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

Results

The analytic sample included 1152 historically redlined Census tracts categorized by HOLC grades A (n = 27), B (n = 99), C (n = 423), and D (n = 603), corresponding to 2.3%, 8.6%, 36.7%, and 52.3% of tracts, respectively. Descriptive statistics for breast cancer screening prevalence and Census tract–level socioeconomic characteristics by redlining grade are presented in Table 1.

Overall, tests indicated that screening prevalence and neighborhood characteristics varied across HOLC grades (all with P < .001) (Table 1). Descriptively, median (IQR) breast cancer screening prevalence was highest among grade A tracts (79.6% [79.0%-80.0%]) and lowest among grade D tracts (75.2% [71.7%-79.2%]). Social vulnerability showed the opposite pattern: all 27 Grade A tracts (100%) were in the lowest SVI quintile, whereas 324 Grade D tracts (53.4%) were in the highest quintile. Table 1 also shows descriptive gradients across HOLC grades in poverty, unemployment, lack of health insurance, crowded housing, and lack of transportation.

To assess the association between historical redlining and breast cancer screening prevalence after accounting for SVI and other covariates, we fitted a bayesian spatiotemporal β regression model with ICAR spatial random effects, AR(1) temporal random effects, and a space-time interaction. Model specifications were compared using the deviance information criterion (DIC), and the model with the lowest DIC was retained for inference. Using this selected model, we observed substantial residual geographic heterogeneity, captured by the spatial random effect us: 711 of 1152 tracts (61.7%) had 95% CrIs for us excluding 0, indicating meaningful spatial clustering in screening prevalence beyond the measured covariates. Temporal dependence was also supported through the AR(1) temporal random effect αt, with the AR(1) coefficient ϕ having a 95% CrI of 0.623 to 0.995, consistent with persistent year-to-year correlation, and the space-time interaction γst was significant for 305 of 1152 tract-year combinations (26.5%), indicating localized departures from the overall temporal pattern. The results are summarized with posterior means, posterior SDs, ORs, and 95% CrIs in Table 2. Because the β regression used a logit link for the mean screening prevalence, ORs were reported as the natural exponentiated form of the fixed-effect coefficients.

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

In the unadjusted bayesian spatiotemporal β regression model including HOLC grade only, grade B (OR, 0.98; 95% CrI, 0.87-1.11), grade C (OR, 0.97; 95% CrI, 0.85-1.11), and grade D (OR, 0.96; 95% CrI, 0.82-1.09) tracts had lower but nonsignificant odds of breast cancer screening compared with grade A tracts. In the adjusted bayesian spatiotemporal model, historical redlining was associated with breast cancer screening. Compared with HOLC grade A, lower odds of screening were observed for grades C (OR, 0.94; 95% CrI, 0.90-0.99) and grade D (OR, 0.94; 95% CrI, 0.89-0.99) tracts after adjustment for SVI; socioeconomic, housing, and transportation covariates; mammography facility distance; and spatial and temporal structures.

Because the primary objective was to estimate the association between HOLC grade and breast cancer screening after adjustment for contemporary neighborhood conditions, conditional associations for SVI and other covariates are presented in Table 2 but were not interpreted as primary findings. The highest SVI quintile was associated with increased screening odds (OR, 1.08; 95% CrI, 1.01-1.14). Lack of high school diploma (OR, 0.94; 95% CrI, 0.92-0.96) and higher mobile home prevalence (OR, 0.98; 95% CrI, 0.97-1.00) were associated with lower screening odds. Although some higher-screening areas appear to be located nearer mammography centers, distance to the nearest facility was not independently associated with screening prevalence after adjustment for redlining, SVI, neighborhood socioeconomic characteristics, and spatial structure. The model also identified meaningful temporal variation in screening prevalence across the study period: temporal random effects indicated screening increased from 2016 to 2017, peaked in 2018 to 2019, remained stable through 2020 to 2023, and declined in 2024 (eFigure 2 in Supplement 1), suggesting time-varying screening levels beyond redlining, SVI, and other covariates.

Beyond temporal patterns, the model revealed substantial residual geographic heterogeneity across the catchment area (Figure 1). The continuous posterior ICAR spatial ORs are shown in Figure 1A, and the corresponding classifications of lower-than-expected, higher-than-expected, and not significant screening areas are shown in Figure 1B. Lower-than-expected screening clusters were observed in areas such as Independence, Blue Valley, North Kansas City, and Armourdale. In contrast, higher-than-expected screening clusters appeared in Midtown/Hyde Park, Waldo/Armour Hills, and Prairie Village. Some lower-than-expected screening clusters were located near mammography facilities, suggesting that proximity alone may not explain these patterns.

Figure. Maps of Model-Adjusted Spatial Patterns in Breast Cancer Screening Prevalence, With Mammography Centers OverlaidThese values reflect model-adjusted residual spatial patterns, not crude screening prevalence. A, Posterior intrinsic conditional autoregressive (ICAR) spatial random effects are expressed as odds ratios (ORs). Values greater than 1.0 indicate higher-than-expected screening after model adjustment, whereas values less than 1.0 indicate lower-than-expected screening after model adjustment. Lighter or yellow areas represent higher residual screening, and darker or purple areas represent lower residual screening. B, Model-adjusted screening areas are based on posterior ICAR spatial random effects. Lower screening indicates Census tracts with significantly lower-than-expected screening after model adjustment, defined by a 95% credible interval (CrI) for the tract-specific spatial OR entirely below 1.0. Higher screening indicates Census tracts with significantly higher-than-expected screening after model adjustment, defined by a 95% CrI entirely above 1.0. Not significant indicates tracts with CrIs crossing 1.0. Screening status was consolidated over the 2016 to 2024 study period. (See the original article for the figure.)

To characterize spatiotemporal patterns in breast cancer screening and inform planning, Figure 2 and Figure 3 present year-specific model-estimated screening prevalences for 2016 to 2024, along with short-term projections for 2025 and 2026. Screening prevalence ranged from 60.2% (693 of 1152) to 83.7% (964 of 1152). Model-estimated screening prevalence closely agreed with observed tract-level screening prevalence (Pearson r = 0.999), indicating close agreement between observed and model-estimated values. Estimated screening prevalence increased from 2017 to 2018, remained relatively stable through 2023, and declined in 2024. Despite these overall shifts, spatial disparities persisted: the same areas repeatedly showed lower estimated screening prevalence across years after adjustment for redlining, SVI, and other tract-level covariates. Model-based projections for 2025 and 2026 indicated that these lower-screening prevalence patterns may continue, including parts of Wyandotte County and eastern Jackson County, supporting geographically targeted outreach and resource allocation. Nationally, the Healthy People 2030 breast cancer screening target is 80.3%. Using 80.3% as a benchmark for progress toward the 2030 goal, each tract-year was classified as at or above the target (≥80.3%) vs below the target (<80.3%) (eFigure 3 in Supplement 1). Across 2016 through 2024, most tract-years in the study area (972 of 1152 [84.4%]) were below this benchmark, indicating that many neighborhoods remain behind the national 2030 screening target. These findings suggested that broad catchment-area efforts may be needed to improve screening overall, while spatial patterns may help identify neighborhoods where additional outreach, navigation, and structural supports should be prioritized.

Figure. Maps of Year-Specific Model-Estimated Breast Cancer Screening Prevalence by Census Tract From 2016 to 2021 (See the original article for the figure.)

Figure. Maps of Year-Specific Model-Estimated Breast Cancer Screening Prevalence by Census Tract From 2022 to 2026 Projected values are shown for 2025 and 2026. (See the original article for the figure.)

To identify pathways linking historical redlining to screening prevalence in 2024, we conducted a tract-level multiple-mediator structural equation modeling comparing HOLC grade D vs grades A to C (eTable and eFigures 4 and 5 in Supplement 1). SVI was not included because it was a composite neighborhood vulnerability measure that overlapped with several individual socioeconomic, housing-, and transportation-related mediators considered separately in the analysis. The total association was explained primarily through socioeconomic mediators, with the largest contributions from poverty (44.5%; 95% CI, 22.7%-76.1%) and lower educational attainment (25.5%; 95% CI, 6.7%-54.0%), while lack of health insurance contributed minimally (1.4%; 95% CI, −11.0% to 15.1%). These findings suggest that socioeconomic conditions, particularly poverty and educational barriers, may be important intervention targets in historically redlined neighborhoods.