The peril of adjusting for baseline when using change as a predictor

Some studies have analyzed the effect of a predictor measured at a later time point (X1), or of the X1-X0 difference, while adjusting for the predictor measured at baseline (X0), on some outcome Y of interest. The present simulation study shows that, if used to analyze the effect of change in X on Y, there is a high risk for this analysis to produce type 1-errors, especially with a strong correlation between true X and Y, when X0 and X1 are not measured with very high reliability, and with a large sample size. These problems are not encountered if analyzing the unadjusted effect of the X1-X0 difference on Y instead, and as this effect exhibits power on par with the adjusted effect it seems as the preferable method when using change between two measurement points as a predictor.