Online Program

Addressing Positivity Violations in the Causal Analysis of Big Data (305264)

Jessica Chubak, Kaiser Permanente Washington 
Rebecca Hubbard, University of Pennsylvania 
Nandita Mitra, University of Pennsylvania 
Jason Roy, Rutgers University 
*Yaqian Zhu, University of Pennsylvania 

Keywords: overlap, trimming, weighting, extrapolation, propensity score, generalizability

In the causal analysis of observational data, the positivity assumption requires that all treatments or exposures of interest be observed in every subgroup. Violations of this assumption are indicated by nonoverlap in the sense that subjects with certain covariate combinations are not observed to receive a treatment of interest, which may arise from contraindications to treatment or small sample size. Here, we emphasize the importance and implications of this often-overlooked assumption for obtaining valid causal inference. We elaborate on challenges nonoverlap poses to estimation and discuss methods to address it including propensity score-based matching, trimming, weighting, and extrapolation. We distinguish between structural and practical violations and provide insight into the suitability of each approach in the context of study objectives. To demonstrate alternative approaches and relevant considerations (including how overlap is defined and the target population to which results may be generalized), we employ an electronic health record-derived data set from a pharmacoepidemiologic study to assess the effects of metformin on colon cancer recurrence among diabetic patients.