Bayesian Propensity Score Matching 

 Many people have realized that conventional propensity score matching (PSM) method does not take into account the uncertainties of estimating propensity scores. In other words, for each observation, PSM assumes that there is only one fixed propensity score. In contrast, Bayesian methods can generate a sample of propensity scores for any observation, by either monitoring the posterior distributions of the estimated propensity scores directly or predicting propensity scores from the posterior samples of the parameters of the propensity score model. 

 Then matching on thus obtained propensity scores, we should expect to get a distribution of estimated treatment effects. This will also provide us with an estimation of the standard error of the treatment effect. The Bayesian S.E. will be larger than the S.E. based on PSM estimate, as it takes into account more uncertainties. This conjecture is indeed confirmed by a recent paper written by Lawrence C. McCandless, Paul Gustafson and Peter C. Austin, "   Bayesian propensity score analysis for observational data   ", which appears in   Statistics in Medicine (2009; 28:94-112)  . The authors show that, the Bayesian 95% credible interval for the treatment effect is 10% wider than conventional propensity score C.I.  

 It seems that we should expect Bayesian propensity score matching (BPSM) perform better than PSM in cases where there are a lot of uncertainties in estimating the propensity scores. Before running into any simulations, however, the question is: what are the sources of the uncertainties in estimating propensity scores? From my point of view, there is at least one source of uncertainties, the uncertainties due to omitted variables. I do not think BPSM can do any better than PSM in solving this issue. But maybe, BPSM can model the error terms and so provide better estimations of the propensity scores? The above authors argue that when the association between treatment and covariates is weak (i.e., when the betas are smaller), the uncertainties in estimating propensity scores are higher. Weak association means smaller R-square or larger AIC, etc. Is this equivalent to larger bias due to omitted variables? 

 Another type of uncertainty related to BPSM, but not to propensity scores, is the uncertainty due to matching procedure. This is avoidable or negligible. Radically, we can just abandon the matching method and resort to linear regression model to predict the outcomes. Or we can neglect the bias from matching procedure, because when we only care about ATT and there is sufficient number of control cases, the bias is negligible, according to Abadie and Imbens 2006. ( "Large Sample Properties of Matching Estimators for Average Treatment Effects." Econometrica 74 (1): 235 - 267. ) 

 Of course, the logit model for the propensity scores could be wrong as well. But this can be manipulated in the simulations. Now my question is: how should we do the simulations to evaluate the performance of BPSM vs. that of conventional PSM?