I am trying to analyze the following data: participants chose over 18 trials whether they would like option 1 or option 2. Option 1 is recoded as 0, option 2 is recoded as 1. Now if I count all answers across 24 trials I for instance get '14', which means that in 14 cases they chose option 2 and in 4 cases option 1. Wilcoxon median test lets me compare with the hypothesized median 9, so far so good.

I have three groups (within subjects, three phases to the experiment). Now I need to know if the difference between the groups is significantly different from 0. So I create Group2-group1 and want to compare it to 0. and Group 3-group2 and compare it to 0. My advisor says to use chi square for this, but I cant find the correct way in SPSS. via non paramatric tests one sample it counts frequencies, not compares averages. via one sample non parametric test I can't insert '0' for some reason as test value.

Any ideas? my data is structured as follows

  • P1 GR1_X1 GR1_X2 GR2_X1 GR2_X2 GR3_X1 GR3_X2
  • SDB 0 1 0 0 0 0
  • tyc 1 0 0 1 1 0

.. and so forth. I created group scores but don't know how to continue.. I know I can't use a normal t-test because the data is binomial and not normally distributed.

  • $\begingroup$ With a dichotomous outcome, I wouldn't use Wilcoxon at all. ... Is the first paragraph its own experiment? If so it lends itself to counting the 1's out of the total count, and comparing that to probability of 0.50 with either a chi-square-goodness-of-fit test, or a binomial test ... I don't quite understand the set-up of the study in the second paragraph. The within-subjects nature might make McNemar or (between two groups) or Cochran's Q (among three groups) a viable option. $\endgroup$ Jan 7, 2020 at 17:43
  • $\begingroup$ The setup was as follows: people received three sets of 18 graphs in which they had to choose for either a statistical forecast or a judgmental forecast. SO I have a long list of 0s (judgmental choice) and 1s (statistical choice). Per person I can count these so they have a number that is maximum 18 per set (which would mean 18 choices of the statforecast)/ I now want to compare whether the difference between the total set scores is significant.. eg person 1 has score SET1 of 12, 13 in set2 and 9 in set 3. I'd say ttest but the data is not normal and I can't find how t o do it with spss.. $\endgroup$
    – ShariDB
    Jan 8, 2020 at 10:40

1 Answer 1


I also don't see how to fit this into a simple chi-square test. For a one-way repeated-measures model with counts or rates as the outcome, have a look at the GEE (generalized estimating equations) models available in the GENLIN procedure (Analyze>Generalized Linear Model>Generalized Estimating Equations in the menus), or generalized linear mixed models in the GENLINMIXED procedure (Analyze>Mixed Models>Generalized Linear) with Poisson loglinear models. For either of these, you need to set the data up with a subject identifier, a time or condition variable (1, 2, 3 for the levels, for example), and the count you're modeling. Some people refer to this data setup as univariate approach setup (as opposed to multivariate), some as narrow (as opposed to wide).

These models do really require decent sample sizes to provide stable estimates.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.