Proposal Description

This session demonstrates how combining survey and publicly available geospatial data can improve the granularity and timeliness of economic welfare measures, and the importance of rigorous evaluation to better the strengths and weaknesses of different approaches in different contexts.

The first talk proposes an integrated modelling framework that jointly models two sample surveys collected at the same time, both estimating a similar asset-based wealth index, while incorporating geospatial covariates as alternative auxiliary information about the population. A key challenge is harmonizing data of varying granularity and quality through modelling, while ensuring statistical validity. This is addressed through a spatial matching procedure. The proposed joint model yields more reliable estimates than conventional univariate approaches, thereby mitigating errors stemming from excessive reliance on geospatial data. The talk concludes by discussing the implications of this small area estimation framework and potential extensions in the modelling and variance estimation, aiming to support more informed policy decisions in data-limited settings.

The second presentation uses data from Mozambique and the UK to examine, both theoretically and empirically, the properties of small-area estimators that use geospatial zonal statistics as predictors to estimate key income parameters. The results show that geospatial data is a reliable source of auxiliary information that can be used to update small-area estimates during intercensal periods. How unit-context models are specified can impact the quality of the estimates. Area-level models using geospatial data as predictors tend to perform better than unit-context models with area-level predictors, and alternative intercensal updating methods, such as Structure Preserving Estimation (SPREE) methods, also offer an alternative, reliable approach.

The third presentation uses data from Laos to examine the performance of remotely sensed data in generating small area estimates of asset wealth and malnutrition. Estimates using only remotely sensed data are generally more accurate for in-sample areas, and in the case of out-of-sample, are more accurate for urban areas in this context. Second, a unit-level model using clusters as units can perform better than area-level model when the outcome is linear, such as average asset wealth, and offers a more balanced trade-off between variance and bias. Finally, using estimates of average asset wealth as a predictor to estimate stunting rates can improve malnutrition estimations, which are generally less precise.

The fourth presentation uses data from Malawi and Mozambique to evaluate modern geospatial foundation models for hyperlocal small-area estimation and nowcasting. It presents a new pre-trained vision transformer model fine-tuned using low rank adaptation techniques. This model outperforms existing foundation models for both cross-sectional and intertemporal prediction of a household asset index. When fine-tuned with two simulated samples a decade apart, predicted R2s across 20 sq km images in the more recent year are 0.62 in Malawi and 0.52 in Mozambique. When predicting changes, prediction accuracy is sensitive to the amount of household data used for fine-tuning. The results demonstrate that foundation models perform well for hyperlocal small area estimation, and that a lagged georeferenced census is a valuable supplement for fine-tuning foundation models when predicting welfare changes.