Feature selection methods and resampling techniques in survival data: determination of risk factors of under-five child mortality

Abstract

The main aim of the research was to identify the risk factors of under-five child mortality using Kenya Demographic Health Survey (KDHS) 2014 data set. Demographic Health Surveys are faced with three main challenges, acute class imbalance, high dimensionality and missing data. The KDHS 2014 data set is made up of 1129 variables and 20964 observations. In addition, the mortality class accounted for 4% of the data while the nonmortality class accounted for the remaining 96%. To determine the risk factors, first we dealt with the missing data by imputation. The class imbalance was handled using three balancing methods: both sampling, under-sampling and over-sampling. We then handled high dimensionality using three filter methods. Random survival forest was used to select highly predictive variables and parameter estimation was done using Cox-PH model. The variables that were found to be significant were child is twin, sex of child, births in the last five years, currently pregnant, wanted pregnancy, living children & current pregnancy, wanted last child, respondent slept under mosquito bed net, ideal number of children, disposal of youngest child’s stools when not using toilet, received polio vaccine and weight for age standard deviation.