Decision rules, thresholds and input uncertainty in robust ward prioritisation for healthcare needs assessment in Uasin Gishu County, Kenya

Abstract

Background. Accessibility maps are often turned into shortlists of areas for follow-up, but the shortlist depends on the decision rule, the threshold, the number of areas selected and the uncertainty considered. We examined how these choices shape a six-ward needs-assessment shortlist in Uasin Gishu County, Kenya. Methods. Ward scores were the modelled population beyond a walking-time threshold to the nearest government (or wider-ownership) dispensary or health centre, under two within-ward population allocations. We compared three rules: the best set under a single reference scenario, the best set by equal-weighted scenario mean, and the minimax-absolute-regret set. Primary analyses used four scenarios, a 60-minute threshold and six wards. Extensions varied the threshold (30, 60, 120 minutes), the shortlist size (3, 6, 9) and the objective (count or proportion). A further analysis added four scenarios in which a slow friction land-cover class in dense cells was replaced. All primary optima were confirmed by enumerating all 593,775 six-ward sets. Results. With four scenarios, maximum regret was 4,923 modelled people for the minimax set, 6,659 for the equal-mean set and 10,319 for the reference set; minimax and equal-mean differed by one ward. When the friction class was replaced, the minimax and equal-mean sets became identical, and Huruma was replaced by Tarakwa. Across all eight scenarios, four wards (Karuna/Meibeki, Ziwa, Kaptagat, Cheptiret/Kipchamo) were among the six highest-scoring wards in at least seven scenarios. Minimax and equal-mean sets differed in 3 of 18 extended configurations. At 120 minutes only one ward had a positive score, so the remaining slots in any fixed-size list were arbitrary. Conclusions. In this setting, input uncertainty, especially in the friction surface, changed the shortlist more than the choice between robust and average decision rules. Shortlists should be reported with their stable core, their specification-dependent members, and any slots filled without supporting evidence. Preprint, not peer reviewed. Companion papers: measurement sensitivity, decision analysis, and external check of inputs (Uasin Gishu County, Kenya). Code and derived data: S1 zip.