Heterogeneity Blindness in ALS Clinical Trials: Power Loss, Estimand Mismatch, and a Latent-Class Alternative

Standard analytical methods in ALS clinical trials—linear mixed models (LMM) and ANCOVA—assume that all patients follow a single disease trajectory. We present simulation evidence from approximately 14,650 synthetic trials demonstrating that this assumption imposes three compounding costs: (1) severe power loss, with LMM detecting class-specific treatment effects in only 12% of trials compared to 90% for a latent-class mixed model (LCMM) pipeline; (2) estimand mismatch, where ANCOVA applied to change scores in the presence of informative dropout produces approximately 40% collider bias under missing-not-at-random (MNAR) conditions; and (3) masking of heterogeneous treatment effects that may explain decades of failed Phase III trials. We propose a two-stage LCMM pipeline—trajectory classification followed by within-class inference with full-pipeline permutation testing—that recovers class-specific effects while maintaining nominal Type I error control. All code and simulation data are publicly available.