The goal of a propensity score model is to efficiently control confounding and not to predict treatment or exposure. The choice of variables to include in the propensity score model should be based on theory. As such, one should avoid using automated statistical procedures (e.g., stepwise methods). In your opinion, why is this important? Isn't the goal of any statistical model to parsimoniously predict outcome? Are there situations that you can think of where automated procedures may be acceptable?