A1543
Title: Cross-cohort data integration for complex surveys with unobserved target outcomes
Authors: Danhyang Lee - Southern Methodist University (United States) [presenting]
Abstract: Large-scale surveys fielded at different time points often share predictors but differ in outcome availability, preventing direct inference for the more recent population. A transportability framework is developed for integrating repeated cross-sectional surveys under complex multistage designs when the target outcome is observed only in the source cohort. Under covariate shift, survey design weights are combined with density-ratio importance weights and a target-directed penalized pseudo-log-likelihood extending importance-weighted cross-validation to complex surveys is proposed, paired with a doubly robust AIPW estimator. When the outcome model itself may have shifted, bridge-calibrated estimation is introduced: shared intermediate outcomes observed in both cohorts proxy for drift in the target outcome relationship. A target-weighted penalized procedure selects informative bridges, and the resulting estimate is embedded in a sensitivity analysis yielding partial identification intervals anchored by observed bridge discrepancies. Simulations evaluate performance under varying covariate shift, outcome-model drift, and bridge informativeness. The framework is applied to Head Start FACES 2009 and 2014 cohorts, producing transportability-adjusted estimates of kindergarten math achievement.
