Dear Editor, We read with great interest the article by Bae et al1 entitled “Prognostic implication of tumor grade in patients with node-negative breast cancer aged ≤50 years with 21-gene recurrence scores of 11–25” recently published in the International Journal of Surgery. The authors investigated the prognostic utility of histological tumor grade among patients with node-negative, hormone receptor–positive, and HER2-negative breast cancers and intermediate Oncotype DX 21-gene recurrence scores (RS) of 11–25. We commend the authors for providing pivotal evidence that clinical risk factors – specifically tumor grade – offer additional prognostic stratification for premenopausal women with intermediate RS2. However, we had several concerns about the statistical methodology, which might affect the validity and interpretability of their findings. First, in the subgroup of patients aged ≤50 with RS of 11–25 incorporated into the Cox regression analysis, the authors might not adhere to the empirical rule of a minimum events-per-variable ratio of 10. The number of covariates entered into the multivariable analysis (n = 9) was described in detail, whereas the total number of recurrence events was not reported, which seemed to be very small as shown in the Kaplan–Meier curves. This was further evidenced by the remarkably wide confidence intervals presented in the tables. To mitigate model overfitting and ensure robust hazard estimates in the scenario, penalized regression, bootstrap validation, or variable selection strategies could be methodological choices and should be detailed in the Methods section, if adopted. Given that this is the key analysis to support the study’s primary finding, the authors might better acknowledge and address the issue of potential overfitting and sparse data bias. Secondly, the stratified analysis by chemotherapy administration could be complicated by collider-stratification bias or indication bias. It was evident that patients with high-grade tumors were significantly more likely to receive chemotherapy, and chemotherapy receipt was a post-exposure variable influenced by the very exposure (tumor grade) under investigation. The authors found that tumor grade was an independent prognostic factor only in patients who did not receive chemotherapy, that is, chemotherapy seemed to attenuate the adverse prognostic impact of high-grade tumors, which is biologically plausible. Nevertheless, stratifying by a variable that lies on the causal pathway between exposure and outcome would introduce collider-stratification bias, which can distort the true exposure–outcome relationship in unpredictable ways. We suggested that a formal interaction term in the Cox model (tumor grade × chemotherapy), instrumental variable approaches or causal mediation analysis should be taken into account. Thirdly, our concern relates to the analytical framework in which propensity score matching and multivariable Cox regression were applied using substantially overlapping variable sets. This might create statistical redundancy and lead to over-adjustment, as the matched sample no longer contains the natural variation needed to estimate their independent effects accurately, artificially narrowing the variance of covariates and potentially masking their true prognostic contributions. The integration of tumor grade, or artificial intelligence-derived histopathological features3 into breast cancer treatment decisions for premenopausal patients with intermediate genomic risk remains an area of active investigation. We hope additional clarifications or further analyses could strengthen the evidence base for this clinically relevant question.
Wu et al. (Wed,) studied this question.