WHEN NOT TO TRUST RWE: POSITIVITY-AWARE CAUSAL INFERENCE AND RCT-CALIBRATED TRANSPORTABILITY OF HYDROCORTISONE IN SEPTIC SHOCK FROM A US ICU DATABASE TO ASIAN CRITICAL CARE
Author(s)
Nikhil Tiwari, Bachelors1, Shivesh Gupta, Bachelors2.
1Frekil, San-Francisco, CA, USA, 2Frekil, San Francisco, CA, USA.
1Frekil, San-Francisco, CA, USA, 2Frekil, San Francisco, CA, USA.
OBJECTIVES: Two landmark trials disagree on adjunctive hydrocortisone in septic shock (ADRENAL, null; APROCCHSS, benefit), and observational analyses risk amplifying confounding by indication. Using overlap weighting as a positivity-robust estimand, we estimated the 90-day mortality effect, calibrated it against both trials, and tested transportability to Asian ICUs.
METHODS: Retrospective cohort of 5,713 adults with Sepsis-3 vasopressor dependent septic shock (MIMIC-IV v3.1); exposure = IV hydrocortisone within 24h (24-hour landmark). A pre-specified DAG informed a 23 covariate gradient-boosted propensity model. After positivity diagnostics, 90-day risk difference (RD) and RMST were estimated under IPTW, overlap weights (ATO), ATT, PS-matching, and equipoise restricted cohorts, with negative-control outcomes and E-values. A weighted effect-modifier surface was calibrated to ADRENAL/APROCCHSS profiles; matching-adjusted indirect comparison (MAIC) transported the effect to four Asian cohorts with effective-sample-size (ESS) gating.
RESULTS: IPTW suggested harm (RD +8.8%) but extrapolated beyond support (treated ESS 17%; 15/31 covariates imbalanced). Pre-specified as primary under this positivity violation, overlap weighting (ATO) achieved full balance (0/31 SMD>0.1; ESS 81%) and gave a non-significant null in the overlap population (RD +4.5%, 95%CI −1.8 to +10.9; HR 1.169). Calibrated effect-modifier predictions were directionally concordant with both trials (ADRENAL HR 1.045 vs 0.95; APROCCHSS 0.872 vs 0.88), nominating respiratory source and vasopressor dose, not SOFA as candidate modifiers; negative controls indicated ~70% residual-bias reduction. MAIC to four Asian cohorts suggested benefit (RD 4.1% to 11.9%) but combined ESS 1.8-3.0% triggered an automatic "exploratory-only" flag.
CONCLUSIONS: Standard IPTW produced qualitatively wrong harm signal that positivity diagnostics and overlap weighting reversed into a trial-concordant null, showing identical data can yield opposite conclusions without positivity checks. With these safeguards, observational data recovered the ADRENAL-APROCCHSS severity gradient and localized hydrocortisone's benefit to refractory, respiratory-source shock. Positivity diagnostics, RCT calibration, and ESS gating should be prerequisites before transporting Western RWE to Asian ICU decisions.
METHODS: Retrospective cohort of 5,713 adults with Sepsis-3 vasopressor dependent septic shock (MIMIC-IV v3.1); exposure = IV hydrocortisone within 24h (24-hour landmark). A pre-specified DAG informed a 23 covariate gradient-boosted propensity model. After positivity diagnostics, 90-day risk difference (RD) and RMST were estimated under IPTW, overlap weights (ATO), ATT, PS-matching, and equipoise restricted cohorts, with negative-control outcomes and E-values. A weighted effect-modifier surface was calibrated to ADRENAL/APROCCHSS profiles; matching-adjusted indirect comparison (MAIC) transported the effect to four Asian cohorts with effective-sample-size (ESS) gating.
RESULTS: IPTW suggested harm (RD +8.8%) but extrapolated beyond support (treated ESS 17%; 15/31 covariates imbalanced). Pre-specified as primary under this positivity violation, overlap weighting (ATO) achieved full balance (0/31 SMD>0.1; ESS 81%) and gave a non-significant null in the overlap population (RD +4.5%, 95%CI −1.8 to +10.9; HR 1.169). Calibrated effect-modifier predictions were directionally concordant with both trials (ADRENAL HR 1.045 vs 0.95; APROCCHSS 0.872 vs 0.88), nominating respiratory source and vasopressor dose, not SOFA as candidate modifiers; negative controls indicated ~70% residual-bias reduction. MAIC to four Asian cohorts suggested benefit (RD 4.1% to 11.9%) but combined ESS 1.8-3.0% triggered an automatic "exploratory-only" flag.
CONCLUSIONS: Standard IPTW produced qualitatively wrong harm signal that positivity diagnostics and overlap weighting reversed into a trial-concordant null, showing identical data can yield opposite conclusions without positivity checks. With these safeguards, observational data recovered the ADRENAL-APROCCHSS severity gradient and localized hydrocortisone's benefit to refractory, respiratory-source shock. Positivity diagnostics, RCT calibration, and ESS gating should be prerequisites before transporting Western RWE to Asian ICU decisions.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
SA1
Topic
Study Approaches
Disease
SDC: Infectious Disease (non-vaccine)