Well this debate is in the context of perfectly measured regressors and satisfaction of no unmeasured confounders. Neither method performs generally better than the other when these conditions aren't met, and both methods can equally accommodate customizations that address those issues.