book (used with author permission),
one with no interaction and another with a strong interaction, and then a slightly unbalanced
version of the latter (Wludyka, 2013).
2. Open ANOM Examples.xlsx, click on Hemoglobin – Normal tab. This data is from a factorial
design to study hemoglobin levels in males (Example 5.1). The factors are Therapy (Chemo or
Radiation) and Drug (Type 1, 2, 3).
3. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Normal Two-Way.
4. Click Next. Select Hemoglobin Level, click Numeric Data Variable (Y) >>; select Therapy, click
Group Category Factor (X1) >>; select Drug, click Group Category Factor (X2) >>. Alpha Level =
0.05, Adjust chart alpha for family-wise error rate is unchecked:
5. Click OK. The ANOM Normal Two-Way Main Effects, Slice charts and Interaction plots are
shown below:
334
6. Since the Interaction is not significant (P-Value is >= 0.1), the Main Effects charts are
highlighted. The decision limits match the manual calculations given in the ANOM book (pages
87, 88). Since the average hemoglobin level for radiation falls below the lower decision limit,
and the average for chemotherapy falls above the upper decision limit, therapy has an (α =
0.05) effect. There is also an effect on hemoglobin level due to the drug since the average
hemoglobin level for drug 3 is above the upper decision line.
Tip: Two-Way ANOM Normal is complimentary to Two-Way ANOVA. Analyzing the above data
using Two-Way ANOVA shows that Therapy is significant with a P-Value = 0.0001, Drug is also
significant with a P-Value = 0.0105 and the Interaction is not significant, with a P-Value =
0.5942.
335
7. Click Recall SigmaXL Dialog menu or press F3 to recall last dialog. Check Adjust chart alpha for
family-wise error rate.
This will apply a Bonferroni correction to the specified alpha values to control the overall
family-wise error rate for the charts. So, the Main Effects alpha will be = 0.05/2 = 0.025 and the
Slice Charts alpha will be = 0.05/2 = 0.025 as well.
336
8. Click OK. The ANOM Normal Two-Way Main Effects and Slice charts are shown below:
9. With the Bonferroni corrected alpha of 0.025, the significant results do not change, but they are
closer to the decision lines. The trade-off for this chart family-wise adjustment is loss of power.
337
10. In ANOM Examples.xlsx, click on Process Yield – Normal tab. This data is from a factorial
design to study the effect of three monomers and two levels of UV exposure (1 = UV; 2 = No
UV) on the percent yield for contact lens production (Example 5.3).
12. Click Next. Select Process Yield, click Numeric Data Variable (Y) >>; select UV, click Group
Category Factor (X1) >>; select Monomer, click Group Category Factor (X2) >>. Alpha Level =
shown below:
338
14. Since the Interaction P-Value is < 0.1, the Slice charts are highlighted. The decision limits for
the Slice charts match the calculations given in Wludyka (2013). See Appendix References for
Analysis of Means (ANOM) Charts. The Slice plot clearly shows the effect of the strong
interaction. When UV = 1 (UV Exposure), the three monomers differ. In particular, the average
for monomer 1 is below the overall average and the average for monomer 3 is above the
overall average. However, for UV = 2 (no UV), the monomers also differ but in an entirely
different manner than in the UV = 1 case, revealing the nature of the interaction.
Tip: As previously noted, Two-Way ANOM Normal is complimentary to Two-Way ANOVA.

Analyzing the above data using Two-Way ANOVA shows that UV is significant with a P-Value =
0.0001, Monomer is significant with a P-Value = 0.0000 (4.4 e-5) and the Interaction is also
significant, with P-Value = 0.0000 (2.2e-6).
339
15. In ANOM Examples.xlsx, click on Process Yield – Normal Unbal tab. This data is from a
factorial design to study the effect of three monomers and two levels of UV exposure (1 = UV; 2
= No UV) on the percent yield for contact lens production, but two of the observations are
missing for A = 2 (no UV) and B = 3 (monomer 3).
17. Click Next. Select Process Yield, click Numeric Data Variable (Y) >>; select UV, click Group
Category Factor (X1) >>; select Monomer, click Group Category Factor (X2) >>. Alpha Level =
shown below:
340
19. The significant results of the experiment have not changed, but the decision limits are now
varying due to the unbalanced data and use critical values for unbalanced data. The decision
limits for the Slice charts match the calculations given in Wludyka (2013).
341
ANOM Binomial Proportions One-Way

1. Open ANOM Examples.xlsx, click on the Test Scores – Binomial Prop tab. This is standardized
math test score data from 10 elementary schools (Example 3.4 from the ANOM book – used
with author permission). We are testing to see if there is a difference between schools at an
alpha = 0.01 level.
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Binomial Proportions
One-Way. Ensure that the entire data table is selected. If not, check Use Entire Data Table.
3. Click Next. Select Proficient, click Numeric Data Variable (Y) >>; select Enrollment, click
Subgroup Column or Size >>; select School, click Optional Group Category (X) >>. Set Alpha
Level = 0.01:
4. Click OK. The ANOM Binomial Proportions One-Way chart is shown below:
342
5. The resulting ANOM decision chart shows that three schools are performing at significantly low
levels and two schools are performing at significantly high levels.
6. SigmaXL automatically checks to see if the sample sizes are large enough for the normal
approximation to the Binomial to be valid, i.e., np and n(p-1) are > 5. Here we see: Warning:
Sample sizes are too small to use the normal approximation to the binomial distribution.
Sample #1 has minimum n(1-p) = 2.0. Note that the warning does not show all occurrences,
just the sample(s) with smallest failed np or n(1-p).
7. This does not mean that the chart results are invalid, an obvious “out” or “in” will not likely be
affected, but the results should be used with caution, and if possible, more data collected.
343
ANOM Binomial Proportions Two-Way

1. Open ANOM Examples.xlsx, click on the Length of Stay - Binomial Prop tab. This is hospital
length of stay (LOS > 6 days) after bypass surgery. The data is from a factorial design with
Factors: Drug (A, B) and Treatment (1, 2, 3). (Example 5.15 from the ANOM book – used with
author permission).
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Binomial Proportions
Two-Way. Ensure that the entire data table is selected. If not, check Use Entire Data Table.
3. Click Next. Select LOS > 6 days, click Numeric Data Variable (Y) >>; select Sample Size, click
Subgroup Column or Size >>; select Drug, click Group Category Factor (X1) >>; select
Treatment, click Group Category Factor (X2) >>. Alpha Level = 0.05, Adjust chart alpha for
family-wise error rate is unchecked:
344
4. Click OK. The ANOM Binomial Proportions Two-Way Main Effects, Slice charts and Interaction
plots are shown below:
5. Since the Interaction P-Value is < 0.1 (automatically determined from Logistic regression), the
Slice charts are highlighted. The Slice plot clearly shows the effect of the strong interaction.
When Drug = A, Treatment 2 results in a significantly lower mean proportion LOS and
Treatment 3 results in a significantly higher LOS. When Drug = B, the Treatment does not have
a significant effect on LOS.
345
ANOM Poisson Rates One-Way

1. Open ANOM Examples.xlsx, click on the Injection Molding-Poisson Rate tab. This is injection
molding data for contact lenses. It is from a study on the extent to which belt speed (cure time)
influences the incidence of nonconformities (such as bubbles and tears) in the lenses. Since
there can be more than one nonconformity per unit, Poisson ANOM is appropriate, similar to
the U Control Chart (Example 2.16 from the ANOM book – used with author permission).
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Poisson Rates One-
Way. Ensure that the entire data table is selected. If not, check Use Entire Data Table.
3. Click Next. Select Nonconformities, click Numeric Data Variable (Y) >>; enter 100 (or select
Sample Size) for Subgroup Column or Size >>; select Belt Speed, click Optional Group Category
(X) >>. Set Alpha Level = 0.05:
346
4. Click OK. The ANOM Poisson Rates One-Way chart is shown below:
5. From the ANOM chart, one sees that a belt speed of 15 results in a significantly low number of
nonconformities and a belt speed of 30 results in a significantly high number of
nonconformities.
347
ANOM Poisson Rates Two-Way

1. Open ANOM Examples.xlsx, click on the ER Visits – Poisson Rate tab. This is Emergency Room
Visits Data. It is a factorial design study to investigate the effect of nurse intervention (1 = home
nurse assigned, 2 = home nurse not assigned) and payor groups (1, 2, 3) on the ER utilization
rates for patients with COPD (a serious lung disease) with an alpha level = 0.01 (Example 5.16
from the ANOM book – used with author permission).
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Poisson Rates Two-
Way. Ensure that the entire data table is selected. If not, check Use Entire Data Table.
3. Click Next. Select ER Visits, click Numeric Data Variable (Y) >>; select Sample Size, click
Subgroup Column or Size >>; select Intervention, click Group Category Factor (X1) >>; select
Payor, click Group Category Factor (X2) >>. Alpha Level = 0.01, Adjust chart alpha for family-
wise error rate is unchecked:
348
4. Click OK. The ANOM Poisson Rates Two-Way Main Effects, Slice charts and Interaction plots are
shown below:
5. Since the Interaction is not significant (P-Value >= 0.1, automatically determined from Poisson
regression), the Main Effects charts are highlighted. Both factors Intervention and Payor are
significant at alpha = 0.01. “Home nurse assigned” results in a significantly lower rate of ER
visits versus “no home nurse assigned.” Payor Type 1 has a significantly lower rate of ER visits
and Type 2 has a significantly higher rate.
349
ANOM Nonparametric Transformed Ranks

ANOM Nonparametric Transformed Ranks is complementary to a Kruskal-Wallis test. Note that it is
not included as an option in the Kruskal-Wallis dialog because there is a difference in the statistic
used (Kruskal-Wallis uses ANOVA on ranks, whereas the ANOM utilizes a normal transformed rank).
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Nonparametric
Transformed Ranks. Ensure that the entire data table is selected. If not, check Use Entire Data
Table.
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>. Set Alpha
Level = 0.05.
4. Click OK. The ANOM Nonparametric Transformed Ranks chart is shown below:
350
5. Here we see that Customer Type 1 mean transformed rank satisfaction score is significantly
below the grand mean and Customer Type 2 is significantly higher. This is consistent with the
results that were observed in the Kruskal-Wallis analysis (see Kruskal-Wallis).
balanced, the decision limit lines will be constant.
7. Open ANOM Examples.xlsx, click on the Exponential Data – TR tab. This is exponential data,
Example 9.16 from the ANOM book with alpha level = 0.01 (used with author permission).
8. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Nonparametric
Transformed Ranks. Ensure that the entire data table is selected. If not, check Use Entire Data
Table.
9. Click Next. Ensure that Stacked Column Format is checked. Select Y, click Numeric Data
Variable (Y) >>; select X, click Group Category (X) >>. Set Alpha Level = 0.01.
351
10. Click OK. The ANOM Nonparametric Transformed Ranks chart is shown below:
11. X = 1 shows as significantly higher than the grand mean transformed rank and X = 3 is
significantly lower (at alpha = 0.01). These results match those given in the ANOM book.
352
12. By way of comparison, rerun the analysis using ANOM One-Way Normal (with alpha = 0.01).
The results are:
13. ANOM Normal fails to detect the significant differences in mean noted above.
353
ANOM Variances
The ANOM Variances chart is complementary to Bartlett’s Test for Equal Variance and is also
available as an option in the Bartlett’s Test dialog. It assumes that the data are normally
distributed.
1. Open Delivery Times.xlsx, click on Sheet 1 tab.
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Variances. Ensure that
the entire data table is selected. If not, check Use Entire Data Table.
3. Click Next. Ensure that Stacked Column Format is checked. Select Delivery Time Deviation,
click Numeric Data Variable (Y) >>; select Floor, click Group Category (X) >>; Select Standard
Deviations; Set Alpha Level = 0.05.
We are analyzing the same normal data used in Bartlett’s Test for Equal Variances. See
Bartlett’s Test.
354
4. Click OK. The ANOM Variances chart is shown below:
5. The ANOM Variances chart visually shows that none of the group standard deviations are
significantly different from the grand mean of all the standard deviations. It is called an ANOM
Variances Chart but displays Standard Deviations for ease of interpretation (similar to a
Standard Deviation S Control Chart). This does however result in non-symmetrical decision
limits. To display Variances rerun the above analysis with option Variances selected.
355
ANOM Levene Robust Variances

The ANOM Levene Robust Variances chart is complementary to Levene’s Test for Equal Variance
and is also available as an option in the Levene’s Test dialog. Since it uses absolute deviations from
the median, it is robust to the assumption of normality.
2. Click SigmaXL > Graphical Tools > Analysis of Means (ANOM) > ANOM Levene Robust
Variances. Ensure that the entire data table is selected. If not, check Use Entire Data Table.
Numeric Data Variable (Y) >>; select Customer Type, click Group Category (X) >>; Set Alpha
Level = 0.05.
We are analyzing the same nonnormal data used in Levene’s Test for Equal Variances. See
Levene’s Test.
356
4. Click OK. The ANOM Levene Robust Variances chart is displayed:
5. The ANOM chart clearly shows Customer Type 1 has significantly higher variance (ADM) than
overall and Customer Type 2 has significantly lower variance.
357
Part Q – Multi-Vari Charts
Multi-Vari Charts
The Multi-Vari chart is a powerful tool to identify dominant Sources of Variation (SOV). The three
major “families” of variation are: Within Unit, Between Unit, and Temporal (Over Time). We will
look at examples of each type of SOV and then use the Multi-Vari Chart to study Overall
Satisfaction in the Customer Data.xlsx file.
1. Open Multi-Vari Data.xlsx, click Sheet Within. Select SigmaXL > Graphical Tools > Multi-Vari
Options.
2. The charts shown will be updated as options are selected. Note that they are for
demonstration purposes and are not based on the Multi-Vari Data.xlsx data (sample charts are
not displayed in Excel for Mac). Ensure that all general Options are selected (Range Line,
Individual Data Points, Min and Max, Standard Deviation Chart). Select Mean Options, ensure
that Show Means, Connect Means, and Group Means are checked. Ensure that Save Defaults
is checked. These settings would be typical for a Multi-Vari chart. (The Median options provide
the ability to display percentiles as an alternative to the Means).
Tip: Multi-Vari Charts can be used to display Confidence Intervals (as we did earlier in Part C).
To do this, check the 95% Confidence Interval.
3. Click Finish. SigmaXL automatically starts the Multi-Vari Chart procedure (this is equivalent to
clicking SigmaXL > Graphical Tools > Multi-Vari Charts).
4. Check Use Entire Data Table. Click Next.
5. Note that the input X’s can be text or numeric but should be discrete. Y’s must be numeric -
typically continuous, but can also be count or proportion data.
358
6. Select Measurement, click Numeric Response (Y) >>; select unit, click Group Category (X1) >>;
select time, click Group Category (X2) >>.
7. Click OK. Resulting Multi-Vari Chart illustrating dominant “Within Unit” Source of Variation:
8. Using Multi-Vari Data.xlsx, click Sheet Between. Select SigmaXL > Graphical Tools > Multi-Vari
Charts. Check Use Entire Data Table. Click Next.
10. Click OK. Resulting Multi-Vari Chart illustrating dominant “Between Unit” Source of Variation:
359
11. Using Multi-Vari Data.xlsx, click Sheet OverTime. Select SigmaXL > Graphical Tools > Multi-
Vari Charts. Check Use Entire Data Table. Click Next.
13. Click OK. Resulting Multi-Vari Chart illustrating dominant “Over Time” Source of Variation:
14. Open Customer Data.xlsx, click Sheet 1. Select SigmaXL > Graphical Tools > Multi-Vari Charts.
360
15. Select Overall Satisfaction, click Numeric Response (Y) >>; select Customer Type, click Group
Category (X1) >>; select Size of Customer, click Group Category (X2) >>; select Product Type,
click Group Category (X3) >>.
16. Click OK. Resulting Multi-Vari chart:
Examining this Multi-Vari chart reveals that the dominant Source of Variation is “within”
Customer Type, followed by “between” Customer Type. Furthermore, it would be worthwhile
to examine the combination of Customer Type 2, Customer Size Large, and Product Type
Consumer.
Other tools that can help us identify potential X factors that may explain some of the large
“Within” variability are the Scatter Plot, Scatter Plot Matrix and Correlation Matrix.
361
Part R – Scatter Plots
Scatter Plots
1. Open Customer Data.xlsx. Click Sheet 1 Tab. Click SigmaXL > Graphical Tools > Scatter Plots; if
necessary, click Use Entire Data Table, click Next.
2. Select Overall Satisfaction, click Numeric Response (Y) >>; select Avg Days Order Time to
Delivery, click Numeric Predictor (X1) >>. Check Trendline, 95% Confidence Interval and 95%
Prediction Interval as shown:
3. Click OK. The resulting Scatter Plot is shown with equation, trendline, 95% confidence interval
(blue lines – for a given X value this is the 95% confidence interval for predicted mean Overall
Satisfaction) and 95% prediction interval (red lines – for a given X value this is the 95%
confidence interval for predicted individual values of Overall Satisfaction).
The equation is based on linear regression, using the method of least squares. R-squared * 100
is the percent variation of Y explained by X (here 10.3%).
362
4. Now we want to redo the Scatter Plot and stratify by Customer Type. Press F3 or click Recall
SigmaXL Dialog to recall last dialog. (Or, Click Sheet 1 Tab; Click SigmaXL > Graphical Tools >
Scatter Plots; click Next.)
5. Select Overall Satisfaction, click Numeric Response (Y) >>; select Average Days Order Time to
Delivery, click Numeric Predictor (X1) >>; select Customer Type, click Group Category (X2) >>,
Uncheck 95% Confidence Interval and 95% Prediction Interval. Click OK:
Clearly, according to the analysis, Customer Type 3 is happier if orders take longer! But, does
this make sense? Of course not! Customer Sat scores should not increase with Order to
Delivery time. What is happening here? This is a coincidental situation. Something else is
driving customer satisfaction. We will now look at the Scatter Plot Matrix to help us investigate
other factors influencing Customer Satisfaction.
Tip: Be careful when interpreting scatter plots (and other statistical tools): Y versus X
correlation or statistical significance does not always mean that we have a causal relationship.
Umbrella sales are highly correlated to traffic accidents, but we cannot reduce the rate of traffic
accidents by purchasing fewer umbrellas! The best way to validate a relationship is to perform
a Design of Experiments (see Improve Phase).
363
Scatter Plot Matrix

> Graphical Tools > Scatter Plot Matrix.
3. Select the variable Avg No. of orders per month; shift-click on Staff Knowledge and click
Numeric Data Variables (Y) >> as shown:
4. Click OK. Resulting Scatter Plot Matrix:
Of particular interest is Overall Satisfaction versus Responsiveness to Calls and Ease of

Communications. These will be explored further with Multiple Linear Regression.
364
Part S – Correlation Matrix
The correlation matrix complements the scatterplot matrix by quantifying the degree of
association. The following table shows the approximate relationship between r, R-squared, and
degree of association:
Pearson Correlation Coefficient R-Squared (%) Degree of association

(r)
0.9 <= |r| <=1 > 80 % Strong
0.7 <= |r| < 0.9 50 % to 80 % Moderate
|r| < 0.7 < 50 % Weak
Pearson Probability, p > 0.05 None
Correlation Matrix
1. Open Customer Data.xlsx. Click Sheet 1 tab (or press F4 to activate last worksheet). Click
SigmaXL > Statistical Tools > Correlation Matrix. Ensure that entire data table is selected. If
not, check Use Entire Data Table. Click Next.
2. Select the variable Avg No. of orders per month; shift-click on Staff Knowledge and click
Numeric Data Variables (Y) >> as shown:
365
3. Click OK. Resulting Correlation Matrix is shown:
Correlations highlighted in red are considered significant (P-Values < .05). The corresponding
correlation coefficients above the P-Values are also highlighted in red. (Compare these to the
Scatterplot Matrix.)
Note that Spearman’s Rank Correlation complements Pearson’s Correlation, in that it provides
a robust measure of association. Spearman’s rank is based on correlated ranks, which are not
sensitive to outliers or departures from normality.
An automatic normality check is applied, utilizing the powerful Doornik-Hansen bivariate

normality test. A yellow highlight recommends Pearson or Spearman correlations be used (but
only if it is significant). Pearson is highlighted if the data are bivariate normal, otherwise
Spearman is highlighted. Always review the data graphically with the scatterplot matrix as well.
Reference
Doornik, J.A. and Hansen, H. “An Omnibus Test for Univariate and Multivariate Normality,” Oxford
Bulletin Of Economics And Statistics, 70, Supplement (2008).
366
Part T – Multiple Regression
Multiple Regression
Multiple Regression analyzes the relationship between one dependent variable (Y) and multiple
independent variables (X's). It is used to discover the relationship between the variables and
create an empirical equation of the form:
Y = b0 + b1*X1 + b2*X2 + ... + bn*Xn
This equation can be used to predict a Y value for a given set of input X values. SigmaXL uses the
method of least squares to solve for the model coefficients and constant term. Statistical tests of
hypothesis are provided for the model coefficients.
1. Open Customer Data.xlsx. Click Sheet 1 Tab (or press F4 to activate last worksheet). Click
SigmaXL > Statistical Tools > Regression > Multiple Regression. If necessary, click Use Entire
Data Table, click Next.
2. Select Overall Satisfaction, click Numeric Response (Y) >>, select Responsive to Calls and Ease of
Communications, click Continuous Predictors (X) >>. Fit Intercept and Display Residual Plots
are checked by default.
Note: Fit Intercept is optional. If unchecked the model will not fit an intercept (constant) for
the model. This is useful when you have strong a priori reasons to believe that Y = 0 when the X
or X’s are equal to 0 and the relationship is linear. An example would be Y=automobile fuel
consumption, X=automobile weight.
367
3. Click OK. The resulting Multiple Regression report is shown:
This model of Overall Satisfaction as a function of Responsiveness to Calls and Ease of

Communications looks very good with an R-Square value of 90%. Both Predictors are shown to
be significant with their respective P-Values < .05. Clearly we need to focus on these two X
factors to improve customer satisfaction.
4. The variance inflation factor (VIF) and Tolerance scores are used to measure multicollinearity or
correlation among predictors. VIF = 1 indicates no relation among predictors (which is highly
desirable – you will see examples of this in Design of Experiments); VIF > 1 indicates that the
predictors are correlated; VIF > 5 indicates that the regression coefficients are strongly
correlated and this will lead to poor estimates of the coefficients. Tolerance = 1/VIF.
5. The Durbin-Watson Test is used to determine if the residuals are autocorrelated. If either P-
Value is < .05, then there is significant autocorrelation. This is likely due to an external bias
factor affecting your process over time (e.g. “warm-up” effect, seasonality). This will also be
evident in the plot of residuals versus observation order. Autocorrelation in the residuals is not
a problem for this model.
6. If the Fit Intercept option is unchecked, the following diagnostics are not reported due to
statistical issues associated with the removal of the constant. These are: R-Square, R-Square
Adjusted, VIF (Variance Inflation Factor) and Tolerance collinearity diagnostics, DW (Durbin-
Watson) autocorrelation report, Anova report for Discrete factors (discrete factors will not be
permitted when "Fit Intercept" is unchecked). All other tests including residual diagnostics are
reported. Users can run the model with the "Fit Intercept" on to get the above statistics and
then run with "Fit Intercept" off.
368
7. A predicted response calculator allows you to enter X values and obtain the predicted response
value, including the 95% confidence interval for the long term mean and 95% prediction
interval for individual values:
8. Click the Mult Reg Residuals Sheet and scroll right to view the Residual Plots shown below:
Histogram of Regular Residuals for: Overall Satisfaction Normal Probability Plot of Regular Residualsfor: Overall Satisfaction
60 3.00
50 2.00
40 1.00
Frequency
NSCORE
30 0.00
20 -1.00
10 -2.00
0 -3.00
14
32
50
68
86
04
22
00
50
00
4
0
.9
.7
.5
.4
.2
.0
.0
.5
0.
0.
0.
0.
0.
1.
1.
0.
0.
1.
-0
-0
-0
-0
-0
-0
-1
-0
Regular Residuals Residuals
Regular Residuals vs Data Order for: Overall Satisfaction Regular Residuals vs Predicted Values for: Overall Satisfaction
1.50 1.500000
1.00 1.000000
0.50 0.500000
Regular Residuals
Regular Residuals
0.00 0.000000
-0.50 -0.500000
-1.00 -1.000000
-1.50 -1.500000
50
00
50
00
50
00
50
00
1
1
21
41
61
81
10
1.
2.
2.
3.
3.
4.
4.
5.
Residuals vs Responsive to Calls Residuals vs Ease of Communications

Response: Overall Satisfaction Response: Overall Satisfaction
1.058332953 1.058332953
0.558332953 0.558332953
Regular Residuals
Regular Residuals
0.058332953 0.058332953
-0.441667047 -0.441667047
-0.941667047 -0.941667047
00
50
00
50
00
50
00
50
00
40
90
40
90
40
90
40
90
1.
1.
2.
2.
3.
3.
4.
4.
5.
1.
1.
2.
2.
3.
3.
4.
4.
Responsive to Calls Ease of Communications
Residuals are the unexplained variation from the regression model (Y - Ŷ). We expect to see the
residuals normally distributed with no obvious patterns in the above graphs. Clearly this is not
the case here, with the Residuals versus Predicted Values indicating there is likely some other X
factor influencing the Overall Satisfaction. It would be appropriate to consider other factors in
the model.
9. SigmaXL also provides Standardized Residuals and Studentized (Deleted t) Residuals. The
standardized residual is the residual, divided by an estimate of its standard deviation. This
makes it easier to detect outliers. Standardized residuals greater than 3 and less than -3 are
considered large (these outliers are highlighted in red). Studentized deleted residuals are
369
computed in the same way that standardized residuals are computed, except that the ith
observation is removed before performing the regression fit. This prevents the ith observation
from influencing the regression model, resulting in unusual observations being more likely to
stand out.
10. Click Recall Last Dialog (or press F3). Change the Residual type to Standardized. Click OK. The
resulting Standardized Residual plots are displayed:
Histogram of Standardized Residuals for: Overall Satisfaction Normal Probability Plot of Standardized Residuals for: Overall Satisfaction
70 3.00
60
2.00
50
1.00
Frequency
40
NSCORE
0.00
30
-1.00
20
-2.00
10
0 -3.00
00
00
00
00
0
65
40
14
89
63
38
12
2
00
00
00
00
00
00
00
.8
.0
.3
.5
.8
.0
.0
.0
.0
.0
0.
1.
2.
2.
3.
4.
5.
0.
1.
2.
3.
4.
5.
6.
-3
-3
-2
-1
-0
-0
-4
-3
-2
-1
Standardized Residuals Standardized Residuals
Standardized Residuals vs Data Order for: Overall Satisfaction Standardized Residuals vs Predicted Values for: Overall Satisfaction
6.00 6.000
5.00 5.000
4.00 4.000
3.00 3.000
Standardized Residuals
2.00 2.000
1.00 1.000
0.00 0.000
-1.00 -1.000
-2.00 -2.000
-3.00 -3.000
-4.00 -4.000
-5.00 -5.000
50
00
50
00
50
00
50
00
1
1
21
41
61
81
10
1.
2.
2.
3.
3.
4.
4.
5.
Residuals vs Responsive to Calls Residuals vs Ease of Communications

Response: Overall Satisfaction Response: Overall Satisfaction
4.179854 4.179854
3.179854 3.179854
2.179854 2.179854
1.179854 1.179854
0.179854 0.179854
-0.820146 -0.820146
-1.820146 -1.820146
-2.820146 -2.820146
-3.820146 -3.820146
00
50
00
50
00
50
00
50
00
40
90
40
90
40
90
40
90
1.
1.
2.
2.
3.
3.
4.
4.
5.
1.
1.
2.
2.
3.
3.
4.
4.
Responsive to Calls Ease of Communications
11. Other diagnostic measures reported but not plotted include Leverage, Cook's Distance
(Influence), and DFITS. Leverage is a measure of how far an individual X predictor value
deviates from its mean. High leverage points can potentially have a strong effect on the
estimate of regression coefficients. Leverage values fall between 0 and 1. Cook's distance and
DFITS are overall measures of influence. An observation is said to be influential if removing the
observation substantially changes the estimate of coefficients. Cook’s distance can be thought
of as the product of leverage and the standardized residual; DFITS as the product of leverage
and the studentized residual.
12. The easiest way to identify observations with high leverage and/or influence is to plot the
measures on a run chart.
370
Multiple Regression with Continuous and Categorical Predictors

13. Click Recall Last Dialog (or press F3). Select Customer Type, click Categorical Predictor (X) >>.
We will not discuss residuals further so you may wish to uncheck Display Residual Plots. Keep
Overall Satisfaction as the Numeric Response (Y) and Responsive to Calls and Ease of
Communications as the Continuous Predictors (X):
14. Click OK. The resulting Multiple Regression report is shown:
371
15. The parameter estimates table now includes Customer Type 2 and Customer Type 3, but where
is Customer Type 1? Since Customer Type is a discrete predictor, SigmaXL applies “dummy
coding” (i.e. Customer Type 1 corresponds to Coded Variable 1 = 0 and Coded Variable 2 = 0;
Customer Type 2 has Coded Variable 1 = 1 and Coded Variable 2 = 0; and Customer Type 3 has
Coded Variable 1 = 0 and Coded Variable 2=1). Hence Customer Type 1 becomes the “hidden”
or reference value. Other statistical software packages may use a different coding scheme
based on -1,0,+1 instead of 0,1. The advantage of a 0,1 coding scheme is the relative ease of
interpretation when making predictions with the model.
16. In the Predicted Response Calculator enter the settings as shown:
Note that for Customer Type 1, you would enter Customer Type 2 = 0 and Customer Type 3 = 0;
for Customer Type 2, as shown, you entered Customer Type 2 = 1 and Customer Type 3 = 0; for
Customer Type 3 you would enter Customer Type 2 = 0 and Customer Type 3 = 1.
17. Note the addition of the Analysis of Variance for Categorical (Discrete) Predictors. The
Customer Type P-Value is .08, so we do not have strong evidence to keep this term in the
model. However, many practitioners will use an alpha value of 0.1 as a criteria for removal.
You are probably wondering why the P-Value for Customer Type is not lower given the results
we saw earlier using ANOVA. The change in P-Value is due to the inclusion of Responsive to
Calls and Ease of Communications in the model. We have higher scores for Ease of
Communications and Responsive to Calls with Customer Type 2. Statistically, Customer Type is
somewhat correlated to Responsive to Calls and Ease of Communications (VIF for Customer
Type = 1.66 and 1.39).
372
Part U – Logistic Regression
Binary Logistic Regression

Binary Logistic Regression is used to analyze the relationship between one binary dependent
variable (Y) and multiple independent numeric and/or discrete variables (X's). It is used to discover
the relationship between the variables and create an empirical equation of the form:
Ln(Py/(1-Py)) = b0 + b1*X1 + b2*X2 + ... + bn*Xn
This equation can be used to predict an event probability Y value for a given set of input X values.
SigmaXL uses the method of maximum likelihood to solve for the model coefficients and constant
term. Statistical tests of hypothesis and odds-ratios are provided for the model coefficients. The
odds-ratios identify change in likelihood of the event for one unit change in X.
An example application from medical research would be Y=Disease (Yes/No) and X’s = Age, Smoker
(Yes/No), Number Years of Smoking and Weight. The model coefficient P-Values would indicate
which X’s are significant and the odds-ratios would provide the relative change in risk for each unit
change in X.
We will analyze the familiar Customer Satisfaction data using Y=Discrete Satisfaction where the
values have been coded such that an Overall Satisfaction score >= 3.5 is considered a 1, and scores
< 3.5 are considered a 0. Please note, we are not advising that continuous data be converted to
discrete data in actual practice, but simply using the Discrete Satisfaction score for continuity with
the previous analysis.
SigmaXL > Statistical Tools > Regression > Binary Logistic Regression. If necessary, click Use
2. Select Sat-Discrete, click Binary Response (Y) >>; select Responsive to Calls and Ease of
Communications, click Continuous Predictors (X) >>; select Customer Type, click Categorical
Predictors (X) >>.
373
Note that Response Count (Y)/Sample Size (Trials) should be used when each record contains
both the number of occurrences along with associated sample size. This is common when
tracking daily quality data or performing design of experiments where each run contains a
response of the number of defects and sample size.
3. Click OK. The resulting Binary Logistic report is shown:
374
4. The Likelihood Ratio P-Value < .05 tells us that the model is significant. The P-Values for the
coefficients in the Parameter Estimates table confirm that Responsive to Calls and Ease of
Communications are significant.
5. The P-Value in Wald Estimates for Categorical (Discrete) Predictors table tells us that Customer
Type is not significant here.
Tip: Significance for categorical predictors should be based on the Wald Estimates not the P-
Values given in the Parameter Estimates table.
6. Note that Customer Type 1 is not displayed in the Parameter Estimates table. This is the
“hidden” reference value for Customer Type. Categorical predictors must have one level
selected as a reference value. SigmaXL sorts the levels alphanumerically and selects the first
level as the reference value.
7. Now we will rerun the binary logistic regression but remove Customer Type as a predictor.
Press F3 or click Recall SigmaXL Dialog to recall last dialog. Remove Customer Type by
highlighting Customer Type and double-clicking (or press the Remove button).
8. Click OK. The resulting Binary Logistic report is shown:
375
9. The Odds Ratios in the Parameter Estimates table tell us that for every unit increase in
Responsive to Calls we are 11.7 times more likely to obtain a satisfied customer. For every unit
increase in Ease of Communications we are 13.3 times more likely to obtain a satisfied
customer.
10. McFadden’s Pseudo R-Square mimics the R-square found in linear regression. This value varies
between 0 and 1 but is typically much lower than the traditional R-squared value. A value less
than 0.2 indicates a weak relationship; 0.2 to 0.4 indicates a moderate relationship; greater
than 0.4 indicates a strong relationship. Here we have an R-square value of 0.66 indicating a
strong relationship. This is also confirmed with the Percent Correctly Predicted value of 96%.
376
11. The Pearson, Deviance and Hosmer-Lemeshow Goodness of Fit tests are used to confirm if the
binary logistic model fits the data well. P-Values < .05 for any of these tests indicate a
significant lack of fit. Here the Hosmer-Lemeshow test is indicating lack of fit. Residuals
analysis will help us to see where the model does not fit the data well.
12. The measures of association are used to indicate the relationship between the observed
responses and the predicted probabilities. Larger values for Goodman-Kruskal Gamma, Somers’
D and Kendall’s Tau-a indicate that the model has better predictive ability.
13. The residuals report is given on the Sheet Binary Logistic Residuals. Three types of residuals
are provided: Pearson, Standardized Pearson and Deviance. The Standardized Pearson
Residual is most commonly used and is shown here plotted on a Run Chart. To create, click
SigmaXL > Graphical Tools > Run Chart, check Use Entire Data Table to select the Residuals
data, click Next, select Std. Pearson Residual as the Numeric Data Variable (Y). Click OK.
2.66
Run Chart - Std. Pearson Residual
0.66 Mean: -0.05
-1.34
-3.34
-5.34
-7.34
0
1
2
3
4
5
6
7
8
109
11
12
13
14
15
16
17
18
19
20
21
22
234
256
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
432
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
809
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
1099
2
2
Any Standardized Pearson Residual value that is less than -3 or greater than +3 is considered
extreme and should be investigated. There are 3 such outliers here: rows 24, 67, and 77 in the
residuals table. The +3.4 value indicates that the predicted event probability was low (.08) but
the actual result was a 1. The -6.6 value indicates that the predicted event probability was high
(.98) but the actual result was a 0. The large negative residuals have high Responsive to Calls
and Ease of Communications but dissatisfied customers. The reasons for these discrepancies
should be explored further but we will not do so here.
14. Reselect the Binary Logistic sheet. Scroll over to display the Event Probability calculator:
377
This calculator provides a predicted Event Probability for given values of X (in this case the
probability of a satisfied customer). Enter the values 3,3; 4,4; 5,5 as shown:
If Responsive to Calls and Ease of Communications are both equal to 3, the probability of a
satisfied customer is only .07 (7%); if Responsive to Calls and Ease of Communications are both
equal to 5, the probability of a satisfied customer is .9995 (99.95%)
15. Note that if the calculator includes predictors that are categorical (discrete), enter a 0 or 1 to
denote the selected level as shown below (using the original analysis which included Customer
Type):
If we wanted to select Customer Type 1, enter a 0 for both Customer Types 2 and 3. Customer
Type 1 is the hidden reference value.
378
Ordinal Logistic Regression

Ordinal Logistic Regression is used to analyze the relationship between one ordinal dependent
variable (Y) and multiple independent continuous and/or discrete variables (X's).
We will analyze the Customer Satisfaction data using Y=Loyalty –Likely to Recommend score which
contains ordinal integer values from 1 to 5, where 5 indicates that the customer is very likely to
recommend us and 1 indicates that they are very likely to not recommend us.
SigmaXL > Statistical Tools > Regression > Ordinal Logistic Regression. If necessary, click Use
2. Select Loyalty – Likely to Recommend, click Numeric Ordinal Response (Y) >>; select Responsive
to Calls and Ease of Communications, click Continuous Predictors (X) >>.
3. Click OK. The resulting Ordinal Logistic report is shown:
379
4. The Likelihood Ratio P-Value < .05 tells us that the model is significant. The low P-Values for
the coefficients confirm that Responsive to Calls and Ease of Communications are significant.
5. The Odds Ratios tell us that for every one-unit increase in Responsive to Calls the chance of a
Loyalty score of 1 versus 2 (or 2 versus 3, etc.) is reduced by a multiple of 0.36. This is not very
intuitive but will be easy to see when we use the Response Outcome Probability calculator.
6. McFadden’s Pseudo R-Square value is 0.185 indicating that this is a weak (but close to
moderate) degree of association. This is also confirmed with the Percent Correctly Predicted
value of 48%.
7. The Pearson and Deviance Goodness of Fit (GOF) tests are used to confirm if the ordinal logistic
model fits the data well. P-Values < .05 would indicate a significant lack of fit. Given that the
GOF P-Values are greater than .05, we conclude that there is no significant lack of fit.
380
8. Scroll across to the Response Outcome Probability calculator. This calculator provides
predicted outcome (event) probabilities for given values of X (in this case the probability of a
satisfied customer). Enter the values 3,3; 4,4; 5,5 as shown:
9. Referring to the Predicted Probability for each Level, if Responsive to Calls and Ease of
Communications are both equal to 3, we would expect to see typical loyalty scores of 3 (57%)
with some at 2 (20%) and 4 (20%); if Responsive to Calls and Ease of Communications are both
equal to 5, we would expect typical loyalty scores of 4 (53 %) with some at 3 (34%) and 5 (9%).
381
SigmaXL: Improve Phase Tools: Design of Experiments
(DOE)

SigmaXL: Improve Phase Tools: DOE
Part A – Overview of Basic Design of Experiments (DOE)

Templates
The DOE templates are similar to the other SigmaXL templates: simply enter the inputs and
resulting outputs are produced immediately. The DOE templates provide common 2-level designs
for 2 to 5 factors. These basic templates are ideal for training, but use SigmaXL > Design of
Experiments > 2-Level Factorial/Screening Designs to accommodate up to 19 factors with
randomization, replication and blocking.
Click SigmaXL > Design of Experiments > Basic DOE Templates to access these templates:
• Two-Factor, 4-Run, Full-Factorial
• Three-Factor, 4-Run, Half-Fraction, Res III
• Three-Factor, 8-Run, Full-Factorial
• Four-Factor, 8-Run, Half-Fraction, Res IV
• Four-Factor, 16-Run, Full-Factorial
• Five-Factor, 8-Run, Quarter-Fraction, Res III
• Five-Factor, 16-Run, Half-Fraction, Res V
After entering the template data, main effects and interaction plots may be created by clicking
SigmaXL > Basic DOE Templates > Main Effects & Interaction Plots. The DOE template must be
the active worksheet.
DOE Templates are protected worksheets by default, but this may be modified by clicking SigmaXL
> Help > Unprotect Worksheet.
Advanced analysis is available, but this requires that you unprotect the DOE worksheet. The
following example shows how to use Excel’s Equation Solver and SigmaXL’s Multiple Regression in
conjunction with a DOE template. Caution: If you unprotect the worksheet, do not change the
worksheet title (e.g. Three-Factor, Two-Level, 8-Run, Full-Factorial Design of Experiments). This
title is used by the Main Effects & Interaction Plots to determine appropriate analysis. Also, do not
modify any cells with formulas.
385
Part B – Three Factor Full Factorial Example Using DOE

Template
1. Open the file DOE Example - Robust Cake.xlsx. This is a Robust Cake Experiment adapted from
the Video Designing Industrial Experiments, by Box, Bisgaard and Fung.
2. The response is Taste Score (on a scale of 1-7 where 1 is "awful" and 7 is "delicious").
3. The five Outer Array Reps have different Cooking Time and Temperature Conditions.
4. The goal is to Maximize Mean and Minimize StDev of the Taste Score.
5. The X factors are Flour, Butter, and Egg. Actual low and high settings are not given in the video,
so we will use coded -1 and +1 values. We are looking for a combination of Flour, Butter, and
Egg that will not only taste good, but consistently taste good over a wide range of Cooking Time
and Temperature conditions.
6. Scroll down to view the Pareto of Abs. Coefficients for Average (Y).
7. The BC (Butter * Egg) interaction is clearly the dominant factor. The bars above the 95%
confidence blue line indicate the factors that are statistically significant; in this case only BC is
significant. Keep in mind that this is an initial analysis. Later, we will show how to do a more
powerful Multiple Regression analysis on this data. (Also, the Rule of Hierarchy states that if an
interaction is significant, we must include the main effects in the model used.)
8. The significant BC interaction is also highlighted in red in the table of Effects and Coefficients:
386
Calculation of Effects and Coefficients for Average (Y):
Constant A B C AB AC BC ABC
Avg(Avg(Y)) @ +1: 4.865 4.79 5.125 4.74 4.755 4.06 4.62
Avg(Avg(Y)) @ -1: 4.5 4.575 4.24 4.625 4.61 5.305 4.745
Effect (Delta): 0.365 0.215 0.885 0.115 0.145 -1.245 -0.125
Coefficient (Delta/2): 4.6825 0.1825 0.1075 0.4425 0.0575 0.0725 -0.6225 -0.0625
SE Coefficient: 0.2325 0.2325 0.2325 0.2325 0.2325 0.2325 0.2325 0.2325
T-value 20.1397849 0.7849 0.4624 1.9032 0.24731 0.31183 -2.6774 -0.2688
P-value 9.6928E-20 0.4383 0.6469 0.066 0.80625 0.75719 0.01161 0.7898
9. The R-Square value is given as 27%. This is very poor for a Designed Experiment. Typically, we
would like to see a minimum of 50%, with > 80% desirable.
R-Square: 27.03%
R-Square Adj.: 11.06%
S 1.4705
The reason for the poor R-square value is the wide range of values over the Cooking
Temperature and Time conditions. In a robust experiment like this, it is more appropriate
to analyze the mean response as an individual value rather than as five replicate values.
The Standard Deviation as a separate response will also be of interest.
10. If the Responses are replicated, SigmaXL draws the blue line on the Pareto Chart using an
estimate of experimental error from the replicates. If there are no replicates, an estimate
called Lenth’s Pseudo Standard Error is used.
11. If the 95% Confidence line for coefficients were to be drawn using Lenth’s method, the value
would be 0.409 as given in the table:
This would show factor C as significant.
12. Scroll down to view the Pareto of Coefficients for StDev(Y).
387
13. The A (Flour) main effect is clearly the dominant factor, but it does not initially appear to be
statistically significant (based on Lenth’s method). Later, we will show how to do a more
powerful Regression analysis on this data.
14. The Pareto chart is a powerful tool to display the relative importance of the main effects and
interactions, but it does not tell us about the direction of influence. To see this, we must look
at the main effects and interaction plots. Click SigmaXL > Design of Experiments > Basic DOE
Templates > Main Effects & Interaction Plots. The resulting plots are shown below:
388
15. The Butter*Egg two-factor interaction is very prominent here. Looking at only the Main Effects
plots would lead us to conclude that the optimum settings to maximize the average taste score
would be Butter = +1, and Egg = +1, but the interaction plot tells a very different story. The
correct optimum settings to maximize the taste score is Butter = -1 and Egg = +1.
16. Since Flour was the most prominent factor in the Standard Deviation Pareto, looking at the
Main Effects plots for StDev, we would set Flour = +1 to minimize the variability in taste scores.
The significance of this result will be demonstrated using Regression analysis.
389
17. Click on the Sheet Three-Factor 8-Run DOE. At the Predicted Output for Y, enter Flour = 1,
Butter = -1, Egg = 1 as shown:
The predicted average (Y-hat) taste score is 5.9 with a predicted standard deviation (S-hat) of
0.68. Note that this prediction equation includes all main effects, two-way interaction, and the
three-way interaction.
Multiple Regression and Excel Solver (Advanced Topics):

18. In order to run Multiple Regression analysis, we will need to unprotect the worksheet. Click
SigmaXL > Help > Unprotect Worksheet.
19. In the Coded Design Matrix, highlight columns A to ABC, and the calculated responses as
shown:
20. Click SigmaXL > Statistical Tools > Regression > Multiple Regression. Click Next.
390
21. Select Average (Y), click Numeric Response (Y) >>; holding the CTRL key, select B, C, and BC;
click Continuous Predictors (X) >> as shown:
22. Click OK. The resulting regression report is shown:
Note that the R-square value of 92.85% is much higher than the earlier result of 27%. This is
due to our modeling the mean response value rather than considering all data in the outer
array. Note also that the C main effect now appears as significant.
23. Click on the Sheet Three-Factor 8-Run DOE.
391
24. With the Coded Design Matrix highlighted as before, click SigmaXL > Statistical Tools >
Regression > Multiple Regression. Click Next.
25. Select StDev (Y), click Numeric Response (Y) >>; select A, click Continuous Predictors (X) >> as
shown:
26. Click OK. The resulting regression report is shown:
Note that Factor A (Flour) now shows as a statistically significant factor affecting the Standard
Deviation of Taste Score.
27. Now we will use Excel’s Equation Solver to verify the optimum settings determined using the
Main Effects and Interaction Plots.
392
28. Click on the Sheet Three-Factor 8-Run DOE. At the Predicted Output for Y, enter 1 for Flour.
We are setting this as a constraint, because Flour = +1 minimizes the Standard Deviation. Reset
the Butter and Egg to 0 as shown:
29. In Excel 2007 click Office Button | Excel Options | Add-Ins > Manage Excel Add-Ins, click Go; in
Excel 2010: File | Options. Ensure that the Solver Add-in is checked. If the Solver Add-in does
not appear in the Add-ins available list, you will need to re-install Excel to include all add-ins.
30. Click OK. Click Data | Analysis | Solver). Set the Solver Parameters as shown:
Cell J11 is the Y-hat, predicted average taste score. Solver will try to maximize this value. Cells
H11 to H13 are the Actual Factor Settings to be changed. Cells I11 to I13 are the Coded Factor
393
settings where the following constraints are given: I11=1; I12 >= -1; I12 <= 1; I13 >= -1; I13 <=1.
31. Click Solve. The solver results are given in the Predicted Output for Y as Butter = -1 and Egg =
1.
32. Solver indicates that a solution is found:
33. Click OK to keep the solution.
394
Part C – Design and Analysis of Catapult Full Factorial

Experiment
Catapults are frequently used in Six-Sigma or Design of Experiments

training. They are a powerful teaching tool and make the learning fun.
If you have access to a catapult, we recommend that you perform the
actual experiment and use your own data. Of course, you can also
follow along using the data provided. The response variable (Y) is
distance, with the goal being to consistently hit a target of 100 inches.
1. Click SigmaXL > Design of Experiments > 2-Level

Factorial/Screening > 2-Level Factorial/Screening Designs.
2. The Number of X Factors can be 2 to 19. Using process knowledge,

we will limit ourselves to 3 factors: Pull Back Angle, Stop Pin and Pin Height. Pull Back will be
varied from 160 to 180 degrees, Stop Pin will be positions 2 and 3 (count from the back), and
Pin Height will be positions 2 and 3 (count from the bottom).
3. Select Number of Factors = 3.
4. The available designs are then given as: 4-Run, 2**(3-1), 1/2 Fraction, Res III and 8-Run, 2**3,
Full-Factorial. If we had more than 5 factors, a Resolution III or Plackett-Burman Screening
design would typically be used. Here we will choose the 8-Run, 2**3, Full-Factorial design.
395
5. Next we will review the Power Information to help us determine the number of replicates. This
design currently shows the following:
These power calculations assume 3 center points:

Very Low Power to detect Effect = 1*StDev (1-Beta < 0.5);
Medium Power to detect Effect = 3*StDev (0.8 <= 1-Beta < 0.95).
Note: the power calculations require an estimate of experimental error, hence an assumption
of 3 center points is used.
We would like to have a reasonable chance (medium power) to detect an Effect = 2*StDev.
Change the Number of Replicates to 2. The Power Information is now:
Medium Power to detect Effect = 2*StDev (0.8 <= 1-Beta < 0.95);
Very High Power to detect Effect = 3*StDev (1-Beta >= 0.99).
We will therefore choose two replicates. The number of replicates will always be a tradeoff
between the desired power and the cost of the experimental runs.
Specify 2 or more blocks if there are constraints such as the number of runs per day or some
other known external “nuisance” variable (like 2 different catapults or 2 operators). Here we
will keep Blocks = 1 (i.e. no Blocking).
Center Points are useful to provide an estimate of experimental error with unreplicated
designs, and allow detection of curvature. Typically 3 to 5 center points are used. Here we will
not use center points because we have replicated the design twice and do not expect significant
396
curvature in the distance response. Furthermore, center points could not be set for Pin Height
and Stop Pin (without drilling additional holes!).
6. Complete the Factor Names, Level Settings and Response Name as shown:
8. You can enter information about the experiment in the fields provided. If you have access to a
catapult, perform the experimental runs in the given randomized sequence, and enter the
distance values in the Distance column.
9. If you are not able to perform the catapult experiment, open the file Catapult DOE V6.xlsx.
397
10. Before we begin the regression analysis we will have a quick look at the Main Effects and
Interaction Plots. Click SigmaXL > Design of Experiments > 2 Level Factorial/Screening > Main
Effects & Interaction Plots. The resulting plots are shown below:
Pull Back Angle is the dominant factor having the steepest slope. We can also see that the
interaction terms are weak with the almost parallel lines.
398
11. Click SigmaXL > Design of Experiments > 2-Level Factorial/Screening > Analyze 2-Level
Factorial/Screening Design.
12. We will use the default analyze settings (all terms in the model) to start. Click OK. The
resulting Analysis report is shown:
399
400
13. The model looks very good with an R-Square value of 99.9%! The standard deviation
(experimental error) is only 1.03 inches. Clearly Pull Back Angle is the most important predictor
(X factor), but all the main effects and two way interaction are significant. However, the three-
way interaction is not significant, so it should be removed from the model.
14. Click Recall Last Dialog (or press F3).
15. Remove the ABC interaction term as shown:
16. Click OK. The revised report is shown below:
401
402
17. All the terms in the model are now significant, and there is no evidence of lack of fit (P-Value
for lack-of-fit is 0.128 which is > .05).
18. The Residual Plots also look very good:

Histogram of Regular Residuals for: Distance Normal Probability Plot of Regular Residuals for: Distance
1.60
5
1.10
4
0.60
Frequency
NSCORE
3 0.10
-0.40
2
-0.90
1
-1.40
0 -1.90
21
62
03
44
00
00
00
00
4
0
.4
.0
.6
.2
.0
.0
.0
0.
0.
1.
1.
0.
1.
2.
3.
-1
-1
-0
-0
-3
-2
-1
Regular Residuals Regular Residuals
Regular Residuals vs Predicted Values for: Distance Regular Residuals vs Data Order for: Distance
2.000 2.00
1.500 1.50
1.000 1.00
Regular Residuals
0.500 Regular Residuals 0.50
0.000
0.00
-0.500
-0.50
-1.000
-1.00
-1.500
-1.50
-2.000
-2.00
00
00
00
00
0
0
.0
.0
.0
0.
0.
0.
0.
40
60
80
9
10
12
14
16
11
13
15
17
Fitted Values Observation Order
19. Scroll up to the Predicted Response Calculator. Enter the predicted values shown. These initial
settings were determined by trial and error.
403
20. Excel’s Solver may also be used to get a more exact solution:
21. The model prediction must then be confirmed with actual experimental runs at the given
settings of Pull Back Angle = 179.5, Stop Pin = 2, and Pin Height = 2.
22. Alternative settings to achieve the target distance may be obtained with Contour/Surface Plots.
Click SigmaXL > Design of Experiments > 2-Level Factorial/Screening > Contour/Surface Plots.
Set the Pin Height to 2 as shown (after clicking OK, you can use Recall SigmaXL Dialog to create
another Contour/Surface plot with Pin Height set to 3) :
404
23. Click OK. The following Contour and Surface Plots are displayed (with Pin Height = 2). Note the
contour line with Catapult target distance = 100 inches. Although pin settings are discrete, they
appear as continuous, so this will be a constraint in our selection of alternative settings. In
addition to Pull Back Angle = 179.5, Stop Pin = 2, Pin Height = 2, we see that Pull Back Angle =
171.4, Stop Pin = 3, Pin Height = 2 is also a valid setting. Alternative setting options are valuable
in a designed experiment because they allow you to select lowest cost optimum settings, or
settings that are easier to control.
Alternative
setting to
achieve a target
distance of 100
inches
Tip: Use the contour/surface in conjunction with the predicted response calculator to
determine optimal settings.
405
Part D – Design and Analysis of Response Surface

Experiment – Cake Bake
1. We will illustrate the use of response surface methods using a layer cake baking experiment.
The response variable is Taste Score (on a scale of 1-7 where 1 is "awful" and 7 is "delicious").
Average scores for a panel of tasters have been recorded. The X Factors are A: Bake Time (20
to 40 minutes) and B: Oven Temperature (350 to 400 F). The experiment goal is to find the
settings that maximize taste score. Other factors such as pan size need to be taken into
consideration. In this experiment, these are held constant.
2. Click SigmaXL > Design of Experiments > Response Surface > Response Surface Designs.
3. The Number of X Factors can be 2 to 5. Select Number of Factors = 2.
4. The available designs are sorted by number of runs. Increasing the number of runs allows for
uniform precision or blocking, but we will select the design with fewest runs, the 10-Run,
Central Composite Design (2 Ctr Pts).
5. As discussed earlier with the catapult experiment, it is a good practice to replicate an

experiment if affordable to do so. Here we will select Number of Replicates = 2 and check
Block on Replicates. Blocking on replicates allows us to perform the experiment over a two
week period, with each block corresponding to week number.
6. Change the Alpha Axial Value option to Face Centered (Alpha = 1.0). This simplifies the design
to a 3 level design, rather than a 5 level design with alpha = 1.414. (The trade-off is that we lose
the desirable statistical property of rotatability for prediction).
Tip: This alpha is the distance from the center point to the end of the axial (star) point. Do not
confuse this with Alpha for Pareto Chart which is the P-Value used to determine statistical
significance. Unfortunately the term “alpha” has been chosen by statisticians to define two
completely different things.
7. Enter Factor Names and Level Settings and Response Name as shown:
406
9. Open the file RSM Example – Cake Bake to obtain response values.
407
10. Click SigmaXL > Design of Experiments > Response Surface > Analyze Response Surface
Design.
11. We will use the default analyze settings (all terms in the model, including the block term) to
start. Click OK. The resulting Analysis report is shown:
408
12. The model looks very good with an R-Square value of 99.09%. The standard deviation
(experimental error) is only 0.19 on a 1 to 7 taste scale. All of the model terms are statistically
significant (P < .05), but the Block term is not, so it should be removed from the model. Note
that AA and BB denote the quadratic model terms.
13. Click Recall Last Dialog (or press F3).
14. Uncheck Include Blocks as shown:
409
15. Click OK. The revised report is shown below:
410
16. To create a contour and surface plot, click SigmaXL > Design of Experiments > Response
Surface > Contour/Surface Plots.
17. Click OK. The following Contour and Surface Plots are displayed.
Approximate
Maximum
Taste Score
411
18. Go to Analyze – 2 Factor RSM tab. Scroll across to the Predicted Response Calculator. Enter
the predicted values shown. These initial settings were determined from the contour plot as an
estimate to yield a maximum taste score.
19. Excel’s Solver may also be used to get a more exact solution:
20. Although the model is predicting values that exceed the maximum taste score of 7, this is
expected to give us the best possible settings for cook time and bake temperature. To quote
the eminent statistician George Box, “All models are wrong, some are useful”. Additional
experimental runs carried out at Time = 23.4 minutes and Temperature = 367.7 F. confirm that
these are ideal settings.
412
SigmaXL: Control Phase Tools: Statistical Process Control
(SPC) Charts

SigmaXL: Control Phase Tools: SPC
Part A - Individuals Charts
Tip: See Part E – Control Chart Selection Tool. The Control Chart Selection Tool makes it easy for
you to select the correct statistical process control chart depending on data type and
subgroup/sample type and size.
Individuals Chart Template

Click SigmaXL > Templates and Calculators > Basic Control Chart Templates >
Individuals Chart.
This template is also located at SigmaXL > Control Charts > Basic Control Chart Templates >
Individuals Chart.
See Measure Phase Part B – Templates and Calculators for Individuals Chart template
example:
Basic Control Chart Templates – Individuals Chart
415
Individuals Chart
1. Open Customer Data.xlsx, click on Sheet 1. Click SigmaXL>Control Charts>Individuals. Ensure
that entire data table is selected. If not, check Use Entire Data Table. Click Next.
2. Select Overall Satisfaction, click Numeric Data Variable (Y) >>. Ensure that Calculate Limits is
selected.
3. Click OK. Resulting Individuals control chart:
4. We have seen this data earlier as a run chart. The Control Chart adds calculated control limits.
Note that the Upper Control Limit exceeds the survey upper limit of 5. Here it would be
appropriate to change the UCL to 5.0. Click Recall SigmaXL Dialog menu or press F3 to recall
last dialog.
5. Select Overall Satisfaction, click Numeric Data Variable (Y) >>, change UCL to 5.
416
6. Click OK. Resulting Individuals chart with modified UCL:
Tip: You should not change the calculated control limits unless you have a legitimate reason
such as a boundary condition. Do not replace the control limits with specification limits – the
control chart will lose its statistical ability to detect assignable causes. We will redo the
Individuals chart for Overall Satisfaction later using the Individuals Nonnormal tool.
7. Open Delivery Times.xlsx. Click Sheet 1 Tab. This data set contains room service delivery time
deviations in minutes. The Critical Customer Requirement is target time +/- 10 minutes.
8. Click SigmaXL > Control Charts > Individuals. Ensure that entire data table is selected. If not,
check Use Entire Data Table. Click Next.
9. Select Delivery Time Deviation, click Numeric Data Variable (Y) >>, check Tests for Special
Causes, check Sigma Zone Lines. Ensure that Calculate Limits is selected.
10. Select Advanced Options, specify LSL = -10, Target = 0, USL = 10 as shown:
417
418
11. Click OK. The resulting chart is shown:
Some tests for special causes are indicated on the chart. If more than one test fails, the number
corresponds to the first failed test.
12. There are no points that exceed the +/- 3 sigma limits on this chart, but we see some indication
of instability with tests for special causes. The Tests for Special Causes report below the chart
provides detailed information about each observation identified as a special cause. Note that
the control chart also shows the +/- 1 sigma and +/- 2 sigma lines to aid in viewing these tests.
419
13. These tests for special causes can have defaults set to apply any or all of Tests 1-8. Test 2 can
be set to 7, 8, or 9 points in a row on same side of CL. Test 3 can be set to 6 or 7 points in a row
all increasing or decreasing. Test 7 can be set to 14 or 15 points in a row within 1 standard
deviation from CL. Click Sheet 1 Tab. Click SigmaXL > Control Charts > “Tests for Special
Causes” Defaults to run selected tests for special causes:
Note that these defaults will apply to Individuals and X-bar charts. Test 1 to 4 settings will be
applied to Attribute Charts.
14. Click on the Tab Display Options for Tests.
15. If you prefer to create a separate worksheet for each Tests for Special Causes report, choose
Create a new sheet option. The default is to display Tests for Special Causes on the same sheet
as the Control Chart. Note that this report will be overwritten when you add data or
recalculate control limits.
16. Click Save.
420
17. Click on Indiv Proc Cap Tab to view the Process Capability Report, which includes potential
(short term) capability indices Cp and Cpk:
18. While this process demonstrated some slight instability on the control charts, the bigger issue
was being late 6 minutes on average and having a Standard Deviation of 7.2 minutes! One
improvement implemented was rescheduling the service elevators so that Room Service and
Maintenance were not both trying to use them during peak times.
421
19. Click on the Indiv sheet. With 725 data points, you may want to have a closer look at the most
recent data. To do this, click SigmaXL Chart Tools > Show Last 30 Points. (If this menu item
does not appear, click on any cell adjacent to the chart.) The resulting chart is shown:
20. To reset the chart, click SigmaXL Chart Tools > Show All Data Points.
21. To enable scrolling, click SigmaXL Chart Tools > Enable Scrolling. A warning message is given:
Scrolling will clear user custom formats, but does not affect Tests for Special Causes.
22. Click OK. A scroll dialog appears allowing you to specify the Start Subgroup and Window
Width:
23. At any point, you can click Restore/Show All Data Points or Freeze Chart. Freezing the chart
will remove the scroll and unload the dialog. The scroll dialog will also unload if you change
worksheets. To restore the dialog, click SigmaXL Chart Tools > Enable Scrolling.
422
24. Click OK. The control chart appears with the scroll bar beneath it. You can also change the
Start Subgroup and Window Width and Update.
423
Individuals Charts: Advanced Limit Options – Historical Groups

1. Open Catapult Data - Before After Improvement.xlsx. Click Sheet 1 Tab. This data set contains
Catapult firing distances. Before_After denotes before improvement and after improvement.
The target distance is 100 inches, with the goal being to hit the target and minimize variation
about the target. We would like to use an individuals control chart with historical groups to
split the limits demonstrating the before versus after improvement.
2. Click SigmaXL > Control Charts > Individuals. Ensure that entire data table is selected. If not,
check Use Entire Data Table. Click Next.
3. Select Distance, click Numeric Data Variable (Y) >>; select Operator, click Optional X-Axis
Labels >>.
4. Click Advanced Options. Select Specify Historical Groups. Select Before_After, click Historical
Groups >>.
Note: Process Capability analysis is not permitted when Historical Groups are used.
424
5. Click OK. The resulting Individuals Control Chart with split limits based on historical groups is
shown, demonstrating a clear process improvement:
425
Individuals & Moving Range Charts

1. Open Customer Data.xlsx, click on Sheet 1. Click SigmaXL>Control Charts>Individuals &
Moving Range. Ensure that entire data table is selected. If not, check Use Entire Data Table.
Click Next.
2. Select Overall Satisfaction, click Numeric Data Variable (Y) >>. Ensure that Calculate Limits is
selected.
3. Click OK. Resulting Individuals & Moving Range control chart:
426
Individuals Charts for Nonnormal Data

(Box-Cox Transformation)
An important assumption for Individuals Charts is that the data be normally distributed (unlike the
X-Bar Chart which is robust to nonnormality due to the central limit theorem). If the data is
nonnormal, the Box-Cox Transformation tool can be used to convert nonnormal data to normal by
applying a power transformation. The Johnson transformation and other distributions may also be
used with Automatic Best Fit (see Measure Phase Tools, Part J – Process Capability, Capability
Combination Report Individuals Nonnormal).
1. Open the file Nonnormal Cycle Time2.xlsx. This contains continuous nonnormal data of
process cycle times. We performed a Process Capability study with this data earlier in the
Measure Phase, Part H.
2. Initially, we will ignore the nonnormality in the data and construct an Individuals Chart. Click
SigmaXL > Control Charts > Individuals. Ensure that entire data table is selected. If not, check
Use Entire Data Table. Click Next.
3. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. Select Calculate Limits. Check
Tests for Special Causes. Click OK.
427
4. The resulting Individuals Chart is shown:
This chart clearly shows that the process is “out-of-control”. But is it really? Nonnormality can
cause serious errors in the calculation of Individuals Chart control limits, triggering false alarms
(Type I errors) or misses (Type II errors).
5. We will now construct Individuals Control Charts for nonnormal data. Select Sheet 1 Tab (or
press F4). Click SigmaXL > Control Charts > Nonnormal > Individuals Nonnormal. Ensure that
the entire data table is selected. If not, check Use Entire Data Table. Click Next.
6. Select Cycle Time (Minutes), click Numeric Data Variable (Y) >>. We will use the default
selection for Transformation/Distribution Options: Box-Cox Transformation with Rounded
Lambda. Check Tests for Special Causes as shown:
428
7. Click OK. The resulting control charts are shown below:
Note that there are no out-of-control signals on the control charts, so the signals observed earlier
when normality was assumed were false alarms.
429
as:
430
Part B - X-Bar & Range/StDev Charts
X-Bar & R Charts

1. Open the file Catapult Data – Xbar Control Charts.xlsx. Each operator fires the ball 3 times.
The target distance is 100 inches. The Upper Specification Limit (USL) is 108 inches. The Lower
Specification Limit (LSL) is 92 inches.
2. Select B2:F22 (if not already selected); here, we will only use the first 20 subgroups to
determine the control limits.
3. Select SigmaXL > Control Charts > X-Bar & R. Do not check Use Entire Data Table!
Variables (Y) >>; select Operator, click Optional X-Axis Labels >>. Check Tests for Special
Causes as shown:
431
432
5. Click OK. Resulting X-bar & R charts:
6. This is currently a stable catapult process. Subgroups 21 to 25 were added afterwards. To add
the additional data to this chart, click SigmaXL Chart Tools > Add Data to this Control Chart as
shown. In Excel 2007+ click:
433
7. Note that the Add Data button does NOT recalculate the control limits. Once control limits are
established, they should only be recalculated when a deliberate process change or
improvement is introduced. (Control Limits can be recalculated using SigmaXL Chart Tools >
Recalculate Control Limits, but a warning message Are you sure that you want to recalculate
control limits? is given.)
8. The Tests for Special Causes report gives us more detail on the recent instability:
The X-bar chart and Tests for Special Causes report clearly shows that this process is now out of
control with an unstable mean. The process must be stopped, and the Out-of-Control Action
Plan must be followed to determine and fix the root cause. In this case, the assignable cause
was a change of rubber band requiring a reset of the pull back angle. The use of tests for
special causes gave us an early warning of this at observation number 22.
Note that the Range chart is in-control even though the X-Bar chart is out-of-control.
9. The tests for special causes can have defaults set to apply any or all of Tests 1-8. Test 2 can be
set to 7, 8, or 9 points in a row on same side of CL. Click SigmaXL > Control Charts > “Tests for
Special Causes” Defaults to run selected tests for special causes. (Note that these defaults will
apply to Individuals and X-bar charts. Test 1 to 4 settings will be applied to Attribute Charts.)
10. To add a comment to a data point, select SigmaXL Chart Tools > Add Data Label. Select Text
Label. Enter a comment as shown. Click on the data point to add the comment. Click Done.
11. Now we will look at Process Capability Indices for this process. Click on Sheet 1 (or press F4 to
activate last worksheet). Click SigmaXL > Control Charts > X-Bar & R. Check Use Entire Data
Table. Click Next. (Alternatively select B2:F27, press F3.)
12. Select Shots 1-3, click Numeric Data Variables (Y) >>.
13. Click Advanced Options. Enter LSL = 92, Target = 100, USL = 108.
434
14. The resulting dialog box settings are shown:
435
15. Click OK. Click X-Bar & R – Proc Cap sheet for the Process Capability report:
Note the difference between Pp and Cp; Ppk and Cpk. This is due to the process instability. If
the process was stable, the actual performance indices Pp and Ppk would be closer to the Cp
and Cpk values.
436
X-Bar & R Charts – Exclude Subgroups

After creating a control chart, you can specify subgroups (or rows) to exclude by using the Exclude
Data tool.
16. Click on Sheet 1 (or press F4 to activate last worksheet). Click SigmaXL > Control Charts > X-Bar
& R. Check Use Entire Data Table. Click Next.
17. Select Shots 1-3, click Numeric Data Variables (Y) >>. Ensure that Calculate Limits is selected.
Click OK.
18. The resulting X-bar & R charts are displayed:
The control limits here were calculated including subgroups 21 to 25 which have a known
assignable cause.
437
19. To calculate the control limits excluding subgroups 21 to 25, click SigmaXL Chart Tools >
Exclude Subgroups. Select Show Highlighted Points for Excluded Subgroups. Enter
21,22,23,24,25 as shown:
20. Click Exclude Subgroups. The control chart limits are recalculated and the excluded points are
highlighted:
Tip: You can also choose to show gaps for excluded subgroups or delete excluded subgroups
from the charts.
438
X-Bar & S Charts

The target distance is 100 inches. The Upper Specification Limit (USL) is 108 inches. The Lower
Specification Limit (LSL) is 92 inches.
2. Select B2:F22; here, we will only use the first 20 subgroups to determine the control limits.
3. Select SigmaXL > Control Charts > X-Bar & S.
4. Do not check Use Entire Data Table!
Variables (Y) >>; select Operator, click Optional X-Axis Labels >>. Check Tests for Special
Causes as shown:
439
6. Click OK. Resulting X-bar & S charts:
7. This is currently a stable catapult process. Subgroups 21 to 25 were added afterwards. To add
the additional data to this chart, click SigmaXL Chart Tools > Add Data to this Control Chart as
shown. In Excel 2007+ click:
440
Part C – Attribute Charts (P, NP, C, U)
P-Charts
1. Open New York Daily Cycle Time – Discrete.xlsx. This is data from the Sigma Savings and Loans
Company, New York location. Each day, the cycle time (in days) for completed loans and leases
was recorded. N indicates the number of loans counted. A Fail was recorded if the cycle time
exceeded the critical customer requirement of 8 days. Note that we are not recommending
that continuous data be converted to discrete data in this manner, but rather using this data to
illustrate the use of P charts for Discrete or Attribute data. P Charts (for Defectives) can have
fixed or varying subgroup sizes.
2. Select SigmaXL > Control Charts > Attribute Charts > P. Ensure that B3:E23 are selected, click
Next.
3. Select Fail, click Numeric Data Variable (Y) >>; select N, click Subgroup Column or Size >>. If
we had a fixed subgroup size, the numerical value of the subgroup size could be entered
instead of Column N. Ensure that Tests for Special Causes is unchecked.
Note: Tests for Special Causes cannot be used with varying subgroup sizes.
441
4. Click OK. The resulting P-Chart is shown:
The moving limits are due to the varying subgroup sizes. While this P-chart shows stability, a
much bigger concern is the average 41% failure rate to deliver the loans/leases in 8 days or less!
P-Charts: Advanced Limit Options – Historical Groups

1. Open New York Daily Cycle Time Discrete – Before After Improvement.xlsx. Before_After
denotes before improvement and after improvement. We would like to use a P-Chart with
historical groups to split the limits demonstrating the before versus after improvement.
2. Click SigmaXL > Control Charts > Attribute Charts > P. Ensure that the entire data table is
selected. If not, check Use Entire Data Table. Click Next.
3. Select Fail, click Numeric Data Variable (Y) >>; select N, click Subgroup Column or Size >>.
Ensure that Tests for Special Causes is unchecked.
4. Click Advanced Options. Select Specify Historical Group Column. Select Before_After, click
Historical Groups >>.
442
5. Click OK. The resulting P-Chart with split limits based on historical groups is shown,
demonstrating a clear process improvement:
443
NP-Charts
1. Open Attribute Data NP Chart Defectives.xlsx. The subgroup size is constant at 50. NP Charts
(for Defectives) require a fixed subgroup size.
2. Select SigmaXL > Control Charts > Attribute Charts > NP. Ensure that the entire data table is
3. Select Defectives, click Numeric Data Variable (Y) >>; enter 50 for Enter Subgroup Size. Check
Tests for Special Causes.
Note: Tests for Special Causes can always be used with NP charts because the subgroup size is
fixed, which results in constant control limits.
4. Click OK. The resulting NP-Chart is shown:
This is a stable process, with no tests for special causes flagged.
444
C-Charts
1. Open Attribute Data C Chart Defects.xlsx. C Charts (for Defects) assume a fixed subgroup size.
2. Select SigmaXL > Control Charts > Attribute Charts > C. Ensure that the entire data table is
3. Select Defects, click Numeric Data Variable (Y) >>. Check Tests for Special Causes.
Note: Tests for Special Causes can always be used with C charts because the subgroup size is
assumed to be fixed, which results in constant control limits.
4. Click OK. The resulting C-Chart is shown:
445
U-Charts
1. Open Attribute Data U Chart Defects.xlsx. U Charts (for Defects) can have fixed or varying
subgroup sizes.
2. Select SigmaXL > Control Charts > Attribute Charts > U. Ensure that the entire data table is
3. Select Defects, click Numeric Data Variable (Y) >>. Select N, click Subgroup Column or Size >>.
Ensure that Tests for Special Causes is unchecked (Tests for Special Causes cannot be used
with varying subgroup sizes).
4. Click OK. The resulting U-Chart is shown:
446
Part D – P’ & U’ Charts (Laney)
P' and U' (Laney) Control Charts are attribute control charts that should be used when the
subgroup/sample size is large and assumptions are not met. Typically you will see that the control
limits do not “look right,” being very tight with many data points appearing to be out-of-control.
This problem is also referred to as overdispersion. This occurs when the assumption of a Binomial
distribution for defectives or Poisson distribution for defects is not valid. Individuals charts are
often recommended in these cases, but Laney’s P’ and U’ charts are a preferred alternative.
References:
[1] Crossley, Mark L.(2007), The Desk Reference of Statistical Quality Methods, Second Edition,
Milwaukee, WI, ASQ Quality Press, pp. 352 – 355.
[2] Laney, David B., “P-Charts and U-Charts Work (But Only Sometimes),” Quality Digest,
http://www.qualitydigest.com/sept07/departments/what_works.shtml.
[3] Laney, David B. (2002), “Improved Control Charts for Attribute Data,” Quality Engineering
14:531–7.
[4] M. A. Mohammed and D. Laney (2006), “Overdispersion in health care performance data:
Laney’s approach,” Qual. Saf. Health Care 15; 383-384.
P’-Charts
1. Open Laney – Quality Digest – Defectives.xlsx. This data is used with permission from David
Laney and was used in the example given in reference [2].
2. We will begin with the creation of a regular P-Chart for this data. Select SigmaXL > Control
Charts > Attribute Charts > P. Ensure that the entire data table is selected. If not, check Use
Entire Data Table. Click Next.
3. Select Defectives as the Numeric Data Variable (Y), N as the Subgroup Column or (Size).
Ensure that Tests for Special Causes is unchecked (Tests for Special Causes cannot be used
with varying subgroup sizes). Click OK. The resulting P-Chart is shown:
447
At first glance, this chart suggests that the process is out of control. The problem here is
actually due to the Binomial assumption not being valid (also known as overdispersion). This
problem becomes evident when the sample sizes are large.
4. Select Sheet Defectives (or press F4). Click SigmaXL > Control Charts > Attribute Charts > P’
(Laney). Click Next.
5. Select Defectives, click Numeric Data Variable (Y) >>; select N, click Subgroup Column or Size
>>. (If we had a fixed subgroup size, the numerical value of the subgroup size could be entered
instead of Column N.)
6. Click OK. The resulting P’-Chart is shown:
448
7. Now we see that the process is actually “in-control.” Laney’s Sigma (Z) is a measure of the
overdispersion. See referenced articles for further details.
U’-Charts
1. Open Attribute Data – U’ Defects - Crossley.xlsx. This data is used with permission from Mark
Crossley and was used in the example given in reference [1]. Note that there are multiple
defects per unit.
2. We will begin with the creation of a regular U-Chart for this data. Select SigmaXL > Control
Charts > Attribute Charts > U. Ensure that the entire data table is selected. If not, check Use
Entire Data Table. Click Next.
3. Select Defects as the Numeric Data Variable (Y), N as the Subgroup Column (Size). Click OK.
The resulting U-Chart is shown:
This chart suggests that the process is out of control, but the problem here is actually due to
the Poisson assumption not being valid (also known as overdispersion). This problem becomes
evident when the sample sizes are large.
4. Select Sheet U’ Chart Defects - Crossley (or press F4). Click SigmaXL > Control Charts >
Attribute Charts > U’ (Laney). Click Next.
449
5. Select Defects, click Numeric Data Variable (Y) >>; select N, click Subgroup Column or Size >>.
(If we had a fixed subgroup size, the numerical value of the subgroup size could be entered
instead of Column N.) Ensure that Tests for Special Causes is unchecked (Tests for Special
Causes cannot be used with varying subgroup sizes).
6. Click OK. The resulting U’-Chart is shown:
7. Now we see that the process is actually “in-control.” Laney’s Sigma (Z) is a measure of the
overdispersion. See reference [1] for further details.
450
Part E – Control Chart Selection Tool
The Control Chart Selection Tool makes it easy for you to select the correct statistical process
control chart depending on data type and subgroup/sample type and size.
Data Types and Definitions

Continuous/Variable: Data that is measured on a continuous scale where a mid-point (or other
subdivision) has meaning. For example, when measuring cycle time, 2.5 days has meaning.
Other examples include distance, weight, thickness, length and cost. Customer Satisfaction on
a 1 to 5 scale can be considered as continuous in that a satisfaction score of 3.5 has meaning.
Continuous data is always in numeric format.
Discrete/Attribute: Data that is categorical in nature. If we have defect types 1, 2, and 3,

defect type 1.5 has no meaning. Other examples of discrete data would be customer
complaints and reasons for product return. Discrete data can be text or integer numeric
format.
Defective: An entire unit that is nonconforming to customer requirements. A unit may be

defective because of one or more defects. For example, an application form is good only if all
critical entry fields are correct. Any error in a critical field is a defect, resulting in a defective
form. A single form can have more than one defect.
Defect: Any specific nonconformity to customer requirements. There can be more than one
defect per unit or area of opportunity, such as the entry errors described above.
Subgroup/Sample: Data for a subgroup are usually collected within a short period of time to
ensure homogeneous conditions within the subgroup (common cause variation), in order to
detect differences between subgroups (special cause variation).
Subgroup/Sample Size: The number of observations within your sample, not the number of
samples. Subgroup sizes of 3 to 5 are common for continuous measures in parts
manufacturing, while individual measurements are common in chemical processes
(temperature, pH) and transactional areas (financial). Subgroup size for discrete data should be
a minimum of 50.
Subgroup/Sample Size is constant: The number of observations within your sample remains
fixed over time.
Subgroup/Sample Size varies: The number of observations within your sample varies over
time.
451
Subgroup/Sample Size is very large and assumptions not met: This applies to discrete data
when the subgroup sizes are approximately 5,000 or higher and the control limits do not “look
right,” being very tight with many data points appearing to be out-of-control. This problem is
also referred to as “overdispersion.” This occurs when the assumption of a Binomial
distribution for defectives or Poisson distribution for defects is not valid. (Note: If the problem
of overdispersion is apparent with your continuous data, use SigmaXL > Control Charts >
Advanced Charts > I-MR-R or I-MR-S).
Control Chart Selection Tool – Individuals Chart

1. Open Customer Data.xlsx, click on Sheet 1. Click SigmaXL > Control Charts > Control Chart
Selection Tool.
2. We would like to create a control chart of the Overall Satisfaction data. Since this can be
considered as continuous data, the data type is Continuous/Variable Data. The
subgroup/sample size is 1 (i.e. there is no subgrouping), so select Individuals (subgroup/sample
size = 1). At this point, we can choose Individuals or Individuals and Moving Range. We will
keep the simpler Individuals selection as shown. (Note that the above data types and
definitions can be viewed by clicking the Data Types and Definitions tab):
3. Click OK. This starts up the Individuals Chart dialog (see Part A – Individuals Charts for
continuation).
452
Control Chart Selection Tool – X-Bar & R Chart

The target distance is 100 inches. Select B2:F22; here we will only use the first 20 subgroups to
determine the control limits.
2. Click SigmaXL > Control Charts > Control Chart Selection Tool.
3. Since catapult shot distance measurement is continuous, we keep the default selection
Continuous/Variable Data. The catapult shot data are in subgroups, so select Subgroups
(subgroup/sample size > 1). The subgroup/sample size is small (3), so we will use the X-Bar &
Range Chart as shown:
4. Click OK. This starts up the X-Bar & Range dialog (see Part B – X-Bar & Range Charts for
continuation).
453
Control Chart Selection Tool – P-Chart

1. Open New York Daily Cycle Time – Discrete.xlsx. This is data from the Sigma Savings and Loans
Company, New York location. Each day, the cycle time (in days) for completed loans and leases
was recorded. N indicates the number of loans counted. A Fail was recorded if the cycle time
exceeded the critical customer requirement of 8 days.
2. Click SigmaXL > Control Charts > Control Chart Selection Tool.
3. Since this data is discrete, select Discrete/Attribute Data. We are looking at Defectives data
since each loan is a pass or fail, so select Defectives (unit is good/bad, pass/fail). The
subgroup/sample size varies day to day so Subgroup/Sample Size varies is selected as shown.
The recommended chart is the P-Chart (proportion defective):
4. Click OK. This starts up the P-Chart dialog (see Part C – P-Charts for continuation).
454
Part F – Advanced Charts: I-MR-R/S
I-MR-R Charts
If the within-subgroup variability is much smaller than between subgroup, the classical X-bar & R
(or S) chart will not work, producing numerous (false) alarms. The correct chart to use, in this case,
is the I-MR-R (or S) chart. The subgroup averages are treated as individual values (I-MR) and the
within subgroup ranges are plotted on the Range chart.
1. Open Multi-Vari Data.xlsx. Select Sheet Between. We saw this data previously using Multi-Vari
charts. First, we will incorrectly use the X-bar & R chart, and then apply the correct I-MR-R
chart.
2. Click SigmaXL > Control Charts > X-bar & R. Check Use Entire Data Table.
3. Click Next. Select Stacked Column Format. Select Measurement, click Numeric Data Variable
(Y) >>; select unit, click Subgroup Column or Size >>.
455
4. Click OK. Resulting X-bar & R chart:
Almost all of the data points in the X-bar chart are out-of-control! This is due to the small
within-subgroup variability (the control limits are calculated from the within-subgroup
variability).
5. Click Sheet Between. Select SigmaXL > Control Charts > Advanced Charts > I-MR-R.
6. Click Next. Select Stacked Column Format. Select Measurement, click Numeric Data Variable
(Y) >>; select unit, click Subgroup Column or Size >>.
456
7. Click OK. Resulting I-MR-R chart:
This chart is much cleaner, showing a stable Individuals and Range chart. The MR chart may be
trending up, but we would want to collect more data before making this conclusion. Typically,
you want at least 20 (30 preferred) subgroups before calculating final control limits.
457
SigmaXL Appendix
SigmaXL: Appendix
SigmaXL Version 8.0 Feature List Summary
460
SigmaXL: Appendix
461
SigmaXL: Appendix
462
SigmaXL: Appendix
463
SigmaXL: Appendix
Descriptive Statistics Options and Tolerance Interval

Calculator (Normal Exact)
Shapiro-Wilk (SW) and Kolmogorov-Smirnov-Lilliefors (KSL)

Normality Test
Shapiro-Wilk [1] is a popular and powerful alternative to the Anderson Darling normality test. It
was originally restricted to sample sizes less than 50. Royston [2] extended the sample size to 2000.
Rahman and Govidarajulu [3] extended the sample size further to 5000 and this has been
implemented in SigmaXL.
For sample sizes greater than 5000, Kolmogorov-Smirnov-Lilliefors (KSL) is used. Lilliefors [4]
modified the Kolmogorov-Smirnov test for the case of unknown mean and standard deviation.
Doornik-Hansen (DH) Normality Test

The Doornik-Hansen Normality Test [5] is a univariate or multivariate variate omnibus test based
on Skewness and Kurtosis. In descriptive statistics, the univariate version is used, which is based on
the method of Shenton and Bowman [6]. The bivariate version of DH is used in the Correlation
Matrix to test for bivariate normality.
This test is best suited for data with ties, i.e. “chunky” data. Anderson-Darling and Shapiro-Wilk are
severely affected by ties in the data and will trigger a low P-Value even if the data are normal.
Confidence and Tolerance Intervals

A confidence interval is described by the percentile of the population and the confidence
probability. Thus, the 95% confidence interval for the 80th percentile of the sample will consist of
a lower bound and an upper bound, where the true poulation 80th percentile will lie 95% of the
time.
A tolerance interval is the interval, described by a lower and upper bound, that a specified
minimum proportion of the sample’s values will lie with a given confidence probability. Thus, as an
example, the 95% tolerance interval for the range in which 80 percent of the sample will fall will
consist of a lower bound and an upper bound.
Notation:
n is the number of observations in the sample.

v is the degrees of freedom in the sample (n − 1) .
X is the mean of the sample.
464
SigmaXL: Appendix
s is the standard deviation of the sample.

p is the quantile (CI), or the percentage of the population in the interval (TI).
Zp is the pth quantile of the standard normal distribution.
α is the alpha probability - typically 0.05.
Tolerance Interval Calculator (Normal Exact)

When the sample can be assumed to be derived from a population that has a normal distribution,
then parametric assumptions can be used to ascertain the Tolerance Interval. For each of the
following, the lower L and upper U bounds are given by:
L = X − ks
U = X + ks
The value of k is shown below:
Normal Tolerance Interval - one sided exact

t1−α ,v ,λ
k=
n
where t1−α ,v ,λ is the 1 − α percentile of the non-central t distribution with v degrees of freedom
and non centrality parameter λ = n Z p .
Normal Tolerance Interval - two sided exact
 v χ p2 ,1,λ 
G ( z , n, p, k ) = 1 − Fv  2 
 k 
where: χ p ,1,λ is the p percentile of the non-central χ 2 distribution with 1 degree of freedom and
2
non centrality parameter λ = z 2 /n , and where Fv is the cdf of the χ 2 distribution with v degrees
of freedom. Then k is the root such that:
∞
2∫ G ( z, n, p, k )φ ( z )dz = 1 − α
0
where φ (z ) is the pdf of the standard normal distribution. SigmaXL uses a 40 point Gauss-Legendre
quadrature to evaluate the integral. See Krishnamoorthy and Mathew [7]. The approximation
method by Howe [8] is used to obtain initial values.
465
SigmaXL: Appendix
Percentile (Nonparametric) Confidence and Tolerance Intervals

This methodology is used when no a priori assumptions can be made regarding the distribution of
the data. Let x be the sorted data. The exact lower L and upper U bounds are given by:
L = xr
U = xs
The interpolated lower L and upper U bounds are given by:

L = λl xr +1 + (1 − λl ) xr
U = λu xs +1 + (1 − λu ) xs
Confidence Interval – two sided exact

r = Bα/2, n , p
s = B1−α/2, n , p + 1
where Bα , n , p is the α percentile of the binomial distribution with parameters n and p .
Confidence Interval – two sided interpolated
Let r and s be the exact order statistics for confidence interval. Then
−1
 r (1 − p )( Bn , p (r ) − α/2) 
λl = 1 + 
( n − r ) p (α /2 − B ( r − 1)) 
 n, p 
−1
 s (1 − p)( Bn , p ( s ) − (1 − α/2)) 
λu = 1 + 
( n − s ) p (1 − α /2 − B ( s − 1)) 
 n, p 
where Bn , p ( k ) is the cdf of the binomial distribution with parameters n and p . See Nyblom [9]
which is an extension of the interpolation method of Hettmansperger and Sheather [10] for median
using a double exponential distribution.
Tolerance Interval - two sided exact

k = B1−α , n , p + 1
r = ( n − k + 1)/2
s = n − r +1
where B1−α , n , p is the 1 − α percentile of the binomial distribution with parameters n and p .
Tolerance Interval - two sided interpolated

Let k , r and s be the two sided exact order statistics for tolerance interval.
1 − α/2 − Bn , p (k − 2)
λ=
Bn , p (k − 1) − Bn , p (k − 2)
466
SigmaXL: Appendix
v1 = λxr + (1 − λ ) xr +1
v2 = λxs + (1 − λ ) xs −1
If xs − v1 < v2 − xr then L = v1 and U = xs ; else L = xr and U = v2 . See Young and Mathew [11].
References for Descriptive Statistics Options and Tolerance

Interval Calculator (Normal Exact)
[1] “Shapiro-Wilk Test” in Wikipedia: The Free Encyclopedia, Wikimedia Foundation Inc.,
https://en.wikipedia.org/wiki/Shapiro%E2%80%93Wilk_test.
[2] Royston, Patrick (September 1992). “Approximating the Shapiro–Wilk W-test for non-
normality”, Statistics and Computing. 2 (3): 117–119.
[3] Rahman and Govidarajulu (1997). “A modification of the test of Shapiro and Wilk for
normality”, Journal of Applied Statistics. 24 (2): 219–236
[4] Lilliefors, H. (June 1967), "On the Kolmogorov–Smirnov test for normality with mean and
variance unknown", Journal of the American Statistical Association, Vol. 62. pp. 399–402.
[5] Doornik, J.A. and Hansen, H. (2008). “An Omnibus test for univariate and multivariate
normality”, Oxford Bulletin of Economics and Statistics, 70, 927-939.
[6] Shenton, L. R. and Bowman, K. O. (1977). “A bivariate model for the distribution of √b1 and
b2”, Journal of the American Statistical Association, Vol. 72, pp. 206–211.
[7] Krishnamoorthy, K. and Mathew, T. (2009), Statistical Tolerance Regions: Theory,
Applications, and Computation, Hoboken, NJ:Wiley.
[8] Howe, W. G. (1969), “Two-Sided Tolerance Limits for Normal Populations - Some
Improvements”, Journal of the American Statistical Association 64, pp. 610-620.
[9] Nyblom, J. (1992). “Note on interpolated order statistics”, Statistics and Probability Letters
14, pp. 129-131.
[10] Hettmansperger, T.P. and S.J. Sheather (1986). “Confidence Intervals Based on Interpolated
Order Statistics”, Statistics and Probability Letters 4, pp. 75-79.
[11] Young, D.S. and Mathew, T. (2014). “Improved nonparametric tolerance intervals based on
interpolated and extrapolated order statistics”, Journal of Nonparametric Statistics 26 (3),
pp. 415-432,
467
SigmaXL: Appendix
Statistical Details for Nonnormal Distributions and

Transformations
Maximum Likelihood Estimation (MLE)

Maximum likelihood estimates of the parameters are calculated by maximizing the likelihood
function with respect to the parameters. The likelihood function is simply the sum of the log of the
probability density function (PDF) for each uncensored observation, and the log of the complement
of the cumulative density function (CDF) for each right censored observation (in Reliability/Weibull
Analysis). Initial estimates are derived using a branch and bound algorithm.
The maximum likelihood estimates are then calculated using the Newton-Raphson method. This is
an iterative process that uses both the first and second derivatives to move to a point at which no
further improvement in the likelihood is possible.
The standard errors of the parameter estimates are derived from the Hessian matrix. This matrix,
which describes the curvature of a function, is the square matrix of second-order partial derivatives
of the function.
For some data sets, the likelihood function for threshold models is unbounded, and the maximum
likelihood methodology fails. In this context, a threshold is estimated using a bias correction
method. This is an iterative methodology that evaluates the threshold based on the difference
between the minimum value of the variate and the prediction for the minimum value, conditional
on the current values of the parameters.
References for MLE and Distributions:

Greene, W.H. Econometric Analysis 4th Ed Prentice Hall, New Jersey.
Johnson, N. L., and Kotz, S. (1990). "Use of moments in deriving distributions and some
characterizations", Mathematical Scientist, Vol. 15, pp. 42-52.
Johnson, N.L., Kotz, S. and Balakrishnan, N. (1994). Continuous Univariate Distributions-Volume 1,
Second Ed., Wiley, New York.
Johnson, N.L., Kotz, S. and Balakrishnan, N. (1995). Continuous Univariate Distributions-Volume 2,
Second Ed., Wiley, New York.
Nocedal, J. and Wright, S.J. (1999). Numerical Optimization, Springer-Verlag, New York.
Sleeper, A. (2006). Six Sigma Distribution Modeling, McGraw-Hill, New York.
468
SigmaXL: Appendix
Distributions
Beta Distribution
PDF
CDF
where B is the Beta function:
Range 0 ≤ x ≤ 1.
Shape1 parameter α > 0.
Shape2 parameter β > 0.
Γ(β) is the Gamma function and is described below under Gamma distribution.
469
SigmaXL: Appendix
Beta Distribution with Lower/Upper Threshold

PDF
CDF
where B is the Beta function.
Range 0 ≤ x ≤ 1.
Shape1 parameter α > 0.
Shape2 parameter β > 0.
Lower threshold θ 1
Upper threshold θ 2
Notes: Estimation of the 4 parameter Beta distribution is undertaken in two parts. In the first part,
initial parameter estimates are derived using the method of moments. The threshold parameters
are then held using these values, and the shape parameters are then estimated using maximum
likelihood.
Reference for Beta Distribution with Lower/Upper Threshold – Parameter Estimation:
Wang, J.Z. (2005). "A note on Estimation in the Four Parameter Beta Distribution", Comm in
Stats Simulation and computation, Vol. 34 pp. 495- 501.
470
SigmaXL: Appendix
Box-Cox Distribution with Threshold

PDF
CDF
Range 0 < x - θ < ∞.
Location parameter, µ, the mean.
Scale parameter, σ > 0, the standard deviation.
Shape parameter λ.
Threshold parameter θ < min(x).
Notes: The concentrated likelihood is used in the ML estimation. This implies that the location and
scale parameters are not estimated freely, but are derived as the mean and standard deviation of
the BoxCox transformed variate. The estimated parameters λ and θ are then used in the Box-Cox
(Power) transformation. See Transformations below.
471
SigmaXL: Appendix
Exponential Distribution
PDF
CDF
Range 0 ≤ x < ∞.
Scale parameter, α > 0, the mean.
Exponential Distribution with Threshold

PDF
CDF
Range 0 ≤ x < ∞.
Scale parameter, α > 0, the mean.
472
SigmaXL: Appendix
Gamma Distribution
PDF
CDF
where Γ(β) is the Gamma function:
Range 0 ≤ x < ∞.
Scale parameter α > 0.
Shape parameter β > 0.
Gamma Distribution with Threshold

PDF
473
SigmaXL: Appendix
CDF
Range 0 ≤ x - θ < ∞.
Scale parameter α > 0.
Half Normal Distribution

PDF
CDF
Range 0 ≤ x < ∞.
Scale parameter, σ > 0, the standard deviation
Reference for Application of Half Normal Distribution:
Chou, C., & H. Liu, (1998). "Properties of the half-normal distribution and its application to quality
control", Journal of Industrial Technology Vol. 14(3) pp. 4-7
474
SigmaXL: Appendix
Largest Extreme Value Distribution

PDF
CDF
Range −∞ < x < ∞.
Location parameter, µ, the mode.
Scale parameter σ > 0.
Logistic Distribution
PDF
CDF
475
SigmaXL: Appendix
Range −∞ < x < ∞.
Loglogistic Distribution
PDF
CDF
Range 0 < x < ∞.
Loglogistic Distribution with Threshold

PDF
CDF
476
SigmaXL: Appendix
Range 0 < x < ∞.
Lognormal Distribution
PDF
CDF
Range 0 < x < ∞.
Scale parameter, µ, the mean of ln(x) .
Shape parameter, σ > 0, the standard deviation of ln(x) .
Lognormal Distribution with Threshold

PDF
CDF
477
SigmaXL: Appendix
Range 0 < x < ∞.
Scale parameter, µ, the mean of ln(x) .
Shape parameter, σ > 0, the standard deviation of ln(x) .

Reference for Lognormal Distribution with Threshold – Parameter Estimation:
Giesbrecht, F. and A.H. Kempthorne (1966). "Maximum Likelihood Estimation in the Three-
parameter Lognormal Distribution", Journal of the Royal Statistical Society, B 38, pp. 257-264.
Normal Distribution
PDF
CDF
Range −∞ < x < ∞.
Scale parameter, σ > 0, the standard deviation.
Note: For consistency with other reports in SigmaXL such as Descriptive Statistics, the standard
deviation is estimated as the sample standard deviation using n-1 (rather than n).
478
SigmaXL: Appendix
Smallest Extreme Value Distribution

PDF
CDF
Range −∞ < x < ∞.
Location parameter, µ, the mode.
Weibull Distribution
PDF
CDF
Range 0 ≤ x < ∞.
Scale parameter, α > 0, the characteristic life.
479
SigmaXL: Appendix
Weibull Distribution with Threshold

PDF
CDF
Range 0 ≤ x < ∞.
Scale parameter, α > 0, the characteristic life.
Reference for Weibull Distribution with Threshold – Parameter Estimation:
Lockhart, R.A. and M.A. Stephens (1994)."Estimation and Tests of Fit for the Three-parameter
Weibull Distribution", Journal of the Royal Statistical Society, Vol.56(3), pp. 491-500.
480
SigmaXL: Appendix
Transformations
Box-Cox (Power) Transformation
λ. ≠ 0
λ. = 0
Range 0 < x < ∞.

Shape parameter λ.
Note: The optimum shape parameter,λ, is derived using a grid search in which the criteria function is the
standard deviation of the standardized transformed variable.
Box-Cox (Power) Transformation with Threshold
λ. ≠ 0
λ. = 0
481
SigmaXL: Appendix
Range 0 < x - θ < ∞.
Shape parameter λ.
Note: The parameters λ and θ are estimated using MLE as described above in Box-Cox Distribution
with Threshold.
Johnson Transformation
The Johnson Transformation selects one of the three families of distribution: SB (bounded), SL
(lognormal), and SU (unbounded) and the associated parameters so as to transform the data to be
normally distributed. The methodology follows Chou et al (1998) and uses the Anderson Darling P-
Value as the normality criteria.
Johnson Transformation - SB
z is N (0, 1)
Range ε < x < ε + λ.
Location parameter η > 0.
Scale parameter λ > 0.
Shape parameter γ unbounded.
Shape parameter ε unbounded.
482
SigmaXL: Appendix
Johnson Transformation - SL
z is N (0, 1)
Range x > ε
Johnson Transformation - SU
z is N (0, 1)
Range −∞ < x < ∞.
Scale parameter λ > 0.
483
SigmaXL: Appendix
References for Johnson Transformation:

Chou,Y., A.M. Polansky, and R.L. Mason (1998). "Transforming Nonnormal Data to Normality in
Statistical Process Control," Journal of Quality Technology, Vol. 30(2), pp. 133-141.
David, H.A. (1981). Order Statistics, John Wiley & Sons, New York
Tadikamalla,P.,R. (1980). "Notes and Comments: On Simulating Nonnormal Distributions",
Psychometrika,, Vol. 45(2), pp. 273-279.
Automatic Best Fit

SigmaXL uses the Anderson Darling P-Value as the criteria to determine best fit. All distributions
and transformations are considered and the model with the highest AD P-Value is initially selected
(denoted as adpvalmax). A search is then carried out for models that are close, having an AD P-
Value greater than adpvalmax - 0.1 (with an added criteria that AD P-Value be > 0.2), but having
fewer parameters than the initial best fit model. If a simpler model is identified, then this is
selected as the best fit.
Since AD P-Values are not available for distributions with thresholds (other than Weibull), an
estimate is obtained by transforming the data to normality and then using a modified Anderson
Darling Normality test on the transformed data. The transformed z-values are obtained by using
the inverse cdf of the normal distribution on the cdf of the nonnormal distribution. The Anderson
Darling Normality test assumes a mean = 0 and standard deviation = 1.
This approach, unique to SigmaXL, is an extension of the Chou methodology used in Johnson
Transformations and allows a goodness of fit comparison across all distributions and
transformations.
Another approach to comparing models is Akaike's information criterion (AIC) developed by

Hirotsugu Akaike. AIC is the MLE log-likelihood with a penalty for the number of terms in the model
(where the penalty factor also depends on the sample size, n). The AD P-Value method used by
SigmaXL has the advantage that it is not limited to models with maximum likelihood parameter
estimation.
References for Automatic Best Fit:
Chou,Y., A.M. Polansky, and R.L. Mason (1998). "Transforming Nonnormal Data to Normality in
Statistical Process Control," Journal of Quality Technology, Vol. 30(2), pp. 133-141.
D'Agostino, R.B. and Stephens, M.A. (1986). Goodness-of-Fit Techniques, Marcel Dekker.
http://en.wikipedia.org/wiki/Akaike_information_criterion
484
SigmaXL: Appendix
Process Capability Indices (Nonnormal)

Z-Score Method (Default)
The Z-Score method for computing process capability indices utilizes the same z-value
transformation to normality described above in Automatic Best Fit. The transformed z-values are
obtained by using the inverse cdf of the normal distribution on the cdf of the nonnormal
distribution. Normal based capability indices are then applied to the transformed z-values. This
approach offers two key advantages: the relationship between the capability indices and calculated
defects per million is consistent across the normal and all nonnormal distributions, and short term
capability indices Cp and Cpk can be estimated using the standard deviation from control chart
methods on the transformed z-values. The Z-Score method was initially developed by Davis Bothe
and expanded on by Andrew Sleeper. For further details, see Sleeper, Six Sigma Distribution
Modeling.
Percentile (ISO) Method
The Percentile method to calculate process capability indices uses the following formulas:
Ppu = (USL – 50th percentile)/(99.865 percentile – 50th percentile)
Ppl = (50th percentile – LSL)/(50th percentile – 0.135 percentile)
Ppk = min(Ppu, Ppl)
Pp = (USL – LSL)/( 99.865 percentile – 0.135 percentile)
References for Process Capability Indices:
Bothe, D.R. (1997). Measuring Process Capability, McGraw-Hill, New York.

Sleeper, A. (2006). Six Sigma Distribution Modeling, McGraw-Hill, New York.
485
SigmaXL: Appendix
Control Charts (Nonnormal)

Individuals – Original Data
as:
Individuals/Moving Range – Normalized Data
SigmaXL’s default setting is to display the two charts: Individuals – Original Data, Individuals &
Moving Range – Normalized Data (with tests for special causes unchecked).
486
SigmaXL: Appendix
Exact and Monte Carlo P-Values for Nonparametric and

Contingency Tests
Introduction
Nonparametric tests are popular because they do not assume that the sample data are normally
distributed, but they do, however, assume that the test statistic follows a normal or chi-square
distribution when computing the P-Value using “large sample” or “asymptotic” theory. Similarly,
the Chi-Square test for independence used in r x c contingency tables assumes that the test statistic
is distributed as chi-square. When the sample size is small or sparse, these assumptions may be
invalid and can result in substantial error in the computed P-Value. For example, Cochrane’s rule
[5] for r x c contingency tables is a commonly used guideline: no more than 20% of the cells should
have an expected value less than 5 and none should have an expected value less than 1.
The solution to the small sample problem is to compute exact P-Values using the true permutation
distribution of the test statistic. However, explicit full enumeration of all possible outcomes quickly
becomes computationally infeasible. For example, a 5x6 contingency table results in 1.6 billion
possible outcomes.
Network algorithms by Mehta & Patel [6, 7] use techniques from operations research such as
backward induction and forward probing to implicitly and efficiently solve the exact P-Value for
contingency tables. These algorithms have also been extended to work with continuous data. For
example, the nonparametric Mann-Whitney can be converted to a weighted singly ordered
contingency table and the exact P-Value solved using the network algorithm.
Some data sets are too large even for the network algorithm to solve in a reasonable time period,
but still do not satisfy Cochrane’s rule. In these cases, Monte Carlo simulation is used to sample
from the full enumeration of possible outcomes. 10,000 replications are typically used and 99%
confidence intervals provided for the Monte Carlo P-Values. If the P-Value is low, 1 million
replications are recommended.
Full Enumeration
To deal with a situation where the size of the data set does not justify the large sample asymptotic
assumption, we will first consider full enumeration. To illustrate, consider the problem of
evaluating the probability of getting a total of 3 or less when tossing two dice.
Evaluated for each of the n = 36 possible outcomes the sum of those situations (m) for which the
sum or the two dice is 3 or less. Since m = 3 {1,1}, {1,2}, {2,1}, the probability is m/n = 3/36 =1/12.
487
SigmaXL: Appendix
Full enumeration is 100% accurate, and always produces the exact P-Value. It works fine provided
n is not too large. However, for 20 dice, n, the number of possible outcomes is 3656 trillion, which
makes full enumeration infeasible. As mentioned above, even a relatively small problem such as a
5x6 contingency table results in 1.6 billion possible outcomes, again making full enumeration
infeasible.
Exact
Instead of evaluating each of the n possible outcomes, shortcuts can be derived which reduces the
number of outcomes that need to be evaluated. Starting with the first dice showing unity, once the
second dice exceeds 2 the remaining four outcomes can be counted as a block, since the statistic
for each case will exceed the test statistic. Thus rather than evaluating each as one would with full
enumeration, the block is evaluated. When the first dice shows two, and once the second dice
exceeds 1, the remaining 5 outcomes can again be counted as a block. Once the first dice shows 3,
all the remaining outcomes can be evaluated as a block, since again every possible outcome will
then exceed 3. Hence instead of enumerating 36 outcomes, one enumerates 6 outcomes {1,1},
{1,2}, {1, 3+}. {2,1}, {2,2+}, {3+,1+}. This type of model is referred to as Enhanced Enumeration. It
uses full enumeration of all possible outcomes node by node, and evaluates the statistic as a block
if the largest statistic possible from the current position exceeds the critical value (in which case the
probability of this path is added to the total probability), or if the smallest statistic possible from
the current position is less than the critical value, implying no augmentation of the total probability
[15].
A second methodology involves creating a network of a set of nodes in a number of stages. The
distances between nodes are defined so that the total distance of a path through the network is
the corresponding value of the test statistic. At each node, the algorithm computes the shortest
and longest path distances for all the paths that pass through that node. This works easily for those
statistics that can be evaluated as the linear sum of partial statistics evaluated at each node. This
methodology is commonly referred to in the literature as the Network Model, after the pioneering
work of Mehta and Patel [1, 6, 7].
The Network Model has advantages over enhanced enumeration in that the problem is broken
down to nodes. For example, if we had 4 dice, then there would be 5 nodes. The advantage of the
network model is that the length (and hence probability) at each node for the remaining nodes can
be estimated using a backward step through the network. Since networks tend to be sparse, the
set of all paths through the network which have the same probability up to the current node can be
retained. Thus, for any set of paths which have the same length up to the current node, one can
quickly read off the longest and shortest remaining sub-paths to the terminal node. Specifically, as
in enhanced enumeration, if every possible outcome does not contribute to the P-Value, the entire
488
SigmaXL: Appendix
set can be eliminated, while if every possible outcome contributes, then the probability of the
entire set can be evaluated as a closed form.
Thus an exact evaluation of a probability is identical to full enumeration, but significantly faster.
Again, as the size of the problem increases, the exact method also becomes infeasible since the
number of outcomes that do have to be evaluated becomes too large. The methodology used to
evaluate each of the exact nonparametric tests is described below.
Monte Carlo
Simulate the throw of the two dice nrep times, and count the number of times (m) that the sum of
the two dice is 3 or less. The Monte Carlo P-Value is then defined as m/nrep; this is an estimate of
the exact P-Value, and can be made more accurate by increasing nrep. Thus, with 20 dice, one
could choose a value of nrep = 100,000, and carry out the simulation to ascertain an unbiased
estimate of the P-Value. Thus the beauty of Monte Carlo is that it can be used on any problem, no
matter how large the total possible outcomes; however it is only asymptotically exact.
Two Tailed (Sided) Tests

A two tailed test is a statistical test in which the critical area of a distribution is two sided and tests
whether a sample is either greater than or less than a certain range of values. A one tailed test, on
the other hand, tests a one sided hypothesis, such as whether a sample is less than a certain range
of values. Thus a two tailed test will be the sum of a tail on the left hand side of a distribution, and
the corresponding tail on the right hand side of the distribution. If a distribution is symmetric, a
two tailed test will simply be twice the value of the one tailed test.
Formally, the two tailed probability is
Pr(|stat – mu| >= |statcrit – mu|)
summed over all possible combinations, where statcrit is the statistic evaluated for each possibility,
and mu is the expected value of the statistic. For symmetric distributions, 2*1-Sided and 2-Sided
will be identical. For non-symmetric distributions, such as the Runs Test and the Mann-Whitney
Test, there will be a difference between 2*1-Sided and 2-Sided.
489
SigmaXL: Appendix
Sample Sizes for Exact

Nonparametric Tests
The following sample size guidelines are based on our recommendations and those given in Siegel
[2] and Gibbons [3] for when exact nonparametric tests should be used rather than asymptotic
(Normal or Chi-Square approximation):
Sign Test: N <= 50 (exact is automatically utilized in the regular Sign Test)
Wilcoxon Signed Rank: N <= 15
Mann-Whitney: Each sample N <= 10
Kruskal-Wallis: Each sample N <= 5
Mood’s Median: Each sample N <= 10
Runs Test (Above/Below) or Runs Test (Up/Down) Test: N <= 50
These are minimum sample size guidelines. It is acceptable to use an exact test for larger sample
sizes, but computation time can become an issue especially for tests with two or more samples. In
those cases, one can always use a Monte Carlo P-Value with 99% confidence interval.
The maximum total sample size (total of all samples) that can be used for Nonparametric Exact or
Monte Carlo is 1000. The Runs Test for Randomness and Wilcoxon Test automatically switch to
large sample asymptotic if N > 1000.
One Way Chi-Square and Contingency Tables
Use Chi-Square Exact/Monte Carlo or Fisher’s Exact/Monte Carlo if more than 20% of the cells have
expected counts less than 5 or if any of the cells have an expected count less than 1 [4,5].
The maximum number of cells in a table is limited to 1000. For a One-Way Chi-Square the
maximum sum of cell counts is 1000. For a Contingency Table, the maximum sum of cell counts is
5000.
Power
change due to inadequate sample size.
490
SigmaXL: Appendix
Validation of SigmaXL Exact P-Values

SigmaXL Exact P-Values are validated by comparison to textbook examples [2, 3], published paper
examples [1, 6, 7, 8, 9] and other exact software such as StatXact, SPSS Exact, SAS and Matlab.
Monte Carlo P-Values are validated using 1e6 replications and compared against exact. Repeated
simulations are used to validate the confidence intervals.
1 Sample Tests
Sign Test: The number of sign changes in one direction is distributed binomial, and so the binomial
distribution function can be used directly to evaluate the tail probability [10].
Runs Test (Above/Below): Since the number of runs cannot exceed the length of the sample, it is
easy to evaluate the probability of r runs when there is a binary variable with n1 and n2 elements
[11].
Runs Test (Up/Down): Given the sample size and the number of up/down runs, the probability is
provided by published tables [12].
Wilcoxon Test: Given the signed rank statistic, look at all combinations of ranks that could
contribute to the observed value of the Wilcoxon critical value. For each of these cases, the
number of combinations that actually contribute is evaluated. These counts are then converted to
probabilities [13].
2 Sample Tests
Mann-Whitney Test: The continuous data is first converted to a weighted singly ordered
contingency table and the exact P-Value is solved using the network algorithm.
Each possible combination of ranks is systematically evaluated, but the process is cut short for any
given path if the largest statistic possible from the current position is less than the critical value (in
which case all paths from the present are added to the feasible combinations), or if the smallest
statistic possible from the current position is greater than the critical value (in which none of the
paths from the present are added to the feasible combinations).
K Sample Tests
Kruskal-Wallis Test: Full enumeration.
491
SigmaXL: Appendix
Mood’s Median Test: This test is evaluated using the Network Model. In each case, a chi-square
statistic is evaluated. The number of combinations at the end of each path that has a chi-square
statistic greater than the critical value is summed over all the enumerations, and is then converted
to a probability.
Contingency Tests
One Way Chi-Square Test: For a one dimension vector of frequencies, and expected probabilities,
the Chi Square test is evaluated using the Enhanced Enumeration model.
Two Way Fisher Exact Test: The exact probability for a two dimension r x c contingency table is
evaluated using the Network model [1]. This is a one tailed test, and includes the probabilities of
all the equivalent tables (those with same row and column counts) that have a lower or equal
probability as the sample table. As in the one dimensional case, tests are made to see if the
shortest path from the current position exceeds the path length of the observed table (in which
case the probability of all paths from the current position are added to the probability, or whether
the longest path from the current position is less than the path length of the current table, implying
no augmentation of the probability.
References for Exact and Monte Carlo P-Values

[1] Cyrus R. Mehta & Nitin R. Patel. (1983). "A network algorithm for performing Fisher's exact
test in r x c contingency tables." Journal of the American Statistical Association, Vol. 78, pp.
427-434.
[2] Siegel, S., & Castellan, N.J. (1988). Nonparametric Statistics for the Behavioral Sciences (2nd
Ed.). New York, NY: McGraw-Hill.
[3] Gibbons, J.D. and Chakraborti, S. (2010). Nonparametric Statistical Inference (5th Edition).
New York: Chapman & Hall.
[4] Yates, D., Moore, D., McCabe, G. (1999). The Practice of Statistics (1st Ed.). New York: W.H.
Freeman.
[5] Cochran WG. “Some methods for strengthening the common [chi-squared] tests.”
Biometrics 1954; 10:417–451.
[6] Mehta, C. R.; Patel, N. R. (1997) "Exact inference for categorical data," unpublished preprint.
See Table 7 (p. 33) for validation of Fisher’s Exact P-Value.
[7] Mehta, C.R. ; Patel, N.R. (1998). "Exact Inference for Categorical Data." In P. Armitage and T.
Colton, eds., Encyclopedia of Biostatistics, Chichester: John Wiley, pp. 1411–1422.
[8] Mehta, C.R. ; Patel, N.R. (1986). “A Hybrid Algorithm for Fisher's Exact Test in Unordered rxc
Contingency Tables,” Communications in Statistics - Theory and Methods, 15:2, 387-403.
492
SigmaXL: Appendix
[9] Narayanan, A. and Watts, D. “Exact Methods in the NPAR1WAY Procedure,” SAS Institute
Inc., Cary, NC.
[10] Myles Hollander and Douglas A. Wolfe (1973), Nonparametric Statistical Methods. New
York: John Wiley & Sons.
[11] https://onlinecourses.science.psu.edu/stat414/node/330
[12] Eugene S. Edgington, (1961). “Probability Table for Number of Runs of Signs of First
Differences in Ordered Series,” Journal of the American Statistical Association, Vol. 56, pp.
156-159.
[13] David F. Bauer (1972), “Constructing confidence sets using rank statistics,” Journal of the
American Statistical Association Vol 67, pp. 687–690.
[14] Hongsuk Jorn & Jerome Klotz (2002). “Exact Distribution of the K Sample Mood and Brown
Median Test,” Journal of Nonparametric Statistics, 14:3, 249-257.
[15] Valz, P. D. and Thompson, M. E. (1994), “Exact Inference for Kendall’s S and Spearman’s Rho
with Extensions to Fisher’s Exact Test in rxc Contingency Tables,” Journal of Computational
and Graphical Statistics, Vol 3(4), pp. 459- 472.
493
SigmaXL: Appendix
Hypothesis Test Assumptions Report
The Hypothesis Test Assumptions Report is an optional text report for the 1 Sample t-Test, Paired t-
Test, 2 Sample t-Test, One-Way ANOVA and Welch’s ANOVA. The cells are color highlighted: Green
(OK), Yellow (Warning) and Red (Serious Violation).
Normality
Each sample is tested for Normality using the Anderson-Darling (AD) test. If the AD P-Value is less
than 0.05, the cell is highlighted as yellow (i.e., warning – proceed with caution). The Skewness and
Kurtosis are reported and a note added, “See robustness and outliers.” An example report is
shown:
If the AD P-Value is greater than or equal to 0.05, the cell is highlighted as green.
A red highlight is not used for the Normality report.
Robustness
A minimum sample size for robustness to nonnormality is determined using the regression
equations in Basic Statistical Templates – Minimum Sample Size for Robust t-Tests and ANOVA.
These were derived from extensive Monte Carlo simulations to determine a minimum sample size
required for a test to be robust, given a specified sample Skewness and Kurtosis.
When there are 2 or more samples, a minimum sample size is calculated for each sample, and the
largest value is (conservatively) considered as the minimum sample size required for the test.
494
SigmaXL: Appendix
If the AD Normality test for every sample is greater than or equal to 0.05, the following message is
shown for robustness:
If each sample size is greater than or equal to the minimum for robustness, the minimum sample
size value is reported and the test is considered to be robust to the degree of nonnormality present
in the sample data:
If any sample size is less than the minimum for robustness, the minimum sample size value is
reported and a suitable Nonparametric Test is recommended. The cell is highlighted in red:
Recommended Nonparametric Tests – One Sample
The following Nonparametric Tests may be recommended for one sample:
1 Sample Sign
1 Sample Wilcoxon
1 Sample Wilcoxon – Exact
The 1 Sample Wilcoxon is more powerful that the 1 Sample Sign, but assumes that the data are
symmetrical. Wilcoxon is therefore the default recommendation, but if the sample is moderately
skewed then the Sign test is recommended.
Robustness to symmetry equations were developed in a manner similar to those used for
robustness to nonnormality above. Monte Carlo simulations were run using sample sizes (N) of 20
to 2000. Observed alpha values were determined empirically from the P-Values of 100,000
replicate Wilcoxon tests. For a given sample size, nonnormal data with Skew = 0.01 to +1.0 was
generated using the DiscoverSim Pearson Family function (see SigmaXL’s DiscoverSim Workbook
for details). The population test median (H0) used for the Wilcoxon test was estimated empirically
by averaging the 100,000 sample medians. The test was considered robust when the simulated
495
SigmaXL: Appendix
observed alpha was between .04 and .06, .i.e. within +/- 20% of the specified alpha of .05.
“Maximum Skewness” occurred when the observed alpha exceeded .04 or .06.
Other alpha values were not considered. Power was not considered.
This simulation was performed for Ha = “Not Equal To” (NE), “Greater Than” (GT), and “Less Than”
(LT). A regression analysis was performed on the Log10 (Maximum Skewness) versus Log10(N),
resulting in the following equations (with R-Square reported):
1. log10(Maximum Skewness NE) = 0.5391 - 0.5028 log10(N) (R-Square = 99.5%)
2. log10(Maximum Skewness GT) = 0.2380 - 0.5725 log10(N) (R-Square = 98.1%)
3. log10(Maximum Skewness LT) = 0.2796 - 0.5459 log10(N) (R-Square = 99.7%)
Note that if Skewness is negative, equation 2 is used for LT and equation 3 is used for GT.
For example, given a sample size of 30 and Ha = “Not Equal To,” solving for Maximum Skewness NE
in equation 1, the value is 0.626. Therefore, if the sample Skewness is less than or equal to 0.626,
Wilcoxon is recommended. If the sample Skewness is greater than 0.626, the Sign Test is
recommended.
If the Sign Test is recommended, Exact is not explicitly recommended since it is already “built in”
the Sign Test for N <= 50.
If the Wilcoxon Test is appropriate and N <= 15, then Wilcoxon – Exact is recommended (using the
rules of thumb in Sample Sizes for Exact).
Recommended Nonparametric Tests – Two Samples
The following Nonparametric Tests may be recommended for two samples:
2 Sample Mann-Whitney
2 Sample Mann-Whitney – Exact
If each sample N <= 10, then Mann-Whitney – Exact is recommended (using the rules of thumb in
Sample Sizes for Exact).
Recommended Nonparametric Tests – Three or More Samples (ANOVA)
The following Nonparametric Tests may be recommended for three or more samples:
Kruskal-Wallis
Kruskal-Wallis - Exact
496
SigmaXL: Appendix
Mood’s Median Test
Mood’s Median Test – Exact
Kruskal-Wallis is more powerful than Mood’s Median Test, but Mood’s Median is robust to outliers.
Kruskal-Wallis is therefore the default recommendation, but if any sample has outliers that are
Likely (2.2*IQR) or Extreme (3*IQR), then Mood’s Median is recommended. See Outliers (Boxplot
Rules) below for further details.
If Kruskal-Wallis is appropriate and each sample N <= 5, then Kruskal-Wallis – Exact is

recommended (using the rules of thumb in Sample Sizes for Exact).
If Mood’s Median is appropriate and each sample N <= 10, then Mood’s Median – Exact is
recommended.
Outliers (Boxplot Rules)

Each sample is tested for outliers using Tukey’s Boxplot Rules: Potential (> Q3 + 1.5*IQR or < Q1 –
1.5*IQR); Likely: 2.2*IQR; Extreme: 3*IQR. If outliers are present, a warning is given and
recommendation to review the data with a Boxplot and Normal Probability Plot and to consider
using a Nonparametric Test.
If no outliers are found, the cell is highlighted as green:
If a Potential or Likely outlier is found, the cell is highlighted as yellow:
Note that upper and lower outliers are distinguished.
497
SigmaXL: Appendix
If an Extreme outlier is found, the cell is highlighted as red:
The Anderson Darling normality test is applied to the sample data with outliers excluded. If this
results in an AD P-Value that is greater than 0.1, a notice is given, “Excluding the outliers, data are
inherently normal." The cell remains highlighted as yellow or red.
Randomness (Independence)
Each sample is tested for randomness (serial independence) using the Exact Nonparametric Runs
Test. If the sample data is not random, a warning is given and recommendation to review the data
with a Run Chart or Control Chart. See Nonparametric Runs Test for Randomness - Exact for
further details. Note that for this report, the option Values Equal to Median is set to both Counted
as Below and Counted as Above, with the reported P-Value being the larger of the two. This
results in a more conservative test to minimize false alarms.
If the Exact Nonparametric Runs Test P-Value is greater than or equal to 0.05, the cell is highlighted
as green:
If the Exact Nonparametric Runs Test P-Value is less than 0.05, but greater than or equal to 0.01,
the cell is highlighted as yellow:
498
SigmaXL: Appendix
If the Exact Nonparametric Runs Test P-Value is less than 0.01, the cell is highlighted as red:
Equal Variance
The test for Equal Variances is applicable for two or more samples. If all sample data are normal,
the F-Test (2 sample) or Bartlett’s Test (3 or more samples) is utilized. If any samples are not
normal, i.e., have an AD P-Value < .05, Levene’s test is used. If the variances are unequal and the
test being used is the equal variance option, then a warning is given and Unequal Variance (2
sample) or Welch’s Test (3 or more samples) is recommended.
If the test for Equal Variances P-Value is >= .05, the cell is highlighted as green:
If the test for Equal Variances P-Value is >= .05, but the Assume Equal Variances is unchecked (2
sample) or Welch’s ANOVA (3 or more samples) is used, the cell is highlighted as yellow:
If the test for Equal Variances P-Value is < .05, and the Assume Equal Variances is checked (2
sample) or regular One-Way ANOVA (3 or more samples) is used, the cell is highlighted as red:
499
SigmaXL: Appendix
If the test for Equal Variances P-Value is < .05, and the Assume Equal Variances is unchecked (2
sample) or Welch’s ANOVA (3 or more samples) is used, the cell is highlighted as green:
500
SigmaXL: Appendix
Attribute Measurement Systems Analysis
Percent Confidence Intervals (Exact Versus Wilson Score)

Confidence intervals for binomial proportions have an "oscillation" phenomenon where the
coverage probability varies with n and p. Exact (Clopper-Pearson) is strictly conservative and will
guarantee the specified confidence level as a minimum coverage probability, but results in wide
intervals. This is recommended only for applications requiring strictly conservative intervals.
Wilson Score has mean coverage probability matching the specified confidence interval. Since the
Wilson Score intervals are narrower and thereby more powerful, they are recommended for use in
Attribute MSA studies due to the small sample sizes typically used [1, 2, 3].
Kappa
Kappa ( κ ) measures the degree of agreement between appraisers when they asses the same
sample. Cohen’s Kappa, used in the Attribute MSA template, is only defined for two raters. Fleiss’
Kappa can be used for multiple assessors [4]. Thus Kappa provides a mechanism for evaluating the
consistency of assessors.
Let P be the measure of agreement between the assessors, and Pe be the measure of agreement
that would occur by chance. Then:
P − Pe
κ=
1 − Pe
The denominator provides the measure of agreement above chance, while the numerator provides
the actual measure of agreement above chance. Thus κ is constrained between -1 and +1, with +1
implying complete consistency or perfect agreement between assessors, zero implying no more
consistency between assessors than would be expected by chance and -1 implying perfect
disagreement.
Rule-of-Thumb Interpretation for Kappa

Fleiss [4, p. 609] gives the following rule of thumb for interpretation of Kappa:
Kappa: >= 0.75 or so signifies excellent agreement, for most purposes, and <= 0.40 or so
signifies poor agreement.
AIAG recommends the Fleiss guidelines [5].
In Six Sigma process improvement applications, a more rigorous level of agreement is commonly
used. Futrell [6] recommends:
501
SigmaXL: Appendix
The lower limit for an acceptable Kappa value (or any other reliability coefficient) varies
depending on many factors, but as a general rule, if it is lower than 0.7, the measurement
system needs attention. The problems are almost always caused by either an ambiguous
operational definition or a poorly trained rater. Reliability coefficients above 0.9 are
considered excellent, and there is rarely a need to try to improve beyond this level.
SigmaXL uses the guidelines given by Futrell and color codes Kappa as follows: >= 0.9 is green, 0.7
to < 0.9 is yellow and < 0.7 is red. This is supported by the following relationship to Spearman Rank
correlation and Percent Effectiveness/Agreement (applicable when the response is binary with an
equal proportion of good and bad parts):
Kappa = 0.7; Spearman Rank Correlation = 0.7; Percent Effectiveness = 85%; Percent
Agreement = 85% (two trials)
Kappa = 0.9; Spearman Rank Correlation = 0.9; Percent Effectiveness = 95%; Percent
Agreement = 95% (two trials)
(Note that these relationships do not hold if there are more than two response levels or the
reference proportion is different than 0.5).
Concordance
Let two assessors rank n subjects, such that x is the rankings of assessor 1, and y is the rankings
assessor 2. If for any pair of observations (i, j ) , the ranking of the two assessors agree - that is
xi > x j and yi > y j , or xi < x j and yi < y j , then the pair of observations are said to be
concordant. Pairs that are not concordant are said to be discordant. If x were a measurement of
exercise, and y a measurement of fitness, then we would expect high levels of concordance.
Kendall’s Coefficient of Concordance

Kendall's Coefficient of Concordance (Kendall's W ) is a measure of association for discrete ordinal
data, used for assessments that do not include a known reference standard. Kendall’s coefficient
of concordance ranges from 0 to 1: A coefficient value of 1 indicates perfect agreement. If the
coefficient = 0, then agreement is random, i.e., the same as would be expected by chance.
There is a close relationship between ρ s , Spearman’s correlation coefficient and W , Kendall’s

coefficient of concordance. W can be calculated directly from the mean ( ρ s ) of the pairwise
Spearman correlations using the relationship [7]:
(k − 1) ρ s + 1
W =
k
where k is the number of trials (within) or trials*appraisers (between) among which Spearman’s
correlation coefficients are computed.
502
SigmaXL: Appendix
Confidence limits for Kendall’s Concordance cannot be solved analytically, so are estimated using
bootstrapping. Ruscio [8] demonstrates the bootstrap for Spearman’s correlation and we apply
this method to Kendall’s Concordance. The data are row wise randomly sampled with replacement
to provide the bootstrap sample (N = 2000). W can be derived immediately from the mean value of
the Spearman’s correlation matrix from the bootstrap sample. The bias corrected values of α and
1 − α are then used to determine the confidence interval from the sorted bootstrap vector W. In
Efron’s bias-corrected and accelerated bootstrap, the value of α is adjusted to account for skew
and/or bias. This has been determined to be a superior bootstrap method compared to the simple
percentile method [9]. A fixed seed is used so that the report results are consistent.
Rule-of-Thumb Interpretation for Kendall’s Coefficient of Concordance
SigmaXL uses the following “rule-of-thumb” interpretation guidelines: >= 0.9 very good agreement
(color coded green); 0.7 to < 0.9 marginally acceptable, improvement should be considered
(yellow); and < 0.7 unacceptable (red).
This is consistent with Kappa and is supported by the relationship to Spearman’s correlation.
Note however that in the case of Within Appraiser agreement with only two trials, then the rules
should be adjusted: very good agreement is >= 0.95 and unacceptable agreement is < 0.85.
Kendall’s Correlation Coefficient

Kendall's Correlation Coefficient (Kendall's tau-b) is a measure of association for discrete ordinal
data, used for assessments that include a known reference standard. Kendall’s correlation
coefficient ranges from -1 to 1: A coefficient value of 1 indicates perfect agreement. If coefficient
= 0, then agreement is random, i.e., the same as would be expected by chance. A coefficient value
of -1 indicates perfect disagreement. Kendall's Correlation Coefficient is a measure of rank
correlation, similar to the Spearman rank coefficient, but uses concordant and discordant pairs and
corrects for ties.
Let two assessors rank n subjects, such that x is the rankings of assessor 1, and y is the rankings of
xi > x j and yi > y j , or xi < x j and yi < y j , then the pair of observations are said to be
concordant. Pairs that are not concordant are said to be discordant. If x were a measurement of
exercise, and y a measurement of fitness, then we would expect high levels of concordance.
Define:
503
SigmaXL: Appendix
Cij = ∑∑xhk + ∑∑xhk

h <i k < j h >i k > j
Dij = ∑∑xhk + ∑∑xhk

h <i k > j h >i k < j
Then the number of concordant ( nc ) and discordant ( nd ) pairs are derived as:
nc = 0.5P where P = ∑xij Cij
i, j
nd = 0.5Q where Q = ∑xij Dij

i, j
and the Kendall Correlation τ b is:

2(nc − nd )
τb =
Dr Dc
where:
Dr = n 2 − ∑Ri2
i
Dc = n 2 − ∑C 2j
j
504
SigmaXL: Appendix
Rule-of-Thumb Interpretation for Kendall’s Correlation Coefficient
SigmaXL uses the following “rule-of-thumb” interpretation guidelines:
>= 0.8 very good agreement (color coded green);
0.6 to < 0.8 marginally acceptable, improvement should be considered (yellow);
< 0.6 unacceptable (red).
These values were determined using Monte Carlo simulation with correlated integer uniform
distributions. They correspond approximately to Spearman 0.7 and 0.9 when there are 5 ordinal
response levels (1 to 5). With 3 response levels, the rule should be modified to 0.65 and 0.9.
References for Attribute Measurement Systems Analysis

[1] Agresti, A. and Coull, B.A.(1998). “Approximate is Better than “Exact” for Interval Estimation
of Binomial Proportions.” The American Statistician, 52, 119–126.
[2] Clopper, C.J., and Pearson, E.S. (1934). “The Use of Confidence or Fiducial Limits Illustrated
in the Case of the Binomial.” Biometrika, 26, 404–413.
[3] Newcombe, R. (1998a). “Two-sided confidence intervals for the single proportion:
Comparison of seven methods.” Statistics in Medicine,17, 857-872.
[4] Fleiss, J.L. (2003). Statistical Methods for Rates and Proportions, 3rd Edition, Wiley & Sons,
NY.
[5] Automotive Industry Action Group AIAG (2010). Measurement Systems Analysis MSA
Reference Manual, 4th Edition, p. 137.
[6] Futrell, D. (May 1995). “When Quality Is a Matter of Taste, Use Reliability Indexes,” Quality
Progress, 81-86.
[7] Siegel, S., & Castellan, N.J. (1988). Nonparametric statistics for the behavioral sciences (2nd
Ed.). New York, NY: McGraw-Hill. p. 262.
[8] Ruscio, J. (2008). “Constructing Confidence Intervals for Spearman’s Rank Correlation with
Ordinal Data: A Simulation Study Comparing Analytic and Bootstrap Methods,” Journal of
Modern Applied Statistical Methods, Vol. 7, No. 2, 416-434.
[9] Bradley Efron, (1987). “Better bootstrap confidence intervals (with discussion),” J. Amer.
Statist. Assoc. Vol. 82, 171-200.
[10] http://en.wikipedia.org/wiki/Kendall_tau_rank_correlation_coefficient
505
SigmaXL: Appendix
Chi-Square Tests and (Contingency) Table Associations
This section describes a number of tests and measures of association that are derived from a
contingency table. These tables depict the observed counts of two variables, shown in r rows and
c columns, with a row sum vector R and a column sum vector C . The total number of cells is n .
In a contingency table, xij represents the number of elements that correspond to row attribute i
and column attribute j .
The contingency table is described using the following summary statistics:

nij , the observed counts or number of elements in each cell.
ni , the sum of the elements in each row.
n j , the sum of the elements in each column.

eij = ni n j /n , the expected count or row-column product for that cell as a proportion of the total
count.
rij = nij − eij , the raw residual or difference between the actual and expected cell count.
sij = rij / eij , the standardized residual.
aij = sij / (1 − ni /n)(1 − n j /n) , the adjusted (standardized) residual expressed as a z score.
sij2 , the cell’s contribution to the chi-squared statistic.
Chi-Square Tests for Nominal Categories

Pearson Chi-squared
Pearson’s chi-squared test [3] is applied to a contingency table to evaluate how likely it is that this
observed table occurs by chance. Given the observed marginals (row and column counts), the
expected cell observation eij = Ri C j /n . The statistic:
χ 2 = ∑( xij − eij ) 2 /eij
i, j
is distributed χ 2 with df = (r − 1)(c − 1) degrees of freedom.
Likelihood Ratio Chi-squared

The likelihood ratio chi-squared test [3] is applied to a contingency table to evaluate the likelihood
of the observed table under the null hypothesis relative to the maximum likelihood. The statistic:
506
SigmaXL: Appendix
χ 2 = 2∑eij ln ( xij /eij )

i, j
is distributed χ with df degrees of freedom.

2
McNemar-Bowker Symmetry
For square tables, the McNemar statistic [11] tests the hypothesis that the contingency table is
symmetric. It is defined as:
( xij − x ji ) 2
M =∑
i< j xij + x ji
for all xij + x ji > 0 . Under the null, M is distributed χ 2 with n(n − 1)/2 degrees of freedom.
Measures of Association for Nominal Categories

Confidence bounds are evaluated as t ± zα s , where z is the inverse of the standard normal cdf, t
is the statistic and s is the asymptotic standard error (ASE1) of the statistic. See [9,12] for ASE1
formulas.
Perason’s Phi
The Phi ( φ ) statistic [1,9] is defined as:
φ = χ 2 /n
where χ 2 is the Pearson Chi-squared statistic. For a 2x2 table, φ is a measure of association for two
binary variables, and equivalent to Pearson Correlation, measures the percent of concentration of
cases on the diagonal. See Cohen [15} for interpretation rules-of-thumb.
Cramer’s V
The Cramer’s V statistic [3, 9] measures the association between two variables as a percentage of
their maximum possible variation. It has a range of 0 ≤ V ≤ 1 . It is defined as:
V= χ 2 /(n(m − 1))
where χ 2 is the Pearson Chi-squared statistic, and m = min (r , c)
Contingency Coefficient
The contingency statistic [3] is defined as:
z = χ 2 /(n + χ 2 )
where χ 2 is the Pearson Chi-squared statistic. Unlike the other statistics, it does not have a
specified range.
507
SigmaXL: Appendix
Cohen’s Kappa
For square tables, Cohen’s κ statistic [3,5] tests the null hypothesis that there is no association
between the row and column variables. It is calculated from the observed and expected
frequencies on the diagonal of a square contingency table. It is frequently used to provide a
mechanism for evaluating the consistency of assessors. It is defined as:
n∑xii − ∑Ri Ci
κ= i i
n 2 − ∑Ri Ci
i
A value of κ = 0 implies no association, while κ = 1 implies perfect association.
Goodman-Kruskal Lambda
Lambda [3] is defined as the proportional reduction in error of prediction of the dependent variable
as a consequence of using the information in the independent variable.
Column as dependent, row as independent:

∑x im − Cm
λr = i
n − Cm
Row as dependent, column as independent:
∑x j
mj − Rm
λc =
n − Rm
Symmetric version:
∑x
i
im + ∑xmj − Cm − Rm
j
λs =
2n − Cm − Rm
where xim and xmj are the largest cell count in row i and column j respectively, and Rm and Cm are
the largest row and column subtotals respectively.
Goodman-Kruskal Tau
Tau [9] is the same as lambda, except the probabilities are specified by marginal proportions.

n∑xij2 /Ri − ∑C 2j
i, j j
τr =
n − ∑C
2 2
j
j
508
SigmaXL: Appendix
n∑xij2 /C j − ∑Ri2
i, j i
τc =
n − ∑R
2
i
2
Thiel’s Uncertainty
Uncertainty [12] is defined as the proportional reduction in the uncertainty (entropy) of the
dependent variable as a consequence of using the information in the independent variable.

U r + U c − U rc
U c| r =
Uc
U r + U c − U rc
U r |c =
Ur
Symmetric version:
2(U r + U c − U rc )
Us =
Ur +Uc
where:
U r = −∑Ri /n ln ( Ri /n)
i
U c = −∑C j /n ln (C j /n)
j
U rc = −∑xij /n ln ( xij /n)

i, j
Tests for Ordinal Categories

Concordant - Discordant
Let two assessors rank n subjects, such that x is the rankings of assessor 1, and y is the rankings of
xi > x j and yi > y j , or xi < x j and yi < y j , then the pair of observations are said to be concordant
[2]. Pairs that are not concordant are said to be discordant. If x were a measurement of exercise,
and y a measurement of fitness, then we would expect high levels of concordance.
509
SigmaXL: Appendix
Define:
Cij = ∑∑xhk + ∑∑xhk
h <i k < j h >i k > j
Dij = ∑∑xhk + ∑∑xhk

h <i k > j h >i k < j
Then the number of concordant ( nc ) and discordant ( nd ) pairs are derived as:
nc = 0.5 P where P = ∑xij Cij
i, j
nd = 0.5Q where Q = ∑xij Dij

i, j
The probability value returned for this test is given in Kendall’s Tau-b below.
Spearman Rank Correlation Coefficient

The Spearman rank correlation coefficient ( ρ ) [3] provides a non-parametric measure of the
association between the two variables. It has a range of − 1 ≤ ρ ≤ 1 . Let RI i and CI j be the row and
column rank scores. Then ρ is evaluated as:
∑x ( RI
i, j
ij i − n/2)(CI j − n/2)
ρ=
(n 3 − ∑R 3 )(n 3 − ∑C 3 ) /12
Under the null hypothesis that there is no correlation,
ρ n−2
t=
1− ρ 2
has a t distribution with n − 2 degrees of freedom.
Measures of Association for Ordinal Categories

Confidence bounds are evaluated as t ± zα s , where z is the inverse of the standard normal cdf, t
is the statistic and s is the asymptotic standard error (ASE1) of the statistic. See [9,12] for ASE1
formulas.
Spearman Rank Correlation Coefficient

Since the Spearman rank correlation coefficient is not distributed normally, the Fisher transform is
used to evaluate the confidence bounds. See Bonnett & Wright [4] for details.
510
SigmaXL: Appendix
Kendall’s Tau-b
The Kendall τ b coefficient [10], like gamma, is a non-parametric measure of association using rank
correlation, but makes a correction for ties. τ b is defined as:
2(nc − nd )
τb =
Dr Dc
where:
Dr = n 2 − ∑Ri2
i
Dc = n 2 − ∑C 2j
j
τ b ranges from +1 to -1. A value of zero implies independence. Given s:
∑x (C
i, j
ij ij − Dij ) 2 − ( P − Q) 2 /n
s=2 under the null hypothesis,
Dr Dc
then τ b /s is distributed standard normal. This is used to provide a P-Value for the Concordant –
Discordant test.
Kendall-Stuart Tau-c
Stuart’s τ c coefficient [1] is a measure of association, but makes a correction for ties as well as table
size. τ c is defined as:
2m(nc − nd )
τc =
n 2 (m − 1)
where m = min (r , c) . τ c ranges from +1 to -1. A value of zero implies independence.
Goodman-Kruskal Gamma
The Goodman-Kruskal’s gamma statistic [8] provides a measure of association using rank
correlation, and is based on the number of concordant and discordant pairs. It is defined as:
nc − nd
γ=
nc + nd
Somer’s D
Somer’s D statistic [1,14] is also a measure of association, similar to τ b , but differs in respect to a
consideration as to whether the row variable is considered the independent variable (and the
column variable dependent), or vice versa.
511
SigmaXL: Appendix
S r = 2( nc − nd )/Dr
S c = 2( nc − nd )/Dc
Symmetric version:
S s = 4( nc − nd )/( Dr + Dc )
References for Chi-Square Tests and (Contingency) Table

Associations
[1] Agresti, A. (2010). Analysis of Ordinal Categorical Data (Second Ed.), New York: John Wiley
and Sons.
[2] Agresti, A. (2013). Categorical Data Analysis (Third Ed.), New York: John Wiley and Sons
[3] Bishop, Y, Fienberg, S and Holland, P. (1975). Discrete Multivariate Analysis: Theory and
Practice. Cambridge, Mass.: MIT Press.
[4] Bonett, D.G, and Wright, T.A. (2000). “Sample size requirements for estimating Pearson,
Kendall and Spearman correlations”, Psychometrica 65 (1),pp 23-28.
[5] Cohen, J. (1960). “A coefficient of agreement for nominal scale”. Educational and
Psychological Measurement 20 (1) pp. 37-46.
[6] Fisher, R.A. (1915). “Frequency distribution of the values of the correlation coefficient in
samples from an indefinitely large population”. Biometrika 10 (4), pp. 507-521.
[7] Fisher, R.A. (1922). “On the interpretation of χ 2 from contingency tables, and the
calculation of P”. Journal of the Royal Statistical Society 85 (1), pp. 87-94.
[8] Goodman, L. A. and W. H. Kruskal, W.H. (1972). “Measures for association for cross-
classification”, Journal of the American Statistical Association 67. pp. 415-421.
[9] IBM SPSS Statistics 24 Algorithms, Crosstabs Algorithms, pp. 161 – 182.
ftp://public.dhe.ibm.com/software/analytics/spss/documentation/statistics/24.0/en/client/
Manuals/IBM_SPSS_Statistics_Algorithms.pdf
[10] Kendall, M. (1938). “A New Measure of Rank Correlation”. Biometrika 30 (1–2) pp. 81-89.
[11] McNemar, Q.(1947).“Note on the sampling error of the difference between correlated
proportions or percentages”. Psychometrika 12 (2). pp. 153-157.
[12] SAS/STAT ® 14.1 Users Guide - Measures of Association,
http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#st
atug_freq_details22.htm
[13] Snedecor, G.W and Cochran, W.G. (1989), Statistical Methods 8th ed. Ames, Iowa.
Blackwell Publishing.
[14] Somers, R. H. (1962). “A new asymmetric measure of association for ordinal variables”.
American Sociological Review 27. pp. 799-811.
512
SigmaXL: Appendix
[15] Cohen, J. (1988), Statistical Power Analysis for the Behavioral Sciences (Second Ed.),
Lawrence Erlbaum Associates, pp. 215 - 227.
513
SigmaXL: Appendix
Multiple Comparison of Means and Variances

(a.k.a. Post-Hoc Tests)
This set of tests is undertaken after an ANOVA or Equal Variance test, and are often referred to as
“post hoc tests”. After the null of equal means or variances has been rejected, the following tests
are computed for each of the .5k (k − 1) sample pairs. Thus this permits the researcher to ascertain
exactly which group(s) are responsible for the rejection of the null.
One-Way ANOVA
Fisher
Fisher’s least significant difference test is a set of individual t-tests on each of the sample pairs, but
uses the pooled standard deviation from all groups, which increases the power. The test statistic is:
yi − y j
t=
s 1/ni + 1/n j
where s is the pooled standard error for all groups, and is assumed distributed Student t.
There is no adjustment for family-wise error rate, so should only be used for k = 3.
Tukey
Tukey’s test is a set of individual t-tests on each of the sample pairs, but uses the pooled standard
deviation from all group. The test statistic is:
2 ( yi − y j )
t=
s 1/ni + 1/n j
and is distributed as the Studentized range [1].
Dunnett
The Dunnett test compares each group to a given control group. For each group, a t statistic is
calculated using the Fisher statistic. The distribution of this statistic is a k − 1 multivariate t, with all
k − 1 arguments being the same Fisher t statistic, and with correlation matrix with off diagonal
elements 2 for the balanced case, and weighted by sample size for the unbalanced case [1].
514
SigmaXL: Appendix
Welch’s ANOVA (Unequal Variance)

Pairwise Welch
This is the pairwise two sample t-test for unequal variances. Like Fisher, there is no adjustment for
family-wise error rate so should only be used for k = 3.
Games-Howell
The Games-Howell test does not assume equal variance for each group. The variance of each group
is used. The test statistic is:
2 | yi − y j |
t=
si2 /ni + s 2j /n j
And is distributed as the Studentized range [2].
Bartlett’s Test for Equal Variance

F-Test
The standard F test statistic for testing the null of equal variances of two groups is:
si2
F=
s 2j
2
where si is the variance of group i , and is distributed as Fn −1,n .
i j −1
There is no adjustment for family-wise error rate so should only be used for k = 3.
F-Test with Bonferroni Correction
The Bonferroni correction multiplies the F-Test probability by a factor of .5k (k − 1) .
Levene’s Test for Equal Variance

Pairwise Levene (a.k.a. Brown-Forsythe)
The pairwise Levene test statistic for testing the null of equal variances of two groups is a simple F
statistic:
515
SigmaXL: Appendix
SSA/(k − 1)
F=
SSE/(n − k )
where n = ni + n j , and SSA and SSE are the sample sum of squares and the error sum of squares of
the transformed variable respectively:
z =| x − median( x) |
The statistic is distributed as F with k − 1 and n − k degrees of freedom [3,4].
There is no adjustment for family-wise error rate, so should only be used for k = 3.
Tukey ADM (Absolute Deviations from Median)
Tukey’s method is applied to the ADM (Absolute Deviations from Median). Conceptually, this is like
an extension of the ANOMV-LEV chart, Analysis of Means version of the Levene Test. See [5]
Nelson, Wludyka and Copeland (2005), page 65 for discussion on robustness.
References for Multiple Comparisons of Means and Variances

[1] J.C. Hsu, J.C. (1966). Multiple Comparisons, Theory and methods. Chapman & Hall.
[2] Games, P. A., and J. F. Howell. 1976. “Pairwise Multiple Comparison Procedures with
Unequal N’s and/or Variances: A Monte Carlo Study”. Journal of Educational Statistics, 1,
113–125.
[3] Levene, H. (1960). In Contributions to Probability and Statistics: Essays in Honor of Harold
Hotelling, I. Olkin et al. eds., Stanford University Press, pp. 278-292.
[4] Brown, M.B., and Forsythe, A.B. (1974). “Robust Tests for Equality of Variances.” Journal of the
American Statistical Association 69:364–367.
[5] Nelson, P.R., Wludyka, P.S. and Karen A. F. Copeland, K.A. (2005) The Analysis of Means: A
Graphical Method for Comparing Means, Rates, and Proportions ASA-SIAM Series on
Statistics and Applied Probability, SIAM, Philadelphia, ASA, Alexandria, VA.
516
SigmaXL: Appendix
Analysis of Means (ANOM) Charts
A comparison of treatment means to determine which are significantly different from the overall
average is commonly required. For example an educational professional might ask which schools in
the system have a higher or lower rate of graduation compared to the average for all schools.
ANOM (Analysis of Means) addresses this problem for a fixed effects model. ANOM identifies the
means that are significantly different from the overall mean. Analysis of means is often used in
process control, since it has a graphical representation that is similar to a Shewhart chart. Typically,
a central line represents the overall average. The treatment means, plotted as deviations from the
overall average are compared with upper and lower decision limits to identify which are
significantly different from the overall mean. The underlying assumptions are that the response
belongs to a given distribution, and that the design is one way or two way.
The lower and upper decision levels (LDL, UDL) are used to ascertain whether the means at the
factor level are different from the grand mean. A factor level mean that lies outside the UDL and
LDL is statistically different from the grand mean at the specified confidence level.
Let n = ∑ jn j is the total number of observations, xij be the i th observation at factor level j , and
the mean for factor level j be:
nj
∑x ij
xj = i =1
nj
The grand mean is:
∑x
i, j
ij
x=
n
ANOM Normal One-Way

For a single factor with k factor levels, and with the response assumed normal, the LDL and UDL
are given by:
LDL j = x − hα s p w j
UDL j = x + hα s p w j
where the pooled variance is:
517
SigmaXL: Appendix
k
s 2p = ∑(n j − 1) s 2j /(n − k )
j =1
2
where s j is the variance of factor j .
where the weighting factor for each group j is:
w j = (n − n j )/(nn j )
which reduces to w j = (k − 1)/n in the balanced case.
where hα is the absolute value of the inverse CDF of the students t distribution at probability α *
with df = n − k degrees of freedom, or is determined from lookup tables [4] (Table B.1 for
balanced, Table B.3 for unbalanced) at the given probability level ( α ),
and where α * = (1 − (1 − α )1/k )/2 if k > 2 and α * = α/2 if k = 2 .
ANOM Normal Two-Way

For a two factor model, with k1 and k 2 factor levels respectively, and with the response assumed
normal, the main effect is given by:
LDL fj = x − hα MSE w fj
UDL fj = x + hα MSE w fj
where LDL fj and UDL fj are the lower/upper bounds for factor f at level j respectively,
where MSE is derived from an ANOVA analysis with a two factor model with cross effects,
w fj = (k f − 1)/(k f n j )
which reduces to w = (k − 1)/n in the balanced case.
where hα is the absolute value of the inverse CDF of the students t distribution at probability α *
with df = k1k 2 (n − 1) degrees of freedom, or is determined from Nelson’s lookup tables at the
given probability level ( α ),
1/k f
and where α * = (1 − (1 − α ) )/2 if k > 2 and α * = α/2 if k = 2 .
The significance of the interaction effect is evaluated using a standard F test:

( S r − Su )/( f r − f u )
f =
S u /f u
518
SigmaXL: Appendix
where Su is the unrestricted sum of squares,

where S r is the restricted sum of squares,
and where f u and f r are the degrees of freedom for the unrestricted and restricted cases
respectively.
f is distributed F with f r − f u and f u degrees of freedom.
BY method
A straight forward method of analyzing a two factor model with k1 and k 2 groups for factor 1 and
2 respectively is to simply evaluate the one factor model with factor 1 for each value of factor 2.
Thus a total of k 2 one way models runs are undertaken.
SLICE method
This method [Wludyka 2013, 2015] is the same as the BY method, but uses the pooled MSE and
degrees of freedom from the full two factor model. The SLICE factor with minimum number of
levels is automatically selected.
ANOM Binomial Proportions One-Way
A binomial distribution can be used for data that consists of proportions. Thus p j = ∑i j xij /n j .
n
For a single factor with k factor levels, and N = ∑ jn j observations, and with the response
assumed binomial, the LDL and UDL are given by:
LDL j = p − hα s p w j
UDL j = p + hα s p w j
where the average proportion is given by:

∑nj
j pj
p=
N

s 2p = p (1 − p )/n j
j

w j = ( N − n j )/( N )
519
SigmaXL: Appendix
and where is the absolute value of the inverse CDF of the normal distribution at probability
hα
α = 1 − α/(2k ) if k > 2 , and α = 1 − α/2 if k = 2 , or is determined from Nelson’s lookup tables at
* *
the given probability level ( α ).
ANOM Binomial Proportions Two-Way

For a balanced two factor model, with k1 and k 2 factor levels respectively, and with the response
assumed binomial, the main effect is given by the one way factorial model, evaluated separately on
each factor. Let k1 be the number of levels of factor 1, with n observation per group.1
LDL = p − hα s p w
UDL = p + hα s p w
where LDL and UDL are the lower/upper bounds respectively,

where the average proportion is given by:
∑j p j
p=
k1

s 2p = p (1 − p/n
where the weighting factor is:

w = ( k1 − 1)/(k1 )
and where is the absolute value of the inverse CDF of the normal distribution at probability
hα
α = 1 − α/(2k ) if k > 2 , and α = 1 − α/2 if k1 = 2 , or is determined from Nelson’s lookup tables at
* *
The significance of the interaction effect is evaluated using a likelihood ratio test:
q = −2(llfu − llf r )
where llfu is the unrestricted log likelihood,

where llf r is the restricted log likelihood
and where the likelihood is specified as:
llf = ∑n ( p ln ( pˆ ) + (1 − p) ln (1 − pˆ ))
1 A similar analysis follows for factor 2.
520
SigmaXL: Appendix
where n is the vector of sample size, p is the vector of probabilities, and p̂ is the predicted
probability given by exp( xβ )/(1 + exp( xβ )) . x is the factor matrix, and β is the maximum likelihood
coefficient vector.
q is distributed χ 2 with degrees of freedom equal to the number of restrictions.
ANOM Poisson Rates One-Way

A Poisson distribution can be used for data that consists of counts. Thus the average count for
group j is c j = ∑in j cij /n j .
For a single factor with k factor levels, and N = ∑ jn j observations, and with the response
assumed distributed Poisson, the LDL and UDL are given by:
LDL j = c − hα s p w j
UDL j = c + hα s p w j
where the average count is given by:

∑n c
j
j j
c=
N

s 2p = c

w j = ( N − n j )/( Nn j )
and where hα is the absolute value of the inverse CDF of the normal distribution at probability
α = 1 − α/(2k ) if k > 2 , and α = 1 − α/2 if k = 2 , or is determined from Nelson’s lookup tables at
* *
ANOM Poisson Rates Two-Way

For a balanced two factor model, with k1 and k 2 factor levels respectively, and with the response
assumed Poisson, the main effect is given by the one way factorial model, evaluated separately on
each factor.
521
SigmaXL: Appendix
The significance of the interaction effect is evaluated using a likelihood ratio test:
q = −2(llfu − llf r )
where llfu is the unrestricted log likelihood,

where llf r is the restricted log likelihood
and where the likelihood is specified as:
llf = ∑ y ln (λ ) − λ − ln ( y!)
where y is the vector of counts, λ = exp( xβ + ln (n)) , n is the vector of sample size, x is the factor
matrix, and β is the maximum likelihood coefficient vector.
q is distributed χ 2 with degrees of freedom equal to the number of restrictions.
ANOM Nonparametric Transformed Ranks

When the distribution of a data set is not known, or cannot be assumed to be normal, Poisson or
binomial, then a non-parametric methodology is used to compare means. Only a single factor (one
way) is permitted.
The rank ( r ) of each element in the data set is derived. Then the rank is transformed using inverse
normal scores:
0.5 + r
x = Φ −1 ( )
2N −1
where N is the number of observations, and Φ −1 is the inverse of the standard normal
distribution function. x is then distributed normally, and the usual ANOM procedure can be
applied.
ANOM Variances
ANOMV can be used to test whether the variance for any sample is significantly different from the
pooled varaiance - it is thus a test of homogeneity. Typically, consistent output is required from a
set of production lines, and ANOMV can be used to identify which line has variability significantly
different from the average.
Let N = ∑ jn j be the total number of observations, xij be the i th observation at factor level j , and
2
the sample variance for factor level j be s j .
One Way balanced

For a single factor with k factor levels, n observations for each group, and with the response
assumed normal, the LDL, CL, and UDL are given by:
LDL = Lα ,k ,n −1k MSe
522
SigmaXL: Appendix
CL = MS e
UDL = U α ,k ,n −1k MSe
where the pooled MSE is:

k
MS e = ∑s 2j
j =1
and where Lα ,k ,n −1 and U α ,k ,n −1 are determined from Nelson’s lookup tables at the given
probability level ( α ) and k and n − 1 degrees of freedom.
One Way unbalanced

For a single factor with k factor levels, n j observations for group j , and with the response
assumed normal, the LDL, CL, and UDL are given by:
N −k
LDL j = L*α ,k ,n MS e
j n −1
j
CL j = MSe
N −k
UDL j = U α* ,k ,n MS e
j n j −1

k
∑( n
j =1
j − 1) s 2j
MS e =
N −k
where L*α ,k ,n is given by:

j
L*α ,k ,n = β −1 (α * , θ j , φ j )
j
where L*α ,k ,n is given by:

j
U α* ,k ,n = β −1 (1 − α * , θ j , φ j )
j
and where α * = (1 − α )/(2k ) , θ j = (n j − 1)/2 , φ j = ∑i ≠ jθ i , and β −1 is the inverse of the cdf of the
beta function.
Large Sample - Balanced

For data sets with large samples - all groups have at least 35 observations, for a single factor with k
factor levels, n observations for each group, and with the response assumed normal, the LDL, CL,
and UDL are given by:
LDL = MSe − hα ,k ,∞ MSe w
523
SigmaXL: Appendix
CL = MS e
UDL = MSe + hα ,k ,∞ MSe w

k
MS e = ∑s 2j
j =1
where w is given by:

k −1 2 γ
( + )
k n −1 n
where hα ,k ,∞ is determined from Nelson’s lookup tables at the given probability level α and k
and ∞ degrees of freedom.
and where γ is an estimate of the common kurtosis - for samples distributed normal, this will be
zero..
A similar formula is used for the large sample unbalanced case - see [4] for details.
ANOM Levene Robust Variances

When the distribution of a data set is not known, or cannot be assumed to be normal, then a non-
parametric methodology is used to compare variances. The methodology is to first transform the
data to the absolute deviation from the median for each group ADM ij :
ADM ij =| xij − md j |
where md j is the median for group j . Then, the ANOM Normal process is used on the ADM.
References for Analysis of Means (ANOM) Charts

[1] Bakir,S.T. (2010) “A Nonparametric Test For Homogeneity Of Variances: Application To GPAs
Of Students Across Academic Majors” American Journal of Business Education, 3(3) pp. 47-
54.
[2] Garrett, L and Nash, J.C. (2001). “Issues in Teaching the Comparison of Variability to Non-
Statistics Students”, Journal of Statistics Education ((2), pp. 1-16.
[3] Nelson, L.S. (1983). "Exact Critical Values for Use with the Analysis of Means", Journal of
Quality Technology, 15, pp. 40-44.
524
SigmaXL: Appendix
[4] Nelson, P.R., Wludyka, P.S. and Karen A. F. Copeland, K.A. (2005) The Analysis of Means: A
Graphical Method for Comparing Means, Rates, and Proportions ASA-SIAM Series on
Statistics and Applied Probability, SIAM, Philadelphia, ASA, Alexandria, VA.
[5] Wludyka, P.S., (2013). “Analyzing Multiway Models with ANOM Slicing” SESUG Working
Paper SD-06 21st Southeast SAS users Group Conference.
[6] Wludyka, P.S., (2015). “Using ANOM Slicing for Multi-Way Models with Significant
Interaction”, Journal of Quality Technology 47(2), pp. 193-203.
[7] Wludyka, P.S. and Noguera, J.G. (2016, November). “Using Anom Slicing For Multiway
Models With Binomial Or Poisson Data,” Poster session presented at INFORMS Annual
Meeting, Nashville, TN.
[8] Wludyka, P.S., Bergen, A. & Liu, X. (2002) “A Monte Carlo Study of Two Normal Based
Homogeneity of Variance Tests for Unbalanced Designs”, Joint Statistical Meetings -
Statistical Computing Section, pp. 3758-3759.
[9] Wludyka, P.S. & Nelson, P.R. (1997). “An Analysis-of-Means-Type Test for Variances from
Normal Populations”, Technometrics, 39(3), pp. 274-285
[10] Wludyka, P.S. & Nelson, P.R. (1999) “Two non-parametric, analysis-of-means-type tests for
homogeneity of variances”, Journal of Applied Statistics, 26:2, pp. 243-256,
525
About SigmaXL, Inc.
Established in 1998, SigmaXL Inc. is a leading provider of user friendly Excel Add-ins for Lean Six
Sigma graphical and statistical tools and Monte Carlo simulation.
Our flagship product, SigmaXL®, was designed from the ground up to be a cost-effective, powerful,
but easy to use tool that enables users to measure, analyze, improve and control their service,
transactional, and manufacturing processes. As an add-in to the already familiar Microsoft Excel,
SigmaXL® is ideal for Lean Six Sigma training and application, or use in a college statistics course.
DiscoverSim™ enables you to quantify your risk through Monte Carlo simulation and minimize your
risk with global optimization. Business decisions are often based on assumptions with a single point
value estimate or an average, resulting in unexpected outcomes. DiscoverSim™ allows you to
model the uncertainty in your inputs so that you know what to expect in your outputs.
Contact Information:
Technical Support: 1-866-475-2124 (Toll Free in North America) or
1-519-579-5877
Sales: 1-888-SigmaXL (888-744-6295)
E-mail: [email protected] Web: www.SigmaXL.com
527

SigmaXL Version 8 Workbook

Uploaded by

Copyright:

Available Formats

SigmaXL Version 8 Workbook

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SigmaXL Version 8 Workbook

Uploaded by

Copyright:

Available Formats

SigmaXL Version 8.

Copyright © 2004-2017, SigmaXL Inc.

Data Preparation – Remove Blank Rows and Columns ......................................................... 30

Boxplots ............................................................................................................................... 119

Part B - EZ-Pivot ......................................................................................................................... 191

Part J – One-Way ANOVA & Means Matrix ............................................................................... 238

Binary Logistic Regression.................................................................................................... 373

Part F – Advanced Charts: I-MR-R/S .......................................................................................... 455

References for Exact and Monte Carlo P-Values ................................................................. 492

ANOM Poisson Rates Two-Way ........................................................................................... 521

Copyright © 2004-2017, SigmaXL Inc.

What’s New in Version 8.0

• Templates and Calculators

o Additional Normality Tests

• Analysis of Means (ANOM) Charts

o Binomial Proportions One-Way

• Welch ANOVA (Assume Unequal Variance)

• Bartlett Equal Variance

• Levene Equal Variance

• Chi-Square Tests & Table Associations

3. Click Next. The activation process continues and is confirmed as shown:

5. In Excel 2011 for Mac, SigmaXL is added to Excel’s menu:

2. Serial Number Used in Previous Activation:

3. Internet Connection Problem:

SigmaXL® Defaults and Menu Options

Clear Saved Defaults

Data Selection Default

Menu Options (Classical or DMAIC)

SigmaXL® System Requirements

Minimum System Requirements:

Memory: 2 GB of RAM or greater.

Hard disk: 200 MB of available hard-disk space.

Drive: CD-ROM or DVD drive.

Display: 1024x768 or higher resolution monitor.

Mac Operating System: OS X Snow Leopard (10.6) or later

Administrative Rights to install software.

Getting Help and Product Registration

Technical support is available by phone at 1-866-475-2124 (toll-free in North America) or 1-519-

Copyright © 2004-2017, SigmaXL Inc.

The Y=f(X) Model

Data Types: Continuous Versus Discrete

It is possible to have various combinations of discrete/continuous X’s and discrete/continuous Y’s.

Examples of Discrete (Category) X and Discrete Y

• X = Customer Type, Y = Number of Complaints

• X = Product Type, Y = Number of Defects

• X = Day Shift vs. Night Shift, Y = Proportion of Defective Units

Examples of Discrete (Category) X and Continuous Y

• X = Customer Type, Y = Customer Satisfaction (1-5)

• X = Before Improvement vs. After Improvement, Y = Customer Satisfaction (1-5)

• X = Location, Y = Order to Delivery Time

Examples of Continuous X and Discrete Y

• X = Responsiveness to Calls (1-5), Y = Number of Complaints

• X = Process Temperature, Y = Number of Defects

Examples of Continuous X and Continuous Y

• X = Responsiveness to Calls (1-5), Y = Customer Satisfaction (1-5)

• X = Amount of Loan ($), Y = Cycle Time (Loan Application to Approval)

Stacked Data Column Format versus Unstacked Multiple Column

within a subgroup, between subgroups, or

To identify special causes.

Copyright © 2004-2017, SigmaXL Inc.