2025/12/15 by Hee Won Choi, Young Lan Kwak, Hyun‐Soo Zhang · 1 voice
Medicine · #Blood transfusion and management #Cardiac, Anesthesia and Surgical Outcomes #Iron Metabolism and Disorders
paper · pdf · doi:10.1111/anae.70095
openalex created_date 2025/12/15 · openalex publication_date 2025/12/15 · openalex updated_date 2026/06/15
We thank Jin et al. [1] for their comments and this gives us the opportunity to clarify several methodological aspects of our study [2]. Although not stated explicitly, our analysis was conceptually equivalent to a landmark analysis set at 2 months postoperatively when haemoglobin status was ascertained, and only patients with available data at this time-point were included. This approach ensures all patients were on an equal footing at the beginning of risk assessment. Because patients who died or lacked haemoglobin data before the landmark were not included, our analyses did not involve any period of guaranteed survival, and thus immortal time bias does not arise within the defined study population. Therefore, the primary estimate is defined explicitly as the association between peri-operative anaemia status and subsequent mortality among survivors with haemoglobin data at 2 months, a limitation acknowledged in our discussion. Regarding the timescale alignment, two analytical choices were possible: redefining follow-up to begin at 2 months and extending to 14 months postoperatively; or retaining follow-up through 12 months after surgery. We adopted the latter approach for clinical relevance and consistency with standard practice, but this decision does not compromise the landmark framework or induce immortal time bias, as all included patients entered the risk set only after exposure ascertainment and followed for the full year. Meanwhile, in accordance with the recommendation to perform multivariate Cox regression with the follow-up beginning from the time-point when postoperative haemoglobin data were included, the results were consistent with those from our original cohort analysis, both in terms of hazard estimates and main associations. To examine potential survivor selection bias, we provide a comparison of baseline characteristics and postoperative outcomes between included and excluded patients (Table 1). Excluded patients were younger and had a lower postoperative transfusion rate yet showed higher rates of mortality (6% vs. 3%) and cardiac arrest (5% vs. 1%), with a mixed profile across other complications, suggesting that any selection bias is minor and bidirectional. 2Concerning the risk of model overfitting, we agree that the limited number of events warrants cautious model construction. The events per variable in the multivariate Cox and logistic regression models were 81/7 = 11.6 and 61/4 = 15.2, respectively, which were within the recommended range of 10–15 [3]. While stepwise selection based solely on p values can lead to overfitting, our approach used the Akaike Information Criterion to guide parsimony, which penalises model complexity and discourages over-parameterisation [4]. In addition, we incorporated the least absolute shrinkage and selection operator as an additional safeguard against overfitting [5]. Our multi-stage procedure was not a complex search for significance but rather a rigorous strategy integrating penalisation-based statistical methods with essential clinical judgement. The parallel use of these methods was intended to ensure consistency and to develop parsimonious models that are not overfitted to data. To confirm the robustness of our models' discriminatory power, we have performed the internal validation that was suggested. Specifically, we conducted a repeated five-fold cross-validation (using 200 repeats with different data shuffles) to generate stable, internally validated estimates. For reference, the resulting mean Harrell's c-index was 0.877 (95%CI 0.866–0.889) for the Cox model, while the mean area under the curve was 0.788 (95%CI 0.760–0.805) for the logistic model. These internally validated results were consistent with the original model performance metrics, supporting the robustness and stability of our findings.