sbrunelli/uganda_child_mortality

Predictive Model for Child Mortality in Uganda # Predicting Mortality of Children under 5 years of age In this project we set 2 goals: 1. **Inference**: Identifying factors that are predictive in nature of child mortality 2. **Predictive**: Build a model that would allow to classify households and assign them a child mortality risk score I worked in collaboration with Living Goods, a San Francisco based NGO, already operative on the field since many year, specifically in Uganda and Kenya. Community Health Providers (CHPs) are identified in each village, they are given a basic education and the tools to visit families in their parishes and provide them with first health assistance. Using a web application developed by Living Goods and deployed on Android phones, each CHP can answer basic questions and gather data as they visit households, but for legal reasons, they cannot record deaths. Our problem was to find a proxy of information that contained both general characteristics of households and mortality data. For that purpose we used data from a Randomized Controlled Trial (RCT) in the form of surveys that targeted the exact same population. ## Inference The first phase of the project was inferential in nature: identify questions that relate to child mortality. The main challenge of the project lied in the data: * pretty **small data** (6700 data points, one data point being a survey) * pretty **fat data** (361 base features, one feature being one question) * extremely **unbalanced classes** (5.33% minority class, minority class being child death) * very **low signal** (almost all questions were categorical in nature, yes/no kind of question, Likert scales when we were lucky) We run each single feature through a separate statistical test: Chi-squared vs ANOVA depending upon the nature of the feature (categorical vs quantitative), and reported the resulting p-values. We chose to set significance at 0.1, which is pretty common in these kind of low signal scenarios, but corrected for the number of tests we …