Letter to the Editor: insights into “Robotic versus laparoscopic surgery for middle and low rectal cancer: a retrospective cohort study emulating the target trial”

HIGHLIGHTS We comment on a target trial emulation of robotic vs. laparoscopic rectal surgery. Post-treatment chemo/RT in PSM and models may introduce overadjustment bias. Clarification is warranted on estimand given conversions and per-protocol labeling. Center-level clustering and heterogeneity could influence inference in multicenter data. Cluster-robust standard errors (SEs) or frailty Cox models may strengthen robustness and transparency. Dear Editor, Minimally invasive surgery for middle and low rectal cancer remains a rapidly evolving field in which the choice of surgical platform may influence long-term oncological outcomes. In this context, we read with great interest the recent study by Wang et al[1], which utilizes a target trial emulation framework to compare long-term oncological outcomes of robotic vs. laparoscopic surgery for middle and low rectal cancer. Using a large multicenter real-world cohort, the authors addressed a clinically important question. To further strengthen the interpretation of the target trial emulation, we would like to respectfully offer several methodological clarifications and suggestions for the authors and readers to consider. This commentary complies with the “2025 TITAN Guideline” for statements and use in managing artificial intelligence[2]. No artificial intelligence tools were used during the writing of this work. First, we noticed that postoperative chemotherapy and radiotherapy were included in both the propensity score calculation and the multivariable adjustment models. In causal inference, postoperative therapies are typically post-treatment variables and may function as mediators rather than baseline confounders[3]. For instance, a more precise surgical approach might facilitate faster recovery or fewer complications, potentially improving the feasibility and completion of adjuvant therapy, which could in turn affect survival. Adjusting for such variables could block part of the causal pathway and introduce over-adjustment bias, depending on the intended estimand. It would be helpful if the authors could clarify whether the primary analysis targeted the direct effect of surgery or the total effect, and report a sensitivity analysis with confounder adjustment restricted to preoperative baseline covariates. Second, we would appreciate clarification on the estimand targeted by the primary analysis and how it aligns with the emulated treatment strategies. Patients were grouped by the initial surgical approach at time zero, and conversions to open surgery were retained in the original groups, which is typically consistent with an intention-to-treat-like estimand. However, the Methods describe the primary outcome analysis as per-protocol. In a target trial framework, a per-protocol effect requires a clearly defined strategy and a prespecified approach to handling departures from that strategy[4]. We would welcome clarification on the strategies being emulated, what was considered a departure from the strategy, and why the primary analysis was described as per-protocol given that conversions were retained. Presenting the primary analysis explicitly as intention-to-treat-like, alongside a clearly defined per-protocol analysis if desired, would improve interpretability. Third, we agree with the authors that surgeon-related factors are difficult to capture in retrospective data, and we note the uneven distribution of procedure volume across centers, including centers performing only laparoscopic surgery. These features raise concerns about center-level correlation and unmeasured heterogeneity. We suggest that inference could be strengthened by accounting for clustering at the hospital level through cluster-robust variance estimation, stratification by center in the Cox model, or a shared frailty approach. We also note that the authors conducted sensitivity analyses addressing center distribution. Incorporating center-level clustering directly in the primary modeling framework could further improve the robustness of confidence intervals and P-values. In addition, an E-value-based sensitivity analysis could help contextualize the potential influence of unmeasured confounding related to surgeon volume, learning-curve effects, or differential selection into robotic surgery[5].