2015 SPSS Exercise

Download as pdf or txt
Download as pdf or txt
You are on page 1of 69

Technische Universitt Mnchen

TUM School of Management


Marketing and Consumer Research

SPSS exercise
As part of Quantitative Research Methods
Wintersemester 2013/2014

Outline:
1 Everything you ever wanted to know about statistics
2 Parameter testing

3 Distribution (non-parametric) testing, t-Test


4 4 Correlation analysis
5.2 Regression analysis
6 Theoretical basis of nonparametric tests

1 Everything you ever wanted to know about statistics


1.1 Preliminary notes about SPSS
SPSS is a statistical programme for the analysis and visualisation of data
It is possible to import data with another format into SPSS, e.g. excel data
can be imported
SPSS is frequently used in companies (e.g. market research)
SPSS enables the user to handle a great number of data (more than with
Excel) and to combine different data sets.

1.2 Change language:


Edit>Options>General (Bearbeiten Optionen Allgemein)
Select Language English under Output as well as User Interface (Sprache
English unter Ausbabe und Benutzeroberflche)

1.3 Information about the data set


Given in data descriptions (own description or external description)
o Sometimes, the data description is already incorporated within SPSS
(via labels and value labels etc.). This information can be printed
o File Display Data File Information Working File OR External File
You should always be very accurate with your data and never work without a
data description
o When you enter your own data, always enter your information in
variable view (e.g. question wordings, answer categories)

1.3 Starting SPSS

When opening SPSS you can select if you want to type in new data or use an
existing data set. If you select cancel you will get to the main window that does
not contain any data.
SPSS uses 3 windows: data-, syntax and SPSS-viewer:
SPSS data-editor (appears when the programme is opened): shows the
data in the data view and the characteristics of the variables under variable
view (file ending .sav)
SPSS Viewer (opens automatically after an analysis): shows the results. The
SPSS viewer shows the results as a hierarchy. One analysis can result in
many tables. Clicking on the + or symbol aggregates or disaggregates the
results. Clicking on one object shows the respective table. (file ending: .spv)
SPSS Syntax-Editor (can be opened with Aufruf ber Men FILE NEW
SYNTAX): instructions are entered in this window. The instructions are
executed by clicking the
(file ending: .sps)

1.3.1 Data Editor

1.3.2 Variable View

In the variable view you can enter important information for every variable
Every variable should have a unique name
o Each variable has to be named in column Name.
The default is VAR00001
The type of data is important as it determines what kind of analyses you can
do with the variable
Different variable types exist
o Numeric
Numbers (e.g. 7, 0, 120)
o String
Letters (e.g. Andy, Idiot)
o Currency
Currency (e.g. 20, $34, 56)
o Date
Dates (e.g. 21-06-1973, 06-21-73, 21-Jun-1973)
If you want to create a string variable:

Alternatively, you can define the variable as date:

For coding/ categorical variables you need to define categories:

Column Type: SPSS can differentiate between different measurement


scales (click on to show selection)
o Most often: numeric = numerical values or string = letters, words
Column labels: allows giving more information about the variable, e.g. the
survey question
Column values: click on to enter more data: value labels:
o Value: 1
o Label = Male Add
o Value : 2
o Label = female Add

1.3.3 The Viewer Window

When you work with several data sets or you work on several days on your
data set but you want to save your output within one output file only, saved it
the first time it opens. When you open SPSS again on another occasion, you
have to open this old output manually and close the new output viewer that
automatically opens. SPSS will then produce the output in the current SPSS
output viewer.
1.3.4 The Syntax Window

1.4 Opening SPSS data

FILE/ OPEN: data/syntax/ output/ script


Important: save all three viewers separately and in between different analyses
and data adjustments

1.5 Missing values


Missing values can be problematic in empirical analyses. SPSS differentiates two
kinds of missing values: system-missing values and user-missing values.
Recoding of missing values (BE CAREFUL)
There are (rare) cases if it is advisable to recode missing values in the data set
It is possible to recode missing values, e.g. to the value 0 or -99999.
Defining missing values under variable view:
Column Missing values: different selection options
Transform/Recode into same variables/ Select variable/ old and new values.../old
value: system-missing/ new value: value: type 0 / Add/ continue/ ok

1.6 Data handling

Data editing only under data view!!


EDIT/ GO TO CASE: easy navigation
DATA/ SORT CASES: allows to order the data according to certain variable of
intrerest in ascending or descending order
Alternative to menu navigation: using the syntax:
o SORT CASES BY XY.
o Important: every command ends with a point!!!
o Instead of entering the syntax by hand we can select the PASTE
button instead of clicking ok paste will enter the correct syntax for
each command we can select in the SPSS menu in the syntax editor:
useful if we need to repeat a certain command several times or we
need to change the variables
MERGING FILES
o DATA / MERGE FILES/ ADD CASES: selection of cases out of an open
or ext
o Also possible to ADD VARIABLES, important to consider key variables
by which the programme should order the variables (make sure that
the cases are ordered the same in both data sets and both data sets
contain the same number of cases)

SELECT CASES
o It can be necessary to restrict the analysis to certain subgroups in the
data set, e.g. only females
o DATA / SELECT CASES: if condition is satisfied: select if gender = 2
continue; output: filter out unselected cases, copy selected cases to
a new data set, delete unselcted cases (ATTENTION, make sure to
have an original hard copy of the data set just in case) ok will
select cases. DATA / SELECT CASES/ RESET to reset the selection
o Note the new filter variable
o Alternative syntax:

1.7 Transformation of variables


In most cases it will be necessary to transform the variables of interest further
to be able to use them in an analysis
EXERCISE: Transformation of variables:
DataSet: ChickFlick.sav
Divide the variable arousal into 3 groups:
o low: physical arousal up to 13
o medium: physical arousal between 14 and 27
o high: physical arousal higher than 27
compute arou=0.
if (arousal le 13) arou=1.
if (arousal ge 14 and arousal lt 28) arou=2.
if (arousal gt 27) arou=3.
execute.
Alternative: Transform Recode into different variables old and new values
DATASET ACTIVATE DataSet1.
RECODE arousal (Lowest thru 13=1) (14 thru 27=2) (28 thru Highest=3) INTO arou3.
EXECUTE.
Compute and execute belong together
TEST: Transformation of variables
Data Set: Record2.sav
Divide the variable airplay into the following categories: hardly ever,
medium, excessive.
o hardly ever: up to 10
o medium: 11-30
o excessive: more than 30 times per week
Assign value labels to the new variable
10

compute air=0.
if (airplay le 10) air=1.
if (airplay ge 11 and airplay le 30) air=2.
if (airplay gt 30) air=3.
execute.
Import of an excel file into SPSS
Copy data Excel Sheet.xlsx from the student server to your personal server
Remember the range of data in Excel (A1:E31)
Check Read variable names from the first row of data
Enter Range
Ok
Save file as .sav file
Before conducting complex analyses, it is reasonable to get a general idea
about the structure and distribution of the data
Good introduction into the data with frequency tables.
Analyze Descriptive Statistics Frequencies Select variable(s); select
display frequency tables
Frequency tables, however, make only sense for variables with a manageable
amount of scores

11

In addition, there are descriptive measures which help getting an idea about
the data
Slides PART II

2 Parameter testing
2.1 Tables & Graphs, descriptive statistics
In order to get a first quick overview of your data and the distribution of a single
variable, it is advisable to produce a tables of descriptive statistics

2.2 Descriptive measures


2.2.1 Measures of location and percentiles

Arithmetic mean (average, average mean): Sum of observations divided by


total number of observations (most used)
Median: Value that divides sample into two halves; Separates the higher half
of a sample from the lower half, lies in the middle of a series of ordered
observations = 50% quantile
Mode: Value that occurs the most frequently, value with the biggest (relative)
frequency; There can be more than one or no mode (if every number in the
set has the same frequency)
Quantiles:
o Value which cuts x% of the distributional area. Below the quantile is a
certain share of all values in a set of ordered observations
o If the distribution is divided in n equal shares, there are n -1 quantiles
o Quartile = sample is divided into 4 equal shares (= 3 quartiles: 25%-,
50%- 75% quartile)
o Quintiles = sample is divided into 5 equal shares (= 4 quintiles)
o Deciles = sample is divided into 10 equal shares (=9 deciles)
o Percentiles = sample is divided into 100 equal shares

2.2.2 Measures of distribution

Variance: main distribution measures; mean square deviation of single values


from the mean
Standard deviation (, S): positive square root of its variance
Variation coefficient: relative standard deviation; ratio of the standard
deviation to the mean,
Range: most simple distribution measure, difference between the highest and
the lowest values in the set
Interquartile range: difference between the 75%-and 25% quartile

12

2.2.3 Measures of shape

Deviation from normal distribution


Skewness: Asymmetric = skewed
o Right-skewed = positive skewed main part of distribution on the left
side
o Left-skewed= negative skewed main part of distribution on the right
side
Kurtosis
o Steepness of a distribution
o Positive excess (sharp): maximum higher compared to normal
distribution (with the same variance)
o Negative excess (flat): maximum lower compared to normal
distribution (with the same variance)
SPSS positive value: left-skewed
SPSS negative value: right-skewed
SPSS = 0 : symmetric distribution
EXERCISE: Generate descriptive statistics (quartiles, mean, standard error of
the mean, median, mode, standard deviation, variance, skewness, kurtosis,
minimum, maximum, sum) for meaninglessness.
Analyze Descriptive Statistics Frequencies Select variable(s); under
statistics select the statistics you want to produce: central tendency,
distribution, dispersion and the 25%, 50%, 75% quartile.
The list of variables shows all numerical and string varaibles
The select the variable of interest and click on the arrow ( ) to remove
variable to Variable(s)
To select certain measures, click on Statistic which opens the following
new window: frequencies: statistic

13

Select the statistics of interest


Consider whether you want to selcet: display frequency tables (!) is not
reasonable in this categroy, only reasonable for variables with manageable
amount of scores (think of income)
The default is: no statistics are calculated

!!Every output table needs an interpretation, also in the exams, although it is


not explicitly requested!!
Table interpretation:
o Mean: Average meaninglessness in the sample
o Median: Value that lies in the middle of an ordered range of variables;
Half of the variable is greater or smaller
o Mode: Meaninglessness value that occurs most often
o Percentiles:
25%: value for meaninglessness below which there are exactly
25% of all values of the distribution, 75% of all scores lie above
this value

14

50%: meaninglessness score below which there are exactly


50% of all meaninglessness scores in the sample, 50% of all
scores lie above this value = median
75%: meaninglessness score below which there are 75% of
meaninglessness scores in the sample, 25% lie above this score
o Variance and standard deviation: Variance is not given in the same
unit than the variable, therefore SD is more suitable
o Skewness:
Pile-up left
o Kurtosis
Rather flat (negative value)

TEST: Descriptive Statistics


Data Set: Cosmetic Surgery.sav
Generate quartiles for relatively happy and unhappy people (quality of
life) after surgery

25% Quartil: quality of life after surgery score, below which there are 25%
of all scores, 75% lie above this value (relatively unhappy people after
surgery)
50% Quartil: Median; quality of life after surgery score, below which there
are 50% of all scores, 50% lie above it
75% Quartil: quality of life score, under which there are 75% of all scpres,
25% lie above (relatively happy people after surgery)
Relatively unhappy people: score below 52.1; relatively happy after
surgery: score above 67

15

2.3 Graphs
2.3.1 The art of presenting data
Graphs should (Tufte, 2001):
show the data.
Induce the reader to think about the data being presented (rather than some
other aspect of the graph).
Avoid distorting the data.
Present many numbers with minimum ink.
Make large data sets (assuming you have one) coherent.
Encourage the reader to compare different pieces of data.
Reveal data.
Why is this graph bad?
The bars have a 3D effect, almost impossible to read
Bars have patterns: distract the eye from the data
Cylindrical bars: difficult to understand
Badly labelled y-axis: number of what
Why is this graph better?

16

18

Number of Obsessive
Thoughts/Actions per Day

16

Thoughts
Actions

Error Bars Show 95% CI

14
12
10
8
6
4
2
0
CBT

BT

No Treatment

Therapy

2-d okit: easier to compare values across therapies


Y-axis has a more informative label
Fewer distractions
Minimum ink: no axis lines, no grid

2.3.2 The Chart Builder

17

2.3.3 Histograms: spotting obvious mistakes

A histogram is a graph in which values of observations are plotted on the


horizontal axis, and the frequency with which each value occurs in the data
set is plotted on the vertical axis.
Histograms plot:
o The score (x-axis)
o The frequency (y-axis)
Histograms help us to identify:
o The shape of the distribution
Skew
Kurtosis
Spread or variation in scores
o Unusual scores
Histogram example
A biologist was worried about the potential health effects of music festivals.
Measured the hygiene of 810 concert-goers over the three days of the
festival.
Hygiene was measured using a standardised technique:
o Score ranged from 0 to 4
0 = you smell like a corpse rotting up a skunks arse
18

4 = you smell of sweet roses on a fresh spring day

19

20

2.3.4 The resulting histogram

The first thing that should leap out at you is that there appears to one case
that is very different to the others
All of the scores appear to be squashed up to one end of the distribution
because they are all less than 5, except for one which has a value of 20
This is an outlier: a score different to the rest
o Outliers bias the mean and inflate the standard deviation
The outlier in this diagram is particularly odd because it has a score of 20,
which is above the top of our scale (hygiene scale ranges from 0 to 4), and so
it must be a mistake
How can we detect the outlier? We can go through the 851 concert goers, or
do a boxplot
2.3.5 Boxplots
Boxplots are made up of a box and two whiskers.
The box shows:
o The median
o The upper and lower quartile
o The limits within which the middle 50% of scores lie.
The whiskers show
o The range of scores
o The limits within which the top and bottom 25% of scores lie

21

22

The outlier that we detected in the histogram is shown up as an asterisk (*) on the
boxplot and next to it is the number of the case (611) that causes this outlier

What does the boxplot show?

This is the boxplots for the hygiene scores on day 1 after the outlier has been
corrected

23

What does the boxplot represent?


It shows us the lowest score (the bottom horizontal line on each plot) and the
highest (the top horizontal line of each plot)
Comparing the males and the females we can see they both had similar low
scores (0, or very smelly) but the women had a slightly higher top score (i.e.
the most fragant female was more hygienic than the cleanest male)
The lowest edge of the tinted box is the lower quartile
Therefore, the distance between the lowest horizontal line and the lowest
edge of the tinted box is the range between which the lowest 25% of scores
fall
This box is slightly larger for women than for men, i.e. there is more variability
in the hygienic score of the bottom 25%
The box shows the interquartile range, i.e. 50% of the scores are bigger than
the lowest part of the tinted area but smaller than the top part of the tinted
area
The top edge of the tinted box shows the upper quartile
In the middle of the box is a slightly thicker horizontal line
It represents the median
The median for females is higher than for males, which tells us that the middle
female scored higher than the middle male
TEST: Graphs
Data Set: Cosmetic Surgery.sav
Produce a histogram for the Beck Depression Inventory (BDI). The histogram
should also display a normal distribution. Does the BDI variable follow a
normal distribution?
Produce a simple boxplot for the BDI separated by gender. Can you detect
any differences?

24

Select BDI and drag it into y-axis?, drag gender on x-axis?


2.3.6 Error bar charts
Bar charts are the typical way for people to display means
Chart Builder is the starting point
The bar (usually) shows the mean score
The error bar sticks out from the bar like a whisker.
The error bar displays the precision of the mean in one of three ways:
o The confidence interval (usually 95%)
o The standard deviation
o The standard error of the mean
(you can select under element properties), error bars represent

25

Bar Chart: One Independent Variable


Bar charts are the usual way for people to display means
How you create these graphs in SPSS depends on how you collected your
data, whether the means come from independent cases and are, therefore,
independent, or came from the same cases and so are related
Other possibilities to create bar charts
o Clustered bar charts for independent means
o Simple bar chart for related means
o Clustered bar chart for related means
o Clustered bar charts for mixed designs

26

One independent variable


o The y-axis needs to be dependent variable, or the thing youve
measured, or more simply the thing for which you want to display the
mean
o The x-axis should be the variable by which we want to split the arousal
data
Finally you can ask SPSS to add error bars to your bar chart to create an
error bar chart by selecting display error bars (under element properties)
Error bars from SPSS are suitable only for normally distributed data

27

This graph displays the mean (and the confidence interval of those means) and
shows us that on average, people were more aroused by Memento than they were
by Bridget Jones Dairy
If we are interested in gender effects: clustered bar charts for independent means

There is now an extra drop zone: cluster on X:


Drag our second grouping variable
o Y-axis: arousal
o X-axis: film
o Cluster on: gender

28

Like the simple bar chart this graph tells us that arousal was overall higher for
Memento than Bridget Jones, but it also splits this information by gender
The mean arousal for Bridget Jones shows that males were actually more
aroused during this film than females
This indicates they enjoyed the film more than the women did
For memento, values are more comparable
But this contradicts the idea of a chick flick
o It actually seems that men enjoy chick flicks more than the chicks
TEST: Graph
Data Set: Heavy Metal.sav
Produce and error bar char for mean drug use for each of the different suicide
risks splitted by martial status.

29

Mean drug use (amounts of joints smoked per week) were overall higher for
suicidal individuals compared to non-suicidal individuals
For non-suicidal individuals, those living together had a lower drug use than
those living separated or divorced
For suicidal people however, mean drug use is slightly higher for those living
together than for those individuals that are separated or divorced

3 Distribution (non-parametric) testing, t-Test


SPSS has different procedures for conducting a t-Test. It depends on the question
which option should be used
Dependent t-test
Compares two means based on related data.
E.g., Data from the same people measured at different times.
Data from matched samples.
Independent t-test
Compares two means based on independent data
E.g., data from different groups of people
30

Significance testing
Testing the significance of Pearsons correlation coefficient
Testing the significance of b in regression.
Dependent t-test specifics
Data arrangement
Same individuals, reactions to different conditions (in different columns)
We will not practice the dependent t-test in class, have a look at Andy Field for an
example
Assumptions of the t-test
Both the independent t-test and the dependent t-test are parametric tests based on
the normal distribution.
Can e.g. be tested with a histogram displaying a nornmal curve
Data are measured at least at the interval level.
The independent t-test, because it is used to test different groups of people, also
assumes:
Variances in these populations are roughly equal (homogeneity of variance).
Will be checked with Levenes test within the SPSS procedure
Scores in different treatment conditions are independent (because they come from
different people).

EXERCISE: Independent t-test


Data Set: 01_ChickFlick.sav
Does the average physical arousal during the movie differ between men and
women? How big is the effect size? Also check the t-test assumptions.
Test for normal distribution
Histogram for arousal, arousal on x-axis, select display normal curve, apply, ok

31

Arousal is approximately normally distributed


Analyze Compare means Independent samples t-test select test
variable (arousal) select grouping variable (gender) define groups (1:1
(male); 2:2 (female) ok

3.1 Testing homogeneity of variance:


Hypotheses Levene test:
Tests whether the variance are equal in the population
H0 Levene Test: variances are equal
H1 Levene Test: variance are not equal
The maximal acceptable level of probability of error (level of significance) is
5% (alpha=0,05).
32

p > 0,05: The H0 Variances are equal cannot be rejected, we can assume
that variance are sufficiently equal to assume variance homogeneity. We
interprete the first line in the output table (line: variances are equal
p < 0,05: The H0 variances are equal has to be rejected. In the ouput table
we interprete the line showing variances are not equal
REPORTING Levenes test: For these data, the Levenes test is
nonsignificant (because p=.357) which is greater than .05 and so we say
that the assumption of homogeneity of variances is met and we work with
the upper line of the t-test.
here: p>.05, i.e. the H0 is accepted and we look at the upper line
REPORTING: The descriptive table shows that means are different in our
sample. Mean arousal for males is 21.5 and for females it is 18.55. The
mean difference is 2.95. In order to test if the difference is significant, the
t-test is conducted
Definition of hypotheses
H0: Physical arousal during the movie does not differ between men and
women
H1: Physical arousal during the movie differs between men and women
If both means have the same size in the population, the t-value is 0
We get a value of 1.13
If there is no difference between the means in the population, the observed tvalue can result with a probability of X
The significance level is 26.6%, this means that when we reject the null
hypothesis we make a mistake of 26.6%
We cannot reject the null hypothesis, there is not significant difference in the
mean arousal
REPORTING: On average, men experienced greater physical arousal during
the film (M = 21.50, SE = 7.42), than do females (M = 18.55, SE = 9.02). This
difference was not significant t(38) = 1.13, p > .05; however, it did represent
a small effect r = .18.
Calculation of r: square root of (t^2/(t^2+df))

t2
t 2 df

=0.18

TEST: t-test
Data Set: Record.sav
Does the number of times a record is played on the radio differ dependeing on
whether the band is attractive or unattractive? Also check the t-test assumptions.
Generate a binary variable separating attractive from unattractive bands (attractive
band: attractiveness le 5; unattractive band: attractiveness gt 5)
33

compute attr=0.
if (attract le 5) attr=1.
if (attract gt 5) attr=2.
execute.

The variable is approximately normally distributed (the t-test is robust enough)


H0: The number of times a record is played on the radio does not differ between
bands that are attractive and those that are not
H1: The number of times a record is played on the radio differs between bands that
are attractive and those that are not

34

Testeing homogeneity of variance:


Hypotheses Levene test:
Tests whether the variance are equal in the population
H0 Levene Test: variances are equal
H1 Levene Test: variance are not equal
p > 0,05: The H0 Variances are equal cannot be rejected, we can assume
that variance are sufficiently equal to assume variance homogeneity. We
interprete the first line in the output table (line: variances are equal
p < 0,05: The H0 variances are equal has to be rejected. In the ouput table
we interprete the line showing variances are not equal
here: p>.05, i.e. the H0 is accepted and we look at the upper line
REPORTING Levenes test: For these data, the Levenes test is
nonsignificant (because p=.442) which is greater than .05 and so we say
that the assumption of homogeneity of variances is met and we work with
the upper line of the t-test.
REPORTING: The descriptive table shows that means are different in our
sample. Records of attractive bands are played more often (28.10 times per
week) compared to those of unattractive bands with 23.50 times per week.
The mean difference is -4.6. In order to test if the difference is significant,
the t-test is conducted
If this means that there is a significant difference in the population, the t-test
is conducted
If both means have the same size in the population, the t-value is 0
We get a value of -1.79
If there is no difference between the means in the population, the observed tvalue can result with a probability of X
The significance level is 7.5%, this means that when we reject the null
hypothesis we make a mistake of 7.5%
We therefore are not safe to reject the null hypothesis, there is no significant
difference in the number of times a record is played on the radio depending
on the attractiveness of a band
REPORTING: On average, unattractive bands have a lower airplay (M =
23.50, SE = 10.78), than attractive bands (M = 28.10, SE = 12.39). This
difference was not significant t(198) = -1.79, p > .05. It did represent a small
effect r = .13.

35

4 Correlation analysis
Theses: Relationship between the value of two variables
H0= there is no linear relationship between the variables
H1= there is a linear relationship between the variables
Measure of the strength and direction of this relationship is the correlation
coefficient
Identifies linear relationships
Expresses the strength of the relationship with a value between -1 and +1
Positive value: positive linear relationship
Negative value: negative linear relationship
Correlation coefficient becomes higher, the stronger the linear relationship
between the variables is
Correlation= 0 : no linear relationship
Orientation guidelines (absolute values)
Absolute value of the correlation Possible interpretation
coefficient
0

No correlation

> 0 bis 0,2

Very weak correlation

0,2 bis 0,4

Weak correlation

0,4 bis 0,6

Medium correlation

0,6 bis 0,8

Strong correlation

0,8 bis < 1

Very strong correlation

Perfect correlation

As the correlation coefficient is restricted to detect linear relationships, also with a


low correlation coefficient we could have a perfect non-linear correlation
To detect these types of effects, it is a good idea to plot the variables against
each other to detect other relationship forms: use a scatter plot: this gives a
first visual impression about a possible relationship
Correlation analysis is suitable for intervall scaled variables

36

4.1 Doing a correlation analysis in SPSS: Analyze Correlate

EXERCISE:
Data Set: 02_Record.sav
Is there a linear relationship between amount spent on advertising and record sales
(number of copies sold). Produce a scatterplot for the two variables.
Graphs Legacy dialog scatter simple scatter select variables for X- and
Y- axis, ok

37

We can see that there is a tendency towards a positive relationship, but we cannot
see how strong this relationship is
Analyze Correlate bivariate, select advertising budget and record sales; keep
default Pearsons correlation coefficient
As a default/ or standard setting, the Pearsons correlation coefficient is
calculated (variables need to have a ratio scale level)
Spearman and Kendall-Tau-B for ordinally scaled variables (variable values
are translated into ranks) rank correlation for ordinally scaled data
Select variables ok

The null hypothesis states: There is no linear relationship between the


variables.
The alternative hypothesis is: There is a linear relationship between the
variables
The maximal acceptable level of significance is 5% (alpha=0,05).

38

The advertising budget was significanlty correlated with record sales,


r=.578, p<.001. The value for r represents a positive and large effect.
Null hypothesis is rejected
The significance test examines whether there is a relationship between the
variables, but does not say something about the strength of the relationship!!
The strength is given with the correlation coefficient
The significance test is the probability at which the correlation
coefficient showing in the sample with the given sample size also shows
with the same size in the population if there is in reality no relationship
in the population
As it was unkown at the beginning of the analysis whether there is a positive
or negative relationship between the variables, a 2-tailed test has been
conducted (independent if there is a negative or positive relationship)

Data Set: 07_Supermodel.sav


Is there a linear relationship between the beauty of a supermodel
(attractiveness in %) and the salary per day? Produce a scatterplot for the
two variables.

Scatter plot is not very useful, seems to be a positive relation, but not very
clear
Based on the scatter plot we cannot determine whether the is a linear
relationship
Procedure: Analyze correlate bivariate
39

The null hypothesis states: There is no linear relationship between the


variables.
The alternative hypothesis is: There is a linear relationship between the
variables
The maximal acceptable level of significance is 5% (alpha=0,05).
The salary per day was not significanlty correlated with the beauty of a
supermodel (attractiveness in %), r=.068, p>.05. The value for r
represents a positive and small effect.
Null hypothesis cannot be rejected
The significance test examines whether there is a relationship between the
variables, but does not say something about the strength of the relationship!!
The strength is given with the correlation coefficient
The significance test is the probability at which the correlation
coefficient showing in the sample with the given sample size also shows
with the same size in the population if there is in reality no relationship
in the population
As it was unkown at the beginning of the analysis whether there is a positive
or negative relationship between the variables, a 2-tailed test has been
conducted (independent if there is a negative or positive relationship)

4.2 Cross tables


Are used to show the joint frequency distribution of two variables
Case groups of variables are examined, that are defined by a combination of
characteristics of both variables
Moreover, there is a statistical test which tests whether or not there is
relationship between the variables
Examines, at which probability the objective relationship of the sample also
exists in the population
Cross tables
40

o Show the joint frequency of two variables


o Not only absolute frequencies (as with the single frequency table)
but also the relative frequencies = expected frequencies are
shown
o Differences in the values of the cross table and the frequency
table can exist, as persons with missing values are not included in
the analysis
o The significance test to analyse possible relationships is based on
the expected frequencies of two variables
Chi-Quadrat-test:
o Examines whether we can infer from the sample observations that
there is relationship between these 2 variables also in the population or
not
Measures of association:
o Aim at expressing the strength of the relationship between 2 variables
with a measure
Categorical variables are very suitable
Analyze Descriptive Statistics Crosstabs

41

EXERCISE:
Data set: 04_CosmeticSurgery.sav
Is there an association between the reason to have cosmetic surgery and
gender?

Table contains the number of cases that fall into each combination of
categories
We can see that 98 of people (35.5% of total) had a cosmetic surgery due
to a physical reason and of these 82 were female (83.7% of the people
having a surgery for a physical reason) and 16 were male (16.3% of all
people with surgery due to a physical reason).
Further, 178 people (64.5% of total) had a cosmetic surgery in order to
change their appearance. Of those having a surgery to change
42

appearance, 75 were female (42.1% of those having a cosmetic surgery


to change appearance) and 103 were male (57.9% of those having a
cosmetic surgery to change appearance).
Of the females, 52.2% had a cosmetic surgery for a physical reason and
47.8% to change appearance. Similarly, of the males, 13.4% had a
cosmetic surgery for a physical reason and 86.6% to change appearance
In summary, for females, the reason to have cosmetic surgery is in equal
parts physical but also to change appearance. For males however, the
vast majority has cosmetic surgery to change appearance
H0: the variables are independent (there is no association between
gender and the reason for cosmetic surgery)
H1: the variables are dependent (there is an association between gender
and the reason for cosmetic surgery)

Checking the assumptions of the Chi-square test: all expected


frequencies should be greater than 5. Smallest count according to
footnote a is 42.25, so the assumption has been met
Pearsons Chi-square test examines whether there is an association between
two categorical variables, test statistic tests whether the two variables are
independent. If the significance value is small enough, we reject the null
hypothesis that the variables are independent
The value of the Chi-squre statistic is 44.464 and the value is highly
significant (p<.001), indicating that gender had a significant effect on the
reason to have cosmetic surgery
Therefore we can conclude, gender is significantly associated to the reason
to have cosmetic surgery, cosmetic surgery to change appearance is much
more likely for males than for females

43

The table with the symmetric measures shows the strength of the
association. For Cramers Phi has a value of 0.401 indicating a medium effect
REPORTING: There was a significant association between gender and the
reason to have cosmetic surgery 2 (1) = 44.464, p< .001. The size of the
effect was medium.
This finding seems to represent the fact that based on the odds ratio, the
odds of having cosmetic surgery for a physical reason were 6.81 times
higher if the people were females comcpared to males
Standardized residuals show that for females the standardized residuals were
significant for both having cosmetic surgery for a physical reason or to
change appearance (both values are bigger than 1.96; z=3.5 and -2.6). The
plus or minus sign tells us that for females significantly more had cosmetic
surgery for a physical reason and significantly less had surgery to change
appearance. For males, significantly less had cosmetic surgery for physical
reason but more to change appearance (z=-4.0 and 3.0, resp.). All cells
contribute significantly to the overall Chi-square statistic

TEST CROSSTABS
Data Set: 08_Facebook.sav
Is there an association between RMiP Exam Result and whether or not
students looked at Facebook during lab classes?

44

Table contains the number of cases that fall into each combination of
categories
We can see that 168 of people (64,6% of total) passed the exam and of
these 44 looked at FB (26.2% of the students passing) and 124 did not
look at FB (73.8% of students passing).
Further, 92 people (35,4% of the total) did fail. Of those failing, 60 did
look at FB (65,2% of those failing) and 32 did not look at FB (34,8% of
those failing).
Of those looking at FB, 42,3% did pass and 57,7% did fail. Similarly, of
those not looking at FB, 79,5% passed and 20,5% did not
In summary, most students passing the exam did not look at FB during
class. Of those students failing the exam, the share of students looking
at FB is higher.
H0: the variables are independent (there is no association between exam
result and looking at FB during class)
H1: the variables are dependent (there is an association between exam
result and looking at FB during the exam)

45

Checking the assumptions of the Chi-square test: all expected


frequencies should be greater than 5. Smallest count according to
footnote a is 36.80, so the assumption has been met
Pearsons Chi-square test examines whether there is an association between
two categorical variables, test statistic tests whether the two variables are
independent. If the significance value is small enough, we reject the null
hypothesis that the variables are independent
The value of the Chi-squre statistic is 37.726 and the value is highly
significant (p<.001), indicating that there is an association between the exam
result and looking at FB during class.
Therefore we can conclude, looking at FB or not has a significant effect on
the result of the exam

The table with the symmetric measures shows the strength of the
association. For Cramers Phi has a value of -0.381 indicating a medium
effect
REPORTING: There was a significant association between looking at FB
during class and the exam result 2 (1) = 37.726, p< .001. The effect size is
medium.
This finding seems to represent the fact that based on the odds ratio, if the
studetns looked at FB during the class the odds of their failing the exam
was 5.23 times higher than if they had not looked

46

Standardized residuals show that for those looking at FB the standardized


residuals were significant for both failing and passing the exam (both values
are bigger than 1.96; z=-2,8 and 3,8). The plus or minus sign tells us that for
those looking significantly more failed and significantly less passed the exam
after looking at FB. For those not looking at FB, significantly more passed but
less failed (z=2,3 and -3.1, resp.). All cells contribute significantly to the
overall Chi-square statistic

5 Regression analysis
5.1 ANOVA
Similar hypotheses to the t-Test:
Are the means of different subgroups of the same variables equal?
Different to the t-Test:
T-test compares 2 means, ANOVA compares more than 2 means
Differentiation between:
One-way ANOVA (1 independent variable) = ANOVA

Multivariate ANOVA
(> 1 independent variable)
One-dimensional ANOVA
(1 dependent variable)

Multidimensional ANOVA (> 1 dependent variable)


Data requirements::
Variable of interest needs to have interval scale
47

Data needs to be normally distributed


Variance homogeneity: the variance of the variable of interestw needs to
be equal among the different categories

EXERCISE:
Data Set: 06_Soya.sav
Run the following analysis excluding those males that ate only one soya meal
per week (filter cases).
Is there a difference in the mean number of sperms depending on the
frequency of soya consumption (no soya meals; four soya meals per week;
seven soya meals per week)?
Select cases
o Data select cases if condition is satisfied: Soya=1 | Soya=3 |
Soya=4 ok
Check if number of soya meals is normally distributed
Graph Legacy Dialogs Histogram, select display normal curve

48

Variable is not normally distributed


One option is to transform the dependent variable
compute sperm2=sperm*sperm.
execute.
compute lnsperm=ln(sperm).
execute.
Check if transformed variables are normally distributed
Analyze Nonparametric tests Legacy dialogs 1 sample KS test
variables: sperm count, sperm2, lnsperm

49

Lnsperm is normally distributed, use this transformed variable to continue


with analysis
Analyze compare means one-way ANOVA select lnsperm as
dependent variable under posthoc select LSD, Tukey, Games-Howell
under options select descriptive, homogeneity of variance ok

Table of descriptives from the one-way procedure


We see that means are different depending on the frequency of soya meals
per week, highest for no soya meals, lowest for 7 times soya per week
REPORTING: The table shows the descriptive statistics of the dependent
variable, the logarithm of sperm count. It shows that means are different,
with the highest means for those eating no soya per week and the lowest
for those eating soya 7 times per week

The next part is a summary of Levenes test to test the assumption of


homogeneity of variances
IMPORTANT: If assumption is broken, select the Welch and BrownForsythe option (robust tests of the equality of means when assumption of
homogeneity of variances is broken) and report those F-ratios and Sig
values as results!
REPORTING: The null hypothesis for Levenes test is: variances are equal.
The alternative hyhpothesis is: variances are not equal. Levenes test for
equality of variances was found to be met for the present analysis,F(2,57) =
1.876, p = .163

50

Main ANOVA summary table


REPORTING: The null hypothesis is: The group means are equal in the
population. Mean sperm count (the ln of it) does not differ depending on
the frequency of soya consumption. The alternative hypothesis is: The
group means are different in the population. Mean sperm count (the ln of
it) does differ depending on the frequency of soya consumption.
Table is divided into between-group effects (effects due to the model the
experimental effect) and within-group effects (this is the unsystematic
variation in the data), sum of squres, df and mean squares are reported
The test of whether group means are the same is represented by the F-ratio
It is greater than 1 indicating a significant result
Finally, SPSS tells us whether this value is likely to have happened by
chance
The final column labelled Sig. indicates the likelihood of an F-ratio of this size
would occur if in reality there was no effect (that is only a 1.6% chance)
Hence, because the observed significance value is less than .05 we can say
the there was a significant effect of soya consumption frequency

51

We asked SPSS to carry out post hoc tests


This Table shows the results of the Tukey test, LSD and Games-Howell test
If we look at the Tukey test first, it is clear from the table that each group of
participants is compared to all of the remaining groups
For each pair of groups the difference between group means is displayed, the
standard error of t hat difference, the significance level of that difference and
a 95% confidence interval
First of all, the no soya meal group is compared to the 4 soya meals per week
and reveals a nonsignificant difference (Sig. is greater than .05), but when
compared to the 7 soya meals per week group, there is a significant
difference (Sig is less than .05)
The 4 soya meals per week is then compared to both the no soya meals and
the 7 soya meals per week groups. The group means differ only between the
4 and 7 group
The rest of the table reports the LSD and Games-Howell test. A quick
inspection reveals the same patterns of results: sperm counts differs between
no and 7 and 4 and 7 but not between no and 4
REPORTING: A one-way ANOVA was used to test the sperm count
differences between three different consumption frequencies of soya
meals. There was a significant effect of soya meals consumption frequency
on sperm count, F(2, 57)=4.43, p<.05.
Post hoc analyses using the Tukey HSD and Games-Howell test indicated
that the average sperm count was lower in the 7 soya meal per week group
(M=.26, SD=1.01) than in the other two conditions (no soya meals: M=1.1,
SD=1.11; 4 soya meals: M=.97, SD=.97), p<.05. The difference between the
no soya and 4 times soya group was not significant.

TEST: Data Set: 07_Supermodel.sav


Generate a new variable by dividing the age of the supermodel into the
following categories:
o youngster: age lt 17 years old
o best age: between 17 and 19 years old
o senior: older than 19 years
Does the average beauty (attractiveness in % judged by a jury) differ by age
category?
compute agecat=0.
if (age lt 17) agecat=1.
if (age ge 17 and age le 19) agecat=2.
if (age gt 19) agecat=3.
execute.

52

Check for normality of dependent variable


Histogram
Variable is normally distributed

Run one-way ANOVA, select posthoc and Welch and Brown-Forsythe (robust
tests of the equality of means when assumption of homogeneity of variances
is broken)

Table shows descriptive statistics, it shows that the mean beauty differs by
age category, especially the beauty of the youngsters is rated lower than that
of the two other categories.
REPORTING: The table shows the descriptive statistics of the dependent
variable, the attractiveness in %. It shows that means are different, with
the highest means for middle age category and the lowest for the youngest
53

The next part is a summary of Levenes test to test the assumption of


homogeneity of variances
REPORTING: The null hypothesis for Levenes test is: variances are equal.
The alternative hyhpothesis is: variances are not equal. Levenes test for
equality of variances was found to be violated for the present
analysis,F(2,228) = 3.304, p < .05.
For this reason, the Robust Tests for Equality of Means are used to interprete
the ANOVA results

IMPORTANT: Because assumption was broken, we selected the Welch and


Brown-Forsythe option (robust tests of the equality of means when
assumption of homogeneity of variances is broken) and report those Fratios and Sig values as results and not the results in this ANOVA output!

Main robust tests of equality of means summary table


REPORTING: The null hypothesis is: The group means are equal in the
population. Mean attractiveness (in %) does not differ depending on the
age category. The alternative hypothesis is: The group means are different
in the population. Attractiveness (in %) does differ depending on the age
category.
The output from the above table is only valid if the equal variance
assumption has been violated. From this example, using the Welch
54

statistic, we find that F(2, 137.78) = 10.061, p <.001. This indicates that there
is a significant difference in the mean attractiveness in % and the age
category.

We asked SPSS to carry out post hoc tests


This Table shows the results of the Tukey test and the Games-Howell test
If we look at the Tukey test first, it is clear from the table that each group of
participants is compared to all of the remaining groups
For each pair of groups the difference between group means is displayed, the
standard error of t hat difference, the significance level of that difference and
a 95% confidence interval

First of all, the group of the youngest is compared to the group of the middle
aged supermodels and reveals a significant difference (Sig. is smaller than
.01), also when compared to the group of the oldest supermodels, there is a
significant difference (Sig is less than .05)
The group of the middle-aged supermodels is then compared to both the
youngest and the oldest group. The group means do not differ between the
middle-aged and the oldest group
The rest of the table reports the Games-Howell test. A quick inspection
reveals the same patterns of results: mean attractiveness in % differs
between the youngest and the middle-aged and the youngest and the oldest,
but not between the middle-aged and the oldest
REPORTING: A one-way ANOVA was used to test the attractivenenss
differences between three different age categories of supermodels. There
was a significant effect of the age category on attractiveness, F(2, 137.78) =
10.061, p <.001.
Post hoc analyses using the Tukey HSD and Games-Howell test indicated
that the mean attractiveness was lower in the group of the youngest
55

supermodels (M=72.87, SD=7.40) than in the other two conditions (middleaged: M=77.44, SD=5.99; oldest: M=77.16, SD=6.05), p<.05. The difference
between the middle-aged and the group of the oldest supermodels was not
significant.

5.2 Regression analysis


Analyses the relationship between a dependent and one or more independent
variables
Goal: To describe relationships quantiatively, to explain the values of the dependent
variable
Application:
Cause analysis
Prediction of the dependent variable
Time series analysis
Principle: A line is searched that minimises the derivations of single values from the
line = Minimisation of squared residuals (least squares)
Difference to correlation analysis: the direction of the relationship needs be clearly
given (e.g. income has an impact on the expenditures for food)

5.2.1 Assumptions:
There must be causality
Variables need to have a metric scale
Introduction of independent nominally-scales variables as dummy variable
Linear relationship between dependent and independent variables
Model assumption (will not be verified in the course)
Normally distributed residuals
Linear relationship between the dependent variable and the independent variables
No multicollinearity
No heteroscedasticity
No autocorrelation

Data set: 03_HeavyMetal.sav


Does drug use (amount of joints smoked in one week) depend on age?
First of all, a scatter plot is useful
GraphsY Legacy dialogs Scatter/Dott Simple scatter defineYAchse drug usex-Achse normelssness
The selection of the axes is important! The y-axis shows the dependent, the x-axis
show the independent variable

56

5.2.2 Model

Y = f(X)
Y = b0+b1*x
Drug use=f(normlessness)
Drug use=3.64+0.50*normlessness+e

B0: Constant, which is the value if X=0.


B1: Regression coefficient determines the slope of the line, thus it determines the
effect of a change in X B1 = Y / X
Conduct a regression analysis
AnalyzeRegressionlinear

57

5.3 Verification of the regression function:


Quality criteria to verify the regression function
Coefficient of determination R2
F-Statistic
Standard error
Measures to verify the regression coefficients
T-value
Beta-value
5.3.1 Coefficient of determination: R2
Analyses the quality of the regression function to empirical data
Measures the share of explained variation on the total variation
R = explained variation / total variation
Measures how cloesly the regression line lies at the points of the scatter
Values between 0 and 1; 1= regression line is consistant with the scatter, the higher,
the higher the share of explained variation on the total variation
The addition of another explaining variable can never worsen the R2, it can only
improve the R2; this can tempt to include as many explaining variables in the
estimation as possible
R depends on the number of regressors. At a given sample size each additional
regressor adds a bigger or smaller share of explanation, which might only be a
coincidence. The value of R can never decline with additional regressors. This is
why the adjusted R2 is used.
5.3.2 Adjusted R
An increase in the number of explaining variables can result in a decline of the
adjusted R
5.3.3 Standard error of the estimation
Measure about how close the predicted values come to the real values. In other
words: What is the mean error that is made from estimating the regression function
to estimate the dependent variable Y.
The R2 of the regression analysis shows that the normlessness explains drug use
with 20.7%.
The explanatory variable available income has an explanatory power of 20.7% for
the dependent variable drug use
The regression model explains the variation of the dependent variable to 20.7%
REPORTING: The quality of the regression is determined by the R2. Drug
use can be explained with 20.7% with the variable normlessness. This
means that the independent variable in the regression explains the
distribution of the dependent variable to 20.7%.
58

Durbin-Watson statistic (not relevant for exam)


DW tests the assumption of independent errors
For any two observations the residual terms should be uncorrelated (or independent)
This eventuality is sometimes described as lack of autocorrelation
This assumption can be tested with the Durbin-Watson test, which tests for serial
correlations between errors
Specifically, it tests whether the adjacent residuals are correlated
The test statistic can vary between 0 and 4 with a value of 2 meaning that the
residuals are uncorrelated
A value greater than 2 indicates a negative correlation between adjacent residuals,
whereas a value below 2 indicates a positive correlation
The size of the Durbin-Watson statistic depends upon the number of predictors in
the model and the number of observations
As a rule of thumb, values less than 1 or greater than 3 are definitely cause for
concern, values around 2 are ok
Here, the DW value is close to 2, ok

5.4 F-Test to test the significance of the regression function


Afterwards, an F-Test is conducted. The F-test checks whether the regression
function is significant, that is if at least 1 coefficient of the model is significant
Can the result from the sample be transfered to the population?
Next to the distribution, also the sample size is considered
H0: There is no linear relationship between the dependent and the independet
variables, all betas =0
H1: There is a linear relationship between the dependent and the independent
variables, betas are 0.
To reject H0 means that there is a linear relationship between the dependent and
independent variable also in the population.
REPORTING: To observe whether the regression model is significantly
better in predicting the model than using the mean, an F-test is conducted.
H0: There is no linear relationship between the dependent and the
independet variable, all betas in the population equal 0; H1: There is a
linear relationship between the dependent and the independent variable,
59

betas are different from 0. Here, the F test statistic value is 32.24 and it is
highly significant (p<.001). This means that at least one coefficient in the
regression function is different from 0.

Coefficients table reports the test results for the single predictors in the model

5.5 Test of single regression coefficients


If the global test of the regression function was significant, that is not all coefficients
are 0, and thus a sigificance between the variables has been shown, the single
regression coefficients needs to be tested next. H0 is: bj = 0. If the empircal value is
different from 0, it is very unlikely that the null hypothesis is correct.
If there is only one independent variable in the model, the F-test of the model is also
a test of the variable whose impact is examined. It would be enough to conduct only
one of the tests.
Using the t-Test, single regression coefficients are tested whether or not they are
different from 0
Contrary to the F-value, a t-value can become negative that is why one has to
consider the absolute value.
T-value higher than 2 for at least 30 observations (circa)
T-value shows the ratio between the coefficient and the standard deviation
Significance level: >0,05 Beta not significant
The impact of the normlessness is significant, as p=0 and therefore lower 0.005
REPORTING: Regression function: Drug Use=3.64+0.50*normlessness+e
The null hypothesis of the constant is: the constant is 0. The alternative
hypothesis is: the constant is not 0. The constant shows the height of drug
is if normless is 0. The t-test statistic has a value of 5.59 and is highly
significant (p<.001). It shows that baseline drug use is 3.64 joints smoked
per week (given that normlessness is 0).
The null hypothesis for normlessness is: beta=0. The alternative hypothesis is: beta
is not 0. There is a linear relationship between drug use and normlessness. The ttest statistic has a value of 5.68 and the value is highly significant (p<.001) meaning
that if normlessness increases about 1 unit, drug use increases about 0.5 units.
Single regression analysis was used to test if normlessness significantly predicted
drug use. The results of the regression indicated the predictor explained 20.7% of
the variance (F(1,119)=32.24, p<.001). It was found that normlessness significantly
predicted drug use (beta= 0.50, p<.001).

60

EXERCISE II:
Data set: 03_HeavyMetal.sav
Test the hypothesis that drug use depends on normlessness and being
young. Discuss the quality of your analysis and report the regression function.
Generate a dummy variable for young using the variable age_group=1.
compute young=0.
if (age_group=1) young=1.
execute.
Why not simply use binary age_group variable? Because it is not a dummy
variable, needs to be 0/1 coded, data set description shows it is 1/2 coded
Dummy Variables:
By using dummies it is possible to also use qualitative normal scale variables. These
normal scaled variables are transferred into binary variables which can be handled
as metric variables. As however, the number of independent variables increases
considerably, this technique is only possible for the dependent variable (but logit
regression).
One of the dummies needs to be excluded and becomes the reference category!

REPORTING: The quality of the regression is determined by the R2. Drug


use can be explained with 21.8% with the variables normlessness and
being young. This means that the independent variables in the regression
explain the distribution of the dependent variable to 21.8%.

61

REPORTING: To observe whether the regression model is significantly


better in predicting the model than using the mean, an F-test is conducted.
H0: There is no linear relationship between the dependent and the
independet variables, all betas in the population equal 0; H1: There is a
linear relationship between the dependent and the independent variables,
betas are different from 0. Here, the F test statistic has a value of 16.44 and
it is highly significant (p<.001). This means that at least one coefficient in
the regression function is different from 0.

REPORTING:
Regression
function:
Drug
Use=3.52+0.48*normlessness+0.45*young+e
The null hypothesis of the constant is: the constant is 0. The alternative
hypothesis is: the constant is not 0. The constant shows the height of drug
if normless and young are 0. The t-test statistic has a value of 5.26 and it is
highly significant (p<.001). It shows that baseline drug use is 3.52 joints
smoked per week (given that the model predictors are 0).
The null hypothesis for normlessness is: beta=0. The alternative hypothesis
is: beta is not 0. There is a linear relationship between drug use and
normlessness. The t-test statistic has a value of 5.37 and the value is highly
significant (p<.001) meaning that if normlessness increases about 1 unit,
drug use increases about 0.48 units.
The null hypothesis for young is: Drug use does not differ between young
and old. The alternative hypothesis is: drug use does differ between young
and old. The t-test statistic has a value of 0.846 and the value is not
significant (p>.05) indicating that there is no difference in drug use
between young and old. (If it would be significant: Young heavy metal
listeners have a drug use which is about 0.45 units higher than the drug
use of the old heavy metal listeners.)
Multiple regression analysis was used to test if normlessness and being
young significantly predicted drug use. The results of the multiple
regression indicated the predictors explained 21.8% of the variance
(F(2,118)=16.44, p<.001). It was found that normlessness significantly
predicted drug use (beta= .48, p<.001) and that being young not
significantly predicted drug use (beta=0.45; p>.05).

62

5.6 Comparison of the standardised and not-standardised betacoefficient:


It would be wrong to infer from the fact that the coefficient of one variable is ten
times higher than the coefficient of another variable that the first variable also has a
ten times higher impact
It needs to be considered that the variables have a different unit, and that is why the
variables are different
To compare the magnitude of the impact of single variables, the standardised
coefficients are to be used (beta coefficients)
The beta coefficients are the coefficients that would result if the dependent and all of
the independent variables would have been standardised (z- transformation)
The interpretation of the beta coefficients is only acceteptable if the independent
variables are not correlated with each other, thus there is no multcollinearity
Regression test:
Data Set: 08_Soya.sav
Does the number of sperms depend on the number of times soya eaten per
week?
Generate dummy variables for each of the 4 categories of soya
Attention: at best, a dummy variable includes not only dummies
Attention II: if the four dummy variables, only three enter the regression (the one
excluded is the reference)
Attention III: DW statistic not optimal, we continue to learn about how to report
regression
Attention IV: F-statistic not significant, the model is not good! Just to see how to
use dummies when we have more than 2 categories!!

REPORTING: The quality of the regression is determined by the R2. Spper


count can be explained with 9.4% with the variables once, seven and four.
63

This means that the independent variables in the regression explain the
distribution of the dependent variable to 9.4%.

REPORTING: To observe whether the regression model is significantly


better in predicting the model than using the mean, an F-test is conducted.
H0: There is no linear relationship between the dependent and the
independet variable, all betas =0; H1: There is a linear relationship between
the dependent and the independent variable, betas are 0. Here, the F test
statistic has a value of 2.64 and it is not significant (p>.05). This means that
the regression model is not a good one (and one should not continue
reporting the results)

REPORTING: Regression function: Sperm count =4.99-0.38*once-0.88*four3.33*seven+e


The null hypothesis of the constant is: the constant is 0. The alternative
hypothesis is: the constant is not 0. The constant shows the height sperm
count if the predictors are all 0. The t-test statistic has a value of 5.40 and
is highly significant (p<.001). It shows that baseline sperm count is 4.99
(millions) (given that the model predictors are 0).
The null hypothesis for once is: Sperm count does not differ between those
eating no soya meal per week and those eating it once. The alternative
hypothesis is: sperm count does differ between no and once. The t-test
statistic has a value of -0.292 and the value is not significant (p>.05)
indicating that there is no difference in sperm ocunt between no and once.
(If it would be significant: Those eating soya meals once a week have a
64

sperm count that is -.382 lower than that of those eating no soya meals per
week)
The null hypothesis for four is: Sperm count does not differ between those
eating no soya meal per week and those eating it four times a week. The
alternative hypothesis is: sperm count does differ between no and four.
The t-test statistic has a value of -0.671 and the value is not significant
(p>.05) indicating that there is no difference in sperm ocunt between no
and once. (If it would be significant: Those eating soya meals four times a
week have a sperm count that is -.671 lower than that of those eating no
soya meals per week)
The null hypothesis for seven is: Sperm count does not differ between
those eating no soya meal per week and those eating it seven times a
week. The alternative hypothesis is: sperm count does differ between no
and seven. The t-test statistic has a value of -2.55 and is significant (p<.05)
indicating that there is a difference in sperm count between no and seven.
Those eating soya meals seven times a week have a sperm count that is 2.55 lower than that of those eating no soya meals per week.

6 Theoretical basis of nonparametric tests


6.1 When to use non-parametric tests:
Non-parametric tests are used when assumptions of parametric tests are not met.
It is not always possible to correct for problems with the distribution of a data set
In these cases we have to use non-parametric tests.
They make fewer assumptions about the type of data on which they can be used.
Non-parmetric tests
Testing of hypotheses about unknown distributions in a population
Application:
If the variables are nominally or ordinally scaled and/ or
If the assumption about the distribution in the population are not sufficient
3 types:
Goodness of fit: Is the hypothetical distribution compatible to an observed
distribution?
Test of independence: Are two characteristics independent of each other? (Chisquare-test with cross tables)
Homogeneity test: Are two or multiple samples from the same population? or Is
there a significant difference between 2 distributions?
Assessing normality: Kolmogorov-Smirnov test
We dont have access to the sampling distribution so we usually test the observed
data
Central Limit Theorem
If N > 30, the sampling distribution is normal anyway
Graphical displays
65

P-P Plot (or Q-Q plot)


Histogram
Values of Skew/Kurtosis
0 in a normal distribution
Convert to z (by dividing value by SE)
Kolmogorov-Smirnov Test
Tests if data differ from a normal distribution
Significant = non-Normal data
Non-Significant = Normal data
Data Set: 06_Soya.sav
Run the following analysis including only those individuals that eat none or one
soya meal per week (filter cases)
Is the sperm count normally distributed?
Filter: Data Select cases If condition is satisfied If Soya <=2 Continue
ok
K-S test: Analyze Descriptive Statistics Explore drag and drop Sperm
count (Millions) to Dependent list Statistics keep default Plots: keep
defaults and select normality plots with tests continue options: default
ok

66

67

The first table produced by SPSS contains descriptive statistics (mean etc.)
The important table is that of the K-S test
This table includes the test statistic itself (D), the degrees of freedom (which should
equal the sample size) and the significance value of this test
A significant value (i.e. less than p<.05) indicates a deviation from normality
The K-S test for the sperm count is highly significant, indicating that the distribution
is not normal
SPSS also produces a normal Q-Q plot for any variables specified, it plots the value
you would expect to get if the distribution were normal (expected values) against the
values actually seen in the data set (observed values)
If the data are normally distributed, then the observed values (the dots) should fall
exactly along the straight line
REPORTING: The sperm count in million, D (40)=0.19, p<.05 is significantly
not normal.

6.2 Mann-Whitney test:


These tests are the non-parametric equivalent of the independent t-test.
Use either to test differences between two conditions in which different participants
have been used
EXERCISE: Mann-Whitney test
Data Set: 06_Soya.sav
Keep the filter from the K-S test before
68

Analyze whether the sperm count in millions is different for people eating no
soya meals per week compared to those eating one soya meal per week.
Data set description: p.559
Analyze Nonpar Legacy dialogs 2 independent samples sperm count
as test variable soya as grouping variable (define groups, group 1: 1 for no
soya meals; group 2:2 for one soya meal per week) test type Mann Whitney U
ok

REPORTING: Sperm count per million of individuals eating no soya meals


per week (Mn rank 20.95) does not differ significantly from individuals
eating one soya meal per week (Mn rank 20.05), U=191, z=-0.243, ns, r=0.04. The relationship is very small.
Calculating the effect size

Z
r
N
R=-0.243/Wurzel aus 40=-0.04

69

You might also like