ABSTRACT
Once we begin looking at the same patient on multiple occasions, we violate the regression assumption that all observations are independent. The technical name for this is pseudoreplication, and it can be addressed properly if you use statistical tests that account for multiple measurements from the same source. We discuss Lord’s paradox where you get different effects if you use change from baseline as an outcome rather than using baseline values as a predictor and post-baseline values as an outcome. We explain why it is inefficient to test each time point separately and recommend using mixed models.
