Table 2.

Analysis considerations for different strategies to handle ICEs.

Chosen intercurrent event strategy	Analysis considerations for time-to-event endpoints
Treatment policy	Events observed after the ICE (e.g., after discontinuation or interruption) are considered in the analysis, that is, data collection of progression or death dates or the corresponding censoring dates required even after a patient experiences such an ICE.
Composite strategy	The ICE (e.g., COVID-19 related death) is considered as event in the definition of time-to-event endpoint.
Hypothetical	For hazard-based quantities, for example, the hazard ratio, assuming absence of informative censoring the relative effect can be estimated through simple censoring at the ICE. If informative censoring cannot be excluded (e.g., patient is censored at the start of new therapy after discontinuation that could be attributed to disease), methods such as inverse probability of censoring weights (IPCW) accounting for that may be indicated (Robins and Finkelstein  2000; Lipkovich, Ratitch, and O’Kelly  2016). These also allow to provide estimates of a hypothetical estimand for survival probabilities.
Principal stratum	Estimation of the treatment effect in the principal stratum such as “patients who would never experience severe impact of COVID-19 infection under either treatment” could be done within the potential outcomes framework. To estimate this effect assumptions will be necessary. A potential assumption that allows for estimation of principal stratum effects is principal ignorability (PI), an assumption similar to the ignorability assumption in propensity score analysis of observational data (Jo and Stuart  2009). PI assumes that, conditional on baseline confounders, the potential outcome (e.g., PFS or OS) for the treated (untreated) is independent of the potential outcome of the COVID-19 status for untreated (treated). Stated differently, once baseline covariates that may confound the relationship between COVID-19 status and the outcome variable are known, knowing the COVID-19 status of the treated (untreated) provides no further information on the outcome for the untreated (treated) and vice versa. Alternatively, Frangakis and Rubin ( 2002) used a joint model for estimation. This requires specifying two models, one for the outcome given the principal stratum and one for the principal stratum membership.
	We emphasize that these assumptions are unverifiable from the collected data. Jo and Stuart ( 2009) and Stuart and Jo ( 2015) described sensitivity analyses for principal ignorability when making the exclusion restriction assumption. As a reviewer pointed out, tipping point analyses can also be used to explore the extent to which inestimable quantities would need to vary to change the conclusion of the analysis. This would be an extension of, for example, the methods proposed in Lou, Jones, and Sun ( 2019) to superiority trials with time-to-event endpoints.