old homes externalities and poor neighborhoods - 6-28...
TRANSCRIPT
Old Homes, Externalities, and Poor Neighborhoods
A Model of Urban Decline and Renewal
By
Stuart S. Rosenthal Department of Economics
426 Eggers Hall Syracuse University
Syracuse, NY 13244-1020 Phone: (315) 443-3809; Fax: (315) 443-1081
Email: [email protected] Home Page: http://faculty.maxwell.syr.edu/ROSENTHAL/
June 28, 2005
Funding for this project from the John D. and Catherine T. MacArthur Foundation, the Ford Foundation, and the Syracuse University Center for Policy Research is gratefully acknowledged. I am also grateful for helpful comments from Jan Brueckner, George Galster, Joe Gyourko, Don Haurin, Matt Kahn, Janet Kohlhase, Chris Mayer, Dan McMillen, and William Strange. I thank seminar participants at the University of British Columbia, University of Connecticut, University of Pennsylvania, University of Wisconsin, Fannie Mae, Federal Reserve Bank of Chicago, Federal Reserve Bank of Kansas City, Homer Hoyt Advanced Studies Institute, and the MacArthur Foundation. Excellent research assistance has been provided by Yong Chen, Ana Dammert, Michael Eriksen, Peter Howe, and Jerry Kalarickal. Any errors are mine alone.
Abstract This paper investigates the extent and nature of urban decline and renewal in the United States using three panels that follow neighborhoods on a geographically consistent basis over extended periods of time. In each decade, neighborhood economic status is measured as a community’s average income relative to that of its MSA. Findings indicate that change in neighborhood economic status is common: on average, neighborhood economic status changes roughly 13 percent per decade and two-thirds of neighborhoods studied in 1950 are of quite different economic status fifty years later. Neighborhood economic status also follows a stationary process, consistent with long-running cycles of decline and renewal. Aging housing stocks and redevelopment contribute to these patterns, as do local externalities. The latter occurs when individuals behave in ways that generate social capital and costs, and when status conscious households care about the attributes of their neighbors. In total, up to one quarter of the one-decade change in neighborhood economic status can be anticipated. Of that amount, externalities account for roughly 80 percent, while age of the housing stock accounts for the remaining twenty percent. Many of the neighborhood factors that drive change at the local level have large and policy relevant effects. Developers and city officials, therefore, can plan for these events.
I. Introduction
Urban decline and renewal are events that evoke passionate responses from neighborhood
residents and city officials alike. Yet although the economic vitality of neighborhoods is central to local
public policy and urban development, evidence on the extent and nature of neighborhood decline and
renewal is limited. This paper examines these issues. I begin with an observation.
Visit nearly any low-income urban neighborhood in the U.S. and it is apparent that poor families
occupy old homes built originally for higher income households.1 This simple fact implies that
neighborhoods cycle through regular periods of decline and renewal. These cycles arise because most
existing housing stocks deteriorate over time until they are replaced with newer dwellings. Housing
demand, meanwhile, increases with income. As a result, higher income families tend to move away from
older neighborhoods while lower-income families move in.2 Contributing to this migration is a second
quite different process based on local externalities. This occurs when individuals behave in ways that
generate social capital and costs (as with gardening versus crime), and when status conscious households
base their migration decisions on the attributes of their neighbors (as with race and income). Together,
aging housing stocks and local externalities imply that change in neighborhood economic status should be
common, at least over a sufficiently long time frame. Moreover, to the extent that change in
neighborhood economic status is rooted in the nature of housing demand, aging of the housing stock, and
social norms, it ought to be possible to anticipate and potentially to influence these events.
This paper explores these issues using a unique set of panels that follow individual
neighborhoods on a geographically consistent basis over extended periods of time. A series of closely
related questions are considered. Perhaps most fundamental, is change in neighborhood economic status
1The primary exceptions are low-income families living in homes built through various federal and local government subsidy programs. As of 1990, for example, there were roughly 1.3 million low income public housing units, but most of these had already begun to age. More recently, just over one million moderate income housing units have been constructed under the federal Low Income Housing Tax Credit (LIHTC) program between 1986 and 2004. But these units are too expensive to accommodate very low income families. With few exceptions, the private market does not build unsubsidized fixed site low-income housing (see Baer (1986), for example).
2
common or rare? Do neighborhoods cycle up and down in economic status? Can change in
neighborhood economic status be anticipated, and if so, what is the role of aging of the housing stock and
redevelopment versus externalities arising from social interactions? What are the neighborhood factors
that contribute most to change in neighborhood economic status? Each of these questions is considered.
We begin by examining evidence on whether change in neighborhood economic status is
common or rare. Tables 1a through 1c shed light on this issue. The tables present transition rates of
neighborhoods between different levels of economic status using a balanced panel of census tracts for 36
MSAs that are followed on a consistent geographic basis from 1950 to 2000. Each census tract is treated
as a separate neighborhood. Status is measured based on the average income of a neighborhood relative
to the average income of all census tracts in the balanced panel for the MSA and year in which the
neighborhood is observed. Neighborhood geography is coded to year 2000 census tract boundaries for all
years. In addition, both here and elsewhere in the paper, neighborhoods are classified into four groups
based on whether neighborhood relative income levels are in the first through fourth quartiles, referred to
below as low income, lower middle-income, upper middle-income and high-income, respectively.
Additional details on the data construction and sample composition are provided in Appendix A.
A striking pattern emerges from all three tables. For the 36 MSAs included in the sample,
roughly two-thirds of all low-income neighborhoods in 1950 (Table 1a, column 1) are of higher income
status in 2000. In Philadelphia (Table 1b) and Chicago (Table 1c), similar magnitudes are evident.
Among higher income tracts in 1950, change is also the dominant pattern in all three tables. For all 36
MSAs (Table 1a), only 27.62 percent of upper-middle income tracts in 1950 were still of upper-middle
income status in 2000. Overall, there is overwhelming evidence that most lower-income neighborhoods
tend to move up the economic ladder while most higher-income neighborhoods move down.
By how much does neighborhood economic status change with each passing decade? Table 2
provides evidence on this point by displaying the average absolute value of the change in Tract relative
2This “filtering” process has long been recognized as the dominant market source of low-income housing (e.g. Sweeney (1974)).
3
income on a decade-by-decade basis. Once again, a balanced panel of neighborhoods is used in
constructing these measures. For the 36 cities followed from 1950 to 2000, the ten-year change in
relative income status ranges from a high of 15.6 percent between 1950 and 1960, to a low of 12.8
percent between 1970 and 1980. From 1970 forward, 325 MSAs throughout the United States are
followed using data obtained from Geolytics Inc. (see the data appendix, Appendix A for details). For
this expanded cross-section, the same pattern holds: in each decade, the average change in neighborhood
relative income status is roughly 12.5 to 13 percent. These patterns make clear that change is the norm
among urban neighborhoods in the United States. Yet, dramatic as these patterns are, to my knowledge
this is the first time that these patterns have been reported. Possibly that is because most families remain
in their homes (and neighborhoods) far less than ten years, a horizon too brief for the change in
neighborhood economic status to be readily apparent to transient residents.3
Having established that most urban neighborhoods are in a constant state of transition, the
remainder of this paper is organized as follows. Details on the neighborhood panels used for the analysis
are provided in Appendix A, along with additional summary measures. Although creation of the data was
a major undertaking, it is sufficient here to indicate that for each panel, all of the neighborhoods are
followed on a geographically consistent basis over time, as in Tables 1 and 2. This leaves the body of the
paper free to focus on analysis and results.
Section II explores the time series properties associated with change in neighborhood economic
status. If neighborhoods cycle up and down around a stable long run mean, that would imply that
neighborhood economic status is a stationary process as opposed to a random walk. This idea is tested in
Section II. Results strongly indicate that neighborhood economic status is a stationary variable,
consistent with the idea that neighborhoods follow regular cycles of decline and renewal over extended
periods of time. Further analysis based just on Philadelphia County from 1900 to present indicates that a
typical neighborhood in that county takes roughly 100 years to complete a full cycle of change up and
3The median renter, for example, moves roughly every one to two years, while the median homeowner moves in six to seven years (Rosenthal (1988)).
4
down in relative income. That extended horizon is broadly consistent with the 10-year rates of change
displayed in Table 2.
Section III elaborates on the mechanisms that drive neighborhood economic change including its
cyclical nature. Two mechanisms are highlighted, aging of the housing stock along with redevelopment,
and local externalities arising from social dynamics. This section also describes the variables and the
empirical model used for the subsequent analysis of the determinants of neighborhood change. As will
become apparent, the empirical model builds off of the unit root regressions from Section II primarily by
adding controls for the age distribution of the housing stock and neighborhood socioeconomic factors.
Section IV presents estimates from the model. Results support the anticipated role of both
filtering and social dynamics. Among homes of different ages, “middle aged” housing, has the most
negative impact on the future economic status of a neighborhood, presumably because it is neither new
enough to attract higher income families nor older enough to be ripe for redevelopment. This is
consistent with the idea that aging of the housing stock contributes to the cyclical nature of neighborhood
decline and renewal. Even after controlling for the age distribution of the housing stock, however,
neighborhood economic status continues to display mean-reverting tendencies. This suggests that income
sorting also contributes to neighborhood economic cycles. One possible explanation is that households
tend to seek out neighborhoods with incomes higher than their own. Results further indicate that filtering
and social dynamics have largely independent effects, with social dynamics accounting for roughly four
times more of the one-decade change in neighborhood economic status as compared to aging of the
housing stock.
Estimates of the individual covariates are also important. For example, if all families in a renter-
occupied neighborhood were transformed into owner-occupiers, the neighborhood’s economic status
would rise by 22.7 percent over the following decade, all else equal. This is consistent with the idea that
because homeowners invest in their neighborhoods they have an incentive to behave in ways that benefit
5
the community.4 On the other hand, findings also suggest that if 10 percent of a middle-income
neighborhood’s housing stock was suddenly transformed into Low Income Housing Tax Credit (LIHTC)
housing, the economic status of the neighborhood would fall by 1.9 percent over the following decade,
ceteris paribus. This reinforces concerns that place-based subsidized rental housing and related spatial
concentration of poverty has adverse effects on neighborhoods.5
Controlling for these and other covariates in the model, up to one-quarter of the one-decade
change in neighborhood economic status can be anticipated. Policy makers and developers, therefore, can
plan for these events. Additional implications are outlined in Section 5, which concludes. We turn now
to the time series properties of neighborhood economic status.
II. Neighborhood Economic Cycles and Mean Reversion
It is clear that change in neighborhood economic status is the norm, not the exception.
Accordingly, this section examines the time series properties of neighborhood economic status. As will
become apparent, results shed further light on the cyclical nature of urban decline and renewal.
Consider first Table 3 which reports results from three regressions that characterize the degree to
which growth in neighborhood economic status is serially correlated. As before, neighborhood economic
status (yit) for neighborhood i in period t, is measured as the ratio of average income in the neighborhood
relative to average income in its MSA.6 For the first two regressions, the models are estimated using data
for just Philadelphia County for various decades from 1900 to 2000. All of the data for these regressions
are coded to the year 1900 Ward boundary geography, and there were 39 such Wards in Philadelphia in
4Partly in the hope of such outcomes the Clinton and George W. Bush administrations have both pushed to expand homeownership. 5Partly in response to these and related concerns, HUD has also developed voucher programs in which low-income family’s seek private market housing and use the government housing voucher to cover a portion of their rent. 6To be precise, let yit be the relative income of neighborhood i (i = 1, …, I) in period t. In addition, yit is defined to be Yit / tY , where Yit is the average level of income in tract i in period t, while tY is the city-wide average level of income in period t. By construction, the expected value of y over all neighborhoods in period t equals 1.
6
1900. It should also be noted that income data for 1900 and 1920 is based on a constructed measure
obtained from the IPUMs (see Appendix A for details) and is subject to more measurement error than the
income data from 1950 forward.7 Bearing that in mind, in the first regression, the period length is set to
50 years, and log(y2000) – log(y1950) is regressed on log(y1950) – log(y1900). In the second regression the
period length is set to 30 years, and log(y1980) – log(y1950) is regressed on log(y1950) – log(y1920). However,
for the third regression, the 36 MSA panel of census tracts is used, where each tract is followed from
1950 to 2000. In this regression, the period length is set to 10 years, and log(yt) – log(yt-1) is regressed on
log(yt-1) – log(yt-2).
Moving from left to right in Table 3, the period length for the corresponding regressions goes
from longest to shortest. In that order, the coefficients on the lagged dependent variables are -0.95, -0.45,
and -0.03, respectively, with t-ratios of 6.2 and 3.2, and 2.91.8 For the typical neighborhood in
Philadelphia County, a neighborhood’s relative economic status in 2000 is 95 percent back to where it
began 100 years earlier in 1900. As the period length narrows in the second and third regressions,
neighborhood cycles are less complete and the coefficients on the lagged dependent variables are reduced.
These results lend support for the idea that neighborhood relative income cycles up and down over
extended periods of time.
An implication of these results is that log(yit) has finite variance and a stable long-run mean even
when t is large – in other words, log(yit) is stationary. To test that idea, consider the following equation,
1 , 1log( ) log( )it o i t ity y e−= θ + θ + , (2.1)
where log(yit) is expressed as a function of a constant and its one period lag. A necessary condition for
log(yit) to be stationary is that 1 1.θ < Moreover, assuming log(yit) cycles up and down over time, with a
7Specifically, I used the IPUMs variable OCCSCORE to measure each individual’s income in 1900 and 1920. That variable is an estimate of the income the individual would have earned in 1950 given their actual occupation. 8The 36 MSA regression was also run including county fixed effects. This mirrors the specification for the Philadelphia County models for which county-wide effects are captured by the constant. Including the county fixed effects, the coefficient on the lagged dependent variable was equal to -0.06 with a t-ratio of 14.58.
7
sufficiently long time between periods t and t-1, 1θ should approach zero. This is because serial
correlation between current and future economic status will eventually damp out and the best estimate of
a neighborhood’s relative status far in the future is the mean of log(yi) which is captured by 0θ .9
However, under the null that θ1 equals 1, OLS estimates of (2.1) are biased.10 Accordingly, subtracting
log(yi,t-1) from both sides of (2.1),
1 , 1log( ) log( )it o i t ity y e−∆ = θ + θ + (2.2)
where , 1log( ) log( ) log( )it it i ty y y −∆ ≡ − and 1 1 1θ = θ − . The null of stationarity can now be expressed as
the hypothesis that 1 0θ ≠ .11 This is the standard Dickey-Fuller test for a unit root, OLS estimates of
which are consistent provided the error term is white noise. Observe also that 1 1θ = − implies that the
time span between periods t and t-1 is sufficiently long that past shocks have damped out entirely, in
which case the period t-1 value for neighborhood status has no impact on the contemporaneous value.
Table 4 presents four different estimates of equation (2.2). The first regression uses the
Philadelphia data from 1900 to 2000 and sets the period length to 50 years.12 The three remaining
regressions all use the 36-MSA sample from 1950 to 2000 with a period length of 10 years. Each of these
regressions also includes a time trend and census tract fixed effects.13 The regressions differ, however, in
the number of lags of the dependent variable included in the model: zero, one, and two lags, respectively.
9Allowing for the non-linearity of the log function, the long run mean for 0θ will also approach zero since the long run mean of yi is 1 by construction. 10When θ1 equals 1, log(yit) is sensitive to shocks far in the past, causing log(yit) to have an unstable mean and exploding variance as t becomes large. 11In the discussion below I emphasize equation (2.2), partly because that specification yields consistent estimates of
1θ even when θ1 equals 1, but also because equation (2.2) models changes in neighborhood relative income.
12In this case, each Ward contributes two observations to the model. Given the small number of time periods for each Ward, the regression in column one of the table does not include Ward fixed effects. When those fixed effects were added to the model the coefficient on log(yt-1) increased in magnitude to -1.8 and remained significant. 13Results are essentially unchanged with the time trend is omitted.
8
The lags are included to soak up serial correlation in the error term that might otherwise bias the
estimates. This is the standard Augmented Dickey-Fuller test for a unit root.
For each of the four regressions in Table 4, the estimate for 1θ is well away from zero and highly
significant. The evidence, therefore, soundly rejects the null of a unit root in favor of the alternative of
stationarity. This is consistent with the idea that neighborhoods cycle up and down in relative economic
status around a stable long-run mean.
III. Determinants of Change in Neighborhood Economic Status: Model
The discussion thus far has established two stylized facts: (i) change in neighborhood economic
status is common, and (ii) neighborhoods tend to cycle in economic status around a long-run mean over
extended periods of time. This section models the determinants of these patterns. At the outset, it should
be emphasized that not all factors that drive change in neighborhood economic status necessarily imply
mean reversion and cycling. Instead, as suggested earlier, two broad mechanisms likely contribute to
change in neighborhood economic status, aging housing stocks along with periodic redevelopment, and
local externalities that arise from social interactions. As will become apparent, there are compelling
reasons to anticipate that aging housing stocks contribute systematically to neighborhood economic
cycles. However, whereas local externalities also drive change at the neighborhood level, most do not
imply mean reversion, at least not directly. The primary exception may be that of income sorting to the
extent that families seek out neighborhoods with incomes higher than their own.
To allow for these different sources of neighborhood change, expression (2.2) from the previous
section is modified to include control variables for the age distribution of the housing stock and
socioeconomic factors that proxy for social interactions. Each of these sets of covariates is discussed
below, followed by a brief presentation of the full empirical model. I begin with aging of the housing
stock and neighborhood redevelopment.
9
Aging of the Housing Stock
Consider that cities tend to develop from the center outwards over time. Absent depreciation and
redevelopment, the oldest dwellings would be found in the city centers, with the youngest structures on
the urban/rural frontier. But this would lead to unusual sky-lines since, at the time of a city’s initial
development, land tends to be cheap and developers construct low-rise land-intensive buildings. The
more familiar sky-lines of our larger cities owe their form to periodic waves of redevelopment.
Two factors drive the urban redevelopment cycle. First, from the hedonic house price literature,
there is compelling evidence that most homes slowly deteriorate with time.14 As homes depreciate, their
flow of services declines and existing homes are eventually demolished and the site redeveloped. This
differs from economic “obsolescence.” As time passes, existing housing capital is at risk of becoming
economically obsolete even if it remains physically sound. This tends to occur in areas subject to sharp
increases in real land prices that create pressure to increase the level of housing capital installed on the
site. Under these conditions, physically sound buildings may be demolished and replaced with more
expansive structures (see Rosenthal and Helsley (1994), for example).
Because housing is a normal good, the slow decay (or increasing obsolescence) of existing
housing stock encourages higher income families to move away from neighborhoods filled with older
housing, ceteris paribus. Lower income families then take their place. This feature of urban housing
markets is at the core of what has come to be referred to as the “filtering” model. The central role of
filtering as the primary market source of low-income urban housing has been appreciated for many years.
The seminal theoretical work is often attributed to Sweeney (1973), while Rothenberg et al. (1991)
provides by far the most ambitious effort to examine filtering and related phenomena at an empirical
14Harding, Rosenthal, and Sirmans (2004), for example, recently examined depreciation of housing capital controlling for maintenance. They estimate the annual rate of depreciation for single family homes is roughly 1.5 percent per year. Other hedonic price studies that do not strip out the influence of maintenance typically estimate housing depreciation rates between 0.5 and 1 percent per year (e.g. Margolis (1982).
10
level. Nevertheless, it should be emphasized that despite broad consensus that filtering is the private
market source of low-income housing, there is an amazing dearth of careful empirical work in this area.15
Several implications follow from these arguments. First, land within a given neighborhood tends
to be developed in roughly the same time period, bearing in mind that time in this context is measured in
decades. This suggests that there should be more variation in the age of the housing stock throughout an
MSA than within individual neighborhoods. To confirm that premise, for each census tract in 2000,
estimates were developed of the standard deviation of the age of the housing stock within the tract
( HouseAgetractσ ) as well as throughout the MSA ( HouseAge
MSAσ ) to which the tract belongs.16 The distribution of
these measures across tracts and also the distribution of the difference between HouseAgetractσ and HouseAge
MSAσ
are reported in Table 5. At the median, HouseAgetractσ is 3.65 years less than HouseAge
MSAσ ; at the 10th percentile,
the differential is 9.47 years. In comparison, the average age of the U.S. housing stock in 2000 was 33.09
years. These estimates, therefore, confirm that homes within a given neighborhood tend to be of more
similar age relative to the homes throughout the broader MSA to which a tract belongs.
Because homes within a given neighborhood are of relatively similar vintage, this implies that
entire neighborhoods become ripe for redevelopment at about the same time. As a result, as homes within
a higher-income neighborhood age, the neighborhood itself slowly filters down. That decline, however,
is punctuated by increases in economic status as obsolete and dilapidated structures are demolished and
replaced with new units attractive to higher income families. These ideas imply a testable set predictions.
Specifically, the presence of newly built housing stock should elevate the future economic status of a
15Additional important theoretical papers on filtering include Ohls (1974), Brueckner (1977, 1980), Sands (1979), Bond and Coulson (1989), Braid (1995), and Arnott and Braid (1996). Important empirical studies include Jones (1978), Weicher and Thibodeau (1985), Baer (1986), Coulson and Bond (1990), and Aaronson (1999). Galster (1996) provides a nice review of several of these papers, along with a number of additional studies in this area. Related work by Brueckner and Rosenthal (2005) considers how aging of the housing stock and periodic redevelopment affect where high- and low-income families live within a given metropolitan area. Glaeser, Kahn, and Rappaport (2001) consider a different set of mechanisms that govern the location of poor families, while Glaeser and Gyourko (2001) study the effect of durable housing stocks on housing supply and rents. 16Estimates were formed using year 2000 data because particularly detailed information is available on the age distribution of the housing stock in that year.
11
neighborhood as higher income families are attracted by the high flow of housing services. “Middle-
aged” housing should be associated with future decline in neighborhood economic status since these
homes are not ripe for redevelopment but their service flow is declining. “Old” housing, in contrast, is
often ripe for redevelopment and the imminent construction of new housing. For this reason, old housing
should be associated with an increase in the future status of the neighborhood. Evidence in support of
these patterns would imply that filtering and redevelopment contribute to systematic cycles of
neighborhood decline and renewal. These ideas will be tested shortly.
Externalities
Neighborhood externalities also contribute to migration and related change in neighborhood
economic status. This likely occurs for two reasons: (i) certain types of families may behave in ways that
generate social capital and costs for the neighborhood, influencing demand for that location, and (ii)
families may choose to migrate into or out of a neighborhood based on the characteristics of their
prospective neighbors, irrespective of how those neighbors may behave. In this latter case, families are
concerned with the social status of the neighborhood. Each of these mechanisms is now considered.
Social Capital and Costs
If certain types of households behave in ways that provide social capital or costs, their presence
will affect the propensity of prospective neighbors to migrate into and out of the neighborhood. On the
positive side, three types of households seem especially likely to impart social capital on a neighborhood
in a manner that would attract higher income families, and thereby, elevate the neighborhood’s future
economic status. These are the presence of prime aged workers, the presence of educated individuals, and
the presence of homeowners.
12
Prime aged workers likely bring a stock of financial resources to a neighborhood as compared to
the elderly or the young. Such resources make it easier for families to maintain their homes, fund local
public projects, and in general provide a positive resource for the community. The presence of educated
individuals also brings resources into the neighborhood, except that in this case it is in the form of human
capital as opposed to financial wealth, all else equal. It is well known, for example, that educated
individuals commit fewer crimes and are more likely to be employed. Educated individuals also have
skills that can potentially be used to help address challenges that may be faced by the community. For all
of these reasons, children of educated individuals may be attractive playmates to parents deciding whether
to move into or out of a neighborhood.
A recent literature has also argued that homeowners make better citizens (see Rhoe 2000), for a
review).17 DiPasquale and Glaeser (1999), for example, provide evidence that homeowners are more
likely to know the name of their congressional representatives, are more likely to belong to neighborhood
groups, volunteer their time, garden, etc.18 More generally, homeowners by definition invest in their
neighborhoods. As such, they have a financial stake in preserving the overall quality of the neighborhood
to a degree that renters do not. Homeowners, therefore, have an incentive to engage in behavior that may
delay downward movement of higher income neighborhoods and accelerate upward movement of lower
income neighborhoods. To the extent that such outcomes are deemed desirable, this suggests a powerful
motive for encouraging homeownership in central city neighborhoods. Indeed, a host of public policy
programs strongly advocate homeownership as a means of invigorating hard pressed neighborhoods.19
17Green and White (1997) also argue that children of homeowners do better in school and have fewer teen pregnancies. Haurin and Haurin (2002) examine related issues. 18DiPasquale and Glaeser (1999) recognize that homeowners stay in their homes and neighborhoods longer than renters, and that length of stay could actually be the salient factor rather than homeownership. However, when DiPasquale and Glaeser control for length of stay, they still find evidence that homeowners pay more attention to their local communities than do renters. 19Cummings, DiPasquale, and Kahn (2002) describe a number of such programs as part of their study of the effects of Philadelphpia Nehemiah programs designed to promote homeownership in the center of the most hard-pressed neighborhoods. Cummings et al. note that the City of Philadelphia has “… long encouraged homeownership as part of its overall community development strategy …” Further, a primary goal stated in the strategic plan of the Office
13
Yet despite a long tradition in the United States of viewing homeownership as a public good, direct
evidence that homeownership produces spillover effects has been quite limited.
On the negative side, crime clearly imposes a social cost on a community. To the extent that
certain types of families commit more crimes, the presence of such households will discourage investment
in the community. That, in turn, will lower property values, attract lower income households, and reduce
the economic status of the neighborhood. Recent papers by DiPasquale and Glaeser (1999) and Glaeser
and Sacerdote (1999) suggest that cities tend to be subject to higher crime rates in part because criminals
are less likely to be apprehended in populous areas. More generally, this implies that densely developed
neighborhoods may be more subject to crime and, therefore, suffer lower expected future economic
status.
To allow for each of these factors, in the regression models to follow, controls are provided for
the density of development in the neighborhood, homeownership rate, the distribution of education
(percent less than high school, high school degree, some college, college degree or more), and the age
distribution of the population (percent age 15 or less, age 16 to 29, age 30 to 54, and age 55 and over).
Also included in the model is the neighborhood income inter-quartile range. For any given average level
of income in the neighborhood, greater variation in neighborhood income implies the presence of higher
income (and lower income) individuals. To the extent that higher income individuals serve as role
models and provide valuable contacts and information for the community, they generate social capital that
could elevate a neighborhood’s future economic status. It is partly for that reason that both HUD and
other philanthropic groups (e.g. the MacArthur Foundation) have begun to actively experiment with
programs designed explicitly to foster mixed income communities.20
of Housing and Community Development (OHCD) of the City of Philadelphia is “promoting homeownership and housing preservation. … to more effectively support economic development and reinvestment in Philadelphia, the City will continue to emphasize homeownership and preservation of the existing occupied housing stock” (OHCD 1997, p. 9; Cummings et al, 2002, p. 332). 20Also included is the percent of adults that are married since married families may view their neighborhoods differently than single-headed households, although whether this occurs and related effects seems ambiguous.
14
Social Status
If families choose to migrate into or out of a neighborhood because they care about a
neighborhood’s social status, this could further affect the future economic standing of the neighborhood.
Three instances in which this may occur stand out. These pertain to the racial composition of the local
population, the presence of various types of place-based subsidized rental housing, and also the level of
income in the community.
“White flight” was first used to describe the huge numbers of white central city households who
moved to the suburbs following the race riots of the 1960s. Implicit in the phrase is the idea that white
families did not want to live in close proximity to African Americans. Because minorities tend to be of
lower economic status than whites, sorting by race and ethnicity has an indirect effect on neighborhood
economic status. This process is often thought to be particularly sensitive to “tipping” points as discussed
by Schelling (1971) and others. To be precise, suppose there is a tipping point at which the white
majority begins to feel uncomfortable with the number of minority families in the community. Once that
tipping point is reached, a small further increase in minority presence could set off a self-reinforcing
chain of moves whereby white families leave the neighborhood to be replaced by minority households.
This could cause the racial composition of the neighborhood to shift dramatically towards minorities,
lowering the economic status of the community. To allow for such effects, in the regressions to follow,
controls are provided for the percent of the neighborhood’s population that is African American and the
percent that is Hispanic.21 In some of the specifications, additional controls are added to check for
nonlinearities.
A similar argument could apply to the presence of place-based subsidized rental housing if there
is a stigma associated with proximity to such projects. This could arise, in part, because federal law
mandates that these projects serve primarily low-income families, and as a result, higher income
households may shy away from neighborhoods in which such housing is situated. Bearing these issues in
21In the 1950 and 1960 data, Hispanics are defined as individuals whose county of origin was Mexico.
15
mind, controls are provided for the presence of two different subsidized housing programs that are
otherwise quite visible, traditional public housing and the more recent Low Income Housing Tax Credit
(LIHTC) program. The presence of each of these types of subsidized housing is expected to deter in-
migration of higher income families and reduce the future economic status of a neighborhood.22
Income sorting is another instance in which families choose neighborhoods based on the
attributes of their prospective neighbors. Beginning with Tiebout (1956) and Hamilton (1976), a large
predominantly theoretical literature has examined conditions under which low-income families attempt to
move to higher income neighborhoods in order to benefit from fiscal surpluses. More generally, if
individuals are status conscious, families of all income levels may have a tendency to seek out
neighborhoods with incomes higher than their own. Absent other considerations, such behavior would
contribute to neighborhood economic cycles as low- and high-income families migrate towards and away
from each other in succession. However, as noted by Hamilton (1976), residents eager to preserve the
status quo have an incentive to enact zoning laws that restrict in-migration of comparatively lower income
families (e.g. minimum lot size and square footage requirements). Enacting such laws requires
coordination and resources and is likely more common among the higher income communities.
These ideas suggest two testable hypotheses. First, even after controlling for the age distribution
of the housing stock and other socioeconomic factors, neighborhood economic status should continue to
display mean reverting tendencies. Second, mean reversion should be less pronounced among high-
income neighborhoods that have a greater ability to exclude comparatively lower-income households. To
22The public housing program began in 1937 and was the major federally subsidized construction program for low-income rental housing up until the mid-1980s. Federal law limits access to public housing to families whose incomes are low relative to their MSA, with the precise definition of “low” having changed somewhat over the years. As of 1990, over one million public housing units were in place. The LIHTC program, in contrast, began in 1986 and is now the major source of newly built federally subsidized low-income rental housing. As of 2000, there were roughly 850,000 LIHTC units in place nationwide; that number climbed to over one million by 2005. Under the LIHTC program, private developers receive a construction subsidy in exchange for a commitment to rent out a minimum percentage of units to low-income families. The minimum number can vary, but in practice, data from HUD’s website indicate that roughly 85 percent of all LIHTC projects are 100 percent low-income. To date, therefore, the LIHTC program has been a predominantly low-income as opposed to mixed-income housing program.
16
test these ideas, the lagged log level of neighborhood relative income, log(yt-1), is included in the
regressions to follow as in Section 2. The coefficient on this variable is expected to be negative as before.
In addition, some of the regressions to follow are further stratified by income quartile groups as in Tables
1a through c. Upon doing so, the influence of log(yt-1) is expected to be less for higher income groups.
Empirical Model
To evaluate the determinants of change in neighborhood economic status the unit root regression
in equation (2.2) is modified by adding controls for the various factors discussed above. The
specification is as follows:
, 1 , 1 2 , 1
3 1 , 1 , 1
log( )
log( ) log( )it MSA t i t i t
i i t i t it
y b HouseAge b SES
b Distance y y e− −
− −
∆ = δ + +
+ + θ + ∆ + (3.1)
where i denotes the census tract, and t is the “current” decade. As with (2.2), the dependent variable,
, 1log( ) log( / )it it i ty y y −∆ ≡ , is the one-decade change in economic status. The lagged level of
neighborhood economic status (log(yi,t-1)) is included to allow for mean reversion, while one lag of the
dependent variable is included to soak up serial correlation in the error term. HouseAget-1 is a vector that
describes the one decade lagged level age distribution of the housing stock, and SESt-1 is a vector of one
decade lagged level socioeconomic attributes of the neighborhood that control for local externalities.
Distance is included in (3.1) to allow for correlation in the location and timing of development, where
distance measures the number of miles to the census tract with the highest population density in year
2000. The model also includes a set of MSA fixed effects, δMSA,t, for each decade. These fixed effects
strip away unobserved factors common to tracts in a given MSA and given decade.23 Identification,
therefore, is based on within-MSA variation across census tracts for a given decade.
23This would include the city-wide level of income, the extent of racial segregation, fiscal policy, and broader macroeconomic conditions that affect immigration, job turnover, household mobility, and more. Racial segregation and discrimination, for example, has been well documented in the United States (e.g., Cutler, Glaeser, and Vigdor (1999)). If African Americans are restricted to a small number of neighborhoods, then such neighborhoods may not cycle in economic status in a manner comparable with majority white neighborhoods.
17
Three final points should be emphasized regarding the regression model in (3.1). First, the model
assumes that the covariates are exogenous. That assumption rests on the predetermined nature of the
covariates, including HouseAge and SES.24 Second, relatively few families remain in their homes and
neighborhoods longer than ten years.25 Accordingly, the use of ten-year lagged covariates helps to ensure
that the estimated effects are driven primarily by in and out migration as opposed to change in the well-
being of existing residents. Third, social norms can change over time, affecting the coefficients on the
SES variables. In addition, data on the age distribution of the housing stock is more detailed in the most
recent decades and the LIHTC program only began in 1986. For these reasons, the regressions presented
below will focus on the one decade change in neighborhood economic status from 1990 to 2000 for all
census tracts in defined MSAs throughout the United States. This provides the most up to date
assessment of what is driving change in neighborhood economic status today. Additional regressions that
draw on the longer 36 MSA panel of census tracts followed from 1950 forward are reported in Appendix
B. In most but not all cases, results are quite similar to estimates from the broader set of MSAs in the
1990 to 2000 period.
IV. Determinants of Change in Neighborhood Economic Status: Evidence
As a starting point, it is useful to consider the degree to which filtering and social dynamics have
independent effects on change in neighborhood economic status. To consider this question, Table 6
24I cannot rule out the possibility, of course, that one-decade lagged attributes may somehow be endogenous despite their predetermined nature. For that reason, in earlier versions of the paper I also estimated the models using two- and even three-decade lagged attributes of the neighborhoods as proxies for the one-decade lagged regressors. This further ensures that the regressors are exogenous, but this approach also introduces a considerable degree of noise. Not surprisingly, results were qualitatively similar to the models presented here but the estimated coefficients were generally smaller in magnitude. Estimates from a two-stage least squares model using the vectors of two-decade lagged attributes of HouseAge and SES as instruments also tended to produce very noisy results. Ultimately, the tradeoff appears to be noise on the part of the more deeply lagged instruments versus the possibility that one-decade lagged attributes may somehow be endogenous. Because the one-decade lagged attributes are predetermined, I emphasize those results in the paper. 25Recall that he median stay in the home for renters is roughly one to two years, while for homeowners the median is between six and seven years.
18
reports results from two estimates of model (3.1). Both models include the age distribution of the housing
stock (HouseAge) with the omitted category specified as the percent of homes age 50 or older. The
second regression also includes controls for the full set of SES variables, but the first regression omits
those control measures.26 To simplify review, only the coefficients on the house age variables are
provided. Comparing estimates across models, it is clear that omitting the SES control measures has little
impact on the coefficients associated with the age distribution of the housing stock. This suggests that the
influence of filtering and redevelopment on change in neighborhood economic status is largely
independent of the impact of SES factors and related local externalities. Considering the very different
conceptual foundations for how each influences a neighborhood – the degree to which housing is a
normal good as compared to the influence of social interactions – this result is appealing.
Table 7 reports the complete set of estimates for the model when both HouseAge and SES
controls are included in the regression. These results are provided in the first column. Also presented in
columns 2 through 5 are estimates from separate regressions having stratified census tracts into four
groups based on the economic quartile to which they belong in 1990: first quartile (low income), second
quartile (lower-middle income), third quartile (upper-middle income), and fourth quartile (high-income).
These groupings are defined in the same manner as in Table 1 at the start of the paper. Because in most
instances the coefficients in Table 7 are very precisely estimated, the discussion below will emphasize
primarily the patterns and magnitudes of the coefficients rather than statistical significance.
Consider first the estimates from column 1 for the All Tracts model and focus initially on the
coefficients on the age distribution of the housing stock. As before, the omitted category for this group is
the percentage of homes age 50 or older. Bearing that in mind, notice that the coefficient on homes age 0
to 4 years is positive 0.06, while the coefficients on the remaining age groups are all negative and vary in
magnitude from -0.04 up to -0.099. The dominant pattern, therefore, is that newly built housing enhances
the future economic status of a neighborhood, “middle-aged” housing has a negative effect, and old
26Both regressions also include controls for Distance, the lagged level of neighborhood economic status, and one lag of the dependent variable, as outlined in expression (3.1).
19
housing has an influence that is increasingly similar to new housing, presumably because it is ripe for
redevelopment. This is exactly what the filtering/redevelopment model predicts.
Observe next that density, measured as 1,000 housing units per square mile, has a negative impact
on the future economic status of neighborhoods with a coefficient of -0.003. To put that in perspective,
the average density across census tracts in 2000 was just over 2,000 housing units per square mile (see
Table A.3 in Appendix A). For the average census tract, therefore, density reduces the future economic
status of the neighborhood by only six tenths of one percent. But, for more urbanized areas with densities
often up to 10 times the national average, this effect rises to 6 percent. This is consistent with the idea
that high density development contributes to crime and other deviant behaviors that impose a social cost
on the community, lowering its future economic status.
In a similar vein, although public housing does not have a significant effect (this will differ in
some of the regressions to follow), the presence of LIHTC housing has a negative and significant impact
on the future economic status of neighborhoods. In Table 7, the presence of 100 LIHTC units would
reduce the one-decade ahead economic status of a neighborhood by 1.19 percent. Note also that the
median census tract in 2000 contained roughly 1,600 housing units, so 100 LIHTC units would be
equivalent to roughly 6.25 percent of the housing stock in the median neighborhood. The implied
elasticity of a neighborhood’s future economic status with respect to the introduction of LIHTC units is
19 percent (1.19 divided by 6.25), a large effect. On balance, these findings lend support to concerns that
construction of place-based subsidized rental housing hurts the future economic status of a neighborhood.
Given that the model controls for the neighborhood’s initial economic status and other socioeconomic
factors, this could reflect a stigma effect since LIHTC housing developments are quite visible in a
community.
The presence of homeowners increases the future economic status of a neighborhood.
Transforming a renter-occupied neighborhood into owner-occupied, all else equal, would cause the one-
20
decade ahead economic status of the neighborhood to rise by 22.7 percent. This is consistent with the
idea that homeowners invest in their communities and have incentives to behave in ways that enhance the
future economic status of the neighborhood, and their property values.
Analogous results are evident for individuals that bring resources to a community. Observe that
the presence of college degree individuals has a sharply positive impact on the future economic status of
the neighborhood: for every 10 percentage points of the local population with a college degree, the future
economic status of the community increases by 4.9 percent. Similarly, observe that the omitted age
category for the population is age 30 to 54, prime age workers. Relative to that group, all of the included
age categories have strongly negative effects. The presence of young workers (age 16 to 29) is especially
pronounced, with a coefficient of -0.629. The neighborhood income inter-quartile range is also positive
and significant. Compared to a neighborhood in which everyone has the same income, an income inter-
quartile range equal to $10,000 (in year 2000 dollars) would increase the one-decade ahead economic
status of the community by 1.8 percent.27 This is consistent with the idea that the presence of individuals
far above local median income may be a source of information (e.g. job networks) and resources for the
community. That idea underlies recent efforts by the Department of Housing and Urban Development
(HUD) and others to promote mixed income communities.28 Overall, these findings are consistent with
the premise that individuals that bring capital into a neighborhood impart positive local spillovers.
The influence of race is also striking. For both African American and Hispanics, the presence of
minorities significantly reduces the future economic status of a neighborhood. For every 10 percent of
the population that is black, the future economic status of the neighborhood falls by 10.8 percent; the
analogous estimate for Hispanics is 3.7 percent. Given the well-known lower average incomes of
27The average income inter-quartile range across tracts in 2000 was $45,902 (see Table A.3 in Appendix A). 28The LIHTC program, for example, stipulates that developers must rent at least 40 percent of their project units to lower income families. Similarly, the moving to opportunity (MTO) experiments conducted by HUD in five cities in the 1990s required low-income housing voucher beneficiaries to locate in middle income neighborhoods. Nehimiah experiments such as the ones in New York and Philadelphia also foster mixed income development by placing middle class homeowners in the middle of low-income neighborhoods. All of these programs have been developed in part with the hope that mixed income development will alleviate poverty traps.
21
minorities, these effects likely reflect the continuing influence of longstanding racial biases that affect the
location decisions of white and minority families.
Consider next the lagged log level of neighborhood relative income. The coefficient is negative
0.397 and is highly significant. As suggested earlier, having conditioned out the influence of the age
distribution of the housing stock and other socioeconomic factors, this is consistent with the idea that
families may seek opportunities to reside in communities with incomes above their own. Moreover,
looking across the quartile-stratified regressions in columns 2 through 5, it is apparent that the coefficient
on log(yt-1) declines monotonically in magnitude with the economic quartile to which a neighborhood
belongs. This is consistent with the idea that higher income neighborhoods are more able to enact local
zoning laws that restrict in-migration of comparatively lower income families.
House Age and SES Factors Stratified by Economic Quartile
Many of the remaining coefficients in columns 2 through 5 are also of interest. These are plotted
in Figures 1 through 6 to facilitate review. Although in most cases results are consistent with those from
the All Tracts model, as will become apparent, important differences also exist.
Figure 1 displays the coefficients on the age distribution of the housing stock as reported in each
column of Table 7. A separate line is provided for each sample group, including the “All Tracts” sample
from column 1, with the age of the housing stock along the horizontal axis. For each quartile, the U-
shaped pattern outlined earlier holds: new housing enhances the future economic status of a community,
middle aged housing hurts, and old housing has an effect that becomes increasingly similar to that of
newly produced stock, presumably because of the increasing likelihood of imminent redevelopment. The
main difference across quartiles is that the presence of middle-aged housing has a more negative impact
on the higher-income communities.
22
Figure 2a displays estimates for several of the socioeconomic factors, including the presence of
homeowners, college educated individuals, individuals age 16 to 29 (relative to individuals age 30 to 54),
African American, and Hispanic. A separate line is provided for each of these factors with neighborhood
economic quartile on the horizontal axis.
Once again, the broad results are consistent with those discussed earlier for the All Tracts model.
In addition, the estimated coefficients associated with the presence of homeowners and prime aged
workers are similar across quartiles. But for the other covariates differences are evident. Observe that
effects associated with the presence of college educated individuals diminish sharply with the income
quartile to which a neighborhood belongs, from over 0.6 in the first quartile to roughly 0.2 percent in the
fourth quartile. This suggests that those neighborhoods with the most to gain from the presence of
educated individuals are communities with the lowest initial economic status. This seems intuitive.
Differences with regard to the influence of race are also striking. Among low income
neighborhoods (the bottom quartile), the presence of African American families has a small negative
influence on the neighborhoods, while the presence of Hispanic families has a moderate sized positive
effect. For both groups, the influence of minority presence becomes increasingly negative with an
increase in the economic quartile to which a neighborhood belongs. For the highest income communities
(4th quartile), an increase in minority presence from 0 to 10 percent of the population would lower the
future economic status of the neighborhood by roughly 32.4 percent in the case of blacks and 37.9 percent
in the case of Hispanics. These estimates are consistent with racial bias in household location decisions
and the historic tendency of higher income white families to flee communities with an increasing minority
presence.
Figure 2b plots the coefficients for three additional variables of interest in Table 7, the density of
development, the presence of public housing, and the presence of LIHTC housing. Here too differences
appear across quartiles. High-density development has a similarly negative impact on the bottom three
quartiles with a magnitude roughly twice the impact reported for the All Tracts model. For the highest
23
income communities, however, density has a small positive effect. This may indicate that higher income
communities are able to shelter themselves from the negative impacts of high density development.
Similarly, public housing has a negative and significant impact on the lowest quartile neighborhoods: 100
additional public housing units reduces the future economic status of such neighborhoods by four tenths
of one percent. This translates into an elasticity of 6.4 percent.29 But that effect is not present for the
higher income communities. In sharp contrast, the influence of recently developed LIHTC housing has a
negative effect on all neighborhoods that increases sharply in magnitude and significance with the income
quartile to which a neighborhood belongs. Not surprisingly, this suggests that the negative effect of
place-based subsidized rental housing on a neighborhood’s future economic status increases with the
degree to which the project introduces families below the community’s mean income.
Nonlinear Effects
As discussed earlier, the influence of socioeconomic factors on the future economic status of a
neighborhood could be non-linear, as with “tipping” models. To explore that possibility, all of the
models in Table 7 were re-estimated converting a subset of the SES covariates into a series of 1-0 dummy
variables. This was done for the presence of African Americans, Hispanics, and the local homeownership
rate. It should be emphasized that these transformed regressors were all included in the regressions at
once, not sequentially.30 Results for the African American, Hispanic, and Homeownership measures are
plotted in Figures 3, 4, and 5, respectively. In each figure, a separate line is provided for each income
quartile group, while the intensity with which a given covariate is present is measured along the
horizontal axis. Estimates for the other covariates in the model were little changed from Table 7 and are
not reported.
29Recall that there are roughly 1,600 units in each census tract across MSAs in the U.S., on average. Then 100 public housing units is roughly 6.25 percent of the stock. Dividing 0.004 by 0.0625 gives 0.064. See also Carter, Schill, and Wachter (1996) for related work on the degree to which public housing affects nearby communities. 30While other covariates could similarly have been discretized, these three are likely to be of particular interest.
24
In Figure 3, observe that the African American variable is decomposed into a series of dummy
variables based on the share of the tract population that is black. The omitted category is 0 to 10 percent
black population, with additional dummy variables for 10 to 20 percent, 20 to 30, 30 to 40, and 50 percent
or more. For the first three economic quartile groups there is a downward sloping, convex pattern to the
plot, indicative of declining marginal effects from further increases in the presence of African Americans.
For the fourth quartile, however, effects appear to accumulate in a relatively linear fashion. Broadly
analogous patterns are evident with respect to the presence of Hispanic individuals as displayed in Figure
4. The exception is that the presence of Hispanics never appears to have a negative impact on first
quartile neighborhoods, and is positive for very high concentrations of Hispanics. On balance, these
estimates indicate that for most but not all communities, the initial presence of minority families has a
more pronounced impact on the economic status of the neighborhood than subsequent further arrivals
Figure 5 displays corresponding estimates for the local homeownership rate. The omitted
category in this case is 75 to 100 percent homeownership, while the included categories are 0 to 25
percent, 25 to 50, and 50 to 75. For low-income neighborhoods (1st quartile), future economic status rises
sharply and in a nearly linear manner with an increase in the neighborhood homeownership rate. Among
middle income communities, these effects increase at an increasing rate, although the pattern is largely
linear once ownership rates exceed 50 percent. Among high income communities (4th quartile), the effect
is approximately linear throughout. For policy makers, probably the important point to take note of here
is that introducing even a small concentration of homeowners into a low-income community is likely to
have a positive impact on the neighborhood’s future economic status. That result offers some
encouragement for those who hope to use homeownership as a mechanism to elevate the status of low-
income communities.
Forecasting Change in Neighborhood Economic Status
25
As a final exercise, this section considers our ability to forecast the one-decade change in
neighborhood economic status, and further decomposes that forecast into the contribution from filtering
and externalities arising from social dynamic factors. For each income quartile group, three different
regressions were compared. The first regression omits the HouseAge and SES variables while retaining
the other variables specified in the model.31 The second regression adds in the age distribution of the
housing stock (HouseAge), while the third is the complete model with the SES variables included.
Displayed in Figure 6 for each income quartile is the increase in adjusted R-square when the age
distribution of the housing stock was added to the second regression, and the further increase in adjusted
R-square when the SES factors were added. The total contribution from HouseAge and SES is the sum of
these two lines. The ratio of the contribution from HouseAge to that sum is also plotted as this provides a
measure of the relative impact of the age distribution of the housing stock in comparison to the impact of
the SES factors.
Several points are evident from Figure 6. First, the influence of the age distribution of the
housing stock edges up with the income quartile to which a neighborhood belongs, but only slightly.
Second, the influence of SES factors is highest for the middle income communities. Third, looking across
the quartile groups, on average, SES factors explain roughly 80 percent of the variation accounted for by
filtering and SES factors, with aging of the housing stock making up the remaining 20 percent.
Taking the complete set of results above together, it is clear that systematic factors enable one to
anticipate change in a neighborhood’s economic status. In Table 7, for example, the adjusted R-squares
indicate that among low-income neighborhoods (1st quartile), nearly one-quarter of the variation in one-
decade change in economic status can be forecasted. For the other quartile groups the analogous value
ranges between 14 and 15 percent. While a considerable amount of variation is not accounted for, these
R-square values are high enough to warrant the attention of policy makers and investors alike.
31These include Distance, the lagged log of neighborhood relative income, the lagged change in log relative income, and the MSA fixed effects.
26
V. Conclusions
Although urban decline and renewal shape U.S. cities as we know them, evidence on the extent
and nature of these events is limited. This paper has sought to fill part of that gap. The paper began with
an observation: Most low-income urban families in the U.S. occupy old homes built originally for higher
income households. That simple observation implies regular cycles of neighborhood decline and renewal
as homes age and are replaced. Superimposed on those patterns are the effects of a largely independent
process, neighborhood externalities. This occurs when certain types of individuals behave in ways that
generate social capital or costs, affecting demand for the neighborhood, and when individuals base
migration decisions on the attributes of their prospective neighbors. To explore these patterns, I
assembled a unique set of panels that follow neighborhoods on a geographically consistent basis over
extended periods of time.
Findings yield several stylized facts. First, change in neighborhood economic status is common.
Roughly two-thirds of urban neighborhoods studied in 1950, for example, were of quite different
economic status in 2000; on average, urban neighborhoods shift 13 percent in economic status with each
decade. Neighborhoods also cycle up and down over extended periods of time. Thus, most urban
neighborhoods are not static but will be of fundamentally different economic status twenty to forty years
in the future. Moreover, up to one quarter of the change in neighborhood economic status can be
anticipated. Policy makers and developers, therefore, can plan for these events.
Results also support the anticipated roles of filtering and local externalities that arise from social
dynamics. For example, a central prediction of the filtering/redevelopment model is that aging housing
stocks should cause a neighborhood’s economic status to cycle in systematic fashion. This is confirmed
in the data: relative to old housing, newly built homes boost the future economic status of a neighborhood
while middle aged housing has a negative effect. This likely reflects that middle aged housing is neither
new enough to attract higher income families, nor older enough to be ripe for redevelopment. Evidence
also suggests that even after controlling for the age distribution of the housing stock, neighborhood
27
economic status continues to display mean reverting tendencies that decline with the broader economic
class to which a neighborhood belongs. One possible explanation offered for this pattern is that status
conscious households may seek opportunities to live in neighborhoods with incomes above their own, and
that high income neighborhoods tend to adopt zoning laws that effectively restrict entry by lower-income
families.
Many of the estimated effects of socioeconomic factors are also economically important. For
example, if all families in a low-income renter-occupied neighborhood were transformed into owner-
occupiers, the neighborhood’s economic status would rise by 22.5 percent over the following decade, all
else equal. This is consistent with recent efforts by both Federal and local government to promote low-
income homeownership, in part motivated by the hope that homeownership will help to strengthen
communities.32 In contrast, if 10 percent of the current housing stock were transformed into Low Income
Housing Tax Credit (LIHTC) housing, neighborhood economic status would fall by 19 percent over the
following decade for the typical neighborhood, and by even more if the LIHTC housing was sited in
middle income communities. Additional sizeable effects are evident with regard to the presence of high
density housing, minorities, educated individuals, and prime aged workers.
As a final perspective, it is useful to take note of several issues that follow directly from this
paper but which are not addressed here. Evidence in this paper, for example, points to several policy
mechanisms that could potentially be used to boost a neighborhood’s economic status (e.g. promoting
homeownership, curtailing placed-based subsidized rental housing). However, it is important to
recognize that raising a neighborhood’s economic status does not necessarily alleviate poverty. Instead,
such efforts could simply force existing low-income residents to relocate to other parts of a city. This
32See, for example, Cummings, DiPasquale, and Kahn (2002) who analyze the results of recent Nehemiah programs in Philadelphia referenced earlier in the paper. In addition, a host of recent mortgage related policies, some in conjunction with the Community Reinvestment Act (CRA), are designed to boost homeownership among lower income urban neighborhoods.
28
issue recently received considerable public press when President Clinton located his office in Harlem.33
Similarly, policy makers are increasingly seeking ways to foster mixed income communities as part of the
broader set of anti-poverty programs. But results from this paper hint that many mixed income
neighborhoods may be unstable and are simply in transition from low to high, or high to low.34 Finally,
because cities tend to develop and redevelop from the center outwards over time, aging of the housing
stock has direct implications for the location of low- and high-income neighborhoods (e.g. central city
versus suburbs). Brueckner and Rosenthal (2005) have begun to explore this issue. While these topics
are all of pressing interest for the health and vitality of urban areas, they go beyond the scope of this
paper and are left for future research.
33While many of Harlem’s residents welcomed the President’s move as part of a recent revival of the area, others have voiced concerns about skyrocketing rents that are forcing long-term poor residents from their homes. See the USA Today (July 31, 2001, page 1), “I’m home’: Clinton opens Harlem Office,” by Charisse Jones. 34In this regard, it is useful to note that numerous theoretical papers on neighborhood sorting tend to focus on equilibrium relationships without modeling dynamics. See, for example, recent work by Epple and Romer (1991), Epple and Sieg (1999), and Epple, Romer, and Seig (2001), and related studies of sorting on the basis of income and tastes for local public goods.
29
References
Aaronson, Daniel, “Neighborhood Dynamics,” Journal of Urban Economics, 49, 1-32 (2001). Arnott, Richard and Ralph Braid, “A Filtering Model With Steady-State Housing,” Regional Science and Urban Economics, 27, 515-546 (1997). Baer, William, “The Shadow Market in Housing,” Scientific American, 255 (5), 29-35 (Nov. 1986). Bogue, Donald, CENSUS TRACT DATA, 1960: ELIZABETH MULLEN BOGUE FILE [Computer file]. ICPSR version. University of Chicago, Community and Family Study Center [producer], 1975. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2000. Bogue, Donald, CENSUS TRACT DATA, 1950: ELIZABETH MULLEN BOGUE FILE [Computer file]. ICPSR version. University of Chicago, Community and Family Study Center [producer], 1975. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2000. Bogue, Donald, CENSUS TRACT DATA, 1940: ELIZABETH MULLEN BOGUE FILE [Computer file]. ICPSR version. University of Chicago, Community and Family Study Center [producer], 1975. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2000. Bond, Eric and Edward Coulson, “Externalities and Neighborhood Change,” Journal of Urban Economics, 26, 231-249 (1989). Braid, Ralph, “The Comparative Statics of a Filtering Model of Housing with Two Income Groups,” 16, 437-448 (1986). Brueckner, Jan, “The Determinants of Residential Succession,” Journal of Urban Economics, 4, 45-59 (1977). Brueckner, Jan, “Residential succession and Land-Use Dynamics in a Vintage Model of Urban Housing,” Regional Science and Urban Economics, 10, 225-240 (1980). Brueckner, Jan and Stuart Rosenthal, “Gentrification and Neighborhood Cycles in a Model with Aging Housing Stocks: Low-Income Suburbs and Wealthy Cities?”, Working paper, 2005. Carter, William, Michael H. Schill, and Susan M. Wachter, “Polarization and Public Housing in the United States,” University of Pennsylvania, Wharton Real Estate working paper, (1996). Coulson, Edward and Eric Bond, “A Hedonic Approach to Residential Succession,” Review of Economics and Statistics, 433-443, (1990). Cummings, Jean L., Denise DiPasquale, and Matthew Kahn, “Measuring the Consequences of Promoting Inner City Homeownership,” Journal of Housing Economics, 11(4), 330-359 (2002). Cutler, David M., Edward L. Glaeser, and Jacob L. Vigdor, “The Rise and Decline of the American Ghetto,” Journal of Political Economy, 107 (3), 455-506 (1999). DiPasquale, Dennis, and Edward Glaeser, Incentives and Social Capital: Are Homeowners Better Citizens? Journal of Urban Economics 45(2): 354-84, (1999).
30
Glaeser, Edward, and Bruce Sacerdote, “Why is there More Crime in Cities?” Journal of Political Economy, 107 (6), 225-258, (1999). Epple, Dennis and Tom Romer, “Mobility and Redistribution,” Journal of Political Economy, 99 (4), 828-858, (1991). Epple, Dennis, Tom Romer, and Holger Sieg, “Interjurisdictional Sorting and Majority Rule: An Empirical Analysis,” Econometrica, 69 (6), 1437-1465, (2001). Epple, Dennis and Holger Sieg, “Estimating Equilibrium Models of Local Jurisdictions,” Journal of Political Economy, 107 (4), 645-681 (1999). Galster, George, “William Grigsby and the Analysis of Housing Sub-markets and Filtering,” Urban Studies, 33, 1979-1805 (1996). Geolytics website: www.geolytics.com . Glaeser, Edward L. and Joseph Gyourko, “Durable Housing, Urban Dynamics, and the Long, Slow Death of Declining Cities,” 2001. Glaeser, Edward L., Matthew E. Kahn, and Jordan Rappaport, “Why Do the Poor Live in Cities?,” National Bureau of Economic Research working paper 7636, April 2000. Green, Richard K. and Michelle White, “Measuring the Benefits of Homeowning: Effects on Children,” Journal of Urban Economics, 41, 441-461 (1997). Hamilton, Bruce W., “Capitalization Of Interjurisdictional Differences In Local Tax Prices,” American Economic Review 66, (1976), 743-753. Harding, John, C. F. Sirmans, and Stuart S. Rosenthal, “Depreciation of Housing Capital and the Gains from Homeownership.” Syracuse University working paper, 2004. Haurin, Donald R., Tom Parcel, and Jean Haurin, “Does Homeownership Affect Child Outcomes?” Real Estate Economics, 30, (2002). Jenkins, Robert M., Procedural History of the 1940 Census of Population and Housing, Madison: University of Wisconsin Press, 1985 Jones, Colin, “Household Movement, Filtering and Trading up Within the Owner Occupied Sector,” Regional Studies, 12, 551-561 (1978). Margolis, Stephen, “Depreciation of Housing: An Empirical Consideration of the Filtering Hypothesis,” The Review of Economics and Statistics, 64 (1), 90-96 (1982). National Historical Geographic Information Systems website: http://www.nhgis.org . Office of Housing and Community Development (OHCD), City of Philadelphia, 1997, “Year 23 Consolidated Plan (Fiscal Year 1998). City of Philadelphia Report: 9.
31
Ohls, James C., “Public Policy Toward Low Income Housing and Filtering in Housing Markets,” Journal of Urban Economics 2, 144-71 (1975). Rohe, W., G. McCarthy, and S. Van Zandt. 2000. The Social Benefits and Costs of Homeownership: A Critical Assessment of the Research. Research Institute for Housing America Working Paper No. 00-01 (May). Rosenthal, Stuart. “A Residence Time Model of Housing Markets,” Journal of Public Economics, 36, 87-109 (1988). Rosenthal, Stuart and Robert Helsley. “Redevelopment and the Urban Land Price Gradient,” Journal of Urban Economics, 35, 182-200 (1994).
Rothenberg, Jerome, George C. Galster, Richard V. Butler, and John Pitkin, The Maze of Urban Housing Markets: Theory and Evidence. Chicago, IL: The University of Chicago Press (1991). Ruggles, Steven and Matthew Sobek, Integrated Public Use Microdata Series: Version 2.0, Minneapolis: Historical Census Projects, University of Minnesota, 1997. Sands, Gary, “A Model for the Evaluation of Filtering,” Growth and Change, 20-24, (October 1979).
Schelling, T., “Dynamic Models of Segregation,” Journal of Mathematical Sociology, 1, 143-186 (1971). Shelley, Michael H., Ward Maps of United States cities: a selective checklist of pre-1900 maps in the Library of Congress, Library of Congress, Geography and Map Division, 1984 Sweeney, James L., “A Commodity Hierarchy Model of the Rental Housing Market,” Journal of Urban Economics 1, 288-323, (1974) Tiebout, C.M., “A Pure Theory of Local Expenditures,” Journal of Political Economy, 64, 416-424 (1956). Weicher, John, and Thomas Thibodeau, “Filtering and Housing Markets: An Empirical Analysis,” Journal of Urban Economics, 23, 21-40 (1988).
32
Table 1a. Transition Probabilities of Census Tract Relative Income Status Between 1950 and 2000 (Estimates are Based on 11,394 Census Tracts from a Balanced Panel for 36 MSAs)*
Low Income
in 1950 Lower-Middle Income in 1950
Upper-Middle Income in 1950
High Income in 1950
Low Income in 2000 35.71 17.71 20.79 20.38 Lower-Middle Income in 2000 28.02 25.04 23.52 18.50 Upper-Middle Income in 2000 22.82 31.41 27.53 21.17 High Income in 2000 13.45 25.84 28.17 39.95 Total Percent 100 100 100 100 *Tracts with median income less than the city-wide 25th percentile in the given year are defined as low income. Tracts with median income between the 25th and 50th percentiles are defined as lower-middle income. Tracts with median income between the 50th and 75th percentiles are defined as upper-middle income. Tracts with median income above the 75th percentile are defined as upper income.
Table 1b. Transition Probabilities of Census Tract Relative Income Status Between 1950 and 2000 (Estimates are Based on 1,292 Census Tracts from a Balanced Panel for PHILADELPHIA MSA)*
Low Income
in 1950 Lower-Middle Income in 1950
Upper-Middle Income in 1950
High Income in 1950
Low Income in 2000 28.40 13.13 17.59 16.05 Lower-Middle Income in 2000 27.78 30.31 27.78 19.75 Upper-Middle Income in 2000 28.70 32.19 29.01 24.07 High Income in 2000 15.12 24.38 25.62 40.12 Total Percent 100 100 100
100
*Tracts with median income less than the city-wide 25th percentile in the given year are defined as low income. Tracts with median income between the 25th and 50th percentiles are defined as lower-middle income. Tracts with median income between the 50th and 75th percentiles are defined as upper-middle income. Tracts with median income above the 75th percentile are defined as upper income.
33
Table 1c. Transition Probabilities of Census Tract Relative Income Status Between 1950 and 2000 (Estimates are Based on 1,477 Census Tracts from a Balanced Panel for CHICAGO MSA)*
Low Income
in 1950 Lower-Middle Income in 1950
Upper-Middle Income in 1950
High Income in 1950
Low Income in 2000 39.64 14.88 15.41 28.57 Lower-Middle Income in 2000 33.94 22.98 22.95 18.29 Upper-Middle Income in 2000 18.22 38.38 25.57 18.29 High Income in 2000 8.20 23.76 36.07 34.86 Total Percent 100 100 100 100 *Tracts with median income less than the city-wide 25th percentile in the given year are defined as low income. Tracts with median income between the 25th and 50th percentiles are defined as lower-middle income. Tracts with median income between the 50th and 75th percentiles are defined as upper-middle income. Tracts with median income above the 75th percentile are defined as upper income.
Table 2: Average Absolute Value of Percentage Change in Census Tract Relative Income By Decadea
Year 36 MSAs 1950 to 2000 331 MSAs 1970 to 2000 1950 - - 1960 15.9 - 1970 13.7 - 1980 12.8 12.6 1990 14.0 13.2 2000 13.3 12.5 aTract relative income in a given decade is equal to the average income in a census tract divided by the average income in its MSA.
34
Table 3: Serial Correlation in Growth in Neighborhood Relative Income (t-ratios based on robust standard errors in parentheses)*
Voting Wards in Philadelphia Countya Census Tracts in 36 MSAs, 1950-2000b
50-Year Period Length Dependent Variable: log(y2000/y1950)
30-Year Period Length Dependent Variable: log(y1980/y1950)
10-Year Period Length Dependent Variable: log(yt/yt-1)
log(y1950/y1900)
-0.9465 (-6.17)
log(y1950/y1920) -0.4538 (-3.20)
log(yt-1/yt-2) -0.0282 (-2.91)
Constant -0.0617 (-1.10)
Constant
-0.0423 (-0.82)
Constant -0.0219 (-28.35)
Observations 39 Observations 39 Observations 57,310 R-square 0.4329 R-square 0.1749 R-square 0.0008 aAll data were coded to year 1900 Ward boundaries. See Appendix A for details. bAll data were coded to year 2000 census tract boundaries. See Appendix A for details.
Table 4: Panel ADF Tests for Unit Roots in Neighborhood Relative Income Dependent variable: 1log( / )
t ty y−
(t-ratios based on robust standard errors in parentheses)
Philadelphia County Ward Panel
50-Year Period Lengtha 36 MSA Census Tract Panel
10-Year Period Lengthb
Zero Lags Zero Lags One Lag Two Lags
1log( )ty − -1.2139 (-6.22)
-0.5593 (-162.11)
-0.6492 (-120.88)
-0.7892 (-87.34
1log( )ty −∆ - - -0.0617 (-12.57)
-0.0806 (-9.63
2log( )ty −∆ - - - -0.0524 (-8.57)
Constant -0.0552 (-1.55)
- - -
Tract Fixed Effects - 18,369 17,316 16,625
Observations 78 75,193 57,310 40,213
Adj. R2 0.3772 0.3498 0.3750 0.4338 aAn alternate model with Ward fixed effects was also estimated. The coefficient on log(yt-1) was -1.8. bThe 36 MSA regression models include a time trend as well. Results based on models that omit the time trend were nearly identical to those reported here.
35
Table 5: Standard Deviation of Age of the Housing Stock in 2000 (in Years)a
Percentile Within Individual Census
Tracts Within MSA to Which a
Tract Belongs
Difference Between Individual Tract and the
MSA to Which it Belongs 1 5.37 13.78 - 13.70 10 9.87 17.06 - 9.47 25 13.28 18.87 -6.69 50 16.99 21.02 -3.65 75 20.35 22.86 -0.79 90 22.98 22.86 1.38 99 25.99 23.78 4.67 aThe average age of the housing stock in 2000 was 33.09 years.
36
Table 6: Independence of Filtering and SES Effects (Dependent Variable: 1-Decade Change in Log Relative Income ; t-
ratios based on robust standard errors in parentheses)
Covariates as of Decade t-1 Exclude SESa Include SESa
% Homes 0-4 yrs 0.0563 0.0643 (6.66) (7.35) % Homes 5-9 yrs -0.1435 -0.0994 (-11.67) (-8.32) % Homes 10-19 yrs -0.0797 -0.0412 (-10.49) (-5.46) % Homes 20-29 yrs -0.1613 -0.0947 (-19.45) (-11.66) % Homes 30-39 yrs -0.0925 -0.0582 (-11.73) (-7.33) % Homes 40-49 yrs -0.0793 -0.0586 (-6.18) (-4.72) Observations 48,630 48,628 MSA*Year Fixed Effects 325 325 Adj R-square 0.064 0.161 aThe model specification in the first column includes all of the variables reported in Table 7 except for the SES factors. The specification in the second column adds in the SES variables and is the same regression as reported in Table 7. Variables other than the house age distribution are not reported to conserve space. The omitted house age category is the percent of home 50 years and older.
37
Table 7: 1990 to 2000 Change in Neighborhood Economic Status By Income Quartile (Dependent Variable: 1-Decade Change in Log Relative Income;
t-ratios based on robust standard errors in parentheses) Neighborhood Economic Status in in 1990
Covariates as of Decade t-1 All Tracts 1st Quartile 2nd Quartile 3rd Quartile 4th Quartile% Homes 0-4 yrs 0.0643 0.0637 0.0315 0.0016 0.0296 (7.35) (2.07) (1.71) (0.10) (1.81)% Homes 5-9 yrs -0.0994 -0.0798 -0.0796 -0.1504 -0.1813 (-8.32) (-2.38) (-3.34) (-7.38) (-8.17)% Homes 10-19 yrs -0.0412 -0.0924 -0.0379 -0.0934 -0.1095 (-5.46) (-4.63) (-2.65) (-6.87) (-6.98)% Homes 20-29 yrs -0.0947 -0.0705 -0.0976 -0.1624 -0.1647 (-11.66) (-3.30) (-6.43) (-11.46) (-9.55)% Homes 30-39 yrs -0.0582 -0.0743 -0.0748 -0.1478 -0.0846 (-7.33) (-3.65) (-5.46) (-10.53) (-4.46)% Homes 40-49 yrs -0.0586 -0.1147 -0.1173 -0.0939 -0.0262 (-4.72) (-4.44) (-5.50) (-3.98) (-0.84)Density (1,000 units/mile) -0.0030 -0.0069 -0.0063 -0.0057 0.0009 (-12.05) (-11.50) (-11.17) (-8.97) (2.36)Public Housing (100s) -0.0002 -0.0040 0.0006 -0.0022 0.0018 (-0.26) (-3.37) (0.22) (-0.64) (0.29)LIHTC Housing (100s) -0.0119 -0.0068 -0.0118 -0.0243 -0.0344 (-3.24) (-1.14) (-1.82) (-3.09) (-2.47)Homeownership Rate 0.2266 0.2398 0.2433 0.2366 0.1608 (34.19) (15.23) (19.66) (18.33) (10.13)% High School or Some College -0.0149 0.2019 -0.0730 -0.0569 -0.1451 (-1.09) (7.17) (-2.74) (-1.83) (-3.17)% College Degree or More 0.4904 0.6643 0.4866 0.3471 0.1974 (38.06) (19.92) (18.62) (13.28) (5.50)% Adults Married -0.1712 -0.0087 0.0244 -0.1033 -0.2417 (-11.74) (-0.28) (0.87) (-3.57) (-6.67)% Age 15 or under -0.4287 -0.7512 -0.7633 -0.6669 -0.3214 (-13.63) (-11.49) (-11.55) (-9.96) (-4.02)% Age 16 to 29 -0.6288 -0.6720 -0.6188 -0.6450 -0.6404 (-31.78) (-15.68) (-15.07) (-15.57) (-15.30)% Age 55 or over -0.3811 -0.4270 -0.4990 -0.5212 -0.4618 (-19.17) (-9.21) (-12.66) (-13.91) (-10.47)% African American -0.1077 -0.0369 -0.0897 -0.1481 -0.3241 (-19.97) (-3.51) (-9.67) (-13.24) (-15.77)% Hispanic -0.0370 0.0914 -0.0424 -0.1557 -0.3787 (-4.02) (5.36) (-2.36) (-6.34) (-8.43)Income Inter-quartile Range ($1,000) 0.0018 -0.0001 0.0013 0.0015 0.0014 (22.68) (-0.24) (6.44) (9.26) (11.87)Distance to CBD (miles) 0.0012 0.0009 0.0008 0.0009 0.0014 (13.34) (4.52) (5.73) (5.24) (5.67)Log(relative income) -0.3967 -0.5846 -0.4345 -0.3763 -0.2529 (-60.32) (-34.51) (-17.45) (-15.63) (-20.65)∆ Log(relative income) -0.0864 -0.0972 -0.0328 -0.0672 -0.1356 (-14.70) (-7.69) (-2.56) (-5.42) (-12.48)Observations 48,628 11,982 12,036 12,349 12,261MSA Fixed Effects 325 325 325 325 325Adj R-square 0.161 0.242 0.145 0.152 0.142
38
Figure 2a: Impact of Socioeconomic Factors on One-Decade Change in Neighborhood Economic Status
-0.80
-0.60
-0.40
-0.20
0.00
0.20
0.40
0.60
0.80
1st 2nd 3rd 4th
Neighborhood Economic Status By Quartile
Mod
el C
oeffi
cien
t
Homeownership Rate % College Deg vs Less Than High School
% Age 16 to 29 vs Age 30 to 55 % African American% Hispanic
Figure 1: Impact of Age of Housing Stock on One-Decade Change in Neighborhood Economic Status
-0.20
-0.15
-0.10
-0.05
0.00
0.05
0.10
0 to 4 5 to 9 10 to 19 20 to 29 30 to 39 40 to 49 50 or more
Age of Housing Stock
Effe
ct R
elat
ive
to %
Hou
sing
> 5
0 Ye
ars
All Tracts 1st Quartile 2nd Quartile 3rd Quartile 4th Quartile
39
Figure 3: Nonlinear Effects of African American Concentration on One-Decade Change in Neighborhood Economic Status
-0.20
-0.18
-0.16
-0.14
-0.12
-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0 to 10 10 to 20 20 to 30 30 to 40 40 to 50 50 or more
Percent of Group in Neighborhood
Effe
ct R
elat
ive
to 0
to 1
0 Pe
rcen
t
All Tracts 1st Quartile 2nd Quartile 3rd Quartile 4th Quartile
Figure 2b: Impact of Density, Public, and LIHTC Housing on One-Decade Change in Neighborhood Economic Status
-0.040
-0.035
-0.030
-0.025
-0.020
-0.015
-0.010
-0.005
0.000
0.005
1st 2nd 3rd 4th
Neighborhood Economic Status By Quartile
Mod
el C
oeffi
cien
t
Public Housing (100s) LHITC Housing (100s) Density (1,000 units/sq mile)
40
Figure 4: Nonlinear Effects of Hispanic Concentration on One-Decade Change in Neighborhood Economic Status
-0.20
-0.15
-0.10
-0.05
0.00
0.05
0.10
0 to 10 10 to 20 20 to 30 30 to 40 40 to 50 50 or more
Percent of Group in Neighborhood
Effe
ct R
elat
ive
to 0
to 1
0 Pe
rcen
t
All Tracts 1st Quartile 2nd Quartile 3rd Quartile 4th Quartile
Figure 5: Nonlinear Effects of Homeownership on One-Decade Change in Neighborhood Economic Status
-0.14
-0.12
-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0 to 25 25 to 50 50 to 75 75 to 100
Percent Homeownership in Neighborhood
Effe
ct R
elat
ive
to 7
5 to
100
Per
cent
All Tracts 1st Quartile 2nd Quartile 3rd Quartile 4th Quartile
41
Figure 6: Relative Impact of Age of the Housing Stock (AHS) and SES Factorson Change in Economic Status as Measured by Effect on Adjusted R-Square
0.00
0.05
0.10
0.15
0.20
0.25
0.30
1st 2nd 3rd 4th
Neighborhood Economic Status By Quartile
Cha
nge
in A
djus
ted
R-s
quar
e
AHS Impact SES Impact AHS Share = AHS/(AHS + SES)
42
Appendix A: Data Creation and Sample Means
Analysis of the change in neighborhood economic status requires consistent data for small areas
within cities over an extended period of time. Such data have been lacking. To address that need, I
constructed a long-term panel that follows the characteristics of neighborhoods over time based on an
inter-temporally consistent set of geographic boundaries. These data combine census tract-level data on a
decade-by-decade basis from 1950 through 2000 for 36 metropolitan areas in the United States.35 An
additional panel was obtained from Geolytics Inc. that follows all identified census tracts in the U.S. on a
consistent geographic basis from 1970 to 2000. For this latter panel, it is important to recognize that
census tract coverage of the U.S. expanded over time and covered the entire country for the first time in
1990. These data were further augmented by a third panel for just Philadelphia County from 1900 to
2000. That panel follows each of the 39 year 1900 voting wards in Philadelphia on a consistent
geographic basis over time.36
Assembling these three panels was a large and complicated effort that centered on two tasks.
First, it was necessary to obtain information about the census tract characteristics in each year. This
feature of the data compilation effort was relatively straightforward and the various sources of data are
described in Table A-1 in this appendix.
Of a more challenging nature was the procedure used to create consistent geographies over time.
This enables one to follow individual neighborhoods from one decade to the next. This is necessary
because census tract geographic boundaries change with each decade.37 For 1950, 1960, and 1970,
correspondence tables from Census data pamphlets are available in printed format for each city in the
sample (see Table A1 for details). Those correspondences indicate the set of tracts from one decade that
35Beginning with 1950, Census adopted the census tract as the geographic unit used to organize the data. For additional details, see Procedural History of the 1940 Census of Population and Housing, by Jenkins, 1985. 36Related work coding census data to common geographic boundaries by the National Historical Geographic Information Systems project can be found at their website: http://www.nhgis.org. The data being processed as part of that project are still in Beta form and limited in scope.
43
comprise a tract in the following decade. The Census tables do not, however, report the degree to which a
tract from an earlier year contributes to one from a later year, only whether tracts from different years
overlap. Accordingly, all tracts from a prior decade that contribute to a current-year tract are given equal
weight when constructing the correspondence tables that map 1950 tracts to 1960 tract geography, and
1960 tracts to 1970 geography. This is an approximation, but it is the best that can be done for these
years.
For the years from 1970 through 2000, a commercially available data product from Geolytics Inc.
was used (www.geolytics.com). Geolytics has coded decennial census tract data for 1970, 1980, 1990,
and 2000 in terms of year-2000 tract boundary geography. This was accomplished by matching
geography using block-level boundary files available from Census for the last few decades. Because
block-level geography is considerably smaller than that of individual tracts, the correspondences for the
later decades are more precisely measured than for the earlier years.
Combining the Geolytics data with those that I created for 1950, 1960, and 1970 yields two
distinct sets of data: tract data for the years 1970 through 2000 coded to year-2000 Census tracts, and
tract data for the years 1950 and 1960 coded to year-1970 Census tract boundaries. It was necessary,
therefore, to convert the 1970 tract boundaries to those of year 2000 so that the entire dataset can be
specified based on the same set of geographic areas. Although Geolytics does not make available the
weight matrixes to make this conversion, they do make available an electronic map of the 1970 Census
tracts. Combining that map with an electronic map of the year-2000 Census tract boundaries from Census
it was possible to use mapping software to compute weights that convert 1970 tract boundaries into year-
2000 tracts. This was done using MapBasic, the programming language that underlies MapInfo, in
conjunction with MapInfo itself to match the geographies for 1970 and 2000.
The procedure above was used to produce data based on consistent census tract geography for
MSAs across the United States, with all tracts specified in terms of year-2000 tract geography. An
37The most common of these changes occurs when an existing tract is subdivided in subsequent years to allow for increased population.
44
unbalanced panel for these 36 metropolitan areas was then created using tracts that were identified by
Census for each given decade, from 1950 to 2000. This panel allows for changes in the geographic
boundaries of the MSAs, and also an expansion of the Census tract coverage that has progressed over the
years; 1990, for example, was the first year that the Census specified census tracts for the entire U.S. The
complete list of the metropolitan areas included among the 36 cities is provided in Table A2, along with
the number of census tracts identified in each metropolitan area in 1950 and 2000.
In addition to the 36 city panel described above, a second panel was created using all MSAs
identified in the Census tract data from 1970 to 2000. This panel has a more limited time span, but
provides a greatly expanded set of MSAs. This panel draws primarily on the Geolytics Inc. data file.
Sample means for the neighborhood attributes uses in the analysis from 1970 to 2000 for each of the two
primary panels are provided in Table A-3 for each decade separately.
The final data creation effort focused on Philadelphia County. For this one county, a balanced
panel of neighborhood attributes was created for various decades from 1900 to 2000, where
neighborhoods for this panel were converted to 1900 Ward-level geographic boundaries.
Correspondences that link the early decades with those from more recent years was made possible by very
detailed raster image maps that are available for the 1899 and 1914 Ward boundaries in Philadelphia
County (see Table A-1 for details). Those maps were converted to electronic boundary files using
MapInfo. Correspondences were then developed using MapBasic and MapInfo in order to code data for
all of the decades used in the Philadelphia County analysis from Section 2 (1900, 1920, 1950, 1980,
2000) to year 1900 Ward boundary geography.
An additional challenge in working with the 1900 and 1920 decades for Philadelphia is that
Ward-level income was not directly reported by Census. Instead, as a proxy, I used a constructed variable
created by the Integrated Public Use Microdata Series (IPUMs) (www.ipums.org) project referred to as
OCCSCORE at the IPUMs. This variable is an estimate of the income that an individual in a given
decade would have earned given their occupation if they were otherwise observed in 1950. This variable
45
was used to compute the relative Ward-level income in comparison to the Philadelphia County average
income for both 1900 and 1920. From 1950 and beyond, income was directly reported in the data and no
such approximation was required. Nevertheless, given the approximate nature of the income data from
1900 and 1920, some caution should be used in reviewing the Philadelphia County analysis in Section 2.
Additional Details
It should be noted that for many cities there was a major overhaul of the tract numbering
system prior to the 1970 census. Tract names had included letter suffixes prior to this time. For
the 1970 census and thereafter tracts were given numerical designations only. Another overhaul
of the tract numbering system occurred prior to the 1960 census. All tracts with multiple letter
suffixes used in the 1950 census were renumbered to eliminate them at this time. While no
documentation of the letter to number coding was found, in the 1960 data from ICPSR #2932
(see Table A-1) both forms of tract codes were provided. This allowed me to decipher the
format used for coding the letters into numbers in the other data sets. Letters “a” through “z”
were given values of 01 through 26, “aa” through “zz” were given values of 27 through 52, “aaa”
through “zzz” was assigned numbers 53 through 78. No letter suffix’s greater than “sss” was
found in any of the cities used in the sample.
Further details for each of the steps necessary to create the 1950 to 2000 data set are provided in
sequence below.
46
Tract Correspondences and Tract Attributes for 1950 and 1960
Tract level data for 1950 and 1960 were taken from the Bogue files (see Table A1). As part of
those files, each tract was referenced by an ID. Those codes were then used in conjunction with
comparability tables in the MSA-specific Census pamphlets (see Table A1) to recode the data to a
consistent set of geographic boundaries. For 1950, the first two columns of the Bogue files contain the
codes that identify the 1950 tracts. These tract ID columns look something like the following:
00090 9 00100 10 00110 11 00120 12 00130 13 00140 14 00150 15
This number is 17A in the tract comparability charts 171 17
This number will be read as 17B in the tract comparability charts. 00172 18 00173 19
The first column is the column of census tract reference ID numbers, referred to in the 1950
Bogue files as the “tractid.” The last number in the tractid is a subscript to the tract ID code. If the value
of the last number is 0, then it can be ignored. For instance, 140 is tractid 14. However, if the last
number is non-zero, then in the Census MSA pamphlets the non-zero code is represented in alphanumeric
code as follows: 171 in the Bogue files appears as 17A in the Census pamphlet, while 172 is 17B, and so
on. Note also, that the second column of numbers above are referred to as the “tractseq” in the 1950
Bogue files. These numbers are also used to identify the individual tracts in the Bogue data for 1950 but
they are not codes provided by Census per se. In the 1960 Bogue files census tracts are coded in a
manner very similar to the structure of the 1950 data.
Census Bureau pamphlets (see Table A1) were then used to identify the “tractid”
correspondences for those tracts whose boundaries changed during the decade. Tracts not listed in these
tables did not change boundaries during the decade and the 1960 tract remains the same as the
47
comparably named 1950 tract. These data were then used to convert all of the 1950 tract attributes into
1960 tract-level geography.
Tract Correspondences for 1960 and 1970
Converting data from 1960 tract geography to 1970 tract boundaries was much easier than
converting from 1950 to 1960 geography. In part, this is because comparability charts from the MSA-
specific census tract pamphlets for 1970 (see Table A1) provide the complete set of correspondences
between the tracts for 1970 and 1960.
Tract Correspondences and Attributes for 1970 to 2000
Data for 1970, 1980, 1990, and 2000 were obtained from the Neighborhood Change Database
Census CD marketed by Geolytics Inc. (www.geolytics.com). As discussed above, these data were all
precoded by Geolytics to year-2000 tract boundaries. In addition, as noted earlier, Geolytics did not
provide the correspondence weights necessary to convert year 1970 data to year 2000 tract geography.
However, Geolytics did provide electronic boundary files for 1970 tracts. These boundary files were
matched to the year 2000 tract boundary files to create the correspondences using mapping software
(MapBasic and MapInfo). This produced the correspondences that convert 1970 tracts to year 2000.
Ward Correspondences for 1900 and 1920
For Philadelphia County, I started with the Ward districts for 1900 and 1920. Raster image maps
showing the ward boundaries for these years were available over the web (see Table A1). These maps
were imported into MapInfo and oriented using county borders as references by overlaying a boundary
file for counties in the United States. The county boundary file was obtained from MapInfo (and
ultimately, Census). Ward boundaries were then converted to data using drawing tools in MapInfo.
48
Correspondence to 2000 tracts were then computed using mapping techniques in conjunction with year
2000 tract boundary files obtained from the Census website (see Table A1).
Public Housing and Low Income Hosing Tax Credit (LIHTC) Housing
Data on public housing units in 1990 were obtained from an analyst at the Department of
Housing and Urban Development (HUD). It is important to note that almost all of the public housing
stock ever built was still in existence in that year. The 1990 tract in which each public housing unit is
found was reported in the data, and subsequently mapped to year-2000 tract geography using 1990 to
2000 tract correspondence weights available at the Census website (see Table A1). Also reported in the
data was the year in which each unit was first available. Drawing on that information, it was possible to
determine the number of public housing units added to the stock in each decade, and as a result, the total
number of public housing units present in each decade for each census tract. Data for the Low Income
Housing Tax Credit (LIHTC) project were downloaded from HUD’s website (see Table A1). These data
also provide the year in which the units were completed, enabling me to determine once more the number
of such units present in each tract in each decade. The important caveat here is that the LIHTC program
only began in 1986 and, for that reason, the LIHTC variable is zero for all decades prior to 1990, as noted
in Table A-3.
49
Table A1: Data Sources Variables MSAs Data source 1900 Census Ward-level traits
Philadelphia IPUMS Minnesota Population Center, University of Minnesota, http://www.ipums.umn.edu
1920 Census Ward-level traits
Philadelphia IPUMS Minnesota Population Center, University of Minnesota, http://www.ipums.umn.edu
1950 Census tract-level traits
All 36 MSAs in Table A1
ICPSR #2931 Elizabeth Mullen Bogue file, http://www.icpsr.umich.edu/index.html
1960 Census tract-level traits
All 36 MSAs in Table A1
ICPSR #2932 Elizabeth Mullen Bogue file, http://www.icpsr.umich.edu/index.html
1970 to 1990 Census tract-level traits
All 36 MSAs in Table A1
Geolytics, 40-years CD, www.geolytics.com
1950 to 1960 census tract correspondence
All 36 MSAs in Table A1
U.S. Censuses of Population and Housing: 1960 Census Tract Final Reports. Washington, DC: U.S. Dept. of Commerce, Bureau of the Census
1960 to 1970 census tract correspondence
All 36 MSAs in Table A1
U.S. Census of Population and Housing: 1970 Census Tract Final Reports. Washington, DC: U.S. Dept. of Commerce, Bureau of the Census
1990 to 2000 census tract correspondence
All MSAs Census tract relationship files on the U.S. Census website: P:\urban_ssrosent\Filtering\correspondences\CensusCorr-1990-2000\Census 2000 Census Tract Relationship Files.htm
1900 and 1920 Ward maps in Raster format
Philadelphia http://www.library.upenn.edu/census/images/1899a.jpg
1970 boundary files for census tracts
All 36 MSAs in Table A1
Geolytics, 40-years CD, www.geolytics.com
1990 boundary files for census tracts
All MSAs U.S. Census Bureau Cartographic Boundary Files: http://www.census.gov/geo/www/cob/
2000 boundary files for census tracts
36 MSAs in Table A1
Geolytics, 40-years CD, www.geolytics.com
Public Housing data for 1940 through 2000
All MSAs These data were obtained from the Department of Housing and Urban Development (HUD).
LIHTC Housing Data All MSAs HUD website: http://www.huduser.org/datasets/lihtc.html
50
Table A2: 36 MSAs from 1950 to 2000
MSA Name Code Num Tracts in
1950 Num Tracts in
2000 Akron 80 93 166 Atlanta 520 170 660 Baltimore 720 492 621 Birmingham 1000 94 196 Boston 1120 639 699 Buffalo 1280 201 294 Chicago 1600 1610 1862 Cincinnati 1640 274 403 Cleveland 1680 525 706 Columbus 1840 314 372 Dallas 1920 351 698 Dayton 2000 196 241 Denver 2080 201 510 Detroit 2160 992 1265 Duluth 2240 44 83 Flint 2640 120 131 Hartford 3280 166 288 Houston 3360 673 772 Indianapolis 3480 216 339 Kansas 3760 205 493 Louisville 4520 98 241 Memphis 4920 4 272 Milwaukee 5080 366 414 New Orleans 5560 345 393 Philadelphia 6160 1549 1308 Pittsburgh 6280 369 702 Portland 6440 29 421 Providence 6480 62 259 Richmond 6760 75 252 Rochester 6840 112 264 St Louis 7040 356 524 Seattle 7600 322 527 Syracuse 8160 145 208 Toledo 8400 127 163 Trenton 8480 32 73 Washington 8840 399 1030
51
Table A3: Sample Means
36 MSA Sample All MSA Sample 1970 1980 1990 2000 1970 1980 1990 2000Relative Income 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000Log(Relative Income) -0.044 -0.058 -0.095 -0.098 -0.041 -0.050 -0.070 -0.073∆ Log(Relative Income) -0.022 -0.006 -0.034 -0.003 -0.017 -0.001 -0.022 -0.004% Homes 0-4 yrs 0.176 0.136 0.096 0.075 0.183 0.157 0.110 0.087% Homes 5-9 yrs 0.130 0.120 0.076 0.056 0.149 0.136 0.094 0.065% Homes 0-9 yrs 0.306 0.256 0.172 0.131 0.332 0.292 0.204 0.152% Homes 10-19 yrs 0.225 0.191 0.183 0.120 0.235 0.198 0.211 0.143% Homes 20-29 yrs 0.118 0.176 0.161 0.157 0.122 0.172 0.157 0.177% Homes 30-39 yrs 0.000 0.111 0.163 0.144 0.000 0.106 0.146 0.140% Homes 40-49 yrs 0.000 0.000 0.097 0.155 0.000 0.000 0.087 0.137% Homes 30 or more yrs 0.351 0.378 0.484 0.592 0.311 0.337 0.428 0.529% Homes 40 or more yrs 0.000 0.267 0.321 0.448 0.000 0.231 0.282 0.389% Homes 50 or more yrs 0.000 0.000 0.224 0.293 0.000 0.000 0.195 0.253Density (1,000 units/mile) 2.077 2.192 2.216 2.270 1.969 2.137 1.885 2.024Public Housing (100s) 0.115 0.162 0.179 0.182 0.097 0.140 0.154 0.156LIHTC Housing (100s) 0.000 0.000 0.031 0.165 0.000 0.000 0.026 0.133Homeownership Rate 0.676 0.655 0.641 0.650 0.673 0.653 0.654 0.662% Less Than High School 0.445 0.312 0.247 0.199 0.439 0.315 0.271 0.219% High School or Some College 0.431 0.504 0.511 0.515 0.440 0.511 0.519 0.534% College Degree or More 0.124 0.184 0.241 0.286 0.121 0.174 0.210 0.247% Adults Married 0.662 0.594 0.539 0.503 0.663 0.608 0.563 0.525% Age 15 or under 0.298 0.225 0.212 0.213 0.293 0.225 0.215 0.211% Age 16 to 29 0.236 0.274 0.230 0.204 0.236 0.273 0.227 0.204% Age 30 to 54 0.297 0.304 0.354 0.378 0.293 0.299 0.342 0.367% Age 55 or over 0.170 0.196 0.204 0.205 0.177 0.203 0.216 0.218% Not African American or Hispanic 0.866 0.813 0.772 0.711 0.852 0.812 0.794 0.744% African American 0.111 0.152 0.179 0.211 0.091 0.116 0.124 0.142% Hispanic 0.023 0.035 0.049 0.078 0.057 0.072 0.082 0.114Income Inter-Quartile Range ($1,000)a 38.522 36.375 43.604 49.657 37.166 35.099 41.013 45.902Distance to CBD (miles) 12.633 13.081 13.615 13.606 11.846 12.254 12.655 12.658Number of Tracts 16205 17238 17678 17709 44212 50906 64134 64328aData from each decade are in year 2000 dollars.
52
Appendix B Supplemental Tables
Table B1: One Decade Change in Neighborhood Economic Status for
36 MSA Panel, 1970 to 2000 (Dependent Variable: One-Decade Change in Log Relative Income ; t-
ratios in parentheses)
Covariates as of Decade t-1 Exclude SES Include SES% Homes 0-9 yrs 0.0628 0.0261 (3.24) (4.68)% Homes 10-19 yrs -0.0469 -0.0497 (-8.16) (-8.50)% Homes 20-29 yrs -0.1050 -0.0647 (-14.12) (-8.85)Density (1,000 units/mile) -0.0047 (-13.59)Public Housing (100s) -0.0081 (-9.42)LIHTC Housing (100s) -0.0015 (-0.28)Homeownership Rate 0.2277 (35.99)% High School or Some College 0.1674 (14.94)% College Degree or More 0.6779 (56.63)% Adults Married -0.2787 (-20.79)% Age 15 or under -0.4911 (-15.45)% Age 16 to 29 -0.7154 (-31.29)% Age 55 or over -0.4208 (-17.53)% African American -0.1418 (-30.56)% Hispanic -0.0159 (-1.33)Income Inter-quartile Range 0.0014 (19.43)Distance to CBD (miles) 0.0025 0.0020 (5.48) (18.59)Log(relative income) 0.0241 -0.3960 (7.77) (-55.63)∆ Log(relative income) -0.1197 -0.0531 (-20.44) (-9.11)Observations 47,948 47,837MSA*Year Fixed Effects 108 108Adj R-square 0.046 0.165