Keywords: propensity scores, bias, outcome data
Room: Wilson C
One of the primary benefits of propensity score analysis is the ability to design a non-randomized study in a way that incorporates covariates without making use of the outcome data for their selection, thus avoiding data-dependent biases. However, often there are too many covariates for the available data resulting in complete or semi-complete separation of the treatment groups. In this situation, it would be desirable to remove covariates that are not related to outcome as these would not be relevant to the analysis. We propose that this should be done in a structured way to avoid introducing unnecessary bias. Note, however, that there may be covariates which are both related to outcome and do not have sufficient overlap between the treatment groups. In this case, a meaningful comparison cannot be correctly made between the treatment groups.