CONTROLLING FOR WHAT COMES NEXT: ESTIMATING GENERALIZED DIRECT EFFECTS OF FIRST-LINE THERAPY ON REAL-WORLD OVERALL SURVIVAL USING COUNTERFACTUAL SECOND-LINE DISTRIBUTIONS IN ADVANCED NON-SMALL CELL LUNG CANCER
Author(s)
Kenneth A. Taylor, DPT, PhD1, Philani Mpofu, PhD2, Lockwood Taylor, PhD, MPH2, Raina Mathur, MPH2.
1Flatiron Health, Durham, NC, USA, 2Flatiron Health, New York, NY, USA.
1Flatiron Health, Durham, NC, USA, 2Flatiron Health, New York, NY, USA.
OBJECTIVES: First-line (1L) therapy effects on real-world overall survival (rwOS) operate partly through second-line (2L) therapies; when 2L practice differs across populations or over time, the estimated 1L effect on rwOS may differ. We demonstrate estimation of a generalized direct effect (GDE) of 1L therapy on rwOS under three pre-specified 2L counterfactual distributions.
METHODS: Retrospective cohort study using the US-based Flatiron Health Research Database. Eligible patients were adults with stage IV NSCLC diagnosed between January 2017 and February 2026, who initiated 1L immunotherapy (IO) or platinum-based chemotherapy (PBCT), with no pre-1L targetable biomarker alterations. We estimated the rwOS difference between 1L treatments under three counterfactual 2L distributions (chemo-heavy, IO-heavy, PBCT-heavy) via parametric mediational g-computation. We adjusted for demographic, clinical (PD-L1, ECOG, metastatic burden, histology), and treatment-setting baseline confounders, with time-varying progression status as an intermediate confounder, time modeled as a natural cubic spline, and interactions of 1L therapy with 2L class and with time in the rwOS hazard. As reference, we estimated the naive adjusted 1L effect standardizing over baseline confounders only. Bootstrapping was used to calculate 95% CI.
RESULTS: We analyzed 20,309 patients (IO: 10,176; PBCT: 10,133). The naive adjusted 1L IO−PBCT absolute difference in rwOS (95% CI) at 12, 24, and 36 months was 8.7% (7.2%-10.3%), 10.7% (9.3%-12.1%), and 9.8% (8.4%-11.1%). Under a chemotherapy-heavy 2L distribution, the GDE was 10.5% (8.5%-12.6%), 10.7% (8.7%-12.6%), and 8.9% (7.2%-10.7%); IO-heavy and PBCT-heavy alternatives gave similar estimates (range ≤1.2% at each horizon).
CONCLUSIONS: We demonstrated feasibility of GDE estimation with EHR-derived data using parametric mediational g-computation. The three counterfactual 2L distributions yielded estimates similar to one another and to the naive 1L contrast, consistent with the overlap among feasible 2L distributions in this cohort; GDE estimation enables investigation of 1L effects in populations whose subsequent treatment distributions differ more substantially.
METHODS: Retrospective cohort study using the US-based Flatiron Health Research Database. Eligible patients were adults with stage IV NSCLC diagnosed between January 2017 and February 2026, who initiated 1L immunotherapy (IO) or platinum-based chemotherapy (PBCT), with no pre-1L targetable biomarker alterations. We estimated the rwOS difference between 1L treatments under three counterfactual 2L distributions (chemo-heavy, IO-heavy, PBCT-heavy) via parametric mediational g-computation. We adjusted for demographic, clinical (PD-L1, ECOG, metastatic burden, histology), and treatment-setting baseline confounders, with time-varying progression status as an intermediate confounder, time modeled as a natural cubic spline, and interactions of 1L therapy with 2L class and with time in the rwOS hazard. As reference, we estimated the naive adjusted 1L effect standardizing over baseline confounders only. Bootstrapping was used to calculate 95% CI.
RESULTS: We analyzed 20,309 patients (IO: 10,176; PBCT: 10,133). The naive adjusted 1L IO−PBCT absolute difference in rwOS (95% CI) at 12, 24, and 36 months was 8.7% (7.2%-10.3%), 10.7% (9.3%-12.1%), and 9.8% (8.4%-11.1%). Under a chemotherapy-heavy 2L distribution, the GDE was 10.5% (8.5%-12.6%), 10.7% (8.7%-12.6%), and 8.9% (7.2%-10.7%); IO-heavy and PBCT-heavy alternatives gave similar estimates (range ≤1.2% at each horizon).
CONCLUSIONS: We demonstrated feasibility of GDE estimation with EHR-derived data using parametric mediational g-computation. The three counterfactual 2L distributions yielded estimates similar to one another and to the naive 1L contrast, consistent with the overlap among feasible 2L distributions in this cohort; GDE estimation enables investigation of 1L effects in populations whose subsequent treatment distributions differ more substantially.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR200
Topic
Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Oncology