Liu & Wu have developped the macro %GLIMMROC in SAS (http://ferran.torres.name/download/shared/roc/roc.pdf) for the estimation of the area under the ROC curve for repeated measures.

There are R package for censored data that should be used for longitudinal data (survAUC, survivalROC etc.)

Thank you both for the suggestions :)

This really does depend on the question you are asking. One way of thinking about this is that you will produce a one dimensional summary (either explicitly or implicitly) of the multivariate (i.e. repeated measures) response.

That summary might be the biomarker at a particular time point, a trend in the biomarker, the asymptotic level of the biomarker or it could even be a (non) linear combination of the time points chosen to maximise some index of group separation (perhaps even the ROC area).

Or is your problem that you are concerned about the multiplicity issues with repeated testing (i.e. at each time point leading to type I error inflation)? If the latter, there are a number of strategies you can adopt.

If you post more details we can give a more helpful response.

Thank you Mervyn. The ultimate goal is to obtain the best estimate reflecting the expression of the biomarker (BM) level over a period of 2-3 weeks measured repeatedly for each subject. This estimate, then, could be used in the ROC analysis combined with a binary outcome of interest.. You pose interesting questions. To date I looked at the change from baseline for last BM measurement, change from baseline using an averagr for the last 3 measurements, and an overall average while on treatment change from baseline. I have several BMs so I also performed FDR analysis to select the top candidates for the ROC analysis. Thank oyu for your suggestions and continued support