Bias Amplification with Propensity Scores: A Reason for Parsimony?

One of the major arguments for the use of propensity score adjustment, rather than multivariable regression adjustment, in nonrandomized studies of exposures is that propensity score models can be estimated precisely when outcome events are rare but exposure is not. Therefore, there often is no upper limit to the number of covariates that may be included in the propensity score model, particularly in the context of large electronic healthcare datasets with hundreds of thousands of patients. However, recent work has shown that adjusting for some pre-treatment variables, specifically instrumental variables or variables that are associated with exposure but not with outcome except through exposure, can increase the bias of exposure effect estimates in the presence of residual confounding. In the research presented here, we consider the implications of these findings in the context of propensity scores, where many covariates may be adjusted for simultaneously, and some variables may be instrumental for the exposure-outcome association after conditioning on other variables in the model. In addition, we explore the potential for bias amplification in a study of hip fracture and mortality among patients initiating statins versus glaucoma medications, where prior glaucoma diagnosis is believed to be an instrument.