Comparison between Semi-Parametric Cox and Parametric Survival Models in Estimating the Determinants of Under-Five Mortality in Nigeria: Application in Nigerian Demographic and Health Survey
Publication Date: 23/11/2019
Author(s): Anthony I. Wegbom, Prof. Victor A. Kiri, Prof. Isaac D Essi.
Volume/Issue: Volume 2 , Issue 2 (2019)
Abstract:
The increasing application of survival analysis in estimating the determinants of under-five mortality, call for the need to use the most flexible and efficient model to estimate reliable determinants free from bias. Most population health and medical research that involve survival analytical techniques are based on the semi-parametric Cox models instead of the parametric versions because of the flexibility associated with the former which does not require specification of the baseline hazard function. In this study, we compare the semi-parametric Cox model and parametric models in estimating the determinants of under-five mortality (U5M) in Nigeria. The 2013 Nigeria Demographic and Health Survey data was used for this study. In order to identify determinants of U5M, the Cox and parametric (exponential, Weibull and Gompertz) models were fitted to the data. The Akaike’s information criterion (AIC) and the Cox-snell residual were employed to find the best model. Based on the AIC and Cox-snell residual, the Cox model had the poorest and Gompertz model had the best fit to data. After controlling for demographic, socioeconomic and healthcare-related variables, residing in the North-west (HR=1.40, 95% CI=1.03-1.90) as well as residing in the South-east (HR=2.14, 95% CI=1.46-3.14); married women (HR=0.65, 95% CI=0.48-0.89); living in the rural areas (HR=1.46, 95% CI=1.14-1.86) and delivered in the health facility (HR=0.76, 95% CI=0.59-0.97) had significant effects on U5M in Nigeria. We conclude that for our data, the parametric models out performed that of the semi-parametric Cox model. We therefore recommend that researchers should conduct formal assessment of the goodness of fit for each of the models under consideration, to inform their decisions on most appropriate final model. As part of this process, we also recommend the involvement of a statistician for the selection of candidate models for their data.
