Amyotrophic lateral sclerosis (ALS) clinical trials have experienced a failure rate exceeding 97% over the past two decades. Standard primary endpoints assume homogeneous, linear progression. We present six simulation experiments totalling approximately 14,650 simulated trials showing that this assumption carries a quantifiable statistical cost.

Key findings: Linear mixed models require approximately 4× the sample size of class-aware analyses for subgroup-specific treatment effects. ANCOVA targets the survivor-average treatment effect, overestimating the population-average by ~36–42% under informative dropout — a structural estimand mismatch from conditioning on survival (collider bias), confirmed by closed-form derivation. A two-stage LCMM pipeline with pseudo-class inference achieves 76–100% power across most stress conditions while LMM achieves 8–22%. Stress testing across 11 data degradation conditions confirms robustness. Permutation calibration maintains approximate nominal Type I error control.

All simulation code, pre-registration records, and adversarial deliberation transcripts are openly available.