Response

I read with great interest the paper by Tsai et al that investigated the impact of antipsychotic augmentation on suicide risk in patients with treatment-resistant depression using a target trial emulation framework.^{Reference Tsai, Yang, Wong, Chuang, Cheng and Shen1} The authors used this robust methodology, including the clone-censor-weight approach, to address potential biases. The authors’ study represents an advancement in psychiatric research, an area often complicated by various biases in observational studies.^{Reference Narita and Furukawa2} However, I would like to offer a suggestion regarding their presentation of the hazard ratio.

The authors state in their statistical analysis section that the outcome model included an ‘interaction term between exposure status and follow-up time’. The inclusion of such an interaction term implies that they assume a time-dependent hazard, meaning that the hazard ratio can vary over the follow-up period. Despite this, the paper presents a single hazard ratio as the main finding. It would be beneficial if the authors could clarify how the single hazard ratio was derived from a model that assumes a time-dependent hazard. Without such clarification, readers may mistakenly interpret the hazard ratio as constant over time. Presenting only a single hazard ratio is typically based on the proportional hazards assumption with built-in selection bias; the selection of participants at higher risk of events earlier in the study distorts the estimate.^{Reference Hernán3,Reference Dumas and Stensrud4} This is why it is important to evaluate the time-dependent hazard ratios and provide cumulative incidence over the follow-up period.^{Reference Murray, Caniglia and Petito5,Reference Narita, Shinozaki, Goto, Hori, Kim and Wilcox6}

Since presenting only a single hazard ratio could be misleading, presenting estimates that are more interpretable to readers, such as risk difference and risk ratio at clinically relevant follow-up points, would be beneficial. Such estimates could be derived using a counterfactual approach like the g-formula, which aligns well with the target trial emulation framework. For example, risk difference and risk ratio could be presented at the midpoint and at the end of the follow-up, with the follow-up period specified, thereby directly highlighting the clinical relevance of the findings.

Clarification on how the single hazard ratio was calculated from a model assuming time-dependent hazards would enhance the transparency of their important findings. Clinically relevant estimates at specific follow-up points would enhance the interpretability and practical application of the findings. A summary contrast could be informative if its computation is clearly described and it is presented alongside time-specific estimates. This study makes a valuable contribution to understanding treatment options for treatment-resistant depression, and further elucidation on these points would be welcome.