Historical Redlining and Spatiotemporal Patterns in Breast Cancer Screening · Rahman MA, Ratnayake I, Pepper S et al.

Methods

A longitudinal ecological study was conducted to examine Census tract–level breast cancer screening rates in the KUCC catchment area (123 counties, including 105 in Kansas and 18 in Missouri) from 2016 to 2024. In accordance with University of Kansas Medical Center institutional policy, this study was exempt from ethics review and informed consent requirement because it used publicly available, deidentified, aggregate data. We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cohort studies.

Annual Census tract–level breast cancer screening prevalence estimates were obtained from the CDC PLACES database and linked to historical redlining and present-day neighborhood characteristics via the Census tract geographic identifier using publicly available datasets.20 The CDC PLACES breast cancer screening measure represents the estimated percentage of women aged 50 to 74 years who report receiving a mammogram within the recommended screening interval and should not be interpreted as annual mammography use or individual-level adherence to yearly screening.

Historical redlining was measured using digitized HOLC residential security grades (grades A to D) from the Mapping Inequality project.10 HOLC grade assignment was harmonized to 2010 Census tract boundaries because the HOLC tract–based linkage file and historical redlining overlap calculations were based on 2010 tract geography. Digitized HOLC polygons were intersected with 2010 Census tract boundaries, and the percentage of each Census tract area overlapping each HOLC grades A to D was calculated. For tracts overlapping multiple HOLC grades, the final tract–level HOLC grade was assigned using the grade with the largest percentage area overlap. Present-day social context was characterized using the SVI, which was categorized into national quintiles, with the first quintile indicating the lowest vulnerability and the fifth quintile indicating the highest.17 Additional Census tract–level covariates captured socioeconomic conditions, household composition, racial and ethnic diversity and language status, and housing and transportation constraints (Table 1).21 The SVI was included as a composite measure of overall neighborhood vulnerability, whereas selected individual covariates were included to represent specific socioeconomic, housing, and transportation pathways that may link historical redlining to screening prevalence; conditional estimates for these covariates were interpreted cautiously because of conceptual overlap with the SVI. The screening outcome was measured annually from 2016 to 2024, whereas HOLC grade, SVI, neighborhood covariates, and mammography facility distance were treated as Census tract–level contextual measures that were fixed over the study period or measured using the closest available data source. Geographic access to screening was measured as straight-line distance (miles) from each centroid to the nearest Mammography Quality Standards Act–certified mammography facility using the US Food and Drug Administration facility database.26 Rurality was not included because the analytic sample comprised urban tracts only and lacked meaningful variation.

Table. Breast Cancer Screening Prevalence and Census Tract–Level Socioeconomic Characteristics by HOLC Redlining Grade (See the original article for the full table.)

A total of 9071 Census tracts were initially considered for inclusion across the study period. Of these tracts, 2519 were identified based on the availability of HOLC grades (A to D). Among these, 224 tracts were excluded because of missing analytic variables, yielding 2275 tracts. To ensure 1 record per Census tract in each year, duplicate tract-year records were resolved after spatial linkage between Census tract boundaries and HOLC polygons. These duplicates occurred when a Census tract overlapped more than 1 HOLC polygon or grade. For each tract-year, the percentage of tract area overlapping each HOLC grade was calculated, and the tract was assigned the HOLC grade with the largest percentage area overlap. After this assignment, duplicate tract-year records were collapsed to 1 record per census tract-year. This process resulted in a final analytic sample of 1152 Census tracts for the bayesian spatiotemporal β regression analyses. The data cleaning and selection process is shown in eFigure 1 in Supplement 1. The initial count reflected tract-year records from 2016 to 2024 rather than unique Census tracts. A formal comparison of HOLC-graded and non–HOLC-graded tracts was not performed because HOLC maps were created primarily for historically urban areas and do not cover the full catchment area. Therefore, findings were interpreted as applicable to historically HOLC-graded urban tracts rather than all Census tracts in the catchment area.

Statistical Analysis

Descriptive characteristics were summarized by HOLC grade using medians and IQRs or counts and percentages. Group differences were assessed using Kruskal-Wallis tests for continuous variables and Pearson χ2 tests for categorical variables.

Breast cancer screening prevalence, modeled as a proportion bounded between 0 and 1, was analyzed using bayesian spatiotemporal β regression with a logit link.27,28 The model included fixed effects for HOLC grade; SVI quintile; Census tract–level socioeconomic, housing, and transportation covariates; and mammography facility distance, along with intrinsic conditional autoregressive (ICAR) spatial random effects, first-order autoregressive (AR[1]) temporal random effects, and space-time interaction terms.29,30,31,32,33 Fixed-effect coefficients were exponentiated and reported as odds ratios (ORs) with 95% credible intervals (CrIs). For bayesian model results, statistical significance was defined as a 95% CrI excluding 1.00 for ORs. Full model equations, adjacency definitions, prior distributions, and computation details are provided in eMethods 1 in Supplement 1.34,35 Models were fit in R using the NIMBLE package (R Project for Statistical Computing), with 3 Markov Chain Monte Carlo of 10 000 iterations each, a 2000-iteration burn-in, and thinning of 2, yielding 12 000 posterior samples for inference.36,37 Posterior ICAR spatial random effects were exponentiated and mapped as tract-level spatial ORs. Tracts were classified as higher than expected, lower than expected, or not significant based on whether the 95% CrI for the tract-specific spatial OR was entirely above 1.0, entirely below 1.0, or crossed 1.0. To decompose the association between historical redlining and breast cancer screening prevalence into direct and indirect components, we conducted a tract-level multiple-mediator analysis restricted to 2024 within a structural equation modeling framework.38,39,40,41 The exposure contrasts HOLC grade D with grades A to C, and screening prevalence was modeled on the logit scale. Mediators included tract-level socioeconomic, housing, and transportation indicators (Table 1), and indirect associations were estimated as the product of the exposure-mediator and mediator-outcome paths, with uncertainty quantified using nonparametric bootstrap resampling.40,42,43,44,45 Corresponding 95% CIs were reported, and intervals excluding 0 were interpreted as evidence of an indirect association.

Two-sided P < .05 was considered statistically significant. Data analysis was performed from July to December 2025 using R, version 4.5.2 (R Project for Statistical Computing). Full model equations, mediator definitions, and implementation details are described in eMethods 2 in Supplement 1.