What Statistical Analysis Should I Use?: Sunday, June 4, 2017 04:22 AM

What statistical analysis
should I use?
1
Sunday, June 4, 2017 04:22 AM
2
Introduction
For a useful general guide see
Policy: Twenty tips for interpreting scientific claims : Nature N
ews & Comment
William J. Sutherland, David Spiegelhalter and Mark Burgman
Nature Volume: 503, Pages: 335337 Date published: (21
November 2013).
Some criticism has been made of their discussion of p values,

see Replication, statistical consistency, and publication bias G.
Francis, Journal of Mathematical Psychology, 57(5) (2013), pp.
153169.
Index End
3
Introduction
These examples are loosely based on a UCLA tutorial sheet. All
can be realised via the syntax window, appropriate command
strokes are also indicated. The guidelines to the APA reporting
style is motivated by Using SPSS for Windows and Macintosh:
Analyzing And Understanding Data Samuel B. Green and Neil J.
Salkind.
These pages show how to perform a number of statistical tests

using SPSS. Each section gives a brief description of the aim of
the statistical test, when it is used, an example showing the
SPSS commands and SPSS (often abbreviated) output with a
brief interpretation of the output.
Index End
4
About the A data file
Most of the examples in this document will use a data file called A, high
school and beyond. This data file contains 200 observations from a
sample of high school students with demographic information about the
students, such as their gender (female), socio-economic status (ses) and
ethnic background (race). It also contains a number of scores on
standardized tests, including tests of reading (read), writing (write),
mathematics (math) and social studies (socst).
5
Syntax:-
display dictionary
/VARIABLES id female race ses schtyp prog read write math science socst.
Variable Position Label Value Label
id 1
female 2 .00 Male
1.00 Female
1.00 Hispanic
race 3 2.00 Asian
3.00 african-amer
4.00 White
1.00 Low
ses 4 2.00 Middle
3.00 High
1.00 Public
schtyp 5 type of school
2.00 private
1.00 general
prog 6 type of program 2.00 academic
3.00 vocation
read 7 reading score
write 8 writing score
math 9 math score
science 10 science score
socst 11 social studies score
6
Index End 7
One sample t-test
A one sample t-test allows us to test whether a sample mean (of a
normally distributed interval variable) significantly differs from a
hypothesized value. For example, using the A data file, say we wish to
test whether the average writing score (write) differs significantly
from 50. Test variable writing score (write), Test value 50. We can do
this as shown below.
Menu selection:- Analyze > Compare Means > One-Sample T test
Syntax:- t-test
/testval=50
/variable=write.
8
One sample t-test
Note the test value of 50 has been selected

9
One sample t-test
One-Sample Statistics
N Mean Std. Deviation Std. Error Mean

The mean of the variable write
for this particular sample of
writing score 200 52.7750 9.47859 .67024
One-Sample Test students is 52.775, which is

Test Value = 50 statistically significantly
Mean different from the test value of
t df Sig. (2-tailed) Difference
50. We would conclude that this
group of students has a
writing score 4.140 199 .000 2.77500
One-Sample Test
significantly higher mean on the
Test Value = 50
writing test than 50. This is
95% Confidence Interval of the
Difference consistent with the reported
Lower Upper confidence interval (1.45,4.10)
writing score 1.4533 4.0967 that is (51.45,54.10) which
excludes 50, of course the mid-
Confidence interval point is the mean.
Crichton, N.
Journal Of Clinical Nursing 8(5) 618-618 1999
10
One sample t-test
Effect Size Statistics
SPSS supplies all the information necessary to compute an effect size, d ,
given by:
d = Mean Difference / SD
where the mean difference and standard deviation are reported in the SPSS
output. We can also compute d from the t value by using the equation
t
d
N
where N is the total sample size. d evaluates the degree that the mean on
the test variable differs from the test value in standard deviation units.
Potentially, d can range in value from negative infinity to positive infinity. If
d equals 0, the mean of the scores is equal to the test value. As d deviates
from 0, we interpret the effect size to be stronger. What is a small versus a
large d is dependent on the area of investigation. However, d values of .2, .5,
and .8, regardless of sign, are by convention interpreted as small, medium,
and large effect sizes, respectively.
11
One sample t-test
An APA Results Section
A one-sample t test was conducted to evaluate whether the mean of the
writing scores was significantly different from 50, the accepted mean. The
sample mean of 52.78 ( SD = 9.48) was significantly different from 50, t
(199) = 4.14, p < .001. The 95% confidence interval for the writing scores
mean ranged from 51.45 to 54.10. The effect size d of .29 indicates a
medium effect.
Index End
12
One sample median test
A one sample median test allows us to test whether a sample median
differs significantly from a hypothesized value. We will use the same
variable, write, as we did in the one sample t-test example above, but
we do not need to assume that it is interval and normally distributed
(we only need to assume that write is an ordinal variable).
Menu selection:- Analyze > Nonparametric Tests > One Sample
Syntax:- nptests
/onesample test (write) wilcoxon(testvalue=50).
13
14
Choose customize analysis
15
Only retain writing score
16
Choose tests tick compare median and enter 50 as the desired

value.
Finally select the run button
17
We would conclude that this group of students has a significantly

higher median (calculated median 54) on the writing test than 50.
Index End
18
Binomial test
A one sample binomial test allows us to test whether the proportion of
successes on a two-level categorical dependent variable significantly
differs from a hypothesized value. For example, using the A data file, say
we wish to test whether the proportion of females (female) differs
significantly from 50%, i.e., from .5. We can do this as shown below.
Two alternate approaches are available.
Either
Menu selection:- Analyze > Nonparametric Tests > One Sample
Syntax:- npar tests

/binomial (.5)=female.
19
Binomial test
20
Binomial test
Choose customize analysis
21
Binomial test
Only retain female
22
Binomial test
Choose tests tick compare observed and under options
23
Binomial test
enter .5 as the desired value.

Finally select the run button
24
Binomial test
Or
Menu selection:- Analyze > Nonparametric Tests > Legacy Dialogs > Binomial
Syntax:- npar tests

/binomial (.5)=female. 25
Binomial test
Select female as the test variable, the default test proportion is .5

Finally select the OK button
26
Binomial test
Binomial Test
Observed Exact Sig.
Category N Prop. Test Prop. (2-tailed)
female Group 1 Male 91 .46 .50 .229
Group 2 Female 109 .54
Total 200 1.00
The results indicate that there is no statistically significant difference

(p= 0.229). In other words, the proportion of females in this sample
does not significantly differ from the hypothesized value of 50%.
27
Binomial test
We hypothesized that the proportion of females is 50%. A two-tailed,

binomial test was conducted to assess this research hypothesis. The
observed proportion of .46 did not differ significantly from the
hypothesized value of .50, two-tailed p = .23. Our results suggest that
the proportion of females do not differ dramatically from males.
Index End
28
Chi-square goodness of fit
A chi-square goodness of fit test allows us to test whether the observed
proportions for a categorical variable differ from hypothesized
proportions. For example, let's suppose that we believe that the general
population consists of 10% Hispanic, 10% Asian, 10% African American
and 70% White folks. We want to test whether the observed
proportions from our sample differ significantly from these
hypothesized proportions. Note this example employs input data
(10, 10, 10, 70), in addition to A.
Menu selection:- At present the drop down menus cannot provide this
analysis.
Syntax:- npar test

/chisquare=race
/expected=10 10 10 70.
29
Chi-square goodness of fit
race
Observed Expected
N N Residual These results show that racial
hispanic 24 20.0 4.0 composition in our sample does not
asian
african-
11
20
20.0
20.0
-9.0
.0
differ significantly from the
amer hypothesized values that we
white 145 140.0 5.0 supplied (chi-square with three
Total 200
degrees of freedom=5.029,
Test Statistics
p= 0.170).
race
Chi-Square 5.029a
df 3
Asymp. .170
Sig.
a. 0 cells (.0%) have expected frequencies less
than 5. The minimum expected cell frequency is
20.0.
Index End
30
Two independent samples
t-test
An independent samples t-test is used when you want to compare the
means of a normally distributed interval dependent variable for two
independent groups. For example, using the A data file, say we wish to
test whether the mean for write is the same for males and females.
Menu selection:- Analyze > Compare Means > Independent Samples T test
Syntax:- t-test groups=female(0 1)

/variables=write.
31
t-test
32
t-test
33
t-test
Do not forget to define those pesky groups.
34
Levene's test
In statistics, Levene's test is an inferential statistic used to assess the
equality of variances in different samples. Some common statistical
procedures assume that variances of the populations from which different
samples are drawn are equal. Levene's test assesses this assumption. It tests
the null hypothesis that the population variances are equal (called
homogeneity of variance or homoscedasticity). If the resulting p-value of
Levene's test is less than some critical value (typically 0.05), the obtained
differences in sample variances are unlikely to have occurred based on random
sampling from a population with equal variances. Thus, the null hypothesis of
equal variances is rejected and it is concluded that there is a difference
between the variances in the population.
Levene, Howard (1960). "Robust tests for equality of variances". In Ingram

Olkin, Harold Hotelling, et al. Stanford University Press. pp. 278292.
35
t-test
Because the standard deviations for
the two groups are not similar (10.3 and
8.1), we will use the "equal variances
not assumed" test. This is supported by
the Levenes test p = .001).
The results indicate that there is a

statistically significant difference
between the mean writing score for
males and females (t=-3.656,
p<.0005). In other words, females
have a statistically significantly higher
mean score on writing (54.99) than
males (50.12).
This is supported by the negative

confidence interval (male - female).
36
t-test
Does equality of variances matter in
this case?
37
Two independent samples t-
test - Effect Size Statistic
Eta square, 2 , may be computed . An 2 ranges in value from 0 to 1. It is
interpreted as the proportion of variance of the test variable that is a
function of the grouping variable. A value of 0 indicates that the
difference in the mean scores is equal to 0, whereas a value of 1 indicates
that the sample means differ, and the test scores do not differ within
each group (i.e., perfect replication). You can compute 2 with the following
equation: 2
2 t
2
t N1 N 2 2
What is a small versus a large 2 is dependent on the area of investigation.
However, 2 of .01, .06, and .14 are, by convention, interpreted as small,
medium, and large effect sizes, respectively.
38
Two independent samples t-
test - An APA Results Section
An independent-samples t test was conducted to evaluate the hypothesis
that the mean writing score was gender dependent. The test was
significant, t (198) = -3.656, p < .0005. Male students scored lower (M =
50.12, SD = 10.31), on average, than females (M = 54.99, SD = 8.13). The
95% confidence interval for the difference in means was quite wide,
ranging from -7.44 to -2.30. The eta square index indicated that 7% ( 2 = .
066) of the variance of the writing score was explained by gender.
Index End 39
Wilcoxon-Mann-Whitney
test
The Wilcoxon-Mann-Whitney test is a non-parametric analog to the
independent samples t-test and can be used when you do not assume that the
dependent variable is a normally distributed interval variable (you only assume
that the variable is at least ordinal). You will notice that the SPSS syntax for
the Wilcoxon-Mann-Whitney test is almost identical to that of the
independent samples t-test. We will use the same data file (the A data file)
and the same variables in this example as we did in the independent t-test
example above and will not assume that write, our dependent variable, is
normally distributed.
Menu selection:- Analyze > Nonparametric Tests

> Legacy Dialogs > 2 Independent Samples
Syntax:- npar test

/m-w=write by female(0 1).
Mann-Whitney test Crichton, N. Journal Of Clinical Nursing 9(4) 583-583 2000

40
test
The Mann-Whitney U: A Test for Assessing Whether Two Independent Sa
mples Come from the Same Distribution
Nadim Nachar
Tutorials in Quantitative Methods for Psychology 2008 4(1) 13-20
It is often difficult, particularly when conducting research in psychology,

to have access to large normally distributed samples. Fortunately, there
are statistical tests to compare two independent groups that do not
require large normally distributed samples. The Mann Whitney U is one of
these tests. In the work, a summary of this test is presented. The
explanation of the logic underlying this test and its application are also
presented. Moreover, the forces and weaknesses of the Mann Whitney
U are mentioned. One major limit of the Mann Whitney U is that the type
I error or alpha () is amplified in a situation of heteroscedasticity.
41
test
The Wilcoxon-Mann-Whitney test is sometimes used for comparing the
efficacy of two treatments in trials. It is often presented as an
alternative to a t test when the data are not normally distributed.
Where as a t test is a test of population means, the Mann-Whitney test

is commonly regarded as a test of population medians. This is not strictly
true, and treating it as such can lead to inadequate analysis of data.
Mann-Whitney test is not just a test of medians: differences in spread c

an be important
Anna Hart
British Medical Journal 2001 August 18; 323(7309): 391393.
As is always the case, it is not sufficient merely to report a p value. In

the case of the Mann-Whitney test, differences in spread may
sometimes be as important as differences in medians, and these need to
be made clear. 42
test
43
test
Note that Mann-Whitney has been selected.
44
test
45
test
Ranks
Mean Sum of
female N Rank Ranks
writing male 91 85.63 7792.00
score female 109 112.92 12308.00 The results suggest that there is a
Total 200 statistically significant difference
between the underlying distributions
Test Statisticsa
of the write scores of males and the
writing
score write scores of females (z=-3.329,
Mann-Whitney U 3606.000 p=0.001).
Wilcoxon W 7792.000
Z -3.329
Asymp. Sig. (2- .001
tailed)
a. Grouping Variable: female
46
Wilcoxon-Mann-Whitney test
- An APA Results Section
A Wilcoxon test was conducted to evaluate whether writing score was

affected by gender. The results indicated a significant difference,
z = -3.329, p = .001. The mean of the ranks for male was 85.63, while the
mean of the ranks for female was 112.92.
Index End
47
Chi-square test
(Contingency table)
A chi-square test is used when you want to see if there is a relationship
between two categorical variables. It is equivalent to the correlation
between nominal variables.
A chi-square test is a common test for nominal (categorical) data. One

application of a chi-square test is a test for independence. In this case,
the null hypothesis is that the occurrence of the outcomes for the two
groups is equal. If your data for two groups came from the same
participants (i.e. the data were paired), you should use the
McNemar's test, while for k groups you should use Cochrans Q test.
48
Chi-square test
(Contingency table)
In SPSS, the chisq option is used on the statistics subcommand of the
crosstabs command to obtain the test statistic and its associated p-
value. Using the A data file, let's see if there is a relationship between
the type of school attended (schtyp) and students' gender (female).
Remember that the chi-square test assumes that the expected value for
each cell is five or higher. This assumption is easily met in the examples
below. However, if this assumption is not met in your data, please see
the section on Fisher's exact test, below.
Two alternate approaches are available.
Either
Menu selection:- Analyze > Tables > Custom Tables
Syntax:- crosstabs
/tables=schtyp by female
/statistic=chisq phi. 49
Chi-square test
50
Chi-square test
Drag selected variables to the row/column boxes

51
Chi-square test
Select
chi-squared
Alternately
52
Chi-square test
Menu selection:- Analyze > Descriptive Statistics > Crosstabs

Syntax:- crosstabs
/tables=schtyp by female
/statistic=chisq. 53
Chi-square test
Select row and column variables.
54
Chi-square test
Select Chi-square and Cramers V
55
Chi-square test
Case Processing Summary
Cases
Valid Missing Total
N Percent N Percent N Percent
type of school * 200 100.0% 0 .0% 200 100.0%
female These results indicate
type of school * female Crosstabulation that there is no
Count
statistically significant
Female
Male female Total relationship between the
type of
school
public 77 91 168
type of school attended
private 14 18 32
Total 91 109 200 and gender (chi-square
with one degree of
Chi-Square Tests
Asymp. Sig. Exact Sig. Exact Sig. freedom=0.047,
Pearson Chi-Square
Value
.047a
Df
1
(2-sided)
.828
(2-sided) (1-sided)
p=0.828).
Continuity Correctionb .001 1 .981
Note 0 cells have

Likelihood Ratio .047 1 .828
Fisher's Exact Test .849 .492
Linear-by-Linear
Association
.047 1 .829
expected count less than
N of Valid Cases 200 5. If not use
Fisher's exact test.
a. 0 cells (.0%) have expected count less than 5. The minimum expected count is 14.56.
b. Computed only for a 2x2 table
56
Chi-square test
A two-way contingency table analysis was conducted to evaluate

whether type of school exhibited a gender bias. School and gender
were found to not be significantly related, Pearson 2 (1, N = 200) = .
047, p = .828, Cramrs V = .015. The proportions of males were .85
and females .84.
57
Chi-square test
Let's look at another example, this time looking at the relationship
between gender (female) and socio-economic status (ses). The point of
this example is that one (or both) variables may have more than two
levels, and that the variables do not have to have the same number of
levels. In this example, female has two levels (male and female) and ses
has three levels (low, medium and high).
Menu selection:- Analyze > Tables > Custom Tables

Using the previous menus.
Syntax:- crosstabs
/tables=female by ses
/statistic=chisq phi.
58
Chi-square test
Cases
Valid Missing Total
N Percent N Percent N Percent Again we find that there is no
female * ses 200 100.0% 0 .0% 200 100.0%
female * ses Crosstabulation relationship between the
Count
ses
variables (chi-square with two
low middle high Total degrees of freedom=4.577,
female male
female
15
32
47
48
29
29
91
109
p=0.101).
Total 47 95 58 200
Chi-Square Tests
Note the absence of Fishers
Asymp. Sig. Exact Test!
Value df (2-sided)
Pearson Chi-Square 4.577a 2 .101
Likelihood Ratio 4.679 2 .096
Linear-by-Linear 3.110 1 .078
Association
N of Valid Cases 200
a. 0 cells (.0%) have expected count less than 5. The
minimum expected count is 21.39.
59
Chi-square test
A two-way contingency table analysis was conducted to evaluate

whether gender was related to social economic status (SES).
Gender and SES were not found to be significantly related,
Pearson 2 (2, N = 200) = 4.58, p = .101, Cramrs V = .151. The
proportions of males in low, middle and high SES were .32, .50, and
.50, respectively.
Index End
60
Phi coefficient
The measure of association, phi, is a measure which adjusts the chi
square statistic by the sample size. The phi coefficient is the equivalent
of the correlation between nominal variables.
It may be introduced at the same time as the Chi-square.
Select Phi
61
Phi coefficient
The p values for the tests are
identical.
2 = 4.577
2 4.577
= 0.151 0.151
n 200
n = 200
Sometimes 2 is used as a measure of

association.
Index End
62
Fisher's exact test
The Fisher's exact test is used when you want to conduct a chi-square
test but one or more of your cells has an expected frequency of five or
less. Remember that the chi-square test assumes that each cell has an
expected frequency of five or more, but the Fisher's exact test has no
such assumption and can be used regardless of how small the expected
frequency is. In SPSS you can only perform a Fisher's exact test on a
2x2 table, and these results are presented by default. Please see the
results from the chi-square example above.
63
Fisher's exact test
A simple web search should reveal specific tools developed for different
size tables. For example
Fisher's exact test for up to 66 tables

For the more adventurous
For those interested in more detail, plus a worked example see.
Fisher's Exact Test or Paper only
When to Use Fisher's Exact Test

Keith M. Bower
American Society for Quality, Six Sigma Forum Magazine, 2(4) 2003, 35-37.
64
Fisher's exact test
For larger examples you might try
Fisher's Exact Test
Algorithm 643
FEXACT - A Fortran Subroutine For Fishers Exact Test On Unordered

R x C Contingency-Tables
Mehta, C.R. and Patel, N.R.
ACM Transactions On Mathematical Software 12(2) 154-161 1986.
A Remark On Algorithm-643 - FEXACT - An Algorithm For

Performing Fishers Exact Test In R x C Contingency-Tables
Clarkson, D.B., Fan, Y.A. and Joe, H.
ACM Transactions On Mathematical Software 19(4) 484-488 1993.
Index End
65
One-way ANOVA
A one-way analysis of variance (ANOVA) is used when you have a
categorical independent variable (with two or more categories) and a
normally distributed interval dependent variable and you wish to test for
differences in the means of the dependent variable broken down by the
levels of the independent variable. For example, using the A data file, say
we wish to test whether the mean of write differs between the three
program types (prog). The command for this test would be:
Menu selection:- Analyze > Compare Means > One-way ANOVA
Syntax:- oneway write by prog.
Information point: Analysis of variance (ANOVA)

Crichton, N.
Journal Of Clinical Nursing 9(3) 380-380 2000
66
One-way ANOVA
67
One-way ANOVA
68
One-way ANOVA
ANOVA
writing score
Sum of Mean
Squares df Square F Sig.
Between 3175.698 2 1587.849 21.275 .000
Groups
Within Groups 14703.177 197 74.635
Total 17878.875 199
The mean of the dependent variable differs significantly among the

levels of program type. However, we do not know if the difference is
between only two of the levels or all three of the levels.
69
One-way ANOVA
To see the mean of write for each level of program type,
Menu selection:- Analyze > Compare Means > Means
Syntax:- means tables=write by prog.
70
One-way ANOVA
71
One-way ANOVA
72
One-way ANOVA Case Processing Summary
Cases
Included Excluded Total
N Percent N Percent N Percent
writing score * type of 200 100.0% 0 .0% 200 100.0%
program
Report
writing score
type of Std.
program Mean N Deviation
general 51.3333 45 9.39778
academic 56.2571 105 7.94334
vocation 46.7600 50 9.31875
Total 52.7750 200 9.47859
From this we can see that the students in the academic program have
the highest mean writing score, while students in the vocational
program have the lowest. For a more detailed analysis refer to
Bonferroni for pairwise comparisons .
73
One-way ANOVA
For 2 need to run a general linear model, an alternate approach
unianova write by prog

/method=sstype(3)
/intercept=include
/print=etasq
/criteria=alpha(.05)
/design=prog.
Analyze > General Linear Model > Univariate
74
One-way ANOVA
75
One-way ANOVA
76
One-way ANOVA
A one-way analysis of variance was conducted to evaluate the relationship

between writing score and the type of program. The independent variable,
the type of program, included three levels, general, academic and vocation.
The dependent variable was the writing score. The ANOVA was significant
at the .05 level, F (2, 197) = 21.28, p < .0005. The strength of relationship
between the writing score and the type of program, as assessed by
2(called partial eta squared), was strong, with the writing score factor
accounting for 18% of the variance of the dependent variable.
Index End 77
Kruskal Wallis test
The Kruskal Wallis test is used when you have one independent
variable with two or more levels and an ordinal dependent variable.
In other words, it is the non-parametric version of ANOVA and a
generalized form of the Mann-Whitney test method since it permits
two or more groups. We will use the same data file as the one way
ANOVA example above (the A data file) and the same variables as in
the example above, but we will not assume that write is a normally
distributed interval variable.

> Legacy Dialogs > k Independent Samples
Syntax:- npar tests

/k-w=write by prog (1,3).
78
Kruskal Wallis test
79
Kruskal Wallis test
80
Kruskal Wallis test
81
Kruskal Wallis test
Ranks
type of Mean
program N Rank
writing general 45 90.64
score academic 105 121.56 If some of the scores receive tied
vocation 50 65.14 ranks, then a correction factor is
Total 200
used, yielding a slightly different
value of chi-squared. With or without
Test Statisticsa,b
writing
ties, the results indicate that there
score is a statistically significant
Chi-Square 34.045 difference (p < .0005) among the
df
Asymp.
2
.000
three type of programs.
Sig.
a. Kruskal Wallis Test
b. Grouping Variable: type of
program
82
Kruskal Wallis test
A Kruskal-Wallis test was conducted to evaluate differences among the

three types of program (general, academic and vocation) on median change
in the writing score). The test, which was corrected for tied ranks, was
significant, 2 (2, N = 200) = 34.045, p < .001. The proportion of variability
in the ranked dependent variable accounted for by the type of program
variable was .17 (2/(N-1)), indicating a fairly strong relationship between
type of program and writing score.
Index End
83
Paired t-test
A paired (samples) t-test is used when you have two related
observations (i.e., two observations per subject) and you want to see if
the means on these two normally distributed interval variables differ
from one another. For example, using the A data file we will test
whether the mean of read is equal to the mean of write.
Menu selection:- Analyze > Compare Means > Paired-Samples T test
Syntax:- t-test pairs=read with write (paired).
84
Paired t-test
85
Paired t-test
86
Paired Samples Statistics
Paired t-test
Std. Std. Error
These results indicate that the mean of

Mean N Deviation Mean
Pair 1 reading score 52.2300 200 10.25294 .72499
writing score 52.7750 200 9.47859 .67024
read is not statistically significantly
Paired Samples Correlations
Correlatio
different from the mean of write
Pair 1 reading score &
N
200
n
.597
Sig.
.000
(t=-0.867, p=0.387).
writing score
Paired Samples Test The confidence interval includes the origin

(no difference).
Paired Differences
Std. Std. Error
Mean Deviation Mean
Pair 1 reading score - writing -.54500 8.88667 .62838
score
Paired Samples Test

Paired Differences
95% Confidence Interval of
the Difference
Lower Upper t
Pair 1 reading score - writing -1.78414 .69414 -.867
score
Paired Samples Test

Sig. (2-
df tailed)
Pair 1 reading score - writing 199 .387
score
87
Paired t-test
A paired-samples t test was conducted to evaluate whether reading and
writing scores were related. The results indicated that the mean score
for writing (M = 52.78, SD = 9.48) was not significantly greater than the
mean score fro reading ( M = 52.23, SD = 10.25), t (199) = -.87, p = ..39.
The standardized effect size index, d , was .06. The 95% confidence
interval for the mean difference between the two ratings was -1.78 to .
69.
t
Recall d
N
Index End
88
Wilcoxon signed rank sum
test
The Wilcoxon signed rank sum test is the non-parametric version of a
paired samples t-test. You use the Wilcoxon signed rank sum test when
you do not wish to assume that the difference between the two
variables is interval and normally distributed (but you do assume the
difference is ordinal). We will use the same example as above, but we will
not assume that the difference between read and write is interval and

> Legacy Dialogs > 2 Related Samples
Syntax:- npar test

/wilcoxon=write with read (paired).
Wilcoxon signed rank test

Crichton, N.
Journal Of Clinical Nursing 9(4) 584-584 2000 89
test
90
test
Select Wilcoxon
91
test
Ranks
Mean Sum of
N Rank Ranks The results suggest that
reading score - writing Negative 97a 95.47 9261.00 there is not a
score Ranks
Positive Ranks 88b 90.27 7944.00
Ties 15 c difference (p= 0.366)
Total 200 between read and write.
a. reading score < writing score
b. reading score > writing score
c. reading score = writing score
Test Statisticsb
reading score
- writing
score
Z -.903a
tailed)
a. Based on positive ranks.
b. Wilcoxon Signed Ranks Test
Index End
92
Sign test
If you believe the differences between read and write were not ordinal
but could merely be classified as positive and negative, then you may
want to consider a sign test in lieu of sign rank test. The Sign test
answers the question How Often?, whereas other tests answer the
question How Much?. Again, we will use the same variables in this
example and assume that this difference is not ordinal.

> Legacy Dialogs > 2 Related Samples
Syntax:- npar test

/sign=read with write (paired).
93
Sign test
94
Sign test
Select Sign
95
Sign test
Frequencies
N We conclude that no statistically
writing score - reading Negative 88 significant difference was found
score Differencesa (p = 0.556).
Positive 97
Differencesb
Tiesc 15
Total 200
a. writing score < reading score
b. writing score > reading score
c. writing score = reading score
Test Statisticsa
writing score
- reading
score
Z -.588
tailed)
a. Sign Test
96
Sign test
A Wilcoxon signed ranks test was conducted to evaluate whether
reading and writing scores differed. The results indicated a non-
significant difference, z = -.59, p = .56. The mean of the negative
ranks were 95.47 and the positive were 90.27.
Index End
97
McNemar test
McNemar's test is a statistical test used on paired nominal data. It is applied to
22 contingency tables with a dichotomous trait, with matched pairs of
subjects, to determine whether the row and column marginal frequencies are
equal (that is, whether there is marginal homogeneity). For k groups use
Cochrans Q test.
You would perform McNemar's test if you were interested in the marginal
frequencies of two binary outcomes. These binary outcomes may be the same
outcome variable on matched pairs (like a case-control study) or two outcome
variables from a single group. Continuing with the A dataset used in several
above examples, let us create two binary outcomes in our dataset: himath and
hiread. These outcomes can be considered in a two-way contingency table.
The null hypothesis is that the proportion of students in the himath group is
the same as the proportion of students in hiread group (i.e., that the
contingency table is symmetric).
Menu selection:- Transform > Compute Variable

Analyze > Descriptive Statistics > Crosstabs
The syntax is on the next slide. 98
McNemar test
Syntax:- COMPUTE himath=math>60.
COMPUTE hiread=read>60.
EXECUTE.
CROSSTABS
/TABLES=himath BY hiread
/STATISTICS=MCNEMAR
/CELLS=COUNT.
99
McNemar test
First the transformation
100
McNemar test
Which is utilised twice, for math and read

101
McNemar test
102
McNemar test
Now the test
103
McNemar test
Select McNemar
104
McNemar test
McNemar's chi-square
Cases
Valid Missing Total statistic suggests that
N Percent N Percent N Percent there is not a statistically
himath *
hiread
200 100.0% 0 .0% 200 100.0%
significant difference in
the proportion of students
himath * hiread Crosstabulation in the himath group and
Count
hiread the proportion of students
.00 1.00 Total in the hiread group.
himath .00 135 21 156
1.00 18 26 44
Total 153 47 200 Alternately accessing the
command directly.
Chi-Square Tests
Exact Sig.
Value (2-sided)
McNemar Test .749a
N of Valid 200
Cases
a. Binomial distribution used.
105
McNemar test
Menu selection:- Analyze > Nonparametric Tests > Legacy Dialogs
> 2 Related Samples
Syntax:- NPAR TESTS

/MCNEMAR=himath WITH hiread (PAIRED)
/MISSING ANALYSIS
/METHOD=EXACT TIMER(5).
106
McNemar test
107
McNemar test
Select McNemar and Exact
108
McNemar test
McNemar's chi-square
statistic suggests that there
is not a statistically
significant difference in the
proportion of students in the
himath group and the
hiread group.
109
McNemar test
Proportions of student scoring high in math and reading were .22
and .24, respectively. A McNemar test, which evaluates
differences among related proportions, was not significant, 2 (1,
N = 200) = .10, p = .75.
Index End
110
Cochrans Q test
In the analysis of two-way randomized block designs where the response
variable can take only two possible outcomes (coded as 0 and 1), Cochran's Q
test is a non-parametric statistical test to verify whether k treatments have
identical effects. Your data for the k groups came from the same participants
(i.e. the data were paired).
You would perform Cochrans Q test if you were interested in the marginal
frequencies of three or more binary outcomes. Continuing with the A dataset
used in several above examples, let us create three binary outcomes in our
dataset: himath, hiread and hiwrite. The null hypothesis is that the proportion
of students in each group is the same.
Menu selection:- Transform > Compute Variable
Analyze > Nonparametric Tests

> Legacy Dialogs > K Related Samples
The syntax is on the next slide. 111
Cochrans Q test
Syntax:- COMPUTE himath=math>60.
COMPUTE hiread=read>60.
COMPUTE hiwrite=write>60.
EXECUTE.
NPAR TESTS
/COCHRAN=himath hiread hiwrite
/MISSING LISTWISE
/METHOD=EXACT TIMER(5).
112
Cochrans Q test
First transform
113
Cochrans Q test
Which is utilised three times, for math, read and write.

114
Cochrans Q test
115
Cochrans Q test
Select Friedman, Kendalls W and Cochrans Q also Exact
116
Cochrans Q test
Cochrans Q statistic (which is a chi-

squared statistic) suggests that
there is not a statistically
significant difference in the
himath, hiread and hiwrite groups.
117
Cochrans Q test
Necessary for summary
118
Cochrans Q test
Proportions of student scoring high in math, reading and writing were .
22, .24, and .25, respectively. A Cochran test, which evaluates
differences among related proportions, was not significant, 2 (2, N =
200) = .60, p = .74. The Kendall coefficient of concordance was .002.
Index End
119
Cochrans Q test
When you find any significant effect, you need to do a post-hoc test (as
you do for ANOVA). For Cochran's Q test: run multiple McNemar's
tests and adjust the p values with Bonferroni correction (a method used
to address the problem of multiple comparisons, over corrects for Type
I error).
Cochran, W.G. (1950).

The Comparison of Percentages in Matched Samples Biometrika, 37,
256-266.
Index End
120
About the B data file
We have an example data set called B, which is used in Roger E.
Kirk's book Experimental Design: Procedures for Behavioral
Sciences (Psychology) (ISBN 0534250920).
Syntax:- display dictionary

/VARIABLES s y1 y2 y3 y4.
Variable Position Measurement Level

s 1 Ordinal
y1 2 Scale
y2 3 Scale
y3 4 Scale
y4 5 Scale
121
About the B data file
Index End
122
One-way repeated
measures ANOVA
You would perform a one-way repeated measures analysis of variance if you
had one categorical independent variable and a normally distributed interval
dependent variable that was repeated at least twice for each subject. This
is the equivalent of the paired samples t-test, but allows for two or more
levels of the categorical variable. This tests whether the mean of the
dependent variable differs by the categorical variable. In data set B, y (y1
y2 y3 y4) is the dependent variable, a is the repeated measure (a name you
assign) and s is the variable that indicates the subject number.
Menu selection:- Analyze > General Linear Model > Repeated Measures
Syntax:- glm y1 y2 y3 y4
/wsfactor a(4).
123
One-way repeated
measures ANOVA
124
One-way repeated
measures ANOVA
You chose the factor name a which you then Add.
125
One-way repeated
measures ANOVA
You chose the factor name a which you then Add.
126
One-way repeated
measures ANOVA
Finally
127
One-way repeated
measures ANOVA
Tests of Within-Subjects Effects
Within-Subjects
Loads of output!!
Measure:MEASURE_1
Factors Type III Sum Mean
Measure:MEASURE Source of Squares df Square F Sig.
_1
A Sphericity 49.000 3 16.333 11.627 .000
Dependent Assumed
a Variable
Greenhouse- 49.000 1.859 26.365 11.627 .001
1 y1 Geisser
2 y2
Huynh-Feldt 49.000 2.503 19.578 11.627 .000
3 y3
Lower-bound 49.000 1.000 49.000 11.627 .011
4 y4
Error(a) Sphericity 29.500 21 1.405
Assumed
Multivariate Testsb Greenhouse- 29.500 13.010 2.268
Hypothesis Geisser
Effect Value F df Error df Sig. Huynh-Feldt 29.500 17.520 1.684
a Pillai's Trace .754 5.114a 3.000 5.000 .055 Lower-bound 29.500 7.000 4.214
a
Wilks' Lambda .246 5.114 3.000 5.000 .055
Hotelling's Trace 3.068 5.114a 3.000 5.000 .055
Tests of Within-Subjects Contrasts
Roy's Largest 3.068 5.114a 3.000 5.000 .055 Measure:MEASURE_1
Root
Type III Sum Mean
a. Exact statistic Source a of Squares df Square F Sig.
b. Design: Intercept A Linear 44.100 1 44.100 19.294 .003
Within Subjects Design: a
Quadrati 4.500 1 4.500 3.150 .119
c
Mauchly's Test of Sphericityb Cubic .400 1 .400 .800 .401
Measure:MEASURE_1
Error(a) Linear 16.000 7 2.286
Within Subjects Mauchly's Approx. Chi-
Quadrati 10.000 7 1.429
Effect W Square df Sig.
c
a .339 6.187 5 .295
Cubic 3.500 7 .500
Mauchly's Test of Sphericityb

Measure:MEASURE_1 Tests of Between-Subjects Effects
Epsilona Measure:MEASURE_1
Transformed Variable:Average
Within Subjects Greenhouse- Huynh- Lower-
Effect Type III Sum Mean
Geisser Feldt bound
Source of Squares df Square F Sig.
A .620 .834 .333
Intercep 578.000 1 578.000 128.444 .000
Tests the null hypothesis that the error covariance matrix of the
t
orthonormalized transformed dependent variables is
Error 31.500 7 4.500
proportional to an identity matrix.
a. May be used to adjust the degrees of freedom for the
averaged tests of significance. Corrected tests are displayed in
the Tests of Within-Subjects Effects table.
128
One-way repeated
measures ANOVA
Tests of Within-Subjects Effects
Measure:MEASURE_1 You will notice that this output
Type III Sum Mean
gives four different
Source of Squares df Square F Sig.
A Sphericity 49.000 3 16.333 11.627 .000 p-values. The output labelled
Assumed sphericity assumed is the p-
Greenhouse- 49.000 1.859 26.365 11.627 .001 value (<0.0005) that you would get
Geisser
Huynh-Feldt 49.000 2.503 19.578 11.627 .000 if you assumed compound
Lower-bound 49.000 1.000 49.000 11.627 .011 symmetry in the variance-
Error(a) Sphericity 29.500 21 1.405 covariance matrix. Because that
Assumed
Greenhouse- 29.500 13.010 2.268
assumption is often not valid, the
Geisser three other p-values offer
Huynh-Feldt 29.500 17.520 1.684 various corrections (the Huynh-
Lower-bound 29.500 7.000 4.214
Feldt, H-F, Greenhouse-Geisser,
G-G and Lower-bound). No matter
which p-value you use, our results
indicate that we have a
statistically significant effect of
a at the .05 level.
129
One-way repeated
measures ANOVA
A one-way within-subjects ANOVA was conducted with the factor being the
number of years married and the dependent variable being y. The means and
standard deviations for scores are presented above . The results for the ANOVA
indicated a significant time effect, Wilkss = .25, F (3, 21) = 11.63, p < .0005,
multivariate 2 = .75 (1-).
Index End
130
Bonferroni for pairwise
comparisons
This is a minor extension of the
previous analysis.
Menu selection:-
Analyze
> General Linear Model
> Repeated Measures
Syntax:-
GLM y1 y2 y3 y4
/WSFACTOR=a 4 Polynomial
/METHOD=SSTYPE(3)
Only the additional outputs are

presented. 131
comparisons
This table simply provides

important descriptive statistics
for the analysis as shown below.
132
comparisons
Using post hoc tests to examine

whether estimated marginal
means differ for levels of
specific factors in the model.
133
comparisons
The results presented in the
previous Tests of Within-Subjects
Effects table, the Huynh-Feldt
(p < .0005) informed us that we
have an overall significant
difference in means, but we do not
know where those differences
occurred.
This table presents the results of

the Bonferroni post-hoc test, which
allows us to discover which specific
means differed.
Remember, if your overall ANOVA

result was not significant, you
should not examine the Pairwise
Comparisons table. 134
comparisons
We can see that there was a
significant difference between 1
and 4 (p = 0.017), while 2 and 4
merit further consideration.
In true SPSS style the results are

duplicated.
135
comparisons
The table provides four variants of the F test. Wilks' lambda is the most
commonly reported. Usually the same substantive conclusion emerges from
any variant. For these data, we conclude that none of effects are
significant (p = 0.055). See next slide.
136
comparisons
Wilks lambda is the easiest to understand and therefore the most frequently used. It
has a good balance between power and assumptions. Wilks lambda can be interpreted as
the multivariate counterpart of a univariate R-squared, that is, it indicates the
proportion of generalized variance in the dependent variables that is accounted for by
the predictors.
Correct Use of Repeated Measures Analysis of Variance E. Park, M. Cho and C.-S. Ki. Korean J. Lab. Med. 2009 29 1-9
Wilks' lambda performs, in the multivariate setting, with a combination of

dependent variables, the same role as the F-test performs in one-way analysis of
variance. Wilks' lambda is a direct measure of the proportion of variance in the
combination of dependent variables that is unaccounted for by the independent variable
(the grouping variable or factor). If a large proportion of the variance is accounted for
by the independent variable then it suggests that there is an effect from the grouping
variable and that the groups have different mean values.
Information Point: Wilks' lambda Nicola Crichton Journal of Clinical Nursing, 9, 381-381, 2000.
Index End 137

About the C data file
The C data set contains 3 pulse measurements from each of 30 people

assigned to 2 different diet regiments and 3 different exercise
regiments.
Syntax:- display dictionary

/VARIABLES id diet exertype pulse time highpulse.
Variable Position
id 1
diet 2
exertype 3
pulse 4
time 5
highpulse 6
138
About the C data file
Index End 139

Repeated measures logistic
regression
If you have a binary outcome measured repeatedly for each subject and
you wish to run a logistic regression that accounts for the effect of
multiple measures from single subjects, you can perform a repeated
measures logistic regression. In SPSS, this can be done using the
GENLIN command and indicating binomial as the probability distribution
and logit as the link function to be used in the model. In C, if we define a
"high" pulse as being over 100, we can then predict the probability of a
high pulse using diet regime.
Menu selection:- Analyze > Generalized Estimating Equations
However see the next slide.
140
regression
While the drop down menus can be employed to set the arguments it is
simpler to employ the syntax window.
Syntax:- GENLIN highpulse (REFERENCE=LAST)

BY diet (order=DESCENDING)
/MODEL diet
DISTRIBUTION=BINOMIAL
LINK=LOGIT
/REPEATED SUBJECT=id CORRTYPE=EXCHANGEABLE.
For completeness the drop down menu saga is shown, some 9 slides!
141
regression
142
regression
143
regression
144
regression
145
regression
146
regression
147
regression
148
regression
149
regression
150
regression
Goodness of Fitb
Model Information
Value
Dependent Variable highpulsea
Quasi Likelihood under 113.986
Probability Distribution Binomial
Independence Model
Link Function Logit
Criterion (QIC)a
Subject 1 id
Corrected Quasi Likelihood 111.340
Effect
under Independence Model
Working Correlation Matrix Structure Exchangeable
Criterion (QICC)a
a. The procedure models .00 as the response, treating 1.00 as the
Dependent Variable: highpulse
reference category.
Model: (Intercept), diet
a. Computed using the full log quasi-
Case Processing Summary likelihood function.
N Percent b. Information criteria are in small-is-
Included 90 100.0% better form.
Exclude 0 .0%
d Tests of Model Effects
Total 90 100.0% Type III
Loads of output!!
Wald Chi-
Correlated Data Summary Source Square df Sig.
Number of Levels Subject id 30 (Intercept) 8.437 1 .004
Effect diet 1.562 1 .211
Number of Subjects 30 Dependent Variable: highpulse
Number of Minimum 3 Model: (Intercept), diet
Measurements per Maximum 3
Subject
Parameter Estimates
Correlation Matrix Dimension 3
95% Wald Confidence Interval
Parameter B Std. Error Lower Upper
Categorical Variable Information
(Intercept) 1.253 .4328 .404 2.101
N Percent [diet=2.00] -.754 .6031 -1.936 .428
Dependent highpulse .00 63 70.0% [diet=1.00] 0a . . .
Variable 1.00 27 30.0% (Scale) 1
Total 90 100.0%
Factor diet 2.00 45 50.0%
1.00 45 50.0%
Total 90 100.0%
151
regression
Parameter Estimates
Hypothesis Test
Wald
Chi-
Parameter Square df Sig.
(Intercept) 8.377 1 .004
[diet=2.00] 1.562 1 .211
[diet=1.00] . . .
(Scale)
Dependent Variable: highpulse
Model: (Intercept), diet
a. Set to zero because this parameter is
redundant.
These results indicate that diet is not statistically significant (Wald

Chi-Square=1.562, p=0.211).
Index End
152
Factorial ANOVA
A factorial ANOVA has two or more categorical independent variables
(either with or without the interactions) and a single normally distributed
interval dependent variable. For example, using the A data file we will look
at writing scores (write) as the dependent variable and gender (female)
and socio-economic status (ses) as independent variables, and we will
include an interaction of female by ses. Note that in SPSS, you do not
need to have the interaction term(s) in your data set. Rather, you can
have SPSS create it/them temporarily by placing an asterisk between the
variables that will make up the interaction term(s). For the approach
adopted here, this step is automatic. However see the syntax example
below.
Menu selection:- Analyze > General Linear Model > Univariate
Syntax:- glm write by female ses.
153
Factorial ANOVA
Alternate
Syntax:- UNIANOVA write BY female ses
/METHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/CRITERIA=ALPHA(0.05)
/DESIGN=female ses female*ses.
Note the interaction term, female*ses.
154
Factorial ANOVA
155
Factorial ANOVA
156
Factorial ANOVA
These results indicate that
the overall model is
(F=5.666, p<0.0005). The
variables female and ses are
also statistically significant
(F=16.595, p<0.0005 and
F=6.611, p=0.002,
respectively). However, note
that interaction between
female and ses is not
(F=0.133, p=0.875).
Index End
157
Friedman test
You perform a Friedman test when you have one within-subjects
independent variable with two or more levels and a dependent variable
that is not interval and normally distributed (but at least ordinal). We will
use this test to determine if there is a difference in the reading, writing
and math scores. The null hypothesis in this test is that the distribution
of the ranks of each type of score (i.e., reading, writing and math) are the
same. To conduct a Friedman test, the data need to be in a long format (
see the next topic).

> Legacy Dialogs > K Related Samples
Syntax:- npar tests

/friedman = read write math.
158
Friedman test
159
Friedman test
160
Friedman test
Ranks
Mean Friedman's chi-square has a value of 0.645 and a
Rank
reading 1.96
p-value of 0.724 and is not statistically
score significant. Hence, there is no evidence that the
writing score 2.04 distributions of the three types of scores are
math score 2.01
different.
Test Statisticsa
N 200
Chi- .645
Square
df 2
Asymp. .724
Sig.
a. Friedman Test
Index End
161
Reshaping data
This example illustrates a wide data file and reshapes it into long form.
Consider the data containing the kids and their heights at one year of age
(ht1) and at two years of age (ht2).
FAMID BIRTH HT1 HT2
1.00 1.00 2.80 3.40

1.00 2.00 2.90 3.80
1.00 3.00 2.20 2.90
2.00 1.00 2.00 3.20
2.00 2.00 1.80 2.80
2.00 3.00 1.90 2.40
3.00 1.00 2.20 3.30
3.00 2.00 2.30 3.40
3.00 3.00 2.10 2.90
Number of cases read: 9 Number of cases listed: 9
This is called a wide format since the heights are wide. We may want the
data to be long, where each height is in a separate observation.
162
Reshaping data
FAMID BIRTH AGE HT
We may want the data to be long, where
1.00 1.00 1.00 2.80
1.00 1.00 2.00 3.40 each height is in a separate observation.
1.00 2.00 1.00 2.90
1.00 2.00 2.00 3.80
Data may be restructured using the point

1.00 3.00 1.00 2.20
1.00 3.00 2.00 2.90
2.00
2.00
1.00
1.00
1.00
2.00
2.00
3.20 and click function in SPSS, or pre-
2.00
2.00
2.00
2.00
1.00
2.00
1.80
2.80 processing with Excel.
2.00 3.00 1.00 1.90
2.00 3.00 2.00 2.40
3.00 1.00 1.00 2.20
3.00 1.00 2.00 3.30
3.00 2.00 1.00 2.30
3.00 2.00 2.00 3.40
3.00 3.00 1.00 2.10
3.00 3.00 2.00 2.90
Number of cases read: 18 Number of cases listed: 18
Index End
163
Ordered logistic regression
Ordered logistic regression is used when the dependent variable is
ordered, but not continuous. For example, using the A data file we will
create an ordered variable called write3. This variable will have the
values 1, 2 and 3, indicating a low, medium or high writing score. We do
not generally recommend categorizing a continuous variable in this way;
we are simply creating a variable to use for this example.
Menu selection:- Transform > Recode into Different Variables
Syntax:- if write ge 30 and write le 48 write3 = 1.

if write ge 49 and write le 57 write3 = 2.
if write ge 58 and write le 70 write3 = 3.
execute.
164
First compute the new variable.
165
Choose a sensible name.
166
Add to create rules and finally Change
167
finally continue
168
use change to execute
169
We will use gender (female), reading score (read) and social studies score
(socst) as predictor variables in this model. We will use a logit link and on
the print subcommand we have requested the parameter estimates, the
(model) summary statistics and the test of the parallel lines assumption.
Menu selection:- Analyze > Regression > Ordinal
Syntax:- plum write3 with female read socst

/link = logit
/print = parameter summary tparallel.
170
171
172
173
The results indicate that the overall model is statistically
significant (p < .0005), as are each of the predictor
variables (p < .0005). There are two thresholds for this
model because there are three levels of the outcome
variable. We also see that the test of the proportional
odds assumption is non-significant (p= 0.563). One of the
assumptions underlying ordinal logistic (and ordinal probit)
regression is that the relationship between each pair of
outcome groups is the same. In other words, ordinal logistic
regression assumes that the coefficients that describe the
relationship between, say, the lowest versus all higher
categories of the response variable are the same as those
that describe the relationship between the next lowest
category and all higher categories, etc. This is called the
proportional odds assumption or the parallel regression
assumption. Because the relationship between all pairs of
groups is the same, there is only one set of coefficients
(only one model). If this was not the case, we would need
different models (such as a generalized ordered logit
model) to describe the relationship between each pair of
outcome groups.
Index End
174
Factorial logistic regression
A factorial logistic regression is used when you have two or more categorical
independent variables but a dichotomous dependent variable. For example,
using the A data file we will use female as our dependent variable, because it
is the only dichotomous variable in our data set; certainly not because it is
common practice to use gender as an outcome variable. We will use type of
program (prog) and school type (schtyp) as our predictor variables. Because
prog is a categorical variable (it has three levels), we need to create dummy
codes for it. SPSS will do this for you by making dummy codes for all
variables listed after the keyword with. SPSS will also create the interaction
term; simply list the two variables that will make up the interaction
separated by the keyword by.
Menu selection:- Analyze > Regression > Binary Logistic
Simplest to realise via the syntax window.
Syntax:- logistic regression female with prog schtyp prog by schtyp

/contrast(prog) = indicator(1). 175
176
Note that the identification of prog as the categorical variable is made

below.
177
Use Ctrl with left mouse key to select two variables then >a*b> for the
product term.
178
Define categorical variables
179
Indicator(1) identifies value 1 as the (first) reference category
180
Loads of output!!
181
The results indicate that the overall model is not statistically

significant (Likelihood ratio Chi2=3.147, p=0.677). Furthermore, none
of the coefficients are statistically significant either. This shows that
the overall effect of prog is not significant.
Index End
182
Correlation
A correlation (Pearson correlation) is useful when you want to see the
relationship between two (or more) normally distributed interval variables.
For example, using the A data file we can run a correlation between two
continuous variables, read and write.
Menu selection:- Analyze > Correlate > Bivariate
Syntax:- correlations
/variables = read write.
Understanding and Interpreting Correlations - an Interactive Visualization
183
Correlation
184
Correlation
Select Pearson
185
Correlation
In the first example above, we see that the correlation between read
and write is 0.597. By squaring the correlation and then multiplying by
100, you can determine what percentage of the variability is shared,
0.597 when squared is .356409, multiplied by 100 would be 36%. Hence
read shares about 36% of its variability with write.
No need for diagonal entries, nor to duplicate the results!
186
Correlation
As a rule of thumb use the following guide for the absolute value of correlation (r):
.00-.19 very weak
.20-.39 weak
.40-.59 moderate
.60-.79 strong
.80-1.0 very strong
Which is based on the coefficient of

determination (r2). Which indicates the
proportion of variance in each of two
correlated variables which is shared by
both.
An index of the degree of lack of

relationship is also available. It is the
square root of the proportion of
unexplained variance and is called the
coefficient of alienation (1-r2). This in
turn leads to an estimate of error
reduction 1-(1-r2).
A Graphic and Tabular Aid To Interpreting Correlation Coefficients J.F. Voorhees Monthly Weather Review 54 423
1871926.
Correlation
In the second example, we will run a correlation between a dichotomous
variable, female, and a continuous variable, write. Although it is assumed
that the variables are interval and normally distributed, we can include
dummy variables when performing correlations.
Syntax:- correlations
/variables = female write.
188
Correlation
In the output for the second example, we can see the correlation
between write and female is 0.256. Squaring this number yields .
065536, meaning that female shares approximately 6.5% of its
variability with write.
189
Correlation
An APA Results Section For
A Table Of Correlations
Correlation coefficients were computed among the five self-concept
scales. Using the Bonferroni approach to control for Type I error
across the 10 correlations, a p value of less than .005 (.05/10 = .005)
was required for significance. The results of the correlational analyses
presented in the table show that 7 out of the 10 correlations were
statistically significant and were greater than or equal to .35 (the
critical value at p = .005 and N-2 degrees of freedom).
Index End
190
Simple linear regression
Simple linear regression allows us to look at the linear relationship
between one normally distributed interval predictor and one normally
distributed interval outcome variable. For example, using the A data file,
say we wish to look at the relationship between writing scores (write) and
reading scores (read); in other words, predicting write from read.
Menu selection:- Analyze > Regression > Linear Regression
Syntax:- regression
/missing listwise
/statistics coeff outs ci(95) r anova
/criteria=pin(.05) pout(.10)
/noorigin
/dependent write
/method=enter read.
Regression analysis Crichton, N. Journal Of Clinical Nursing 10(4) 462-462 2001
Information point: coefficient of determination, R2 Crichton, N. Journal Of Clinical Nursing 8(4)

379-279 1999 191
192
193
194
We see that the relationship
between write and read is positive
(.552) and based on the t-value
(10.47) and p-value (<0.0005), we
would conclude this relationship is
statistically significant. Hence, we
would say there is a statistically
significant positive linear
relationship between reading and
writing.
Take care with in/dependent

assumptions.
195
Index End
A linear regression analysis was conducted to evaluate the prediction of

writing score from the reading score. The scatterplot for the two variables, is
shown, indicates that the two variables are linearly related such that as
reading score increases the writing score increases. The regression equation
for predicting the writing score is
Writing score = .552 Reading score + 23.959
The 95% confidence interval for the slope, .448 to .656 does not contain the
value of zero, and therefore overall strength is significantly related to the
overall injury index. The correlation between the reading and writing scores
was .60. Approximately 36% of the variance of the writing score was
accounted for by its linear relationship with the reading score. 196
Non-parametric correlation
A Spearman correlation is used when one or both of the variables are
not assumed to be normally distributed and interval (but are assumed to
be ordinal). The values of the variables are converted to ranks and then
correlated. In our example, we will look for a relationship between read
and write. We will not assume that both of these variables are normal
and interval.
Syntax:- nonpar corr

/variables = read write
/print = spearman.
197
198
Select Spearman
199
The results suggest that the relationship between read and write
(=0.617, p < 0.0005) is statistically significant.
200
Spearmans correlation works by calculating Pearsons correlation on
the ranked values of this data. Ranking (from low to high) is obtained
by assigning a rank of 1 to the lowest value, 2 to the next lowest and
so on. Thus the p value is only correct if there are no ties in the
data. In the event that ties occur an exact calculation should be
employed. SPSS does not do this. However the estimated value is
usually reliable enough.
Comparison Of Values Of Pearsons And Spearmans Correlation Coef

ficients On The Same Sets Of Data
Jan Hauke, Tomasz Kossowski

Quaestiones Geographicae 30(2) 87-93 2011
Index End
201
Simple logistic regression
Logistic regression assumes that the outcome variable is binary (i.e.,
coded as 0 and 1). We have only one variable in the A data file that is
coded 0 and 1, and that is female. We understand that female is a silly
outcome variable (it would make more sense to use it as a predictor
variable), but we can use female as the outcome variable to illustrate
how the code for this command is structured and how to interpret the
output. The first variable listed after the logistic command is the
outcome (or dependent) variable, and all of the rest of the variables
are predictor (or independent) variables. In our example, female will be
the outcome variable, and read will be the predictor variable. As with
ordinary least squares regression, the predictor variables must be
either dichotomous or continuous; they cannot be categorical.
Syntax:- logistic regression female with read.

202
203
204
Loads of output!!
205
The results indicate that reading score
(read) is not a statistically significant
predictor of gender (i.e., being
female), Wald= 0.562, p=0.453.
Likewise, the test of the overall model
is not statistically significant,
likelihood ratio Chi-squared = 0.564,
p=0.453.
Index End
206
Multiple regression
Multiple regression is very similar to simple regression, except that in
multiple regression you have more than one predictor variable in the
equation. For example, using the A data file we will predict writing score
from gender (female), reading, math, science and social studies (socst)
scores.
Menu selection:- Analyze > Regression > Linear Regression
Syntax:- regression variable = write female read math science socst

/dependent = write
/method = enter.
207
Multiple regression
208
Multiple regression
Note additional independent variables within box
209
Multiple regression
Variables Entered/Removedb
Variables Variables
Model Entered Removed Method
1 social studies . Enter
score, female,
science score,
The results indicate that the

math score,
reading score
a. All requested variables entered.
b. Dependent Variable: writing score
overall model is statistically
Model Summary
Adjusted R Std. Error of the
significant (F=58.60,
Model
1
R
.776a
R Square
.602
Square
.591
Estimate
6.05897
p<0.0005). Furthermore, all of
a. Predictors: (Constant), social studies score, female, science score,
math score, reading score the predictor variables are
ANOVAb statistically significant except
Model
1 Regression
Sum of Squares
10756.924
df
5
Mean Square
2151.385
F
58.603
Sig.
.000a for read.
Residual 7121.951 194 36.711
Total 17878.875 199
a. Predictors: (Constant), social studies score, female, science score, math score, reading score
b. Dependent Variable: writing score
Coefficientsa
Standardized
Unstandardized Coefficients Coefficients
Model B Std. Error Beta t Sig.
1 (Constant) 6.139 2.808 2.186 .030
female 5.493 .875 .289 6.274 .000
reading score .125 .065 .136 1.931 .055
math score .238 .067 .235 3.547 .000
science score .242 .061 .253 3.986 .000
social studies score .229 .053 .260 4.339 .000
a. Dependent Variable: writing score
210
Multiple regression
A multiple regression analysis was conducted to evaluate how well gender, reading and
math scores predicted writing score. The predictors were the three indices, while the
criterion variable was the writing score. The linear combination of measures was
significantly related to the score, F (5, 194) = 58.6, p <.0005. The sample multiple
correlation coefficient was .78, indicating that approximately 60% of the variance of
the writing score in the sample can be accounted for by the linear combination of the
other measures. The table presents indices to indicate the relative strength of the
variables.
211
Multiple regression -
Alternatives
There are problems with stepwise model selection procedures, these notes
are a health warning.
Various algorithms have been developed for aiding in model selection. Many
of them are automatic, in the sense that they have a stopping rule
(which it might be possible for the researcher to set or change from a
default value) based on criteria such as value of a t-statistic or an F-
statistic. Others might be better termed semi-automatic, in the sense
that they automatically list various options and values of measures that
might be used to help evaluate them.
Caution: Different regression software may use the same name (e.g.,
Forward Selection or Backward Elimination) to designate different
algorithms. Be sure to read the documentation to know find out just what
the algorithm does in the software you are using - in particular, whether it
has a stopping rule or is of the semi-automatic variety.
212
Alternatives
The reasons for not using a stepwise procedure are as follows.
There is a great deal of arbitrariness in the procedures. Forwards and

backwards stepwise methods will in general give different best
models. There are differing criteria for accepting or rejecting a
variable at any stage and also for when to stop and declare the current
model best.
The process gives a false impression of statistical sophistication. Often

a complex stepwise analysis is presented, when no proper thought has
been given to the real issues involved.
213
Alternatives
Stepwise regressions are nevertheless important for three reasons.
First, to emphasise that there is a considerable problem in choosing a
model out of so many, so considerable that a variety of automated
procedures have been devised to help. Second to show that while
purely statistical methods of choice can be constructed, they are
unsatisfactory. And third, because they are fairly popular ways of
avoiding constructive thinking about model selection, you may well come
across them. You should know that they exist and roughly how they
work.
Stepwise regressions probably do have a useful role to play, when there

are large numbers of xvariables, when all prior information is taken
carefully into account in inclusion/exclusion of variables, and when the
results are used as a preliminary sifting of the many x-variables. It
would be rare for a stepwise regression to produce convincing evidence
for or against a scientific hypothesis.
214
Alternatives
... perhaps the most serious source of error lies in letting
statistical procedures make decisions for you.
Good P.I. and Hardin J.W., Common Errors in Statistics (and How
to Avoid Them), 4th Edition, Wiley, 2012, p. 3.
Don't be too quick to turn on the computer. By passing the brain

to compute by reflex is a sure recipe for disaster.
Good P.I. and Hardin J.W., Common Errors in Statistics (and How
to Avoid Them), 4th Edition, Wiley, 2012, p. 152.
215
Alternatives
We do not recommend such stopping rules for routine use since
they can reject perfectly reasonable sub-models from further
consideration. Stepwise procedures are easy to explain,
inexpensive to compute, and widely used. The comparative
simplicity of the results from stepwise regression with model
selection rules appeals to many analysts. But, such algorithmic
model selection methods must be used with caution.
Cook R.D. and Weisberg S., Applied Regression Including

Computing and Graphics, Wiley, 1999, p. 280.
216
Alternatives
Stopping stepwise: Why stepwise and similar selection methods a
re bad, and what you should use
A common problem in regression analysis is that of variable

selection. Often, you have a large number of potential
independent variables, and wish to select among them, perhaps
to create a best model. One common method of dealing with this
problem is some form of automated procedure, such as forward,
backward, or stepwise selection. We show that these methods
are not to be recommended, and present better alternatives
using PROC GLMSELECT and other methods.
Contains useful references, is based on SAS.

217
Alternatives
In a large world where parameters need to be estimated from
small or unreliable samples, the function between predictive
accuracy and the flexibility of a model (e.g., number of free
parameters) is typically inversely U shaped. Both too few and too
many parameters can hurt performance (Pitt et al. 2002).
Competing models of strategies should be tested for their
predictive ability, not their ability to fit already known data.
Pitt M.A., Myung I.J. and Zhang S. 2002. Toward a method for
selecting among computational models for cognition Psychol.
Rev. 109 472491 DOI: 10.1037/0033-295X.109.3.472
218
Alternatives
What strategies might we adopt?
Heuristics are a subset of strategies; strategies also include

complex regression or Bayesian models. The part of the
information that is ignored is covered by Shah and
Oppenheimers (2008) list of five aspects. The goal of making
judgments more quickly and frugally is consistent with the goal
of effort reduction, where frugal is often measured by the
number of cues that a heuristic searches.
219
Alternatives
Many definitions of heuristics exist Shah and Oppenheimer
(2008) proposed that all heuristics rely on effort reduction by
one or more of the following:
(a) examining fewer cues,

(b) reducing the effort of retrieving cue values,
(c) simplifying the weighting of cues,
(d ) integrating less information,
(e) examining fewer alternatives.
Shah, A.K. and Oppenheimer, D.M. 2008 Heuristics Made Easy:

An Effort-Reduction Framework Psychological Bulletin 134(2)
207-222 DOI: 10.1037/0033-2909.134.2.207
220
Alternatives
Two alternative techniques, dominance analysis (Budescu 1993) and
relative weight analysis (Johnson 2000), have been developed that
permit more accurate partitioning of variance among correlated
predictors. Simulation research clearly shows that these measures of
importance perform quite well across a variety of conditions and are
recommended for this purpose (LeBreton et al. 2004). Despite calls
advocating for the wider use of these indices (Tonidandel and
LeBreton 2011), researchers seem reluctant to do so.
Budescu, D. V. 1993 Dominance analysis: A new approach to the problem of relative importance of predictors in
multiple regression Psychological Bulletin, 114, 542551 DOI: 10.1037/0033-2909.114.3.542.
Johnson, J. W. 2000 A heuristic method for estimating the relative weight of predictor variables in multiple
regression Multivariate Behavioral Research, 35, 119 DOI: 10.1207/S15327906MBR3501_1.
LeBreton, J. M., Ployhart, R. E. and Ladd, R. T. 2004 A Monte Carlo comparison of relative importance
methodologies Organizational Research Methods, 7, 258282 DOI: 10.1177/1094428104266017.
Tonidandel, S., and LeBreton, J. M. 2011 Relative importance analyses: A useful supplement to multiple
regression analyses Journal of Business and Psychology, 26, 19 DOI: 10.1007/s10869-010-9204-3.
Index End 221

Analysis of covariance
Analysis of covariance is like ANOVA, except in addition to the categorical
predictors you also have continuous predictors as well. For example, the one
way ANOVA example used write as the dependent variable and prog as the
independent variable. Let's add read as a continuous variable to this model,
as shown below.
Menu selection:- Analyze > General Linear Model > Univariate
Syntax:- glm write with read by prog.
222
223
To help understand all these

applications, concentrate on the
introduction of the variables etc.
224
Between-Subjects Factors
Value Label N
The results indicate
that even after
type of program 1.00 general 45
2.00 academic 105
3.00 vocation 50 adjusting for the
reading score (read),
Tests of Between-Subjects Effects the writing scores still
Dependent Variable:writing score
significantly differ by
program type (prog),
Type III Sum of
Source Squares df Mean Square F Sig.
Corrected Model 7017.681 a
3 2339.227 42.213 .000 F=5.867, p=0.003.
Intercept 4867.964 1 4867.964 87.847 .000
read 3841.983 1 3841.983 69.332 .000
prog 650.260 2 325.130 5.867 .003
Error 10861.194 196 55.414
Total 574919.000 200
Corrected Total 17878.875 199
a. R Squared = 0393 (Adjusted R Squared = 0383)
Index End
225
Multiple logistic regression
Multiple logistic regression is like simple logistic regression, except that
there are two or more predictors. The predictors can be interval variables or
dummy variables, but cannot be categorical variables. If you have categorical
predictors, they should be coded into one or more dummy variables. We have
only one variable in our data set that is coded 0 and 1, and that is female. We
understand that female is a silly outcome variable (it would make more sense
to use it as a predictor variable), but we can use female as the outcome
variable to illustrate how the code for this command is structured and how
to interpret the output. The first variable listed after the logistic
regression command is the outcome (or dependent) variable, and all of the
rest of the variables are predictor (or independent) variables (listed after
the keyword with). In our example, female will be the outcome variable, and
read and write will be the predictor variables.
Syntax:- logistic regression female with read write. 226

227
228
Block 1: Method = Enter
a
Unweighted Cases N Percent Omnibus Tests of Model Coefficients
Selected Cases Included in Analysis 200 100.0 Chi-square df Sig.
Missing Cases 0 .0 Step 1 Step 27.819 2 .000
Total 200 100.0 Block 27.819 2 .000
Unselected Cases 0 .0
Model 27.819 2 .000
Total 200 100.0
a. If weight is in effect, see classification table for the total number
Model Summary
of cases.
Loads of output!!
-2 Log Cox & Snell R Nagelkerke R
Step likelihood Square Square
Dependent Variable Encoding
1 247.818a .130 .174
Original Value Internal Value
a. Estimation terminated at iteration number 4 because
male 0
parameter estimates changed by less than .001.
female 1
Block 0: Beginning Block Classification Table a
a,b Predicted
Classification Table
female Percentage
Predicted
Observed male female Correct
female Percentage
Observed Correct Step 1 female male 54 37 59.3
male female
female 30 79 72.5
Step 0 female male 0 91 .0
Overall Percentage 66.5
female 0 109 100.0
a. The cut value is .500
Overall Percentage 54.5
a. Constant is included in the model.
b. The cut value is .500 Variables in the Equation
B S.E. Wald df Sig. Exp(B)
Variables in the Equation Step 1a read -.071 .020 13.125 1 .000 .931
B S.E. Wald df Sig. Exp(B) write .106 .022 23.075 1 .000 1.112
Step 0 Constant .180 .142 1.616 1 .204 1.198 Constant -1.706 .923 3.414 1 .065 .182
a. Variable(s) entered on step 1: read, write.
Variables not in the Equation

Score df Sig.
Step 0 Variables read .564 1 .453
write 13.158 1 .000
Overall Statistics 26.359 2 .000
229
Variables in the Equation
B S.E. Wald df Sig. Exp(B)
Step 1a read -.071 .020 13.125 1 .000 .931
write .106 .022 23.075 1 .000 1.112
Constant -1.706 .923 3.414 1 .065 .182
a. Variable(s) entered on step 1: read, write.
These results show that both read and write

are significant predictors of female.
Index End
230
Discriminant analysis
Discriminant analysis is used when you have one or more normally
distributed interval independent variable(s) and a categorical dependent
variable. It is a multivariate technique that considers the latent
dimensions in the independent variables for predicting group membership
in the categorical dependent variable. For example, using the A data file,
say we wish to use read, write and math scores to predict the type of
program a student belongs to (prog).
Menu selection:- Analyze > Classify > Discriminant
Syntax:- Discriminant groups = prog(1, 3)

/variables = read write math.
231
232
Do not forget to define the range for Prog.
233
Analysis Case Processing Summary Wilks' Lambda
Unweighted Cases N Percent Test of Function(s) Wilks' Lambda Chi-square df Sig.
Valid 200 100.0 1 through 2 .734 60.619 6 .000
Excluded Missing or out-of-range 0 .0 2 .995 .888 2 .641
group codes
At least one missing 0 .0
Standardized Canonical Discriminant
discriminating variable
Function Coefficients
Both missing or out-of- 0 .0
Function
range group codes and at
1 2
least one missing
discriminating variable reading score .273 -.410
Total 0 .0 writing score .331 1.183
Total 200 100.0 math score .582 -.656
Group Statistics Structure Matrix
Valid N (listwise) Function
Loads of output!!
type of program Unweighted Weighted 1 2
general reading score 45 45.000 math score .913* -.272
writing score 45 45.000 reading score .778* -.184

writing score .775* .630
math score 45 45.000
Pooled within-groups correlations
academic reading score 105 105.000
between discriminating variables and
writing score 105 105.000
standardized canonical discriminant
functions
vocation reading score 50 50.000
Variables ordered by absolute size of
writing score 50 50.000
correlation within function.
*. Largest absolute correlation between
Total reading score 200 200.000 each variable and any discriminant
writing score 200 200.000 function
Analysis 1 Functions at Group Centroids

Function
Summary of Canonical Discriminant Functions
type of program 1 2
general -.312 .119
Eigenvalues
academic .536 -.020
Canonical
vocation -.844 -.066
Function Eigenvalue % of Variance Cumulative % Correlation
Unstandardized canonical discriminant
1 .356a 98.7 98.7 .513
functions evaluated at group means
2 .005a 1.3 100.0 .067
a. First 2 canonical discriminant functions were used in the analysis.
234
Functions at Group Centroids
Function
type of program 1 2
general -.312 .119
academic .536 -.020
vocation -.844 -.066
Unstandardized canonical discriminant
functions evaluated at group means
Clearly, the SPSS output for this procedure is quite lengthy, and it is
beyond the scope of this page to explain all of it. However, the main point
is that two canonical variables are identified by the analysis, the first of
which seems to be more related to program type than the second.
Index End
235
One-way MANOVA
MANOVA (multivariate analysis of variance) is like ANOVA, except that
there are two or more dependent variables. In a one-way MANOVA,
there is one categorical independent variable and two or more dependent
variables. For example, using the A data file, say we wish to examine the
differences in read, write and math broken down by program type (prog).
Menu selection:- Analyse > General Linear Model > Multivariate
Syntax:- glm read write math by prog.
236
One-way MANOVA
237
One-way MANOVA
238
One-way MANOVA
Between-Subjects Factors
Value Label N
type of program 1.00 general 45
2.00 academic 105
3.00 vocation 50
Multivariate Testsc
Effect Value F Hypothesis df Error df Sig.
Intercept Pillai's Trace .978 2883.051a 3.000 195.000 .000
Wilks' Lambda .022 2883.051a 3.000 195.000 .000
Hotelling's Trace 44.355 2883.051a 3.000 195.000 .000
Roy's Largest Root 44.355 2883.051a 3.000 195.000 .000
prog Pillai's Trace .267 10.075 6.000 392.000 .000
Wilks' Lambda .734 10.870a 6.000 390.000 .000
Hotelling's Trace .361 11.667 6.000 388.000 .000
Roy's Largest Root .356 23.277b 3.000 196.000 .000
a. Exact statistic
b. The statistic is an upper bound on F that yields a lower bound on the significance level.
c. Design: Intercept + prog
Tests of Between-Subjects Effects

Type III Sum of
Source Dependent Variable Squares df Mean Square
Corrected Model reading score 3716.861a 2 1858.431
writing score 3175.698b 2 1587.849
math score 4002.104c 2 2001.052
Intercept reading score 447178.672 1 447178.672
writing score 460403.797 1 460403.797
Concentrate on
math score 453421.258 1 453421.258
prog reading score 3716.861 2 1858.431
writing score 3175.698 2 1587.849
Error
math score
reading score
4002.104
17202.559
2
197
2001.052
87.323 the third table
writing score 14703.177 197 74.635
math score 13463.691 197 68.344
Total reading score 566514.000 200
writing score 574919.000 200
math score 571765.000 200
Corrected Total reading score 20919.420 199
writing score 17878.875 199
math score 17465.795 199
239
One-way MANOVA
Concluding output table.
The students in the different

programs differ in their joint
distribution of read, write and math.
Index End 240

Multivariate multiple
regression
Multivariate multiple regression is used when you have two or more
dependent variables that are to be predicted from two or more
independent variables. In our example, we will predict write and read from
female, math, science and social studies (socst) scores.
Menu selection:- Analyse > General Linear Model > Multivariate
Syntax:- glm write read with female math science socst.
241
regression
242
regression
243
regression
Multivariate Testsb Tests of Between-Subjects Effects
Effect Value F Hypothesis df Error df Sig. Type III Sum of
Intercept Pillai's Trace .030 3.019a
2.000 194.000 .051 Source Dependent Variable Squares df Mean Square
Wilks' Lambda .970 3.019a
2.000 194.000 .051 Corrected Model writing score 10620.092a 4 2655.023
b
Hotelling's Trace .031 3.019a
2.000 194.000 .051 reading score 12219.658 4 3054.915
Roy's Largest Root .031 3.019a 2.000 194.000 .051 Intercept writing score 202.117 1 202.117
female Pillai's Trace .170 19.851a 2.000 194.000 .000 reading score 55.107 1 55.107
Wilks' Lambda .830 19.851a 2.000 194.000 .000 female writing score 1413.528 1 1413.528
Hotelling's Trace .205 19.851a 2.000 194.000 .000 reading score 12.605 1 12.605
Roy's Largest Root .205 19.851a 2.000 194.000 .000 math writing score 714.867 1 714.867
math Pillai's Trace .160 18.467a 2.000 194.000 .000 reading score 1025.673 1 1025.673
2.000 194.000 .000 science writing score 857.882 1 857.882
2.000 194.000 .000 reading score 946.955 1 946.955
Roy's Largest Root .190 18.467a
2.000 194.000 .000 socst writing score 1105.653 1 1105.653
science Pillai's Trace .166 19.366a 2.000 194.000 .000 reading score 1475.810 1 1475.810
2.000 194.000 .000 Error writing score 7258.783 195 37.225
2.000 194.000 .000 reading score 8699.762 195 44.614
Roy's Largest Root .200 19.366a
2.000 194.000 .000 Total writing score 574919.000 200
socst Pillai's Trace .221 27.466a 2.000 194.000 .000 reading score 566514.000 200

2.000 194.000 .000 Corrected Total writing score 17878.875 199

2.000 194.000 .000 reading score 20919.420 199
a
Roy's Largest Root .283 27.466 2.000 194.000 .000
a. Exact statistic
b. Design: Intercept + female + math + science + socst
Concentrate on
the this table
244
regression
Concluding table.
Tests of Between-Subjects Effects
Source Dependent Variable F Sig.
Corrected Model writing score 71.325 .000
reading score 68.474 .000 These results show that all of the
Intercept writing score 5.430 .021 variables in the model have a
reading score 1.235 .268 statistically significant relationship with
female writing score 37.973 .000
the joint distribution of write and
reading score .283 .596
math writing score 19.204 .000
read.
reading score 22.990 .000
science writing score 23.046 .000
socst writing score 29.702 .000
Error writing score
reading score
Total writing score
reading score
Corrected Total writing score
reading score
a. R Squared = 0594 (Adjusted R Squared = 0586) Index End 245
b. R Squared = 0584 (Adjusted R Squared = 0576)
Canonical correlation
Canonical correlation is a multivariate technique used to examine the
relationship between two groups of variables. For each set of variables, it
creates latent variables and looks at the relationships among the latent
variables. It assumes that all variables in the model are interval and
normally distributed. SPSS requires that each of the two groups of
variables be separated by the keyword with. There need not be an equal
number of variables in the two groups (before and after the with). In this
case {read, write} with {math, science}.
Canonical correlation are the correlations of two canonical (latent)

variables, one representing a set of independent variables, the other a set
of dependent variables. There may be more than one such linear
correlation relating the two sets of variables, with each correlation
representing a different dimension by which the independent set of
variables is related to the dependent set. The purpose of the method is to
explain the relation of the two sets of variables, not to model the
individual variables. 246
Canonical correlation analysis is the study of the linear relations between
two sets of variables. It is the multivariate extension of correlation
analysis.
Suppose you have given a group of students two tests of ten questions
each and wish to determine the overall correlation between these two
tests. Canonical correlation finds a weighted average of the questions
from the first test and correlates this with a weighted average of the
questions from the second test. The weights are constructed to maximize
the correlation between these two averages. This correlation is called the
first canonical correlation coefficient.
You can create another set of weighted averages unrelated to the first
and calculate their correlation. This correlation is the second canonical
correlation coefficient. This process continues until the number of
canonical correlations equals the number of variables in the smallest
group. 247
In statistics, canonical-correlation analysis is a way of making sense of cross-
covariance matrices. If we have two vectors X = (X 1, ..., Xn) and Y = (Y1, ..., Ym)
of random variables, and there are correlations among the variables, then
canonical-correlation analysis will find linear combinations of the X i and Yj
which have maximum correlation with each other (Hrdle and Lopold 2007).
T. R. Knapp notes virtually all of the commonly encountered parametric tests
of significance can be treated as special cases of canonical-correlation
analysis, which is the general procedure for investigating the relationships
between two sets of variables. The method was first introduced by Harold
Hotelling in 1936.
Hrdle, Wolfgang and Simar, Lopold (2007). Canonical Correlation Analysis.

Applied Multivariate Statistical Analysis. pp. 321330. Canonical Correlation Analysis -
Springer ISBN 978-3-540-72243-4.
Knapp, T. R. (1978). Canonical correlation analysis: A general parametric
significance-testing system. Psychological Bulletin 85(2): 410416 DOI:
10.1037/0033-2909.85.2.410 .
Hotelling, H. (1936). Relations Between Two Sets of Variates. Biometrika 28
(34): 321377 DOI: 10.1093/biomet/28.3-4.321 . 248
The manova command, for canonical correlation, is one of the SPSS
commands that can only be accessed via syntax; there is not a sequence of
pull-down menus or point-and-clicks that could arrive at this analysis.
Syntax:- manova read write with math science

/discrim all alpha(1)
/print=sig(eigen dim).
249
The output shows
the linear
combinations
corresponding to
the first canonical
correlation. At the
bottom of the
output are the two
canonical
correlations. These
results indicate
that the first
canonical
correlation is .7728.
250
The F-test in this
output tests the
hypothesis that the
first canonical
correlation is not
equal to zero.
Clearly, F=56.4706
is statistically
significant.
However, the
second canonical
correlation of .0235
is not statistically
significantly
different from zero
(F=0.1087,
p=0.742).
Index End 251
Factor analysis
Factor analysis is a form of exploratory multivariate analysis that is used to
either reduce the number of variables in a model or to detect relationships
among variables. All variables involved in the factor analysis need to be interval
and are assumed to be normally distributed. The goal of the analysis is to try
to identify factors which underlie the variables. There may be fewer factors
than variables, but there may not be more factors than variables. For our
example, let's suppose that we think that there are some common factors
underlying the various test scores. We will include subcommands for varimax
rotation and a plot of the eigenvalues. We will use a principal components
extraction and will retain two factors.
Menu selection:- Analyze > Dimension Reduction > Factor

Syntax:- factor
/variables read write math science socst
/criteria factors(2)
/extraction pc
/rotation varimax
/plot eigen.
Factor analysis Crichton, N. Journal Of Clinical Nursing 10(4) 562-562 2001
252
Factor analysis
253
Factor analysis
254
Factor analysis
255
Factor analysis
256
Factor analysis
Communalities
Initial Extraction
reading score
writing score
1.000
1.000
.736
.704
Communality (which is the opposite
math score
science score
1.000
1.000
.750
.849
of uniqueness) is the proportion of
social studies score 1.000
Extraction Method: Principal Component
.900
variance of the variable (i.e., read)
Analysis. that is accounted for by all of the
Total Variance Explained factors taken together, and a very
Extraction
Sums of low communality can indicate that a
Initial Eigenvalues
Squared
Loadings
variable may not belong with any of
Component
1
Total
3.381
% of Variance
67.616
Cumulative %
67.616
Total
3.381
the factors.
2 .557 11.148 78.764 .557
3 .407 8.136 86.900
4 .356 7.123 94.023
5 .299 5.977 100.000
Total Variance Explained

Extraction Sums of Squared
Loadings Rotation Sums of Squared Loadings
Component % of Variance Cumulative % Total % of Variance Cumulative %
1 67.616 67.616 2.113 42.267 42.267
2 11.148 78.764 1.825 36.497 78.764
3
4
5
Extraction Method: Principal Component Analysis.
257
Factor analysis
The scree plot may be useful in determining how many factors to retain.
258
Factor analysis
Component Matrixa
Component
reading score
1
.858
2
-.020
From the component matrix table, we can see
writing score
math score
.824
.844
.155
-.195
that all five of the test scores load onto the
science score
social studies score
.801
.783
-.456
.536
first factor, while all five tend to load not so
Extraction Method: Principal Component heavily on the second factor. The purpose of
rotating the factors is to get the variables to
Analysis.
a. 2 components extracted.
Rotated Component Matrixa load either very high or very low on each factor.
Component
1 2
In this example, because all of the variables
reading score .650 .559 loaded onto factor 1 and not on factor 2, the
rotation did not aid in the interpretation.
writing score .508 .667
math score .757 .421
Instead, it made the results even more difficult

science score .900 .198
social studies score .222 .922
to interpret.
Extraction Method: Principal Component
Analysis.
Rotation Method: Varimax with Kaiser
Normalization.
a. Rotation converged in 3 iterations.
Component Transformation Matrix

Component 1 2
1 .742 .670
2 -.670 .742
Extraction Method: Principal
Component Analysis.
Rotation Method: Varimax with Index End
Kaiser Normalization.
259
Normal probability
Many statistical methods require that the numeric
variables we are working with have an approximate
normal distribution. For example, t-tests,
F-tests, and regression analyses all require in some
sense that the numeric variables are approximately
260
Normal probability
Following a thorough review, it is suggested
a compelling rationalization of the widely noted
robustness of the fixed-effects ANOVA to non-
normality.
It has often been reported that violation of the
normality assumption should be of little concern.
Glass, G.V., P.D. Peckham, and J.R. Sanders. 1972.

Consequences of failure to meet assumptions
underlying fixed effects analyses of variance and
covariance. Rev. Educ. Res. 42: 237-288. DOI:
10.3102/00346543042003237
261
Normal probability
Another review which also suggests alternate tests
non-normality appeared to have little effect on Type I
error performance but could have more serious
implications for statistical power.
Lix, L.M., J.C. Keselman, and H.J. Keselman. 1996.

Consequences of assumption violations revisited: A
quantitative review of alternatives to the one-way
analysis of variance F test. Rev. Educ. Res. 66: 579-619.
DOI: 10.3102/00346543066004579
262
Normal probability plot
Tools for Assessing Normality include
Histogram and Boxplot
Normal Quantile Plot (also called Normal
Probability Plot)
Goodness of Fit Tests such as

Anderson-Darling Test
Kolmogorov-Smirnov Test
Lillefors Test
Shapiro-Wilk Test
Problem: they dont always agree!

263
You could produce conventional descriptive
statistics, a histogram with a superimposed normal
curve, and a normal scores plot also called a
normal probability plot.
The pulse data from data set C is employed.
264
Analyze
> Descriptive Statistics
> Explore
Under plots select histogram,

also normality plots with tests,
descriptive statistics and
boxplots are default options.
265
266
267
These tests are

considered in the
next section.
268
269
If the data is normal the non-linear vertical axis in a

probability plot should result in an approximately linear
scatter plot representing the raw data. 270
Detrended normal
P-P plots depict
the actual
deviations of data
points from the
straight horizontal
line at zero. No
specific pattern in
a detrended plot
indicates normality
of the variable.
271
272
Graphs
> Legacy Dialogs
> Histogram
Tick display normal

curve
273
274
275
Graphs
> Legacy Dialogs
> Line
Select Simple and

Groups of Cases the
use Define to choose
the variable and select
cum %.
276
277
278
If you wish to superimpose a normal curve, it is probably

simpler in Excel!
This approach underlies the Kolmogorov-Smirnov test. 279
You are seeking to assess the normality of the
data. The pulse data from data set C is employed.
The P-P plot is a normal probability plot with the

data on the horizontal axis and the expected z-
scores if our data was normal on the vertical axis.
When our data is approximately normal the
spacing of the two will agree resulting in a plot
with observations lying on the reference line in
the normal probability plot.
280
281
Index End
Skewness
The skewness is the third centralised normalised moment.
If skewness is positive, the data are positively skewed or skewed right, meaning
that the right tail of the distribution is longer than the left. If skewness is
negative, the data are negatively skewed or skewed left, meaning that the left tail
is longer.
If skewness = 0, the data are perfectly symmetrical. But a skewness of exactly zero
is quite unlikely for real-world data, so how can you interpret the skewness number?
Bulmer (1979) suggests this rule of thumb:
If skewness is less than 1 or greater than +1, the distribution is highly skewed.
If skewness is between 1 and or between + and +1, the distribution is
moderately skewed.
If skewness is between and +, the distribution is approximately symmetric.
Bulmer, M. G. 1979. Principles of Statistics. Dover.
282
Skewness
But what do I mean by too much for random chance to be the explanation? To answer that,
you need to divide the sample skewness G 1 by the standard error of skewness (SES) to get the
test statistic, which measures how many standard errors separate the sample skewness from
zero:
6n(n 1)
SES
(n 2)(n 1)(n 3)
test statistic: ZG1 = G1/SES where
This formula is adapted from page 85 of Cramer (1997).

The critical value of ZG1 is approximately 2. (This is a two-tailed test of skewness 0 at roughly
the 0.05 significance level.)
If ZG1 < 2, the population is very likely skewed negatively (though you dont know by how much).
If ZG1 is between 2 and +2, you cant reach any conclusion about the skewness of
the population: it might be symmetric, or it might be skewed in either direction.
If ZG1 > 2, the population is very likely skewed positively (though you dont know by how much).
Cramer, Duncan 1997

Basic Statistics for Social Research, Routledge.
283
Skewness
It is available via Analyze > Descriptive Statistics > Descriptives
The syntax is
DESCRIPTIVES VARIABLES=x3 x2 x rt_x ln_x x_1 x_2 x_3

/STATISTICS=SKEWNESS.
284
Skewness
Descriptive Statistics
N Skewness
Std.
Statistic Statistic Ratio
Error
x3 90 2.208 0.254 8.69
x2 90 1.885 0.254 7.42
x 90 1.550 0.254 6.10
rt_x 90 1.380 0.254 5.43
ln_x 90 1.208 0.254 4.76
x_1 90 -0.864 0.254 -3.40
x_2 90 -0.526 0.254 -2.07
x_3 90 -0.203 0.254 -0.80
285
Index End
Kurtosis
The kurtosis is the fourth centralised normalised moment.
The question is similar to the question about skewness, and the answers
are similar too. You divide the sample excess kurtosis by the standard
error of kurtosis (SEK) to get the test statistic, which tells you how
many standard errors the sample excess kurtosis is from zero:
test statistic: ZG2 = G2 / SEK where n 2 1

SEK 2 SES
(n 3)(n 5)
The formula is adapted from page 89 of Cramer (1979).
Cramer, Duncan 1997

Basic Statistics for Social Research, Routledge.
286
Kurtosis
The critical value of ZG2 is approximately 2. (This is a two-tailed test of
excess kurtosis 0 at approximately the 0.05 significance level.)
If ZG2 < 2, the population very likely has negative excess kurtosis
(kurtosis <3, platykurtic), though you dont know how much.
If ZG2 is between 2 and +2, you cant reach any conclusion about the
kurtosis: excess kurtosis might be positive, negative, or zero.
If ZG2 > +2, the population very likely has positive excess kurtosis
(kurtosis >3, leptokurtic), though you dont know how much.
Cramer, Duncan. 1997.

Basic Statistics for Social Research. Routledge.
287
Kurtosis!
The rules for determining the type of distribution based on skewness and
kurtosis may however vary among statisticians. Evans (2007) for instance,
suggested that distribution with skewness value of greater than 1 or less than -1
could be considered as highly skewed. Those with skewness value of between 0.5
and 1 or between -1 and -0.5 is said to have moderately skewed distribution
whereas a value between 0.5 and -0.5 indicates relative symmetry. Brown (1997)
on the other hand, proposed to the practitioners to make use of the standard
error of skewness (SES) and standard error of kurtosis (SEK) in deciding
whether the tested data could be assumed to come from a normal distribution.
He suggested that the data could be assumed as normally distributed if the
skewness and kurtosis values lie within the range of 2SES and 2SEK,
respectively. Some practitioners favour one and some favour the others.
Nonetheless, skewness and kurtosis do not provide conclusive information about
normality. Hence, it is always a good practice to supplement the skewness and
kurtosis coefficients with other methods of testing normality such as the
graphical methods and formal tests of normality.
J.R. Evans, Statistics, data analysis and decision making, 3rd edition, Prentice
Hall, pp. 60, 2007.
J.D. Brown, Skewness and kurtosis, Shiken: JALT Testing & Evaluation SIG
288
Newsletter, vol. 1, pp. 18 20, 1997.
Kurtosis
It is available via Analyze > Descriptive Statistics > Descriptives
The syntax is
DESCRIPTIVES VARIABLES=x3 x2 x rt_x ln_x x_1 x_2 x_3

/STATISTICS=KURTOSIS.
289
Kurtosis
Descriptive Statistics
N Kurtosis
Std.
Statistic Statistic Ratio
Error
x3 90 4.776 0.503 9.50
x2 90 3.356 0.503 6.67
x 90 2.162 0.503 4.30
rt_x 90 1.651 0.503 3.28
ln_x 90 1.201 0.503 2.39
x_1 90 0.480 0.503 0.95
x_2 90 -0.001 0.503 0.00
x_3 90 -0.256 0.503 -0.51
290
Index End
Does It Really Matter?
Students t test and more generally the ANOVA F test
are robust to non-normality (Fayers 2011).
However
Thus a clearer statement is that t tests and ANOVA

are robust against type-I errors. This of course accords
with the enthusiasm that many researchers have in
obtaining significant p values.
The aim of the article (see next slide) is to show that
type-II errors can be substantially increased if non-
normality is ignored. (Fayers 2011).
291
Does It Really Matter?
Alphas, betas and skewy distributions: two ways of getting the wrong
answer
Peter Fayers
Adv. Health Sci. Educ. Theory Pract. 2011 16(3) 291296 DOI:
10.1007/s10459-011-9283-6
Introduction to Robust Estimation and Hypothesis Testing (2nd ed.).

Wilcox, R. R., 2005, Burlington MA:
Elsevier Academic Press. ISBN 978-0-12-751542-7.
Robustness to Non-Normality of Common Tests for the Many-

Sample Location Problem
Khan A. and Rayner G.D.
Journal Of Applied Mathematics And Decision Sciences, 2003, 7(4),
187:206 DOI: 10.1155/S1173912603000178
Index End 292
Tukey's ladder of powers
Tukey has designed a family of power transformations
(close cousin to the Box-Cox transformations, but with a
visual aspect useful to find the appropriate transformation
3 y3 to promote symmetry and linearity relationships.
2 y2
1 y1 These transformations preserve order, preserve
proximities and are smooth functions (not producing jumps
y
ln(y or peaks). y1 is the untransformed (raw) variable, y0 is
0 replaced by the logarithm that provides the appropriate
)
transformation between the square root and the reciprocal.
-1 y-1
-2 y-2 You can also use lower and higher powers as listed, as well
-3 y-3 intermediate ones, i.e. y2.5 will be stronger than y but less
than y.
Cartoon
Tukey, J. W. (1977) Exploratory Data Analysis. Addison-
Wesley, Reading, MA.
293
A transformation is simply a means of representing the data in a different
coordinate system. In addition to restoring normality, the transformation
often reduces heteroscedasticity. (Non-constancy of the variance of a
measure over the levels of the factor under study.) This is important, because
constant variance is often an assumption of parametric tests. Subsequent
statistical analyses are performed on the transformed data; the results are
interpreted with respect to the original scale of measurement.
Achieving an appropriate transformation is a trial-and-error process. A

particular transformation is applied and the new data distribution tested for
normality; if the data are still non-normal, the process is repeated.
Nevertheless, there are certain generalities that can be used to direct your
efforts, as certain types of data typically respond to particular
transformations. For example Square-root transforms are often appropriate
for count data, which tend to follow Poisson distributions. Arcsine (sin -1)
transforms are used for data that are percentages or proportions, and tend
to fit binomial distributions. Log and square-root transforms are part of a
larger class of transforms known as the ladder of powers.
294
Transform > Compute Variable
See normal probability plot section for graphical options. 295

SORT CASES BY pulse(D).
COMPUTE ID=$casenum.
EXECUTE.
Compute work = $sysmis.

The pulse data from data set
if id=1 work=pulse.
Execute.
C is employed.
RMV /pmax=SMEAN(work).
It is scaled by these first
COMPUTE pulse1=1 + pulse/pmax.
EXECUTE. compute statement to aid
COMPUTE y3=pulse1 ** 3.
EXECUTE.
interpretation of the plots.
COMPUTE y2=pulse1 ** 2.
EXECUTE.
COMPUTE y=pulse1 . This step is non-essential,
EXECUTE.
COMPUTE rt_y=SQRT(pulse1). only aiding the graphical
EXECUTE.
COMPUTE ln_y=LN(pulse1). presentation.
EXECUTE.
COMPUTE y_1=pulse1 ** -1.
EXECUTE.
The x variables have not
EXECUTE. been scaled, being simply
EXECUTE. powers of pulse.
EXAMINE VARIABLES=y2 y rt_y y_1 y_2 y_3
/COMPARE VARIABLE
/PLOT=BOXPLOT
/STATISTICS=NONE
/NOTOTAL
/MISSING=LISTWISE.
296
Normal
x3 x2 x1
30 20 20
15 10 10
0 0 0
Frequency
x0 x 1/2 x-1
20 16 16
10 8 8
0 0 0
x-2 x-3
16
10
8 5
0 0
Which appears most normal?

x-2 x-3
297
Normal - 95% CI
x3 x2 x1
99.9 99.9 99.9
99 99 99
90 90 90
50 50 50
10 10 10
1 1 1
0.1 0.1 0.1
0 2000000 4000000 0 10000 20000 50 100 150
x0 x 1/2 x-1
99.9 99.9 99.9
99 99 99
Percent
90 90 90
50 50 50
10 10 10
1 1 1
0.1 0.1 0.1
4.0 4.5 5.0 8 10 12 0.005 0.010 0.015
x-2 x-3
99.9 99.9
99 99
90 90
50 50
10 10
1 1
0.1 0.1
0.0000 0.0001 0.0002 0.000000 0.000001 0.000002
Which appears most normal?

x-2 x-3
298
x3 x2 x1 x0 x x-1 x-2 x-3

Mean 106124 10158 99.7 4.592 9.96 0.0102 0.0001 1.12E-06
5 2 06
StDev 567251 3310 14.86 0.137 0.711 0.0012 2.52E- 3.75E-07
96 05
N 90 90 90 90 90 90 90 90
AD 8.615 6.523 4.635 3.048 3.799 1.824 0.988 0.529
p- <0.005 <0.00 <0.00 <0.00 <0.00 <0.005 0.013 0.172
value 5 5 5 5
In general if the normal distribution fits the data, then the

plotted points will roughly form a straight line. In addition the
plotted points will fall close to the fitted line. Also the Anderson-
Darling statistic will be small, and the associated p-value will be
larger than the chosen -level (usually 0.05). So the test rejects
the hypothesis of normality when the p-value is less than or equal
to .
299
To test for normality in SPSS
you can perform a Kolmogorov-
Smirnov Test,
Analyze
> Nonparametric tests
> Legacy Dialogs
> 1-Sample Kolmogorov-
Smirnov Test
300
301
302
One-Sample Kolmogorov-Smirnov Test
x3 x2 x rt_x ln_x x-1 x-2 x-3
N 90 90 90 90 90 90 90 90
Normal Mean 1061244.56 10158.41 99.7000 9.9599 4.5924 .0102 .0001 .0000
Parametersa,b 67 11
Std. Deviation 567251.119 3309.533 14.8584 .71097 .13698 .00130 .00003 .00000
96 01 7
Most Extreme Absolute .255 .221 .192 .178 .163 .133 .105 .079
Differences Positive .255 .221 .192 .178 .163 .063 .060 .054
Negative -.173 -.139 -.108 -.094 -.080 -.133 -.105 -.079
Kolmogorov-Smirnov Z 2.422 2.099 1.821 1.684 1.544 1.263 .993 .745
Asymp. Sig. (2-tailed) .000 .000 .003 .007 .017 .082 .278 .635
a. Test distribution is Normal.
b. Calculated from data.
The Asymp. Sig. (2 tailed) value is also known as the p-value. This tells
you the probability of getting the results if the null hypothesis were
actually true (i.e. it is the probability you would be in error if you
rejected the null hypothesis).
303
Despite the scaling
the log. transform
spoils the final plot.
304
You are seeking
the most normal
data visually.
Probably one of
y_1, y_2 or y_3
transform.
305
Many statistical methods require that the numeric
variables you are working with have an approximately
normal distribution. Reality is that this is often times
not the case. One of the most common departures from
normality is skewness, in particular, right skewness.
When the data is plotted vs. the expected z-scores the

normal probability plot shows right skewness by a
downward bending curve.
When the data is plotted vs. the expected z-scores the

normal probability plot shows left skewness by an
upward bending curve.
306
Tukeys Ladder of Powers
Tukey (1977) describes an orderly way of re-
expressing variables using a power transformation.
If a transformation for x of the type x, results
in an effectively linear probability plot, then we
should consider changing our measurement scale
for the rest of the statistical analysis. There is
no constraint on values of that we may consider.
Obviously choosing =1 leaves the data unchanged.
Negative values of are also reasonable. Tukey
(1977) suggests that it is convenient to simply
define the transformation when =0 to be the
logarithmic function rather than the constant 1.
307
In general if the normal distribution fits the data,
then the plotted points will roughly form a
straight line. In addition the plotted points will
fall close to the fitted line. Also the Anderson-
Darling statistic will be small, and the associated
p-value will be larger than the chosen -level
(usually 0.05). So the test rejects the hypothesis
of normality when the p-value is less than or equal
to .
308
To test for normality is SPSS you can perform a
Kolmogorov-Smirnov Test
Analyze > Nonparametric tests

> 1-Sample Kolmogorov-Smirnov Test
The Asymp. Sig. (2 tailed) value is also known as the

p-value. This tells you the probability of getting the
results if the null hypothesis were actually true (i.e.
it is the probability you would be in error if you
rejected the null hypothesis).
309
The hypothesis are
H0 the distribution of x is normal

H1 the distribution of x is not normal
If the p-value is less than .05, you reject the normality

assumption, and if the p-value is greater than .05, there
is insufficient evidence to suggest the distribution is not
normal (meaning that you can proceed with the
assumption of normality.)
In summary if the test is significant (lower than or equal

to 0.05) implies the data is not normally distributed. 310
To read more about Normality tests.
Comparisons Of Tests For Normality With A Cautionary Note

Dyer, A.R.
Biometrika 61(1) 185-189 1974 DOI: 10.1093/biomet/61.1.185
A Comparison Of Various Tests Of Normality

Yazici, Berna and Yolacan, Senay
Journal Of Statistical Computation And Simulation 77(2) 175-183 2007
DOI: 10.1080/10629360600678310
Comparisons of various types of normality tests

B. W. Yap and C. H. Sim
Journal of Statistical Computation and Simulation 81(12) 2141-2155
2011 DOI: 10.1080/00949655.2010.520163
311
Index End
Median Split
There is quite a literature to suggest that, even though
it is nice and convenient to sort people into 2 groups and
then use a t test to compare group means, you lose
considerable power. Cohen (1983) has said that breaking
subjects into two groups leads to the loss of 1/5 to 2/3
of the variance accounted for by the original variables.
The loss in power is equivalent to tossing out 1/3 to 2/3
of the sample.
The Cost Of Dichotomization or paper

Cohen, J.
Applied Psychological Measurement 7(3) 249-253 1983
312
Index End
Likert Scale
These are not exhaustive notes, rather some thoughts on preparing
a Likert Scale.
Statements are often rated on a five point Likert (Likert 1932)

scale. There has been much research done to demonstrate that a
five-point scale can lead to extremes, and therefore a seven-point
scale is preferable (Payton et al. 2003).
Likert, R. (1932). A technique for the measurement of attitudes.

Archives of Psychology, 22(140) 55pp.
Payton, M.E., Greenstone, M.H. and Schenker, N. (2003).

Overlapping confidence intervals or standard error intervals: What do
they mean in terms of statistical significance?
. Journal of Insect Science, 3(34), 6pp.
313
Likert Scale
A 7-point Likert scale is recommended to maximise the sensitivity of
the scale (Allen and Seaman 2007, Cummins and Gullone 2000).
Allen, I.E., and Seaman, C.A. (2007). Likert scales and data analyses.
Quality Progress, 40(7), 64-65.
Cummins, R.A., and Gullone, E. (2000).

Why we should not use 5-point Likert scales: The case for subjective
quality of life measurement
. In Proceedings, second international conference on quality of life in
cities (p74-93). Singapore: National University of Singapore.
Leung, S-O. (2011). A Comparison of Psychometric Properties and

Normality in 4-, 5-, 6-, and 11-Point Likert Scales, Journal of Social
Service Research 37(4) 412-421 DOI:
10.1080/01488376.2011.580697 an odd place for a technical paper.314
Likert Scale
It is unclear why different scales may have been used in the same
experiment, but it has been shown that data is still comparable when
it has been re-scaled (Dawes 2008).
Dawes J. (2008)
Do data characteristics change according to the number of scale poin
ts used? An experiment using 5 point, 7 point and 10 point scales
International Journal of Market Research. 51 (1) 61-77.
315
Likert Scale
It may be better if the scale contained a neutral midpoint (Tsang
2012). This decision (an odd/even scale) depends whether
respondents are being forced to exclude the neutral position with an
even scale.
Tsang K.K 2012

The use of midpoint on Likert Scale: The implications for educationa
l research
Hong Kong Teachers Centre Journal 11 121-130.
316
Likert Scale
An odd number of points allow people to select a middle option. An
even number forces respondents to take sides. An even number is
appropriate when you want to know what direction the people in the
middle are leaning. However, forcing people to choose a side, without
a middle point, may frustrate some respondents (Wong et al. 1993).
Wong, C.-S., Tam, K.-C., Fung, M.-Y., and Wan, K. (1993).

Differences between odd and even number of response scale: Some e
mpirical evidence
. Chinese Journal of Psychology, 35, 75-86.
317
Likert Scale
Since they have no neutral point, even-numbered Likert scales force
the respondent to commit to a certain position (Brown, 2006) even if
the respondent may not have a definite opinion.
There are some researchers who prefer scales with 7 items or with
an even number of response items (Cohen, Manion, and Morrison,
2000).
Brown, J.D. (2000).

What issues affect Likert-scale questionnaire formats? JALT
Testing and Evaluation SIG, 4, 27-30.
Cohen, L., Manion, L. and Morrison, K. (2000). Research methods in

education (5th ed.). London: Routledge Falmer.
318
Likert Scale
The change of response order in a Likert-type scale altered
participant responses and scale characteristics. Response order is the
order in which options of a Likert-type scale are offered (Weng 2000).
How many scale divisions or categories should be used (1 to 10; 1 to 7;

-3 to +3)?
Should there be an odd or even number of divisions? (Odd gives
neutral centre value; even forces respondents to take a non-neutral
position.)
What should the nature and descriptiveness of the scale labels be?
What should the physical form or layout of the scale be? (graphic,
simple linear, vertical, horizontal)
Should a response be forced or be left optional?
Li-Jen Weng 2000 Effects of Response Order on Likert-Type Scales,

Educational and Psychological Measurement 60(6) 908-924 DOI:
10.1177/00131640021970989
Index End 319
Winsorize
Winsorising or Winsorization is a transformation by limiting extreme
values in the data to reduce the effect of possibly spurious outliers. It
is named after the engineer-turned-biostatistician Charles P. Winsor
(1895-1951).
The computation of many statistics can be heavily influenced by

extreme values. One approach to providing a more robust computation
of the statistic is to Winsorize the data before computing the
statistic.
Apart from confusion about the correct spelling. There is the

ambiguity about where the precise percentile sits.
320
Winsorize
To Winsorize the data, tail values are set equal to some specified
percentile of the data. For example, for a 90% Winsorization, the
bottom 5% of the values are set equal to the value corresponding to
the 5th percentile while the upper 5% of the values are set equal to
the value corresponding to the 95th percentile.
Just because a method exists

does not necessarily mean its
a great idea!!
321
Winsorize
The pulse data from data set C is
employed.
Analyze > Descriptive Statistics

> Frequencies
322
Winsorize
Select statistics
323
Winsorize
Add desired percentiles, 5 then 95
324
Winsorize
For brevity do not display frequency tables
325
Winsorize
Note the percentiles and enter them into the next slide.
326
Winsorize
Transform > Compute Variable
327
Winsorize
Choose a
sensible new
name
Select Old and New Values
328
Winsorize
Then Add
329
Winsorize
Then Add
330
Winsorize
Retain all
other values
Then Add
331
Winsorize
Finally, continue then OK
332
Winsorize
To check your results
Analyze > Descriptive Statistics

> Descriptives
333
Winsorize
OK
334
Winsorize
As desired.
335
Winsorize
Syntax:- freq var pulse /format = notable /percentiles = 5 95.
compute winsor = pulse.

if pulse <= 83 winsor = 83.
if pulse >= 137.25 winsor = 137.25.
descriptives variables=pulse winsor

/statistics=mean stddev min max.
336
Winsorize
Trimming and Winsorization: A review
W. J. Dixon and K. K. Yuen
Statistische Hefte
June 1974, Volume 15, Issue 2-3, pp 157-170
This paper provides a literature review for robust statistical

procedures trimming and Winsorization that were first proposed for
estimating location, but were later extended to other estimation and
testing problems. Performance of these techniques under normal and
long-tailed distributions are discussed.
337
Winsorize
Winsorisation for estimates of change
Daniel Lewis
Papers presented at the ICES-III, June 18-21, 2007, Montreal, Quebec, Canada paper
Outliers are a common problem in business surveys which, if left untreated, can have a
large impact on survey estimates. For business surveys in the UK Office for National
Statistics (ONS), outliers are often treated by modifying their values using a treatment
known as Winsorisation. The method involves identifying a cut-off for outliers. Any
values lying above the cut-offs are reduced towards the cut-off. The cut-offs are
derived in a way that approximately minimises the Mean Square Error of level estimates.
However, for many surveys estimates of change are more important. This paper looks at
a variety of methods for Winsorising specifically for estimates of change. The measure
of change investigated is the difference between two consecutive estimates of total.
The first step is to derive potential methods for Winsorising this type of change. Some
of these methods prove more practical than others. The methods are then evaluated,
using change estimates derived by taking the difference between two regular
Winsorised level estimates as a comparison.
338
Winsorize
Speaking Stata: Trimming to taste
Cox, N.J.
Stata Journal 2013 13(3) 640-666
Trimmed means are means calculated after setting aside zero or more
values in each tail of a sample distribution. Here we focus on trimming
equal numbers in each tail. Such trimmed means define a family or
function with mean and median as extreme members and are attractive
as simple and easily understood summaries of the general level
(location, central tendency) of a variable. This article provides a
tutorial review of trimmed means, emphasizing the scope for trimming
to varying degrees in describing and exploring data. Detailed remarks
are included on the idea's history, plotting of results, and confidence
interval procedures.
Note uses Stata! Index End 339

General Linear Models
Generally, the various statistical analyses are taught independently
from each other. This makes it difficult to learn new statistical
analyses, in contexts that differ. The paper gives a short technical
introduction to the general linear model (GLM), in which it is shown
that ANOVA (one-way, factorial, repeated measure and analysis of
covariance) is simply a multiple correlation/regression analysis (MCRA).
Generalizations to other cases, such as multivariate and nonlinear
analysis, are also discussed. It can easily be shown that every popular
linear analysis can be derived from understanding MCRA.
They present the identities shown on the next two slides.
General Linear Models: An Integrated Approach to Statistics

Sylvain Chartier and Andrew Faulkner
Tutorials in Quantitative Methods for Psychology 2008 4(2) 65-78
340
341
Index End
342
Centre Data
Sometimes to facilitate analysis it is necessary to centre a data set,
for instance to give it a mean of zero. In this case consider the
reading score in data set A.
343
Centre Data
Why center? Different authors have made different recommendations
regarding the centring of independent variables. Some have
recommended mean-centring (i.e., subtracting the mean from the value
of the original variable so that it has a mean of 0); others z-
standardization (which does the same, and then divides by the
standard deviation, so that it has a mean of 0 and a standard deviation
of 1); others suggest leaving the variables in their raw form. In truth,
with the exception of cases of extreme multi-collinearity (which may
arise in multiple-regression/correlation), the decision does not make
any major difference. For instance the p value for an interaction term
any subsequent interaction plot should be identical whichever way it is
done (Dalal and Zickar 2012; Kromrey and Foster-Johnson 1998).
Dalal, D. K. and Zickar, M. J. 2012 Some common myths about centering predictor variables in
moderated multiple regression and polynomial regression Organizational Research Methods, 15,
339362 DOI: 10.1177/1094428111430540
Kromrey, J. D. and Foster-Johnson, L. 1998 Mean centering in moderated multiple regression: Much
ado about nothing Educational and Psychological Measurement, 58, 4267 DOI:
10.1177/0013164498058001005 344
Centre Data
To check we have achieved our
goal we generate descriptive
statistics.
345
Centre Data
To check we have achieved our
goal we generate descriptive
statistics.
Syntax
EXAMINE VARIABLES=read
/PLOT BOXPLOT STEMLEAF
/COMPARE GROUPS
/STATISTICS
DESCRIPTIVES
/CINTERVAL 95
/MISSING LISTWISE
/NOTOTAL.
346
Centre Data
Note, the mean is 53.23.
347
Centre Data
To create a column of means.
348
Centre Data
To create a column of means.
Syntax
AGGREGATE
/OUTFILE=*
MODE=ADDVARIABLES
/BREAK=
/read_mean=MEAN(read).
349
Centre Data
Finally compute the desired variable.
350
Centre Data
Finally compute the
desired variable.
Syntax
COMPUTE read_centre=read - read_mean. 351

EXECUTE.
Centre Data
Repeat descriptive statistics on the new variable. Now zero mean.
Index End 352

Correlation - Comparison
If simply comparing two correlations, a simple web search should
reveal numerous analytic tools. Alternately see, for example,
Kanji, G.K. 1999 100 Statistical Tests London: SAGE.
Test 14 which includes a worked example.
353
Several procedures that use summary data to test hypotheses about
Pearson correlations and ordinary least squares regression
coefficients have been described in various books and articles. No
single resource describes all of the most common tests.
Furthermore, many of these tests have not yet been implemented in
popular statistical software packages such as SPSS. The article
(next slide) describes all of the most common tests and provide
SPSS programs to perform them. When they are applicable, the
code also computes 100 (1 )% confidence intervals corresponding
to the tests. For testing hypotheses about independent regression
coefficients, they demonstrate one method that uses summary data
and another that uses raw data (i.e., Potthoff analysis). When the
raw data are available, the latter method is preferred, because use
of summary data entails some loss of precision due to rounding.
354
SPSS and SAS programs for comparing Pearson correlations and
OLS regression coefficients
Bruce Weaver and Karl L. Wuensch
Behavior Research Methods

September 2013, Volume 45, Issue 3, pp 880-895
DOI: 10.3758/s13428-012-0289-7
For the code access. Not the earlier version on the journal web site.
355
Index End
Sobel Test
The Sobel test will assess whether a mediator (see more extensive notes on
Mediation on the main web page) carries the influence of an independent
variable to a dependent variable.
The Sobel test works well only in large samples. It is recommended using this
test only if the user has no access to raw data. If you have the raw data,
bootstrapping offers a much better alternative that imposes no distributional
assumptions. Consult Preacher and Hayes (2004, 2008) for details and easy-to-
use macros that run the necessary regression analyses for you:
Preacher, K. J., & Hayes, A. F. (2008) Asymptotic and resampling strategies
for assessing and comparing indirect effects in multiple mediator models
Behavior Research Methods, 40, 879-891 DOI: 10.3758/BRM.40.3.879.
Preacher, K. J., & Hayes, A. F. (2004) SPSS and SAS procedures for
estimating indirect effects in simple Mediation models Behavior Research
Methods, Instruments, & Computers, 36, 717-731 DOI: 10.3758/BF03206553.
See How can I perform a Sobel test on a single mediation effect in SPSS?
But its not simple! 356
Index End
Structural Equation Modelling
This tutorial begins with an overview of structural equation modeling (SEM) that
includes the purpose and goals of the statistical analysis as well as terminology
unique to this technique. I will focus on confirmatory factor analysis (CFA), a
special type of SEM. After a general introduction, CFA is differentiated from
exploratory factor analysis (EFA), and the advantages of CFA techniques are
discussed. Following a brief overview, the process of modeling will be discussed
and illustrated with an example using data from a HIV risk behavior evaluation
of homeless adults (Stein & Nyamathi, 2000). Techniques for analysis of
nonnormally distributed data as well as strategies for model modification are
shown. The empirical example examines the structure of drug and alcohol use
problem scales. Although these scales are not specific personality constructs,
the concepts illustrated in this article directly correspond to those found when
analyzing personality scales and inventories. Computer program syntax and
output for the empirical example from a popular SEM program (EQS 6.1; Bentler,
2001) are included.
Structural equation modeling: Reviewing the basics and moving forward

Ullman, Jodie B.
Journal Of Personality Assessment 87(1) 35-50 2006
DOI: 10.1207/s15327752jpa8701_03 357
Structural Equation Modelling
Stein, J. A., & Nyamathi, A. M. (2000). Gender differences in behavioural and
psychosocial predictors of HIV testing and return for test results in a high-
risk population. AIDS Care, 12, 343356. DOI: 10.1080/09540120050043007
Bentler, P.M. (2001). EQS 6 structural equations program manual. Encino, CA:
Multivariate Software.
Download Eqs 6 software
This page serves as a gateway to a tutorial on structural equation modeling or

SEM.
Index End 358

Does It Always Matter?
Scientists think in terms of confidence intervals they
are inclined to accept a hypothesis if the probability
that it is true exceeds 95 per cent. However within the
law beyond reasonable doubt appears to be a claim that
there is a high probability that the hypothesis the
defendants guilt is true.
A Story Can Be More Useful Than Maths

John Kay
Financial Times 26 February 2013
359
we slavishly lean on the crutch of significance testing
because, if we didnt, much of psychology would simply
fall apart. If he was right, then significance testing is
tantamount to psychologys dirty little secret.
Significance tests as sorcery: Science is empirical -

significance tests are not
Charles Lambdin
Theory and Psychology 22(1) 6790 2012
360
The first rule of performing a project
1 The supervisor is always right
The second rule of performing a project
2 If the supervisor is wrong, rule 1 applies
361
Does It Always Matter? Probably!
Estimation based on effect sizes, confidence intervals, and meta-
analysis usually provides a more informative analysis of empirical
results than does statistical significance testing, which has long
been the conventional choice in psychology. The sixth edition of the
American Psychological Association Publication Manual now
recommends that psychologists should, wherever possible, use
estimation and base their interpretation of research results on point
and interval estimates.
The statistical recommendations of the American Psychological

Association Publication Manual: Effect sizes, confidence intervals,
and meta-analysis
Geoff Cumming, Fiona Fidler, Pav Kalinowski and Jerry Lai
Australian Journal of Psychology 2012 64 138146
362
Does It Always Matter? Probably!
The debate is ongoing!
Cumming, G. 2014 The new statistics: Why and how. Psychological

Science, 25, 7-29 DOI: 10.1177/0956797613504966
Savalei V. and Dunn E. 2015 Is the call to abandon p-values the red
herring of the replicability crisis? Frontiers in Psychology 6:245.
DOI: 10.3389/fpsyg.2015.00245
Index End 363

SPSS Tips
Now you should go and try for yourself.
Each week our cluster (5.05) is booked for 2 hours

after this session. This will enable you to come and
go as you please.
Obviously other timetabled sessions for this module

take precedence.
364
Index

What Statistical Analysis Should I Use?: Sunday, June 4, 2017 04:22 AM

Uploaded by

Copyright:

Available Formats

What Statistical Analysis Should I Use?: Sunday, June 4, 2017 04:22 AM

Uploaded by

Document Information

Original Description:

Original Title

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

What Statistical Analysis Should I Use?: Sunday, June 4, 2017 04:22 AM

Uploaded by

Copyright:

Available Formats

What statistical analysis

Some criticism has been made of their discussion of p values,

These pages show how to perform a number of statistical tests

Menu selection:- Analyze > Compare Means > One-Sample T test

Note the test value of 50 has been selected

N Mean Std. Deviation Std. Error Mean

One-Sample Test students is 52.775, which is

Menu selection:- Analyze > Nonparametric Tests > One Sample

Choose customize analysis

Only retain writing score

Choose tests tick compare median and enter 50 as the desired

We would conclude that this group of students has a significantly

Two alternate approaches are available.

Syntax:- npar tests

Choose customize analysis

Only retain female

Choose tests tick compare observed and under options

enter .5 as the desired value.

Syntax:- npar tests

Select female as the test variable, the default test proportion is .5

The results indicate that there is no statistically significant difference

We hypothesized that the proportion of females is 50%. A two-tailed,

Syntax:- npar test

Syntax:- t-test groups=female(0 1)

Do not forget to define those pesky groups.

Levene, Howard (1960). "Robust tests for equality of variances". In Ingram

The results indicate that there is a

This is supported by the negative

Menu selection:- Analyze > Nonparametric Tests

Syntax:- npar test

Mann-Whitney test Crichton, N. Journal Of Clinical Nursing 9(4) 583-583 2000

It is often difficult, particularly when conducting research in psychology,

Where as a t test is a test of population means, the Mann-Whitney test

Mann-Whitney test is not just a test of medians: differences in spread c

As is always the case, it is not sufficient merely to report a p value. In

Note that Mann-Whitney has been selected.

Do not forget to define those pesky groups.

A Wilcoxon test was conducted to evaluate whether writing score was

A chi-square test is a common test for nominal (categorical) data. One

Two alternate approaches are available.

Drag selected variables to the row/column boxes

Menu selection:- Analyze > Descriptive Statistics > Crosstabs

Select row and column variables.

Select Chi-square and Cramers V

Note 0 cells have

A two-way contingency table analysis was conducted to evaluate

Menu selection:- Analyze > Tables > Custom Tables

A two-way contingency table analysis was conducted to evaluate

It may be introduced at the same time as the Chi-square.

Sometimes 2 is used as a measure of

Fisher's exact test for up to 66 tables

For those interested in more detail, plus a worked example see.

Fisher's Exact Test or Paper only

When to Use Fisher's Exact Test

Fisher's Exact Test

FEXACT - A Fortran Subroutine For Fishers Exact Test On Unordered

A Remark On Algorithm-643 - FEXACT - An Algorithm For

Menu selection:- Analyze > Compare Means > One-way ANOVA