Maternal Morbidity and Mortality Through Open Data: A County-Level Association Framework Using Conventional Regression and Interpretable Machine Learning, United States, 2018–2022

This registration describes a county-level study of which community conditions are associated with serious complications of childbirth in the United States, using only data that anyone can download. The outcome is the share of births in a county with at least one of the five maternal morbidity items recorded on the US birth certificate: blood transfusion, admission to intensive care, ruptured uterus, unplanned hysterectomy, and severe perineal tear. It is analyzed for the 570 counties with 100,000 or more residents, the counties for which federal birth data are published at county level, with births pooled over 2018 to 2022 (about 14.5 million births, 79% of all US births in those years). This composite is a measure of complications recorded at delivery; it is not the CDC severe maternal morbidity measure and is not treated as equivalent to maternal mortality. The purpose is twofold. First, to estimate, with a plan fixed before any county analysis, how strongly six community conditions are associated with the county complication rate: the percent of the population that is Black, obstetric access by the March of Dimes designation, the share of women aged 19 to 44 without health insurance, the unemployment rate, household broadband adoption, and the percent of the population that is Hispanic. Five of these carry a stated direction and one does not. Second, to show whether an interpretable machine-learning method and a spatial model, applied to the same counties, add to what conventional regression shows, without letting any comparison decide which result the paper reports. The analysis runs in four phases. Phase 1 uses negative binomial regression, a count model that allows counties to differ more than chance alone would produce: a screening pass across 22 candidate conditions, a stability analysis that records how often each condition is selected across 100 random half-splits of the counties, one confirmatory regression on the six conditions with a separate baseline for each state, and one extended regression that adds any condition surviving the stability step by a rule fixed in advance. Phase 2 fits an Explainable Boosting Machine, a model that learns a readable curve for each condition and a limited number of pairwise interactions, and a spatial regression in which the average level of each condition in neighbouring counties enters as an additional term, with a test of whether neighbouring counties still resemble each other after the conditions are accounted for. Phase 3 evaluates every model the same way, holding out one state at a time and comparing each model's expectations with what happened in the held-out counties, and publishes the full comparison table; no model is selected on the basis of it. Phase 4 tests, rather than assumes, that the complication measure tracks maternal mortality, using published state mortality rates for 39 states, county death counts for the 85 largest counties, and exact county counts from Florida. If the state-level test fails, the paper reports its findings as complication findings and removes all mortality-proxy language. Expected outcomes are: a rate ratio with a confidence interval for each of the six conditions, reported against a reading rule fixed here; screening results for all 22 candidates, published whatever they show; the shape of each relationship and any interactions found by the machine-learning model; a test of residual geographic dependence; one table placing all models side by side on identical held-out states; the result of the mortality validation and its pre-committed consequence; and a plain-language summary for a policy audience that keeps predictive importance, statistical association, and causal effect apart. The study makes association claims only and uses no causal-inference method. One expected finding is already visible in the data structure: the 570 counties with published county birth data contain no maternity care deserts and no county in the four most rural categories, so a county-level tool built on open US data cannot see the rural places that policy most needs to reach, and the paper reports that absence as a result. The analysis plan, every model specification, and the consequence of each test failing are registered before any association between the outcome and any candidate condition is computed, and prior exploratory analyses by a co-author on other outcomes are disclosed in the registration itself.