At this moment, I am busy running a Generalized Estimating Equations model in SPSS. Unfortunately, I cannot use an ordinary logistic regression as the conditions are repeated measures.

I asked people to indicate whether they should click on a search engine result. These results were manipulated by position (low = 0, high = 1), description (short = 0, long = 1) and type of result (non-sponsored = 0, sponsored = 1). Click intention is measured by either clicking (1) or not clicking (0) on the result. These were added to the model, including two interactions terms (PositionType and DescriptionType).

Underneath the parameter estimates of the GEE.

enter image description here

I know it's important to look at the interaction terms first: position*type seems to be significant. However, I would like to make a chart where I point out the differences in position and description by type. Unfortunately, I do not have a clue where to start and how to interpret these numbers.

  • 1
    $\begingroup$ The trick to understanding GEE is that what it estimates is the same as what a linear model would estimate. If the response is binary and you are using a logit model, the output can be interpreted just like a logistic regression. Is that what you are asking? $\endgroup$
    – AdamO
    Apr 24, 2018 at 17:32
  • 1
    $\begingroup$ There are many ways of handling repeated measures with logistic regression that are better than GEE. Look at Markov models and mixed effects models for starters. See hbiostat.org/proj/covid19 $\endgroup$ Dec 2, 2021 at 16:13

1 Answer 1


I originally had trouble with this too but here is what is going on. First, for some reason instead of decimal places your output has commas, not sure why that is happening but you can still interpret it. All of the betas are part of a regression equation, however because you are using binary data the program cannot solve it without a reference group. So SPSS chose 1 as your reference group for everything. As such, if the main effect or interaction has a 1 in it your beta will be zero. If you run the estimated marginal means for the model you will notice the marginal mean is the same as the intercept. To calculate all other marginal means you just have to add the betas to the intercept as in a regular regression model, this will give you the estimated marginal means. As far as interpretation of the betas alone this is the same as in a regression model. For example, position high is the referent as such with a negative beta weight people that were in the low position were less likely to click on the search engine results. One thing to be careful of is to look at the Test of Model Effects for your categorical variables. This will tell you if the interaction was significant (similar to an ANOVA looking at the interaction effects then looking at the simple effects within the interaction). Hope this helped.

  • 3
    $\begingroup$ Gene a, in many parts of mainland Europe, commas and periods are swapped when presenting numbers (i.e. commas are used to denote decimal places less than 1 and periods are used to separate digits in large numbers) $\endgroup$
    – TPM
    Jan 26, 2018 at 19:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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