Wilcoxon test in SPSS (Practical)

Centre for
Multilevel
Modelling
The development of this E-Book has been
supported by the British Academy.

Wilcoxon Sign Rank Test practical

In this practical we are going to investigate how to perform a Wilcoxon test using SPSS. A Wilcoxon test is used when we have two interval
or ratio level variables measured for a set of observations and we want to test if the distribution is different for the two variables but we are
unable to assume normality for one or both of the variables. It can also be to compare ordered categorical variables. It is the non-
parametric equivalent of the paired t-test but unlike the t-test it tests differences in the median rather than the mean. The test does not
assume any distribution for the two variables. To run a single test in SPSS requires that your dataset has two column containing the two
variables to be compared.

The 2015 PISA survey collected information on students' degree of interest in a range of science-related topics. In this example, we will test
whether 15 year olds in England are more interested in learning about topics related to the universe and its history (INT_UNIV) or related to
how science can help prevent disease (INT_DISEASE). Students were asked to rate their interest in both topics on a four-point Likert scale,
with responses ranging from Not interested to Highly interested. Both measures are ordinal variables and each participant contributes a
pair of scores to the data, so a non-parametric test of difference is an appropriate method to use to explore differences in the distribution
of responses on the two topics.

Before we can perform this test we need to check whether the differences between INT_UNIV and INT_DISEASE are normally distributed.
First we need to create a difference variable as follows:

Select Compute from the Transform menu.

Type DIFF_INT_UNIV_INT_DISEASE into the Target Variable box.
Type INT_UNIV - INT_DISEASE into the Numeric Expression box.
Click on the OK button.

We can now use this new generated variable to perform normality checks. Do this as follows:

Select Descriptive Statistics from the Analyze menu.

Select Explore from the Descriptive Statistics sub-menu.
Click on the Reset button.
Copy the DIFF_INT_UNIV_INT_DISEASE variable into the Dependent List: box.
Click on the Plots... button.
On the screen that appears select the Histogram tick box.
Unselect the Stem and leaf button.
Select the Normality plots with tests button.
Click on the Continue button.
Click on the OK button.

We will first look at a histogram of the variable, DIFF_INT_UNIV_INT_DISEASE. This can be found in amongst the set of output objects
and looks as follows:
Ideally for a normal distribution this histogram should look symmetric around the mean of the distribution, in this case .0152. This
distribution appears to be significantly skewed to the left (negatively skewed).

We will next look at a statistical test to see if this backs up our visual impressions from the histogram.

The Kolmogorov-Smirnov test is used to test the null hypothesis that a set of data comes from a Normal distribution.

Tests of Normality

Kolmogorov-Smirnova Shapiro-Wilk

Statistic df Sig. Statistic df Sig.

DIFF_INT_UNIV_INT_DISEASE .253 4730 .000 .903 4730 .000

a. Lilliefors Significance Correction

The Kolmogorov Smirnov test produces test statistics that are used (along with a degrees of freedom parameter) to test for normality. Here
we see that the Kolmogorov Smirnov statistic takes value .253. This has degrees of freedom which equals the number of data points,
namely 4730.

Here we see that the p value (quoted under Sig. for Kolmogorov Smirnov) is .000 (reported as p < .001) which is less than 0.05. We
therefore have significant evidence to reject the null hypothesis that the variable follows a normal distribution.

Although the Kolmogorov Smirnov statistic tells the researcher whether the distribution followed by a variable is statistically significantly
different from a normal distribution one should take care in not overinterpreting such findings. Significance will be strongly effected by the
number of observations and so only a small discrepancy from normality will be deemed significant for very large sample sizes whilst very
large discrepancies will be required to reject the null hypothesis for small sample sizes.

SPSS also supplies QQ plots to assist in looking at normality but for brevity we do not show them here.
We will next move on to the Wilcoxon test itself and will test for a difference in distribution between the two variables, INT_UNIV and
INT_DISEASE.

Below you will see instructions on how to perform the Wilcoxon test in SPSS. If you follow the instructions you will see the two tabular
outputs that are embedded in the explanations below.

Select Non Parametric Tests from the Analyze menu.

Select Legacy Dialogs from the Non Parametric Tests sub-menu.
Select 2 Related Samples... from the Legacy Dialogs sub-menu.
Click on the Reset button.
Copy the Interest in universe and its history[INT_UNIV] variable into the Test Pairs: box.
Copy the Interest in how to prevent disease[INT_DISEASE] variable into the Test Pairs: box.
Click on the Exact... button.
On the screen that appears select the Exact button.
Click on the Continue button.
Click on the OK button.

The first SPSS output table contains a summary of the rankings for the difference variable. Here observations are split into three types
depending on whether the value of INT_UNIV is bigger than INT_DISEASE (negative ranks), the value of INT_DISEASE is bigger than
INT_UNIV (positive ranks), and finally where both variables take the same value (ties). These can be seen below:

Ranks

N Mean Rank Sum of Ranks

Interest in how to prevent disease - Interest in universe and its history Negative Ranks 1265a 1171.67 1482167.00

Positive Ranks 1142b 1239.81 1415861.00

Ties 2323c

Total 4730

a. Interest in how to prevent disease < Interest in universe and its history

b. Interest in how to prevent disease > Interest in universe and its history

c. Interest in how to prevent disease = Interest in universe and its history

The Wilcoxon test works by firstly assigning a sign (or a tie) to the difference between each pair of observations. Here we have worked on
INT_DISEASE - INT_UNIV so that positive ranks are when INT_DISEASE > INT_UNIV. Here we see that there are 1265 negative ranks,
1142 positive ranks and 2323 ties.

Having worked out which observed pairs result in which sign for their difference, the magnitude (excluding the sign) of these differences is
calculated and these are then ranked in order (excluding ties). We now see that the total of the ranks for the negative differences is
1482167.00 resulting in a mean rank of 1171.67 whilst the total of the ranks for the positive differences is 1415861.00 resulting in a mean
rank of 1239.81. Here the mean of the positive ranks is larger than that for negative ranks suggesting that values for INT_DISEASE are
generally larger than for INT_UNIV.
The Wilcoxon test will now decide whether this difference in mean ranks is significant or not as is illustrated in the second table.

The second SPSS output table contains details of the test itself and can be seen below:
 Test Statistics

Interest in how to prevent disease - Interest in universe and its history

Z -1.018b

Asymp. Sig. (2-tailed) .309

Exact Sig. (2-tailed) .309

Exact Sig. (1-tailed) .154

Point Probability .000

b. Based on positive ranks.

The output here consists of test statistics and their significance as calculated in several ways. We are considering the Wilcoxon statistic
which is calculated from the ranks and is not shown explicitly by SPSS but is used to calculate a Z score. Here we see that Z = -1.018 and
this can be compared with a standard normal distribution to test whether there are significant differences between the groups.

Here we see that the p value (quoted next to Asymp. Sig. (2-tailed)) is .309 which is greater than 0.05 and therefore we cannot reject the
null hypothesis that the medians of the two groups are the same. The normal approximation used above is only an approximation to the p
value and it is possible to construct the exact p value. This is given in the next row and we see that the exact p value is .309 whilst the
asymptotic p value is .309. The exact p value agrees with the asymptotic p value that the null hypothesis cannot be rejected. For
completeness the table also gives a p value for a 1-sided test and a point probability but we will ignore these here.
In conclusion, we could report this to a reader as follows:

A comparison of the mean of the distribution of the variables INT_UNIV and INT_DISEASE was desired but due to the non-normality of
the variables a Wilcoxon signed rank test was carried out. The mean of the positive ranks is larger than that for negative ranks suggesting
that values for INT_DISEASE are generally larger than for INT_UNIV. The Wilcoxon Signed rank test results in a Z statistic of -1.018 which
results in an exact p value of .309. This is not significant and we cannot reject the null hypothesis of equal medians for the 2 variables.
The results suggest that students are equally interested in learning about topics related to the universe and its history and those related to
how science can help prevent disease. It does not appear that, overall, students in England are drawn more to one of these types of
scientific inquiry.