In 2007 a lender called Bharatha Swamukti Samsthe began offering microfinance in villages around Bangalore. It told a few "leaders" in each village (teachers, shopkeepers, self-help group heads) and let word of mouth do the rest. Banerjee, Chandrasekhar, Duflo and Jackson had surveyed every household's social ties beforehand, so they could watch the product move through 43 real networks. Their paper, The Diffusion of Microfinance (Science, 2013), fitted a model in which every informed household tells each neighbour with some probability per period: qP if it took a loan, qN if it did not. The published fit was qN = 0.10 and qP = 0.50. People who adopt talk about it far more.
We opened a research programme on that data today (research/civ-seeding). The question is whether a product like this spreads as a simple contagion (one informed friend is enough) or a complex one (you need several adopting friends), and whom a rollout should seed if the data can't tell. The application we have in mind is the first households in a solar-home-system or cookstove rollout. Before testing anything new, we ported the authors' archived Matlab estimation code to Python and re-ran it on their own data (CC0).
What reproduces. Non-leader take-up averages 18.7% across the 43 villages; the paper says 18.5%. Leaders: 24.7% against 24.8%. Adopters pass the word on more than non-adopters in 98.4% of 1,000 bootstrap draws. Silence the non-adopters (set qN to 0) and simulated take-up falls by a third (0.337). The paper says it falls from 20.2% to 13.7%, a drop of 0.32.
What does not. The best-fitting point, with the archived grid, moments and weights, is qN = 0.05 and qP = 0.30, the same for three simulation seeds. The published point scores 11 times worse on the same criterion. The qP direction is nearly flat: the bootstrap's 95% interval runs from 0.15 to 1.0, and the criterion stays within a factor of 2 of its minimum from 0.25 to 0.50. qN = 0.10 is clearly worse than 0.05 at every value of qP. Two checks we added after seeing the gap point the same way, not toward 0.50: two-step optimal weights give qP 0.20 to 0.25, and four-month periods replaced by quarters give 0.15 (months give qN 0.01, qP 0.10).
We don't know yet why. The published table may use moments, weights or a finer grid that the archived script doesn't default to. The Science version, which we did not open, may differ from the 2012 working paper we read. Or our port differs in a way the ten tests miss. The honest reading is narrower than "the paper is wrong". The data and the qualitative story hold up. The second parameter is pinned down much less tightly than one number suggests.
That weak identification is the programme's real subject. If the strength of word of mouth from adopters can't be pinned down from who ended up with a loan, it's an open question whether the same data can tell news-like spreading from reinforcement-driven spreading. And those two give opposite seeding advice: central people for one, tight clusters for the other. Next comes fitting a reinforcement model to the same villages and asking how often the moments pick the right model on villages simulated from each.
Code, data and the pre-registered log: research/civ-seeding. Produced with AI assistance; no affiliation with the study's authors is implied.