Adaptive Sampling Design

Context For a multi-year agricultural livelihoods evaluation spanning two countries, I was asked to design and deploy the endline survey, building on a sampling frame established at baseline several years earlier.

The Challenge Between baseline and endline, the on-the-ground reality had shifted substantially in both countries: population growth in one had changed the underlying sampling universe, while changes to the programme’s operating footprint in the other meant the original frame no longer matched who the programme was actually reaching. Applying the original sampling approach unchanged risked comparing the wrong things at endline, undermining the evaluation’s ability to say anything credible about change over time.

What I Did I designed a hybrid panel and cross-section sampling approach in Stata, preserving a genuine panel element where possible while adjusting for the areas where the original frame no longer held. This kept the evaluation’s before/after comparisons statistically valid despite two years of real-world change. Analysis was then carried out in R, an inherited codebase, using AI-assisted debugging to work efficiently through it, with every analytical step sense checked to preserve rigour.

Outcome The evaluation was able to draw a valid before/after comparison despite significant contextual change in both countries, something a mechanical reapplication of the original design would not have delivered.