Unveiling non-small cell lung cancer treatment effect heterogeneity: a comparative analysis of statistical methods.

For patients with advanced non-small cell lung cancer lacking targetable genomic alterations, the impact of clinico-genomic characteristics on the effectiveness of combining chemotherapy with immunotherapy is unclear. We evaluated four statistical methods for detecting heterogeneous treatment effects (HTE) related to clinical factors, including programmed death-ligand 1 (PD-L1) expression, tumor mutation burden (TMB), and stage at diagnosis, using the American Association for Cancer Research Project GENIE BPC dataset supplemented with institutional data collected under the same data curation model. A two-sided p-value ≤0.05 was used to denote statistical significance for all analyses. The mixture model revealed two latent subgroups: in one subgroup, there was no meaningful treatment effect, with average PFS only 5% longer with immunotherapy alone (95% confidence interval [CI] -19%, 35%); in the second subgroup, immunotherapy alone was associated with a 35% decrease in average PFS (95% CI -59%, 2%), corresponding to a ratio in treatment effects of 1.62 (95% CI 1.02, 2.57). There was a marginal association between lower TMB levels and membership in the subgroup with improved PFS following receipt of chemoimmunotherapy. The causal survival forest highlighted the importance of TMB (variable importance ranking: 1) and PD-L1 (variable importance ranking: 3) when assessing heterogeneity. In contrast, the accelerated failure time and Cox proportional hazards models did not detect any statistically significant HTE. In simulations, the mixture model identified HTE more frequently than other methods, especially with weak covariate relationships, demonstrating its utility for informing personalized treatment approaches.