Full text of Working Papers (Federal Reserve Bank of Chicago) : School Choice Through Relocation : Evidence From the Washington, D.C. Area, Working Paper 1999-07

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Working Papers Series

School Choice Through Relocation:
Evidence from the Washington, D.C. Area
By: Lisa Barrow

Working Papers Series
Research Department
WP 99-7

School Choice Through Relocation: Evidence from the Washington, D.C. Area

Lisa Barrow
Federal Reserve Bank of Chicago
November 1997
Revised: March 1999

I would like to thank Orley Ashenfelter, Kristin Butcher, David Card, Seth Carpenter, Anne Case, Angus
Deaton, Eugena Estes, Henry Farber, Pinelopi Goldberg, Michael Greenstone, Maria Hanratty, Bo Honoré,
Alan Krueger, Helen Levy, Cecilia Rouse, Lara Shore-Sheppard, and Douglas Staiger for helpful conversations
and advice. I am also grateful to Kimberly Sked for research assistance, Ronald Jantz for assistance with
mapping data and software, and Robert Mann of the District of Columbia Public Schools and William
Chappell of Prince George's County Public Schools for allowing me access to the school boundary maps for
their districts. Seminar and workshop participants at Princeton University have also provided useful comments.
Finally, I thank the National Science Foundation Graduate Fellowship Program and the Industrial Relations
Section at Princeton University for financial support. The views expressed in this paper are those of the author
and are not necessarily those of the Federal Reserve Bank of Chicago or the Federal Reserve System. All errors

are my own.

Abstract
In this paper I show how the monetary value that parents place on school quality may
be inferred from their choice of residential location. The method identifies the valuation that
parents place on school quality from the differential effect that measures of school quality
have on the residential choices of households with and without children. I implement the
method with data from the U.S. Census for Washington, D.C. using residential location
decisions in 1990. For whites I find that school quality is an important determinant of
residential choices and that households with children in the top income quintile are willing to
pay $3,300 for schools that generate a 100 SAT point advantage. The evidence does not
indicate that the choices of African Americans are influenced by school quality, which
suggests that this group may be constrained in their location choices.

Lisa Barrow
Economic Research Department
Federal Reserve Bank of Chicago
230 S. LaSalle St.
Chicago, IL 60604
(312) 322-5073
lbarrow@frbchi.org

1
I. Introduction
School choice is among the most controversial issues in the United States today. At the center of
the debate is the question of whether parents can, or should, be able to choose among alternative publiclysupported schools for their children. In 1996, Presidential candidate Bob Dole found the issue important
enough to raise in his nomination acceptance speech, arguing that a wider choice of schools should be
available to all families, not just the children of high-income families. In the midst of this debate, a variety
of programs aimed at increasing parents’ school choices have been adopted across the country. For
example, Minnesota enacted an open enrollment policy to allow students to attend public schools outside of
their own district; the city of Milwaukee established an experimental education voucher program to enable
low-income families to enroll their children in private schools; and, nationwide, so-called charter schools
have been set up within many school districts to offer programs that depart from the current offerings of the
public school system.1
Popular discussions of school choice often seem to overlook the primary mechanism for school
choice emphasized by economic models of public-good provision, namely, the choice of residential location.
In a seminal 1956 paper, Tiebout argued that competition among local jurisdictions would lead to the
efficient provision of a range of public goods including public schooling.2 In the spirit of Tiebout’s

1

See Armor and Peiser (1997) for a general description of the Minnesota plan and charter school
programs. I know of no evaluation studies of the Minnesota program. See Wells et al. (1998) for an
examination of the politics of charter school reform. Rouse (1998) provides an evaluation of the effects of
the Milwaukee program on student achievement, and Witte, Sterr, and Thorn (1995) and Witte et al.
(1994) contain more detailed descriptions of the program.
2

While not many papers have directly looked at households choosing schools through location,
Lankford and Wyckoff (1997) examine the effect of racial composition of schools and neighborhoods on
white families location decisions. Several other studies implicitly assume that households choose schools
through location choice. Hoxby (1994) assumes that households choose location based, at least in part, on
school district quality, and thus school districts in metropolitan areas with lower enrollment concentration
face more competition. Black (1997) and Bogart and Cromwell (1997) both look at differences in house
prices across school boundaries and within the same neighborhood for evidence of how much good schools
are worth; thus, they assume that households choose locations within a neighborhood based on school

2
hypothesis, nearly one-half of parents in the 1993 National Household Education Survey reported that their
decision of where to live was influenced by where their child would go to school (McArthur, Colopy, and
Schlaline (1995)). The importance of this type of choice is also evident in the real estate market. Real
estate agencies routinely provide school information to families with young children. For example, Century
21 South Lakes real estate agency, operating in Northern Virginia, provides average Scholastic Aptitude
Test (SAT) scores for school districts in their area on their World Wide Web home page. 3
In this paper I examine school choice embedded in the choice of residential location. By measuring
the sensitivity of residential location choices to school quality, I can assess whether some school choice
already exists in some areas of the U.S. A key difficulty that arises in studying the effect of school quality
on residential location choice is that unobserved characteristics of different jurisdictions may be correlated
with local school quality. To overcome possible biases arising from such unobservable factors, I compare
the residential location choices of households with children to households without children. I assume that
any unobservable attributes are equally valued by both types of households. Thus, I measure the excess
sensitivity of location decisions of households with children to differences in school quality. In the end, I
use the results to estimate how much households are willing to pay for quality education, and how
increasing the school quality of various location choices would change demands for other communities.
I find that among white households in the District of Columbia, households with income above the
bottom quintile seem to exercise school choice through location choice. One implication of these findings is
that public schools already face some competition from other public schools in the area. Thus, the
promised benefits of competition from moving to complete choice with a full voucher program are
diminished. Among African-American households, I do not find evidence that households with children

quality.
3

Century 21 South Lakes’ home page: http://www.northernva.com/c21/index.html

3
locate in areas with higher school quality than households without children. These results give less support
to the idea that public schools are competing with one another. More particularly, they suggest that
African-American communities may not be competing with each other on the quality of public schools.
The paper is organized as follows. Section II presents the theoretical background and model,
Section III discusses the data, and Section IV presents estimation results. In Section V, I interpret the
results in terms of changes in location distribution associated with improvements in school quality, and I
calculate estimates of the willingness to pay for school quality, as measured by the average SAT score for
the area high schools. Section VI concludes.

II. Theoretical Background and Model
In a very influential paper, Tiebout (1956) hypothesized that households choose from alternative
locations based on preferences for the local services provided by different communities. Tiebout’s analysis
arose in response to arguments by Musgrave (1939) and Samuelson (1954 and 1955) that socially optimal
levels of public goods cannot be obtained because people do not have the incentive to reveal their true
preferences. Tiebout argued, instead, that if people choose locations based on local public service
provision, then they reveal their preferences by where they move. This paper builds on Tiebout’s theory
and attempts to observe differences in preference for school quality, revealed through location choice,
arising from differences in household composition.
The conceptual experiment motivating the analysis considers a household living in a certain city
and deciding whether to relocate. The household may compare employment opportunities and amenities
offered by a set of community choices and choose the location that best matches its ideal. This may mean
staying in the current location, moving to a nearby suburb, or relocating to a different labor market entirely.
Assuming school quality is more important to households with children than households without children,
households with children should be more likely to select residential locations that offer higher school

4
quality. The degree to which they are observed to trade off higher school quality for other amenities (such
as lower rent) provides a measure of the relative valuation placed on school quality.

A. Multinomial Logit Model
More formally, I assume households maximize indirect utility of the form:
Uhj'Vhj%ghj ,

(1)

where Vhj is the component of indirect utility that can be explained by observable household and location
characteristics and ghj is an unobserved component. A household then chooses to locate in the community
that maximizes its utility. Assuming that the ghj are independently and identically extreme value
distributed, the probability that household h chooses community j is:
Phj'

e

Vhj

je
K

,
Vhk

(2)

k' 1

where K represents the maximum number of potential communities. As is well known, the multinomial
logistic model (2) has a number of advantages over other possible specifications. These include the ease of
computation and the fact that choice among a subset of options is governed by a similar logistic model. In
addition, estimates from the logistic model easily can be used to simulate the effect of adding another
choice option. On the other hand, the extreme value distributional assumption implies that the relative
probability of choosing one location over another is unaffected by the characteristics of other choices in the
choice set, commonly referred to as independence of irrelevant alternatives. Clearly, this assumption is not
innocuous. However, preliminary attempts to relax the extreme value assumption have made little
difference in the estimation results.
I further assume that the observed component of indirect utility, Vhj , can be approximated by a

5
linear combination of choice specific attributes and interactions of choice specific attributes with household
characteristics, i.e.
Vhj'Xj

%Wh·Zj

(3)

where Xj are attributes of the community, Wh are household characteristics that do not vary across
communities but are interacted with community specific attributes, Zj.

and

are unknown parameters to

be estimated.

B. Identification Strategy
Throughout, I assume that households' preferences for local public goods differ with the
composition of the household. Specifically, I hypothesize that households with children will value public
education more than households without children because households without children do not directly
benefit from higher quality public schools. An identifying assumption I make is that while school quality
differentially affects households with and without children, other local amenities affect location
probabilities of all households similarly. Thus, consider rewriting equation (3) as follows with school
quality as the only observed amenity:
Vhj'

% 1Sj·1(children under 18) % ej,

1Sj

(4)

where Sj is school quality, 1(children under 18) is an indicator variable taking on values of one for
households with children and zero for households without children, ej is a local amenity unobserved by the
econometrician, and
without children is
estimating

1

and

,

1

1

.

1

, and are unknown parameters. The value of school quality for households

1

and the value for households with children is

% 1. Therefore, I am interested in

1

6
Suppose that the value of the unobservable amenity is linearly related to observable school quality,
i.e.:
ej'd@Sj

(5)

Thus, equation (4) can be rewritten as,
Vhj

' 1Sj% 1Sj·1(children under 18)% (d·Sj)
' [ 1% ·d]Sj% 1Sj·1(children under 18),

(6)

Thus, the coefficient on school quality alone will be biased by the omitted variable by the amount @d, but
the school quality-child interaction coefficient will not. This results from the key assumption that omitted
location characteristics have the same effect on households with and without children. Note in particular
that if households without children put no value on school quality per se, then the estimated direct effect of
school quality depends only on the correlation of school quality with the unobserved amenity, while the
estimated interaction of school quality with the presence of children provides an unbiased estimate of the
true valuation of school quality by households with children.
More generally, if the unobservable amenity is projected on the observable characteristic with some
error, i.e. ej ' d·Sj% j, then equation (1) becomes,

Uhj'[ 1% ·d] Sj% 1Sj@1(kids under 18)%
'Vhj% hj,
where

hj

equals

hj

@ j plus a household and location specific error term. Assuming the combined errors

are independently and identically extreme value distributed, the location probabilities are defined as in
equation (2).

C. Empirical Specification
In order to estimate the model, I need a sensible specification for the indirect utility function. I

(7)

hj

7
model household indirect utility in each possible location as a function of school quality, other local
amenities, household income, and the cost of choosing a particular community. Since local public goods
are predominantly financed with a local property tax,4 the cost of locating in a particular district j can be
considered the property tax payment, tj pHH, where tj is the local property tax, which varies across counties
only, pH is the unit price of housing, and H is the number of housing units purchased. Additionally, it is
possible that zoning regulations put binding constraints on the quantity and quality of housing available for
purchase. Because households that buy more and better quality housing pay a larger share of the tax
burden, community residents have an incentive to limit the development of lower quality and/or smaller
sized housing units. Zoning restrictions of this type may lead to an increase in the size and quality of the
median apartment available in the local market. In this way, potential residents may have to buy more
housing than they would with no restrictions in order to locate in a particular community. To attempt to
capture both the tax cost and differences in the housing stock available in each community, I include
median rent as a measure of the tax price of the location combined with some indication of the type of
housing available.
Using median rent as a measure of the price of a location raises immediate concerns about the
possibility of problems with endogeneity. An analogous issue is commonly discussed in the empirical
industrial organization literature in which consumer level data is often unavailable and market shares and
prices are simultaneously determined (See Berry (1994)). Relative to a market-level analysis, the use of
household level data reduces the simultaneity problem. Nevertheless, if median rents are determined by
local market forces and the sample households are representative of net market demand shifts, then the
problem still exists. However, there are several reasons to think simultaneity is less of a problem for the

4

See U.S. Bureau of the Census (1993). In the case of the Washington, D.C. area, property taxes
are a major source of revenue in all the surrounding communities, e.g., the property tax is 36.5 percent of
tax revenue in D.C., 57 percent of local tax revenue in Maryland, and 79 percent of local tax revenue in
Virginia.

8
case at hand. First, the household group of interest—those living in D.C. in 1985 and remaining in the area
in 1990—is very small relative to the population of the area. As a result, there are likely other groups of
households that are making off-setting location decisions, such as older households relocating for
retirement. Furthermore, to the extent that the land in these areas in not fully developed, housing supply
may be very elastic, and hence demand shifts would have little if any price effect.5
An additional source of endogeneity to consider when using median rent involves the error
component related to unobserved location characteristics that are observed by the households. If these
unobserved characteristics are valued similarly by all households, then the market rent is expected to reflect
these unobserved amenities. As discussed above, this will lead to biased coefficients on variables such as
rent but will not affect the coefficient on the school quality-child interaction term.

III. Data
My sample consists only of households that lived in the District of Columbia in 1985. I choose to
focus on D.C. for several reasons. First, although D.C. is a majority African-American city, its suburbs
include both predominantly white and predominantly African-American communities. In particular, Prince
George's County, MD on the eastern border of the District has a majority African-American population,
mainly located in communities near the District. Second, there has been a great deal of concern about the
quality of the public schools in D.C. The District of Columbia Public School System spends more per
pupil (on average) than school districts of similar size and demographic make-up6 while scoring below

5

In fact, I investigated this problem empirically and found that the proportion of households
locating in each area for the sample of interest is negatively correlated with the growth in median rent by
area from 1980 to 1990. However, when I use all households in the metropolitan area in 1990 that move
between 1985 and 1990, the proportion locating in each area is positively correlated with growth in median
rent from 1980 to 1990.
6

The peer districts as defined by the District of Columbia Financial Responsibility and
Management Assistance Authority are: Baltimore, Boston, Charlotte-Mecklenburg, Chicago, Cleveland,
Detroit, Memphis, Milwaukee, New Orleans, and Newark.

9
national and peer averages on the Comprehensive Test of Basic Skills (CTBS) and the SAT (District of
Columbia Financial Responsibility and Management Assistance Authority (1997)). Third, and perhaps
most importantly, Washington, D.C. is one of the few areas in which this study could be undertaken with
Census Public Use Microdata Samples (PUMS) because the school districts in the area are defined at the
county and independent city level and thus coincide fairly well with the geographic information on the
Census files.7
The main source of data for this project is the 1990 Decennial Census Public Use Microdata
Sample (PUMS), 5% Sample A files. These data provide detailed information on individuals’ households,
where they were living in 1985, where they live in 1990, income, age, education, race, and public and
private school attendance. From these data, I select households in which the householder reported living in
the District of Columbia in 1985 and which were still located in the D.C. Metropolitan Statistical Area
(MSA) in 1990 as defined by the 1990 MSA definitions.8 To these data I have matched information on
local characteristics from Census Summary Tape files and other various sources. The 1990 Census PUMS
files allow me to identify in which Public Use Microdata Area (PUMA) or sub-PUMA a household is
located (See Census of Population (1990b) for more details on PUMAs.). Thus, in the D.C. MSA there are
26 location choices including 5 choices in the District of Columbia proper. A map of the PUMA and subPUMA boundaries in the Washington, D.C. MSA is shown in Figure 1.9

7

As discussed below, the smallest identifiable area on the Census PUMS files is a Public Use
Microdata Area (PUMA) or sub-PUMA of at least 100,000 people. Several counties and independent
cities in the D.C. Metropolitan Statistical Area have fewer than 100,000 residents and therefore cannot be
identified separately from one or more nearby school districts.
8

The householder in the Census files, is "the person, or one of the persons, in whose name the
house is owned, being bought or rented and who is listed in column 1 of the census questionnaire." Census
of Population and Housing (1990b).
9

Some multi-county PUMAs include counties that are not contained in the Washington, D.C. MSA
1990 definition; however, I include the PUMA in my sample as long as at least one county in the PUMA is
in the MSA.

10
An important component of the data is the school quality measure, specifically average SAT scores
by location. I collected high school SAT averages as well as school attendance area maps from the local
school districts. For D.C., Montgomery County, and Prince George’s County where the location choices
are smaller than school districts, I was able to use high school attendance area maps in order to calculate
local SAT averages.10 For all other areas, I use average SAT scores by school district or school district
group.
As described above, I focus on the location decisions of households living in the city. Column (1)
of Table 1 presents selected mean household characteristics for the 12,805 households in the Census
PUMS sample that were living in the District of Columbia in 1985.11 Compared to the nation as a whole,
these households have higher income, are more educated, are less likely to have children under 18, and are
more likely to be African-American. By 1990, 1,351 of these households had moved out of the D.C.
metropolitan area, 1,568 had moved out of the city to the surrounding suburban communities in Maryland
and Virginia, and 9,886 remained in the District of Columbia. In column (2) of Table 1, I summarize
characteristics of the households remaining in the D.C. metropolitan area. Again, the households have
higher income, more education, are less likely to have children under 18 years of age, and are more likely to
be African-American than the average U.S. household.
Sample means by race are presented in the last two columns of Table 1. Means for the white
households are presented in column (3), and means for the African-American households are presented in
column (4).12 These calculations reveal striking differences. Most notably, average household income for

10

More details on the data are contained in an appendix available from the author.

11

This excludes individuals living in group quarters and households whose head reported being on
active military duty. Note that this is not the sample I use in estimation.
12

The sample referred to as white is actually 90.6 percent white; 4.0 percent Asian; 0.6 percent
Native American, Eskimo, or Aleut; and 4.8 percent other.

11
white households is more than double the average for African-American households. The white households
are also younger and have an average of three years more education than the African-American households.
In addition, whites are nearly half as likely to have children in the household, but conditional on having
children, they are more than three times as likely to enroll at least one child in private school.
The D.C. Public School District persistently ranks below average on student outcome measures
such as average SAT scores. In fact, the D.C. district-wide SAT average is more than 100 points below
any other school district in the D.C. metropolitan area, and only two of the locations within the District
have SAT averages above 700 points.13 Thus, the best way for parents to improve public school quality
for their children is to move out of D.C. Of course, private schools may be an option for some households
as well. Indeed, households with children are more likely to be living in the suburbs in 1990 than
households without children, and this is true for both white and African-American households. Of the
households still living in the District of Columbia in 1990, 15 percent of white households and 34 percent
of African-American households had at least one child under 18 years of age, while 29 percent of white
households and 45 percent of African-American households in the suburbs had children.
The patterns of location, both inside and outside the District of Columbia, vary dramatically by
race and reflect the high racial segregation of many of the areas. Eighty-five percent of the households
remaining in the D.C. metropolitan area still live in the District in 1990, with the highest percentage located
in Anacostia and the lowest in the Northwest part of the city. Breaking this down by race, the majority of
white households locate in the Northwest part of D.C. (49.3 percent), and the fewest locate in Anacostia
(2.9 percent). For African-American households the reverse is true; the fewest number locate in the
Northwest part of D.C. (1.7 percent), and the greatest number locate in Anacostia (37.3 percent).
Unfortunately, I cannot identify in which sub-PUMA households were located in 1985 so it is impossible to

13

SAT scores range from 400 to 1600 points. The national average for 1989, the year used in the
estimation, is 903.

12
say exactly how the distribution of households across locations within the District has changed.
Of the households moving out of the District, over 90 percent move to one of four school
districts—two in Maryland and two in Virginia—and 75 percent move to either Montgomery County, MD
or Prince George's County, MD.14 Once again, the location patterns vary by race. For white households,
53.5 percent move to either Montgomery or Prince George's Counties, with 42.3 percent locating in
Montgomery. The next most frequent county choice for whites is Arlington, VA with 20.1 percent locating
there. For African-American households, over 90 percent move to either Montgomery or Prince George's
Counties, with 77.7 percent moving to Prince George's and 13.7 percent moving to Montgomery. No more
than 4 percent of African-American households move to any area outside of Prince George's and
Montgomery Counties. Clearly, African-American and white households do not choose the same
communities. This outcome is consistent with many explanations including differences in income by race,
differences in preferences, and housing discrimination.
To measure school quality, I use average SAT scores for each location choice.15 While average
SAT scores are not a perfect measure of school quality, they likely capture information about true school
quality as well as information about peer group quality. It is widely acknowledged that average SAT
scores vary across schools, school districts, and states due to differences in the participation rate among
potential test-takers; however, schools, school districts, and real estate agencies frequently cite SAT

14

Appendix Tables B3a-B3d give means and standard deviations for characteristics of the sample
households by location and are included in the appendix available from the author. The most interesting
facts to note are that average annual income of sample households by location choice ranges from $26,812
in the Stafford County group to $129,014 in Calvert and St. Mary’s Counties, MD and that the households
locating in Northwest D.C. are the least likely to have children under 18 years of age.
15

I have also tried using pupil-teacher ratios to measure school quality. In these estimations,
households with children are shown to prefer lower quality schools. However, I do not believe pupilteacher ratios can be considered a plausible measure of school quality because the District of Columbia
Public School System has the lowest pupil-teacher ratios in the area but is not considered to provide the
best public education. The low pupil-teacher ratios in D.C. may very well arise due to small special and
remedial classes.

13
averages as evidence of school quality. Thus, it is likely that parents perceive SAT scores as an indicator
of school quality. It is also reasonable to think that parents have access to this information since real estate
agencies make it available to potential home buyers, and The Washington Post publishes many of the local
averages every fall. Within the District of Columbia, SAT averages range from 631 in Anacostia to 826 in
Northwest D.C. The Anacostia SAT average is the lowest over all of the communities while the Bethesda/
Chevy Chase area (in Montgomery County, MD) has the highest average at 1045. In fact, each of the six
locations in Montgomery County has SAT averages at least 100 points higher than the highest SAT
average in the District and more than 300 points higher than Anacostia.

IV. Multinomial Logit Estimation Results
As noted above, the Washington, D.C. area is segregated, and household characteristics vary
dramatically by race. A likelihood ratio test rejects that the coefficients for white and African-American
households from a multinomial logit estimation are equal,

(15) ' 561.46, hence I conduct my analysis by

2

race. The results for white and African-American households are presented in subsections A and B,
respectively. Subsection C contains results allowing the effect of school quality to vary with income for
both races, and subsection D includes estimates allowing some private school choice.

A. Estimates for White Households
Table 3 presents indirect utility coefficient estimates from a multinomial logit model for white
households. The specifications model the probability of choosing each of the 26 choices of location in
1990 conditional on living in the District of Columbia in 1985. These 26 choices are public school choices,
and hence any household choosing to enroll children in private school is not included in the estimation
sample.16 Note that if independence of irrelevant alternatives holds, excluding the private school choices

16

Means of household characteristics for all of the estimation samples are presented in Appendix
(continued...)

14
from the model has no effect on the coefficient estimates.
The model in column (1) includes only school quality as measured by average SAT points, median
rent as the proxy for the cost of choosing a given location, and the number of housing units to account for
differences in the size of the areas. Results from this parsimonious specification suggest that households
are more likely to locate in areas with higher median rent and that high SAT scores reduce the probability
that a household locates to a given area. It is important to bear in mind that both median rent and SAT
scores are likely to be correlated with other location characteristics. As a result, the models in columns (2)
and (3) include controls for other location choice characteristics.
Characteristics contained in both the column (2) and column (3) specifications are: the location’s
distance in miles from central D.C., the crime rate, the number of D.C. Metro stations, total per capita
county and state expenditure17, the population density, the poverty rate, and the proportion of housing that
is owner occupied. As households may have strong preferences for living in areas with households of the
same race and the same class, in column (3) I additionally include the proportion of whites in a location and
indicators for the education level of the householder interacted with the proportion of persons in the
community aged 25 and older with the same education level.18
Across all specifications the effects of most location amenities do not change and are generally
consistent with expectations. The number of D.C. Metro stations in a location has a significant, positive
effect on location probabilities. Adding one Metro station increases location probability by as much as 4

16

(...continued)
Table B1.
17

State expenditures are included to make the Maryland and Virginia public expenditures
comparable to D.C. which is both the state and local government.
The education levels are education #12 years of schooling and no high school diploma, high
school graduate, some college education, and bachelor’s degree or higher.
18

15
percent.19 Per capita state and local expenditure has a positive effect on location probability. To the
extent that expenditure reflects provision of other local public goods, households seem to value these public
goods. Increasing per capita expenditure for all locations in the District by $500 (from $4,926) would
increase the average probability of locating in Northwest D.C. from 0.40 to 0.45.
Population density also has a positive effect on the probability that a household chooses a
particular location. This measure was included as a proxy for the presence of open spaces, such as parks,
as well as the idea that households might move out of the city for more housing space. While this sample
of households does not seem to prefer low population density areas, this is not entirely surprising given that
these households exhibit some preference for city dwelling in 1985 and that the majority choose to stay in
the city. Finally, as one might expect, the proportion of persons in poverty decreases the probability of
location. The poverty coefficient is large in absolute value and implies a 16 percentage point decrease in
location probability per 5 percentage point increase in the poverty rate.
The coefficients on the race and class variables in column (3) reflect the observed segregation in
the D.C. area. White households are significantly more likely to locate in areas with higher percentages of
whites. For all but the most educated, the proportion of people with the same level of education increases
the probability of location; however, the negative result for householders with a bachelor’s degree or more
education is not statistically significant. Only the effects of three variables—distance, crime, and
proportion of housing that is owner occupied—are sensitive to including race and class in the specification.
Turning to the variables of interest, both the direct effect of average SAT scores and the interaction
effect of SAT scores on households with children rise noticeably when controls for other local amenities are

The change in probability of moving to location j with a change in Xj is @Pj@(1-Pj) where is the
coefficient on Xj in the indirect utility function and Pj is the probability of moving to location j. Thus the
maximum change in probability is 0.25@ which occurs when Pj ' 0.5. Since the average predicted
probability of locating in Northwest D.C. is 0.41, the change in probability of locating in Northwest D.C.
from the change in one of its own characteristics will approximately equal the maximum change.
19

16
included. The direct SAT effect becomes positive and significant. The coefficient on average SAT scores
implies that a 100 point increase in average SAT scores increases the probability of location choice by a
maximum of 13 percentage points for households without children. For households with children, the main
effect combined with the interaction term implies that a 100 point increase in average SAT scores increases
the probability of location by a maximum of 25 percentage points. Note, however, that I assume the main
effect of SAT scores cannot be interpreted directly since I rely on being able to identify the effect of school
quality on location choice using the difference in the effect between households with and without children.
Thus, if the true valuation of schools by households without children is zero, then only the net 12
percentage point increase in location probability for households with children can be attributed to the
change in school quality.
The effect of median rent on location probabilities becomes negative when controlling for the other
local amenities, and the negative effect becomes significant when controlling for race and class. The
coefficient in column (3) implies that a $100 per month increase in the median rent leads to a maximal 5
percentage point decrease in the probability of location.
In columns (1)-(3) of Table 3, I have allowed preference differences between households with and
without children to enter only through the school quality-child interaction. To examine whether the
coefficient estimate on the SAT-child interaction term is upwardly biased, I present estimates fully
interacting all right-hand-side variables with the indicator for whether a household has children under 18.
The coefficient estimates for the direct effects of all variables are listed in column (4) of Table 3, and the
coefficient estimates for the interaction effects are listed in column (5). Looking at the SAT coefficient
estimates, the estimates tell a remarkably similar story to the estimates in column (3). The SAT-child
interaction coefficient estimate in column (5) is 0.424 compared to the coefficient estimate of 0.474 on the
SAT-child interaction in column (3). Since I am primarily interested in the school quality variable, I
interpret this as evidence that it is not unreasonable to restrict the differences for households with and

17
without children to enter only through the school quality measure.20 Although I can reject that the child
interaction coefficients on all but school quality are jointly zero ( 2(14) '170), individually, most are not
statistically significant. Thus, I continue to use the simpler specification of column (3) that is easier to
interpret and has more precision.

B. Estimates for African-American Households
Table 4 presents the same specifications as above estimated using the African-American household
sample. The base SAT coefficient in the simplest specification is again negative, but more importantly, the
SAT-child interaction has a negative and statistically significant effect on location probability. Median
rent, which is expected to be biased in a positive direction due to omitted amenity variables in this
specification, has a negative effect on location probabilities.
The specifications in Columns (2) and (3) of Table 4 include controls for other location amenities
that might be correlated with SAT scores and/or median rent. Several amenities have surprising coefficient
signs. The crime rate and the proportion of persons in poverty have positive effects on location
probabilities, and the number of metro stops has a negative effect on location probabilities. These results
may arise because the African-American households largely choose to locate in relatively few places. To
the extent that these African-American communities have higher crime and poverty rates and fewer metro
stops than the white communities, these results may reflect the differences between the de facto choice set
for African-Americans and the full set of choices in the model rather than true preferences. I find some
evidence that this may be the case when I restrict the choice set to communities with more than 30 percent

20

Note, however, that if the SAT-child interaction coefficient estimate is upwardly biased by an
unobserved amenity that is highly correlated with the SAT-child interaction but orthogonal to the other
amenity-child interaction terms, then this test will not help reveal the upward bias.

18
African-American populations.21
African-American households’ location decisions reflect the existing segregation in the area,
similar to the estimates for white households. African-Americans are much more likely to move to
locations with higher proportions of African-Americans, and all education groups tend to move to
communities with more households of their same education level.
Looking at the school quality and cost measures, there is little evidence that SAT scores
significantly affect the location probabilities of African-American households. I hypothesized that SAT
points would have a positive effect on location probabilities for households with children. The effect of
SAT points on the location decisions of households with children is negative for both the column (2) and
column (3) specifications; however, the coefficient loses statistical significance with race and class
controls. The base effect of SAT scores becomes positive with the inclusion of race and class controls, but
again, the result is not statistically significant. The cost measure enters as predicted. Median rent has a
significant negative effect on location probabilities, implying up to a 10 percentage point decrease in
location probability per $100 increase in median rent per month.
Columns (4) and (5) of Table 4 estimate the column (3) specification allowing the effect of all
variables to differ by child status. As with the white households, the results are largely consistent with the
estimates of column (3); allowing the difference in preferences by child status to enter only through the
school quality measure seems to have little affect on the school quality coefficient estimate.
The findings for African-American households are puzzling. I have explored potential
explanations for why these results do not square with economic theory such as income effects, higher costs
of relocation, and constrained choice.22 While I do not present the findings here, they suggest that there is

21

Results and discussion are contained in the appendix.

22

Specifically, I try reweighting the data for African-American households to reflect the distribution
(continued...)

19
no simple explanation for the persistent negative coefficient on the SAT-child interaction for AfricanAmerican households. The explanation is likely to require a more complex combination of income and
social factors than I can accommodate in the model at this time. However, one set of estimates suggests
that a more restricted choice set may be appropriate for African-American households. Restricting the
choice set to locations with at least 30 percent of the population being African-American, the coefficient
estimates on the base SAT effect, the crime rate, the proportion of persons in poverty, and the number of
metro stations all have the predicted sign in this specification while in most other specifications the
coefficient signs are counterintuitive.23 This result could arise if the white communities are effectively not
available to the African-American households and the white communities have better than average
amenities. If this is the case, it would appear that African-American households have different preferences
than expected when, in fact, among the available choices African-American households prefer lower crime,
better metro access, etc.

C. Allowing School Quality Effects to Vary with Household Income
The estimations in Tables 3 and 4 assume that indirect utility is linear in household income, and
thus income has no effect on the probability of location choice because it does not vary across the choices.
However, poorer households are less able to relocate. To the extent that they are located in areas with the
lowest quality schools in 1985, I should not observe them exhibiting preference for school quality if they

22

(...continued)
of income or income and household structure of the white households. I also try estimating the model using
only the wealthiest African-American households, using only households that moved since 1985, or using
only the subset of choices with at least 30 percent of the population African-American. See the appendix
available from the author for the complete set of results.
23

Restricting the white sample to the choice set of communities with more than 30 percent of the
population white decreases the base SAT effect and increases the SAT-child interaction coefficient. The
effect of the proportion of persons in poverty becomes positive and statistically insignificant. All other
results remain relatively unchanged.

20
are unable to trade other consumption, perhaps necessities, for school quality. To explore this possibility, I
allow the effect of school quality to vary with income. Table 5 presents estimates including interactions of
income quintiles with SAT and with the SAT-child interaction omitting the lowest quintile.24 Results for
white households are presented in column (1), and results for African-American households are presented
in column (2).
For both race groups, the results on the amenity, race, and class variables are substantially
unchanged from the corresponding estimates in Tables 3 and 4, so I focus on the results for the variables of
interest. For whites the base effect of average SAT scores increases monotonically with income quintile,
becoming significantly different from zero for the third through fifth quintiles. For the SAT-child
interaction, there is no general pattern by income quintile. Although all coefficients are positive, only the
coefficient for the third quintile is statistically significant. However, restricting the SAT-child-income
interactions to equal zero can be rejected by a likelihood ratio test,

2

(4)'11.53. The net effects imply that

a 100 point increase in SAT scores maximally increases the probability of location by 14 percentage points
for households with children in the lowest income quintile to 30 percentage points for households with
children in the third income quintile. For households without children, the maximal increase in location
probability associated with a 100 point increase in SAT points ranges from 9 percentage points for the
lowest quintile to 16 percentage points for the top quintile.
For African-American households, the base effect of average SAT scores does not vary with
income quintile. When interacted with the SAT-child interaction, the coefficients are all positive, and they
are statistically significant for households in the second and fourth income quintiles. However, the net
effect of SAT scores on location probabilities is only positive for households in the fourth income quintile

24

The quintiles in $1995 are: household income less than or equal to $14,880; between $14,880 and
$29,940; between $29,940 and $46,980; between $46,980 and $77,750; and greater than $77,750. They
are constructed from the 1990 Census Summary Tape File 3A using all of the households living in the
District of Columbia in 1990.

21
which implies a maximal 3.3 percentage point increase in location probability for a 100 point increase in
SAT points. Compared to households without children, households with children in both the second and
fourth income quintiles are more likely to locate in areas with higher SAT scores.
For both white and African-American households, the results suggest that there are some
differences in school quality effects by income. Although many coefficients are not individually significant,
likelihood ratio tests easily reject that the income-SAT and income-SAT-child interactions are jointly equal
to zero. One could conclude from the results that lower income households have less preference for school
quality; however, if poor households cannot relocate due to income constraints, one would observe the same
outcome. As a whole, these results suggest that wealthier households, particularly wealthier white
households, seem to be able to exercise school choice through the choice of residential location.

D. Private School Choice
One omission from the above analysis is that some parents choose to send their children to private
school. Forty-five percent of households with children in the Northwest section of D.C. have a child aged
6 to 17 enrolled in private school. Similarly, over 15 percent of households with children living in
Damascus/Poolesville or Bethesda/Chevy Chase have at least one child enrolled in private school (See
Appendix Tables B3a-B3d). I have considered several ways of incorporating private school choice into
the analysis, although none is perfectly satisfactory. One possibility is to add a private school option for
households with children for each location choice. This increases the choice set for households with
children to 52 choices. However, within location the characteristics for all variables other than school
quality and cost, namely private school tuition, are identical. Additionally, any private school could be
chosen by the household so it is unclear what quality and cost measures should be assigned to the private

22
school alternative.25
In an attempt to assess how private school choices may be affecting the above results, I re-estimate
the specification of Table 5 allowing five private school choices—one for each of the locations in the
District of Columbia. As a result, any household with children that sends a child to private school and
locates in the District is added to the estimation sample. These households account for 86 percent of the
white households choosing a private school option and 93 percent of the African-American households
choosing a private option. In Figure 2, I present the proportion of households with children locating in
D.C. that have at least one child enrolled in private school, by race and by income.26 For either race, any
differences in SAT-child effects that are found are likely to occur in the upper tail of the income
distribution where higher proportions of households are sending their children to private school. This is
particularly true for white households since the private school enrollment rate increases to 36 percent in the
top decile. The private school enrollment rate increases more gradually with income for African-American
households in the sample; however, the school quality will be more understated (when not allowing for
private choices) to the extent that African-Americans are living in lower quality school districts on average.
Of course, any systematic differences in the quality of private schools attended by white and AfricanAmerican children make this comparison less clear and are not taken into account in the specifications.
Columns (1) and (2) of Table 6 re-estimate columns (1) and (2) of Table 5 including the
households that locate in D.C. and choose a private school. In these columns, I have simply treated the
households choosing private school as though they chose the public school option. I do this to see more
clearly which changes in estimates arise from changes in the estimating sample and which arise from the

25

Another possibility is to control for private school enrollment on the right hand side of the
equation. The problem with this option is that the decisions of where to live and whether to send children to
private school are likely joint decisions for many households.
26

The income deciles were created using the full sample of PUMS households living in the District
of Columbia in 1990.

23
change in the choice set. Very little difference results from estimating the model using the larger sample.
Columns (3) and (4) of Table 6 re-estimate columns (1) and (2) expanding the choice set to 31 for
households with children. The five "new" choices are each of the location choices in the District with their
respective location characteristics but using private school quality and cost measures. Since I do not
observe the actual private school a child attends, only that he or she is enrolled in private school, I begin by
assigning the quality and tuition cost of the average private school student in the U.S. For private school
quality, I assign each of the private school choices an SAT score equal to the average SAT score for all
private test-takers (962).27 Private school cost is median monthly rent plus one-twelfth of the enrollment
weighted average tuition for private secondary schools of all types (average annual tuition ' $4,708)
(National Center for Education Statistics (1995)). Thus, all private school choices have the same SAT
score, but cost will vary across choices with differences in median rent. Within a location, the only
difference between the public and private choice characteristics are the quality and cost variables. Because
private school choices are only included for District of Columbia locations, the school quality of the private
choice is always higher than that of the public choice. As a result, for all households choosing private
schools in the columns (1) and (2) estimations, the SAT score is understated relative to the private school
average.
The estimates in column (3) for white households show a noticeable decrease in the base SAT
effect from positive and statistically significant to negative and statistically significant. For AfricanAmerican households the base SAT coefficient also decreases and becomes negative and statistically
significant. The SAT-income quintile interactions change little for either sample; however, both samples
show a large increase in the SAT-child-income interaction effect for households in the highest income
quintile. This result is consistent with the earlier observation that school quality would be understated for

27

This average is from 1997 for religious and independent school test-takers, converted to the old
SAT score scale. (The College Board (1997a, 1997b)).

24
all households choosing private school and living in the District. For white households with children, all
income quintiles show net positive effects of SAT scores on location probability relative to households
without children. For African-American households, the same is true for all but the third income quintile;
however, any positive effect of SAT for households with children relative to households without children is
still quite small.
Looking at a guidebook for private schools in the Washington, D.C. area, the private secondary
schools located in the District of Columbia have higher than average tuition (See Coerper and Mersereau
(1995)). Assuming higher average tuition reflects higher average quality, columns (5) and (6) simply reestimate the specification of columns (3) and (4) using a higher quality and higher cost private school
option. The hypothetical high quality, high cost option gets an average SAT score of 1100 and tuition of
$8,000. The SAT score was chosen to be above the highest quality public school choice in the area,
Bethesda/Chevy Chase, and the tuition is approximately equal to the median tuition for private secondary
schools in D.C.28 For both white and African-American households the results show decreased importance
of school quality for all but the bottom income quintile for households with children. However, the monthly
cost coefficient also declines so I will compare the estimates more directly below using the willingness-topay calculations.

V. Interpretation of the Results
These multinomial logit estimates can be used to look at the results in several ways. First, I
consider the effect on location choice of improving the quality of public schools in D.C. Northwest D.C.
has the highest predicted probability of location for white households with and without children, followed
by Central D.C. The next largest probability prediction for whites with children is Bethesda/Chevy Chase.
For African-American households, the largest predicted probabilities are for the Anacostia and Northeast

28

Estimated using tuition data and enrollment information from Coerper and Mersereau (1995).

25
sections of D.C. However, I focus on the effects of improving school quality for whites since the AfricanAmerican results imply a net negative effect on location probabilities for all but the fourth income quintile.
Consider improving the quality of public schools in Northwest D.C. In 1997 a 100 point increase
in SAT score from 50 points below the national mean to 50 points above the mean moves a student from
the fortieth to the fifty-ninth percentile (The College Board (1997)). An increase of 100 points in average
SAT scores in Northwest D.C. would increase the average predicted probability for white households with
children by 13.2 percentage points. For households without children the predicted increase is smaller at 8.8
percentage points.29 Bethesda/Chevy Chase has the highest location probability outside of the District, and
increasing the school quality of Northwest D.C. decreases the probability that households with children
locate in Bethesda/Chevy Chase by 1.9 percentage points.30
Given the coefficient estimates from the indirect utility functions, one can calculate an estimate of
the average willingness to pay for school quality, in this case SAT points.31 Rewrite the estimated indirect
utility function as follows:
Vhj'fh(SAT)%gh(median rent)%X ,

(8)

where fh(SATj) is the piece of indirect utility that is a function of SAT scores, gh(median rentj) is the piece

29

For the District of Columbia, improving the quality of the public schools may have the additional
positive effect of helping to increase the tax base of the District. The population of D.C. has decreased
over the last several years. Although not all of the decrease may be attributed to poor school quality, these
results suggest that improving the quality of the public schools may encourage more households to remain
in the District.
30

The effects on location probabilities for all other choices will be even smaller because the
predicted location probabilities are small initially.
31

This method of calculating willingness-to-pay fails to recognize the possibility that households
may change their choices if one location improves its school quality. McFadden (1995) proposes
calculating mean willingness-to-pay using equivalent variation and suggests a way for measuring
willingness-to-pay by equivalent variation when indirect utility is non-linear in income. At this point, I
have not compared my calculations to the McFadden technique.

26
of indirect utility that is a function of median rent, and X represents everything else. I subscript the
functions by h since they will vary with household composition and income. Next, I calculate the
willingness to pay for SAT points as

Mmedian rent/MSAT using the implicit function theorem. This equals

!fh’(SAT)/gh’(median rent). Using the estimates from column (1) of Table 5,
!
!

fh)(SAT)
)

'

gh(median rent)
0.369%0.196 @ 1(child<18)%0.254 @ Q5%0.328 @ 1(child<18)@Q5
&0.191

(9)

for a household with children with income in quintile 5, where Q5 is an indicator for a household being in
the top income quintile. Since I am interested in the net willingness-to-pay relative to households without
children, I subtract off the willingness-to-pay for households without children. The following table
summarizes calculations of net willingness-to-pay for households with children for the top and bottom
income quintiles using the specifications with and without private school choice. I also include the average
estimated willingness-to-pay when school quality effects are not allowed to vary with income. The
discussion focuses on the results for white households because African-Americans in the bottom and top
income quintiles appear to have negative willingness-to-pay for school quality in the specification without
private school choice.32
For the estimates by income quintile, the standard errors are quite large and thus I cannot reject
that willingness-to-pay for school quality is equal to zero.33 This reflects the imprecision of the underlying
estimates, however, and when willingness-to-pay for school quality is calculated from the estimation
without variation by income quintile, the standard error is much smaller.

32

None of the willingness to pay calculations for African-American households are large or
statistically significant.
33

Standard errors were estimated using the delta method.

27
What Households with Children Are Willing to Pay per Year
for 100 SAT Points (Standard Errors)
White
Households

No Private
School
Choice

Allowing
Private
School
Choice

Income
quintile 1

$1,231
(1359)

Income
quintile 5

$3,298
(4026)

Average

$2,646
(1108)

Income
Quintile 1

$163
(5764)

Income
Quintile 5

$3,572
(2148)

Average

$1,700
(471)

Compared to households in the fifth income quintile without children, households in the fifth
income quintile with children would be willing to pay on average $274 more per month for 100 additional
SAT points, or $3,292 per year. This same calculation using the estimates from column (3) of Table 3
implies an average willingness-to-pay for SAT points of $2,646 per year per 100 points for households
with children in all income quintiles. This amount is net of the average willingness-to-pay for SAT points
for households without children. Households with children in the lowest quintile are estimated to be willing
to pay $1,231 per year more than households in the same income group without children. When the private
school choices within D.C. are included, net willingness-to-pay for school quality is estimated to be only
$163 per year for households with children in the bottom quintile and rises to $3,572 for households in the
top income quintile.34 For comparison, the average willingness to pay for school quality is $1,700 per year

34

For private schools I use the coefficient estimates from column (3) of Table 6 to calculate
willingness-to-pay for the first and fifth income quintiles.

28
per 100 SAT points.35
One way to consider the plausibility of these estimates is to look at the tuition costs and SAT
averages of private schools in the D.C. area. Private high school tuition is often more than three times this
$3,292 estimate and only falls below $4,000 for two high schools inside the District of Columbia. (Coerper
and Mersereau (1995)) When I include high schools located outside the District, there are several more
schools with tuition under $4,000, but many are still much higher. While I do not have SAT information
for these specific schools, nationally the SAT average for private independent schools is 93 points higher
than the national public school SAT average, and the SAT average for religious private schools is only 29
points higher (The College Board (1997a, 1997b)). D.C. Public Schools have average SAT scores below
the national public average so the gains to sending a child to private school may be closer to 160 points.
Given these facts, it is not implausible that parents in the top income quintile are willing to pay nearly
$3,300 to gain 100 SAT points.

VI. Conclusion
White households seem to exercise some school choice through location choice. On average,
households with children have a greater likelihood of moving to an area with higher SAT scores than
households without children. When I allow these effects to vary with income quintile, the coefficients are
not individually statistically significant, but I can reject that the income quintile effects for households with
children are jointly zero. While it is impossible to say whether the observed amount of choice is the "right"
amount of choice, the willingness-to-pay estimates seem plausible given the tuition costs of private schools
in the area.
The story for African-American households is much less clear. The counterintuitive results cannot

35

The estimates allowing for private school choice without variation by income category are not
shown in the paper, but they are consistent with the column (3) results of Tables 3 and 4.

29
be explained by simple income effects and likely result from a complex combination of income and
effective choice sets. While on the surface these results may suggest that an education voucher program is
useful, several other factors including tuition cost and race need to be taken into consideration.
In the fall of 1997, Congress was considering an education voucher program for D.C. for
households earning less than 185 percent of the poverty line (approximately $30,000 for a family of four in
199736). The District of Columbia Student Opportunity Scholarship Act of 1997 would have provided
vouchers paying a maximum of $3,200 toward tuition and fees (including transportation) at public, private,
or independent schools located in D.C., Montgomery County, Prince George’s County, Arlington County,
Fairfax County, Alexandria City, or Falls Church City. As noted above, there are only two high schools in
the District with tuition below $4,000 and only one has tuition below the $3,200 limit. While there are
more high schools in this tuition range outside of the District, transportation costs will not be insignificant.
In sum, it would have been difficult for the low income households that qualified for the program to find
affordable schools using the tuition voucher. In addition, if the African-American results on race reflect
preferences for living in areas with high percentages of African-Americans, these preferences may also
carry over to preferences in the racial composition of schools. If this is the case, expanding school choice
to more public and private schools in the area may not add many choices that parents really prefer.

36

Department of Health and Human Services (1997)

30
References
ADC of Alexandria, Inc. 1996. A Guide to Montgomery County Public Schools, School Years 1996-1997:
Secondary.
Armor, David J. and Brett M. Peiser. 1997. Competition in Education: A Case Study of Interdistrict
Choice. Boston: Pioneer Institute for Public Policy Research.
Berry, Steven T. 1994. "Estimating Discrete-Choice Models of Product Differentiation," RAND Journal of
Economics. 25(2) pp. 242-262.
Black, Sandra. 1997. "Do Better Schools Matter? Parental Valuation of Elementary Education." Federal
Reserve Bank of New York Research Paper No. 9729.
Bogart, William T. and Brian A. Cromwell. 1997. "How Much More is a Good School District Worth?"
National Tax Journal. 50(2) pp. 215-232.
Census of Population and Housing, 1990a: Public Use Microdata Samples U.S. [machine-readable data
files] Prepared by the Bureau of the Census. Washington, D.C.: The Bureau, 1992.
Census of Population and Housing, 1990b: Public Use Microdata Sample U.S. Technical Documentation.
Prepared by the Bureau of the Census. Washington, D.C.: The Bureau, 1992.
Department of Health and Human Services. 1997. "Update of the HHS Poverty Guidelines," Federal
Register. 62(46) pp. 10856-10859.
The College Board. 1997a. "High School Information," Profile of College-Bound Seniors National Report.
http://www.collegeboard.org/index_this/sat/html/cbttyp97.html.
The College Board. 1997b. "SAT I-Composite (V+M) Test Score Distributions and Percentiles for 1997
College-Bound Seniors, Total Test Takers," Profile of College-Bound Seniors National Report.
http://www.collegeboard.org/index_this/sat/html/totlvm97.html.
Coerper, Lois H. and Shirley W. Mersereau. 1995. Independent School Guide for Washington, D.C. and
Surrounding Area. Chevy Chase, MD: Independent School Guides.
District of Columbia Financial Responsibility and Management Assistance Authority.1997. Report on
D.C. Public Schools Performance. http://www.nubian.com/dcfra/newinfo/edreform3.html
Hausman, Jerry and Daniel McFadden. 1984. "Specification Tests for the Multinomial Logit Model,"
Econometrica 52(5) pp. 1219-1240.
Hoxby, Caroline Minter. 1994. "Does Competition Among Public Schools Benefit Students and
Taxpayers?," NBER Working Paper # 4979.
Lankford, Hamilton and James Wyckoff. 1997. "The Effect of School Choice and Residential Location on
the Racial Segregation of Elementary Students." Working Paper, State University of New York-

31
Albany.
Maryland Comptroller of the Treasury, Income Tax Division. 1984. 1984 Summary Report.
Maryland Department of Fiscal Services. 1989. Local Government Finances in Maryland, Year Ended
June 30, 1989.
McArthur, Edith, Kelly W. Colopy, and Beth Schlaline. 1995. "Use of School Choice," Education Policy
Issues: Statistical Perspectives. National Center for Education Statistics.
McFadden, Daniel. "Modelling the Choice of Residential Location," in Spatial Interaction Theory and
Planning Models. ed. by A. Karlqvist, L. Lundqvist, F. Snickars, and J. Weibull. New York:
North-Holland Publishing, 1978.
McFadden, Daniel. 1981 "Econometric Models of Probabilistic Choice," in Structural Analysis of Discrete
Data. Ed. By Charles Manski and Daniel McFadden. Cambridge: MIT Press.
McFadden, Daniel 1995. "Computing Willingness-to-Pay in Random Utility Models," Working Paper.
Department of Economics, University of California, Berkeley.
Musgrave, Richard A. 1939. "The Voluntary Exchange Theory of Public Economy," The Quarterly
Journal of Economics. February. 213-237.
National Center for Education Statistics, 1995. Digest of Education Statistics. Washington, D.C.: U.S.
Department of Education, Office of Educational Research and Improvement.
Office of Criminal Justice Plans and Analysis. 1990. 1990 Crime and Justice Report for the District of
Columbia. Government of the District of Columbia.
Park, Jin Heum. 1994. "Estimation of Sheepskin Effects and Returns to Schooling Using the Old and the
New CPS Measures of Educational Attainment" Industrial Relations Section Princeton University
Working Paper #338.
Rouse, Cecilia Elena. 1998. "Private School Vouchers and Student Achievement: An Evaluation of the
Milwaukee Parental Choice Program" Quarterly Journal of Economics. 113(2): 553-602.
Samuelson, Paul A. 1954. "The Pure Theory of Public Expenditure," The Review of Economics and
Statistics. 36(4) pp.387-389.
Samuelson, Paul A. 1955. "Diagrammatic Exposition of a Theory of Public Expenditure," The Review of
Economics and Statistics. 67(4) pp. 350-356.
Tiebout, Charles M. 1956. "A Pure Theory of Local Expenditures," Journal of Political Economy. 64(5)
pp. 416-424.
Wells, Amy Stuart, Cynthia Grutzik, Sibyll Carnochan Catalán, Julie Slayton, and Ash Vasudeva. 1998.
"Underlying Policy Assumptions of Charter School Reform: The Multiple Meanings of a

32
Movement," University of California, Los Angeles, mimeo.
Witte, John F., Cristopher A. Thorn, Kim M. Pritchard, and Michele Clairbourn. 1994. "Fourth-Year
Report: Milwaukee Parental choice Program," University of Wisconsin mimeo.
Witte, John F., Troy D. Sterr, and Christopher A. Thorn. 1995. "Fifth-Year Report: Milwaukee Parental
Choice Program." University of Wisconsin mimeo.
U.S. Bureau of the Census. 1987. 1987 Census of Governments Compendium of Government Finances.
Vol. 4. No. 5 U.S. Department of Commerce, Bureau of the Census.
U.S. Bureau of the Census. 1993. Government Finances: 1990-1991, Series GF/91-5. Washington, D.C.:
U.S. Government Printing Office.

33

Table 1
Selected Household Characteristics

D.C. in 1985 and D.C. in 1985 and
District of
D.C. MSA in
D.C. MSA in
Columbia 1985
1990
1990: Whites
(1)
(2)
(3)
Household Income

D.C. in 1985 and
D.C. MSA in
1990: AfricanAmericans
(4)

52,573
[53044]

52,813
[53782]

81,804
[73437]

38,911
[33131]

48.3
[17.3]

49.7
[17.2]

49.1
[17.5]

50.1
[17.1]

Householder’s education

13.2
[3.5]

13.0
[3.5]

15.1
[3.2]

12.0
[3.2]

% with children under 18

29.4
[45.6]

30.1
[45.9]

18.3
[38.7]

35.8
[47.9]

% with children under 10

22.6
[41.8]

22.9
[42.0]

14.2
[35.0]

27.0
[44.4]

% with children under 6

15.6
[36.3]

15.5
[36.2]

10.5
[30.7]

17.9
[38.4]

% African-American

63.2
[48.2]

67.6
[46.8]

% White

33.6
[47.2]

29.4
[45.5]

Householder age

% of households with
children in private school b
N

—
90.6a
[29.1]

—
—

10.2
[30.3]
(N=3880)

10.6
[30.9]
(N=3530)

25.3
[43.5]
(N=718)

7.1
[25.6]
(N=2812)

12,805

11,454

3,865

7,589

Notes: Household income in 1995 dollars. All means are weighted using household weights from the Census.
Columns (3) and (4) are not the estimation samples. See Appendix Table B1 for statistics on the samples used for
estimation. Standard deviations are in brackets.
a

The remaining households are: 4.0 percent Asian; 0.6 percent Native American, Eskimo, or Aleut; and 4.8
percent other.
b

Percentages are conditional on having children under 18 in the household. Number of households with children
is given in parentheses.

34
Table 2
Multinomial Logit Estimation of Location Probabilities for White Households
Only Interacting SAT with Child
Indicator

Interacting Child Indicator with
All Right-Hand-Side Variables
Direct Effects
(4)

Interaction Effects
(5)

0.525
(0.175)

0.431
(0.317)

0.474
(0.046)

—

—

-0.085
(0.087)

-0.215
(0.090)

-0.169
(0.105)

-0.217
(0.211)

—

-0.026
(0.093)

-0.309
(0.098)

-0.121
(0.113)

-0.546
(0.230)

Crime rate
(per 1 million people)

—

-0.038
(0.020)

-0.021
(0.020)

-0.054
(0.024)

0.091
(0.045)

Number of D.C. Metro
Stations

—

0.183
(0.037)

0.154
(0.038)

0.225
(0.046)

-0.222
(0.082)

Per capita county expend
net of education ($1000)

—

1.650
(0.078)

1.498
(0.096)

1.603
(0.112)

-0.554
(0.221)

Population per km2
(1000's of people)

—

0.264
(0.058)

0.290
(0.059)

0.293
(0.069)

0.068
(0.133)

Proportion of persons in
poverty

—

-15.486
(1.700)

-13.136
(1.783)

-16.001
(2.041)

7.911
(4.514)

Proportion owner
occupied housing units

—

-2.545
(0.443)

-1.504
(0.464)

-2.580
(0.561)

3.854
(1.017)

Proportion of the
population that is white

—

—

1.918
(0.213)

1.424
(0.250)

0.207
(0.477)

1(ed<h.s. graduate) x
proportion non-h.s. grads.

—

—

6.564
(0.772)

5.681
(0.871)

0.763
(1.792)

1(ed=h.s. graduate) x
proportion h.s. grads.

—

—

5.190
(1.021)

4.361
(1.139)

3.211
(2.704)

1(ed=some college) x
proportion some college

—

—

3.050
(1.482)

2.725
(1.710)

5.387
(3.789)

1(ed=college graduate)x
proportion college grads.

—

—

-0.227
(0.309)

-0.344
(0.336)

1.771
(0.899)

Number housing units
(100,000's of units)

0.364
(0.030)

0.478
(0.045)

0.450
(0.046)

0.455
(0.058)

0.003
(0.094)

Log Likelihood

-11441

-7105

-6971

SAT
(100’s of points)

(1)
-0.794
(0.029)

(2)
0.658
(0.139)

(3)
0.537
(0.147)

SAT x 1(child < 18)
(100’s of points)

0.299
(0.042)

0.358
(0.045)

Median Rent
($100)

0.275
(0.021)

Distance from D.C. in
miles

-6886

Notes: The dependent variable is an indicator for location choice. There are 26 choices and 3,685 households in
each estimation. Estimation results in columns (4) and (5) are from one estimation with all right-hand-side
variables interacted with an indicator for the household having children under 18. Standard errors are in

35
parentheses.

36
Table 3
Multinomial Logit Estimation of Location Probabilities for African-American Households
Only Interacting SAT with Child
Indicator

Interacting Child Indicator with
All Right-Hand-Side Variables
Direct Effects

Interaction Effects

SAT
(100’s of points)

(1)
-1.178
(0.041)

(2)
-0.520
(0.094)

(3)
0.027
(0.141)

(4)
0.133
(0.185)

(5)
-0.009
(0.293)

SAT x 1(child < 18)
(100’s of points)

-0.071
(0.034)

-0.084
(0.037)

-0.044
(0.037)

—

Median Rent
($100's)

-0.379
(0.027)

-0.387
(0.051)

-0.411
(0.055)

-0.504
(0.072)

0.071
(0.123)

Distance from D.C. in
miles

—

-0.948
(0.116)

-0.917
(0.122)

-1.203
(0.200)

0.356
(0.260)

Crime rate
(per one million people)

—

0.040
(0.017)

0.077
(0.020)

0.100
(0.027)

0.011
(0.043)

Number of D.C. Metro
Stations

—

-0.056
(0.032)

-0.094
(0.035)

-0.141
(0.049)

-0.001
(0.073)

Per capita county expend.
net of education ($1000)

—

0.229
(0.053)

0.600
(0.093)

0.763
(0.122)

-0.494
(0.192)

Population per km2
(1,000's of people)

—

-0.014
(0.038)

-0.068
(0.042)

-0.096
(0.058)

-0.062
(0.088)

Proportion of persons in
poverty

—

3.593
(1.238)

1.394
(1.269)

0.305
(1.824)

6.844
(2.703)

Proportion owner
occupied housing units

—

0.937
(0.438)

0.204
(0.479)

0.268
(0.673)

0.607
(1.003)

Proportion of the pop.
African-American.

—

—

0.958
(0.336)

0.563
(0.455)

1.105
(0.683)

1(ed<h.s. graduate) x
proportion non-h.s. grads.

—

—

3.667
(0.443)

3.595
(0.561)

0.185
(0.929)

1(ed=h.s. graduate) x
proportion h.s. graduates

—

—

4.785
(0.483)

4.433
(0.631)

-0.256
(0.997)

1(ed=some college) x
proportion some college

—

—

8.849
(0.909)

9.043
(1.210)

-1.272
(1.863)

1(ed=college graduate) x
proportion college grads.

—

—

1.826
(0.255)

1.764
(0.304)

-0.277
(0.603)

Number of housing units
(100,000's of units)

0.560
(0.055)

0.338
(0.079)

0.298
(0.079)

0.290
(0.105)

0.043
(0.159)

Log Likelihood

-14214

-13753

-13532

—

-13412

Notes: The dependent variable is an indicator for location choice. There are 26 choices and 7,374 households in
each estimation. Estimation results in columns (4) and (5) are from one estimation with all right-hand-side
variables interacted with an indicator for the household having children under 18. Standard errors are in

37
parentheses.

38
Table 4
Multinomial Logit Estimation of Location Probabilities Allowing for Differences
in School Quality Effect by Household Income Quintile

White
Households
(1)

AfricanAmerican
Households
(2)

SAT
(100’s of points)

0.369
(0.162)

-0.017
(0.148)

SAT x 1(child < 18)
(100’s of points)

0.196
(0.199)

-0.368
(0.100)

SAT x Q2
(100’s of points)

0.093
(0.088)

-0.128
(0.068)

SAT x Q3
(100’s of points)

0.164
(0.083)

0.074
(0.064)

SAT x Q4
(100’s of points)

0.204
(0.080)

0.010
(0.066)

SAT x Q5
(100’s of points)

0.254
(0.078)

0.121
(0.074)

SAT x 1(child<18) x
Q2
(100’s of points)

0.168
(0.245)

0.507
(0.128)

SAT x 1(child<18) x
Q3
(100’s of points)

0.488
(0.229)

0.201
(0.128)

SAT x 1(child<18) x
Q4
(100’s of points)

0.050
(0.221)

0.506
(0.123)

SAT x 1(child<18) x
Q5
(100’s of points)

0.328
(0.209)

0.203
(0.140)

Median Rent
($100’s)

-0.191
(0.090)

-0.402
(0.055)

Number of households

3,685

7,374

Log Likelihood

-6953

-13511

Notes: The dependent variable is an indicator for location choice. There are 26 choices for columns (1) and (2). In
all cases, SAT scores are measured in hundreds of points. Standard errors are in parentheses. Q2-Q5 are
indicators for the second through fifth income quintile. The quintiles are: household income less than or equal to
$14,880; between $14,880 and $29,940; between $29,940 and $46,980; between $46,980 and $77,750; and greater
than $77,750. The specifications in columns (1) and (2) also include: distance from central D.C., the crime rate,
number of D.C. Metro stations, per capita county and state expenditure, population density, the poverty rate, the
rate of owner occupancy, number of housing units, and the race and education interactions of Tables 3 and 4.

39
Table 5
Multinomial Logit Estimates Expanding the Choice Set to Include Some Private School Choice
Including HHs Choosing
Private in DC, Assigning
Them the Public Choice

Average Cost and
Average Quality
Private Option

High Cost and High
Quality Option

White HHs
(3)

AfricanAmerica
n HHs
(4)

White
HHs
(3)

AfricanAmerica
n HHs
(4)

White
HHs
(5)

AfricanAmerica
n HHs
(6)

SAT
(100’s of points)

0.381
(0.162)

-0.032
(0.148)

-0.395
(0.115)

-0.220
(0.082)

-0.366
(0.116)

-0.003
(0.086)

SAT x 1(child < 18)
(100’s of points)

0.195
(0.195)

-0.372
(0.099)

0.035
(0.169)

-0.396
(0.080)

0.141
(0.149)

-0.307
(0.073)

SAT x Q2
(100’s of points)

0.094
(0.088)

-0.125
(0.068)

0.094
(0.088)

-0.124
(0.068)

0.094
(0.088)

-0.126
(0.068)

SAT x Q3
(100’s of points)

0.163
(0.083)

0.077
(0.064)

0.161
(0.083)

0.081
(0.064)

0.161
(0.083)

0.077
(0.064)

SAT x Q4
(100’s of points)

0.201
(0.081)

0.012
(0.066)

0.190
(0.080)

0.017
(0.066)

0.190
(0.080)

0.014
(0.066)

SAT x Q5
(100’s of points)

0.251
(0.078)

0.124
(0.074)

0.229
(0.077)

0.131
(0.074)

0.230
(0.078)

0.126
(0.074)

SAT x 1(child<18) x Q2
(100’s of points)

0.152
(0.238)

0.486
(0.127)

0.145
(0.210)

0.439
(0.102)

0.103
(0.181)

0.370
(0.092)

SAT x 1(child<18) x Q3
(100’s of points)

0.435
(0.224)

0.164
(0.127)

0.408
(0.200)

0.227
(0.100)

0.232
(0.171)

0.174
(0.089)

SAT x 1(child<18) x Q4
(100’s of points)

0.003
(0.216)

0.473
(0.122)

0.072
(0.189)

0.441
(0.098)

0.029
(0.165)

0.341
(0.089)

SAT x 1(child<18) x Q5
(100’s of points)

0.133
(0.203)

0.164
(0.136)

0.504
(0.178)

0.416
(0.106)

0.387
(0.155)

0.327
(0.096)

Monthly Cost
($100’s)

-0.177
(0.089)

-0.390
(0.055)

-0.258
(0.049)

-0.418
(0.048)

-0.231
(0.049)

-0.352
(0.046)

Number of households

3,840

7,573

3,840

7,573

3,840

7,573

Notes: The dependent variable is an indicator for location choice. There are 26 choices in columns (1) and (2), and
31 choices for households with children and 26 choices for households without children for columns (3)-(6).
Monthly cost is monthly median rent for columns (1) and (2) and monthly median rent plus monthly tuition
payment for (3)-(6). Q2-Q5 are indicators for the second through fifth income quintile. The quintiles are:
household income less than or equal to $14,880; between $14,880 and $29,940; between $29,940 and $46,980;
between $46,980 and $77,750; and greater than $77,750. Each estimate also includes: distance from central D.C.,
the crime rate, number of D.C. Metro stations, per capita county and state expenditure, population density, the
poverty rate, the rate of owner occupancy, number of housing units, and the race and education interactions of
Tables 3 and 4. Standard errors are in parentheses.

Figure 1
Washington, D.C. Area PUMAs

DiscrK:t of Colombia

Calvert & St Mary's

40

41

Figure 2
Private School Enrollment Rates for D.C. Households by Race and by Income Decile

0.5

Proportion Enrolling Private

0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
1

2

3

4 5 6 7
Income Decile

8

9

Whites (DC Only)
African-Americans (DC Only)

10

42
Appendix A: Exploring the African-American Results
The findings for African-American households presented in section IV of the paper are quite
puzzling. In Table A1, I explore potential explanations for why these results do not square with economic
theory. As seen in Table 1 of the paper, the African-Americans have lower household income on average
than the whites; thus, the specifications in Table 4 may not fully capture the effect of income on location
decisions. Columns (1)-(3) of Table A1 explore the possibility that the results of Table 4 reflect the effects
of income. If the effect of SAT scores on location probabilities increases with income for households with
children, then the coefficient estimate for white households will be greater than that estimated for AfricanAmerican households because white households in the sample tend to be in the upper portion of the income
distribution. In column (1) of Table A1, the income deciles of the white households have been used to reweight the African-American households thus giving more weight to high income households. Comparing
the results to Column(3) of Table 4, the negative school quality effect only gets stronger and thus cannot be
explained by a simple income effect.
The income difference between the white and African-American samples may also reflect
differences in household structure. Thirty-six percent of African-American households in the sample are
male or female headed family households while only 8 percent of the white households fall in this
category.37 Column (2) of Table A1 combines the income deciles with household type to form weights,
again using the distribution of the white sample, to give more weight to married couple, non-family, and
wealthier households. Once again, the school quality results are counterintuitive. In column (3) I explore
the effects of income in a third way by estimating the specification only for households in the top 60
percent of the D.C. income distribution (53 percent of the African-American sample households). Again,
SAT scores as a measure of school quality have no significant effect on location probability.
Columns (4) and (5) explore the African-American results from two additional angles. Relocation
may be more costly for African-American households, perhaps because of factors such as housing
discrimination or the availability of public housing. Looking only at households that actually relocate,
either within or outside the District of Columbia, one might see that school quality affects location
probabilities. Column (4) re-estimates the model including only households that are living at a different
address than they were in 1985. This sub-sample contains 35 percent of the households in the original
estimation sample. Here, the direct effect of SAT becomes positive and statistically significant while the
SAT-child interaction becomes more strongly negative and statistically significant. Finally, column (5)
simply restricts the choice set to communities with more than 30 percent of the population identified as
African-American. This restriction eliminates 16 locations from the choice set but fewer than 5 percent of
the households from the original sample.38 In this estimate, the direct effect of SAT becomes large and
statistically significant—a maximal 40 percentage point increase in location probability per 100 SAT
points—but the SAT-child interaction remains negative and statistically significant.

37

Married couple households make up 27 percent of the African-American sample and 34 percent
of the white sample. The rest are non-family households.
38

Two location-specific explanatory measures have to be dropped from this estimation because
there are only 10 choices in the restricted choice set. In the specification shown, I have dropped population
density and proportion of housing that is owner occupied because the coefficient estimates are statistically
insignificant in the Table 4, column (3) results. The coefficient estimate on the SAT-child interaction is not
sensitive to the choice of excluded variables among those that have statistically insignificant coefficient
estimates in Table 4, column (3).

43
The results in Table A1 suggest that there is no simple explanation for the persistent negative
coefficient on the SAT-child interaction for African-American households. The explanation is likely to
require a more complex combination of income and social factors than I can accommodate in the model at
this time. However, the coefficient estimates on amenities in the column (5) specification suggest that the
more restricted choice set may be appropriate for African-American households.39 The coefficients on the
base SAT effect, the crime rate, the proportion of persons in poverty, and the number of metro stations all
have the predicted sign in this specification while in most other specifications the coefficient signs are
counterintuitive. This result could arise if the white communities are effectively not available to the
African-American households and the white communities have better than average amenities. If this is the
case, it would appear that African-American households have different preferences than expected when, in
fact, among the available choices African-American households prefer lower crime, better metro access,
etc.
Finally, I also estimate the Table 5, column (2) specification restricting the choice set to the
communities with at least 30 percent of the population African-American. These coefficient estimates are
presented in Table A2. The results are similar to the column (2), Table 5 results, but the base effect of
SAT becomes strongly positive and statistically significant and the SAT-child coefficient and SAT-childincome quintile interaction coefficients approximately double in magnitude. As a result, all of the SATchild-income interactions are positive and statistically significant. The second, fourth, and fifth income
quintiles for households with children show a small but net positive effect of school quality on location
probability relative to households without children. These results suggest that a combination of income and
limits on potential choice may explain the counterintuitive results for African-American households with
children although the evidence is not very strong.

39

Restricting the white sample to the choice set of communities with more than 30 percent of the
population white decreases the base SAT effect and increases the SAT-child interaction coefficient. The
effect of the proportion of persons in poverty becomes positive and statistically insignificant. All other
results remain relatively unchanged.

44
Table A1
Multinomial Logit Estimation of Location Probabilities for African-American Households
Household
Household
Income Above Moved Since
1985
40th Percentile
(3)
(4)
-0.196
0.314
(0.168)
(0.155)

Weight by
Income
(1)
-0.271
(0.182)

Weight by
Income &
Type
(2)
-0.051
(0.169)

SAT x 1(child < 18)
(100’s of points)

-0.087
(0.048)

-0.112
(0.055)

-0.022
(0.047)

-0.168
(0.045)

-0.109
(0.055)

Median Rent
($100)

-0.158
(0.072)

-0.289
(0.072)

-0.329
(0.068)

-0.445
(0.082)

-0.770
(0.269)

Distance from D.C. in
miles

-0.879
(0.179)

-0.992
(0.174 )

-1.157
(0.175)

-1.153
(0.135)

-1.023
(0.758)

Crime rate
(per one million people)

0.042
(0.026)

0.082
(0.024)

0.069
(0.024)

0.160
(0.026)

-0.068
(0.065)

Number of D.C. Metro
Stations

-0.099
(0.046)

-0.151
(0.045)

-0.127
(0.042)

-0.220
(0.040)

0.044
(0.033)

Total per capita county
exp. net of education

0.757
(0.123)

0.766
(0.118)

0.439
(0.110)

-0.200
(0.105)

1.010
(0.276)

Population per km2
(1000’s of people)

-0.010
(0.057)

-0.051
(0.055)

-0.083
(0.052)

-0.193
(0.049)

Proportion of persons in
poverty

0.342
(1.641)

0.866
(1.641)

2.644
(1.571)

8.489
(1.601)

Proportion owner
occupied housing units

0.675
(0.613)

0.647
(0.610)

1.078
(0.570)

0.833
(0.555)

Proportion of the pop.
African-American

0.766
(0.450)

0.524
(0.443)

0.791
(0.392)

1.494
(0.345)

1.393
(0.406)

1(ed<h.s. graduate) x
proportion non-h.s. grads.

3.368
(0.657)

3.110
(0.698)

3.476
(0.658)

4.269
(0.628)

3.067
(0.547)

1(ed=h.s. graduate) x
proportion h.s. graduates

4.955
(0.718)

5.335
(0.754)

4.437
(0.683)

3.377
(0.719)

4.929
(0.527)

1(ed=some college) x
proportion, some college

6.748
(1.236)

6.356
(1.234)

6.291
(1.116)

6.166
(1.124)

10.224
(0.999)

1(ed=college graduate) x
proportion college grads.

2.900
(0.303)

2.860
(0.291)

2.584
(0.316)

2.344
(0.411)

0.939
(0.326)

Number of housing units

0.505
(0.081)

0.527
(0.069)

0.299
(0.093)

0.255
(0.078)

3.006
(1.032)

Number of households

7,374

7,374

3,901

2,583

7,034

SAT
(100’s of points)

Community Pop.
$ 30% AfricanAmerican
(5)
1.342
(0.420)

—
-6.648
(2.722)
—

Notes: The dependent variable is an indicator for location choice. There are 26 choices in each of columns (1)-(4).
Column (5) includes only 10 choices. Standard errors are in parentheses.

45
Table A2
Multinomial Logit Estimation of Location Probabilities Allowing for Differences
in School Quality Effect by Household Income Quintile:
Restricted Choice Set
African-American
Households, Areas
with $ 30% AfricanAmerican
SAT
(100’s of points)

1.336
(0.424)

SAT x 1(child < 18)
(100’s of points)

-0.736
(0.138)

SAT x Q2
(100’s of points)

-0.168
(0.094)

SAT x Q3
(100’s of points)

0.104
(0.093)

SAT x Q4
(100’s of points)

-0.024
(0.097)

SAT x Q5
(100’s of points)

-0.123
(0.124)

SAT x 1(child<18) x
Q2
(100’s of points)

0.876
(0.179)

SAT x 1(child<18) x
Q3
(100’s of points)

0.443
(0.181)

SAT x 1(child<18) x
Q4
(100’s of points)

0.912
(0.177)

SAT x 1(child<18) x
Q5
(100’s of points)

0.784
(0.209)

Median Rent
($100’s)

-0.772
(0.138)

Number of households

7,034

Log Likelihood

-11465

Notes: The dependent variable is an indicator for location choice. There are 10 choices. SAT scores are measured
in hundreds of points. Standard errors are in parentheses. Q2-Q5 are indicators for the second through fifth
income quintile. The quintiles are: household income less than or equal to $14,880; between $14,880 and $29,940;
between $29,940 and $46,980; between $46,980 and $77,750; and greater than $77,750. The specification also
includes: distance from central D.C., the crime rate, number of D.C. Metro stations, per capita county and state
expenditure, the poverty rate, number of housing units, and the race and education interactions of Table A1.

46
Appendix B: Data
Selected mean characteristics for the estimation samples are listed in Table B1. Location
characteristics and households characteristics are listed by location choice in appendix tables B2a-B2e and
B3a-B3d, respectively. The sources for these data are described below.
1990 Census Public Use Micro Sample
Data on household income, householder education, household type, presence and age of children in
the household, public and private school enrollment, and place of birth come from the 1990 Census PUMS
5% Sample files. Household income is in 1995 dollars, and place of birth is an indicator for a choice being
located in the householder’s state of birth. Householder education in years is converted to a continuous
variable using the suggestion by Park (1994).
The 1990 Census PUMS files allow me to identify in which Public Use Microdata Area (PUMA)
or sub-PUMA a household is located.40 Most counties with populations greater than 100,000 can be
identified on the PUMS files, and for households living in many large population counties, one can identify
residential location by a smaller sub-PUMA area. Households living in counties with fewer than 100,000
residents can only be identified as living in a PUMA that consists of a group of counties and/or
independent cities. In the DC area, school districts are defined at the county or independent city level so for
large counties I can identify the exact public school district in which a household is living while for smaller
counties and cities I can identify a household as living in one of two or more public school districts. Three
large district–Washington, DC; Montgomery County, MD; and Prince George's County, MD–can be
broken up into between 5 and 7 sub-PUMAs each. Given their large populations, these counties are likely
to have a great deal more variation in community type and school quality within the school district; thus, in
my model I allow households to choose to locate to one of the several sub-PUMA areas in each of these
school districts. The maps in figures B1 and B2 show the boundaries of these areas with the names I have
associated with each. Fairfax County, VA is also large in population with approximately 819,000
inhabitants, yet, no sub-PUMA areas are defined, and in fact, the county is grouped with two independent
cities, Fairfax and Falls Church.
1990 Census Summary Tape File 1A
Data on percent of housing that is owner occupied, percent of the population that is AfricanAmerican, and percent of the population that is white were obtained from the 1990 Census Summary Tape
File 1A (STF1A) files. For sub-PUMAs, data were aggregated from the census tract level using census
tract to PUMA mappings from MABLE/Geocorr.41 For single- and multi-county PUMAs I use 1990
Census STF1A data at the county level.
1990 Census Summary Tape File 3A
Data on the percent of persons below poverty, median household income, median gross rent, and
education category proportions come from Census STF3A files. Again MABLE/ Geocorr census tract to
PUMA mappings were utilized in aggregating data to the sub-PUMA level and data for other PUMAs was
constructed from county level data. In calculating the various percentiles, I assumed households were

40

See Census of Population (1990b) for more details on PUMAs.

41

These mappings were developed by John Blodgett, senior Programmer/Analyst at the University
of Missouri, St. Louis, under contract with CIESIN/SEDAC.

47
uniformly distributed across the income and rent ranges used in the census.
SAT Scores
SAT score averages are for 1989 and come from the school districts themselves as well as annual
articles published in The Washington Post. For single district PUMA’s I use the district SAT average. For
location choices containing more than one school district, I use enrollment-weighted averages. In the case
of the Loudoun County group, scores were only obtained for the Fredericksburg City and Spotsylvania
County districts so the average for the group is based on those two districts alone. Finally, for sub-PUMA
areas I gathered information on high school boundary areas and assigned census tracts in whole or in part
to a school based on the school boundary definitions and the sub-PUMA area definitions. I assumed
population was uniformly distributed across a census tract and that high school students were similarly
distributed. Thus, I use the census tract and census tract portion populations to calculate a weighted
average of the high schools’ average SAT scores by sub-PUMA.
MABLE/Geocorr
Distance from D.C.: All distances are population weighted averages from 38.89° latitude, -77.02°
longitude.
Population Density: Land area in square kilometers calculated by MABLE/Geocorr.
Other Data
Crime Rates: FBI Uniform Crime Reports data. For all areas but D.C., crime rates are the average
from 1987 to 1990 of serious crimes per 100,000 population. Serious crimes include: murder and nonnegligent manslaughter, forcible rape, robbery, aggravated assault, burglary, larceny-theft, and motor
vehicle theft. These county-level data came from USA COUNTIES 1996 CD-ROM. For D.C. PUMAs,
PUMA-level crime rates were calculated from 1990 census tract level crime index data from Office of
Criminal Justice, Plans and Analysis, Government of the District of Columbia, 1990. "1990 Crime and
Justice Report for the District of Columbia."
Per capita county plus state expenditures: This is per capita total direct expenditure for individual
counties which is equal to aggregate county, municipal, and township expenditures less expenditures on
education.
Metro Stations: Counts are based on stations in existence in 1990. Stations that were within,
roughly, 0.5 miles of a border were assigned split shares in each bordering PUMA.
Private school tuition: Enrollment weighted, private school tuition was calculated using data from
Coerper and Mersereau (1995). The high schools and their enrollment and tuition are listed in appendix
Table B4.

Figure B1
Washington, D.C. sub-PUMAs

Northwest DC

Howard U./
GeorgiaAv.

48

49

Figure B2
Montgomery Co., Prince George’s Co., and Washington, D.C. sub-PUMAs

Olney

Takoma Pmk/Silver Spring

50
Table B1. Household Characteristics for Estimation Sample.
White Households

African-American Households

Public Only

Public and DC
Private

Public Only

Public and DC
Private

Household Income

77,842
[ 68979]

81,341
[72923]

38,284
[32491]

38,852
[33106]

Householder age

49.2
[17.8]

49.1
[17.5]

50.2
[17.2]

50.1
[17.1]

Householder’s education

15.1
[ 3.1]

15.1
[ 3.1]

11.9
[ 3.2]

12.0
[ 3.2]

% Education # 12 years, no
diploma

7.4
[26.1]

7.3
[26.0]

29.6
[45.7]

29.3
[45.5]

% High school graduates

11.4
[31.8]

11.0
[31.3]

35.4
[47.8]

35.3
[47.8]

% Some college

16.3
[37.0]

16.0
[36.6]

19.8
[39.8]

19.9
[39.9]

% Bachelor’s degree or
higher

64.9
[47.7]

65.7
[47.5]

15.2
[35.9]

15.5
[36.2]

% with children under 18

14.4
[35.1]

17.8
[38.3]

34.1
[47.4]

35.6
[47.9]

% with children under 10

12.3
[32.9]

13.9
[34.6]

26.1
[43.9]

26.9
[44.3]

% with children under 6

10.0
[29.9]

10.4
[30.5]

17.8
[38.3]

17.9
[38.3]

% of potential sample1

95.3

99.4

97.2

99.8

N

3685

3840

7374

7573

Notes: Standard deviations are in brackets. Public Only includes all households without children and all
households with children that have no children enrolled in private school. Public and DC Private includes
all of the households in the Public Only sample plus households with children locating in DC that enroll at
least one child in private school.
1
The potential sample is all households (white or African-American) that lived in DC in 1985 and in the DC
Metropolitan area in 1990.

51
Table B2a. Characteristics of Location Choices: Washington, D.C. PUMAs.

Northwest

Howard U. /
Georgia Av.

Northeast

Anacostia

Central DC

SAT

826

670

656

631

741

Median Rent

814

509

508

473

626

Distance

4.28

3.65

2.50

3.30

0.95

Crime rate per
100,000 persons

6,666

7,928

10,019

7,604

22,093

4

0.5

4

3.5

11.5

4925.61

4925.61

4925.61

4925.61

# of DC Metro
Stations
Total per capita
county expenditure

4925.61

Population Density

2632

5946

3596

3773

4511

% person in poverty

7.0

15.5

14.4

23.6

19.5

% owner occupied
housing units

50.7

44.2

47.8

29.1

28.5

% African-American

8.5

80.9

82.8

94.2

46.9

% white

85.8

11.4

15.7

4.8

44.8

% persons w/# 12
yrs educ, no diploma.

5.69

34.0

31.8

37.0

22.8

% persons w/ high
school diploma.

8.4

22.8

25.2

32.0

15.1

% persons w/ some
college.

14.8

21.2

19.7

21.1

15.8

% persons w/ a BA
degree or more

71.1

22.0

23.3

9.9

46.3

Median Household
Income

63,225

33,793

36,584

29,038

37,260

Tax rate

0.894

0.894

0.894

0.894

0.894

103,662

110,780

115,316

160,407

116,735

Population

Notes: The column headings are intended to give the reader a sense for the location of the PUMAs. They do not
reflect any "official" names of areas in the District. See Figure B1 for a map of their locations. Northwest is
PUMA 101, Howard U./Georgia Av. is PUMA 102, Northeast is PUMA 103, Anacostia is PUMA 104, and Central
DC is PUMA 105. SAT scores are population weighted averages of school averages for 1989. Income and rent
figures are in $1995. Tax rates are dollars per $100 valuation. Distance is measured in miles. Population density

52
is in persons per km2.

53
Table B2b. Characteristics of Location Choices: Montgomery County, MD.

Olney

Fairland /
White Oak

SAT

946

951

958

Median Rent

892

875

848

Distance

14.97

12.22

Crime rate per 100,000
persons

4127.5

4127.5

# of DC Metro Stations

0.0

0.0

Total per capita county
expenditure
Population Density
% person in poverty

Gaithersburg / Takoma Pk /
Rockville
Silver Spring

19.46

Damascus /
Poolesville

Bethesda /
Chevy Chase

962

992

1045

768

894

898

8.66

20.26

9.51

4127.5

4127.5

4127.5

4127.5

2.0

1.5

0.5

5.0

2583.59

2583.59

2583.59

2583.59

2583.59

2583.59

986

916

1327

2522

130

1446

2.7

4.0

5.2

6.5

2.2

3.6

% owner occupied
housing units

76.9

66.2

62.6

54.6

88.2

70.1

% African-American

11.5

21.5

11.2

21.5

5.3

3.4

% white

77.7

65.1

76.6

65.4

86.1

88.4

9.2

8.7

10.4

15.1

7.4

5.1

% persons w/ high
school diploma.

20.0

17.5

18.8

19.1

15.6

10.1

% persons w/ some
college.

26.3

25.2

27.0

23.5

24.4

17.8

% persons w/ a BA
degree or more

44.5

48.6

43.8

42.4

52.6

67.0

68,344

67,751

59,880

51,504

86,887

82,833

% persons w/# 12 yrs
educ, no diploma

Median Income
Tax Rate
Population

0.716
102,285

0.716
103,695

0.716
186,953

0.716
127,132

0.716
100,984

0.716
135,978

Notes: The PUMAs have been named to help give the reader a general sense of location. Refer to the maps in Figure
B2 for their location. Olney is PUMA 1201, Fairland/White Oak is PUMA 1202, Gaithersburg/Rockville is PUMA
1203, Takoma Pk/Silver Spring is PUMA 1204, Damascus/Poolesville is PUMA 1205, and Bethesda/Chevy Chase is
PUMA 1206. SAT scores are population weighted averages of school averages for 1989. Income and rent figures are
in $1995. Distance is measured in miles. Tax rates are dollars per $100 valuation. Population density is in persons
per km2.

54
Table B2c. Characteristics of Location Choices: Prince George’s County, MD.

Mt.Rainier
/Hyattsville

CollegePk/
New
Carrollton

Seat Pleasant
/ Capitol Hts

Laurel /
Greenbelt

Bowie/ Upper
Marlboro

Suitland /
Oxon Hill

Andrews
AF.B. and
south

SAT

782

803

750

851

842

749

781

Median Rent

685

776

708

793

907

745

851

13.37

13.02

Distance
Crime rate
# of DC Metro
Stations

5.81
5906

8.10
5906

6.88
5906

5.84

11.27

5906

5906

5906

5906

0.0

0.0

0.0

0.0

0.33

1.33

2.83

2697.69

2697.69

2697.69

2697.69

2697.69

2697.69

2697.69

2446

1759

1689

629

342

1517

181

8.1

8.6

9.9

4.2

2.0

5.5

2.2

% owner occ.
housing units

45.7

49.9

52.0

55.8

84.7

47.9

81.8

% Af.-Amer.

51.0

36.3

89.4

19.5

40.7

73.7

41.6

% white

35.9

55.5

9.3

72.3

55.8

22.0

53.6

%# 12 yrs ed., no
diploma

24.3

15.8

24.7

12.2

10.2

17.1

13.8

% high school
diploma.

28.1

27.8

36.1

23.5

25.5

34.0

30.1

% persons w/ some
college.

24.1

26.5

27.2

27.8

30.9

31.1

31.8

% persons w/ a BA
or more

23.6

30.0

11.9

36.5

33.4

17.9

24.3

42,628

47,851

44,357

56,077

71,335

48,643

65,944

Total per capita
county exp.
Population Density
% person in
poverty

Median Income
Tax Rate
Population

0.912
101,186

0.912
102,598

0.912
107,869

0.912
104,090

0.912
103,361

0.912
109,135

0.912
101,029

Notes: The PUMAs have been named to help give the reader a general sense of location. Refer to the maps in Figure B2 for
more detail. Mt.Rainier/Hyattsville is PUMA 1301, CollegePk/New Carrollton is PUMA 1302, Seat Pleasant /Capitol Hts is
PUMA 1303, Laurel/Greenbelt is PUMA 1304, Bowie/Upper Marlboro is PUMA 1305, Suitland/ Oxon Hill is PUMA 1306,
and Andrews AF.B. and south is PUMA 1307. SAT scores are population weighted averages of school averages for 1989.
Income and rent figures are in $1995. Distance is measured in miles. Tax rates are dollars per $100 valuation. Population

55
density is in persons per km2.
Table B2d. Characteristics of Location Choices: Other Maryland Locations
Calvert & St.
Mary’s Co.s

Charles
County

Frederick
County

SAT

888

872

924

Median Rent

652

805

651

Distance

40.78

22.07

43.37

Crime rate

2831

4299

3079

0.0

0.0

0.0

Total per capita
county expenditure

1934.95

1971.70

2170.39

Population Density

85

85

87

% person in poverty

6.5

5.0

4.8

% owner occ.
housing units

75.8

75.7

70.8

% AfricanAmerican

14.4

18.2

5.3

% white

84.0

79.3

93.1

%# 12 yrs ed., no
diploma

22.0

19.0

19.6

% high school
diploma.

36.1

36.6

34.0

% persons w/ some
college.

24.8

28.2

24.4

% persons w/ a BA
or more

17.2

16.2

22.0

Median Income

50,392

57,045

50,859

0.640

0.837

0.789

127,346

101,154

150,208

# of DC Metro
Stations

Tax Rate
Population

Notes: Calvert and St. Mary’s Counties, MD are PUMA 400, Charles County is PUMA 700, and Frederick County
is PUMA 900. See the map in Figure 1 of the paper for the exact locations. SAT scores are enrollment weighted
averages of school district averages for 1989. Income and rent figures are in $1995. Distance is measured in
miles. Tax rates are dollars per $100 valuation. Population density is in persons per km2.

56
Table B2e. Characteristics of Location Choices: Virginia Locations.

Alexandria

Fairfax Co.
Fairfax City
Falls Church
City

Stafford
King George
Fredericksburg
Caroline
Spotsylvania

909

928

983

903

867

818

894

667

12.65

48.14

Arlington Co.

Loudoun
Manassas
Manassas Park
Prince Williams

SAT

966

Median Rent

820

Distance
Crime rate
# of DC Metro
Stations

4.54

22.68

6.11

6049

3637

6880

3715

3032

5.0

0.0

2.5

1.5

0.0

Total per capita
county expenditure

2498.89

2056.73

2698.47

2326.07

1730.31

Population Density

2550

149

2811

815

47

% person in poverty

7.1

3.3

7.1

3.6

6.3

% owner occ.
housing units

44.6

71.2

40.5

70.5

74.9

% African-Amer.

10.5

10.3

21.9

7.6

14.4

% white

76.6

85.0

69.1

81.5

83.6

%# 12 yrs ed., no
diploma

12.5

13.1

13.1

8.7

24.6

% high school
diploma.

14.8

27.0

15.6

17.0

33.0

% persons w/ some
college.

20.5

31.4

22.7

25.4

22.9

% persons w/ a BA
or more

52.3

28.5

48.5

48.9

19.5

54,814

61,030

50,969

72,458

47,870

Median Income
Tax Rate
Population

0.690
170,936

1.181
336,506

0.930
111,183

1.012
847,784

0.730
170,410

Notes: Arlington Co. is PUMA 800; Loudoun Co., Manassas City, Manassas Park City, and Prince Williams Co. are PUMA
900; Alexandria City is PUMA 1000; Fairfax Co., Fairfax City, and Falls Church City are PUMA 1100; and Stafford Co.,
King George Co., Fredericksburg City, Caroline Co., and Spotsylvania Co. are PUMA 2200. See Figure 1 in the paper for
the locations. SAT scores are enrollment weighted averages of school district averages for 1989. Income and rent figures are
in $1995. Distance is measured in miles. Tax rates are dollars per $100 valuation. Population density is in persons per km2.

57
Table B3a. Characteristics of the Sample Households: Washington, D.C. PUMAs.

Northwest

Howard U. /
Georgia Av.

Northeast

Anacostia

Central DC

Household income

94,948
[85815]

44,391
[40891]

42,081
[35714]

34,797
[28628]

51,716
[53510]

Householder
education

15.51
[ 2.72]

12.09
[ 3.61]

11.94
[ 3.55]

11.69
[ 2.70]

13.30
[ 3.92]

%AfricanAmerican

7.54
[26.42]

86.79
[33.86]

89.38
[30.82]

96.77
[17.69]

50.28
[50.01]

% white

88.64
[31.74]

9.65
[29.54]

9.85
[29.80]

2.52
[15.68]

44.27
[49.68]

Householder age

54.54
[17.52]

54.24
[16.95]

54.79
[16.79]

48.49
[16.19]

49.41
[17.18]

%with children
under 18

16.06
[36.73]

29.44
[45.59]

29.34
[44.06]

42.67
[49.47]

18.96
[39.21]

%with children
under 10

10.83
[31.08]

22.06
[41.48]

19.99
[40.01]

31.65
[46.52]

14.46
[35.18]

% with children
enrolled in private
school a

45.36
[49.88]
(N=275)

12.13
[32.68]
(N=540)

9.03
[28.69]
(N=577)

5.64
[23.09]
(N=1137)

10.06
[30.12]
(N=351)

1758

1796

1931

2530

1871

# of households

Notes: The column headings are intended to give the reader a sense for the general area defined by the PUMAs.
They do not reflect any "official" names given to areas in the District. Refer to the map in Figure B1 for their
locations. Northwest is PUMA 101, Howard U. / Georgia Av. is PUMA 102, Northeast is PUMA 103, Anacostia is
PUMA 104, and Central DC is PUMA 105. Standard deviations are in brackets. Means are weighted using the
Census household weights. Income is in $1995.
a
Percentages are conditional on having children under 18 in the household. Number of households with children
is given in parentheses.

58
Table B3b. Characteristics of the Sample Households: Montgomery County, MD PUMAs.

Olney

Fairland /
White Oak

Gaithersburg/
Rockville

Takoma Pk /
Silver Spring

Damascus /
Poolesville

Bethesda /
Chevy Chase

55,934
[42843]

58,907
[38601]

52,241
[34235]

62,562
[47196]

120,489
[68364]

118,715
[87366]

Householder Education

14.14
[ 2.94]

14.71
[ 2.21]

14.05
[ 2.33]

14.68
[ 3.20]

15.35
[ 2.45]

16.23
[ 1.74]

%African- American

48.19
[50.70]

68.28
[47.10]

20.52
[40.74]

39.51
[49.04]

14.99
[36.54]

3.01
[17.15]

% White

44.22
[50.39]

27.57
[45.23]

65.06
[48.10]

54.52
[49.95]

71.66
[46.12]

95.62
[20.53]

Householder age

40.60
[14.73]

33.23
[11.35]

38.92
[15.70]

37.94
[12.26]

39.68
[10.63]

40.27
[12.53]

%with Child Under 18

46.67
[50.62]

40.51
[49.69]

32.79
[47.36]

37.37
[48.53]

58.55
[50.42]

40.12
[49.18]

%with Child Under 10

32.79
[47.63]

31.72
[47.10]

29.18
[45.87]

33.81
[47.46]

47.54
[51.11]

29.64
[45.82]

% with children
enrolled in private
school a

6.25
[24.91]
(N=18)

8.29
[28.38]
(N=18)

0
[.]
(N=0)

3.26
[17.91]
(N=59)

17.60
[39.78]
(N=12)

17.92
[38.69]
(N=58)

35

42

57

158

22

146

Household Income

# of Households

Notes: The PUMAs have been named to help give the reader a general sense of their location. Refer to the maps in Figure B2 for more detail. Olney is
PUMA 1201, Fairland/White Oak is PUMA 1202, Gaithersburg/Rockville is PUMA 1203, Takoma Pk/Silver Spring is PUMA 1204, Damascus/Poolesville is
PUMA 1205, and Bethesda/Chevy Chase is PUMA 1206. Standard deviations are in brackets. Means are weighted using the Census household weights.
Income is in $1995.
a
Percentages are conditional on having children under 18 in the household. Number of households with children is given in parentheses.

59
Table B3c. Characteristics of the Sample Households: Prince George’s County, MD PUMAs.

Mt.Rainier
/Hyattsville

CollegePk/
New
Carrollton

Seat
Pleasant /
Capitol Hts

Laurel /
Greenbelt

Bowie/
Upper
Marlboro

Suitland /
Oxon Hill

Andrews
AF.B. and
south

43,516
[21782]

45,829
[27548]

44,729
[25018]

51,393
[32397]

67,819
[31123]

43,025
[24153]

65,174
[24424]

Householder Education

13.52
[ 2.70]

13.14
[ 2.35]

12.80
[ 1.92]

14.42
[ 2.21]

13.99
[ 2.56]

12.81
[ 1.85]

14.37
[ 2.27]

%African- American

87.06
[33.65]

83.22
[37.59]

98.75
[11.15]

64.27
[48.69]

82.08
[38.74]

96.72
[17.86]

81.63
[39.19]

% White

5.89
[23.61]

12.56
[33.33]

0.91
[9.55]

32.01
[47.40]

17.92
[38.74]

2.40
[15.36]

14.18
[35.31]

Householder Age

36.97
[11.18]

37.16
[13.63]

35.36
[ 9.98]

34.10
[ 8.22]

41.60
[14.70]

36.72
[12.22]

42.08
[13.14]

%w/ Children Under 18

41.63
[49.43]

45.38
[50.09]

57.75
[49.54]

39.21
[49.60]

41.18
[49.71]

42.77
[49.63]

40.06
[49.60]

%w/ Children Under 10

35.55
[47.99]

41.16
[49.51]

47.57
[50.08]

33.81
[48.06]

31.89
[47.07]

33.70
[47.41]

25.56
[44.15]

% with children enrolled
in private school a

5.75
[23.43]
(N=82)

4.71
[21.47]
(N=40)

3.00
[17.14]
(N=106)

0
[.]
(N=14)

9.38
[29.83]
(N=22)

4.88
[21.69]
(N=76)

5.90
[24.24]
(N=18)

187

84

174

32

51

163

42

Household Income

# Households

Notes: The PUMAs have been named to help give the reader a general sense of their location. Refer to the maps in Figure B2 for more detail.
Mt.Rainier/Hyattsville is PUMA 1301, CollegePk/New Carrollton is PUMA 1302, Seat Pleasant /Capitol Hts is PUMA 1303, Laurel/Greenbelt is PUMA
1304, Bowie/Upper Marlboro is PUMA 1305, Suitland/ Oxon Hill is PUMA 1306, and Andrews AF.B. and south is PUMA 1307. Standard deviations are in
brackets. Means are weighted using the Census household weights. Income is in $1995.

60
a

Percentages are conditional on having children under 18 in the household. Number of households with children is given in parentheses.

61
Table B3d. Characteristics of the Sample Households: Other Maryland PUMAs and Virginia PUMAs.

Calvert & St.
Mary’s Co., MD Charles County
Household Income

Frederick
County

Arlington Co.

Loudoun
Manassas
Manassas Park
Prince
Williams

Alexandria

Fairfax Co.
Fairfax City
Falls Church

Stafford
King George
Fredericksburg
Caroline
Spotsylvania

76,309
[92359]

52,715
[36341]

61,271
[58333]

68,328
[44266]

53,902
[30824]

54,856
[26340]

84,321
[71379]

29,169
[15352]

Householder
Education

13.11
[ 2.42]

12.23
[ 2.37]

14.53
[ 2.43]

15.52
[ 2.43]

14.33
[ 2.16]

15.20
[ 2.01]

15.06
[ 2.58]

12.99
[ 3.32]

%African-American

36.07
[55.45]

27.56
[46.25]

18.38
[41.84]

9.57
[29.53]

27.33
[45.57]

22.74
[42.27]

26.11
[44.10]

78.05
[44.71]

% White

63.93
[55.45]

72.44
[46.25]

81.62
[41.84]

81.77
[38.76]

72.67
[45.57]

70.02
[46.20]

64.42
[48.07]

21.95
[44.71]

Householder Age

48.18
[15.70]

54.39
[19.19]

51.02
[19.03]

37.35
[12.23]

38.43
[10.22]

34.70
[10.91]

40.36
[11.98]

41.15
[9.04]

%with Children
Under 18

0

29.53
[47.22]

10.29
[32.82]

19.51
[39.77]

30.67
[47.15]

10.26
[30.60]

43.26
[49.74]

65.85
[51.22]

%with Children
Under 10

0

29.53
[47.22]

10.29
[32.82]

19.05
[39.42]

26.00
[44.85]

10.26
[30.60]

36.63
[48.37]

51.22
[53.99]

% with children
enrolled in private
school a

0

0
[.]
(N=5)

0
[.]
(N=1)

6.07
[24.33]
(N=28)

0
[.]
(N=7)

0
[.]
(N=7)

8.52
[28.17]
(N=54)

0
[.]
(N=5)

# Households

4

15

7

134

23

60

125

7

Notes: Maryland Counties: Calvert and St. Mary’s are PUMA 400, Charles is PUMA 700, and Frederick is PUMA 900. Virginia Counties: Arlington is PUMA
800; Loudoun, Manassas City, Manassas Park City, and Prince Williams are PUMA 900; Alexandria City is PUMA 1000; Fairfax County, Fairfax City, and
Falls Church City are PUMA 1100; and Stafford, King George, Fredericksburg City, Caroline, and Spotsylvania are PUMA 2200. See the map in Figure 1 of
the paper for the PUMA locations. Standard deviations are in brackets. Means are weighted using the Census household weights. Income is in $1995.

62
a

Percentages are conditional on having children under 18 in the household. Number of households with children is given in parentheses.

63
Table B4. Private Secondary Schools in Washington, D.C.
School Name

Annual Tuition

Archbishop Carroll H.S.

Grade Range

Total Enrollment

$4,000

9-12

700

$11,200

6-12

245

$8,000

9-PG

150

Field School

$11,500

7-12

180

Georgetown Day School

$12,925

9-12

1000

Georgetown Visitation Prep.

$8,400

9-12

403

Gonzaga College H.S.

$7,350

9-12

750

$16,565

K-12

250

Maret School

$12,820

K-12

500

National Cathedral School

$13,500

4-12

550

Nationhouse Watoto School

$2,750

N-12

85

Oakcrest School

$6,200

7-12

130

Parkmont School

$10,950

6-12

65

Sidwell Friends School

$13,020

5-12

1053

St. Albans School

$12,499

4-12

550

St. Anselm’s Abbey School

$9,100

6-12

198

St. John’s College H.S.

$6,290

7-12

510

Washington Academy

$6,800

9-PG

30

Washington Ethical H.S.

$9,700

9-12

40

$11,650

N-12

655

Edmund Burke School
Emerson Preparatory School

Lab School of Washington

a

Washington International School

Notes: All data in this table come from Coerper and Mersereau (1995). Where multiple tuition levels are given in
Coerper and Mersereau (1995), the highest tuition is listed above. Generally, tuition is higher for secondary school
students and higher for non-Catholic students at Catholic schools. MacArthur School is also located in D.C., but it
is not listed since it primarily serves students boarding at the Psychiatric Institute of Washington.
a
Lab School of Washington specializes in education for above average, learning disabled students, and tuition may
be funded through D.C. or Maryland if approved by the public school system. (Coerper and Mersereau (1995)).