Problems With Interpreting Treatment Effects in Observational Competing Risks Data—Reply

Abstract

To the Editor I read with interest the article by Sun et al, 1 which addresses an important topic in head and neck cancer treatment. One issue is that the authors state they did not have patient-level human papillomavirus status. However, it is possible to obtain p16 data for many patients in the US Veterans Affairs (VA) database. 2 It would be helpful to know if their findings hold in patients with known p16 status (regardless of tumor site). Second, nonspecificity is a known problem in comparative effectiveness studies in competing risk populations. 3 It is important that researchers show, as a routine practice, comparative effects of treatments on competing (noncancer) mortality, to avoid misleading conclusions. Often, one can find that survival differences are attributable to effects on competing mortality, owing to residual confounding, which is not possible to eradicate entirely with propensity scores 3 ; however, using correction methods can partially attenuate this problem when present. 3 The authors should have also addressed the potential effect of temporal trends in treatment allocation. Lastly, in the VA data (and, presumably, the overall population), there is a great deal of heterogeneity in the baseline relative risk for competing events among older patients, 4 which can affect the propensity to benefit from chemotherapy. 5 It would help to know if this baseline risk (which can be measured using validated risk scores) 5 was evenly distributed across the carboplatin-based vs cetuximab groups. Due to the many complexities with interpreting treatment effects in observational competing risks data, randomized clinical trials are required to resolve which regimen should constitute the standard definitive treatment for cisplatin-ineligible head and neck cancer.