Presentation (CTI)

OBJECTIVES: Per-protocol real-world evidence (RWE) analyses may inform treatment and reimbursement decisions, but their conclusions can be fragile when protocol deviation is related to prognosis. This study developed a diagnostic and sensitivity-reporting bundle to quantify when a per-protocol conclusion is supported by measured data and when it depends on residual assumptions about deviators.
METHODS: The Pressure, Symptoms, and Residual Uncertainty (PSR) bundle was evaluated in Monte Carlo simulations with N=2,000 and 200 replicates across 18 deviation mechanisms. Pressure was summarized by SPD(t), the log hazard ratio for protocol deviation per 1-SD higher standardized prognostic score within the risk set. Symptoms were summarized by risk-set relative effective sample size and upper-tail weight concentration. Residual uncertainty was summarized by delta-star, the smallest log hazard-multiplier for residual non-ignorability among deviators that reversed the sign of the primary per-protocol log risk ratio. Scenarios varied the true effect, model misspecification, and residual non-ignorability.
RESULTS: Under a non-null per-protocol effect, inverse probability of censoring weighting had near-nominal coverage (0.94), providing a stable primary reference in the simulated setting. However, this apparent stability did not eliminate residual sensitivity. Under a constant residual sensitivity model, inverse probability weighted conclusions crossed zero for delta values from 0 to 2 in 21% of non-null and 11% of null replicates. Median finite delta-star values were 1.50 and 1.25, respectively. Operating maps showed that residual departures could overturn conclusions even when conventional weight diagnostics appeared stable.
CONCLUSIONS: PSR separates three questions that are often mixed together in per-protocol RWE: whether prognosis drives deviation, whether weights show instability, and how strong residual non-ignorability would need to be to change the conclusion. These summaries may help decision makers identify when a per-protocol RWE finding should be treated as fragile rather than decision-ready.