The goal of BACps is to estimate the average causal effect accounting for two sources of uncertainty:

• uncertainty regards the propensity score. The propensity score is a quantity that we do not observe and thus we have to estimate. So, the idea is to account for the fact that we are not using the true propensity score and thus we can be making a mistake

• uncertainty regards the model. This is related to the uncertainty that we face when we decide the variables that we include in the model. Instead of fixing one model associated with one set of covariates/features, we consider all possible models.

If every model is given the same weights, most of the time instrumental variables will be included in the propensity score since instrumental variables are associated to the exposure variable. However, the literature has shown that including these variables might increase the variance and amplify the bias of the estimate (Pearl 2010 and Brookhart et al. 2006). The strategy that we apply to limit the instrumental variables in the propensity score model is to incorporate a informative prior. The informative prior over the model indicator of the propensity score play the role to shape the posterior model distribution of the propensity score so that this posterior gives less weights to models that include instrumental variables and more weights to those models that include confounders and predictors of the outcome variable.

You can install the package version from GitHub with:

# install.packages("devtools")
devtools::install_github("pablogonzalezginestet/BACps")

See the vignette for details: online vignette

Brookhart MA, Schneeweiss S, Rothman KJ, Glynn RJ, Avorn J, Stürmer T. Variable selection for propensity score models. Am J Epidemiol. 2006 15;163(12):1149-56.

Pearl, J., P. Grunwald, and P. Spirtes. “Proceedings of the Twenty-Sixth Conference on Uncertainty in Artificial Intelligence (UAI 2010).” (2010): 417-24.