Survival analysis of patients with multidrug-resistant tuberculosis in KwaZulu-Natal, South Africa : a comparison of cox regression and parametric models

Abstract

Researchers in medical sciences often prefer the Cox semi-parametric model instead of parametric models because of its restrictive distributional assumptions, but under certain circumstances, parametric models estimate the parameters more efficiently and powerful than the Cox model. The objective of this study was to compare the Cox and parametric models by studying a dataset of patients diagnosed with multidrug-resistant tuberculosis (MDR-TB). A total of 1 542 patients were included in the study from four decentralised sites located in rural areas and one centralised hospital in KwaZulu-Natal, South Africa from 1 July 2008 to 30 July 2012. Out of 1 542 patients with MDR-TB, 886 (57.5%) were cured and 245 (15.9%) died. According to the AIC, the Lognormal and Weibull regression models were the best fitting to data and the Cox regression model was the weakest. According to the results from parametric models, baseline weight of patients had an increased risk of death in both univariate and multivariate analysis. Patients with ages 31 – 40, 41 - 50 and >50 years at diagnosis had an increased risk for death in Cox proportional hazards model. In univariate analysis the data strongly supported the Lognormal regression among parametric models, while in multivariate analysis Weibull and Lognormal are approximately similar, according to Akaike Information Criterion. Although it seems that there may not be a single model that is substantially better than others, Lognormal is the most favorable as an alternative to Cox for identifying risk factors for patients with MDR-TB.