Comparing Cox Regression and Parametric Models for Survival Analysis of Patients with Gastric Cancer

Although Cox regression is commonly used to detect relationships between patient survival and demographic/clinical variables, there are situations where parametric models can yield more accurate results. The objective of this study was to compare two survival regression methods, namely Cox regression and parametric models, in patients with gastric carcinoma registered at Taleghani Hospital, Tehran. Using data from 746 patients who had received care at Taleghani Hospital from February 2003 through January 2007, we compared survival rates between different patient groups with both parametric methods and Cox regression models. The former group included Weibull, exponential and log-normal regression; we used the Akaike Information Criterion (AIC) and standardized parameter estimates to compare the efficiency of various models. All the analyses were performed with the SAS software and the level of significance was set at P< 0.05. The results showed a significantly higher chance of survival in the following subgroups: those with age at diagnosis < 35 years, lower tumor size and those without metastases (P< 0.05). According to AIC, Cox and exponentials model are similar in multivariate analysis but in univariate analysis parametric models are more efficient than Cox, except in the case of tumor size. Log-normal appears to be the best model. Cox and exponential models have similar performance in multivariate analysis. However, it seems that there is no single model that performs substantially better than others in univariate analysis. The data strongly supported the log-normal regression among parametric models; it can give more precise results and can be used as an alternative for Cox in survival analysis of patients with gastric cancer.