A girls-education charity working in three regions of Ghana came to us in early 2023 with a question funders kept asking — does the programme actually keep girls in school? — and a year of administrative data nobody had cleaned.
The brief
The charity had run its literacy-and-mentorship programme for four years across fourteen primary schools. Anecdotally, headteachers loved it. Statistically, the picture was murkier — drop-out rates had moved, but so had national averages, school catchments, and at least three competing interventions in the same districts.
Our brief, plainly: find out whether the programme works, separate from everything else moving in the background. The funder needed a defensible answer before its 2025 grant cycle.
Our approach
We proposed a quasi-experimental evaluation using a difference-in-differences design — comparing drop-out trajectories in programme schools against matched non-programme schools in the same districts, before and after the programme started.
Three things made the design defensible:
- Pre-programme parallel trends established across four years of administrative data.
- Match on six covariates including baseline drop-out, gender ratio, and distance to the nearest secondary school.
- Robustness checks against three plausible threats — selection, spillover, and concurrent interventions.
We don't lead with method. We lead with the question. The method is in service of an honest answer — never the other way around.
What we found
Drop-out fell by 38 percentage points more in programme schools than in matched comparison schools over the 2020–2023 window. The effect was concentrated in girls aged 11–13 — the age band the programme had quietly been most invested in, though no-one had said so on paper.
What changed for the charity
The funder tripled its grant in the next cycle. Two new philanthropic funders signed on the strength of the brief. Most usefully, the implementing team finally had language for what they'd been doing intuitively for four years — and used it to redesign year-five around the 11–13 cohort.
What we'd do differently
We'd have built the qualitative interviews into the design from the start. The quantitative result was clean; the mechanism — why the programme worked for that age band specifically — became clear only in the final month, and only after a hurried week of school visits. Next time, listen earlier.
Methodology, instruments, and full data tables available on request. We share the working with anyone seriously evaluating a programme of their own.