A1297
Title: Prognostic diagnostics of positivity violation in causal inference for longitudinal multilevel data
Authors: Huixia Savannah Wang - Umeå University (Sweden) [presenting]
Abstract: Positivity is a fundamental assumption in causal inference, requiring that all individuals have a nonzero probability of receiving each exposure level given their covariates. While the propensity score is commonly used to assess this assumption through balancing covariates with respect to exposure, the prognostic score instead summarizes covariate associations with potential outcomes. Prognostic score adjustment reduces outcome variation between groups, leaving only variation due to the exposure effect and random noise. Although the prognostic score is naturally aligned with the estimation of the average exposure effect under no exposed group, its use is extended as a diagnostic and trimming tool to support the estimation of the average exposure effect in a longitudinal setting. The approach applies prognostic score adjustment to ensure comparability, then fits a random-intercept Bayesian additive regression tree (riBART) model for flexible outcome modeling, and finally estimates the average exposure effect via g-computation. This implementation helps mitigate challenges arising from the positivity violation. The approach is further applied to SHARE data, examining the causal effect of depression on human cognitive function.
