Historical Bias in Mortgage Lending, Redlining, and Implications for the Uncertain Geographic Context Problem: A Study of Structural Housing Discrimination in Dallas and Boston

Abstract

According to the uncertain geographic context problem, a lack of temporal information can hinder measures of bias in mortgage lending. This study extends previous methods to: (1) measure the persistence of racial bias in mortgage lending for Black Americans by adding temporal trends and credit scores, and (2) evaluate the continuity of bias in discriminatory areas from 1990 to 2020. These additions create an indicator of persistent structural housing discrimination. We studied the Boston-Cambridge-Newton and Dallas-Fort Worth metropolitan statistical areas to examine distinct historical trajectories and urban development. We estimated the odds of mortgage denial for census tracts. Overall, all tracts in Boston-Cambridge-Newton (N = 1003) and Dallas-Fort Worth (N = 1312) displayed significant change, with greater odds of bias over time in Dallas-Fort Worth and lower odds in Boston-Cambridge-Newton. Historically redlined areas displayed the strongest persistence of bias. Results suggest that temporal data can identify persistence and improve sensitivity in measuring neighborhood bias. Understanding the temporality of residential exposure can increase research rigor and inform policy to reduce the health effects of racial bias.

Introduction

In the United States (US), Black Americans have long been discriminated against during mortgage lending. Black Americans are more likely to be denied loans or receive higher interest rates than white populations, even when controlling for income and credit score [1]. Institutional racism is associated with poorer health outcomes and limited access to resources [2]. Historically, the Home Owners’ Loan Corporation (HOLC) demarcated areas (i.e., historical redlining), primarily based on racial composition, deeming them high-risk for loans and influencing neighborhood-level socioeconomic and health outcomes in US cities [3]. Homeownership and accumulation of generational wealth is linked to better health outcomes [4–6]. However, Black Americans are systematically denied homeownership or charged higher interest increasing financial strain, housing instability, and stress [1, 7].

Beyer et al. used Home Mortgage Disclosure Act (HMDA) data to derive a spatially continuous index measure of racial bias in mortgage lending (hereafter bias) and historical redlining [8]. Some previous studies incorporating this measure found strong associations between cancer mortality and contemporaneous bias and historical redlining [9, 10], while others did not [11, 12]. This inconsistency may reflect temporal variations and choice of the exposure period (i.e., how latency effects of racism can vary based on the life-course exposure period [13, 14]. This measure can lose sensitivity if temporality and change in economic barriers are not considered [15]. Mei Po Kwan first identified the ‘uncertain geographic context problem’ in 2012 [13]. This concept highlights an issue with area-based measures: the choice of geographic definitions can influence study outcomes if these definitions fail to accurately represent the area or timeframe where exposure impacts the study population. Furthermore, human activities’ spatial and temporal variability has significant implications for any research investigating how contextual factors affect health. To address these contextual uncertainties, scholars call for a dynamic conceptualization of place-based exposure [16]. The persistence of bias over time can be an important indicator of areas experiencing historic and continued discrimination in housing practices. In addition, the seasonal fluctuations the housing market is subject to create an important consideration for housing discrimination index calculations [17]. By incorporating both temporality and seasonality, we can better identify areas that have persistently experienced housing discrimination, thus improving measurement of the exposure.

Here, we extend Beyer et al.’s measure by considering temporality. Our study aims are to: (1) replicate Beyer et al.’s method for estimating bias while accounting for temporal trends and credit score differences; (2) analyze trends in mortgage lending bias over a three-decade period in two Metropolitan Statistical Areas (MSAs) in the US; and (3) evaluate continuity of mortgage lending bias in historically redlined areas over three decades, as an indicator of persistent structural housing discrimination. Data on residential exposure measures will clarify the impact of temporal variation and inform future measurement use and policies to reduce structural racism effects.

Methods

Study Area

To understand geographic elements of structural housing discrimination, we studied two areas: Boston-Cambridge-Newton MA-NH (BCN) and Dallas-Fort Worth TX (DFW); they provide contexts with distinct historical trajectories of urban development, racial and ethnic composition, and patterns of displacement and gentrification. We used boundary definitions from the 2010 Decennial Census of Census Metropolitan Statistical Areas and New England Central Township Areas to select our geographic coverage.

Data

We downloaded HMDA data for 1990–2020 from the National Archives [18] (1990–2006) and the Federal Financial Institutions Examination Council [19] (2007–2020). For HMDA data dated 2010–2020, we excluded data collected in 2017 from calculations due to changes in reporting requirements that were subsequently revoked in 2018. It is widely accepted that the 2017 HMDA data are not comparable to data from other years [20]. To compensate for these missing data, we instead included data from 2020. HMDA data includes race and ethnicity of the primary applicant, sex of the primary applicant, loan size, applicant income, action (e.g., bank denied or approved, applicant declined or accepted), type of loan (e.g., primary mortgage, pre-application, home improvement), reason for denial if applicable (e.g., credit score, loan-to-income ratio), and census tract FIPS code of the mortgage application address. Given the potential for measurement bias seen in existing racial discrimination in race-blind algorithmic underwriting systems, we excluded applications denied for a low credit score [17, 21–23]. We obtained boundaries for census tracts from the TIGER/LINE archives, using the web for 1990 decade, and from R’s tidycensus package API for 2000 and 2010 decades [24]. To identify historically redlined areas, we downloaded HOLC tract-level measures in 2010 Census boundaries from Noelke et al.’s database [25], which classifies tracts using the majority coverage area’s credit default risk in the 1930s (e.g., a tract with ≥ 50% with a grade of D = ‘Hazardous’ was classified as “Mainly D – Hazardous”; note other grades were A = ‛Best’; B = ‛Still Desirable’; C = ‛Definitely Declining’).

Processing

We followed Beyer et al.’s methods for creating the bias index using an adaptive spatial filter [8, 26–30] to count how many Black and white applicants in that tract were approved versus denied for a primary mortgage for any reason other than a low credit score. We created a grid of points, spaced 2.5 miles apart, across our two study areas. At each grid point, the adaptive spatial filter included the closest tract centroid to the grid point by Euclidean distance and counted how many Black and white applicants in that tract were approved versus denied for a primary mortgage for any reason other than a low credit score. If summed total for the four outcomes (i.e., approved Black applicant, approved white applicant, denied Black applicant, denied white applicant) was < 5, the filter included successively further away tract centroids until these cumulative totals were ≥ 5. We estimated a logistic regression model for the odds of mortgage denial, accounting for applicant sex, race, loan-to-income ratio, and application year. We interpolated the odds ratios (OR) of application denial for Black race from the logistic regression at each grid point to a continuous surface using inverse distance weighting. The continuous surface was summarized at the tract-level using the median value of raster cells and 2010 boundaries to create a continuous bias measure. Processing was completed in R. Maps were created using ArcGIS Pro 3.0.3.

Analysis

The continuous tract-level bias score was dichotomized using an OR cutoff of ≥ 2.5. We defined persistent bias as those tracts with OR ≥ 2.5 across all three-decade time series (i.e., 1990–1999, 2000–2009, 2010–2020). We calculated a measure of percent change in the odds of bias from the 1990–1999 estimate to the 2010–2020 estimate for each tract and averaged across all areas. We assessed each MSA as a whole and subsets of those tracts that had bias present in the 1990s and were identified as historically redlined (i.e., HOLC grade “hazardous”). We analyzed each geographic and subsample area to calculate median, interquartile range, and coefficient of variation. We used the differential local Moran’s Index (I) to test the spatial autocorrelation change over time. A differential Moran’s Index value indicates the spatial autocorrelation of changes between two time points, with positive values suggesting similar changes cluster together in space, negative values indicating dissimilar changes cluster together, and values close to zero suggesting changes are randomly distributed.

Results

Boston-Cambridge-Newton MA-NH

BCN has a history of bias, leading to disparities across localities (Fig. 1) 1owever, no tracts showed persistent bias over all three decades (Fig. 2). Table 1 summarizes changes in the odds of bias over time (1990–2020). Generally, all BCN tracts displayed decreased odds of bias over time (median =  − 0.68%). Additionally, when testing the percent change of bias among those tracts that had early bias in the 1990s, there was a decrease in bias (median =  − 17.02%). When we focus on historically redlined tracts (i.e., “hazardous”), the odds of Black applicants experiencing bias increased by 6.34% over time.

Fig. 1.

Bias in mortgage lending for Boston-Cambridge-Newton New England Central Township Area in (A) 1990–1999, (B) 2000–2009, and (C) 2010–2020*. Notes: * For HMDA data dated 2010-20, we excluded 2017 from calculations due to changes in reporting requirements that were revoked in 2018. It is widely accepted that the 2017 HMDA data are not comparable to data from other years. We instead included data from 2020. OR = odds ratio

Fig. 2.

Bias in mortgage lending for Boston-Cambridge-Newton New England Central Township Area in 1990–2020*, (A) Persistence in mortgage lending bias of Boston-Cambridge-Newton, (B) Overlap of Home Owner’s Loan Association historical redlining and mortgage lending bias in downtown Boston, MA. Notes: * For HMDA data dated 2010-20, we excluded 2017 from calculations due to changes in reporting requirements that were revoked in 2018. It is widely accepted that the 2017 HMDA data are not comparable to data from other years. We instead included data from 2020. OR = odds ratio

Table 1.

Percent change in the census tract-level odds of bias over time overall and subsampled by the presence of contemporaneous or historical bias

Percent change in the odds of bias over time (%)
Geographic areas and subsamples	N	Median (IQR)	Range	Coefficient of variation	Differential Moran’s I
Boston-Cambridge Newton
All census tracts	1003	− 0.68% (31.75)	− 66.57, + 253.76	8.68	0.62 (p < 0.001)
Tracts with bias present in the 1990s*	354	− 17.02% (22.15)	− 66.57, + 83.36	− 1.45	0.51 (p < 0.001)
Tracts historically redlined**	117	+ 6.34% (20.12)	− 35.98, + 183.62	2.85	0.75 (p < 0.001)
Dallas-Fort Worth
All census tracts	1312	+ 3.02% (46.00)	− 75.59, + 309.13	4.48	0.73 (p < 0.001)
Tracts with bias present in the 1990s*	390	− 25.90% (26.90)	− 75.59, + 123.31	− 1.06	0.58 (p < 0.001)
Tracts historically redlined**	46	+ 29.40% (99.80)	− 38.73, + 153.67	1.61	0.91 (p < 0.001)

^*Bias present in the earliest time period was defined as a tract having an OR ≥ 2.5 during 1990–1999

^**Historically redlined tracts were those that were D-Graded as “hazardous” by the HOLC in the 1930s

IQR, interquartile range

Spatial clustering tests revealed a moderate amount of neighborhood bias autocorrelation for all tracts over time, which is explained by the bias existing in prior decades (I = 0.62). Alternatively, while bias from previous decades had a lesser influence on predicting future neighborhood bias in tracts that were biased in the 1990s (I = 0.51), it was more predictive for neighborhoods that had historically been redlined (I = 0.75).

Dallas-Fort Worth TX

DFW also has a history of racial bias in mortgage lending, leading to disparities across various neighborhoods and cities (Fig. 3). We found 19 tracts that showed persistent bias across all three decades (Fig. 4). Generally, all DFW tracts displayed an increase in the odds of bias over time (median =  + 3.02%). However, there was a moderate decrease in bias (median =  − 25.90%) among those tracts with existing bias in the 1990s. Among those historically redlined tracts, the odds of Black applicants experiencing bias increased a median of 29.40% over time.

Fig. 3.

Bias in mortgage lending for Dallas-Fort Worth, TX Metropolitan Statistical Area in (A) 1990–1999, (B) 2000–2009, and (C) 2010–2020*. Notes: * For HMDA data dated 2010-20, we excluded 2017 from calculations due to changes in reporting requirements that were revoked in 2018. It is widely accepted that the 2017 HMDA data are not comparable to data from other years. We instead included data from 2020. OR = odds ratio

Fig. 4.

Bias in mortgage lending for Dallas-Fort Worth Metropolitan Statistical Area in 1990–2020*, (A) Persistence in mortgage lending bias of Dallas-Fort Worth, (B) Overlap of Home Owner’s Loan Association historical grades and mortgage lending bias in downtown Dallas, TX. Notes: * For HMDA data dated 2010-20, we excluded 2017 from calculations due to changes in reporting requirements that were revoked in 2018. It is widely accepted that the 2017 HMDA data are not comparable to data from other years. We instead included data from 2020. OR = odds ratio

Like BCN, there was a high amount of autocorrelation in neighborhood bias for all tracts over time (I = 0.73) with similar patterns of predicting future neighborhood bias (e.g., prior bias was a better predictor of future neighborhood bias for historically redlined tracts; I = 0.91).

Discussion

This study extended Beyer et al.’s [8] method by adding temporal trends and credit score differences to the measure of racial bias in mortgage lending for Black Americans and examined if bias changed over three decades in two distinct MSAs, BCN and DFW. Our results highlight spatiotemporal variation in racial bias in mortgage lending in both contexts. Interestingly, the study areas had converse change over time, with DFW showing increasing bias and BCN showing decreasing bias. In addition, both MSAs displayed increasing structural housing discrimination among those historically redlined areas. Some DFW tracts showed persistent structural housing discrimination from 1990 to 2020. The persistence of bias was seen strongly in redlined areas, which increased over time in both contexts.

The enduring implications of racism in neighborhood structures, particularly those redlined areas, necessitate that measurements consider the persistence of biased practices in historically oppressed areas [3]. Accounting for the residual impact of past bias can allow future research to capture a portion of the systemic bias engrained in an area.

This study has some limitations. We relied on HMDA data, which may not capture all relevant factors that contribute to structural housing discrimination [2]. Our study focused on only two MSAs in the US. Our results underscore the importance of considering urban processes and history when interpreting and generalizing results to other major cities. Another limitation of this study is the absence of applicant age data. In our analysis, we partially mitigated this limitation by incorporating related variables such as loan-to-income ratio, which is correlated with age. However, the lack of direct age data may affect the precision of our findings. Future studies could enrich our understanding by incorporating age data as it becomes more extensively available, allowing for a more nuanced analysis of its impacts. Lastly, historical boundaries used by the HOLC do not align with modern boundary definitions (i.e., census tracts). To robustly measure the exposure of modern neighborhoods to historical redlining, we used the database extracted by Noelke et al. This measure was evaluated across 54 classifications to select the classification with high predictive validity, and it was tested in cross-validation and cross-classification with health and socio-economic outcomes [25].

Policy development and future research studying relationships between systemic racism in housing and health outcomes can benefit from this new measure, which incorporates historical spatial and temporal patterns of racial bias in mortgage lending. Equity in home ownership is associated with health disparities and should be considered in both spatial and temporal contexts [13].

Conclusion

Findings advance understanding of discrimination against Black Americans in mortgage lending and provide a new measure to study health and socioeconomic outcomes. Using this temporally informed measure may resolve the inconsistent patterns observed between bias in mortgage lending and health outcomes and uncover how systemic denial of homeownership for Black Americans can lead to persistent negative health effects through neighborhood oppression.

Funding

This work was supported by funding from the National Cancer Institute (UM1CA221940-05S1, U01CA253912-02S2) and the American Cancer Society (CRP-22-080-01-CTPS). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data availability

All data generated or analyzed during this study are publicly available.