Modelling child mortality via discriminant analysis and logistic regression

Abstract

Prevalence of deaths of children has particularly become a global concern in strategic decision making in the field of health sector. In Sri Lanka, the risk of deaths at childhood period was higher during the past few decades. Many studies have concerned about the child mortality in various perspectives. The purpose of this study is to find the significant factors on under-five mortality and to recommend a most suitable statistical model to predict the child mortality, under aged 0-5 years of age. The secondary data was collected from the demographic and health survey (2016) conducted by the Department of Census and Statistics (DCS), Sri Lanka. Two types of statistical models: linear discriminant model and binary logistic model are statistically evaluated. Two models were evaluated with classification accuracy, ROC curve, sensitivity/ specificity and sample size variations. Both methods found that, gender of child, marital status, mother’s literacy, status of antenatal care, delivery type, pregnancy duration and decision-making ability are significantly influential variables (p < 0.05) on the status of child mortality. According to the classification results produced by models, discriminant model correctly classified the 89.6% of grouped cases and binary logistic regression model correctly classified the 94.6% of grouped cases irrespective of the status of child mortality. With respect to the all seven indicators, it was found that binary logistic regression model was more efficient and more effective than linear discriminant model. The inferences derived can be effectively used for strategic decision making in the health sector for reducing the child mortality in the future.