A machine learning approach to assessing multidimensional poverty and targeting assistance among forcibly displaced populations

• Data science enables poverty targeting beyond PMT and unidimensional expenditure metrics. • Machine learning can identify refugee households most in need using flexible, multidimensional measures. • Geospatial and place-based factors are essential for accurate poverty prediction. • Traditional targeting methods may overlook food-insecure households in need of aid. • Regular updating poverty metrics and predictive models improves targeting over time. Increasing trends in forced displacement and poverty are expected to intensify in coming years. Data science approaches can be useful for governments and humanitarian organizations in designing more effective targeting mechanisms. This study applies machine learning techniques and combines geospatial data with survey data collected from Syrian refugees in Lebanon over the last four years to help develop more effective and efficient targeting strategies. Our proposed approach helps: (1) identify the households most in need of assistance based on a flexible, multidimensional poverty metric and (2) operationalize this method without resorting to impractical and expensive data collection procedures. Our findings highlight the importance of a comprehensive and versatile framework that captures other poverty dimensions along with the commonly used expenditure metric, while also allowing for regular updates to keep up with (rapidly) changing contexts over time. The analysis also points to geographical heterogeneities that are likely to impact the effectiveness of targeting strategies. The insights from this study have important implications for agencies seeking to improve targeting and increase the efficiency of shrinking humanitarian funding.