Estimating causal effects on quality of life under treatment discontinuation: the ALTA-1L trial

Time to worsening in health-related quality of life (HRQoL) is increasingly used in oncology trials. Treatment discontinuation poses a challenge: once discontinued, patients cease HRQoL assessments, precluding outcome observation. Standard survival analyses censor at discontinuation, assuming noninformative censoring-an assumption often violated when discontinuation relates to disease progression or toxicity. Bridging the ICH E9(R1) estimand framework with causal inference methods clarifies how to define and estimate treatment effects in such settings. We reanalyze time to worsening in global health status from the ALTA-1L trial (brigatinib vs crizotinib in ALK + non-small-cell lung cancer), integrating both frameworks.

Following Young et al's (2020) causal framework for competing events, we defined two estimands structured according to ICH E9(R1): (1) a controlled direct effect under a hypothetical strategy, envisioning a scenario where treatment discontinuation would not occur, estimated using inverse probability of censoring weighted Kaplan-Meier to adjust for informative censoring; and (2) a total effect under a while-on-treatment strategy, with discontinuation as a competing event, estimated using the Aalen-Johansen estimator. Risk ratios (RRs) were estimated at 36 months with bootstrap CIs.

The original ALTA-1L analysis reported hazard ratio = 0.69 (95% CI: 0.49, 0.98), censoring at discontinuation and assuming noninformative censoring. Deriving the RR at 36 months from Kaplan-Meier curves yields 0.75 (95% CI: 0.59, 0.97). After adjusting for informative censoring, the controlled direct effect was RR = 0.89 (95% CI: 0.65, 1.26). The total effect was RR = 1.03 (95% CI: 0.76, 1.40), reflecting the competing risk structure: earlier discontinuation in the crizotinib arm (discontinuation RR = 0.54; 95% CI: 0.38, 0.72) reduced observed worsening events. These different estimates illustrate how different estimands address distinct clinical questions.

This study bridges the ICH E9(R1) estimand framework with causal inference methods for time-to-event HRQoL analysis when discontinuation precludes observation. By quantifying bias from standard approaches, we provide methodological clarity for applied researchers. To facilitate practical implementation, we translate these insights into a decision flowchart for estimand specification and method selection when intercurrent events (ICEs) act as competing events. Future trials should prespecify ICE-handling strategies and consider data collection beyond ICEs to support treatment policy estimation.