Author's reply

Abstract

We would like to reply to the comments by Weng et al. regarding our recent study, in which we reported that the Selvester QRS score could predict future cardiac events in patients with non-ischemic dilated cardiomyopathy (NIDCM), reflecting myocardial fibrosis assessed by collagen volume fraction [[1]Hiraiwa H. Okumura T. Sawamura A. Sugiura Y. Kondo T. Watanabe N. et al.The Selvester QRS score as a predictor of cardiac events in nonischemic dilated cardiomyopathy.J Cardiol. 2018; 71: 284-290Abstract Full Text Full Text PDF PubMed Scopus (14) Google Scholar]. We appreciate the attentive examination of the statistical methods used in our study and the constructive suggestion regarding proper statistical analyses. First, we performed standard univariate and multivariate Cox proportional hazard regression analyses, which are common practices in clinical research. Actually, in our study, a total of 20 cardiac events occurred among 91 NIDCM patients and we indeed applied four covariates with a p-value of <0.01 to the multivariate model in order to avoid overfitting. Basically, the full-model approach in which all the candidate variables are included is claimed to avoid overfitting and selection bias, although it is often impractical to include all candidate covariates [[2]Royston P. Moons K.G. Altman D.G. Vergouwe Y. Prognosis and prognostic research: developing a prognostic model.BMJ. 2009; 338: b604Crossref PubMed Scopus (769) Google Scholar]. However, there is no consensus on the best method for selecting variables. The best approach is considered to be the selection of all potential risk factors for cardiac events established in previous studies and the application of significant factors to the multivariate analysis. However, the risk factors for cardiac events in our study population have not been fully established. In addition, the cut-off p-value of variables in applying to the multivariate model is controversial. Therefore, in our study, we adopted covariates with a p-value of <0.01 but not <0.2 in the univariate analysis to apply to the multivariate analysis, and we avoided a more overfitting model in comparison with the case that we used covariates with a p-value of <0.2. Furthermore, we performed bivariate Cox proportional analyses to confirm the predictive value of the Selvester QRS score independent of any of confounding factors. The Selvester QRS score increased the risk of cardiac events independent of the variables that we detected as significant prognostic determinants in the univariate analyses. As a result, our conclusion is reasonable even in this small sample size. Second, as Weng et al. pointed out, we fully agree with the need to test for the proportional hazards assumption. Actually, we confirmed the constancy of the proportional hazards assumption by the Schoenfeld residuals plot and Schoenfeld residuals test in advance using R version 3.1.2 (R Foundation for Statistical Computing, Vienna, Austria) [[3]Klein J.P. Moeschberger M.L. Survival analysis: techniques for censored and truncated data. Springer, New York2003Crossref Google Scholar]. Thus, we do not believe that any major violations of the proportional hazards assumption were present. Finally, our study contained some limitations, as we mentioned in the article. However, as described in this reply, our conclusion that the Selvester QRS score is useful to predict future cardiac events in patients with NIDCM is not an overoptimistic expectation but a statistically understandable consequence.