I understand that when applying GLMMs (e.g. in logistic mixed effects regression), the interpretation of the coefficients for the fixed effects is that they are also conditional on the random effects (e.g. a unit change in a predictor marks the log odds change in an outcome variable only for subjects with the same random effects). These are very likely quite different from the marginal change for the population. I am looking for a way to estimate and test this marginal change.

I read this answer and the related links. It is great, but the package used here is restricted to using a single grouping variable. My model has random effects both for subject and for item (subjects are measured over different items), so I believe it does not apply.

I would also be happy using GEEs if it can help me, but my specification includes also random slopes (and 2 grouping variables), so I think they do not apply either. Can someone recommend a way to estimate the marginal effects for such data? I am less interested in the conditional effects.

  • $\begingroup$ How many items do you have, and are you certain that you want to include them as a random effect? $\endgroup$ – Dimitris Rizopoulos Jan 1 '19 at 15:16
  • $\begingroup$ I have 6 or 12 items (two experiments). It makes a lot of sense to include them as random effects because they are really of little specific interest in by themselves, and are supposed to represent something much more general. $\endgroup$ – Cuenco Jan 1 '19 at 16:57
  • $\begingroup$ Having 6 or 12 items sounds too few to consider them as a random effect, even if they come from a pull of items. Often you need more than15-20 levels in your grouping variable to get a stable estimate of the variance across the different levels. $\endgroup$ – Dimitris Rizopoulos Jan 1 '19 at 17:51
  • $\begingroup$ Hmm... Do you have a reference for that? I learnt that 5-6 items are already enough. $\endgroup$ – Cuenco Jan 2 '19 at 6:31
  • $\begingroup$ Check here: stats.stackexchange.com/questions/7004/… . Of course, in mixed models you can gain information from the repeated measurements but still you need sufficient number of levels. $\endgroup$ – Dimitris Rizopoulos Jan 2 '19 at 7:23

Your Answer

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

Browse other questions tagged or ask your own question.