This study aimed at identifying the determinants of Under Five Child mortality (U5CM) based on Kenya Demographic and Health Survey (KDHS, 2014). One of the key challenges with Demographic and Health Survey datasets involves extreme imbalance between the mortality and non-mortality classes. In this particular research only 6.4% of children experienced under five years mortality while 94.6% survived beyond five years. To establish the determinants of U5CM, we opted to handle the class imbalance using four different balancing techniques: Random Under-sampling, Random Over-sampling, Both-sampling, and Synthetic Minority Over-sampling technique. We then did variable selection using Random Survival Forests following the four techniques. The variables selected from each of the four datasets were then used in a Cox-PH regression to determine the effect of select covariates on child mortality, after conducting appropriate model diagnostics. After the analysis, the variables which resulted in increased hazard of child mortality include V206 (Sum of demised sons), V207 (Sum of demised daughters), V203 (Sum of daughters living at home), V218 (Sum of existing children), V238 (Number of deliveries in the last 3 years), HW72 (Weight for height standard deviations) and interaction between B1 (Child’s Month of birth) and V206. Based on model selection indices,Under-sampling balancing schemes performed well for identification of U5CM determinants. By grouping these variables, this study identified birth characteristics of the child (such as age at birth), reproduction factors of the mother (such as number of siblings born before), feeding conditions and anthropometric measurements as key determinants of U5CM.
展开▼
