Challenges in handling allogeneic stem cell transplantation in randomized clinical trials.

[BACKGROUND] In randomized clinical trials (RCTs) for hematological malignancies, patients may undergo allogeneic hematopoietic stem cell transplantation (allo-HSCT) as part of standard clinical pathways. Allo-HSCT is a potentially curative but high-risk procedure performed after randomization and thus constitutes an important intercurrent event that can substantially influence survival outcomes. However, its handling in statistical analyses is not standardized.

[OBJECTIVE] To review current statistical methods used to handle postrandomization allo-HSCT as an intercurrent event in RCTs, and to highlight how each method corresponds to a different estimand, reflecting distinct clinical questions.

[METHODS] We reviewed 93 RCTs published between January 1, 2014, and April 1, 2024 that reported survival outcomes with postrandomization allo-HSCT.

[RESULTS] Three different statistical methods were employed to estimate the treatment effects: censoring at the time of allo-HSCT (64 analyses), a time-dependent covariate in a Cox model (24 analyses), or ignoring allo-HSCT status (17 analyses). Each method estimates the treatment effect in response to a different clinical question and estimand, with specific assumptions that must be considered when interpreting the results. Censoring corresponds to the "hypothetical" estimand, but its validity requires 2 things: first, that the likelihood of receiving allo-HSCT is similar across treatment arms; and second, that patients who undergo transplantation have a similar prognosis to those who do not. Time-dependent covariate incorporates the effect of allo-HSCT but is not associated with a specific estimand and requires careful interpretation. Ignoring allo-HSCT corresponds to the "treatment policy" strategy, of comparing the treatment strategy, whichever allo-HSCT or not, without additional assumptions.

[CONCLUSION] There is no consensus on handling allo-HSCT as an intercurrent event in survival analyses. Censoring, although common, may introduce bias if treatment or prognostic covariates influence allo-HSCT use. The treatment policy estimand should be preferred when allo-HSCT is part of the therapeutic strategy.