Mixed effect linear model with binary outcome - model and coding

[Robert Azzopardi]

Hi all,

Hoping for some advice re a model for this scenario:
Look at group of patients who either have disease A or dont - variable ptype = 0 (no A) or 1 (has A)
Patients evaluated at two visits giving baseline values and then follow up values for a series of outcomes being investigated - variable visit=0 or visit=1
Outcomes being various different types - binary, ordinal, catergorical and continiuous.
Potential confounding factors gender and diabetes I have also aimed to include in the model.
There are only ~60 patients equally split in terms of those having disease or not. I am treating patienst as the random effects component of the mixed effect model.

I'm wondering for a binary outcome 'Slow', in which patients are either deemed slow (1) or not (0) at both baseline and follow up, whether this code gives me the effect of having disease A from baseline to follow up:

 melogit slow i.visit##i.(ptype gender diabetes)|| study_id: , or

 lincom 1.ptype + 1.ptype#1.visit , or

Any help is much appreciated

[Joseph Coveney]

I'm wondering . . . whether this code gives me the effect of having disease A from baseline to follow up

I'm not quite sure that I understand what you mean by "the effect of having disease A" at the two observation time points, but you might be looking more for something like

lincom 1.ptype#1.visit, or

that is, the interaction term alone (without the main effects term for disease at baseline).

You've formed interaction terms involving visit also for patient's sex and diabetes mellitus status. Do you expect either of these to change over the observation interval? (Were you looking instead for a three-way interaction of each and disease A & visit?)

If "slow" is a sign of disease A (like Parkinson's disease), then do you risk selection effects? It would make the choice of comparison group curious, too.

Your post's title says, "Mixed effect linear model . . .", but you're fitting a mixed effects logistic regression model. Is that just a typo, or are you trying to fit a so-called linear probability model (all the predictors are categorical).

[Robert Azzopardi]

Thank you for your quick answer!

I've probably worded it poorly Joseph. The intention is to find out the effect of having disease A on outcome slow developing over time (ie. when visit equals 1 compared to visit equals 0). Or in laymens language does our outcome develop at a faster rate in our disease A group rather then our control group (with knowledge we only have time point 0 (visit=0) and time point 1 (visit=1)).

I dont expect sex or diabetes to change over observational period so will remove them from the interaction term. We are investigating whether slow developing is related to our disease. It is not a known sign of the disease.

It should have been a mixed effects logistic regression model (is a typo). I am also curious if this is okay to use if some individuals start off with the outcome present? Ie. some people start with slow=1 and either have slow=1 or slow=0 at visit=1 time point.