Hello,

I have panel patient dataset in which I am trying to predict invalidity (0/1) based on lagged variables. In one model, I use a single binary predictor lagged by 1 visit to predict invalidity in the following one. I predict in the training set and plot ROC curve in the test set. Using an example cross-sectional situation from my dataset, I get an ROC (rocgold command) curve with one single point, which is expected. However, using mixed-effect multilevel univariate logistic regression in the entirety of the panel dataset, even with a single binary predictor variable, I obtain a smooth ROC curve. Is it possible this palette of points on the sensitivity-specificity plane is due to the varying error terms? I am rather a beginner in statistics so it might be I am missing some basic thing.

Figure: The green and the yellow curves come from a binary predictor

Thank you

Jan