SESUG Paper 172-2019

Probability Plots for Exploratory Data Analysis

Dennis J. Beal, Leidos, Oak Ridge, Tennessee

ABSTRACT

Probability plots are used in statistical analysis to check distributional assumptions, visually check for
potential outliers, and see the range, median, and variability of a data set. Probability plots are an important
statistical tool to use for exploratory data analysis. This paper shows SAS® code that generates normal and
lognormal probability plots using the Output Delivery System (ODS) on a real environmental data set using
PROC UNIVARIATE and interprets the results. This paper is for beginning or intermediate SAS users of
Base SAS® and SAS/GRAPH®.

Key words: normal probability plots, lognormal probability plots, PROC UNIVARIATE

INTRODUCTION
When the data analyst receives a data set, the first step to understanding the data is to conduct an
exploratory data analysis. This first step might include a PROC CONTENTS to list the variables that are in
the data set and their data types (numeric, character, date, time, etc.). The next step might be to run PROC
FREQ to see the values of all the variables in the data set. Another common practice is to run PROC
UNIVARIATE to see summary statistics for all numeric variables. These steps help guide the data analyst
to see whether the data needs cleaning or standardizing before conducting a statistical analysis. In addition
to summary statistics, PROC UNIVARIATE also generates histograms and normal probability plots for
numeric variables. In SAS v. 9.3 and 9.4 PROC UNIVARIATE generates generic histograms and normal
probability plots for the numeric variables with the ODS.

In addition, PROC GPLOT also produces normal and lognormal probability plots. Producing graphs in PROC
GPLOT gives the data analyst the most control of the details of the graph, such as customizing symbols,
fonts, font sizes, line thickness and type, text boxes, titles, data labels, tick marks, X and Y axis scales, and
colors. Multiple plots can be combined into a single plot when there are several variables to overlay. The
annotation facility can also be used to further customize the graph. These options are best used in SAS v.
9.3 without ODS.

Examples using real environmental data are shown by plotting concentrations for the metal lead in two soil
depth intervals: surface soil and subsurface soil. In addition to quantitative distributional tests such as the
Shapiro-Wilk test to test the null hypothesis of normality or lognormality, the probability plots confirm the
results of these tests visually.

This paper will show the SAS code that generates high quality normal and lognormal probability plots using
PROC UNIVARIATE. The SAS code presented uses the SAS System for personal computers version 9.3
and 9.4 running on a Windows 7 Professional platform.

CALCULATING PROBABILITY PLOTS

In order to understand probability plots, some calculations must first be performed if using PROC GPLOT.
PROC UNIVARIATE performs these calculations automatically. For normal probability plots, the variable to
be examined is plotted on a linear scale on the vertical Y axis. The X axis shows plotting positions from the
inverse standard normal cumulative distribution function (cdf). For a data set with n observations, the data
are sorted from smallest to largest. We denote the data as xi for i = 1 to n, where x1 is the smallest
observation and xn is the largest observation. Though there are several plotting position functions in the
statistical literature, the most common plotting position used for probability plots is the Blom plotting position
(Helsel 2005, p. 48).
Probability Plots for Exploratory Data Analysis SESUG 2019

For each observation xi (i = 1, 2, …, n), the plotting position pi is calculated as shown in Equation 1.

𝑖−0.375
𝑝𝑖 = (1)
𝑛+0.25

Note that all pi must be exclusively between 0 and 1 (0 < pi < 1). Each pi represents the cumulative area
(percentile) under the standard normal cdf for each xi. The xi are then plotted on the Y axis, while the pi are
plotted on the X axis for each ordered pair (pi, xi) for i = 1, 2, …, n.

A linear regression line is then fit to the data. If the data plot approximately linearly along its regression line,
then the data will be approximately normally distributed on a normal probability plot (where Y has a linear
scale). Any curvature, breaks in the distribution or inflection points will indicate deviations from normality.
Potential outliers can also be identified on both the right and left tails of the distribution. The slope of the
regression line is a measure of the standard deviation of the data. A steep regression line indicates large
variability, while a flatter regression line indicates low variability. The median is the data point that is
associated with pi = 50th percentile. The range is the difference between the maximum data point (xn) and
the minimum data point (x1).

PROBABILITY PLOTS USING PROC UNIVARIATE

Using PROC UNIVARIATE in SAS v. 9.4, the SAS code to generate normal probability plots is simple and
generates normal probability plots, as shown in Figures 1 and 2. The probplot option in PROC
UNIVARIATE creates the normal probability plot without any additional calculations programmed by the
user. The example data set metal consists of real surface and subsurface soil concentrations for a site
analyzed for total metals.

data metal;
input MEDIA $1-15 Lead;
UNITS = ‘mg/kg’;
label lead = 'Lead (mg/kg) in Surface Soil';
datalines;
Subsurface Soil 38.5
Subsurface Soil 47.65
Subsurface Soil 48.55
Subsurface Soil 55.9
Subsurface Soil 58.1
Subsurface Soil 58.7
Subsurface Soil 59.9
Subsurface Soil 61.9
Subsurface Soil 63.8
Subsurface Soil 71.3
Subsurface Soil 74.8
Subsurface Soil 79.9
Subsurface Soil 80.2
Subsurface Soil 93.1
Subsurface Soil 99.1
Subsurface Soil 103
Subsurface Soil 106
Subsurface Soil 106
Subsurface Soil 122
Subsurface Soil 125
Subsurface Soil 140
Subsurface Soil 150
Subsurface Soil 153
Subsurface Soil 163
Subsurface Soil 174
Subsurface Soil 185
Subsurface Soil 220

2
Probability Plots for Exploratory Data Analysis SESUG 2019

Subsurface Soil 251

Subsurface Soil 462
Subsurface Soil 505
Surface Soil 36.4
Surface Soil 37.3
Surface Soil 45.7
Surface Soil 47
Surface Soil 48.4
Surface Soil 50.8
Surface Soil 52.2
Surface Soil 55.9
Surface Soil 65.9
Surface Soil 66.1
;
run;

ods graphics on;

proc univariate data=metal; ** for Fig. 1;
where media = 'Surface Soil';
var Lead;
probplot Lead / normal(mu=est sigma=est)
odstitle='Normal Probability Plot for Lead in Surface Soil';
output out=statsout min=min mean=mean median=median skewness=skewness
max=max std=std cv=cv probn=probn; run; quit;
ods graphics off;

proc print data=statsout; run;

data metal2;
set metal;
label lead = 'Lead (mg/kg) in Subsurface Soil';
run;

ods graphics on;

proc univariate data=metal2; ** for Fig. 2;
where media = 'Subsurface Soil';
var Lead;
probplot Lead / normal(mu=est sigma=est)
odstitle='Normal Probability Plot for Lead in Subsurface Soil';
run; quit;
ods graphics off;

3
Probability Plots for Exploratory Data Analysis SESUG 2019

Figure 1. Normal Probability Plot for Lead in Surface Soil

Figure 1 shows 10 lead concentrations in surface soil. The data plot approximately linearly along its linear
regression line, which indicates the data are approximately normally distributed. The regression line is
generated using the normal(mu=est sigma=est) option, which estimates the two parameters mean
(mu) and standard deviation (sigma) from the data for an assumed normal distribution. Summary statistics
such as the minimum, mean, median, maximum, standard deviation, skewness and coefficient of variation
were also produced by PROC UNIVARIATE and output into the SAS data set statsout. Figure 1 shows
the median concentration is 49.6 mg/kg at the 50th percentile. The Shapiro-Wilk (SW) test p-value calculated
using the probn option within PROC UNIVARIATE confirms the surface soil is normally distributed with a
p-value of 0.4689.

Figure 2 shows 30 lead concentrations in subsurface soil. The distribution has significant curvature
compared to its linear regression line and a break in the distribution between the second and third highest
concentrations. The two highest concentrations appear as possible outliers given that they stand out from
the rest of the distribution. The plot indicates a statistical outlier test should be conducted for the two highest
concentrations to determine if these concentrations are statistical outliers at a specified significance level.
Figure 2 shows the median concentration is 101.5 mg/kg at the 50 th percentile. The SW p-value for
subsurface soil is <0.0001, which indicates the data are not normally distributed.

Because the subsurface soil data in Figure 2 are not normally distributed due to significant curvature, a
break in the distribution and two possible outliers in the right tail, the data should be tested to see if it is
lognormally distributed. One way to do that is to calculate the natural logarithm (ln) of each data point and
create a normal probability plot on the ln transformed concentrations. If the ln transformed concentrations
plot approximately linearly along its linear regression line, then the data are approximately lognormally
distributed.

4
Probability Plots for Exploratory Data Analysis SESUG 2019

Figure 2. Normal Probability Plot for Lead in Subsurface Soil

The SAS code to transform the data and produce a normal probability plot on the ln transformed data is
shown next. The SAS function log calculates the natural logarithm on positive concentrations. A normal
probability plot on the ln transformed concentrations is shown in Figure 3.

data metal3;
set metal;
where media = 'Subsurface Soil';
if lead > 0 then Ln_Lead = log(lead);
label Ln_Lead='Ln(Lead) in Subsurface Soil [ln(mg/kg)]';
run;

ods graphics on ;
proc univariate data=metal3; ** for Fig. 3;
var Ln_Lead;
probplot Ln_Lead / normal(mu=est sigma=est)
odstitle='Normal Probability Plot for Ln(Lead) in Subsurface Soil';
output out=statsout min=min mean=mean median=median skewness=skew
max=max std=std cv=cv probn=probn;
run; quit;
ods graphics off;
proc print data=statsout; run;

5
Probability Plots for Exploratory Data Analysis SESUG 2019

Figure 3. Normal Probability Plot for Ln (Lead) in Subsurface Soil

Figure 3 shows the ln transformed concentrations plot approximately linearly along its linear regression line,
which indicates the data are lognormally distributed. The SW test p-value is 0.1555, which confirms the
data are lognormally distributed at the 0.05 significance level. The Y axis in Figure 3 is in ln transformed
units and is on a linear scale. The two highest concentrations still appear to be possible outliers. The median
transformed concentration is 4.62 ln(mg/kg), so the median untransformed concentration is exp(4.62) =
101.5 mg/kg. When the X and Y axes are on a linear scale, the probability plot is a normal probability plot.
When the X or Y axis is on a log10 scale using untransformed concentrations, the probability plot is a
lognormal probability plot.

Another way to test if the data are lognormally distributed without first transforming the concentrations is to
use the lognormal option within PROC UNIVARIATE. However, a lognormal distribution has more
parameters than a normal distribution. The three parameters that must be estimated in order to generate a
linear regression line through the data are scale, sigma, and theta. Sigma is the shape parameter, while
theta is the threshold parameter. The following SAS code produces the lognormal probability plot shown in
Figure 4. The odstitle option can be used to change the title of the graph.

ods graphics on ;
proc univariate data=metal2;
where media = 'Subsurface Soil'; ** for Fig. 4;
var Lead;
probplot Lead / lognormal(scale=est sigma=est theta=0)
odstitle='Lognormal Probability Plot for Lead in Subsurface Soil';
run; quit;
ods graphics off;

6
Probability Plots for Exploratory Data Analysis SESUG 2019

Figure 4. Lognormal Probability Plot for Lead in Subsurface Soil

Figure 4 shows the data plot very nicely along the regression line except the second highest concentration.
The parameter estimates from the data are for the threshold parameter theta (33.003), the scale parameter
(4.1307), and shape parameter sigma (0.999066). Figure 4 is a lognormal probability plot because the scale
on the X axis assumes percentiles from a lognormal distribution.

Probability plots for many other continuous distributions (such as beta, exponential, Gumbel, Pareto, power,
Rayleigh, and Weibull) can also be generated for testing the distribution of data simply by specifying these
distributions and their appropriate parameters as options in the probplot statement. The advantage of
generating normal and lognormal probability plots using PROC UNIVARIATE is the simplicity of the code.

CONCLUSION
Probability plots are a useful exploratory data analysis tool that data analysts can use for exploring data.
Probability plots reveal whether the data follow a parametric distribution such as normal or lognormal,
identify possible outliers, show variability in the data and central tendency estimates. SAS code was
presented to easily produce normal or lognormal probability plots using the probplot option within PROC
UNIVARIATE using ODS.

REFERENCES

Helsel, Dennis R. 2005. Nondetects and Data Analysis: Statistics for Censored Environmental Data.
Hoboken, NJ: John Wiley & Sons.

7
Probability Plots for Exploratory Data Analysis SESUG 2019

CONTACT INFORMATION
Your comments and questions are valued and encouraged. Contact the author at:

Dennis J. Beal, Ph.D.

Senior Statistician/Risk Scientist
Leidos
301 Laboratory Road
P.O. Box 2502
Oak Ridge, Tennessee 37831
e-mail: [email protected]

SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of
SAS Institute Inc. in the USA and other countries. ® indicates USA registration.
Other brand and product names are registered trademarks or trademarks of their respective companies.