Combining satellite imagery and machine learning to target social protection in Pakistan

In Pakistan, the newly established Sindh Social Protection Strategy Unit (SPSU) is in charge of implementing three cash transfer programs: (i) a conditional cash transfer for pregnant and lactating women, (ii) an unconditional cash transfer for female agricultural workers and (iii) an unconditional cash transfer for food insecure households.

This project aims to help the government design a dynamic, shock responsive system for targeting these interventions. To do this, the team will use machine learning to create small area estimates of socioeconomic conditions across Sindh province. This will highlight the villages or ‘clusters’ of greatest (or least) need, allowing them to be targeted for aid from SPSU.

The team will validate the predictive accuracy of their approach in two ways. First, they will use in-sample validation by dividing a very large sample of existing household survey data into a training and validation set. They will train the network on one set, then estimate for the second, and evaluate the accuracy of their predictions. Given the large volume of data available, this should allow a high degree of confidence in the accuracy metrics. Second, the team will use out-of-sample validation by conducting a new household survey in areas for which they do not have ground-truthing data. They will make estimates of village/cluster poverty scores using machine learning and compare with the novel survey data.