I am trying to run a linear mixed effects model on some longitudinal data (2 sessions) using 3dLMEr. Participants did both sessions, though some participants are missing one of the sessions due to high movement etc. As this is a preliminary analysis, my n is quite small (~ 40). I tried to fit the model with random intercepts for Subject to account for the repeated measures (1|Subj) -- this was the only random effect I included in the model. Unfortunately, I'm getting the message
boundary (singular) fit: see ?isSingular
I was wondering whether it would be more appropriate to remove the Subject random intercept from the model, or to leave it in with the caveat that the model fit is singular. I feel wary of running a model on this kind of paired/related data without accounting for Subject with a random intercept, but I also don't want to the model to be grossly overfitted.
Alternatively, is there another analysis approach that could handle the repeated measures structure without having this issue (or is it kind of a moot point and the only way to fix this is to increase my n)?
The warning alone doesn't necessarily indicate a big issue. Unusual data, such as voxels near the brain boundaries, could trigger it. Could you share the model specification line from your 3dLMEr script?
the design was within-subject so all participants experienced both levels of Session (though some have missing data in the actual model because of quality), TwoAG.cen is a (centred) continuous variable, and Motion is the number of TRs censored.
The difference between these approaches might be small. The reason for these options is briefly discussed in this blog post, and explained in more detail in this manuscript.
Does the variable Motion vary across sessions? Consider whether it is appropriate to include it as a covariate based on the provided material.
Gang Chen
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.