A modified weighted log-rank test for confirmatory trials with a high proportion of treatment switching

TL;DR

This paper introduces a modified weighted log-rank test designed for confirmatory cancer trials with high treatment switching, aiming to improve statistical power and reduce bias in overall survival analysis.

Contribution

The paper proposes a novel mWLR test that predicts hazard ratios to adaptively weight events, enhancing analysis accuracy in trials with substantial treatment switching.

Findings

The mWLR test outperforms the standard log-rank test in simulations.

It effectively accounts for treatment switching in survival analysis.

The method incorporates prior trial information for better weighting.

Abstract

In confirmatory cancer clinical trials, overall survival (OS) is normally a primary endpoint in the intention-to-treat (ITT) analysis under regulatory standards. After the tumor progresses, it is common that patients allocated to the control group switch to the experimental treatment, or another drug in the same class. Such treatment switching may dilute the relative efficacy of the new drug compared to the control group, leading to lower statistical power. It would be possible to decrease the estimation bias by shortening the follow-up period but this may lead to a loss of information and power. Instead we propose a modified weighted log-rank test (mWLR) that aims at balancing these factors by down-weighting events occurring when many patients have switched treatment. As the weighting should be pre-specified and the impact of treatment switching is unknown, we predict the hazard ratio function and use it to compute the weights of the mWLR. The method may incorporate information from previous trials regarding the potential hazard ratio function over time. We are motivated by the RECORD-1 trial of everolimus against placebo in patients with metastatic renal-cell carcinoma where almost 80\% of the patients in the placebo group received everolimus after disease progression. Extensive simulations show that the new test gives considerably higher efficiency than the standard log-rank test in realistic scenarios.