I read with great interest this article by Nowacki et al.
First off, I thank the authors for demonstrating the limitations of propensity scores in prediction modelling and highlighting important misconceptions. The results presented confirm that addition of a propensity score does not improve model concordance over simply adjusting for the factors included in the score.
Another potential advantage of propensity scores would be to remove individuals with extremely high or low treatment probabilities. For example, in the surgical site infection example, one could feasibly remove individuals with outlying probabilities of receiving open surgery.
Even outside the context of causal effect modelling, where one is focusing on treatment, it would be interesting to examine how removal of outliers and addition of the score would affect prediction. Did the authors consider using propensity scores in this way, either through matching or a simple cutoff?