The Lingering Legacy of Redlining on School Funding, Diversity, and Performance Between 1935-1940 the Home Owners' Loan Corporation (HOLC) assigned A (minimal risk) to D (hazardous) grades to neighborhoods that reflected their "mortgage security" and visualized these grades on color-coded maps used by banks and other mortgage lenders to provide or deny home loans within residential neighborhoods. In this study, we leverage a spatial analysis of 144 HOLC-graded core-based statistical areas (CBSAs) to understand how historic HOLC maps relate to current patterns of school funding, racial diversity, and performance. We find districts and schools located today in historically redlined neighborhoods have less district-level per-pupil revenues, larger shares of Black and non-White student bodies, less diverse student populations, and worse average test scores relative to those located in A, B, and C neighborhoods. These nationwide results are consistent by region and when controlling for CBSA. Finally, we document a persistence in these patterns across time, with overall positive time trends regardless of HOLC grade but widening gaps between D vs. A, B, and C outcomes. These findings suggest that education policymakers need to consider the historical implications of redlining and past neighborhood inequality on neighborhoods today when designing modern interventions focused on improving life outcomes of students of color. Suggested citation: Lukes, Dylan, and Christopher Cleveland. (2021). The Lingering Legacy of Redlining on School Funding, Diversity, and Performance. (EdWorkingPaper: 21-363). Retrieved from Annenberg Institute at Brown University: https://doi.org/10.26300/qeer-8c25 VERSION: March 2021 EdWorkingPaper No. 21-363 Dylan Lukes Harvard University Christopher Cleveland Harvard University
85
Embed
The Lingering Legacy of Redlining on School Funding ... · Redlining, where the Home Owners’ Loan Corporation (HOLC) and the Federal Housing Administration (FHA) assigned A-D security
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
The Lingering Legacy of Redlining on School Funding, Diversity, and Performance
Between 1935-1940 the Home Owners' Loan Corporation (HOLC) assigned A (minimal risk) to D (hazardous) grades to neighborhoods that reflected their "mortgage security" and visualized these grades on color-coded maps used by banks and other mortgage lenders to provide or deny home loans within residential neighborhoods. In this study, we leverage a spatial analysis of 144 HOLC-graded core-based statistical areas (CBSAs) to understand how historic HOLC maps relate to current patterns of school funding, racial diversity, and performance. We find districts and schools located today in historically redlined neighborhoods have less district-level per-pupil revenues, larger shares of Black and non-White student bodies, less diverse student populations, and worse average test scores relative to those located in A, B, and C neighborhoods. These nationwide results are consistent by region and when controlling for CBSA. Finally, we document a persistence in these patterns across time, with overall positive time trends regardless of HOLC grade but widening gaps between D vs. A, B, and C outcomes. These findings suggest that education policymakers need to consider the historical implications of redlining and past neighborhood inequality on neighborhoods today when designing modern interventions focused on improving life outcomes of students of color.
Suggested citation: Lukes, Dylan, and Christopher Cleveland. (2021). The Lingering Legacy of Redlining on School Funding, Diversity, and Performance. (EdWorkingPaper: 21-363). Retrieved from Annenberg Institute at Brown University: https://doi.org/10.26300/qeer-8c25
VERSION: March 2021
EdWorkingPaper No. 21-363
Dylan LukesHarvard University
Christopher ClevelandHarvard University
1
The Lingering Legacy of Redlining on School Funding, Diversity, and Performance
bracket, 1939 rent bracket, rental demand, and predicted rent trend. 3) New Construction
captured the number of new properties built within the past year, the prices for these units, and
how they were selling. 4) Overhang of Home Properties captured unsold HOLC properties. 5)
Sale of Home Properties captured sold HOLC properties 6) Mortgage Funds captured mortgage
funds’ availability. 7) Total Tax Rate per $1000 captured the local tax rate. 8) Description and
Characteristics of Area captured other qualitative detail about the terrain and population.2 These
2 The area descriptions for Los Angeles A-1 and D-1 can be viewed in Appendix A2.
6
grades were solidified as color-coded maps in which D-graded areas were colored red and
prompted the later coinage of the term “redlining.”
A current debate exists as to whether HOLC A-D grade assignments were racially biased
or merely a geographic snapshot of an outcome caused by America’s long history of racial
discrimination before it (Fishback et al., 2020). Similarly, there is an ongoing debate about how
the specific HOLC maps affected lending practices. Some authors argue that access to HOLC
maps was limited, while others argue that access to the maps was widespread. Regardless, it is
clear that HOLC encouraged the general practice of using maps to classify the creditworthiness
of neighborhoods (Hoffman et al., 2020). There is evidence that predominantly non-white
Northern European neighborhoods were more likely to receive lower grades as explicit factors in
the grading process. Broadly the HOLC maps are considered to have redirected both public and
private capital and homeownership for intergenerational wealth-building to native-born white
families and away from African American and immigrant families (Appel & Nickerson, 2016;
Krimmel, 2018; Anders, 2018; Aaronson et al., 2021).
Modern Outcomes Linked to HOLC Maps
A literature base that leverages both descriptive and causal empirical strategies is
developing that links neighborhoods’ HOLC grades to various modern outcomes.
Mitchell and Franco (2018) descriptively analyze the modern demographic and
residential patterns of HOLC-graded areas. Most of the neighborhoods (74%) that the HOLC
graded as high-risk or “Hazardous” eight decades ago are low-to-moderate income today.
Additionally, most of the HOLC graded “Hazardous” areas (nearly 64%) are minority
neighborhoods now. There is significantly greater economic inequality in cities where more
7
HOLC-graded high-risk or “Hazardous” areas are currently minority neighborhoods.
Gentrification is associated with a greater economic change in the HOLC highest risk
“Hazardous” neighborhoods and higher levels of interaction between black and white residents,
but also greater economic inequality in cities. Cities in the South showed the least change in the
HOLC-evaluated “Hazardous” neighborhoods that today have lower incomes and higher
populations of majority-minority residents. The Midwest closely followed the South in the
persistence of low-to-moderate income (LMI) neighborhoods and HOLC “Hazardous” areas.
Two papers have highlighted the impacts of the HOLC maps on housing outcomes.
Appel and Nickerson (2016) use a spatial regression discontinuity boundary design to study
long-term impacts on home prices. They show that housing characteristics varied smoothly at the
boundaries when the maps were created. Despite this initial smoothness, they find that "redlined"
neighborhoods have 4.8% lower home prices, fewer owner-occupied homes, and more vacant
structures relative to adjacent areas. Similarly, Aaronson, Faber, Hartley, Mazumder, and
Sharkey (2021) use a spatial regression discontinuity boundary propensity score design to study
the effects of the HOLC maps on the long-run trajectories of neighborhoods. They find that the
maps led to reduced homeownership rates, house values, and rents and increased racial
segregation in later decades. They conclude that the HOLC maps had meaningful and lasting
effects on urban-neighborhood development through reduced credit access and subsequent
disinvestment (Aaronson et al., 2021).
Other papers have explored a broader set of outcomes connected to the HOLC maps.
Jacoby, Dong, Beard, Wiebe, and Morrison (2018) use descriptive spatial analysis to evaluate the
relationship between HOLC grades and modern firearm violence in Philadelphia. After adjusting
for socio-demographic factors when the map was created from the 1940 Census, firearm injury
8
rates are highest in historically red-zoned areas of Philadelphia (Beard et al., 2017; Jacoby et al.,
2018). Nardone, Casey, Rudolph, Karasek, Mujahid, and Morello-Frosch (2020) descriptively
explore how birth outcomes within California vary based upon HOLC grade. They find that a
lower HOLC grade is associated with adverse birth outcomes, although this relationship was less
clear after propensity score matching and stratifying by metropolitan area (Nardone et al., 2020).
Hoffman, Shandas, and Pendleton (2020) use descriptive spatial analysis to study the association
between a neighborhood's HOLC grade and current surface temperatures. They find that 94% of
studied areas display consistent city-scale patterns of elevated land surface temperatures in
formerly redlined areas relative to their non-redlined neighbors by as much as 7 °C. Regionally,
Southeast and Western cities display the greatest differences while Midwest cities display the
least. Nationally, land surface temperatures in redlined areas are approximately 2.6 °C warmer
than in non-redlined areas.
These recent studies suggest a strong association between historical practices for
assigning grades to neighborhoods and modern outcomes. The HOLC maps and related historical
policies have contributed to racial disparities in neighborhood diversity, income, health care,
access to healthy food, incarceration, and public infrastructure investment. A current debate
exists as to whether HOLC A-D grade assignments were racially biased or merely a geographic
snapshot of an outcome caused by America’s long history of racial discrimination prior to it
(Fishback et al., 2020). Similarly, there is an ongoing debate about how the specific HOLC maps
affected lending practices. Some authors argue that access to HOLC maps was limited, while
others argue that access to the maps was widespread. Regardless, it is clear that HOLC
encouraged the general practice of using maps to classify the creditworthiness of neighborhoods
(Hoffman et al., 2020). We consider this paper’s main contribution its exploration of the
9
relationship between HOLC map grades and present-day educational outcomes of school
funding, school diversity, and school performance as these outcomes are intricately connected to
housing as discussed in the following section and have not yet been explored.
The Relationship Between Schools, Neighborhoods, and Student Outcomes
Education funding in the U.S. operates at the levels of federal, state, local, and within-
district. The two primary federal revenue sources are Title I and IDEA that generally distribute
dollars based on student population size and poverty concentration. At the state level, each state
has different mixes of revenue streams dedicated to education and implements a unique funding
formula to direct dollars to school districts. At the local level, districts primarily use property
taxes to generate revenue. Within a district, funding is generally distributed through a traditional
centralized model in which the district deploys resources to schools in the form of staff,
programs, and services (Urban Institute, 2017; Roza, Hagan & Anderson, 2020; Baker et al.,
2018; Brittain, Willis & Cookson, 2019; Green et al., 2021).
Researchers are finding that school funding reform has demonstrable positive impacts on
student achievement. For example, Jackson et al. (2016) examine school spending impacts using
variation due to court-ordered school finance reforms between 1972 and 1990. Event study and
instrumental variable models reveal that a 10% increase in per-pupil spending each year for all
12 years of public-school leads to 0.31 more completed years of education, about 7% higher
wages, and a 3.2 percentage point reduction in the annual incidence of adult poverty; effects are
much more pronounced for children from low-income families. Exogenous spending increases
were associated with notable improvements in measured school inputs, including reductions in
student-to-teacher ratios, increases in teacher salaries, and longer school years. Additionally,
10
Lafortune et al. (2018) study the impact of post-1990 school finance reforms on absolute and
relative spending and achievement in low-income school districts. Using an event-study research
design that exploits the apparent randomness of reform timing, they find that a one-time $1,000
increase in per-pupil annual spending sustained for ten years increased test scores by between
0.12 and 0.24 standard deviations.
Interrelatedly, there has been a growing focus on race- and income-based integration
within districts and schools to improve student outcomes. Fahle et al. (2020) use national student
test scores from the last decade and find a strong link between racial school segregation and
academic achievement gaps. Studies of the desegregation regulations of the South have found
that desegregation had a positive impact on black students and no negative impact on white
students (Ashenfelter, Collins & Yoon, 2006; Guryan, 2004; Johnson, 2019). Billings et al.
(2014) study the end of race-based busing in Charlotte-Mecklenburg schools (CMS). They find
that both white and minority students score lower on high school exams, decreases in high school
graduation and four-year college attendance for whites, and large increases in crime for minority
males when assigned to schools with more minority students. Dalane and Marcotte (2020) find
that within-school (i.e., classroom) segregation has risen by about 10 percent between 2007 and
2014 in elementary and middle schools they study. Further, they find that segregation of
economically disadvantaged students within schools is correlated with the level of segregation
between schools in districts, and this relationship grew stronger over the panel. These findings of
increasing segregation in certain parts of the country have been highlighted in several other
recent studies (Clotfelter et al., 2019; Alcaino & Jennings, 2020; Clotfelter et al., 2020;
Monarrez et al., 2020).
11
Research is also increasingly documenting how neighborhoods affect the individuals who
grow up within them. The salience of neighborhoods became well documented through the
efforts of Massey & Denton (1993) to link persistent poverty among blacks in the United States
to the unparalleled degree of deliberate segregation they experience in American cities. Chetty,
Hendren, and Katz (2016) find from the Moving to Opportunity experiment that moving from a
high-poverty housing project to a lower-poverty neighborhood did not improve scores on reading
and math tests for Black children who were less than 13 years of age at random assignment but
did improve early adult outcomes measured by subsequent educational attainments and labor
market earnings. Separately, List et al. (2020) estimate the spillover effects from a large-scale
early childhood intervention on the educational attainment of over 2,000 disadvantaged children
in the United States. They document large spillover effects on both treatment and control
children who live near treated children. The spillover effects are localized, decreasing with the
spatial distance to treated neighbors. The spillover effect on non-cognitive scores operates
through the child's social network, while parental investment is an important channel through
which cognitive spillover effects operate. More broadly, the recent development of the
Opportunity Atlas has illuminated the average outcomes in adulthood of people who grew up in
each Census tract to trace the roots of outcomes such as poverty and incarceration back to where
kids grew up (Chetty et al., 2018; Chetty & Hendren, 2018a; Chetty & Hendren, 2018b;
Murnane, 2021).
Another recent study has illuminated the connection between school quality and
neighborhood value. Bayer et al. (2020) conduct a national study of the causal impact of school
spending and local taxes on housing prices by pairing variation induced by school finance
reforms with 25 years of national data on housing prices. The results indicate that households
12
highly value school spending and spending on teachers and staff salaries. They also find that
salary spending is provided at inefficiently low levels throughout much of the United States, as
increases in salary spending within a school district funded entirely by local taxes would
generally raise house prices.
This academic literature highlights the importance of school funding, school diversity,
and broader neighborhood context on students' long-term outcomes. This paper's primary
contribution is to extend the roots of perceived issues in school funding, diversity, and
performance to the inception of redlining to demonstrate the intergenerational relationship
between HOLC A-D security ratings and present-day educational outcomes.
DATA
In what follows, we provide a succinct overview of the district, and school-level data
sources we used in our analysis, in addition to the digitized Home Owners’ Loan Corporation
(HOLC) maps created by the “Mapping Inequality” project (Nelson et al., 2020). For additional
details on each data set, please refer to Appendix A1.
HOLC Data
The prominent data for this study is the 1935-1940 HOLC maps preserved by the U.S.
National Archives and recently digitized by a team of researchers at the University of New
Richmond, Virginia Tech, the University of Maryland, and Johns Hopkins University as part of
the “Mapping Inequality” project (Nelson et al., 2020). This data is freely available online and
contains digitized versions of every available city-level HOLC map from 1935-1940 in Shapefile
or GeoJSON format. The entirety of their spatial data includes over 7,000 neighborhoods and
nearly 240 unique cities across the United States. For each HOLC city, this data contains HOLC
13
Geographic Information System (GIS) polygons, HOLC grade assignment text (i.e., A, B, C, D),
and detailed area description transcriptions (Nelson et al., 2020). For our study, we aggregate the
data to 144 unique core-based statistical areas (CBSAs), which span the United States and
include all four US Census Bureau Regions (i.e., Midwest, Northeast, South, West). These
CBSAs are captured below in Figure 1. Finally, we include all CBSAs with at least one HOLC
grade in our analysis sample to maximize geographic coverage. As a result, certain CBSAs do
not contain the full A-D HOLC rating set.
[Insert Figure 1 here]
NCES Data
We combined the HOLC data with geospatial and non-geospatial district and school-level
data from the National Center for Education Statistics (NCES). Our geospatial data includes both
district boundaries and public school-level latitude and longitude point locations. At the district
level, each year, the NCES Education Demographic and Geographic Estimate (EDGE) program
develops updated school district boundary files for public elementary, secondary, and unified
school districts. The development of unified district boundaries is done in collaboration with the
U.S. Census Bureau’s School District Boundary Review program (SDBR), where the NCES
EDGE program works jointly with the U.S. Census Bureau’s Education Demographic,
Geographic, and Economic Statistics (EDGE) Branch to develop annual composite school
district files. The 2018-2019 district boundary inputs and data layers used in our paper ultimately
derive from Census TIGER/2019 Line geospatial data (U.S. Census Bureau, 2020).3
3 We supplement this data with 2018-2019 NYC public school district boundaries from the NYC Department of City Planning
(DCP). This supplemental geospatial boundary data was required since all NYC public school districts were reported as one
unified district in the NCES EDGE geospatial data. Thus, combining the NYC DCP and NCES EDGE geospatial data allow us to
account for each of NYCDOE’s 32 school districts and more accurately capture the within district variation in HOLC grades
across the city (NYC Department of City Planning, 2020).
14
To complete our school-level analyses, we use the 2018-19 NCES EDGE school-point
location data, which provides latitude and longitude coordinates for public elementary and
secondary schools from the NCES EDGE Common Core of Data (CCD) (U.S. Department of
Education, 2020). While the NCES does have a few school-level boundary datasets for 2013-14
and 2015-16 school years from their experimental School Attendance Boundary Survey (SABS),
we opted to use the more exhaustive and annually updated NCES EDGE school point location
data.4 Like the district boundary datasets, the NCES EDGE school point locations data was
created with the help of the U.S. Census Bureau’s SDBR program. The NCES non-geospatial
data is also at the district and school level. This data includes both fiscal and non-fiscal data,
which we leverage in our analyses below on district-level financing (i.e., local, state, federal,
overall) and school diversity by HOLC grade. For the non-fiscal district and school data, we use
NCES 2017-2018 and NCES 2018-19 datasets.5 This paper leverages the enrollment by
race/ethnicity data for its analyses.
For the fiscal data, we leveraged the most recent NCES 2017-18 F-33 survey data, which
provides general financing information (e.g., revenue and expenditure totals and subtotals) at the
district-level. This data is provided by the NCES through their Common Core of Data (CCD)
(U.S. Department of Education, 2020). For district-level revenues, we use F-33 reported total
general revenues and their associated first-level subtotals (e.g., local, state, federal revenues). In
addition to these first-level subtotals, there are a variety of additional district-level finance fields
4 SABS was an experimental survey led by NCES and supported by the U.S. Census Bureau to collect school boundaries for the
2013-14 and 2015-16 school years. This effort led to the collection of over 70,000 school boundaries across the United States, but
is now discontinued (U.S. Department of Education, 2020). 5 Data were customized and pulled using the NCES Elementary/Secondary Information System’s (Elsi) “tableGenerator” tool.
Non-fiscal field options include school or district 1) general information, including name, NCES agency ID, county FIPS, state
FIPS; 2) characteristics, including agency type, core-based statistical area (CBSA), school status; 3) enrollments & enrollment
details, including total enrollments, enrollments by race/ethnicity; 4) teacher and staff information such as full-time equivalent
(FTE) teachers, pupil/teacher ratio, total staff.
15
available. In this paper, we chose to primarily fix our analyses on district-level revenue
outcomes. Finally, to facilitate an apples-to-apples comparison across districts, we transform all
district-level financial data into per-pupil terms.6
Stanford Education Data Archive (SEDA) Data
We used school-level SEDA data for all our school-performance analyses (Reardon et al.,
2021).7 This data provides us with students’ academic outcomes grades 3-8 spanning 2009-2018
and includes students’ average test scores, test score trends, and learning rates. Average test
scores and test score trends reflect educational opportunities and changes therein, while the
learning-rates outcome measures the amount learned per grade (Reardon et al., 2021). We rely
on SEDA’s cohort standardized (CS) scale achievement estimates based on the Spring 2009 4th
grade cohort for each of these student achievement measures. The CS scale achievement
estimates are measured in SD units relative to the national average and are calculated using OLS
and Empirical Bayes (Reardon et al., 2021).8 Pairing these measures with 1935-1940 HOLC
maps allows us to look at how redlining maps from over eight decades ago are associated with
the current state of educational opportunity and student learning rates today. In the following
section, we describe our approach for both the SEDA and NCES data.
METHODS
We use an approach motivated by Hoffman, Shandas, and Pendleton (2020), which
documents the impacts of redlining policies on resident exposure to intra-urban heat. While at
6 Since all fiscal data was transformed into per pupil terms, district enrollment weights were required when calculating per pupil
weighted averages by HOLC grades. See “Analysis” section for additional details. 7 At the time of writing, Version 4.0 was the most recent SEDA data available. 8 As per SEDA, OLS estimates are more appropriate here than EB estimates since we use precision weights in our regression
models. Regardless, our results are robust to the underlying estimation procedure.
16
times our papers’ methods deviate, the overarching analytic templates are similar.9 In what
follows, we describe our paper’s methods beginning with an overview of how each district and
school-level analysis sample was constructed and ending with our analytic approach, which
remains the same regardless of the unit of analysis or outcome.
School-Level Analysis Samples
A key piece of this paper is the connection between the geospatial HOLC maps and the
geospatial NCES public district and school-level data. For the NCES school-level geospatial
data, our approach is straightforward. First, we subset our data to include only those cities with at
least one assigned HOLC grade. This exercise gives us 144 CBSAs across 38 U.S. states. Next,
we use the “Join Features” tool in ERSI’s ArcGIS Summarize Data toolbox to map A-D HOLC
ratings to all U.S. public primary and secondary schools. Specifically, we execute a “one to
many” join via the ArcGIS “completely contains” spatial relationship option. This option
matches features from disparate data sets only if a target layer features from one dataset
completely contain join layer features from another. For this exercise, HOLC rating and public
primary and secondary schools are the target and join layers, respectively.10 This match leads to
n = 10,013 U.S. public primary and secondary schools mapped to a unique A-D HOLC rating in
our sample. Once NCES public primary and secondary schools are mapped to HOLC grades, we
merge it with the NCES 2018-19 school-level data. Once data merges are complete, our analysis
9 This is largely an artifact of outcomes analyzed. Hoffman et al. (2020) map intra-urban land surface temperature anomalies to
HOLC ratings, whereas this paper maps schools and districts to HOLC ratings. 10 Recall, the NCES 2018-19 public school geospatial data is composed of latitude and longitude point locations, so when we
execute a “one to many” join via the ArcGIS “completely contains” option we are simply mapping those latitude and longitude
points to the HOLC grade it is contained within. If a school falls outside an HOLC grade, it will not receive a grade.
17
sample includes n = 10,010 U.S. public primary and secondary schools mapped to unique HOLC
grades.11 This analysis sample is only for the school-level diversity outcomes (Table 1).
We also need to link our dataset to the 2009-2018 SEDA student achievement data for
the school-level student performance analysis. Doing so reduces the sample for the school-level
performance analysis by just over one quarter, giving us a total of n = 7,303 U.S. public primary
and secondary schools mapped to unique HOLC Grades (Table 1). Nearly all unmatched schools
are due to grade-level mismatches (i.e., SEDA data only covers grades 3-8). However, some
unmatched schools are also due to school year coverage (i.e., SEDA data only covers schools up
to 2018). As a result, all grade levels K-2 and 9-12 and any new public 3-8 grade level schools
created after 2018 are excluded from the school-level student performance analyses.
Importantly, the distribution of HOLC A-D grades remains quite stable across samples,
both nationwide and by CBSA. At the national level HOLC A-D distributions are nearly
identical varying only by a few percentage points. Comparing samples by CBSA we see the
exact same set of CBSAs represented across the school-level diversity and SEDA samples. In
addition, while there is a bit more variation in HOLC A-D distributions across samples at the
CBSA level, these are generally driven by CBSAs with already low district counts in the full
school-level diversity sample. For a more thorough overview of the school-level analysis
samples, please refer to Appendix A1, where we have included sample breakdowns by CBSA
and HOLC security ratings.
District-Level Analysis Sample
11 The initial merge with 2018-19 NCES school-level data matches 98.5% of schools. The remaining unmatched 1.5% of schools
are subsequently matched with NCES 2017-18 data. Only 1 school remains after matching 2018-19 orphans with 2017-18 NCES
schools. This school is dropped from the data giving n = 10,012. Finally, two schools mapped to the E HOLC rating, where E
represents “other”, were also dropped giving us a total of n = 10,010 observations for the school-level diversity analysis sample.
18
For the NCES district-level geospatial data, our approach is slightly more nuanced. As
before, we first subset our data to include only those cities with at least one HOLC rating, which
leads to 144 CBSAs across 38 states. Next, we use the “Summarize Within” tool in ERSI’s
ArcGIS Summarize Data toolbox to calculate the area of each HOLC grade polygon within each
NCES public school district, with school district boundaries set as the polygon layer and HOLC
grades set as the summary layer. By design, this tool also calculates the total area of each public
school district included in the NCES 2018-19 geospatial boundary data. Finally, we calculate the
HOLC grade weighted average for each public school district with weights based on each HOLC
grade polygon area. To view HOLC maps overlaid with NCES district boundaries, please refer to
Appendix A3 where we provide NCES district boundary maps superimposed on 1935-1940
HOLC maps for a variety of CBSAs found in our samples (e.g., Cleveland Municipal School
District, Los Angeles Unified School District).
Following this exercise, each public school district included in the NCES 2018-19
geospatial data now has a HOLC grade assigned to it; however, not all public-school districts
have polygon boundaries (e.g., charter schools). To capture these omitted districts, we use NCES
2018-19 school district geospatial data with district latitude and longitude point locations.
Following a similar procedure outlined above for our school-level point location mapping to
HOLC grades, we obtain HOLC grades for each previously omitted district point location. Next,
we append the polygon district datasets and point location district datasets, each with their
respective HOLC grade mappings, to create our n = 4,220 district to HOLC grade mapping
dataset12. Finally, we merge this dataset with the 2017-18 NCES district-level F-33 fiscal data to
get our final district-level analysis sample of n = 3,930 and d = 2,135 unique districts (Table 1).
12 Given the natural overlap between the NCES 2018-19 district polygon boundary dataset and the NCES 2018-19 district point
location dataset, duplicates from the point location dataset were removed once appended.
19
Nearly all unmatched NCES districts are due to a mismatch in school year coverage (i.e., F-33
data only covers the 2017-18 school year). As before, we provide a more thorough overview of
analysis samples in Appendix A1.
[Insert Table 1 here]
District & School-Level Analytic Approach
Our primary analytic method is based on a difference of means equation that accounts for
urban area variation in our sample. Specifically, we compare how different (e.g., more diverse or
less diverse schools, more per pupil district revenue or less per pupil district revenue), on
average, a particular homogenous set of HOLC ratings (e.g., all A’s) are from another
homogenous set (e.g., all D’s) within a given urban area. This is similar to the approach taken by
Hoffman, Shandas, and Pendleton (2020), however, instead of calculating 𝛿 for each HOLC
grade polygon and subsequently deriving the average 𝛿 by HOLC rating for a given urban area,
we move straight to the final step. The differencing equation is as follows:
As seen above, 𝛿𝑌𝑖ℎ has two components. The first, 𝑌𝑖ℎ�̇� is the average of our outcome variable
Y across all k = 1, 2, …, n HOLC polygons of type h = A, B, C, D in the ith urban region. The
second, 𝑌𝑖ℎ�̇̇�
is the average of our outcome variable Y for the entire ith urban area. We use the
results from this differencing exercise to compare how 𝛿𝑌𝑖ℎ varies by HOLC grade by urban
area. These comparisons are done both nationwide as well as by region. In addition to these
comparisons, we also look at how the unadjusted means, without differencing out urban area
averages, vary by HOLC grades. Given there is very little difference between the CBSA adjusted
20
and unadjusted procedures, we have chosen to elevate the latter for ease of interpretation. All
CBSA adjusted results are included in Appendix C.
We take a couple of different approaches to facilitate these comparisons. On the
descriptive side, for each of our outcome variables, we create density plots by HOLC grade both
nationwide and by the region. For a given outcome, to determine whether differences in the
HOLC grade distributions are statistically significant, we use the One-Sample Kolmogorov–
Smirnov (K-S) test. Distributional plots by HOLC grade and their associated K-S equality tests
are in the “V. Results” section below. Next, to determine whether differences in our outcome by
HOLC grade are statistically significant, we apply a Tukey’s Honest Significant Differences
(HSD) post hoc multiple comparison test. Tukey’s HSD test is executed after first running a
weighted OLS regression of a given outcome on indicator variables for HOLC grades with
weights equal to the school or district total student populations.
Finally, for our school-diversity analysis, we also need to calculate the within-school-
between-student diversity metric. To do so, we opt for Simpson’s Diversity Index (1-D), which
calculates the probability that two randomly selected students from a given ith school will belong
to different racial groups (Simpson, 1949; Hirschman, 1964). Simpson’s Diversity Index ranges
from 0 to 1 with larger values representing greater within-school-between-student racial diversity
and is calculated as follows:
where pr represents the probability of that two randomly selected students from a given ith school
will belong to the same race. Additionally, r represents the 7 commonly used racial categories by
the U.S. Census Bureau including American Indian or Alaska Native, Asian, Black or African
21
American, Hispanic, Native Hawaiian or Other Pacific Islander, White, and Multiple Races.
About three percent of the schools in our sample were missing student body breakdowns by race
due to data quality flags or failure to report. These schools were omitted from the analysis. As a
complement to Simpson’s Diversity Index, we also consider the Exposure Index, a popular
segregation index in the social sciences and one that approximates between-group contact and
interaction for a given race-group pair (Massey & Denton, 1988). We use this index to evaluate
segregation for all schools in our sample by A-D HOLC grade for various majority-minority
race-group pairs, including White-Asian, White-Black, White-Hispanic and White-Non-White.
RESULTS
To facilitate a more meaningful discussion of our results, we partition them by the three
overarching outcomes addressed in this paper: 1) district financing, 2) school-level diversity, and
3) school-level student performance. Each section includes density plots and margins plots, both
overall and by region, for each respective outcome by HOLC rating. Each section also discusses
results from various statistical tests we use to compare outcomes by HOLC A-D grade, including
the K-S equality of distribution test and the post hoc Tukey HSD test (Appendix B). Finally, for
both the district financing and school-level diversity outcomes, we conclude with a time series
analysis that looks at how the relationship between HOLC A-D grades and respective outcomes
has changed over the past three decades.13
Overall, our results shed light on a hitherto unexplored relationship between historic
1935-1940 HOLC A-D maps and present-day and historical district and school-level outcomes.
Notably, we find that more often than not, the modern-day district-level financing, school-level
13 Our time series analyses does not include student achievement outcomes as SEDA school-level student achievement data only
spans 2009-2018.
22
diversity, and school-level student performance outcomes are worse for those schools and
districts located today in the least favorable redlined D neighborhoods. Our longitudinal analysis
confirms these findings hold across the three decades we consider but also highlights a widening
gap in average outcomes between D vs. A, B, and C schools across time. This especially true for
D vs. A and B schools and districts. In the sections that follow, we present the district-level
financing results first and subsequently move to the school-level analyses, covering school
diversity and school student performance.
District Finances
Current Outcomes
Before describing the statistical test results of this section, we first graphically present the
district finance data for each outcome.14 Looking at the distribution plots (Figure 2) below,
nationally, one can see distributional variation by HOLC grade across all F-33 district-level
finance outcomes, where districts with higher HOLC grade ratings (i.e., A, B) often have better
distributional per-pupil finance outcomes relative to districts with lower HOLC grade ratings
(i.e., C, D). This is especially true for districts with D ratings, although some favorable
distributional patterns emerge for per-pupil state and federal revenues (Figure 2.II-2.III).
Regionally, while there are some similarities with the national narrative, one does see
differentiated region-specific trends across these outcomes. For example, Midwest and Northeast
regions closely mirror the nationwide trends described above, especially when comparing
districts with D ratings to their higher rated A, B, and C counterparts. In comparison, the South
14 While not included here, we also evaluated other district finance outcomes, including total per-pupil expenditures, per-pupil
instructional salaries, and per-pupil total benefits. Since each exhibited patterns similar to those described below for per-pupil
total revenue, they were omitted.
23
and West regions exhibit much greater distributional parity by HOLC grade. This is true across
nearly all F-33 district-level finance outcomes reported. An important caveat – the South region
results are driven, at least in part, by the fact so few districts with A and B HOLC ratings exist in
the sample. This is confirmed by the relatively low densities of A districts compared to C and D
districts in the density plots.
Looking more closely at overall per-pupil revenue, both nationally and by region, we see
material distributional differences by district HOLC grade where districts with higher HOLC
grade ratings are often shifted to the right. What is driving this? A more cohesive story emerges
after breaking out total per-pupil revenue by its constituent per-pupil local, state, and federal
revenue components. At the national level, we see that districts with higher HOLC grade ratings
(i.e., A, B) exhibit a rightward distributional shift in local revenues relative to districts with
lower HOLC grade ratings (i.e., C, D). While state and federal per-pupil revenue distributional
differences favor C and D districts relative to their A and B counterparts, they are not large
enough to offset the initial deficit in local funding. Nationally, all distributional differences
between D districts and their higher-rated A, B, and C counterparts are statistically significant
(Table B1.1).
Across regions, we see similar patterns play out, albeit with some variation. For example,
for the Northeast, distributional outcomes for A-D districts are nearly identical at the state and
local level. However, at the federal level, distributions of districts with higher HOLC grade
ratings (i.e., A, B) are shifted to the right relative to districts with lower HOLC grade ratings
(i.e., C, D). For the Northeast, distributional differences in per-pupil federal revenues across
HOLC grade drive those distributional differences seen for overall per-pupil revenue. In
comparison, the West and South have near distributional parity by HOLC grades across all F-33
24
district-level finance outcomes, while the Midwest almost perfectly mirrors the overall
nationwide results. For the Midwest and Northeast, most distributional differences between D
districts and their higher-rated A, B, and C counterparts are statistically significant. This is far
less the case for the South and West regions (Table B1.1).
[Insert Figure 2 here]
The general patterns described above are further explored here using margins plots and
Tukey HSD tests. Table B1.2 provides weighted averages and standard errors of each per-pupil
district finance variable, and Table B1.3 provides average differences and Tukey HSD tests
across all A-D HOLC grade pairs. Each is broken out by region and includes results at the
national level. We supplement these tables with margins plots (Figure 3), which include
weighted averages by HOLC grade, 95% confidence interval bands, and weighted grand means.
Below we discuss a few key trends that emerge from these analyses.
To start, both nationwide and by region, average total per-pupil revenues are lower for
districts with D HOLC grades relative to those with A, B, and C grades. Tukey HSD tests
confirm these differences are statistically significant at the national level. However, results are
inconsistent across regions, particularly the South and West. Next, comparing average per-pupil
total revenues across regions by A-D HOLC grade, the Northeast outperforms all others
independent of A-D status. In contrast, the South exhibit some of the lowest per-pupil total
revenue averages of all regions.
At the state and federal level, our results corroborate the density-plot narrative above.
Namely, while C and D district per-pupil averages outpace A and B district averages in a
statistically significant way, this redistributive state and federal gain are not enough to close the
large and statistically significant gap in average local per-pupil revenues that favors A and B
25
districts relative to C and D. This holds for the Midwest and Northeast regions, with inconsistent
patterns across the South and West.
[Insert Figure 3 here]
Longitudinal Outcomes
In this section, we expand our analysis to look at how, if at all, the relationship between
HOLC A-D grades and the finance outcomes changed over time to shed light on whether
historical school finance reforms alleviated or exacerbated gaps between those districts in
historically assigned HOLC A-D neighborhoods and district-level resources. To do so, we create
a panel dataset that spans nearly three decades and includes district-level finance data from the
late 1980s to the late 2010s. This exercise reduces our original cross-sectional district-level
analysis sample (d = 2,135) by half (dp = 1,113).15 To ensure those districts that remain in our
panel dataset are representative of those contained in the original 2017-18 cross-sectional
sample, we perform a few robustness checks across samples. While the distribution of HOLC A-
D grades across districts is slightly different between the 2017-18 cross-sectional and panel
samples, we find no evidence to suggest that average outcomes by A-D HOLC grades vary by
them (Figure B1.1, Table B1.4).16 With these details in mind, we discuss the time series results
below.
First, to better understand how district funding changed over time by HOLC A-D grade,
in Figure 4, we show A-D averages across time for each educational finance outcome, weighted
15 For the original 2017-18 cross-sectional data set we have n = 3,930. This translates to d = 2,135 and c = 144 distinct districts
and CBSAs, respectively. For the 1989-90 through 2017-18 panel data set we have np = 2,818. This translates to d = 1,113 and c
= 102 CBSAs, respectively. 16 For each F-33 finance outcome we consider in this paper, we fail to reject the null hypothesis of equality of common
coefficients across models using 1) the full cross-sectional 2017-18 sample, and 2) the partial panel 2017-18 sample, where each
regression model consists of a given finance outcome regressed on an A-D HOLC indicator variable. This result suggests no
statistically significant differences between the weighted average A-D HOLC grades across these two samples for any of our F-
33 educational finance variables.
26
by district enrollment. Each outcome is adjusted for inflation and denominated in 2018 USD.
Starting with average per-pupil total revenue, one can see near-parallel lines across time with
only marginal differences in slopes by HOLC A-D grade. In addition, compound annual growth
rates (CAGRs), calculated from 1989-90 to 2017-18, differ little by HOLC security rating and
hover around four percent regardless of HOLC grade. While equality in growth rates across time
is encouraging, it is less so after considering level differences in average per-pupil total revenue
between D districts and their A, B, C counterparts (Figure 5, Table B1.5). Here we see an initial
and subsequently increasing gap in per-pupil total revenue that favors districts located in
historically non-redlined neighborhoods (i.e., A, B, C). Thus, while growth rates are similar
across HOLC A-D security ratings, per-pupil funding gaps are not. This result is a direct
consequence of initial per-pupil total revenue gaps by HOLC A-D grade in the late 1980s paired
with near-identical A-D growth rates across time. Findings for per-pupil local revenue mirror
those for per-pupil total revenue but are even more pronounced, with larger initial funding gaps
in the late 1980s and smaller growth rates.
In contrast to per-pupil total and local revenue trends, per-pupil state and federal revenues
favor those districts with lower HOLC security ratings (i.e., C, D) relative to those with higher
(i.e., A, B). Like our 2017-18 cross-sectional findings, state and federal redistributive policies
appear to benefit those districts most in need and have consistently done so over the three
decades we consider here. For per-pupil state revenues, funding gaps in the late 1980s favor C
and D districts relative to A and B districts. Growth rates lead to C and D district convergence,
with B districts remaining mostly in parallel; however, there is a distinct increase in the gaps
between B, C, and D districts and their highest-rated A counterparts. These results are further
27
confirmed in Figure 5, where there is limited variation across years for D vs. C and D vs. B
comparisons, but a clear monotonic increasing relationship for the D vs. A group.
Finally, per-pupil federal revenues match much of the findings for per-pupil state revenues.
However, there are some notable wrinkles. First, there is a large positive spike in per-pupil federal
funding for the 2009-10 school year, so much so that the 10-year CAGR from 1999-2009 surpasses
ten percent for all HOLC A-D grade districts, with A districts outstripping all others at 12.2
percent. This precipitous increase in federal funding is consistent with the surge in education
funding from the American Recovery and Reinvestment Act of 2009 (ARRA), which allocated
around $100 billion to the U.S. Department of Education who subsequently distributed it to states
through the State Fiscal Stabilization Fund (SFSF), Title I – Part A, and the Individuals with
Disabilities Education Act (IDEA) – Part B (U.S. Department of Education, 2009). In line with
nationwide trends in per-pupil federal funding, in the decade that follows 2009-10 per-pupil federal
funding decreases across all HOLC A-D grades although remains above historical averages from
the late 1980s and 1990s. Next, while level differences in average per-pupil federal revenues
existed before 2009-10, gaps between A districts and their lower B, C, and D counterparts clearly
widen during this decade. This is true even though A districts exhibited the highest 10-year CAGR
from 1999-2009. Finally, unlike per-pupil state revenues, the D vs. A gap is not monotonically
increasing. Although we see widening gaps from 1989-90 through 2009-10 between A vs. B, C,
and D districts, they shrink in the years following the 2009-10 ARRA funding surge. This is
exhibited for the D vs. A group in Figure 5.
Overall, these results reflect broader trends in U.S. K-12 education finance, namely
increases in state and federal revenues over time resulting from a bevy of state and federal
programs that support financially hamstrung districts located in poorer communities. We show for
28
the three decades considered in this paper that, on average, those districts most in need of additional
resources due to inequality in local funding are those districts that also receive the most state and
federal resources. Notably, these districts are on the whole in neighborhoods today that were
historically redlined in the 1930s and 1940s. Thus, we find evidence of a long-standing association
between historical neighborhood inequality and present-day district-level finance inequality.
Given the recent research that shows redlining's negative impact on homeownership rates, home
values and rents (Aaronson et al., 2021), these results should not be too surprising given the close-
knit relationship between school funding and local property values.
[Insert Figure 4 here]
[Insert Figure 5 here]
School Diversity
Current Outcomes
Once again, we first present distribution plots of our key outcomes before moving to the
margins and Tukey HSD results. Looking at the distribution plots (Figure 6) below, nationally,
one can see distributional differences by HOLC grade across all four school-level diversity
outcomes. These differences manifest not in right or left distributional shifts but in differences in
densities for fixed values along the x-axis. Overall, districts with higher HOLC grade ratings
(i.e., A, B) often have more(less) density relative to their lower HOLC district peers (i.e., C, D)
at low(high) levels of percent Black and percent Non-White, and high(low) levels of percent
White and the Simpson's Diversity Index. This is especially true when comparing D districts
with A districts. There is limited regional variation across these outcomes.
29
Looking more closely at percent Black and Simpson's Diversity Index (Figure 6), the
above-described pattern manifest in distinct humps at the extreme ends of each respective
outcome density plot, with greater densities for districts with higher HOLC ratings at lower
values for percent Black and higher values of Simpson's Diversity. Notably, this bimodality is
not present for A schools with a single peak (and the greatest density of all HOLC grades) at
high levels of the Simpson's Diversity Index. While this bimodality does not exist in the percent
White and percent non-White, it is clear from the density plots that districts with D ratings have
significantly more shares of their student body composed of Black and Non-White students
relative to their A and B peers. Nearly all distributional differences, both nationally and by
region, between D districts and their higher-rated A, B, and C counterparts are statistically
significant (Table B2.1).
These findings lend themselves to a more nuanced narrative than, for example, D schools
have predominantly Black students whereas A schools have mostly White. While it is true that
schools with A HOLC ratings have far less percent Black than D schools, they also have, on
average, greater diversity. This is further confirmed from the margins and Tukey HSD test
results below (Table B2.2, Table B2.3). Pairing these results with our above district-finance
results provides evidence that schools with A HOLC grades might have more diversity because
they can afford to. For example, given the rising importance placed on school diversity in
schools and districts across the nation, it is perhaps unsurprising that those districts and schools
with the most available resources (and thus the greatest ability to enact diversity mandates
charged to them) also have the most diverse student populations than those schools with less.
[Insert Figure 6 here]
30
The general patterns described above are further explored here using margins plots and
Tukey HSD tests. Table B2.2 provides weighted averages and standard errors of each per-pupil
school diversity variable, and Table B2.3 provides average differences and Tukey HSD tests
across all A-D HOLC grade pairs. Each is broken out by region and includes results at the
national level. We supplement these tables with margins plots (Figure 7), including weighted
averages by HOLC grade, 95% confidence bands, and weighted grand means. Below we discuss
a few key trends that emerge from these analyses.
To start, both nationwide and by region, schools with D HOLC grades relative to those
schools with A, B, and C grades have larger shares of Black and Non-White students in their
student body populations. Tukey HSD tests confirm these differences are all statistically
significant, except for the West region's percent Black outcome. For the percent White outcome,
the inverse is true, such that D schools have much smaller shares of White students than their A,
B, and C counterparts. Once again, these differences are statistically significant, both national
and across all regions. Finally, the demographic differences in student body populations between
D schools and A, B, C schools are not uniform across HOLC grades. Notably, the gaps are
widest for D vs. A school comparison and narrowest for D vs. C school comparison.
Looking across regions, the West has the lowest average percent Black student body
composition, which is true regardless of HOLC grade. In contrast, the Midwest, Northeast, and
South all have similar percent Black student body compositions, each of which is around three
times larger than the West. Next, for percent White, the Midwest region has the highest averages,
whereas the Northeast, South, and West closely mirror one another by HOLC grades.
The final outcome variable, and perhaps the most interesting when framed in the above
context, is the Simpson's Diversity Index. Moving down from A to D HOLC grades, average
31
Simpson's Diversity Index monotonically decreases, such that A schools have on average the
most diverse student populations and D schools have the least. Differences between D schools
and their higher HOLC grade counterparts (i.e., A, B, C) are statistically significant both
nationwide and across all regions. This finding corroborates the density plot results above.
Namely, even though A schools have the highest percent of White students and lowest percent of
Black students, they also have, on average, the highest probability that two randomly selected
students will be of different races.
[Insert Figure 7 here]
Longitudinal Outcomes
In this section, we expand our analysis to look at how, if at all, the relationship between
HOLC A-D grades and the diversity outcomes changed over time to shed light on whether
historical school reforms alleviated or exacerbated gaps between those schools in historically
assigned HOLC A-D neighborhoods and school-level diversity. Like the educational finance
time series analysis, we first create a panel dataset that spans nearly three decades and include
school-level student demographic data from the late 1980s to the late 2010s. This exercise
reduces our original cross-sectional school-level analysis sample (n = 10,010) by over half (np =
4,693).17 To check if those schools that remain in our panel dataset are representative of those in
the original 2018-19 cross-sectional sample, we perform a few robustness checks across samples.
Overall, the distribution of HOLC A-D grades across schools varies little between the 2018-19
cross-sectional and panel samples. However, we find evidence that average outcomes by A-D
17 For the original 2018-19 cross-sectional data set we have n = 10,010. This translates to d = 2,037 and c = 141 distinct districts
and CBSAs, respectively. For the 1988-89 through 2018-19 panel data set we have np = 4,693. This translates to d = 478 and c =
31 CBSAs, respectively.
32
HOLC grades differ by sample.18 While often small, these differences are almost always
statistically significant for A-D HOLC grades across each diversity outcome except for
Simpson’s Diversity Index, which has only statistically significant differences between samples
for the “B” security rating (Figure B2.1, Table B2.4). With these details in mind, we discuss the
time series results below.
First, to better understand how school student demographics and diversity changed over
time by HOLC A-D grade, in Figure 8 and Figure 9, we show A-D averages and differences
across time for each diversity outcome, weighted by school enrollment. All mean outcomes and
differences are captured in Table B2.5. Starting with the percent Black outcome, overall, we see
negative downward sloping convex trends from 1988-89 through 2018-19. For those schools
located in historically rated A and B neighborhoods, there is a small uptick in percent Black from
1988-89 to 1998-99 reaching 24.8 and 26.5 percentage points, but this overwhelmed by negative
trends in the two decades that follow, so much so that both groups end up below their original
1988-89 shares. Positive gaps between D vs. A, B, and C schools exhibit an initial downward
trend from 1988-89 to 2008-09 and flattens out over the final decade. The gap in average shares
of Black students is largest between the D vs. A group beginning at 13.4 percentage points in
1988-89 and shrinking to 10.3 percentage points by 2018-19. In comparison, the gap between D
vs. B and D vs. C schools begins at 10.8 and 7.8 before decreasing by a few percentage points
over the following decades to end at 7.4 and 6.5 percentage points, respectively.
18 For each diversity outcome we consider in this paper, we reject the null hypothesis of equality of common coefficients across
models using 1) the full cross-sectional 2018-19 sample, and 2) the partial panel 2018-19 sample, where each regression model
consists of a given diversity outcome regressed on an A-D HOLC indicator variable. This result suggests statistically significant
differences between the weighted average A-D HOLC grades across these two samples for each of our diversity outcomes. The
exception being the Simpson’s Diversity outcome, which fails to reject the null hypothesis of equality of common coefficients for
A, C, and D HOLC grades across samples.
33
For the percent White outcome, we see parallel downward sloping lines across time with
negative gaps between D vs. A, B, and C counterparts remaining mostly flat from 1988-89
through 2018-19. Thus, although student bodies have become less White over time across all
HOLC A-D grades in our time series sample, they have done so at similar rates. There is a clear
rank order by HOLC A-D grade to the lines in Figure 8 with “A” schools having, on average, the
largest share of White students, “B” the second, “C” the third, and “D” the fourth and smallest.
Thus, gaps are largest between A and D schools, with differences that hover around 35
percentage points. For the percent Non-White outcome, we see the same patterns as the percent
White outcome, except in reverse such that we observe upward sloping parallel lines across time
with positive gaps between D vs. A, B, and C counterparts remaining relatively flat from 1988-
89 through 2018-19.
Finally, for the Simpson’s Diversity (1-D) outcome, we see uniform increases across all
HOLC A-D grades from 2008-09 to 2018-19 with nearly imperceptible gaps between A and B
schools but notable negative differences between D vs. A, B, and C counterparts. However, these
patterns are not consistent across time. For example, in 1988-89, average diversity levels hover
around 0.35 across all HOLC A-D grades with gaps near zero between D vs. A, B, and C
schools. From this point forward, the diversity index steadily increases for those schools in the
highest-rated HOLC neighborhoods (i.e., A, B) while remaining relatively flat for those schools
in the lowest-rated HOLC neighborhoods (i.e., C, D). Thus, by 2008-09 there are notable
negative gaps in the diversity index between D vs. A, B, and C schools that continue to grow into
2018-19. This pattern is especially true for D vs. A, B comparisons where once near-zero gaps in
the diversity index in 1988-89 surpass -0.08 in 2018-19.
[Insert Figure 8 here]
34
[Insert Figure 9 here]
These findings reflect broader trends in U.S. K-12 public school demographics that have
led to a more racially and ethnically diverse school-age population today. For example, over the
past two decades, shares of White 5-17-year-olds have decreased from 62 percent to just over 50
percent, whereas shares of Hispanic 5-17-year-olds have increased to 25 percent from 16 percent
(NCES, 2019). These public-school demographic shifts mirror the increasing diversity of the
U.S. population, driven in part by diversifying urban demographics resulting from nationwide
migration patterns that brought White families into cities from the suburbs and Black, Hispanic,
and Asian families out to them (Wells et al., 2020). Even so, more diverse populations may not
always translate to more diverse schools. Since the 1990s, court-ordered desegregation plans
from the 1960s and 1970s have been gradually lifted, leading to increased school segregation
(Lutz, 2011; Reardon et al., 2012; Reardon et al., 2019). Also, intergroup exposure between
Whites and Non-White students has decreased since the 1990s, with Non-White students
attending schools with fewer shares of White students (Fiel, 2013).19 These countervailing forces
could overwhelm, or at a minimum, limit some of the benefits that a more diverse U.S.
population has on school diversity.
With the above in mind, our paper shows student diversity increasing over time for all
HOLC A-D grades. Notably, we see little movement in Simpson's Diversity Index from the late
1980s through the late 2000s for those in historically redlined D neighborhoods. This might once
again reflect the waning influence of court-order desegregation plans starting in the 1990s, and
standalone could be a harbinger of resegregation in the years to follow. However, this trend
19 Importantly, Fiel (2013) finds this result was due to a growing share of the minority population relative to whites, not from
increasing between-group segregation. This is reflected in the negative trends in percent White and positive trends in percent
Hispanic and percent Non-White we present in this paper.
35
reverses and spikes upward in the last decade, joining already upward sloping trend lines for A,
B, and C schools. While these patterns could reflect more recent efforts that target school
diversity through new avenues such as SES integration instead of historical policies based on
race integration (Wells et al., 2020), it may also be an artifact of our chosen diversity measure.
To complement the Simpson’s Diversity Index results, we alternatively look at the
Exposure Index for various race-group pairings. This is a well-known segregation index used
frequently across the social sciences and differs from Simpson’s Diversity Index by looking at
exposure to different races via race-pair groupings. Looking at Figure B2.2 and Table B2.6, one
can see an increase in White vs. Non-White exposure over time, driven largely by the increasing
White-Hispanic Exposure Index, although the White-Asian Exposure Index does increase
moderately over time, too. This further corroborates that U.S. schools, at least those in urban
areas we consider in our sample, are beneficiaries of shifting U.S. demographics. Also, the
decreasing White-Black Exposure Index closely mirrors nationwide trends and lends evidence to
the White-Black resegregation narrative resulting from a reversal of desegregation efforts in the
1960s and 1970s.
Finally, over the three decades we consider in our time series analysis, we observe gaps
between D vs. A, B, and C grades that are persistent and often growing over time. These
inequalities between those schools located today in what were historically the best-rated
neighborhoods and those located today in what were the worst-rated neighborhoods highlight the
potentially stubborn historical legacy of HOLC A-D map grades.
Student Performance
36
Current Outcomes
As before, we first introduce distribution plots of the outcomes and then move to
describing the results of the margins and Tukey HSD analyses. The distribution plots (Figure 10)
below, both nationally and by region, indicate no distributional differences by HOLC rating for
two of our three student performance outcomes, namely the measures of average student learning
and average student test score changes. For average student test scores, we see distributional
variation, where the distributions of schools with higher HOLC grade ratings (i.e., A, B) are
shifted to the right relative to those distributions of schools with lower HOLC grade ratings (i.e.,
C, D). This distributional shift is most apparent for schools with D ratings relative to schools
with A ratings. There is no variation between regions across each of the student performance
groups at the regional level, and overall, patterns mirror those nationwide. Nationally and by
region, all distributional differences between D districts and their higher-rated A, B, and C
counterparts are statistically significant for the average student test score outcome. While there
are some statistically significant results for the other variables, they are often minimal and not
uniformly significant across regions (Table B3.1).
[Insert Figure 10 here]
The general patterns described in the density plots above are further explored here using
margins plots and Tukey HSD tests. Table B3.2 provides weighted averages and standard errors
of each per-pupil school performance variable, and Table B3.3 provides average differences and
Tukey HSD tests across all A-D HOLC grade pairs. Each is broken out by region and includes
results at the national level. We supplement these tables with margins plots (Figure 11),
including weighted averages by HOLC grade, 95% confidence bands, and weighted grand
means. Below we discuss a few key trends that emerge from these analyses.
37
To start, both nationwide and by region, there are virtually no statistically significant
differences across HOLC grades for the outcomes of average student learning and average
student test score changes. Each of their respective HOLC grade averages, both nationwide and
by region, is not statistically different from their respective grand means. One can see this below
in Figure 11 and for each pairwise comparison in Table B3.3. In contrast, the average student
Math and ELA score outcome exhibit statistically significant differences between HOLC grades
both nationwide and across all regions. This is true for all pairwise comparisons, including D
schools vs. their higher A, B, C counterparts. Also, moving from A to B, B to C, and C to D,
average student test scores decrease monotonically such that the gap between A and D is the
widest among all pairwise comparisons. This is once again true both nationwide and across each
region.
These results tell us that while learning rates and changes in educational opportunity are,
on average, the same across all HOLC grades, overall educational opportunity is not.
Specifically, those schools located in historically D assigned neighborhoods have less
educational opportunity than those located in A, B, and C neighborhoods. For example,
nationwide A and D schools are separated by 0.53 SD units or just over 1.5 grade levels. These
gaps are present across all regions and widen to as much as 0.80 SD units or just under 2.5 grade
levels in the West region and shrink to 0.42 SD units or around 1.25 grade levels in the
Northeast. Comparing B and D schools paints a similar picture as above, albeit somewhat muted,
with schools separated by 0.27 SD units or just under one grade level. These gaps favoring B and
D schools are also exhibited within each region, increasing to as much as 0.55 SD units in the
West and decreasing to as little as 0.19 SD units in the Midwest. Finally, C and D schools show
the greatest similarity of all D vs. A, B, and C comparisons with gaps shrinking to single digits
38
nationwide and for some regions. Overall, C and D schools are separated by 0.09 SD units or just
over a quarter grade level. Differences in educational opportunity are still largest for the West
region with a gap of 0.17 SD units, while the Northeast gap is only 0.06 SD units. Both
nationwide and by region, these differences in D vs. A, B and C educational opportunity are
nearly all statistically significant.
What do these findings mean? First, and importantly, they are only a snapshot of present-
day differences in educational performance measures by HOLC A-D grade, and thus cannot
speak to whether gaps have risen or fallen over the past several decades. Without a more
exhaustive accounting of these trends over time, we cannot measure progress nor bring historical
context to bear. For example, if trends in educational opportunity favored D vs. A, B, and C
schools over the past half century, one might view the current gaps in this outcome as a historical
lower-bound and vice versa for a historical upper bound. Unfortunately, apart from the current
2009-2018 SEDA panel, we lack historical data on educational performance measures. Even so,
we can make prognostications on what might be if the status quo remains. Given the large gaps
in educational opportunity by HOLC A-D security rating and the near-zero SD unit changes in it
for each HOLC A-D grade, the educational opportunity gap is expected to remain unabated into
the future. The equality exhibited in average learning rates and average educational opportunity
changes by HOLC A-D grade, which standalone might be a positive finding, will lead to a
continued inequality in average educational opportunity across them given the large and existing
gaps in educational opportunity by HOLC A-D grade.
Finally, this equilibrium could have a silver lining if school and later life outcomes only
depend on meeting a minimum educational opportunity threshold. For example, while gaps in
educational opportunity would remain constant over time, a positive average change in
39
educational opportunity uniform across A-D HOLC grades, could in time, raise all schools to and
above the minimum educational opportunity threshold. Unfortunately, our findings do little to
support this claim, as changes in educational opportunity, while positive, are very small and
often fall below 0.01 SD units.
[Insert Figure 11 here]
CONCLUSION
Between 1935-1940, the Home Owners' Loan Corporation (HOLC) assigned A (minimal
risk) to D (hazardous) grades that encouraged or discouraged banks and other mortgage lenders
from providing home loans within residential neighborhoods. With the release of newly digitized
HOLC maps from the University of New Richmond lead “Mapping Inequality Project,” there has
been a recent surge in research quantifying the negative impacts of redlining on long-term social
and economic outcomes (Appel & Nickerson, 2016; Krimmel, 2018; Anders, 2018; Aaronson et
al., 2021). However, to the best of our knowledge, this effort has yet to extend to K-12 public
school educational outcomes.
This paper examines how historic HOLC A-D maps explain current district funding
patterns, school diversity, and school performance. We employ a novel approach of mapping
1935-1940 HOLC A-D neighborhood grades to present-day districts and schools. At the district-
level, we also find that those districts with the worst HOLC grades have, today, the least
favorable district-finance outcomes relative to their higher-rated peers.20 The finance results
show in stark relief how inequality in per-pupil funding at local levels drives inequality in per-
pupil financing overall, but also highlight the mitigating effects of redistributive federal and state
20 Recall, discrete A-D HOLC grades for districts are based on A-D HOLC weighted averages where weights were derived from
A-D HOLC grade polygon areas.
40
policies on funding gaps generated by local differences. The per-pupil federal and state funding
results of this paper, which indicate more favorable outcomes for C and D districts, are mildly
encouraging and highlight a redistributive system that is correctly targeting those districts most
in need. However, those districts in our sample most in need are often those located today in
historically redlined neighborhoods. This finding hints at the intergenerational transmission of
neighborhood inequality. Finally, much of these results persist across time, with overall positive
time trends in outcome measures regardless of HOLC A-D grade but widening gaps between D
vs. A, B, and C districts.
We also find those schools located today in historically redlined residential
neighborhoods to have, on average, larger shares of Black and non-White student bodies and less
diverse student populations, and worse average ELA and math scores. For the diversity
outcomes, these differences are persistent and growing over time, albeit for a smaller, less
representative sample.21 However, while A assigned public schools have the highest percent
White student populations, they also exhibit the highest student diversity levels via Simpson’s
Diversity Index. That said, A schools have the lowest Exposure Index values across all HOLC
A-D grades and race-group pairings (i.e., White-Asian, White-Black, White-Hispanic, White-
Non-White). Finally, there is virtually no difference in both average learning rates and trends in
test scores across A-D schools.
These results suggest that education policymakers need to consider the historical
implications of past neighborhood inequality on present-day neighborhoods when designing and
implementing complex modern interventions that target inequitable outcomes between students
of different socioeconomic and racial groups.
21 We do not have longitudinal data for the student-performance outcomes.
41
References
Aaronson, D., Faber, J., Hartley, D., Mazumder, B., & Sharkey, P. (2021). The long-run effects
of the 1930s HOLC redlining maps on place-based measures of economic opportunity and
socioeconomic success. Regional Science and Urban Economics, 86, 103622, 1-15.
Alcaino, M., & Jennings, J. (2020). How Increased School Choice Affects Public School
Enrollment and School Segregation. EdWorkingPaper No. 20-258. Providence, RI:
Annenberg Institute at Brown University. Retrieved December 8, 2020, from
https://www.edworkingpapers.com/ai20-258.
Anders, J. (2018). The Long Run Effects of De Jure Discrimination in the Credit Market: How
Redlining Increased Crime. Working Paper. Retrieved March 1, 2021, from
Figure 1. 1935-1940 HOLC CBSAs by US Census Bureau Region
Notes: The above pink dots represent each unique CBSA in our analysis sample. These total to n =144 present-day CBSAs mapped to the 1935-
1940 HOLC Residential Security Maps and are broken down by regions as follows: Northeast (n=30), Midwest (n = 53), South (n = 46), and
West (n = 15).
Table 1. Summary of Analysis Samples by Outcome Groups
Outcome Group Outcome Variable Level Year Sample Size Districts (d), CBSAs (c) Source
Fiscal Data (1) Per-Pupil Revenue
– Total, Local, State,
Federal
District 2017-18 n = 3,930 d = 2,135 c = 144
NCES F-33
Survey
Student Diversity % Black, % White, %
Non-White,
Simpson’s Diversity
Index
School 2018-19 n = 10,010 d = 2,037 c = 141
NCES
Student
Data
Student Performance
Average Test Scores,
Student Learning,
Trends in Test Scores
School 2009-19 n = 7,303 d = 1,371 c = 140
SEDA
Version 4.0
Notes: For a more detailed breakdown of outcome group samples by CBSA by HOLC A-D grade please refer to Appendix A1 where we provide
school district counts and HOLC A-D percentages by CBSA across samples.
49
Figure 2. District Per-Pupil Revenue – Total, Federal, State & Local, Overall & by Region, 2017-18 [Unadjusted]
Notes: [I] Per-pupil total revenue; [II] Per-pupil federal revenue [III] Per-pupil state revenue; [IV] Per-pupil local revenue. Epanechnikov
kernel used for univariate kernel density estimation. Per-pupil total revenues are restricted to <$40K to remove outliers. Per-pupil federal
revenues are restricted to <$6K to remove outliers. Per-pupil state revenues are restricted to <$30K to remove outliers. Per-pupil local revenues are restricted to <$40K to remove outliers.
I
III
IV
II
50
Figure 3. District Per-Pupil Revenue – Total, Federal, State, & Local, Overall & by Region, 2017-18 [Unadjusted]
Notes: [I] Per-pupil total revenue; [II] Per-pupil federal revenue [III] Per-pupil state revenue; [IV] Per-pupil local revenue. For each plot, both
overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a
regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
I
II
III
IV
51
Figure 4. HOLC A-D Averages Over Time, Finance Outcomes (1989-2018) [USD 2018] [Unadjusted]
Notes: [top left] Per-Pupil Total Revenue; [top right] Per-Pupil Federal Revenue; [bottom left] Per-Pupil State Revenue; [bottom right] Per-Pupil Local Revenue. All values represented in USD 2018. Weighted averages are from a regression of a given finance outcome on HOLC B-D
indicators with student enrollment as a precision weight. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator. Although not included above, they are included below in Appendix Table B1.5.
Figure 5. HOLC Pairwise Comparisons Over Time, Finance Outcomes (1989-2018) [USD 2018] [Unadjusted]
Notes: [top left] Per-Pupil Total Revenue; [top right] Per-Pupil Federal Revenue; [bottom left] Per-Pupil State Revenue; [bottom right] Per-
Pupil Local Revenue. All values represented in USD 2018. Weighted averages are from a regression of a given finance outcome on HOLC B-D
indicators with student enrollment used as precision weights. Brackets represent Tukey HSD 95% confidence intervals. While all HOLC A-D pairwise comparisons were made using the Tukey HSD post hoc multiple comparisons procedure, only D vs. A, B and C groups are highlighted above.
52
Figure 6. School Student Body % Black, % White, % Non-White, & Simpson’s Diversity, Overall & by Region,
2018-2019 [Unadjusted]
Notes: [I] Percent Black [II] Percent White [III] Percent Non-White [IV] Simpson’s Diversity (1-D). For each outcome, the Epanechnikov kernel was used for univariate kernel density estimation.
I
II
III
IV
53
Figure 7. School Student Body % Black, % White, % Non-White, & Simpson’s Diversity, Overall & by Region,
2018-2019 [Unadjusted]
Notes: [I] Percent Black [II] Percent White [III] Percent Non-White [IV] Simpson’s Diversity (1-D). For each plot, both overall and by region,
the horizontal line represents the grand mean, weighted by student enrollment. Each black dot represents the weighted average outcome by
HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a regression of the above
I
II
III
IV
54
outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Figure 8. HOLC A-D Averages Over Time, Diversity Outcomes (1988-2018) [Unadjusted]
Weighted averages are from a regression of a given diversity outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator. Although not included above, they are included below in Table B2.5.
Figure 9. HOLC Pairwise Comparisons Over Time, Diversity Outcomes (1988-2018) [Unadjusted]
Weighted averages are from a regression of a given diversity outcome on HOLC B-D indicators with student enrollment used as precision
weights. Brackets represent Tukey HSD 95% confidence intervals. While all HOLC A-D pairwise comparisons were made using the Tukey HSD post hoc multiple comparisons procedure, only D vs. A, B and C groups are highlighted above.
55
Figure 10. School-Level Average Student Math & ELA Scores, Student Learning Per Year & Average Student Test
Score Change, Relative to U.S. National Average, Grades 3-8, 2009-2018 (SD Units) [Unadjusted]
Notes: [I] Average Student Math & ELA Test Scores [II] Average Student Learning Per Year [III] Average Student Test Score Change.
Epanechnikov kernel used for univariate kernel density estimation. SEDA OLS estimation used above. Results are not sensitive to Empirical Bayes estimation.
I
II
III
56
Figure 11. School-Level Average Student Math & ELA Scores, Student Learning Per Year & Average Student Test
Score Change, Relative to U.S. National Average, Grades 3-8, 2009-2018 (SD Units) [Unadjusted]
Notes: [I] Average Student Math & ELA Test Scores [II] Average Student Learning Per Year [III] Average Student Test Score Change. For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot represents the
weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived
from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
I
II
III
57
APPENDIX A. CBSA and HOLC Details
Appendix A1. CBSA Summary Statistics by 1935-1940 HOLC Grade & Outcome Group
APPENDIX B. MAIN TEXT SUPPLEMENTAL TABLES & FIGURES
Appendix B1. District Finance Outcomes
Table B1.1 Kolmogorov-Smirnov Equality of Distribution Test, Finance Outcomes [Unadjusted]
U.S. Census Bureau Regionb
Overall Midwest Northeast South West
District PPR – Total,
2017-18 ($USD)
D vs. Aa -0.28*** -0.44*** -0.29** -0.17 -0.27
D vs. B -0.33*** -0.28*** -0.30*** -0.36** -0.17
D vs. C -0.14*** -0.10* -0.31*** -0.15 -0.26**
District PPR –
Federal, 2017-18
($USD)
A vs. D 0.51*** 0.48*** 0.59*** 0.54*** 0.61***
B vs. D 0.38*** 0.38*** 0.43*** 0.19 0.29
C vs. D 0.11*** 0.11** 0.22*** 0.19* 0.12
District PPR – State,
2017-18 ($USD)
A vs. D 0.39*** 0.55*** 0.22* 0.23 0.30
B vs. D 0.18*** 0.25*** 0.04 0.24 0.18
C vs. D 0.05 0.16*** <0.01 0.13 0.07
District PPR – Local,
2017-18 ($USD)
D vs. A -0.41*** -0.60*** -0.36*** -0.31 -0.42*
D vs. B -0.38*** -0.40*** -0.20*** -0.4** -0.26
D vs. C -0.16*** -0.22*** -0.12 -0.16 -0.22
Notes: a – Tests the null hypothesis that the first HOLC grade distribution contains smaller values than the second HOLC grade distribution. For
example, D vs. A test the null hypothesis that D distribution contains smaller values than A distribution; b – The values seen here, both for overall and each region, represent the largest difference between distributions in the direction as specified from the “Group” row input. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Table B1.2 Margins – Overall and by Region, Finance Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
District
PPR –
Total,
2017-18
($USD)
A $16,938
[$1,778]
$16,612
[$995]
$23,035
[$1,086]
$12,884
[$1,742]
$15,672
[$591]
B $17,662
[$583]
$16,579
[$782]
$21,573
[$415]
$15,563
[$1,370]
$12,715
[$1,214]
C $15,435
[$399]
$15,849
[$348]
$23,268
[$618]
$12,203
[$385]
$15,455
[$389]
D $14,453
[$472]
$14,413
[$414]
$19,933
[$912]
$11,728
[$271]
$13,730
[$806]
District
PPR –
Federal,
2017-18
($USD)
A $705
[$133]
$547
[$98]
$450
[$43]
$962
[$144]
$765
[$268]
B $942
[$52]
$838
[$106]
$903
[$90]
$1,133
[$113]
$989
[$113]
C $1,430
[$49]
$1,496
[$104]
$1,489
[$141]
$1,409
[$59]
$1,357
[$128]
D $1,136
[$52]
$1,212
[$89]
$1,197
[$81]
$1,126
[$106]
$991
[$108]
62
District
PPR –
State,
2017-18
($USD)
A $4,498
[$571]
$5,411
[$689]
$4,559
[$210]
$3,023
[$218]
$6,511
[$1,068]
B $6,236
[$424]
$6,580
[$510]
$6,931
[$564]
$4,097
[$1,411]
$6,639
[$723]
C $7,272
[$378]
$7,744
[$341]
$11,544
[$728]
$4,752
[$364]
$8,268
[$667]
D $7,514
[$426]
$8,081
[$394]
$9,311
[$1,674]
$5,802
[$203]
$7,522
[$601]
District
PPR –
Local,
2017-18
($USD)
A $11,735
[$1,610]
$10,654
[$1,509]
$18,026
[$1,161]
$8,899
[$1,988]
$8,396
[$1,211]
B $10,484
[$604]
$9,161
[$745]
$13,739
[$633]
$10,333
[$1,785]
$5,088
[$700]
C $6,733
[$296]
$6,608
[$536]
$10,235
[$719]
$6,041
[$430]
$5,830
[$558]
D $5,803
[$300]
$5,120
[$493]
$9,425
[$1,095]
$4,801
[$408]
$5,217
[$667]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by
region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table B1.3 Tukey HSD – Overall and by Region, Finance Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
District
PPR –
Total,
2017-18
($USD)
B vs. A $724
[$771]
-$33
[$963]
-$1,461
[$992]
$2,679
[$1,752]
-$2,957
[$1,433]
C vs. A -$1,504
[$712]
-$764
[$914]
$233
[$954]
-$681
[$1,506]
-$217
[$1,302]
D vs. A -$2,486**
[$800]
-$2,200
[$981]
-$3,102**
[$1,125]
-$1,156
[$1,740]
-$1,942
[$1,426]
C vs. B -$2,228***
[$357]
-$731
[$381]
$1,694***
[$457]
-$3,360***
[$985]
$2,740***
[$676]
D vs. B -$3,210***
[$509]
-$2,166***
[$522]
-$1,640
[$752]
-$3,835**
[$1,315]
$1,015
[$893]
D vs. C -$982*
[$415]
-$1,436***
[$424]
-$3,335***
[$702]
-$475
[$963]
-$1,725**
[$661]
District
PPR –
Federal,
2017-18
($USD)
B vs. A $237**
[$88]
$291
[$168]
$454**
[$151]
$171
[$218]
$223
[$290]
C vs. A $725***
[$82]
$949***
[$160]
$1,040***
[$145]
$447*
[$187]
$592
[$263]
D vs. A $431***
[$92]
$665***
[$171]
$747***
[$171]
$163
[$217]
$225
[$288]
C vs. B $488***
[$41]
$658***
[$67]
$586***
[$70]
$276
[$123]
$369**
[$137]
D vs. B $194***
[$58]
$374***
[$91]
$293*
[$114]
-$8
[$164]
$2
[$180]
D vs. C -$294***
[$48]
-$284***
[$74]
-$292**
[$107]
-$284*
[$120]
-$366**
[$134]
District
PPR –
B vs. A $1,738**
[$586]
$1,169
[$658]
$2,372
[$1,177]
$1,074
[$943]
$128
[$1,546]
63
State,
2017-18
($USD)
C vs. A $2,773***
[$541]
$2,333***
[$625]
$6,985***
[$1,132]
$1,729
[$810]
$1,757
[$1,403]
D vs. A $3,016***
[$607]
$2,670***
[$671]
$4,752***
[$1,336]
$2,779**
[$936]
$1,011
[$1,538]
C vs. B $1,035***
[$271]
$1,164***
[$261]
$4,613***
[$543]
$655
[$530]
$1,629
[$729]
D vs. B $1,278***
[$387]
$1,501***
[$357]
$2,380**
[$893]
$1,705*
[$708]
$883
[$963]
D vs. C $243
[$315]
$337
[$290]
-$2,233**
[$833]
$1,050
[$518]
-$746
[$713]
District
PPR –
Local,
2017-18
($USD)
B vs. A -$1,251
[$691]
-$1,493
[$1,029]
-$4,287***
[$1,202]
$1,433
[$1,803]
-$3,308
[$1,784]
C vs. A -$5,002***
[$639]
-$4,046***
[$976]
-$7,792***
[$1,156]
-$2,858
[$1,550]
-$2,565
[$1,620]
D vs. A -$5,933***
[$717]
-$5,535***
[$1,047]
-$8,601***
[$1,364]
-$4,098
[$1,791]
-$3,178
[$1,776]
C vs. B -$3,751***
[$320]
-$2,553***
[$407]
-$3,505***
[$555]
-$4,291***
[$1,014]
$742
[$842]
D vs. B -$4,681***
[$457]
-$4,041***
[$557]
-$4,314***
[$912]
-$5,532***
[$1,354]
$129
[$1,112]
D vs. C -$930*
[$372]
-$1,489***
[$453]
-$810
[$851]
-$1,241
[$992]
-$613
[$823]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Figure B1.1 HOLC A-D Coefficient Plots, Full vs. Panel Sample, 2017-18, Finance Outcomes (USD 2018)
Notes: [top left] Per-Pupil Total Revenue; [top right] Per-Pupil Federal Revenue; [bottom left] Per-Pupil State Revenue; [bottom right] Per-
Pupil Local Revenue. All values represented in USD 2018. Coefficients (solid dots) are HOLC A-D weighted averages from regressions of
respective finance outcomes on HOLC A-D indicators without a constant term. The 95% confidence intervals (lines) are calculated using Huber-White heteroskedasticity-consistent standard errors. Regressions are run using both the 2017-18 full sample and the 2017-18 data from the panel
sample. The full sample is the original 2017-18 cross-sectional sample and includes all d = 2,135 districts. The panel sample is the time series
sample spanning 1989-90 through 2017-18 school years and includes dp = 1,113 districts.
64
Table B1.4 HOLC A-D Means, SEs and Differences, Full vs. Panel Sample, Finance Outcomes (USD 2018)
Sample Comparisons
Full Panel Diff. p-value
District PPR –
Total, 2017-18
($USD)
A $16,938
[$1,778]
$17,059
[$2,787]
-$120 0.91
B $17,662
[$583]
$17,738
[$906]
-$75 0.82
C $15,435
[$399]
$15,526
[$627]
-$92 0.70
D $14,453
[$472]
$14,269
[$666]
$184 0.63
District PPR –
Federal, 2017-18
($USD)
A $705
[$133]
$695
[$208]
$10 0.9
B $942
[$52]
$932
[$81]
$10 0.74
C $1,430
[$49]
$1,411
[$74]
$18 0.57
D $1,136
[$52]
$1,088
[$89]
$47 0.24
District PPR –
State, 2017-18
($USD)
A $4,498
[$571]
$4,481
[$874]
$17 0.95
B $6,236
[$424]
$6,231
[$655]
$6 0.98
C $7,272
[$378]
$7,281
[$593]
-$9 0.97
D $7,514
[$426]
$7,315
[$536]
$199 0.59
District PPR –
Local, 2017-18
($USD)
A $11,735
[$1,610]
$11,883
[$2,535]
-$147 0.87
B $10,484
[$604]
$10,575
[$937]
-$91 0.79
C $6,733
[$296]
$6,834
[$465]
-$101 0.57
D $5,803
[$300]
$5,866
[$519]
-$63 0.79
Notes: Means are HOLC A-D weighted averages from regressions of respective finance outcomes on HOLC B-D indicators. Standard errors are
Huber-White heteroskedasticity-consistent and are in brackets. Regressions are run using both the 2017-18 full sample and the 2017-18 data from the panel sample. The full sample is the original 2017-18 cross-sectional sample and includes all d = 2,135 districts. The panel sample is the time series sample spanning 1989-90 through 2017-18 school years and includes dp = 1,113 districts.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
65
Table B1.5 Margins & Tukey HSD Over Time, Finance Outcomes (1989-2018) [Unadjusted]
Margins a Tukey HSD b
A B C D D vs. A D vs. B D vs. C
District
PPR –
Total (USD
2018)
1989-90 $5,934 [$506]
$6,079 [$232]
$5,333 [$95]
$4,765 [$145]
-$1,169*** [$313]
-$1,313*** [$187]
-$567*** [$154]
1999-00 $9,010 [$594]
$9,139 [$278]
$8,370 [$115]
$7,909 [$186]
-$1,101** [$418]
-$1,230*** [$257]
-$461 [$213]
2009-10 $13,808 [$1,287]
$14,222 [$469]
$12,963 [$268]
$11,698 [$314]
-$2,110** [$784]
-$2,524*** [$500]
-$1,265** [$415]
2017-18 $17,059 [$1,851]
$17,738 [$602]
$15,526 [$416]
$14,269 [$442]
-$2,789** [$1,013]
-$3,469*** [$661]
-$1,257 [$550]
District
PPR –
Federal (USD
2018)
1989-90 $143 [$25]
$209 [$16]
$378 [$18]
$254 [$17]
$110** [$37]
$45 [$22]
-$124*** [$18]
1999-00 $266 [$56]
$415 [$28]
$703 [$32]
$513 [$23]
$247*** [$64]
$98* [$39]
-$190*** [$33]
2009-10 $844 [$118]
$1,212 [$71]
$1,909 [$73]
$1,412 [$68]
$569*** [$180]
$200 [$115]
-$497*** [$95]
2017-18 $695 [$138]
$932 [$54]
$1,411 [$49]
$1,088 [$59]
$393*** [$111]
$156 [$73]
-$323*** [$60]
District PPR –
State (USD 2018)
1989-90 $1,632 [$285]
$1,776 [$101]
$2,513 [$136]
$2,196 [$90]
$563* [$228]
$419** [$136]
-$317** [$112]
1999-00 $2,699 [$218]
$3,205 [$207]
$4,205 [$159]
$4,055 [$151]
$1,356*** [$357]
$849*** [$220]
-$151 [$182]
2009-10 $3,639 [$239]
$4,682 [$289]
$5,632 [$289]
$5,410 [$220]
$1,771** [$586]
$728 [$374]
-$222 [$310]
2017-18 $4,481 [$581]
$6,231 [$435]
$7,281 [$394]
$7,315 [$356]
$2,834*** [$760]
$1,085 [$496]
$34 [$413]
District
PPR – Local (USD
2018)
1989-90 $4,159 [$605]
$4,093 [$238]
$2,442 [$104]
$2,316 [$167]
-$1,842*** [$350]
-$1,777*** [$209]
-$126 [$172]
1999-00 $6,045 [$782]
$5,519 [$280]
$3,462 [$165]
$3,342 [$184]
-$2,704*** [$447]
-$2,177*** [$276]
-$120 [$228]
2009-10 $9,325 [$1,402]
$8,328 [$438]
$5,422 [$256]
$4,876 [$275]
-$4,449*** [$700]
-$3,452*** [$447]
-$546 [$371]
2017-18 $11,883 [$1,684]
$10,575 [$623]
$6,834 [$309]
$5,866 [$345]
-$6,017*** [$896]
-$4,710*** [$585]
-$968 [$486]
Notes: a – Weighted averages are from a regression of a given finance outcome on HOLC B-D indicators with student enrollment used as
precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator. Means are
provided first, with robust standard errors below in brackets; b – Although all HOLC A-D pairwise comparisons were made using the Tukey
HSD post hoc multiple comparisons procedure, only D vs. A, B and C groups are highlighted above. Average differences are provided first, with standard errors below in brackets. All values are measured in USD 2018.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
66
Appendix B2. School Diversity Outcomes
Table B2.1 Kolmogorov-Smirnov Equality of Distribution Test, Diversity Outcomes [Unadjusted]
U.S. Census Bureau Region b
Overall Midwest Northeast South West
% Black A vs. D a 0.26*** 0.34*** 0.32*** 0.43*** 0.17**
B vs. D 0.16*** 0.21*** 0.25*** 0.29*** 0.12***
C vs. D 0.13*** 0.13*** 0.2*** 0.17*** 0.03
% White D vs. A -0.54*** -0.54*** -0.58*** -0.44*** -0.65***
D vs. B -0.34*** -0.34*** -0.36*** -0.23*** -0.47***
D vs. C -0.16*** -0.17*** -0.19*** -0.10*** -0.19***
% Non-White A vs. D 0.54*** 0.54*** 0.58*** 0.44*** 0.65***
B vs. D 0.34*** 0.34*** 0.36*** 0.23*** 0.47***
C vs. D 0.16*** 0.17*** 0.19*** 0.10*** 0.19***
Simpson's Diversity
(1-D)
D vs. A -0.18*** -0.26*** -0.18*** -0.22*** -0.41***
D vs. B -0.12*** -0.17*** -0.10*** -0.14*** -0.31***
D vs. C -0.09*** -0.12*** -0.11*** -0.08** -0.17***
Notes: a – Tests the null hypothesis that the first HOLC grade distribution contains smaller values than the second HOLC grade distribution. For
example, D vs. A test the null hypothesis that D distribution contains smaller values than A distribution; b – The values seen here, both for overall and each region, represent the largest difference between distributions in the direction as specified from the “Group” row input.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Table B2.2 Margins – Overall and by Region, Diversity Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
% Black A
0.22
[0.013]
0.24
[0.021]
0.24
[0.025]
0.22
[0.034]
0.08
[0.015]
B 0.26
[0.008]
0.32
[0.016]
0.27
[0.013]
0.32
[0.022]
0.08
[0.008]
C 0.27
[0.005]
0.36
[0.011]
0.26
[0.008]
0.38
[0.018]
0.09
[0.004]
D 0.36
[0.007]
0.45
[0.014]
0.35
[0.010]
0.52
[0.018]
0.09
[0.005]
% White A
0.46
[0.015]
0.54
[0.024]
0.42
[0.025]
0.42
[0.039]
0.43
[0.035]
B 0.33
[0.008]
0.39
[0.016]
0.31
[0.012]
0.23
[0.018]
0.36
[0.021]
C 0.21
[0.005]
0.26
[0.009]
0.20
[0.007]
0.15
[0.011]
0.17
[0.009]
67
D 0.13
[0.004]
0.18
[0.009]
0.13
[0.006]
0.12
[0.009]
0.10
[0.010]
% Non-
White A 0.54
[0.015]
0.46
[0.024]
0.58
[0.025]
0.58
[0.039]
0.57
[0.035]
B 0.67
[0.008]
0.61
[0.016]
0.69
[0.012]
0.77
[0.018]
0.64
[0.021]
C 0.79
[0.005]
0.74
[0.009]
0.80
[0.007]
0.85
[0.011]
0.83
[0.009]
D 0.87
[0.004]
0.82
[0.009]
0.87
[0.006]
0.88
[0.009]
0.90
[0.010]
Simpson's
Diversity
(1-D)
A 0.49
[0.011]
0.48
[0.019]
0.54
[0.017]
0.40
[0.025]
0.55
[0.017]
B 0.48
[0.006]
0.47
[0.012]
0.51
[0.009]
0.40
[0.015]
0.49
[0.017]
C 0.45
[0.004]
0.42
[0.008]
0.51
[0.005]
0.34
[0.012]
0.42
[0.010]
D 0.41
[0.005]
0.38
[0.010]
0.49
[0.006]
0.34
[0.012]
0.33
[0.012]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by
region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table B2.3 Tukey HSD – Overall and by Region, Diversity Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
% Black B vs. A 0.05***
[0.015]
0.08**
[0.028]
0.03
[0.023]
0.10**
[0.036]
0.01
[0.016]
C vs. A 0.06***
[0.014]
0.12***
[0.026]
0.02
[0.022]
0.16***
[0.034]
0.01
[0.015]
D vs. A 0.14***
[0.014]
0.21***
[0.027]
0.11***
[0.022]
0.30***
[0.034]
0.01
[0.015]
C vs. B 0.01
[0.008]
0.04*
[0.017]
-0.01
[0.012]
0.06*
[0.024]
<0.01
[0.008]
D vs. B 0.09***
[0.009]
0.13***
[0.018]
0.08***
[0.013]
0.20***
[0.024]
0.01
[0.009]
D vs. C 0.08***
[0.007]
0.09***
[0.015]
0.09***
[0.011]
0.14***
[0.021]
0.01
[0.007]
% White B vs. A -0.13***
[0.012]
-0.15***
[0.022]
-0.11***
[0.019]
-0.19***
[0.024]
-0.07*
[0.029]
C vs. A -0.25***
[0.011]
-0.27***
[0.020]
-0.23***
[0.018]
-0.27***
[0.022]
-0.27***
[0.027]
D vs. A -0.33***
[0.011]
-0.36***
[0.021]
-0.30***
[0.019]
-0.30***
[0.022]
-0.34***
[0.028]
C vs. B -0.12***
[0.007]
-0.12***
[0.013]
-0.12***
[0.010]
-0.08***
[0.016]
-0.20***
[0.015]
D vs. B -0.20***
[0.007]
-0.21***
[0.014]
-0.19***
[0.011]
-0.12***
[0.016]
-0.27***
[0.016]
D vs. C -0.07***
[0.006]
-0.09***
[0.011]
-0.07***
[0.009]
-0.04**
[0.014]
-0.07***
[0.012]
68
% Non-
White
B vs. A 0.13***
[0.012]
0.15***
[0.022]
0.11***
[0.019]
0.19***
[0.024]
0.07*
[0.029]
C vs. A 0.25***
[0.011]
0.27***
[0.020]
0.23***
[0.018]
0.27***
[0.022]
0.27***
[0.027]
D vs. A 0.33***
[0.011]
0.36***
[0.021]
0.30***
[0.019]
0.30***
[0.022]
0.34***
[0.028]
C vs. B 0.12***
[0.007]
0.12***
[0.013]
0.12***
[0.010]
0.08***
[0.016]
0.20***
[0.015]
D vs. B 0.20***
[0.007]
0.21***
[0.014]
0.19***
[0.011]
0.12***
[0.016]
0.27***
[0.016]
D vs. C 0.07***
[0.006]
0.09***
[0.011]
0.07***
[0.009]
0.04**
[0.014]
0.07***
[0.012]
Simpson's
Diversity
(1-D)
B vs. A -0.01
[0.010]
-0.01
[0.020]
-0.02
[0.015]
<0.01
[0.024]
-0.06
[0.029]
C vs. A -0.04***
[0.01]
-0.06***
[0.018]
-0.03
[0.014]
-0.06*
[0.023]
-0.13***
[0.027]
D vs. A -0.08***
[0.010]
-0.10***
[0.019]
-0.05***
[0.014]
-0.06*
[0.023]
-0.22***
[0.028]
C vs. B -0.03***
[0.006]
-0.05***
[0.012]
<0.01
[0.008]
-0.05***
[0.016]
-0.07***
[0.015]
D vs. B -0.07***
[0.006]
-0.09***
[0.013]
-0.02**
[0.008]
-0.05***
[0.016]
-0.16***
[0.016]
D vs. C -0.04***
[0.005]
-0.05***
[0.010]
-0.02**
[0.007]
<0.01
[0.014]
-0.09***
[0.012]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Table B2.4 HOLC A-D Means, SEs and Differences, Full vs. Panel Sample, Diversity Outcomes
Sample Comparisons
Full Panel Diff. p-value
% Black A 0.22
[0.013]
0.20
[0.015]
0.018 0.047**
B 0.26
[0.008]
0.23
[0.009]
0.036 <0.001***
C 0.27
[0.005]
0.23
[0.007]
0.037 <0.001***
D 0.36
[0.007]
0.30
[0.01]
0.056 <0.001***
% White A 0.46
[0.015]
0.48
[0.020]
-0.026 0.004***
B 0.33
[0.008]
0.37
[0.011]
-0.042 <0.001***
C 0.21
[0.005]
0.25
[0.008]
-0.041 <0.001***
D 0.13
[0.004]
0.15
[0.008]
-0.023 <0.001***
% Non-White A 0.54
[0.015]
0.52
[0.020]
0.026 0.004***
69
B 0.67
[0.008]
0.63
[0.011]
0.042 <0.001***
C 0.79
[0.005]
0.75
[0.008]
0.041 <0.001***
D 0.87
[0.004]
0.85
[0.008]
0.023 <0.001***
Simpson's
Diversity (1-D)
A 0.49
[0.011]
0.50
[0.014]
-0.006 0.329
B 0.480
[0.006]
0.50
[0.008]
-0.018 <0.001***
C 0.45
[0.004]
0.45
[0.007]
-0.005 0.246
D 0.41
[0.005]
0.42
[0.009]
-0.007 0.295
Notes: Means are HOLC A-D weighted averages from regressions of respective diversity outcomes on HOLC B-D indicators. Standard errors
are Huber-White heteroskedasticity-consistent and are in brackets. Regressions are run using both the 2017-18 full sample and the 2017-18 data
from the panel sample. The full sample is the original 2017-18 cross-sectional sample and includes all n = 10,010 schools. The panel sample is the time series sample spanning 1988-89 through 2018-19 school years and includes np = 4,693 schools.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Table B2.5 Margins & Tukey HSD Over Time, Diversity Outcomes (1988-2018) [Unadjusted]
Margins a Tukey HSD b
A B C D
D vs. A D vs. B D vs. C
% Black 1988-89 0.22 [0.018]
0.25 [0.011]
0.28 [0.008]
0.36 [0.012]
0.13*** [0.019]
0.11*** [0.013]
0.08*** [0.011]
1998-99 0.25 [0.018]
0.27 [0.011]
0.28 [0.008]
0.35 [0.011]
0.10*** [0.019]
0.09*** [0.013]
0.07*** [0.011]
2008-09 0.24 [0.016]
0.26 [0.01]
0.27 [0.008]
0.33 [0.011]
0.09*** [0.018]
0.07*** [0.013]
0.06*** [0.011]
2018-19 0.20 [0.014]
0.23 [0.008]
0.23 [0.007]
0.30 [0.009]
0.10*** [0.016]
0.07*** [0.011]
0.07*** [0.010]
% White 1988-89 0.63 [0.021]
0.54 [0.013]
0.41 [0.009]
0.26 [0.010]
-0.37*** [0.020]
-0.28*** [0.014]
-0.15*** [0.012]
1998-99 0.57 [0.021]
0.47 [0.012]
0.33 [0.009]
0.20 [0.009]
-0.37*** [0.019]
-0.26*** [0.013]
-0.13*** [0.011]
2008-09 0.53 [0.019]
0.41 [0.011]
0.28 [0.008]
0.16 [0.008]
-0.37*** [0.017]
-0.25*** [0.012]
-0.11*** [0.011]
2018-19 0.48 [0.018]
0.37 [0.010]
0.25 [0.007]
0.15 [0.007]
-0.33*** [0.015]
-0.21*** [0.011]
-0.09*** [0.009]
% Non-
White
1988-89 0.37 [0.021]
0.46 [0.013]
0.59 [0.009]
0.74 [0.010]
0.37*** [0.020]
0.28*** [0.014]
0.15*** [0.012]
1998-99 0.43 [0.021]
0.53 [0.012]
0.67 [0.009]
0.80 [0.009]
0.37*** [0.019]
0.26*** [0.013]
0.13*** [0.011]
2008-09 0.47 [0.019]
0.59 [0.011]
0.72 [0.008]
0.84 [0.008]
0.37*** [0.017]
0.25*** [0.012]
0.11*** [0.011]
2018-19 0.52 [0.018]
0.63 [0.010]
0.75 [0.007]
0.85 [0.007]
0.33*** [0.015]
0.21*** [0.011]
0.09*** [0.009]
1988-89 0.34 [0.014]
0.36 [0.008]
0.37 [0.006]
0.36 [0.008]
0.02 [0.014]
<0.01 [0.009]
-0.01 [0.008]
70
Simpson's
Diversity (1-D)
1998-99 0.38 [0.014]
0.38 [0.008]
0.37 [0.006]
0.35 [0.008]
-0.02 [0.014]
-0.03*** [0.009]
-0.02 [0.008]
2008-09 0.43 [0.013]
0.42 [0.008]
0.39 [0.006]
0.35 [0.008]
-0.07*** [0.013]
-0.06*** [0.009]
-0.03*** [0.008]
2018-19 0.50 [0.012]
0.50 [0.007]
0.45 [0.006]
0.42 [0.008]
-0.08*** [0.013]
-0.08*** [0.009]
-0.04*** [0.008]
Notes: a – Weighted averages are from a regression of a given diversity outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator. Means are
provided first, with robust standard errors below in brackets; b – Although all HOLC A-D pairwise comparisons were made using the Tukey
HSD post hoc multiple comparisons procedure, only D vs. A, B and C groups are highlighted above. Average differences are provided first, with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Figure B2.1 HOLC A-D Coefficient Plots, Full vs. Panel Sample, Diversity Outcomes
full and panel samples here. Coefficients (solid dots) are HOLC A-D weighted averages from regressions of respective diversity outcomes on HOLC A-D indicators without a constant term. The 95% confidence intervals (lines) are calculated using Huber-White heteroskedasticity-
consistent standard errors. Regressions are run using both the 2017-18 full sample and the 2017-18 data from the panel sample. The full sample
is the original 2017-18 cross-sectional sample and includes all n = 10,010 schools. The panel sample is the time series sample spanning 1988-89 through 2018-19 school years and includes np = 4,693 schools.
Figure B2.2 HOLC A-D Averages Over Time, Exposure Index (1988-2018) [Unadjusted]
Notes: [top left] White-Asian Exposure Index; [top right] White-Black Exposure Index; [bottom left] White-Hispanic Exposure Index; [bottom right] White-Non-White Exposure Index. Exposure index calculations for each race pair factor in all other student race groups.
71
Table B2.6 HOLC A-D Averages Over Time, Full and Panel Samples, Exposure Index (1988-2018) [Unadjusted]
HOLC Security Rating
Overall A B C D
White-Asian a 1988-89 4.4% 4.1% 4.6% 4.4% 4.2%
1998-99 5.4% 5.0% 5.7% 5.5% 4.8%
2008-09 6.2% 5.6% 6.4% 6.5% 5.3%
2018-19 6.8% 6.1% 7.1% 7.1% 6.0%
2018-19 b 7.5% 6.5% 7.6% 8.0% 7.3%
White-Black a 1988-89 15.9% 13.0% 14.7% 14.9% 22.2%
1998-99 16.3% 13.8% 15.3% 15.3% 23.1%
2008-09 16.4% 13.9% 14.9% 16.3% 22.8%
2018-19 13.8% 11.5% 12.5% 13.8% 19.5%
2018-19 b 15.0% 11.5% 13.3% 14.9% 20.1%
White-Hispanic a 1988-89 8.9% 5.1% 6.4% 9.8% 13.3%
1998-99 11.3% 6.5% 8.8% 12.6% 16.4%
2008-09 14.5% 9.2% 11.7% 16.2% 21.4%
2018-19 18.8% 12.3% 15.5% 21.0% 27.3%
2018-19 b 19.9% 12.8% 16.3% 22.0% 25.8%
White-Non-White a
1988-89 29.8% 22.5% 26.3% 29.8% 40.5%
1998-99 33.7% 25.8% 30.5% 34.1% 45.5%
2008-09 38.4% 29.9% 34.3% 40.3% 51.1%
2018-19 45.9% 36.0% 41.8% 48.5% 59.3%
2018-19 b 48.7% 36.9% 43.6% 51.1% 59.3%
Notes: a – Signifies the student race pair used in the exposure index calculation. For example, White-Black provides the White-Black exposure
index, while factoring in all other student race groups. Exposure index calculations do vary if all other race groups apart from the relevant race
pair are ignored; b – Exposure index values derived using the full 2018-19 cross-sectional data set. All other within-group (i.e., White-Black) values across time derived from the 1988-89 through 2018-19 panel data set.
72
Appendix B3. School Performance Outcomes
Table B3.1 Kolmogorov-Smirnov Equality of Distribution Test, Student Performance Outcomes [Unadjusted]
U.S. Census Bureau Regionb
Overall Midwest Northeast South West
Avg. Student Math &
ELA Scores
D vs. Aa -0.47*** -0.52*** -0.44*** -0.49*** -0.67***
D vs. B -0.25*** -0.25*** -0.22*** -0.24*** -0.45***
D vs. C -0.12*** -0.16*** -0.11*** -0.12*** -0.18***
Student Learning
(Annual)
D vs. A -0.07 -0.06 -0.11 -0.07 -0.12
D vs. B -0.05* -0.02 -0.09** -0.07 -0.12*
D vs. C -0.03 -0.02 -0.05 -0.08 -0.03
Avg. Student Test
Score Change
D vs. A -0.08** -0.15** -0.14** -0.1 -0.1
D vs. B -0.05** -0.07 -0.05 -0.08 -0.13**
D vs. C -0.04** -0.07** -0.06** -0.06 -0.03
Notes: a – Tests the null hypothesis that the first HOLC grade distribution contains smaller values than the second HOLC grade distribution. For example, D vs. A test the null hypothesis that D distribution contains smaller values than A distribution; b – The values seen here, both for
overall and each region, represent the largest difference between distributions in the direction as specified from the “Group” row input.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
Table B3.2 Margins – Overall and by Region, Student Performance Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
Avg.
Student
Math &
ELA
Scores
A 0.235
[0.030]
0.154
[0.051]
0.273
[0.055]
0.215
[0.062]
0.347
[0.074]
B -0.030
[0.016]
-0.215
[0.028]
0.056
[0.026]
-0.106
[0.036]
0.102
[0.040]
C -0.205
[0.011]
-0.301
[0.018]
-0.087
[0.017]
-0.260
[0.025]
-0.277
[0.027]
D -0.296
[0.014]
-0.409
[0.030]
-0.145
[0.022]
-0.342
[0.029]
-0.453
[0.024]
Student
Learning
(Annual)
A 0.004
[0.004]
0.010
[0.007]
0.007
[0.009]
-0.016
[0.007]
0.015
[0.010]
B 0.010
[0.002]
0.013
[0.004]
0.007
[0.004]
-0.007
[0.006]
0.035
[0.006]
C 0.011
[0.002]
0.022
[0.002]
0.002
[0.003]
0.004
[0.005]
0.017
[0.004]
D 0.007
[0.002]
0.017
[0.003]
0.002
[0.004]
-0.012
[0.004]
0.021
[0.005]
Avg.
Student
Test
Score
Change
A 0.003
[0.002]
-0.002
[0.003]
0.006
[0.004]
-0.001
[0.004]
0.014
[0.006]
B 0.002
[0.001]
-0.006
[0.002]
0.001
[0.002]
-0.001
[0.003]
0.019
[0.003]
C 0.002
[0.001]
-0.002
[0.001]
0.003
[0.002]
-0.005
[0.002]
0.011
[0.002]
73
D <0.001
[0.001]
-0.008
[0.002]
0.002
[0.002]
-0.007
[0.003]
0.014
[0.002]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table B3.3 Tukey HSD – Overall and by Region, Student Performance Outcomes [Unadjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
Avg.
Student
Math &
ELA Scores
B vs. A -0.265***
[0.030]
-0.368***
[0.051]
-0.217***
[0.055]
-0.322***
[0.059]
-0.245***
[0.072]
C vs. A -0.440***
[0.028]
-0.454***
[0.046]
-0.360***
[0.052]
-0.476***
[0.055]
-0.624***
[0.068]
D vs. A -0.532***
[0.029]
-0.562***
[0.048]
-0.419***
[0.053]
-0.557***
[0.056]
-0.800***
[0.069]
C vs. B -0.175***
[0.017]
-0.086**
[0.030]
-0.143***
[0.028]
-0.154***
[0.039]
-0.379***
[0.039]
D vs. B -0.267***
[0.018]
-0.194***
[0.033]
-0.201***
[0.030]
-0.236***
[0.040]
-0.555***
[0.042]
D vs. C -0.091***
[0.015]
-0.108***
[0.026]
-0.059*
[0.025]
-0.082*
[0.034]
-0.176***
[0.033]
Student
Learning
(Annual)
B vs. A 0.006
[0.005]
0.003
[0.008]
<0.001
[0.009]
0.009
[0.010]
0.020
[0.012]
C vs. A 0.007
[0.004]
0.012
[0.007]
-0.005
[0.009]
0.020
[0.010]
0.001
[0.011]
D vs. A 0.003
[0.005]
0.007
[0.007]
-0.005
[0.009]
0.004
[0.010]
0.005
[0.011]
C vs. B 0.001
[0.003]
0.009
[0.004]
-0.005
[0.005]
0.011
[0.007]
-0.019**
[0.007]
D vs. B -0.003
[0.003]
0.004
[0.005]
-0.005
[0.005]
-0.005
[0.007]
-0.015
[0.007]
D vs. C -0.004
[0.002]
-0.005
[0.004]
<0.001
[0.004]
-0.016**
[0.006]
0.004
[0.006]
Avg.
Student
Test Score
Change
B vs. A -0.001
[0.003]
-0.004
[0.004]
-0.005
[0.005]
<0.001
[0.006]
0.005
[0.006]
C vs. A -0.002
[0.002]
-0.001
[0.004]
-0.003
[0.005]
-0.004
[0.005]
-0.003
[0.005]
D vs. A -0.003
[0.003]
-0.006
[0.004]
-0.004
[0.005]
-0.006
[0.005]
<0.001
[0.006]
C vs. B <0.001
[0.001]
0.004
[0.003]
0.001
[0.003]
-0.004
[0.004]
-0.008*
[0.003]
D vs. B -0.002
[0.002]
-0.002
[0.003]
0.001
[0.003]
-0.005
[0.004]
-0.005
[0.003]
D vs. C -0.002
[0.001]
-0.005*
[0.002]
<0.001
[0.002]
-0.002
[0.003]
0.003
[0.003]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are
provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
74
APPENDIX C. CBSA ADJUSTED RESULTS
Appendix C1. District Finance with CBSA Adjustments
Figure C1.1 District Per-Pupil Revenue – Total, Overall & by Region, 2017-18 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Figure C1.2 District Per-Pupil Revenue – Federal, Overall & by Region, 2017-18 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
75
Figure C1.3 District Per-Pupil Revenue – State, Overall & by Region, 2017-18 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Figure C1.8 District Per-Pupil Revenue – Local, Overall & by Region, 2017-18 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table C1.M Margins – Overall and by Region, Finance Outcomes [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
District
PPR –
Total,
2017-18
($USD)
A $225
[$482]
$1,237
[$544]
-$339
[$602]
$425
[$1,323]
-$400
[$466]
B -$315
[$118]
$1,044
[$344]
-$1,089
[$162]
-$152
[$115]
-$1,189
[$171]
C $167
[$63]
-$167
[$162]
$867
[$128]
$86
[$45]
$202
[$40]
D -$224
[$184]
-$617
[$143]
$1,322
[$669]
-$892
[$439]
-$392
[$98]
District
PPR –
A -$599
[$93]
-$724
[$72]
-$755
[$124]
-$529
[$119]
-$341
[$273]
76
Federal,
2017-18
($USD)
B -$318
[$40]
-$447
[$80]
-$301
[$66]
-$156
[$89]
-$325
[$89]
C $97
[$12]
$144
[$13]
$206
[$44]
$47
[$17]
$62
[$17]
D -$75
[$33]
-$130
[$60]
$85
[$47]
-$116
[$76]
-$109
[$70]
District
PPR –
State,
2017-18
($USD)
A -$1,816
[$400]
-$1,187
[$270]
-$4,003
[$642]
-$833
[$103]
-$911
[$977]
B -$1,547
[$201]
-$709
[$167]
-$2,857
[$405]
-$729
[$292]
-$906
[$398]
C $284
[$60]
$112
[$81]
$1,625
[$160]
$43
[$53]
$8
[$117]
D $466
[$173]
$477
[$168]
-$237
[$517]
$598
[$356]
$1,008
[$379]
District
PPR –
Local,
2017-18
($USD)
A $2,640
[$793]
$3,147
[$646]
$4,418
[$775]
$1,787
[$1,457]
$852
[$789]
B $1,550
[$231]
$2,200
[$451]
$2,068
[$370]
$732
[$434]
$42
[$373]
C -$214
[$55]
-$423
[$93]
-$964
[$139]
-$4
[$69]
$131
[$113]
D -$615
[$271]
-$964
[$119]
$1,475
[$920]
-$1,374
[$740]
-$1,292
[$381]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by
region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table C1.T Tukey HSD – Overall and by Region, Finance Outcomes [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
District
PPR –
Total,
2017-18
($USD)
B vs. A -$540*
[$226]
-$193
[$371]
-$750
[$481]
-$578
[$420]
-$790***
[$232]
C vs. A -$58
[$209]
-$1,404***
[$352]
$1,206**
[$463]
-$339
[$361]
$602**
[$211]
D vs. A -$449
[$233]
-$1,854***
[$377]
$1,662***
[$524]
-$1,317***
[$417]
$7
[$231]
C vs. B $482***
[$104]
-$1,212***
[$147]
$1,956***
[$221]
$239
[$236]
$1,391***
[$110]
D vs. B $91
[$147]
-$1,661***
[$200]
$2,412***
[$331]
-$739*
[$315]
$797***
[$145]
D vs. C -$391***
[$119]
-$450**
[$163]
$456
[$304]
-$978***
[$231]
-$594***
[$107]
District
PPR –
Federal,
2017-18
($USD)
B vs. A $281***
[$37]
$277***
[$64]
$454***
[$68]
$373***
[$88]
$16
[$94]
C vs. A $696***
[$34]
$868***
[$61]
$960***
[$66]
$576***
[$76]
$403***
[$85]
D vs. A $524***
[$38]
$594***
[$66]
$839***
[$75]
$413***
[$88]
$232*
[$93]
C vs. B $415***
[$17]
$590***
[$25]
$506***
[$31]
$203***
[$50]
$387***
[$44]
77
D vs. B $243***
[$24]
$317***
[$35]
$385***
[$47]
$40
[$66]
$216***
[$58]
D vs. C -$172***
[$19]
-$274***
[$28]
-$121**
[$43]
-$163***
[$48]
-$171***
[$43]
District
PPR –
State,
2017-18
($USD)
B vs. A $269
[$197]
$478*
[$208]
$1,146**
[$417]
$104
[$324]
$4
[$410]
C vs. A $2,100***
[$183]
$1,298***
[$197]
$5,627***
[$401]
$876**
[$278]
$919*
[$373]
D vs. A $2,282***
[$204]
$1,663***
[$212]
$3,766***
[$455]
$1,431***
[$322]
$1,919***
[$408]
C vs. B $1,831***
[$91]
$820***
[$82]
$4,481***
[$191]
$772***
[$182]
$915***
[$194]
D vs. B $2,013***
[$128]
$1,185***
[$112]
$2,620***
[$287]
$1,327***
[$243]
$1,914***
[$255]
D vs. C $182
[$104]
$365***
[$91]
-$1,861***
[$264]
$555**
[$178]
$1,000***
[$189]
District
PPR –
Local,
2017-18
($USD)
B vs. A -$1,090***
[$270]
-$948*
[$373]
-$2,350***
[$583]
-$1,055
[$587]
-$810
[$447]
C vs. A -$2,855***
[$250]
-$3,570***
[$354]
-$5,381***
[$561]
-$1,791***
[$505]
-$720
[$406]
D vs. A -$3,255***
[$279]
-$4,111***
[$380]
-$2,943***
[$636]
-$3,161***
[$583]
-$2,143***
[$445]
C vs. B -$1,764***
[$125]
-$2,623***
[$148]
-$3,032***
[$268]
-$736
[$330]
$89
[$211]
D vs. B -$2,165***
[$176]
-$3,164***
[$202]
-$593
[$401]
-$2,106***
[$440]
-$1,334***
[$278]
D vs. C -$400**
[$142]
-$541***
[$164]
$2,438***
[$368]
-$1,370***
[$322]
-$1,423***
[$206]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
78
Appendix C2. School-Level Diversity with CBSA Adjustments
Figure C2.1 School Student Body % Black, Overall & by Region, 2018-2019 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean outcome, weighted by student enrollment. Each
black dot represents the weight average outcome by HOLC grade. Each gray band represents the 95% confidence interval of the weighted average outcome.
Figure C2.2 School Student Body % White, Overall & by Region, 2018-2019 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
79
Figure C2.3 School Student Body % Non-White, Overall & by Region, 2018-2019 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Figure C2.4 School Student Diversity, Overall & by Region, 2018-2019 [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table C2.M Margins – Overall and by Region, Diversity Outcomes [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
% Black A -0.07
[0.008]
-0.12
[0.012]
-0.04
[0.012]
-0.08
[0.022]
<0.01
[0.006]
B -0.04
[0.002]
-0.04
[0.004]
-0.04
[0.001]
-0.07
[0.008]
-0.01
[0.003]
C -0.01
[0.001]
-0.02
[0.001]
-0.02
[0.001]
-0.01
[0.004]
<0.01
[<0.001]
D 0.06
[0.001]
0.09
[0.003]
0.05
[0.001]
0.07
[0.004]
0.01
[0.001]
% White A 0.19
[0.008]
0.18
[0.014]
0.17
[0.013]
0.21
[0.022]
0.25
[0.015]
80
B 0.09
[0.002]
0.09
[0.006]
0.09
[0.003]
0.06
[0.007]
0.12
[0.005]
C -0.02
[0.001]
-0.01
[0.001]
-0.02
[0.001]
-0.02
[0.003]
-0.02
[0.001]
D -0.07
[0.001]
-0.09
[0.002]
-0.06
[0.002]
-0.06
[0.003]
-0.07
[0.003]
% Non-
White
A -0.19
[0.008]
-0.18
[0.014]
-0.17
[0.013]
-0.21
[0.022]
-0.25
[0.015]
B -0.09
[0.002]
-0.09
[0.006]
-0.09
[0.003]
-0.06
[0.007]
-0.12
[0.005]
C 0.02
[0.001]
0.01
[0.001]
0.02
[0.001]
0.02
[0.003]
0.02
[0.001]
D 0.07
[0.001]
0.09
[0.002]
0.06
[0.002]
0.06
[0.003]
0.07
[0.003]
Simpson's
Diversity
(1-D)
A 0.04
[0.006]
0.02
[0.009]
0.03
[0.007]
0.03
[0.015]
0.13
[0.017]
B 0.02
[0.002]
0.01
[0.004]
0.01
[0.002]
0.05
[0.006]
0.04
[0.006]
C 0.01
[0.001]
0.01
[0.001]
<0.01
[0.001]
-0.01
[0.003]
0.02
[0.001]
D -0.03
[0.001]
-0.03
[0.001]
-0.01
[0.001]
-0.03
[0.003]
-0.06
[0.003]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by
region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table C2.T Tukey HSD – Overall and by Region, Diversity Outcomes [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
% Black B vs. A
0.03***
[0.003]
0.07***
[0.006]
<0.01
[0.004]
0.010
[0.01]
-0.01
[0.003]
C vs. A 0.05***
[0.003]
0.10***
[0.005]
0.03***
[0.003]
0.07***
[0.010]
<0.01
[0.003]
D vs. A 0.12***
[0.003]
0.20***
[0.005]
0.10***
[0.004]
0.15***
[0.010]
0.01***
[0.003]
C vs. B 0.03***
[0.002]
0.02***
[0.003]
0.02***
[0.002]
0.06***
[0.007]
0.01***
[0.002]
D vs. B 0.10***
[0.002]
0.13***
[0.004]
0.09***
[0.002]
0.14***
[0.007]
0.02***
[0.002]
D vs. C 0.07***
[0.001]
0.10***
[0.003]
0.07***
[0.002]
0.09***
[0.006]
0.01***
[0.001]
% White B vs. A
-0.10***
[0.003]
-0.09***
[0.006]
-0.08***
[0.005]
-0.15***
[0.009]
-0.12***
[0.007]
C vs. A -0.21***
[0.003]
-0.19***
[0.006]
-0.19***
[0.004]
-0.23***
[0.008]
-0.27***
[0.006]
D vs. A -0.26***
[0.003]
-0.27***
[0.006]
-0.23***
[0.004]
-0.27***
[0.008]
-0.32***
[0.006]
C vs. B -0.10***
[0.002]
-0.09***
[0.004]
-0.11***
[0.002]
-0.08***
[0.006]
-0.14***
[0.003]
81
D vs. B -0.16***
[0.002]
-0.17***
[0.004]
-0.15***
[0.003]
-0.12***
[0.006]
-0.20***
[0.004]
D vs. C -0.05***
[0.002]
-0.08***
[0.003]
-0.04***
[0.002]
-0.04***
[0.005]
-0.05***
[0.003]
% Non-
White B vs. A 0.10***
[0.003]
0.09***
[0.006]
0.08***
[0.005]
0.15***
[0.009]
0.12***
[0.007]
C vs. A 0.21***
[0.003]
0.19***
[0.006]
0.19***
[0.004]
0.23***
[0.008]
0.27***
[0.006]
D vs. A 0.26***
[0.003]
0.27***
[0.006]
0.23***
[0.004]
0.27***
[0.008]
0.32***
[0.006]
C vs. B 0.10***
[0.002]
0.09***
[0.004]
0.11***
[0.002]
0.08***
[0.006]
0.14***
[0.003]
D vs. B 0.16***
[0.002]
0.17***
[0.004]
0.15***
[0.003]
0.12***
[0.006]
0.20***
[0.004]
D vs. C 0.05***
[0.002]
0.08***
[0.003]
0.04***
[0.002]
0.04***
[0.005]
0.05***
[0.003]
Simpson's
Diversity
(1-D)
B vs. A -0.02***
[0.002]
-0.01*
[0.004]
-0.02***
[0.003]
0.02
[0.008]
-0.10***
[0.007]
C vs. A -0.04***
[0.002]
-0.02***
[0.004]
-0.03***
[0.003]
-0.04***
[0.007]
-0.11***
[0.006]
D vs. A -0.07***
[0.002]
-0.05***
[0.004]
-0.04***
[0.003]
-0.06***
[0.007]
-0.19***
[0.006]
C vs. B -0.02***
[0.001]
-0.01*
[0.002]
-0.01***
[0.002]
-0.05***
[0.005]
-0.02***
[0.003]
D vs. B -0.05***
[0.001]
-0.04***
[0.003]
-0.02***
[0.002]
-0.07***
[0.005]
-0.10***
[0.004]
D vs. C -0.03***
[0.001]
-0.04***
[0.002]
-0.01***
[0.001]
-0.02***
[0.004]
-0.08***
[0.003]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level
82
Appendix C3. School-Level Student Performance with CBSA Adjustments
Figure C3.1 School-Level Average Student Math & ELA Scores in Grade Levels Relative to U.S. National
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted
averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Figure C3.2 School-Level Student Learning Per Year Relative to U.S. National Average, Grades 3-8, 2009-2018
(SD Units) [Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
83
Figure C3.3 School-Level Average Student Test Score Change in Grade Levels, Grades 3-8, 2009-2018 (SD Units)
[Adjusted]
Notes: For each plot, both overall and by region, the horizontal line represents the grand mean, weighted by student enrollment. Each black dot
represents the weighted average outcome by HOLC A-D grade. The gray bands represent 95% confidence intervals. HOLC A-D weighted averages are derived from a regression of the above outcome on HOLC B-D indicators with student enrollment used as precision weights. Standard errors are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
Table C3.M Margins – Overall and by Region, Student Performance [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
Avg.
Student
Math &
ELA
Scores
A 0.426
[0.014]
0.422
[0.025]
0.403
[0.023]
0.376
[0.031]
0.571
[0.031]
B 0.155
[0.005]
0.101
[0.011]
0.161
[0.004]
0.118
[0.019]
0.262
[0.011]
C -0.035
[0.001]
-0.022
[0.003]
-0.044
[0.001]
-0.020
[0.006]
-0.044
[0.004]
D -0.129
[0.002]
-0.116
[0.004]
-0.099
[0.003]
-0.156
[0.007]
-0.190
[0.007]
Student
Learning
(Annual)
A 0.002
[0.002]
0.004
[0.004]
0.011
[0.004]
-0.007
[0.005]
-0.010
[0.007]
B 0.005
[0.001]
<0.001
[0.001]
0.008
[0.001]
0.002
[0.002]
0.006
[0.001]
C <0.001
[<0.001]
0.002
[<0.001]
-0.002
[<0.001]
0.005
[0.001]
-0.002
[<0.001]
D -0.004
[<0.001]
-0.004
[0.001]
-0.004
[<0.001]
-0.007
[0.001]
-0.002
[0.001]
Avg.
Student
Test
Score
Change
A 0.006
[0.001]
0.006
[0.001]
0.010
[0.002]
0.004
[0.003]
<0.001
[0.003]
B 0.003
[<0.001]
0.001
[0.001]
0.003
[<0.001]
0.007
[0.001]
0.005
[0.001]
C -0.001
[<0.001]
0.001
[<0.001]
-0.002
[<0.001]
-0.001
[<0.001]
-0.002
[<0.001]
D -0.003
[<0.001]
-0.004
[<0.001]
-0.002
[<0.001]
-0.004
[0.001]
-0.002
[<0.001]
Notes: HOLC A-D weighted averages are derived from a regression of each respective outcome on HOLC B-D indicators, both overall and by
region, with student enrollment used as precision weights. Standard errors are included in brackets below each mean. They are heteroskedasticity-consistent and estimated using the Huber-White sandwich estimator.
84
Table C3.T Tukey HSD – Overall and by Region, Student Performance [Adjusted]
U.S. Census Bureau Region
Overall Midwest Northeast South West
Avg.
Student
Math &
ELA Scores
B vs. A -0.274***
[0.008]
-0.329***
[0.014]
-0.238***
[0.010]
-0.248***
[0.023]
-0.312***
[0.019]
C vs. A -0.465***
[0.007]
-0.450***
[0.013]
-0.446***
[0.009]
-0.398***
[0.022]
-0.624***
[0.018]
D vs. A -0.559***
[0.007]
-0.542***
[0.013]
-0.506***
[0.009]
-0.529***
[0.022]
-0.768***
[0.019]
C vs. B -0.191***
[0.004]
-0.121***
[0.008]
-0.208***
[0.005]
-0.149***
[0.015]
-0.311***
[0.010]
D vs. B -0.285***
[0.005]
-0.213***
[0.009]
-0.268***
[0.005]
-0.280***
[0.015]
-0.456***
[0.011]
D vs. C -0.094***
[0.004]
-0.092***
[0.007]
-0.060***
[0.004]
-0.131***
[0.013]
-0.144***
[0.009]
Student
Learning
(Annual)
B vs. A 0.003***
[0.001]
-0.003
[0.002]
-0.003
[0.001]
0.010***
[0.003]
0.014***
[0.002]
C vs. A -0.002
[0.001]
-0.001
[0.002]
-0.013***
[0.001]
0.013***
[0.003]
0.005*
[0.002]
D vs. A -0.006***
[0.001]
-0.007***
[0.002]
-0.015***
[0.001]
0.001
[0.003]
0.005**
[0.002]
C vs. B -0.005***
[0.001]
0.002
[0.001]
-0.010***
[0.001]
0.003
[0.002]
-0.01***
[0.001]
D vs. B -0.009***
[0.001]
-0.004***
[0.001]
-0.012***
[0.001]
-0.008***
[0.002]
-0.009***
[0.001]
D vs. C -0.004***
[<0.001]
-0.006***
[0.001]
-0.002***
[0.001]
-0.012***
[0.002]
0.001
[0.001]
Avg.
Student
Test Score
Change
B vs. A -0.002***
[0.001]
-0.005***
[0.001]
-0.008***
[0.001]
0.004*
[0.002]
0.006***
[0.001]
C vs. A -0.006***
[0.001]
-0.004***
[0.001]
-0.012***
[0.001]
-0.003
[0.002]
-0.001
[0.001]
D vs. A -0.009***
[0.001]
-0.010***
[0.001]
-0.013***
[0.001]
-0.007***
[0.002]
-0.001
[0.001]
C vs. B -0.004***
[<0.001]
0.001
[<0.001]
-0.004***
[<0.001]
-0.008***
[0.001]
-0.007***
[0.001]
D vs. B -0.006***
[<0.001]
-0.005***
[0.001]
-0.005***
[<0.001]
-0.011***
[0.001]
-0.007***
[0.001]
D vs. C -0.002***
[<0.001]
-0.006***
[<0.001]
<0.001
[<0.001]
-0.004***
[0.001]
<0.001
[0.001]
Notes: Includes all A-D HOLC grade pairwise comparisons from Tukey’s HSD post hoc multiple comparison test. Average differences are
provided first with standard errors below in brackets.
*** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level