Response to: Bayesian ­reappraisal of ACNS0332 and ACNS0334 strengthens ­subgroup treatment effects in high-risk pediatric group 3 medulloblastoma

We read with great interest the letter reporting a Bayesian reanalysis of high-risk Group 3 Medulloblastoma (hr-G3-MB) patient subgroup data from the Children’s Oncology Group ACNS0334 and ACNS0332 studies. We commend the authors’ effort to address the well-known challenges of adequately powering subgroup analyses in pediatric oncology trials and note the reanalysis presented by Mistry and colleagues supports the conclusions of ACNS0334 and ACNS0332. For both studies, the statistical analysis according to protocol design incorporated a one-sided test to evaluate whether intensified therapy improved event-free survival (EFS). A one-sided test was determined to be appropriate, as the study committees hypothesized that intensification would improve and not worsen EFS. Results of both studies showed clinical and statistical significance in hr-G3-MB, 19% improvement in EFS on the carboplatin arm of ACNS0332 (one-sided P = .047)¹ and 36% improvement in EFS on the methotrexate arm of ACNS0334 (one-sided P = .037).² While the methodological innovation of the Bayesian analysis presented by Mistry and colleagues is notable and may improve confidence in the study results, we believe several points merit careful consideration.

Bayesian borrowing is critically dependent on the exchangeability between historical and trial populations. Particularly in rare pediatric populations, even subtle changes in treatment, supportive care, or patient selection across time can result in non-trivial bias. One could consider incorporation of propensity score modeling or Bayesian hierarchical adjustment for baseline characteristics if available. Inclusion of multiple historical comparators or broad composite datasets could further increase robustness. With any model, the clinical interpretability depends on understanding the model assumptions and choices. Bayesian models are especially dependent on the parameters that control how much historical data is borrowed and assumptions of prior probability. Formal sensitivity analysis and simulation studies could better quantify uncertainty. While the Bayesian Fragility Index is an interesting concept, its practical value may be limited unless accompanied by clear guidelines for interpretation. In particular, the fragility threshold must be contextualized with respect to sample size, censoring mechanisms, and the clinical relevance of individual events. It is important to consider how the model performs when there is no treatment effect (evaluation of type I error).

Using historical data with Bayesian methods can be a helpful way to improve future clinical trials, especially in rare diseases with limited sample size. However, this method should be used carefully. Trials should clearly explain when and how they plan to use historical data and how similar historical and current patient groups need to be. In model development, it will be important to test how sensitive the results are to the choices of parameters made in the analysis through simulations prior to initiation of clinical trials, which may rely more heavily on external controls.

Funding

National Institutes of Health (U10CA180899 Statistics & Data Center to G.K., A.O., Y.L.; 5R50CA275857-02 to S.L.).

Conflict of interest statement.

No authors have a conflict of interest related to this work.