Exposure-Adjusted Pedestrian Harm Is 16 Percent Higher per 10 Percentage Points of Neighborhood Poverty in New York City

Pedestrian injury and fatality rates are frequently reported to be elevated in lower-income neigh- borhoods, but population-based rates confound infrastructure hazards with variations in pedestrian exposure. We integrate modeled pedestrian volumes for New York City with street centerline geometry and five years of geocoded police-reported collision records (2019–2023). Crosswalk segments are identified and clustered into 41,937 discrete crossing sites, predominantly intersections, to which collisions are spatially assigned. Across 38,477 sites located in census tracts with complete covariate data, pedestrian injury and fatality collisions increase by 16% per 10-percentage-point increase in the tract poverty rate (rate ratio [RR] 1.16, 95% CI [1.12, 1.20]), conditioning on pedestrian crossing volume, roadway width, lane counts, posted speed limits, tract-level commuting mode shares, land use, and borough fixed effects. Using a model conditioning only on pedestrian volume yields an RR of 1.30 (95% CI [1.26, 1.35]). The estimate is stable across alternative count estimators, site grouping distances, collision assignment radii, temporal partitions, borough exclusions, clustered covariance structures, and designated truck route controls. The estimate attenuates to 1.10 under ZIP-code fixed effects and increases to 1.19–1.22 after correcting for Census sampling error in poverty estimates. Pedestrian fatalities show a similar and slightly steeper gradient (RR 1.29, 95% CI [1.17, 1.42]), which reflects more collisions and, less precisely, a higher probability that a collision is fatal. Conventional models relying on residential population proxies yield attenuated ratios of 1.06–1.09, indicating that exposure adjustment, rather than the scale of aggregation, accounts for the larger estimate. These estimates reflect observational associations in the absence of comprehensive vehicle volume counts. We provide an open-source pipeline to reproduce all data processing, models, and figures from public records.