An Evaluation of a Residual Confounding Quantity of Propensity Score Stratification.

• Activity Number: 56 - Causal Inference
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Yoko Sato* and Eiji Nakatani and Seiichiro Yamamoto
• Companies: Shizuoka Graduate University of Public Health and Shizuoka Graduate University of Public Health and Shizuoka Graduate University of Public Health
• Keywords: Propensity score stratification; Average causal effect; Residual confounding; Gini index; ROC-AUC; Pearson correlation coefficient

Abstract

Propensity score (PS) stratification is potentially an excellent method for estimating the average causal effect (ACE) because it can use off-support cases without waste and is not affected by extreme weights under practical positivity violation. However, the problem of biased ACE inference due to residual confounding remains, and a method to evaluate residual confounding is required. We focused on three factors leading to residual confounding and scales to visualize them as follows: 1) an imbalance in the distribution of PSs within a stratum (visualized scale; Gini index), 2) an association of PSs within a stratum with treatment assignment (receiver operating characteristic area under the curve, ROC-AUC), and 3) an association of PSs within a stratum with outcome (ROC-AUC for binary outcomes, Pearson correlation coefficient for continuous outcomes). We proposed a residual confound quantity (RCQ) based on the within-stratum PS by converting the range of these scales to 0-1, calculating the sum, and averaging for each stratum. An evaluation of RCQ of PS stratification using the often-used quantile-based method, based on the right heart catheterization dataset, is provided.