A taxonomy of treatment effects in data with two waves of measurement and a promotion of triangulation

It is common that researchers estimate the effect of a predictor (X) on a subsequent measure of an outcome (Y2) while adjusting for a prior measure of the outcome (Y1) and to interpret statistically significant effects to indicate increasing or decreasing effects of X on Y, even though this method is known to be susceptible to spurious findings. Here, we show that all combinations of null, positive, and negative estimated effects of X on Y2 when adjusting for Y1 and null, positive, and negative true effects of X on Y are possible. Hence, such adjusted effects, e.g., in cross-lagged panel models, should not be used for causal inference on their own. We recommend triangulation, where the effect of X on the Y2-Y1 difference as well as on Y1 when adjusting for Y2 are estimated in addition to the effect on Y2 when adjusting for Y1. Certain combinations of effects would corroborate (although never definitely prove) causal conclusions while other combinations would suggest that estimated effects may have been spurious and advise caution.