Predicting household poverty with machine learning methods: the case of Malawi

Poverty alleviation continues to be paramount for developing countries. This necessitates the need for poverty tracking tools to monitor progress towards this goal and effect timely interventions. One major way poverty has been tracked in Malawi is by carrying out integrated household surveys every five years to quantify poverty at local and national levels. However, such surveys have been documented as very expensive, tedious, and sparsely administered by many low- and middle-income countries. Therefore, this study looked at whether machinelearning models can be used on existing survey data to predict poor and non-poor households, and whether these models can predict poverty using a smaller number of features than those collected in integrated household surveys. This was achieved by comparing the performance of three off-the-shelf, open-source machinelearning classification algorithms namely Logistic Regression, Extra Gradient Boosting Machine and Light Gradient Boosting Machine, in correctly predicting poor and non-poor households from Malawi survey data. These supervised learning algorithms were trained using 10-fold cross-validation. The experiments were carried out on the full panel of features which represent all the questions asked in a household survey, as well as on smaller feature subsets. The Filter method and SHapley Additive exPlanations method were used to rank the importance of the features, and smaller data subsets were selected based on these rankings. The highest prediction accuracy achieved for the full panel data set of 486 features was 87%. When the Filter method rankings were used, the models' prediction accuracy dropped to 63% for the top 50 features subset. However, using the SHAP method rankings, the maximum prediction accuracy level was maintained and only dropped slightly to 86% with the top 50 feature subset; to 84% with the top 20 features; and 73% for the top 10 features. Area under the Curve, Receiver Operating Characteristic curve, recall, precision, F1 score, Matthews Correlation Coefficient and Cohen's Kappa scores confirmed the classification models' reliability. The study, therefore, established that poverty can be predicted by open-source machine learning algorithms using a substantially reduced number of features with accuracy comparable to using the full feature set. A policy recommendation is to employ only the top explanatory features in surveys. This will enable shorter, lower-cost surveys that can be administered more frequently. The aim is to assist policymakers and aid organisations to make more timely interventions with better targeting of the poorest