Finer partitions, weaker evidence: marginal fairness hides intersectional harm across five Sub-Saharan African Low-and-middle-income Countries

Abstract Background Fairness audits in artificial intelligence (AI) still test one protected attribute at a time. We prove by construction that single-axis parity does not bound intersectional parity, then test this gap empirically on stunting-risk models built on child health data from five low-and-middle-income African countries. A second question follows directly: can that evidence be trusted once the obvious fix, a finer partition, strains a limited sample? Results Auditing more finely does not automatically buy more trustworthy evidence. We apply the counterfactual false positive/false negative rate framework of Wastvedt, Huling and Wolfson (2024) to stunting-risk models on Demographic and Health Survey data from Nigeria (2024; n = 5,747; 30 cells), Ghana (2022; n = 2,833; 73 cells), DR Congo (2023-24; n = 6,678; 112 cells), Liberia (2019-20; n = 1,568; 23 cells), and Ethiopia (2024-25; n = 3,481; 63 cells). In every country, the intersectional maximum pairwise gap (Δ_MAX) exceeds both single-axis Δ_MAX values, on both cFPR and cFNR: Nigeria (0.584, 0.849), Ghana (0.918, 1.000), DR Congo (0.658, 0.906), Liberia (0.398, 0.635), Ethiopia (1.000, 1.000). A permutation u-value test (R = 500, δ = 0.1) tells a sharper story: strong evidence for this excess appears only where the partition is coarse relative to the primary-sampling-unit (PSU) count (Nigeria, u = 0.836–0.866), while the four finer-partitioned countries return markedly weaker evidence for the same, often larger, gap: a partition-granularity effect in which a finer partition on a smaller sample inflates both the gap and its null. Conclusions An intersectional audit is not automatically more trustworthy than a marginal one simply because it is finer. Across five low-and-middle-income country (LMIC) settings, statistical confidence in an intersectional gap tracks partition granularity relative to sample size and PSU count, not gap size alone. Theory and measurement also agree without exception on the prior claim: a fairness audit stopping at single protected attributes understates the disparity a model carries at the intersections. For the Sustainable Development Goal (SDG) 10 of the United Nations, neither finding is a footnote; together they separate a model certified fair from a model that is fair.