We read the paper by Loeffelbein and colleagues, published online in the International Journal of Oral and Maxillofacial Surgery in December 2016, with enthusiasm . The authors aimed to examine tumour-specific and patient-related risk factors in a large single-centre cohort, i.e., factors that have a considerable effect on postoperative survival following ablative tumour surgery with elective neck dissection and immediate microvascular reconstruction in oral squamous cell carcinoma. It was concluded that postoperative survival does not depend only on tumour-related characteristics. The American Society of Anesthesiologists (ASA) status needs to be taken into account during treatment planning, as it is a significant predictor of patient survival. Although the study made considerable contributions to the area, certain methodological points must be considered.
First, the authors applied the Cox proportional hazard regression model to identify predictors of overall survival. Afterwards, covariates that were shown to be associated with overall survival in the univariate analysis ( P ≤ 0.05) were imported into the final multivariate model, which is problematic. In fact, considering P ≤ 0.05 as the basis on which to select the covariates that should be imported into the final multivariate model will result in the covariates with a relatively small effect being overlooked and only covariates with a relatively large effect being considered. With this unsuitable strategy of multivariate model building, the effect of covariates with relatively large and small effects would be over- and under-estimated, respectively, leading to an important bias, named testimation bias . Therefore, a relaxed P -value such as P ≤ 0.2 is proposed for the selection of the variables for multivariate models in order to dilute the aforementioned bias .
Second, the multivariate Cox proportional hazard regression model was used to examine the identified predictors of overall survival; however, the proportional hazard (PH) assumption is violated in the aforementioned study, which is obvious in Figs. 2–4 . In these figures, this important assumption is not established for ASA status, N stage, Union for International Cancer Control (UICC) stage, or alcohol abuse, which using the Cox proportional hazard regression model will be misleading. Hence, it is suggested that a variant of this model, such as the extended or stratified Cox regression model, should be applied to avoid misleading results .
A take home message for readers is that the standard strategy for model building should be used to avoid any improper multivariate models . Also, the PH assumption should be checked initially using time-dependent covariates methods , then the appropriate variant of the Cox regression model must be applied to avoid any misleading results.
The present study was not funded by any organization.