Exposure definition sensitivity unmasks hidden confounding in crystalloid target trial emulation

Target trial emulation (TTE) is increasingly used to estimate causal effects from observational data. We tested whether satisfying standard diagnostics is sufficient by emulating a balanced crystalloid versus 0.9% saline trial in Medical Information Mart for Intensive Care IV (MIMIC-IV) (n = 42,883) with eICU Collaborative Research Database (eICU-CRD) external validation. Two pre-specified exposure definitions-initial assignment and 48-h dominant strategy-yielded directionally opposite major adverse kidney events within 30 days (MAKE-30) estimates (odds ratio [OR]: 0.49 vs. OR: 2.51) despite satisfactory propensity score diagnostics (C-statistic: 0.787; maximum standardized mean difference [SMD]: 0.048). Treatment switching reached 89.4%, reflecting local fluid workflows. Pre-specified negative control outcomes flagged residual confounding but were length-of-stay dependent. The pattern was absent in eICU-CRD, where recorded switching was 1.2%. Standard TTE diagnostics did not detect confounding from time-varying institutional practice. We propose that operationalization sensitivity, switching characterization, and directed acyclic graph (DAG)-guided assessment become routine TTE components.