Improving Propensity Score Methods in Pharmacoepidemiology

Inferences about intended effects of treatments are ideally investigated using randomized control trials (RCTs). Randomization leads to comparability of treated and comparison groups on measured and unmeasured prognostic factors, except the treatment, thereby preventing confounding. RCTs, however, may not always be feasible or necessary, for example, in drug-safety research involving unpredictable adverse events. However, when adverse events are related to the main effects of the therapy, patient’s prognosis and the potential for adverse events guide prescribing behavior leading to systematic differences in prognostic factors between treated groups. Hence, confounding by (contra-)indication is a threat to internal validity of the study results. Therefore, inferences from non-randomized studies require that the studies be designed like RCTs (e.g., using propensity score (PS) matching or weighting, i.e., design and analysis are apart) or confounders are accounted for (PS stratification or covariance adjustment, i.e., analysis stage) or a combination of the two (PS matching/weighting combined with covariate adjustment). The PS is the probability of receiving a particular treatment conditional on measured covariates, i.e., patient’s treatment in daily practice is predicted using prognostic factors. PS methods can be considered the observational study analogues of randomization in RCTs although PS methods attempt to balance only measured factors. Under the assumption of no unmeasured confounding given the PS, PS methods help researchers design and analyze observational studies in a way that mimics RCTs. We studied different aspects of the PS methods in pharmacoepidemiology. First, we focused on covariate selection, assessment as well as reporting of covariate balance using different balance metrics in PS analysis. We provided a checklist for reporting PS analysis and recommended the standardized difference for measuring and reporting covariate balance, PS model selection, and choosing optimal caliper width in PS matching. Next, we demonstrated marginal structural models, whose parameters are estimated using inverse probability of treatment weights, when treatment is time-varying and confounding is time-dependent using clinical examples. Results were compared with those of conventional time-varying Cox regression and PS methods. Further, the assumptions underlying PS methods, particularly of unmeasured confounding, and advantages of PS methods, compared to regression approaches in pharmacoepidemiology were discussed. Moreover, We extended the use of balance measures for quantitative falsification of instrumental variables assumptions thereby helping researchers’ decision on whether to proceed with or refrain from instrumental variable analysis. Finally, we provided a guidance document comprising a step-by-step description of the PS methodology and interpretations of effect estimates from different PS approaches. In conclusion, we argue that PS methods are invaluable tools for estimating treatment effects using observational data; their use is optimal when combined with model-based adjustment. In addition to initial designing of the study, full specification of all analyses performed is required for utilizing full advantage of the PS methods. Furthermore, adequate reporting of aspects of the PS analysis is as crucial as the analysis itself, since readers rely on this information for better appraisal of the quality of the study and validity of the results. Hence, critical items of PS analysis should be incorporated in guidelines on the conduct and reporting of observational studies, such as the STROBE statement and the ENCePP guide on methodological standards in pharmacoepidemiology to improve the quality of conduct and reporting PS based studies.