Propensity Score Matching for Multiple Treatment Comparisons in Observational Studies

s A major limitation of making inference about treatment effect based on observational data from a non-randomized study designs is the treatment selection bias, in which the baseline characteristics of the population under one treatment could dramatically differ from the other one. If not handled properly, such sources of heterogeneity will introduce confounding effects into a causal-effect relationship and result in bias in the estimation of treatment effect. The Propensity Score (PS) method is one of the approaches that have been widely used in practice to correct this selection bias through balancing the observed patients’ characteristics among treatment groups. Until recently, the PS method has been applied exclusively for 2 treatment comparison settings (e.g. treatment vs. control) despite that it is frequently of interest to compare more than 2 treatments or interventions in medical and cancer research. PS covariate adjustment, inverse probability weighting (IPW) estimator, and PS matching are the three PS approaches commonly seen in two treatment comparisons, and among them, PS matching has been shown to have the greatest potential to eliminate the imbalance among covariates. However, not all of them are ready to be applied in the comparison of more than 2 treatments, especially for PS matching. To the best of our knowledge, we have not seen any such extension. In this study, we filled the gap and proposed an analytical approach to generalize PS matching for multiple (>= 2) treatments comparisons. This study was motivated by the desire to address comparisons of no adjuvant therapy, adjuvant chemotherapy alone and chemoradiation therapy in resected pancreatic adenocarcinoma (rPAC) patients in a recent data analysis based on the National Cancer Data Base (NCDB). We present the proposed method and illustrate it in the above case study as well as compare it with other two PS approaches.