Inheriting the Grade: HOLC “Redlining” Maps and Contemporary Neighborhood Crime

The Origins and Legacy of the HOLC Maps

Faced with the threat of the Great Depression, various New Deal programs attempted to protect and promote an American Dream of home ownership, albeit within tightly bounded racial and class lines (Rothstein 2017). The primary goal of the HOLC was to reduce foreclosures by providing affordable and standardized mortgages (Hillier 2003; Jackson 1980, 1987) that would stabilize housing values (Aaronson, Hartley, and Mazumder 2021). Between 1933 and 1936, the HOLC granted 1 million self-amortizing mortgages to homeowners who were in default or who had already lost their homes (Hillier 2003). In 1935, when the HOLC lending program was concluding, HOLC’s parent organization, the Federal Home Loan Bank Board, directed HOLC employees, local lenders, and real estate experts to rank areas within 239 cities to estimate real estate risk (Hillier 2003). Although the quality of the housing stock, amenities, and proximity to industry shaped the neighborhood grades, racial and ethnic composition featured prominently in the HOLC mapping rationale. In some cases, the presence of even one Black family earned neighborhoods the highest risk category (Winling and Michney 2021). In other areas, HOLC appraisers singled out Jewish and Mexican families as well (Mendez-Carbajo 2021). HOLC appraisal manuals and the notes taken on each neighborhood often used explicitly bigoted language to express a fundamental assertion underlying segregation and institutionalized racism in the housing market: that proximity of non-White, and especially Black, households threatened the value of White housing investment (Rothstein 2017).

Social scientists and historians provide conflicting accounts of the maps’ independent impact on actual lending practices, partly because of the overreliance on case studies with one or a small number of cities (Fishback et al. 2022; Hillier 2003; Jackson 1980; Michney and Winling 2020). Although interpretations of the historical record vary, we can categorize them into two broad perspectives. The first articulates an institutionalization hypothesis, which contends that the HOLC’s methodology for grading areas facilitated the adoption of similar appraisals in other private and public lending agencies (Aaronson, Hartley, and Mazumder 2021; Crossney and Bartelt 2005; Greer 2013; Woods 2012). Jackson (1980, 1987), credited as the first to associate HOLC grades with investment based on patterns in two areas (St. Louis city/county and Newark city/Essex county), concluded that the “damage caused by the HOLC came not through its own actions, but through the influence of its appraisal system on the financial decisions of other institutions” (Jackson 1980:430). Among these entities, scholars most often highlight the FHA and the GI Housing Bill, which together accounted for a much greater share of residential loans than the HOLC in the United States at the time. The FHA also constructed its own security maps that clearly guided pro-White/anti-Black lending decisions, and some scholars suggest the work of the HOLC influenced the FHA maps directly (Xu 2022). ¹ Faber (2020) notes that the legacy of the HOLC maps would have been more limited had it not been for the FHA and GI Housing Bill instituting the practice of restricting lending to higher graded areas of cities. Widespread communication between housing agencies, realtors, and lenders federalized the logic of redlining represented by the HOLC maps. Massey and Denton (1993:52) claim the HOLC methodology spread beyond the public sphere to influence private lending entities as well: “[T]he HOLC not only channeled federal funds away from black neighborhoods but was also responsible for a much larger and more significant disinvestment in black areas by private institutions.”

The institutionalization perspective implies that the HOLC maps were an agent of segregation and that their methodology played an important role in setting areas of cities on varied trajectories of investment and development. Most research typically adopts this perspective, either explicitly or implicitly, when explaining how the HOLC maps might correlate with contemporary outcomes (Krieger et al. 2020; Mehranbod et al. 2022; Meier and Mitchell 2022).

An alternative perspective takes a more critical view of the maps’ legacy and begins with claims that the FHA was developing its own lending cartography before the HOLC published its security maps (Fishback et al. 2022). Likewise, Hillier (2003), using a single city (Philadelphia) as a case study, documents that the practice among lenders of avoiding certain areas of cities preceded the creation of the HOLC maps. Furthermore, some scholars now concede that the HOLC (1) did not make their maps widely available to the public (Hillier 2003), (2) made most of their loans before the publication of the maps, and (3) lent to more Black families than previously assumed, although always in ways that reinforced racial segregation (Michney and Winling 2020). Together, these arguments suggest that the HOLC maps and/or methodology had a less direct influence on the lending practices of other public or private agencies than commonly assumed. Based on their own assessment of the maps’ legacy in three cities, Fishback et al. (2022:1) conclude the maps had “little effect on the geographic distribution” of lending from either the HOLC or FHA. Instead, the maps reflected racial and class biases already in play in the 1930s.

Importantly, despite their differences, these two perspectives agree that the HOLC maps provide the most comprehensive and available documentation of the institutional biases underlying the housing market in the late 1930s. Thus, both perspectives imply the HOLC maps may correlate with contemporary patterns of inequality, although through somewhat different historical mechanisms. The institutionalization perspective implies an association between HOLC maps and the inequitable structural development of urban neighborhoods. By influencing the priorities of public and private lending agencies, HOLC maps set areas of cities on unequal paths of investment that likely contributed to subsequent disparities in urban outcomes (Aaronson, Faber, et al. 2021; Aaronson, Hartley, and Mazumder 2021). In contrast, the second perspective implies that any association between the HOLC maps and contemporary outcomes is spurious and contends that while the structural antecedents of the maps may have shaped differential investments across urban space over time, the maps themselves had little direct influence.

Conceptual Model: Linking HOLC Grades to Neighborhood Crime

A handful of recent single-city studies are among the first to find positive associations between HOLC grades and violence in Philadelphia (Jacoby et al. 2018), Louisville (Benns et al. 2020), Boston (Poulson et al. 2021, 2023), Los Angeles (Anders 2023), and Chicago (Powell and Porter 2022) at the census tract level. In one of two studies to explore pathways between HOLC grades and crime, Poulson et al. (2021) clarify that redlined areas have higher contemporary levels of disadvantage, housing insecurity, and residential segregation, which in turn yield higher concentrations of firearm violence in Boston communities. Powell and Porter (2022) find that contemporary levels of concentrated disadvantage mediate the association between HOLC grades and present-day homicide rates in Chicago neighborhoods. Mehranbod et al. (2022) move beyond these single-city analyses and explore relationships between HOLC grades and violence across zip codes in multiple cities. They find that redlined zip codes had higher incidences of firearm deaths than greenlined places. Importantly, these associations varied by city. This latter finding, coupled with the fact that most HOLC studies base their conclusions on one or a small number of cities, underscores the need for multicity analyses.

Building on this recent work, historical accounts of the HOLC and other federal housing programs, research on residential segregation and urban development, and the communities and crime literature, we present a conceptual model linking the HOLC grades to contemporary neighborhood crime (Figure 1). A premise of the framework is that the HOLC gradients both reflected and fueled the relative ranking of areas in the political economy of place (Logan and Molotch 1987). Accordingly, political and economic actors reward historically upgraded neighborhoods with amenities and resources because these areas facilitate the city’s agenda of growth and capital accumulation. In contrast, the agenda neglects and/or further handicaps downgraded neighborhoods, depleting them of the ability to demand resources within the growth machine (Logan and Molotch 1987). Our model assumes that the HOLC map grades (Figure 1, Panel 2) thus helped crystalize the logic that set neighborhoods on different courses of segregation and (dis)investment, contributing to divergent contemporary social and economic fortunes. These contemporary neighborhood conditions—namely, ethnoracial segregation, housing investment and stability, and the concentration of socioeconomic disadvantages (Figure 1, Panel 3)—help to explain variation in neighborhood violent and property crime (Figure 1, Panel 4). Although the model generally is consistent with the institutionalization perspective, we also acknowledge the socioeconomic and racial structural antecedents of the maps (Figure 1, Panel 1), which allows a partial test of the alternative spuriousness perspective. We elaborate on key aspects of the model in the following.

OPEN IN VIEWER

Figure 1.

Conceptual framework linking Home Owner’s Loan Corporation (HOLC) grades to neighborhood crime.

Data and Sample

We compile information from three data sets: (1) the second wave of the National Neighborhood Crime Study (NNCS2; Krivo, Lyons, and Vélez 2022), (2) HOLC maps normalized to census tracts by Meier and Mitchell (2021), and (3) 1940 census data provided by the National Historical Geographic Information System (Manson et al. 2022). ² The NNCS2 includes data on serious crimes known to the police and structural characteristics for census tracts ³ across 85 large cities circa 2010. To our knowledge, it is the only publicly available data set with crime and other data for neighborhoods in a representative set of medium to large U.S. cities. The breadth of coverage, including central cities and suburbs across all regions, makes the NNCS2 uniquely suited for assessing neighborhood inequality in crime.

We began by merging the NNCS2 with the HOLC grades provided by Meier and Mitchell (2021). Differences in missing data across the crime measures result in samples of 52 cities and 4,679 mapped tracts for violence and 53 cities with 4,753 tracts for burglary. We then merge these data with the 1940 census by overlaying shapefiles for 2010 and 1940 census tracts with ArcGIS Pro. ⁴ Because the 1940 data did not cover all cities, our final analytic samples include 3,795 census tracts across 27 cities for violence and 3,867 census tracts across 28 cities for burglary. For a list of the cities and the associated number of census tracts included in the following main analyses, see Appendix 1.

HOLC Grades

The Mapping Inequality Project (Nelson et al. 2021) has digitized most HOLC maps for the nation. Because the HOLC grades do not correspond neatly to census tract boundaries, we use aerial interpolation data from Meier and Mitchell (2021), who normalized the maps from the Mapping Inequality Project to 2010 census boundaries for numerous cities. Meier and Mitchell (2021) first constructed a continuous HOLC score by assigning a numeric value corresponding to each HOLC risk category (A = 1, B = 2, C = 3, and D = 4). They then multiplied these values by the proportion of the census tract area covered by each category and summing (Richardson et al. 2020:16). As they explain, “[I]f 50% of the census tract’s graded area was classified ‘A’ and 50% was ‘B,’ the calculation would be ((1 × 0.50) + (0.50 × 2)) = 1.50” (Richardson et al. 2020:16). Ranging from 1.00 to 4.00, higher scores indicate more disadvantaged HOLC grading. Meier and Mitchell (2021) then rounded up the continuous score to the nearest integer to create four categorical variables that correspond to each grade; we use these intervals in our analyses. Meier and Mitchell (2021) provide HOLC scores only for 2010 tracts that had at least 20 percent of their area covered by the historic HOLC maps, reasoning that tracts not meeting this threshold were “not sufficiently impacted” by the maps (Richardson et al. 2020:16). Most of the 2010 census tracts excluded from Meier and Mitchell’s (2021) scoring did not exist in the 1930s, when the HOLC constructed the maps, or were industrial areas at the time with little residential population. We thus limit our main analyses to tracts that had at least 20 percent of their 2010 area mapped by the HOLC. However, we also report results from supplementary models that include tracts that were not mapped (or were “insufficiently” mapped) to assess the robustness of our findings.

Table 1 presents the distribution of the intervals for the HOLC grades. The HOLC redlined 35 percent of tracts, whereas 43 percent received yellow grades. About 5 percent of tracts were greenlined for investment, with 17 percent receiving the second most advantaged (blue) grade. ⁶

OPEN IN VIEWER

Table 1.

Summary Statistics for Tract and city-level Variables.

	Obs	Mean	Median	Std. Dev.	Variance	Min	Max
Tract-level Variables
Outcome (2010–2013)
Violence	3,795	61.25	45.00	55.09	3034.97	0	513
Burglary	3,867	149.11	120.00	114.34	13072.69	0	1200
HOLC Grades
Green 'A'	3,867	.05		.23		0	1
Blue 'B'	3,867	.17		.38		0	1
Yellow 'C'	3,867	.43		.49		0	1
Red 'D'	3,867	.35		.48		0	1
1940 Census
Disadvantage index	3,867	–.02	–.11	.71		–1.97	3.62
Percent Black	3,867	5.71	.22	16.60		0	99.81
Median Housing values	3,867	3606.39	3341.00	2231.93		0	19803
Neighborhood Segregation
White	3,867	.21		.41		0	1
Black	3,867	.26		.44		0	1
Latino	3,867	.10		.30		0	1
Minority	3,867	.06		.24		0	1
White-Black Integrated	3,867	.08		.27		0	1
White-Latino Integrated	3,867	.08		.27		0	1
Other Integrated	3,867	.21		.40		0	1
Socioeconomic Structural Conditions
Home mortgage dollars per housing unit, logged	3,867	1.91	1.95	1.28		0	5.82
Vacancies	3,867	14.63	11.87	10.24		0	59.74
Foreclosure rate	3,867	43.51	30.94	52.90		0	1916.67
Homeowner rate	3,867	43.02	42.63	21.06		0	100
Disadvantage index	3,867	.23	.36	.92		–1.83	3.15
Additional tract-level controls
Population in 1000s, logged	3,867	1.11	1.16	.48		–1.12	2.76
Foreign born	3,867	17.40	11.07	17.08		0	78.26
Young males	3,867	7.67	7.06	4.60		0	53.44
City-level Variables
Population	28	785677.50	721459.00	830553.50		124879	3804503
Percent Black	28	28.97	32.48	17.77		6.25	81.84
Percent Latino	28	18.61	17.40	15.23		2.19	48.37
Disadvantage index	28	.40	.22	.93		–1.47	2.46
Residential mobility	28	19.73	16.33	3.42		14.13	27.14
Vacancies	28	13.31	13.73	5.41		6.52	28.69
White-Black segregation	28	60.35	66.91	9.87		38.64	82.48
South	28	.21		.42		0	1
West	28	.25		.44		0	1

Note: Summary statistics are based on the analytic sample for burglary. HOLC = Home Owner’s Loan Corporation.

Analytic Strategy

We estimate two-level random intercept models for tracts (Level 1) nested within cities (Level 2). To assess the appropriateness of a multilevel modeling strategy, we first estimated unconditional random intercept models that indicated substantial variation in both violence and burglary across cities (i.e., random intercept variance). Second, we calculated log-likelihood ratio tests comparing the random intercept model to a single-level negative binomial model. These tests indicate that the multilevel model outperforms the single-level model for both outcomes.

Because crimes are relatively rare events within census tracts, we estimate nonlinear negative binomial count models within the multilevel framework. These models account for overdispersion in which the variance of the dependent variable is considerably larger than the mean or median, as is the case with our count data (see Table 1). We specify that crime counts have variable exposure by tract population, which allows us to interpret the analyses as crime rates per tract population. As mentioned, we follow previous work using NNCS2 data and control for logged tract population as a substantive predictor on the right side of the equation as well, given that neighborhood population consistently associates with per capita violence and property crime (Lyons et al. 2022).

Our data do not meet strict requirements for formal mediation analyses over multiple decades. Instead, we assess the statistical significance of changes across models in the coefficients for the HOLC scores using a general framework for comparing marginal effects for nonlinear models developed by Mize, Doan, and Long (2019). The approach transforms parameter estimates into predicted margins and then uses seemingly unrelated estimations to combine the variance and covariance estimates across models with varying specifications. As Mize et al. (2019:152) explain, “marginal effects are a useful method for quantifying effects because they are in the natural metric of the dependent variable and they avoid identification problems when comparing regression coefficients” across models. Their framework thus provides an optimal assessment of the significance of changes in coefficients across nonlinear models in particular (Mize et al. 2019:156). Although not a formal assessment of mediation, the approach is equivalent to what mediation scholars refer to as a “differences in coefficient test” (MacKinnon et al. 2002:87–89), which is one strategy for assessing mediation informally (Mize et al. 2019:186). We estimated these tests using mecompare within Stata’s GSEM, following the example and code written by Mize et al. (2019). ⁹

We start with baseline models that control for only the HOLC grades. We then assess whether the coefficients for the HOLC grades change significantly once we control for 1940 conditions and contemporary tract- and city-level demographics. Subsequent models add in variables measuring racial residential segregation (ethnoracial composition), housing characteristics, and concentrated disadvantage separately to assess their contributions in explaining the relationship between HOLC grades and crime. The final model includes all covariates.

Results

Table 2 provides estimates for multilevel models predicting counts of neighborhood violence. The baseline Model 1 indicates a dose-response relationship between HOLC grades and violence. Neighborhoods graded B (blue), C (yellow), and D (red) have higher levels of violence relative to areas that were graded A (greenlined), and all other comparisons between grades are statistically significant. ¹⁰ However, the predicted gap between redlined and greenlined areas is the most striking, at about 57 violent crimes. Model 2 adjusts for 1940 conditions and other contemporary controls at the tract and city levels and provides little evidence that the relationship between HOLC grades and violence is spurious. Although adjusting for 1940 conditions (disadvantage, percentage Black population, and housing values) reduces the coefficients for the HOLC scores, tracts appraised blue, yellow, and red still have substantially higher levels of violence today than greenlined tracts. The dose-response relationship between HOLC grades and violence, however, breaks down net of controls. Specifically, the difference between redlined and yellowlined places is no longer significant once accounting for 1940 conditions. Of the 1940 historical controls, the index measuring concentrated disadvantage and median housing value both are positively associated with violence circa 2010. Additional tests reveal that the positive coefficient for median housing value is contingent on controlling for 1940 disadvantage; without accounting for disadvantage, the coefficient for median housing value is negative and significant. ¹¹

OPEN IN VIEWER

Table 2.

Multi-level Negative Binomial Models for Violence with Tract Population as Exposure.

	Model 1: Baseline			Model 2: Controls			Model 3: Segregation			Model 4: Housing			Model 5: Disadvantage			Model 6: Full
	B		SE	B		SE	B		SE	B		SE	B		SE	B		SE
HOLC Interval (ref = Green)
Blue	.5243	***	.0635	.5076	***	.0623	.3645	***	.0557	.2104	***	.0554	.2994	***	.0560	.2415	***	.0527
Yellow	.9093	***	.0596	.7862	***	.0636	.4889	***	.0574	.2521	***	.0577	.3620	***	.0579	.2477	***	.0549
Red	1.1120	***	.0604	.7555	***	.0732	.4654	***	.0660	.2451	***	.0659	.3435	***	.0667	.2499	***	.0627
1940 Disadvantage index				.2781	***	.0329	.3437	***	.0288	.1981	***	.0284	.2923	***	.0282	.2158	***	.0273
1940 Percent Black				.0246		.0874	–.4421	***	.0767	–.0622		.0765	–.1868	*	.0773	–.3372	***	.0731
1940 Housing value1				.0025	**	.0008	.0022	**	.0007	.0024	***	.0007	.0044	***	.0007	.0022	***	.0006
Neighborhood Segregation (ref= White)
Black							1.2370	***	.0376							.6817	***	.0492
Latino							.6507	***	.0550							.3287	***	.0604
Minority							1.0801	***	.0544							.6812	***	.0593
White-Black Multiethnic							.7279	***	.0467							.4412	***	.0455
White-Latino Multiethnic							.3359	***	.0490							.1506	**	.0482
Other Multiethnic							.6140	***	.0385							.3798	***	.0388
Home mortgage dollars, logged										–.3097	***	.0133				–.0404		.0211
Vacancies										.0116	***	.0016				.0070	***	.0015
Foreclosure rate										.0025	***	.0004				.0017	***	.0003
Homeowner rate										–.0083	***	.0006				–.0092	***	.0006
Disadvantage Index													.4848	***	.0144	.1908	***	.0258
Tract Controls
Logged population				–.2483	***	.0301	–.2103	***	.0269	–.1338	***	.0269	–.2721	***	.0269	–.1463	***	.0259
Foreign born				–.0100	***	.0010	–.0036	**	.0012	–.0107	***	.0009	–.0098	***	.0009	–.0079	***	.0011
Young males				.0229	***	.0033	.0114	***	.0027	.0003		.0027	.0017		.0027	–.0041		.0026
City Controls
Population2				–.0002		.0001	–.0002		.0001	–.0002		.0001	–.0002		.0001	–.0002		.0001
Percent Black				–.0043		.0092	–.0106		.0088	.0054		.0084	.0016		.0087	–.0020		.0083
Percent Latino				.0049		.0065	–.0026		.0062	.0099		.0059	.0095		.0061	.0025		.0059
Disadvantage index				.0316		.1434	.0509		.1373	–.1947		.1299	–.1964		.1360	–.1277		.1298
Residential mobility				–.0670	**	.0237	–.0416		.0227	–.0434	*	.0214	–.0461	*	.0224	–.0394		.0213
Vacancies				.0458		.0277	.0413		.0266	.0026		.0251	.0376		.0263	.0260		.0251
White-Black segregation				.0263	*	.0107	.0202		.0103	.0225	*	.0097	.0216	*	.0102	.0188		.0097
South				.0003		.1933	.0640		.1851	–.0814		.1748	–.0518		.1830	.0354		.1744
West				.1385		.2478	.1216		.2371	.2217		.2241	.1408		.2346	.0518		.2237
Constant	–4.9251	***	.1133	–5.3617	***	.6457	–5.5900	***	.6164	4.2196	***	.5849	5.2078	***	.6102	4.7974	***	.5830
Random intercept variance	.2613		.0738	.0782		.0239	.0728		.0218	.0644		.0194	.0710		.0240	.0646		.0193
Log Likelihood	–18775.889			–18597.377			–18067.884			–18001.145			–18115.054			–17788.416
AIC	37563.78			37236.75			36189.77			36052.29			36274.11			35640.83
BIC	37601.23			37367.82			36358.29			36208.33			36411.42			35840.56

Notes: N = 3,795 tracts within 27 cities. 1 = estimates multiplied by 100; 2 = estimates multiplied by 1000. *** p <.001; ** p <. 01; * p < .05

Models 3, 4, and 5 assess separately the contribution of tract segregation, housing conditions, and concentrated disadvantage compared to Model 2. As expected, all measures in these separate groups predict violent crime. Furthermore, they also help explain why poorer graded tracts have more violence today. Compared to Model 2, controlling for ethnoracial composition in Model 3 reduces the gaps between green and blue or yellow areas by about 28 percent [(.5076 –.3654)/.5076] and 38 percent [(.7862–.4889)/.7862)], respectively, and the gap between greenlined and redlined areas by about 38 percent [(.7555–.4654)/.7555)]. As a block, contemporary housing conditions, including home mortgage dollars, vacancies, foreclosure rates, and the homeownership rate play an even more pronounced explanatory role in Model 4. Accounting for housing conditions reduces the gap between green and blue, green and yellow, and green and red tracts by 59 percent, 68 percent, and 68 percent, respectively, compared to Model 2. Adjusting for the concentrated disadvantage index yields similar size reductions in coefficients for HOLC grades in Model 5. For example, concentrated disadvantage alone accounts for a 55 percent reduction in the gap between greenlined and redlined tracts compared to Model 2. ¹²

Model 6 includes all covariates. Comparing this full model with the baseline (Model 1) indicates strong support for our general conceptual framework. We are able to explain about 67 percent of the difference in violence between the highest (green) and lowest (red) graded areas, 68 percent of the gap between green and yellowlined tracts, and 52 percent of the disparity in violence between the two highest graded tracts (green and blue). However, the criminogenic influence of the HOLC grades persists even net of the fuller set of controls, indicating that we are not able to specify completely the inheritance of the HOLC maps.

Do these patterns extend to more common forms of property crime? Table 3 provides results for neighborhood burglary. The baseline Model 1 shows a similar dose-response association between HOLC grades and burglary, although the disparities between grades are somewhat smaller for burglary than they are for violence. Tracts with historical B, C, and D scores have higher levels of burglary relative to areas with A scores, and all other comparisons between grades are statistically significant. The unadjusted predicted gap between redlined and greenlined areas is the most notable, with a difference of about 50 burglaries. After accounting for historical and contemporary controls in Model 2, the general dose-response relationship between HOLC grades and burglary persists, suggesting no support for the spuriousness perspective (in fact, the coefficients for the grades increase net of these controls compared to Model 1).

OPEN IN VIEWER

Table 3.

Multi-level Negative Binomial Models for Burglary with Tract Population as Exposure.

	Model 1: Baseline			Model 2: Controls			Model 3: Segregation			Model 4: Housing			Model 5: Disadvantage			Model 6: Full
	B		SE	B		SE	B		SE	B		SE	B		SE	B		SE
HOLC Interval (ref = Green)
Blue	.0995	*	.0401	.1775	***	.0377	.0690	*	.0348	.0557		.0351	.0629		.0347	.0532		.0335
Yellow	.1744	***	.0375	.3114	***	.0386	.1309	***	.0361	.0927	*	.0367	.0874	*	.0361	.0703	*	.0351
Red	.2477	***	.0381	.3586	***	.0448	.1704	***	.0419	.1274	**	.0423	.1290	**	.0418	.1041	*	.0405
1940 Disadvantage index				.0307		.0205	.0558	**	.0188	.0450	*	.0189	.0322		.0184	.0509	**	.0183
1940 Percent Black				–.1158	*	.0546	–.3420	***	.0508	–.1744	***	.0501	–.2476	***	.0499	–.2892	***	.0491
1940 Housing value1				.0012	*	.0005	.0011	*	.0004	.0016	***	.0004	.0022	***	.0004	.0017	***	.0004
Neighborhood Segregation (ref= White)
Black							.6178	***	.0239							.2708	***	.0320
Latino							.2869	***	.0354							.0090		.0391
Minority							.5726	***	.0351							.2731	***	.0390
White-Black Multiethnic							.3857	***	.0295							.2321	***	.0296
White-Latino Multiethnic							.3353	***	.0314							.2031	***	.0313
Other Multiethnic							.3885	***	.0241							.2431	***	.0247
Home mortgage dollars, logged										–.1368	***	.0089				.0008		.0138
Vacancies										.0125	***	.0010				.0094	***	.0010
Foreclosure rate										.0014	***	.0002				.0009	***	.0002
Homeowner rate										.0022	***	.0004				.0022	***	.0004
Disadvantage Index													.2618	***	.0094	.1571	***	.0172
Tract Controls
Logged population2				–.1126	***	.0185	–.0994	***	.0172	–.0685	***	.0175	–.1265	***	.0170	–.0830	***	.0169
Foreign born				–.0162	***	.0006	–.0143	***	.0008	–.0149	***	.0006	–.0169	***	.0006	–.0141	***	.0007
Young males				.0046	*	.0018	.0007		.0016	–.0018		.0016	–.0054	**	.0016	–.0035	*	.0016
City Controls
Population				–.0001		.0001	–.0001		.0001	–.0001		.0001	–.0001		.0001	–.0001		.0001
Percent Black				–.0108		.0088	–.0141		.0084	–.0079		.0074	–.0084		.0079	–.0109		.0075
Percent Latino				–.0041		.0062	–.0089		.0059	–.0019		.0053	–.0021		.0056	–.0053		.0054
Disadvantage index				–.0253		.1378	.0035		.1309	–.0667		.1162	–.1350		.1241	–.0457		.1182
Residential mobility				–.0153		.0218	–.0014		.0207	–.0046		.0184	–.0078		.0197	.0008		.0187
Vacancies				.0722	**	.0270	.0676	**	.0256	.0402		.0227	.0658	**	.0243	.0518	*	.0231
White-Black segregation				–.0039		.0096	–.0057		.0091	–.0023		.0081	–.0039		.0087	–.0039		.0082
South				.3739	*	.1793	.4113	*	.1703	.3396	*	.1511	.3743	*	.1614	.4009	**	.1536
West				–.0231		.2286	.0019		.2172	.0908		.1927	.0358		.2058	.0234		.1959
Constant	–3.0951	***	.1066	–2.9320	***	.6100	–3.1049	***	.5793	–2.8947	***	.5154	–2.8917	***	.5493	–3.2895	***	.5237
Random intercept variance	.2830		.0764	.0785		.0216	.0711		.0196	.0553		.0153	.0636		.0175	.0574		.0159
Log Likelihood	–21301.816			–20923.884			–20575.666			–20571.522			–20575.755			–20395.375
AIC	42615.63			41889.77			41205.33			41193.04			41195.51			40854.75
BIC	42653.19			42021.23			41374.36			41349.55			41333.23			41055.08

Notes: N = 3867 tracts within 28 cities. 1 = estimates multiplied by 100; 2 = estimates multiplied by 1000. *** p <.001; ** p <. 01; * p < .05

Model 3 indicates that contemporary neighborhood segregation accounts for a good portion of the association between HOLC grades and burglary today. Compared to Model 2, adjusting for ethnoracial neighborhood composition reduces the gap between greenlined and both yellowlined and redlined areas by 58 and 52 percent, respectively. The gap in property crime between green- and blue-shaded areas reduces by about 61 percent and is no longer statistically significant. These patterns are similar when we adjust for housing conditions in Model 4 or concentrated disadvantage in Model 5. Accounting for the concentrated disadvantage index alone explains between 64 percent and 72 percent of the disparities between the highest appraised greenlined tracts and other-graded areas compared to Model 2.

Again, comparing the full Model 6 with either Model 1 or 2 illustrates strong support for our conceptual framework. ¹³ Although the gaps between red and green, yellow and green, and red and blue shaded areas persist, accounting for indicators of our conceptual model reduces these disparities substantially. HOLC grades matter for levels of property crime more than 70 years later, and much of the reason is that the grades help distinguish the contemporary structural fortunes of neighborhoods that lead to higher levels of burglary.

Supplemental Analyses

Models provided in the online Supplemental Material indicate the robustness of our findings to various alternative samples and specifications. First, Tables S1A and S1B in the online Supplemental Material reveal similar patterns if we include a continuous measure of the HOLC grading rather than the categorical scale used here. Second, in Tables S2A and S2B in the online Supplemental Material, we compared graded tracts with those that were not assigned a grade by the HOLC. This robustness check helps adjust for the inconsistencies in which tracts were graded across cities. Previous research (Jacoby et al. 2018; Mehranbod et al. 2022) indicates that areas of cities that were not mapped have higher levels of firearm homicide than greenlined areas. We find similar patterns for violence and burglary: Historically unmapped tracts have more crimes today than those greenlined but less crime than redlined tracts. However, including unmapped tracts does little to alter the substantive patterns pertaining to the graded tracts that we report here.

Third, because many of the cities in the NNCS2 that were mapped by the HOLC are not covered in the 1940 census, we ran models without conditioning on the availability of 1940 data, displayed in Tables S3A and S3B in the online Supplemental Material. Doing so increases the sample of tracts and almost doubles the number of cities (4,652 accounting for disadvantage, tracts across 51 cities for violence and 4,727 tracts in 52 cities for burglary). However, the pattern of results is very similar: HOLC grades exhibit a strong dose-response association with crime in unadjusted models, and our contemporary structural controls account for sizeable proportions of the gap between redlined and greenlined areas for violence and burglary. Fourth, given the likely role that redlining played in widening racial gaps in homeownership, we explored a measure of Black-White inequality in homeownership (Tables S4A and S4B in the online Supplemental Material). This variable does not explain significantly more of the gap between graded areas and crime than the models reported in Tables 2 and 3.

Fifth, we assess our decision to control for 1940 rather than 1930 census measures. As shown in Appendix 2, using the 1930 instead of 1940 census severely restricts the number of cities and the breadth of socioeconomic measures available in our analyses. Nonetheless, comparing models for cities that have overlapping data for both 1930 and 1940 permits some assessment of the consequences of our analytic decisions. Only five and six cities (for violence and burglary, respectively) have enough data from both the 1930 and 1940 censuses to allow for comparative analyses. ¹⁴ For violence, the Akaike information criterion and Bayesian information criterion statistics suggest that models with 1930 variables may provide a marginally better fit than those using 1940 variables, as shown in Tables S5A and S5B in the online Supplemental Material. More importantly, the models tell a similar story regardless of the dependent variable: HOLC grades associate with neighborhood violence and burglary net of historical controls, and the contemporary characteristics of neighborhoods largely account for these associations. That the general pattern of results holds even for this much more restricted sample suggests that our decision to rely on 1940 versus 1930 data has little consequence for our substantive conclusions. We thus prefer to employ the 1940 census to maximize the number of cities and tracts in our main analyses (see also Jacoby et al. 2018; Mehranbod et al. 2022).

Finally, Tables S6A and S6B in the online Supplemental Material provide alternative tests Finally, Tables S6A and S6B in the online Supplemental Material provide alternative tests assessing (1) whether HOLC grades contribute to model fit net of all other controls and (2) whether HOLC grades contribute more to model fit than the 1940 census variables.

Discussion and Conclusions

The government-sponsored Home Owner’s Loan Corporation appraised the creditworthiness of neighborhoods across many cities in the 1930s. Their residential security maps render visible the prevailing political economic foundations of the housing market and their attendant racism and classism. The HOLC maps may have also institutionalized and diffused the practice of redlining that discouraged investment in Black, non-White, and poor areas. As such, they present an opportunity to assess common but rarely tested assertions in the communities and crime literature that tie contemporary inequality in the distribution of crime to historical interventions in the housing market (Lyons et al. 2022; Peterson and Krivo 2010; Sampson 2012). Although recent studies trace these maps to various present-day inequalities, few examined the consequences for neighborhood crime across multiple cities. Pairing data from the National Neighborhood Crime Study (Wave 2) with HOLC maps, we assessed the consequences of inheriting HOLC grades for neighborhood crime across at least 27 cities circa 2010. We find that neighborhoods with a relatively poor HOLC grading had more violence and burglary some 70 years later. Our results are robust to various sample sizes and model specifications. Overall, our study joins a growing list of others that suggest the enduring weight of redlining maps for contemporary urban fortunes.

Our findings provide tentative support for an institutionalization (rather than spuriousness) perspective, which posits that the HOLC maps and methodology altered neighborhood trajectories of investment and structural development. We advanced a conceptual framework in which historic redlining shapes contemporary neighborhood crime through subsequent racial residential segregation, homeownership, lending practices, and other socioeconomic dynamics. In support of institutionalization, our multilevel models generally align with these assertions: Contemporary levels of concentrated disadvantage, racial residential segregation, and housing conditions account for a large proportion of the association between historic HOLC grades and crime.

We are able to account for most of the disparity in violence and burglary between redlined and greenlined areas with our modeling framework. Nonetheless, gaps remain among the historic grades even in our most fully specified models. Future work should continue to refine the conceptual model by highlighting additional structural factors that may explain the residual influence of the HOLC grades on serious neighborhood crime. Toward that end, we need to map out comprehensively how communities shifted in the intervening decades in response to housing policy and investment that unfolded after the creation of the HOLC maps (Krieger et al. 2020). Tracing how neighborhood socioeconomic, demographic, and political conditions developed over time within and across the HOLC gradient is necessary for a fuller articulation of the institutionalization perspective. Although our study provides a foundation for future research, we lack information on the structural characteristics of neighborhoods in the decades between the HOLC maps and present-day crime. With detailed longitudinal data, research should be able to reveal whether and how and under what contexts the HOLC grades shifted the fortunes of U.S. neighborhoods over time in ways that influence the patterning of crime today. Such data would also permit fuller analyses of the mediation pathways that the institutionalization perspective hypothesizes. Although we employ methods to test the statistical significance of changes in the HOLC parameter estimates across models that are consistent with the idea of mediation, our data do not lend themselves to a full assessment of causal pathways and counterfactual mediation (VanderWeele 2016). A fuller test of the institutionalization perspective thus awaits more complete data and analyses.

We find that the associations between HOLC grades and violence and property crime are robust to controls for socioeconomic and racial demographic factors measured near the origin of the maps. At face value, this further buttresses the general claims of the institutionalization perspective and limits the plausibility of spuriousness. We caution, however, that we cannot provide a definitive test of spuriousness. Skeptics, for example, might question whether our models miss other important historical antecedents of the maps related to the political economy of place. Although the 1940 census provides more detail on the socioeconomic and demographic character of places than previous decennial census data, the HOLC based their maps on complex and dynamic appraisal methods that are hard to capture with “snapshots” of administrative data.

Studies based on a single or a small number of cities inform the bulk of our understanding of the HOLC appraisal maps and their implications for current levels of inequality. Yet the historical record suggests that the creation and implementation of HOLC maps varied by city. Following Mehranbod et al. (2022), we encourage future work to explore city-level variation in more detail. One particularly fruitful line of inquiry may be to examine cross-level interactions to explore the extent to which broader contemporary and historical city contexts shape the relationship between HOLC grades and crime. In part because political economic growth machine agendas developed differently across the urban landscape, a multilevel framework should provide a fuller and more complex understanding of the legacy of the HOLC map for neighborhood levels of crime across the country.

In conclusion, our study suggests that the historically intertwined racist and classist ideologies and practices underlining the political economy of the housing market cast a long shadow on inequalities in the spatial distribution of crime today. Crime as an enduring consequence of inheriting the HOLC grades highlights the role of institutionalized, government-sponsored housing policy interventions in making or breaking the well-being of neighborhoods. National Science Foundation grants (SES-1357207 and SES-1356252) partially funded this research.

Supplemental Material