I came across the following issue recently. I hope someone will be able to help me understand.

Say that I want to study the effect of a categorical variable X on my outcome y. X is a factor, and I use dummy coding with one level of X (say, x0) as a reference. It is a repeated measure design, where each subject is tested for all levels of the factor X. To proceed with the analysis, I fit a mixed-effect model with random intercept (grouped by subject-ID) using Maximum Likelihood estimation. Hence, the intercept will correspond to the level x0, and the fitting algorithm will estimate a specific intercept for each subject (in addition to the fixed effect of course).

There can be missing data. In particular, let's assume that for some subjects, the outcome y corresponding to the level x0 (ie reference level of X) is missing. How can the algorithm estimate the subject-specific intercepts for the missing values? If there is only one missing value, then the problem should be feasible: since random effects are assumed to be normally distributed, then their sum should be equal to zero, and this constraint helps finding unambiguous estimates. What happen if there is more than one missing value (associated to the level x0)? Doesn't the problem become ill-posed and have multiple solutions? The function lme seems to be able to deal with this (I can fit the model with no warnings and no errors), but how reliable will this model be?

Thanks a lot for your help!

A mixed effect model does not force the random intercepts to sum to zero, it rather assumes they have mean zero on the population level to be able to fit them.

Missing values in the outcome variable are indeed no problem for the inference of mixed effect models, as long as the missingness is not related to unmeasured confounders (Not missing at random, NMAR). The random intercept will be estimated from the measurements of the subject that are not missing, in the meanwhile sharing information with other subjects about the fixed effects.

  • Thanks! Still, doesn't the estimation problem become ill-posed? I have situations in which the predictions of the missing data-point do not make sense (when compared to the other subjects) – Cristiano Sep 29 '17 at 15:14
  • UPDATE: using 'nlm' and 'nlminb' optimizer I have a problem of singular convergence (as if I do not have enough data to run the fit); using 'optim' I can fit the model but some estimate just do not make sense. What does this mean? Why can 'optim' fit and how do I interpret such a fit? – Cristiano Sep 29 '17 at 15:21
  • @Cristiano Mixed effect models are notoriously hard to fit. Some approaches work better than others in some situations, often for unclear reasons. Can you give us some more information on the predictions that "do not make sense"? – Knarpie Oct 2 '17 at 12:21

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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