Understanding the Risks Factors of Under-Five Child Mortality in Kenya: Random Survival Forest and Accelerated Failure Time Shared Frailty Models

Khaoya Mutaki; Nelson Onyango; Rachel Sarguta

Khaoya Mutaki School of Mathematics, University of Nairobi
Nelson Onyango School of Mathematics, University of Nairobi
Rachel Sarguta School of Mathematics, University of Nairobi

Keywords: Risks factors, Shared Frailty, Random Survival Forest, Accelerated Failure Time

Abstract

Under-five mortality rates is one of the health indicators of great importance to any country. Kenya is among the countries in the Sub-Saharan Africa with high Under-Five Child Mortality (U5CM) rates. It is therefore important to apply best statistical approaches to establish which factors influence child mortality. This will go a long way to inform the optimal design of health intervention strategies within the country and globally. In this study, Random Survival Forest (RSF) and Accelerated Failure Time Shared Frailty Models have been used to analyze U5CM based on the Kenya Demographic Health Survey (KDHS, 2014) dataset. Akaike Information Criterion (AIC) statistics was used to select the model of best fit. Results obtained from fitting the AFT-shared frailty model, showed that there was presence of unobserved heterogeneity at community level. However, there was no evidence to conclude the existence of unobserved heterogeneity at the household level. Among the variants of the AFT Shared Frailty models analysed, the Log-logistic AFT- model showed that “the sons who have died,” “daughters who have died,” “duration of breastfeeding,” and “months of breastfeeding” had significant influence on the U5CM (p <0.05). The Log-logistic AFT model with Gaussian frailty emerged to be the best model for the U5CM since it had the least Akaike Information Criterion (AIC) statistic. On the other hand, the results from Random Survival Forest, “sons who have died,” “daughters who have died,” “living children plus current pregnancy,” “sex of child,” “duration of breastfeeding,” “number of living children,” and “months of breastfeeding” were ranked as important factors that have influence on the under-five mortality. Furthermore, this study also found out that there was presence of unobserved heterogeneity at community level of clustering. At the household level however, there was no unobserved heterogeneity, hence there was no need for household frailty term.