I'm doing a meta-analysis which involves some multiple logistic regression. As it's a meta-analysis, I'm compiling data from various studies, which therefore differ slightly in their methodology. One particular aspect of methodology is likely to affect the response variable, and I therefore decided to include it as a random effect in my models.

However, now I'm being asked to provide more information about the likely effect of this predictor. Is it significant? How strong is the effect? The only way I can think to do so is to treat it as a fixed effect and see if adding/removing it improves the model or not, using F tests to compare models (they are quasi-binomial models). However, I had a feeling that treating variables as both random and fixed variables was wrong - it should be one of the other. I've never quite got a grip on the difference between the two, and would appreciate any advice. Thanks

edit: Here's a bit more info about my study. In my meta-analysis I'm comparing the results of other studies to my own study. The response in the model is the similarity in the data. Regarding this particular aspect of methodology, I used "method A". If all the other studies used "method A" it would be fine, but some used "method A", some "method B", some "method C" and some "method D". Each different method will introduce bias, but may do so to a different degree. Hope that makes sense.


1 Answer 1


It sounds as if your variable should be modeled as a fixed effect anyway, since you would presumably expect this particular aspect of methodology to have the same effect on results in all studies where it was applied.

The answers to this question are quite good: What is the difference between fixed effect, random effect and mixed effect models?

  • $\begingroup$ Thanks for your response. There are different categories of this methodology - some categories would be expected to have more of an effect than others (see extra info in the Edit above). Do you still think fixed effect is most appropriate? $\endgroup$
    – rw2
    Commented Dec 8, 2015 at 21:20
  • 1
    $\begingroup$ Yes. You introduce a categorical variable for the method, the parameter estimate on each category in that variable is the estimate of the bias relative to your baseline method. $\endgroup$
    – BjaRule
    Commented Dec 9, 2015 at 22:12

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.