Evaluating area-level features for proxy means test models: evidence from rural, semi-urban and urban districts in poverty targeting

While satellite imagery and administrative records have been widely used for poverty mapping, their role in poverty targeting, particularly within the Proxy Means Test (PMT) framework, remains underexplored. This study evaluates the usefulness of area-level features, such as Night-Time Light intensity, vegetation and built-up indices (NDVI, NDWI, NDBI), land cover types and the number of schools and health facilities, in improving household-level expenditure estimates and reducing targeting errors in PMT models. Using household data from a variety of district types in Indonesia, including rural, semi-urban and urban districts, we assess how these features affect model performance in terms of inclusion and exclusion errors under both 20% and 40% poverty thresholds. Our findings show that area-level features improve PMT model performance primarily in districts with a high number of sub-districts, particularly in semi-urban settings. In contrast, in urban and rural districts with limited sub-district granularity, these features often introduce noise or degrade accuracy. We also observe systematic misclassification patterns: poor households in more developed sub-districts tend to be wrongly excluded, while non-poor households in lower resource areas are wrongly included. Using the Shapley value approach, we show that the predictive effect of features such as NDVI varies by context, for instance, in some districts, NDVI positively correlates with expenditure due to productive croplands, diverging from broader trends. These results underscore the importance of spatial granularity, threshold sensitivity and local context when integrating area-level data into poverty targeting systems.