IPW RMST Validation Stress‐Test

This repository contains a simulation-based stress test of a clone-censor-weighting (CCW) restricted mean survival time (RMST) pipeline under realistic violations of target trial emulation assumptions.

The goal is to answer a single, important scientific question:

When real-world assumptions fail in realistic ways, does CCW-RMST fail catastrophically, misleadingly, or gracefully?

What the Stress Test Shows

The pipeline fails gracefully under epidemiologically plausible departures from ideal conditions:

• Directionally informative — Null scenarios give ≈0 effect; protective scenarios recover a negative ΔRMST
• Stable diagnostics — Weights remain usable, ESS does not collapse, convergence flags work
• Negative lags ≈ 0 — No spurious pre-vaccination effects even with latent confounding
• Placebo centered near 0 — Falsification test behaves as expected

These results provide strong support for using CCW-RMST as a safe, diagnostically reliable exploratory tool in observational vaccine effectiveness studies — especially when only limited measured confounders (e.g., age and sex) are available.

Key Features of the Simulation

• Scenarios

  • Null (true HR ≈ 1)
  • Protective (true HR ≈ 0.7)

• Realistic Data-Generating Process

  • Calendar-time mortality waves (seasonal + epidemic)
  • Phased vaccination rollout with supply shocks
  • Age/sex prioritization
  • Latent unmeasured health confounder (affects both vaccination timing and death hazard)

• Methodological Pipeline

  • Clone-Censor-Weighting with strategy-specific IPCW
  • Calendar-time + relative-time B-splines in propensity and hazard models
  • Stabilized weights (clipped factors before cumprod, percentile truncation)
  • Pooled logistic regression (CLogLog link) on person-day rows with frequency weights
  • Bootstrap CIs + placebo falsification (sampled from numerator inside each replicate)
  • Weighted averaging over t₀ grid by n_eligible(t₀)

• Diagnostics

  • Effective sample size (ESS)
  • Standardized mean differences (SMDs)
  • Positivity violations
  • Censoring rates
  • GLM convergence flags

Value

The simulation demonstrates that CCW-RMST exhibits substantial robustness to realistic violations of sequential ignorability — including latent confounding, nonstationary hazards, and phased treatment assignment.

Negative lags and placebo checks behave as expected falsification tests, and the method recovers direction and rough magnitude even under stress.

This supports the use of CCW-RMST as a safe, diagnostically trustworthy exploratory tool in observational vaccine mortality studies with limited measured confounders.

Status: January 2026 — simulation-based methodological stress test (not proof of causal identification)

Full code: AG) Clone Censor weight RMST Stress-Test simulation.py