[Methodological appraisal of statistical approaches to evaluating the impact of chronic graft-versus-host disease on mortality after hematopoietic stem cell transplantation].

Objective

Using the impact of chronic graft-versus-host disease (cGVHD) on overall survival after hematopoietic stem cell transplantation as an example, this study aims to introduces and critically appraises five statistical approaches for handling time-dependent events in survival analysis.

Methods

This study was based on data from the Center for International Blood and Marrow Transplant Research (CIBMTR) GV18-03 study. A total of 4361 patients with acute myeloid leukemia or myelodysplastic syndrome were included, who were aged ≥40 years and had received an HLA 8/8 matched sibling donor transplant between 2008 and 2017. Five analytical approaches were used, treating cGVHD as: ① a baseline fixed covariate, ② a time-dependent covariate, ③ a time-dependent event via landmarking analysis, ④ a transitional event in a multi-state model, and ⑤ an exposure in the parametric g-formula. For each approach, we presented its core concept and results, and elaborated on its rationale, strengths, limitations, and applicable scenarios.

Results

The five approaches showed significant differences in bias control, modeling flexibility, and clinical interpretability. Method 1 was prone to immortal time bias. Method 2 partially corrected this bias but was vulnerable to informative censoring and estimation instability due to insufficient sample sizes in the early-onset cGVHD group. Method 3 was straightforward but dependent on a predefined landmark time point and unable to account for cGVHD occurring after this point. Method 4 allowed for a comprehensive description of the impact of time-dependent events on survival prognosis by modeling dynamic clinical transitions. Method 5 could simultaneously handle both time-dependent events and confounders, making it suitable for estimating population-level effects of specific exposure strategies.

Conclusion

Researchers should select the appropriate method based on their specific study objectives to ensure that the estimation of the impact of time-dependent events on outcomes is both statistically valid and clinically meaningful.