when to use repeated measures logistic regression

[rebekah shao]

Dear all,

Please could you help me to understand whether I need to use conditional logit/fixed effects logit models (https://www3.nd.edu/~rwilliam/stats3...xedEffects.pdf) for my research?

My data is repeated measures (same people over three time-points; T1, T2 and T3), however, I am planning on only comparing two time-points two each other at a time.

eg. Does T1 exposure predict T3 depression? CONTROLLING for T1 depression

example code: logistic exposure T3_depression T1_depression

another eg. Does T2 exposure predict T3 depression? CONTROLLING for T2 depression

Note: exposure is sometimes continuous, sometimes ordinal. Depression is binary.

Is it ok to still use normal logistic regression here even though I am using repeated measures data i.e. someones baseline depression at T1 and their depression at T3 in the same model? I think that if I wanted to use all three time points in one model then I would definitely need to use xtlogit....

Many thanks!!

[Marcos Almeida]

I think you can use the baseline value as a predictor. This is the principle, say, found in ANCOVA (I'm assuming Yvar is continuous, as you informed in the text). Also, you could use the difference between the time measurement and the baseline as the Y var. Or a proportion of that as well. No doubt, the model should go on a par with the options and assumptions.That said, this may be taken as "downgrading", for you have data which would fit better in a repeated measures model. Selecting just one time measurement, or changing the Y var from continuous to ordinal to binary...well, this could be viewed as a fishing expedition. That's the reason to specify the study design beforehand. Hopefully that helps.

[Clyde Schechter]

I think you mean
logistic T3_depression T1_exposure T1_depression, with T3_depression as the outcome, not as the predictor of T1_exposure.

That's fine.

In fact, -xtlogit- would be inappropriate for testing this hypothesis because the only outcome you are looking at is the T3 outcome. An -xtlogit- model would have to include the depression outcomes at each time and would basically tell you with exposure at any given time predicted depression at the same time. (Or you could modify it slightly to ask whether exposure at each time predicted depression at, say, the next time.) But you can't use -xtlogit- if there is only a single time period where the outcome variable is of interest.

Added: Crossed with #2. Marcos raises some good points about whether it is really appropriate to focus only on T3_depression as the outcome. I regard that as a content-issue that I'm not qualified to comment on. But I certainly agree that if you are just throwing different models at the data without prior specification of hypotheses, then you are either not doing science at all, or are just doing preliminary, exploratory, hypothesis generating analyses.

[rebekah shao]

Thank you both - that makes sense to me now why I would use logistic or xtlogit. And point taken re: different ways of modelling the exposure!