Predicting Digital Inclusion in Somalia Using Machine Learning

Abstract

Digital inclusion affects participation in education, communication, financial services, and the wider digital economy, but individual-level evidence from Somalia remains limited. This study uses microdata from the 2022 Somalia Integrated Household Budget Survey to classify two outcomes among people aged 15 years and older: mobile-phone ownership and internet use during the previous three months. Individual records were merged with household, welfare, and asset information. Logistic regression and random forest classifiers were trained using an 80:20 stratified split and evaluated with accuracy, precision, recall, F1-score, receiver- operating-characteristic curves, and area under the curve (AUC). Random forest outperformed logistic regression for both outcomes. For mobile-phone ownership, random forest achieved accuracy of .841, an F1-score of .903, and an AUC of .860. For recent internet use, it achieved accuracy of .792, an F1-score of .723, and an AUC of .880. Feature-importance results associated mobile-phone ownership mainly with age, household position, assets, expenditure, and household size, whereas internet use was more strongly associated with literacy, education, electricity, assets, and household welfare. The models identify predictive associations rather than causal effects. The findings suggest that policies focused only on device access may be insufficient; broader digital inclusion also requires literacy, education, reliable electricity, affordable connectivity, and improved household economic conditions.