Gamuchirai-Magamba/zw-marriage-risk

District-level child marriage estimates for Zimbabwe, with uncertainty # zw-marriage-risk **Where should child marriage prevention programmes go in Zimbabwe?** District-level prevalence estimates with honest uncertainty. ### 🗺️ Open the map · API docs · Model card --- --- ## The problem **One in three Zimbabwean girls marries before she turns 18.** That figure has barely moved in a decade — 32.4% in the 2015 DHS, 33.7% in the 2019 MICS. An organisation with funding for, say, sixty wards has to choose where to work. Zimbabwe has 91 districts and over 1,900 wards. National statistics cannot answer that question, and neither can provincial ones — a province like Mashonaland West holds two million people. So programmes are targeted largely on judgement. ## The methodological problem The obvious fix is to compute a rate per district. **It does not work.** ``` Direct district estimates (weighted average of each district's own respondents): range 3.3% to 75.1% districts with under 20 women 23 of 91 smallest district 5 women ``` Look at the smallest districts: | district | women | married | "rate" | give or take | |---|---|---|---|---| | kariba | 5 | 3 | 59.3% | **±22 points** | | gokwe south urban | 8 | 5 | 72.8% | ±16 points | | mbire | 11 | 8 | 75.1% | ±13 points | Kariba's five respondents could plausibly represent anything from 15% to 100%. And notice that the districts ranked *worst* are largely the ones sampled *least* — small samples produce extreme numbers. **A map built on this would send money wherever the survey happened to be thinnest.** ## The approach A **multilevel logistic regression** with a random intercept per district, fitted on both surveys pooled (3,487 women aged 20–24, 839 clusters, all 91 districts). Each district gets its own estimate, but those estimates are drawn from a shared distribution. The consequence is **partial pooling**: - a district with 399 respondents moves the model, so its estimate stays close to its own data and its interval is …