Experimental Design
Eligible schools were identified by FIs. (see the attached pre-analysis plans for details on eligibility). Selected schools were randomly assigned to receive an offer of an IfL loan or to the control group, which received the traditional loan offerings from their FI.
Randomization proceeded in three batches between September to December 2025. The reason for three rounds of randomization was because, following the first round of randomization (Batch 1, N = 360), FIs identified some schools in the randomization sample that were not eligible for IfL loans, due to payment history, school closures, and other reasons. Thus, the FIs provided lists of additional eligible schools in November (Batch 2, N = 114) and December (Batch 3, N = 76), which were then randomized into treatment and control groups.
In Batch 1, schools were assigned to one of 12 strata, defined by which FI’s portfolio they fell in, whether the school had baseline assessment scores from participating in the benchmarking exercise, whether the school’s baseline reading performance was above or below the median school’s performance, and whether the school’s average per-term fee was above or below the median school. In Batches 2 and 3, due to the smaller number of schools and the absence of baseline data, strata were only defined by FI portfolio. Within each stratum, 50% of schools were randomly assigned to the treatment group and the remainder to the control group.
This process resulted in a study frame of 510 schools (255 treatment, 255 control). However, we faced two challenges with including all 510 schools in our endline sample. First, take-up was lower than expected. In our sample size calculations, we assumed that take-up would be at least 70%. However, by July 2026, only 48% of treatment schools had taken on an IfL loan. With 48% take-up, minimum detectable effects are roughly twice as large as when take-up is 100%. Since we believe that even modest impacts from IfL loans on learning outcomes are important to identify, this reduction in precision was unacceptable. Second, our budget assumed that we would conduct assessments in 300 schools.
Given these challenges, we sought to remove schools from our sampling frame such that the resulting treatment sample would have higher take-up, without undermining the integrity of random assignment. We describe this process in detail in the attached pre-analysis plan. This process resulted in 267 schools remaining for our endline assessment. These 267 schools are well-balanced across treatment/control overall and within each FI subsample: PC treatment: N = 62, PC control: N = 62; JF treatment: N = 72; JF control: N = 71. Critically, the 134 treatment schools in our endline sample have higher take-up rates (70%), as of July 2026, than the original 255 schools assigned to treatment (48%). We will attempt to conduct assessments in all of these schools; however, based on our experiences during the benchmarking exercise and midline, we expect that a few schools may refuse to participate.