Misinterpretation of statistical results and recurrent intracranial hemangiopericytoma

Abstract

We read with great interest the article written by Tian et al. (J Neurooncol doi: 10.1007/s11060-013-1193-4) about the clinical characteristics and prognostic analysis of recurrent intracranial hemangiopericytoma (HPC) and the conclusions of the study were summarized based on the statistical results. In the study, they reviewed 46 cases of recurrent intracranial HPC with a follow-up duration of 41.6 months. At the final evaluation, 16 cases suffered a second or further recurrence. They used univariate analysis (Kaplan– Meier method, Log-rank test) and identified the adverse predictors of age \35 years and unclear tumor boundary for predicting a second recurrence in the 46 cases. Meanwhile, they confirmed the time interval between the first and second recurrences as a risk factor (Kaplan–Meier method, Log-rank test) for poor prognosis in 16 patients with a second or further recurrence. However, in another 29 patients, they used Kaplan–Meier method (Breslow test) and stated that gross total resection (GTR) with external beam radiotherapy (EBRT) (n = 24) was significantly associated with increased overall survival, compared with GTR alone. As we known, Breslow test is best at detecting early differences and Log-rank method is best at detecting differences between the curves late in the time period of the study. However, Tian et al. did not perform the Logrank test when comparing the difference between GTR?EBRT and GTR alone by Kaplan–Meier method. In addition, if the proportional hazard (PH) assumption is not valid then Breslow (also called Wilcoxon) is the choice. Moreover, they performed a Cox regression (multivariates analysis without details) to verify the independent predictor. As mentioned above, only 16 patients did have the time interval between the first and second recurrences, and only 29 received GTR?EBRT or GTR alone. Patients (n = 30), who did not have a second recurrence, certainly would not have the socalled time interval. Therefore, it was clear that the maximal available number of patients entering the Cox regression was 16 patients and other patients (n = 30) were excluded automatically. We assumed that the result of the Cox regression was obtained only based on the 16 patients but not the 46 patients. Furthermore, they did not correctly interpret the statistical results and described that ‘‘treatment strategy approached significance’’ just based on a p value of 0.037 that was a significant shortcoming of this study and an arbitrary judgment. They did not take the span of 95 % confidence interval into consideration. It is a common sense that when the range of 95 % confidence interval spans the value of 1, the statistical result will not be significant, even if the p value was less than 0.05 [1]. Therefore, the treatment strategy of the study no longer approached significance, because the range of the 95 % confidence interval of the hazard ratio spanned 1 (0.976–12.406). In the abstract of the study, Tian et al. described that ‘‘univariate and multivariate regression analyses showed that factors associated with good prognosis included recurrence age over 35 years, an interval between the first and second recurrence of more than 1 year and a clear boundary of the recurrent tumor’’, but in fact these factors were only significant during univariate analysis rather than during multivariate analysis. This misinterpretation would lead to misunderstanding of the results of the study because the majority of readers including clinicians and patients only review the abstract. Meanwhile, the limitations should be stressed that confounding factors and bias could not be minimized during univariate analysis. D. Li J.-T. Zhang (&) Beijing Tiantan Hospital, Capital Medical University, Beijing, China e-mail: zhangjunting2003@aliyun.com