Using machine learning to predict where to scale girls’ education programs

• In Event: How can education leaders leverage big data and AI when scaling?

Proposal

After successful randomized control trial results in 2018, for its program to enrol girls into government schools and improve learning outcomes in remote rural India; Educate Girls, an Indian NGO, was determined to identify ways to take this successful program to scale with an inclusive approach and in partnership with the state education ministries. Since inception, Educate Girls had enrolled 345,000 girls into government schools, however, it was frustrated by this pace of expansion. By instituting new data and performance management systems during the RCT (undertaken as part of the world’s first Development Impact Bond), Educate Girls and its data and evaluation partner IDinsight began to understand that it should be possible to target interventions better through the use of predictive analytics, thereby reaching more girls, faster.

Traditionally, Educate Girls’ approach involved conducting door-to-door surveys of households in nearly every village in an expansion region, to identify girls and whether they were enrolled in school or not. This process was enormously resource-intensive and made achieving the organisation’s target of reaching 1.56 million out-of-school girls by 2024 a challenge.

To address this issue, in partnership with IDinsight, a machine learning algorithm was developed, to identify concentrations of out-of-school girls. They first combined their historical door-to-door survey data from 3 million households across 8,000 villages with comprehensive public datasets, including the 2011 census and DISE educational data encompassing 313 socio-economic and cultural indicators. They then set up the algorithm to use this data to predict the number of out-of-school girls in each village, and then cluster villages with the highest concentration, together. This allowed Educate Girls’ teams to identify potential hotspots with an above average number of out-of-school girls and target high-burden areas, most in need of educational outreach.

As a consequence, Educate Girls were able to effectively allocate their field resources to these ‘hot spots’ and spot check surveys then allowed them to assess accessibility to villages and plan for reduced travel time within clusters. In other words, by hiring and deploying staff and volunteers to these clusters of villages, rather than working in all the villages in a given administrative district, Educate Girls were able to find villages with 2.5 times as many out-of-school girls per village compared to those found through their traditional survey techniques.

Crucially Educate Girls and IDinsight have also iteratively refined the algorithm to increase precision. For example, adding new public data sources, adding more door-to-door survey data, and updating and re-training the model for new regions. Through this iterative refinement they have applied the model to 50,000 villages across 3 States beyond Rajasthan, namely Madhya Pradesh, Uttar Pradesh and Bihar. This has enabled Educate Girls to prioritize the top 15-20% of villages per district, and thus enrol more girls into school faster.

Since first applying the machine learning model to their work more than 5 years ago, the total impact of this algorithm has amounted to Educate Girls identifying and enrolling more than 500,000 additional out-of-school girls in school than would otherwise have been possible.

Author

• Akhil Agrawal, Educate Girls