NBER WORKING PAPER SERIES
CRIME, URBAN FLIGHT, AND THE CONSEQUENCES FOR CITIES
Julie Berry Cullen Steven D. Levitt
Working Paper 5737
NATIONAL
BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 September 1996
We would like to thank Daron Acemoglu, Philip Cook, David Cutler, Jonathan Gruber, Jen-y Hausman, Lawrence Katz, John Lott, David Mustard, James Poterba, William Wheaton, and seminar participants at Harvard, MIT, and the University of Pennsylvania for helpful comments. We are gratefil to Rich Robinson of CIESIN for providing the Census data extracts. Comments can be addressed either to Julie Berry Cullen, Building E-52, Department of Economics, Massachusetts Institute of Technology, Cambridge, MA 02142, or to Steven Levitt, Harvard Society of Fellows, 78 Mount Auburn Street, Cambridge, MA 02138. e-mail addresses:
[email protected];
[email protected]. edu. The financial support of the National Science Foundation and the National Institute on Aging is gratefully acknowledged. All remaining errors are the sole responsibility of the authors. This paper is part of NBER’s research program in Public Economics. Any opinions expressed are those of the authors and not those of the National Bureau of Economic Research. 01996 by Julie Berry Cullen and Steven D. Levitt. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including O notice, is given to the source.
NBER Working Paper 5737 September 1996 CRIME, URBAN FLIGHT, AND THE CONSEQUENCES FOR CITIES ABSTRACT
This paper demonstrates that rising crime rates in cities are correlated with city depopulation. Instrumental variables estimates, using measures of the certainty and severity of a state’s criminal justice system as instruments for city crime rates, imply that the direction of causality runs from crime to urban flight. Using annual city-level panel data, our estimates suggest that each additional reported crime is associated with a one person decline in city residents. There is some evidence that increases in suburban crime tend to keep people in cities, although the magnitude of this effect is small.
Analysis of individual-level
data from the 1980 census confirms the city-level results and
demonstrates that almost all of the crime-related population decline is attributable to increased outmigration
rather than a decrease in new arrivals to a city. Those households that leave the city
because of crime are much more likely to remain within the SMSA than those leaving the city for other reasons. The migration decisions of high-income households and those with children are much more responsive to changes in crime than other households.
Crime-related mobility imposes costs
on those who choose to remain in the city through declining property values and a shrinking tax base.
Julie Berry Cullen Department of Economics, E52 Massachusetts Institute of Technology Cambridge, MA 02142
[email protected]
Steven D. Levitt Harvard Society of Fellows 78 Mount Auburn Street Cambridge, MA 02138 and NBER
[email protected]. edu
The difficulties
confronting
into large-scale limitations,
urban flight.
Violent
crime
Among
higher than in cities with
considers
that
locking
the greatest
the post-war
populations
faced
populations
below
subsidies,
has pushed
private
expenditures
security
many
in 1993
and seven times
guards,
alarm systems)
cities to cities
when
is
were
greater
on police and the level of victim
trend
tax
by large American
over 500,000
50,000,
are well
suburbanization
federal
public services
difficulties
cities with
decades
crime rates in large cities are even more remarkable
both per capita
doors,
turned
tax base, shrinking
in need of greater
rates in U.S.
rural areas. z Higher
(e.g.
A declining
and city residents
the brink of fiscal crisis.
times
cities in recent
The urban riots of the late 1960s
documented.’
crime.
large American
four
than in
one
precaution
are much greater
in large
cities.3 This paper examines extensive
literature
Kelejian
1973,
relatively studies
analyzing
Frey 1979,
little focus
MarshalI
(e.g. Taeuber
1979,
Marshall
and Taeuber
and O’Flaherty
Bradbury,
Craig,
Downs,
and Lute
Although 1965,
1987),
the phenomenon,
the level of crime as a right-hand
1 See, for instance, and Inman,
this subject
crime and urban flight.4
on the role of crime in explaining
have included
(1 987),
the link between
side variable
and Small (1 982),
there
Bradford there
Gottdeiner
(1 986),
and
has been
A handful
in cross-city
is an
of OLS
Wilson
(1 994).
2 The Uniform Crime Re~o rts, which reflect only those crimes reported to the police, are likely to understate the actual difference in crime rates across city types because of the greater likelihood a crime will be reported to the police in smaller cities (Levitt 1995), 3 In 1990-1991, capita
cities with
on police protection,
expended
an average
4 While correlated
of $97
urban flight with
of city residents
values.
Thus,
per capita
(U.S.
Department
of families
failing to complete proxies
populations
below
below
high school,
for city decline 1
an average 75,000
of Commerce
in its own right, declines
in the percent
urban flight
over one million spent
cities with
is interesting
increases
fraction
population
whereas
of $210
1993).
in city population
the poverty and declines
more generally.
per
residents
are also
line in the city, the in median
housing
estimates
of urban flight.
These
crime rates and urban flight Wooldredge
methodological perspective,
the simplest
associated
with
relationship
Katzman
within
a net decline
1980,
between
Sampson
and
5 Previous
research 1988,
G In a simple
reported
not
on changes
rates,
crime
1985).
between
rising crime
of this effect
rates and
1976-1993.
crime in a central
Higher suburban
the magnitude
any net population
necessarily
Taylor
and Covington
of residents,
model where change.
for future
point,
a positive level,
city is
crime rates tend to
is much smaller
1988).
values
conclusions
relationship
especially
than for
to that
is concerned
with
fully capitalize
regularity
leaving
puzzle
work. 2
population
literature,
(Smith
which
net population
in housing tastes
imply any change
we take the empirical the theoretical
between
changes.
the costs and benefits
by changes
costs and heterogeneous
but do not necessarily
from these
in poor neighborhoods
In contrast
can be captured
Relocation
In this paper,
as a starting
any causa/
our analysis
property
in amenities
of the residents,
of city residents.
draw
has also documented
then changes
fully adjust
focusing
cities over the period
one resident.
rates at the neighborhood
on gross flows
populations
although
one cannot
and crime
composition
correlation
each additional
level,
(O’Brien
and
in crime rates,
in crime reporting
across cities
on the
From a theoretical
that changes
On a practical
using a panel of 137
primarily
are both theoretical
in crime rates.
non-comparabilities
of about
cities,
There
focuses
cities.
Of course,
and Jarjoura
our analysis
model predicts
the strong empirical
declines
in central
choice
practices
for other factors,
with
keep people
location,
a positive
1982,
changes
in city populations
associated
I documents
controlling
focuses
residential
changes
city population
turnover
Grubb
on the topic,
for examining
and police department
Section
crimes
research
justifications
problems
definitions,
central
Frey 1979,
obtain
in crime rates rather than levels.
will influence
mitigates
After
to previous
of changes
levels,
(e.g.
typically
1986).5
In contrast impact
studies
values
affect
of why
the
in the total
of fluctuating property
of a without
number
city values
do not
estimates.
While
it is also possible urban flight. comprised
the most obvious that an omitted
For instance,
positively rates.
with
may be responsible
loss in shrinking
city residents
On the other hand,
city size, city growth
is the case,
for both rising crime
rates of crime victimization),
will rise, even if each resident’s
since crime rates are strongly
is likely to be correlated
OLS may understate
and
cities is disproportionately
(who tend to have lower remaining
is unchanged.
correlated
If that
individuals
crime rate among
risk of victimization
link runs from rising crime rates to urban flight,
third factor
if population
of high income
then the average
causal
the causal
impact
with
rising crime
of crime on city
population. We attempt
to establish
use of instrumental but not otherwise source
variables belong
These
are plausibly city population
changes
variables
excluded
aggregate census
yields
migration census
resulting
is overwhelmed limitation
population
to affect
changes.
a number
it is also possible
equation.
manner,
impact
analysis
in Section
category,
between
3
Ill.
yet
that
changes
is that it addresses
any in city
sample
only of the
We are able to obtain race, and family
increasing
In
of crime on
suggesting
or compositional
data from the 5 percent
cities by income to distinguish
system.
for large cities to have higher crime rates.
panel-data
insights
variables
rate
prison releases
in the predicted
The estimated
the
The logical
justice
and state
larger than the OLS estimates,
Using PUMS
of additional
crime
the crime
changes.
of the criminal
prison commitments
from omitted
of the above
must affect
city population
in the punitiveness
by the tendency
in and out of 81 U.S.
data,
explaining
in state
are slightly
crime and urban flight through
A valid instrument
are demonstrated
using 2SLS
A major
Il.
from the city population
bias in the OLS estimates population
in Section
is changes
we use lagged
per crime.
link between
in the equation
of such instruments
particular,
a causal
estimated
status.
out-migration
1980
and
Using
declining
in-migration
suggest
and to identify
that almost
migration; percent compared income
all of the impact
the link between
of those
people
to forty
Whites
are twice
as responsive
Because
to crime
IV summarizes
accompanied
are likely to have greater
services
is diminished.
declining
housing
prices,
eroding
of cities,
1987),
cities can be caught
and Katz
In the presence
To the extent 1991,
groups will exert
mobility
Benabou a further
1993,
Households
of high
with
While
and city hospitals,
children
rough calculations
those
remaining
returns
spiral in which
such as police protection,
externality
on those
4
is the primary
(Blanchard
those
capital
the exodus left behind.
with
revenue
and Summers
rising tax rates lower
to human
1995)
to
in the city
the ability to provide
tax base, which
and Glaeser
of
In addition
that rising crime rates are associated
are important
of
the rich, rising crime will be
public services
of fiscal increasing
negative
decisions
other costs on city residents.
among
of poverty.
the property
Cutler
the SMSA,
children.
numerous
also suggests
that peer effects
within
out-
Seventy
crime rates than those
to crime.
is concentrated
in a downward
weak.
for cities and their residents.
crime imposes
Our results
from increased
The mobility
without
need for locally provided
Evidence
appears
of the paper and provides
concentrations
source
wages.
as households
public transportation,
results
to changing
similar responses
out-migration
by increasing
public education,
more responsive
and crime-related
crime-related
city out-migrants.
the results
costs of victimization,
who leave cities.
cities due to crime remain
of all central
and blacks show
the costs of crime
in crime and in-migration
central
are five times
the poor.
the direct
leaving
of those
of crime on city population
changes
percent
households
Section
the destinations
real
development
(Case
of the most mobile
Section
1: Correlations
between
We begin our analysis changes
in crime
and suburban described
crime
SMSA,
a simple
in city population.
rates in the specification, socioeconomic,
and state.
Our logic for focusing
the previous
individual’s
residential
decision
economic
(although
The theoretical
prediction
that changes
city population
is strongly
supported
The basic specification
We include
we employ
on changes
potentially
rates.
whereas
This specification
independent
or changes
at the sample
contemporaneously covariates
makes
endogeneity
whether changes
these
in
with
some adjustment
presented
into the lag).
changes
in
later in the paper.
(S TATEPOpit_l)
+ AGEit_l
time,
With
crime variables of changes
the exception
problems.’
covariates
G + At
and regions
+ Yr
(1) + eit
respectively.
are in changes
The population in per capita
in crime on mobility
to be
such as log changes
in levels of both variables,
mean.
in crime rates is that
of
have been incorporated
evidence
other specifications
and will be instrumented,
to reduce
7 It is unclear logic that
the two
the impact
of the level of crime;
and population, evaluated
allows
characteristics
CRIMEit + fi2ASUBURB_CRIMEit + P3 UNEM
1, t, and r index cities,
is in log changes
city
is as follows:
+ B6%BLACKit_1
variable
between
in both central
and not levels of crime should effect
+ ~4 INC;MEit_l + ~~aln
the subscripts
changes
and demographic
by the empirical
Aln (CITY_ POPit) = ~lACITY
where
relationship
a range of other covariates
level of crime in a city will already
choice
in Citv POK)UIation
reduced-form
along with
to capture
equilibrium,
and C hanaes
in Crime
by specifying
and changes
below
the city,
Changes
in both crime
lead to similar results
of the crime variables,
when
which
enter
one year lags are used for all of the other In addition
to economic
controls,
should enter in levels or changes.
rather than levels of crime rates appropriate 5
suggests
a set of
The same changes
for
variables
reflecting
populations, sources
and a variety
U.S.
analysis
of indicator
(1) is estimated
central
cities with
we allow
of compromises migration
variables
because
is unavailable,
and death
rates
gross flows
populations
are included
the analysis so changes
(approximately
due to migration
is performed in overall
(roughly
log-changes
to capture
using data from the decennial city population
0.2
census,
changes
to the use of net migration
in state
other systematic
a Immigration
versus
net migration
other variables
up differential frequently authors)
include
utilized
8 Moreoverr community effectively
they
levels.
movers
affect
little effect
Our results
the importance
to
across
States
over the (1 986),
to the choice
are also robust
Ill. reported
to the
adults
controls move
levels as the correct
of this paper (available
may be picking
six times choice.
on request
as The results from the
on the crime coefficients.
city populations
move to, whereas
moves
are not sensitive
of index crimes
(e.g. young
suggests
An earlier version with
doubles
which
low compared
Given that less than half of all
Re~orts.
across groups
over age 65),
changes
number
Birth
and Wooldredge
On the other hand, the demographic
levels of mobility
as those
that follow
as well.
Crime
in Section
not
are used instead.
Sampson
flows.
a number
are influenced
are relatively
find that their results
the
data on net
rates in the United
per year.
for
Throughout
Annual
which
rates,
1976-1993
necessitate
of the U.S. population
data from the Census
in Uniform
in 1975.
at the city level.
per hundred)
percent
the years
Data limitations
and immigration
The crime rates used are the per capita police and collected
100,000
city population,
six percent
1987)).
are roughly
than
by city size.
1.5 and 0,9
lines each year (Schwartz
of overall
greater
but also by birth, death,
time period examined
these
lagged
using a panel of data covering
for heteroskedasticity
only by migration
county
of the population,
of variation. Equation
137
the age distribution
in both the community
births and deaths of migration 6
affect
in explaining
they
leave and the
only one community. city-level
population
This flows.
victimizations
are reported
by substantial
measurement
level of regions, paper,
estimates
presented
in Section
are available
level of aggregation
Area,
yearly
by state.
associated
year dummies, effects.
census
the left-hand trends
region-year
city,
annually
to linearly
Per capita
over time. interactions
the percent
adopted
is already
In some cases,
7
into violent
and
of suburban
As a general
income
Most
city-
between
the
principle,
we
data.
of numerous
we replace
is available
of a city’s
population
that
the problems indicator
variables:
and city-fixed
a city-fixed
effect
the year and region
for regional
on a
In the OLS
regions,
in differences,
control
crime.
is also available
for minimizing
to the nine census
to better
aggregated
a tradeoff
of the population
is the inclusion
side variable
in the IV
in our regressions.
data rather than using state-level
corresponding
crimes
necessitating
by SMSA.
strategy
crime
particularly
only estimates
of collection.
interpolate
two
similar
panel are less than ideal.
years,
The age distribution
those
the
of this
obtaining
as our measure
source of data available
the primary
region dummies
Because
with
census
between
below
We use crime rates in the rest of the
in a city-level
the lack of good controls
picks up within-city indicators
level.
of this section, with
annual
We choose
is black from decennial regressions
available
and property,
disaggregating
the central
version
the coefficients,
on request.
contaminated
is not available
In a previous
we present
full results
excluding
are available
basis at the state
annually
identify
of the data and the frequency
rates
however,
of collinearity
Consequently
only in decennial
use the most disaggregated Unemployment
Il.
in the tables;
of covariates
Ievel data is collected
data,
into violent
to separately
from the authors
Statistical
The types
crimes
The high degree
it difficult
over all crime categories
Metropolitan
Victimization
to disaggregate
makes
the UCR crime data are potentially
the use of UCR statistics.
on both types.
classifications
property
error.
necessitating
we attempted
coefficients
to the police,
economic
and
demographic regional
shifts
characteristics.
also worth variables
Two
are uncorrelated
results
set of covariates
Summary enter equation
statistics
OLS regression
Column
1 corresponds
3 includes
columns
with
directly
effects
effects
on the crime variables.
reported
straightforward
to demonstrate
9 The larger is the variation year-to-year
crime
utilized
1.
we
here or a greatly
For those
variables
in Table
2 adds city-fixed
effects.
Columns
are in the top two
improves
4-6
that D1 closely
rows.
the fit of the model
are statistically
of interpreting
interpretable.
the closer is the approximation.
in Table
interactions.
the results,
crime is a useful measure
and thus is directly
set of controls,
that
are 2. Column
mirror the first
crime levels added to the specification.
The crime coefficients
For the purposes
better
we use
controls.
Column
and region-year
interactions
each additional
measure
have a much
for a range of specifications
(1).
are
of the crime
for both the levels and changes
on the crime rate changes
and region-year
specifications.
to equation
variables
Ill of the paper where
set of covariates
statistics
or other
as long as the omitted
estimates
in Section
household-level
are presented
once-lagged
The coefficients fixed
results
both city-fixed
Second,
and therefore
summary
estimates,
consistent
for the data are presented
(1 ) as changes,
shown.
variables
we use the limited
including
for climate
points about the lack of good control
be obtained.
whether
over time in tastes
the instruments,
data from the census
similar
expanded
three
with
will nonetheless
individual-level
further
changes
First, in the instrumental
noting,
parameters
obtain
as well as possible
approximates
g The interpretation
city-
but has little effect
significant the change
of the magnitude
Including
in all in city residents
of the effect.
(but slightly
overstates)
for
It is that
of ~z is more complicated,
in crime rates relative to the variation in city populations, In our data, the standard deviation of percent changes in
rates is five to six times
larger than the corresponding
populations. 8
value for city
however,
because
it depends
our data,
suburbs
on average
to determine sample,
the change
suburban rates.
with
crime
translates
unemployment
translates persons
with
substantially
with
to the omitted The coefficient between
0.33
effects
effects
to 0.64.
income
are imprecise.
population
Excluding
growth
this variable,
for each
effect
100
in This
additional estimates
in
on city populations
There
is some evidence faster.
more slowly, in percent
black,
estimated
has little impact
a
the earlier however,
which
of city-fixed
is very large with
that
Cities with
mirroring
The age coefficients,
however,
change
in city population.
“over age 65, ” flip sign with the inclusion state
on
point change
to the state-level
variation
city crime
a ten percent
tend to grow
appear to grow
of our
less than one percent.
decline
are included.
is little within-city
the estimates
on lagged
percent
Thus,
The coefficient
A percentage
is similar in magnitude
higher per capita
There
category
of slightly
manner.
In
as large as that for city crime
5 and 35 city residents
of black residents
of Frey (1 979).
city-fixed
to 0.23
city-fixed
an additional
in city crime rates (roughly
The size of the unemployment
when
in states
initial fraction
findings
estimate
and Katz (1 992),
located
higher
That
about one-sixth
cities.
crime at the mean
of about one resident.
in a plausible
of between
suburban
populations.
as central
to the estimates,
in city population
leads to a 0.035
unemployed.
increases cities
enter
into a decrease
Blanchard
According
change
into a decline
residents
per additional
into an impact
deviation
The other covariates SMSA
as many
in city population
rates translate
A one-standard
in crime),
~z by three.
a decline
sizes of the city and suburban
have three times
in city residents
one must divide
is associated
on the relative
so
are relative effects. elasticities
on the crime
parameters. The specifications population
in columns
occur contemporaneously.
1-3 of Table
2 assume
To test the possibility
9
that changes of partial
in crime
adjustment
and or lagged
responses lagged
to crime,
crime rates
generally Lagged
columns
4-6
include
are statistically
less than a tenth crime changes
the once-lagged
significant,
as large as those
also appear
crime rate in levels.
and the magnitudes associated
to have weak
with
of the coefficients
changes
explanatory
None of the
power,
are
in crime rates. as do /cads of crime
changes. We have also explored shown
in tabular
place of one-year time period, sample.
Estimating
differences
we cannot
Finally,
across
form).
other variants
of the basic specification
the model using fixed
yields similar estimates.
effects
Dividing
reject the null of equal coefficients
splitting
cities greater
numerous
the sample
than or less than 250,000
Sect ion 11:Establishing
or longer differences the sample
across the two
by city size, we cannot
(not
according
in to
parts of the
reject the null of equality
in population.
Causa Iitv in the Relationship
C itv PODulations
between
and Crime
w The preceding
section
crime rates and changes the direction flight,
there
might
affect
urban areas, perhaps
of causality.
suggesting
to increased
anonymity
of big-city
a strong
in city populations, While
are also a number crime rates.
documents
Looking
life, or because
provide
channels
through
in densely
big cities attract
10
between
changes
any clear guidance
which
changes
higher in large
lead to higher crime rates,
settled
areas,
likely criminals
about
lead to urban
population
crime rates are much
in city population
opportunities
correlation
is that rising crime rates
cross-sectionally,
that increases
criminal
but cannot
one possibility
of plausible
negative
due
the increased (e.g. Glaeser
and
in
Sacerdote
1996).’0
understate
the true causal
growth
If that
may overstate third variable correlated Second,
hand, there
between
are a number
the true magnitude that
with
is positively
associated Iinearities
city population
with
increasing
between
if there
changes
concentrations
rates are defined
measurement
the dependent
may since city
measurement
10 Where rates.
error induces
data
Beattie
offenses mostly
against
American
cities
activity, Wilson
basis.
a negative
bias.12
city population
the reported
12 When
crime
an individual
argues
exacerbate
almost
An apparent
this bias. is present
regression
model
may be an incorrect
higher than rural crime
seven times
of the professional reduced
victimization
when
side variables,
1690-1720,
exception
be
for non-
problem
prosecutions greater
is
police force the amount
for
than the
to this pattern
is that the rich are more likely to report
police than the poor, so even if the actual changes,
always
that for the period
the emergence
might
will not bias the coefficients,
latter half of the 19th century appears to have dramatically in large American cities (Monkkonen 1981). to this argument
further
of the right-hand
in the city of London were
of Essex and Sussex,
in the late 1800s;
11 One caveat
(1 987)
Unlike the standard
Finally,
of less mobile
then depopulation
urban crime rates are almost reports
comprised
then a ratio-bias
side variable
in the denominator
First, an omitted
OLS coefficient.
and crime that would
on a per capita
for instance,
property
rural counties
cities is increasingly
of poverty
are available,
(1 995),
the negative
in criminal
that the OLS estimates
in crime rates and negatively
error in city populations,
appears
suggesting of city crime.
changes
crime rates.’l
error in the left-hand
variable
impact
could explain
in shrinking
per capita
is measurement
of factors
with
are also more likely to engage
since crime
Table
section
city crime and depopulation
of the causal
correlated
if the pool of residents
groups that
where
relationship
of the previous
will lead to higher crime rates. On the other
Third,
is the case, the estimates
in the of disorder
crimes
to the
rate rises due to compositional
rate may not.
household’s
8, ratio bias is not a concern.
mobility The results 11
decision
is used as the dependent
are unaffected,
suggesting
variable
ratio-bias
in
is not
denominator tourists,
for crime rates if a large fraction
If many
central
former
city and be victimized
between
city populations We attempt
what
city residents
follows,
crime with
stories
rates
our efforts
(e.g. reverse
coefficients.
crime coefficients
otherwise
will reduce
effect
on city migration
median
Piehl 1991, capture
except
commits
Spelman
impact roughly
1994).
not only incapacitation,
a major problem
correlation
variables. treating
In
suburban
the
and compellin9
for city
we have also experimented
with
must affect
city population
no impact
on the city crime
that the suburban
city crime rates,
changes.
of the state
deterrence
criminal
or incapacitation,
The two
of prisons.
instruments
Prisoner
15 index crimes
Panel-data
in the
instrumented.
of the severity
via crime.
negative
fit, we can never reject
determining
crime either through
crime-reducing
prisoner
first-stage
are measures
applicable
crime rates,
to work
perspective,
Empirically,
and ratio bias).
or
in this paper.
From a theoretical
for the city crime variable
in the equation
which
a spurious
for city crime rates,
are more directly
are equal to zero when
of such instruments
documented
causality
of a weaker
instruments
belong
above
but continue
the use of instrumental
on instrumenting
for both city and suburban
Because
Valid
through
to city population.
discussed
instrumenting
causality
of crime are commuters
to the suburbs
and city crime rates as measured
crime rates as exogenous endogeneity
relocate
therer13 this could induce
to determine
we focus
of the victims
analyses
but also deterrence
The natural justice
but should
suggest
a year prior to incarceration using state-level and replacement
aggregates, effects,
choices
system,
we use exploit
self-reports
but not
have little the well-
that the (Dilulio
and
which yield
here.
13 Census data on place of work suggests that roughly 80 percent of those who work in the central city continue to work there after moving residence to the suburbs. 12
estimates Levitt
of crime
reductions
per prisoner
(Marvell
and Moody
1994,
1996). The particular
rates of state
probation
resulting
used are lagged
per reported
from criminal
and parole violations.14
deterrence
via the certainty
rate implies
punishment
These
prisoners
were
These
flows
in this paper,
the instruments to be strongly
through
their impact
instruments sample
represents
14 When commitments
15 State negative residents
prisons
and
rate translates
into
percent
of state
in 1979.
The state
held in local jails.
crime rates.
in city populations,
Though
The instruments
however,
are
except
likely to be true since the
level rather than the city level, of the population
other than the index
The average
and only eleven
percent
city in our of the
it is located .15
separately,
new commitments
and parole
in the first-stage
and probation-related
regression
and the 2SLS
results
affected. generally
are not located
in big cities which
between
city populations
and commitment
are being sent to prison,
populations.
from
effects
22 percent
or those
with
This is particularly
had similar coefficients
correlation
In 1993,
prisoners
with changes
at the state
in which
not substantially
both
A high commitment
for crimes
only, up from 6.4
are still highly correlated
on crime rates.
entered
both incapacitation
A low release
drug offenses.
charges
only seven percent
crime for the state
include
resulting
in the state.
and conviction.
either federal
correlated
are defined
of punishment
include convictions
notably
held on drug-related
unlikely
Commitments
and commitments capture
and release
per conviction.
prison data also does not include imperfect,
in the commitment
crime in the state.
variables
of detection
prison population
considered
changes
convictions
and severity
a high likelihood
a longer mean
crimes
instruments
prison systems
new prison terms
were
of similar magnitudes
Our reduced
the removal
form estimates,
of those
however, 13
could potentially rates:
offenders
when
many
will reduce
find that commitment
induce city
city
rates and city
a
One potential
problem
with
the instruments
denominated
by crime rates (albeit
crime rates.
In the absence
denominator
of the instrument
functional
form
however,
generally source
is correctly
There
stage regression Empirically,
while
however,
measurement important
using state-level
a small fraction
the negative
while
relationship the effect
of state values
of crimes
error in reported
commitments
we attempt Second,
rates, First, the cities
to lessen this
ratio bias will
and the crime rate in the first-
will be systematically
we find that commitments
crime
if the
data and individual
Nonetheless,
of the instruments.
of releases
of the instrument
than city-level
crime.
for
in the
points to note on this topic.
rather
between
are
they are also instrumenting
error, the presence
With
are two
of bias by using only lagged
exaggerate
is the fact that they
does not result in any contamination
specified.
are constructed
represent
crime rates)
of measurement
ratio bias arises.
instruments
state
selected
and releases
understated.
carry coefficients
of similar
magnitude. Columns covariates
1-3 of Table
used in the previous
expected
the first-stage
section.
In all cases,
sign: a higher rate of prison commitments
are accompanied prison variables Releases
3 present
by crime increases. are jointly
and commitments
prison commitment period.
generally
not strong
predictors
using the same sets of
the instruments reduces
crime
in the bottom
enter with whereas
and economic
correlated,
implying
A one-percentage
variables
included
controlled
that this source
14
more releases the
level for all of the crime regressions.
of crime rates having
results.
the
panel of the table,
point increase
rate per crime shifts city crime by roughly
The demographic
are positively
at the .01
have similar effects.
or release
two-year
populations
significant
As reported
estimates
in the
six percent
in the regressions
over a are
for the other factors.
of bias cannot
explain
our
If prison commitments a negative
causal
city population
obtained
such estimates.
level.
link between
variables
in the first-stage
As predicted,
regressions.
a reduced
Columns enters
The prison variables
the magnitude
for crime rates and there form regression
should yield coefficients
each of the instruments
in the first three columns.
As in the first stage,
are valid instruments
crime and city populations,
on the instrumental
sign to those
observed
and releases
of
are opposite
4 to 6 of Table
significant
on commitments
in
3 present
with the opposite
are jointly
of the coefficients
that
is
sign at the .05 and releases
are comparable. Table endogenous
4 presents
the two-stage
and instrumenting
and releases.
The coefficients
on city crime that we obtain
city crime
compared
to estimates
that
OLS coefficient
the estimates
additional
results
are statistically
in roughly
in city populations
higher crime rates. suggest
that declining deviation
between
one and two
‘G If there
increase
percent.l
are lagged
population
changes,
the variation
are lagged)
from these
variables tendency
the standard
biasing the 2SLS estimates.
level.
estimates
Each
suggest
for large cities to have of Wilson
(1 987)
which
A one-
in city population
are treated
as exogenous,
from the specification, factors
errors
or compositional
into a decline which
in
in population,
and contemporaneous
omitted
15
more negative
crime rather than vice-versa.
that are omitted
(which
reduction
rates as
in commitments
Although
on the arguments
G The other covariates,
variables
with the instruments potentially
by the natural
“cause”
changes
are slightly 2.
city crime
The larger 2SLS
from omitted
cast some doubt
treating
from zero at the .05
in city crime rates translates
correlated
to our instruments
resulting
city populations
standard
different
1.1 from OLS.
is overwhelmed
Our results
in Table
a 1.3 to 2.0 person
of approximately
any bias in the OLS estimates
changes
estimates
with the once- and twice-lagged
each case than the corresponding rise substantially,
least squares
but are crime and
will mistakenly
be attributed
To test this hypothesis,
we
of
carry coefficients
similar to those
P-values bottom .05
level of statistical
fraction
ease in column argument
total
4, but are rejected
Comparing
of cities in our sample
according
of the instruments
in OLS estimates,
for cities that
towards
estimates
suggest
that,
zero because
in column
if anything,
of the inclusion
4, for which
the coefficients
to the
are large relative
4 and 5, the estimated
of their state.
This pattern
are actually
population
the erogeneity reported
pass with
is
slightly
(-1 .12 vs. -1. 19), of the instruments
in columns
is
1-3 are biased
of cities of all sizes.
Sect ion Ill: Usina Census Data to Identifv Income
of the state
restrictions
for cities that
the city crime coefficients
a
made
This lends credence
on city crime in columns
are less than four percent
The larger parameter more likely,
where
that represent
a city has more or less
The overidentifying
is suspect
poor performance
To test this hypothesis,
to whether
5.
in the
are near the
the arguments
are less persuasive.
at the .05 level in column
the coefficients
restrictions
For such cities,
population.
are reported
reason for the relatively
of crime is larger in the cities that are a small fraction
not apparent smaller
population.
4 break the sample
of the state’s
restrictions
the overidentifying
is the inclusion
of the instruments
that the erogeneity
to the state.
cases,
One possible
of the state’s
4 and 5 of Table
than four percent
impact
restrictions
for the erogeneity
columns
In all three
significance.
of the overidentifying
earlier
in the OLS specifications.
from an N*R2 test of the overidentifying
panel of the table.
non-negligible
reported
Differential
Effects
of Crime
on M iaration
bv
Group and Race
estimated variable presented
Table
4 including
to capture
the once-lagged
such omitted
factors.
dependent The results
in the table. 16
variable obtained
as a right-hand are almost
side
identical
to those
Because
the results
has been impossible this section,
to isolate
we perform
Census
of Population
income
group and race.
analysis
two types
and Housing
of analysis
households
as the dependent
advantages
of the latter characteristics
Analvzina
Census
mobility
sample,
identify
movers
remaining PUMS
variable
analysis
the migration
(as opposed
are that it makes
and that
it eliminates
ratio-bias
5 percent
in the PUMS
cities and surrounding
90 cities,
we limit our
but for sub-groups
of
of individual
aggregates). to control
9 are missing
of the
The
for individual
1980
data were
of the U.S.
in and out of 90 of the 137 cities analyzed
Of those
by
concerns.
sample
usable data for 2.5 percent
central
patterns
to the Citv Level
Only half of the observations leaving
it possible
In
data to the city level
decisions
to city-level
it
from the 1980
of 81 cities,
section,
data,
of the population.
migration
the individual-level
employs
population
information
crime-related
aggregates
approach
Data Aaareaated
47 cities,
data.
sub-groups
using mobility
to investigate
The data used are from the PUMS to the city-level.
among
that are similar to the preceding
The second
household
impacts
from aggregate
Since we only have a cross-section
regressions
the population.
differential
The first approach
to OLS.
and estimates
up to this point are derived
census,
aggregated
also included
population.
in the preceding
areas are not separately
in the
We are able to section. identified
data on one or more covariates,
In the in the
leaving
81
cities in our sample. To construct household stayers, within
the net migration
head in 1975 comers,
and 1980.
and goers.
data,
Households
Households
city limits are categorized
we compare
the place of residence
are divided
that do not move
as stayers.
Those
17
into three or change
of the
migration
categories:
addresses
but remain
who arrive in a city or leave the city
are classified comers
as comers
and goers
according
to whether
SMSAS.
Households
are also grouped
in 1979.
Households
with
1993
dollars)
$10,000
percent
with
income
and the remaining
according
below
as “low”
above
households
changes
and goers is computed
according
in income
city, with statistics
by income
as the denominator. over 1,000
patterns
percent.l
Net migration
rates were
17 Given the possible
endogeneity
households
1s This number populations be explained
is greater
by differences
city population of 5,856
are middle 1979
30
income,
income,
we
or to examine
of stayers, within
household
5,
Overall,
comers
the income
observations
cities in our sample.
in Table
per
Summary
net migration
(comers
for the five year period is approximately
of 1979
income,
attainment,
than the observed
for the cities in our sample
income.
Approximately
incomes
the fraction
larger for high income
by educational
between
“middle”
income.
in
and 1979.17
using 1975
population)
$20,000
income
we only have
we have data,
are presented
of 1975
classifying
1975
1979
50 percent
to their initial 1975
for the smallest
minus goers as a fraction E
Because
between
group,
“high”
from all sources
(roughly
are labeled
category,
We have an average
observations
on migration
dollars)
that level are considered
are high income.
with
divide
or across
income
dollars
Households
we further
is within
household
in 1979
in 1993
For each of the 81 cities for which
category
$10,000
income.
of residence
to total
fall into the low income
20 percent
to classify
within-household
their change
($20,000-$60,000
of the households
are unable
income
are classified
and $30,000
Households
In some of the analysis,
or goers respectively.
3 percent rates.
(-1 6,8
percent)
we have also examined
obtaining
over the period
in birth and death
households
between
birth and death changes
rates accounts
and net migration.
aggregate
1975-1980.
Annual
for a three percentage Overall, 18
and
mobility
similar patterns.
national
decline Most
in city
of that gap can
birth rates
(per 100
population) were approximately 1.5 in this time period; comparable death rate statistics were roughly 0.9. Assuming that the same rates apply to these cities, the differential population
-8
the correlation
point gap between between
city
city population
actually while
positive
the non-black
residing
population
in one of these
households with
for low income
households declined
central
left the SMSA;
households
New
percent).
by 10 percent.
cities in 1975
less likely to leave.
high income
(0.5
Blacks flowed
Roughly
had left that
of the new arrivals
60 percent
to central
of households
with
three-quarters
less likely to come to cities.
over 70 percent
30 percent
city by 1980,
arrivals to cities offset
into cities on net,
low income
of these
leavers,
of city out-migrants
cities came from
outside
the
SMSA. OLS estimates changes
in crime
of the relationship
rates are presented
than a panel of data, that
five-year
changes census
rather
in place of residence years,
Therefore,
specifications
in Table
in owner-occupied manufacturing family,
mean temperature
are available
and 1980 Column
combined.
changes
covariates
or the 1980 1 of Table
6 differ
The coefficient
and net-migration
rather
from earlier estimates
of city residents
completing
in the manufacturing
income,
and mean
percent
of housing
in
levels did not materially
affect
the estimates -1.39,
in our sample
is 0.45. 19
sector,
mean
hourly
are multi-
Census.
Using changes
the
percent
All of these
variables
The
1970
between
the crime coefficients.
for all cities and income
is consistent
section,
units that
by the U.S. 6.
in
high school,
annual precipitation.
in Table
on city crime,
is available
used in the previous
are used as controls
6 presents
rates and
to using a cross-section
A richer set of covariates
employed
family
in January,
in Table
in the Cit v and Cou ntv Data Books published
value of the census 1970
city median
In addition
to all the variables
6 add the fraction
in city net-migration
in crime rates are used (since only five-year
are available).
percent
changes
6.
shown
changes
in addition
housing,
wage,
in Table
the specifications than one-year
between
groups
with the estimates
of the
preceding
two
changes,
while
median
sections
family
growth,
and is statistically
carrying income
in a city is a statistically
significantly
Columns income
net migration income
different
response
of low income
crime over the five-year decline 0.8
in high-income
percent
with
decrease
no apparent
without rates.
children. While
families
with
the suburban
Because the specifications including
Families
in Table
a 2.0
the poor. across
High
of city population are not
children
of variables
Columns
the three
is associated
columns
three times
are imprecisely to suburban
is large relative
with
compare
crime
and region dummies to the inclusion
of the influences
in the bottom
percent and a
families
with
and
to city crime
it appears
that
of observations
in
from specifications
panel of the table
variables.
of children
a 4.3
rates.lg
to the number
of control
in city
on racial lines,
as responsive
estimated,
in across
households,
the city crime coefficient
20
change
The
Middle-
specifications
deviation
High
of -2.58.
as great.
5 and 6 divide the sample
are almost
(1 996).
group.
a coefficient
fall in middle-income
The last two
6, we also present
treatment
by income
of the crime coefficients
per capita)
percent
races.
of our results
19 For a more detailed
When
A one standard
crime coefficients
only the crime variables
estimated.
is less than one-fifth
response.
are also more responsive
the number
test of the sensitivity
Glendon
households
(0.01 65 crimes
with
crime
The other covariates
of net migration
we reject the equality
households, among
predictor
to city crime rates with
at the .05 level,
period
difference
children
jointly
groups
significant
estimates
have an intermediate
high and low income
Suburban
from zero.
separate
2-4 are estimated
level.
are imprecisely
and low rainfall.
are most responsive
households
columns
temperatures
2-4 provide
households
at the .10
the same sign as previously,
as are high January
statistically
significant
The estimates
on locational
choice,
as a are
see
similar
in all cases. Further
comers
patterns
and goers,
are revealed
and by whether
in Table
7 for city crime rates.
previous
table.
from Table (row
(row
Almost
possible recent Table
1).
overall
40 percent
remain
rates
The coefficient crime-related
decompose
in the
on goers
impact
the effect
on
by
on goers compared
is that current
residents
in-migrants.
(including
those
arises from increased to comers,
are better
Comparing
in the SMSA
One
informed
the bottom
70 percent
for blacks
out-
and families
who move for reasons
about
two
rows
of
of the time with
children).
other than crime),
in the SMSA.
Analvzin a Individual
Hou sehold
on individual
advantages.
covariates.
on city crime
on mobility
for the rich, more frequently
for all out-migrants
the coefficient rows.
to
is presented
to the categories
we further
cities due to crime remain
In contrast,
Focusing
impact
in crime than are potential leaving
This breakdown
(row 2) equals the total
and goers,
according
7 correspond
in the subsequent
by the large magnitudes
for this patterns
(less frequently
has two
in Table
lines.
effects
or across SMSAS.
all of the crime-related
explanation
7, those
comers
is within
as evidenced
changes
on comers
Within
the migration
migration,
cross SMSA
The columns
is then decomposed
5) minus the coefficient
whether
moves
the net migration
In each case, the first row redisplays
6, which
net-migration
by decomposing
Second,
household
First, it allows it eliminates
population
does not affect
performing
a household-level
Dec isions u sina Ce nsus Data
Mobilitv
mobility
the inclusion
concerns
the dependent analysis
decisions
of household
over ratio-bias variable,
aggregates
characteristics
since measurement
The primary
are the computational
21
rather than city-level
difficulty
costs.
as error in city
associated
For example,
with
to avoid
sample
selection
particular though
city,
bias when
one must include
only a trivial
fraction
We use the results choices
in reducing
demonstrates increased
that
on the decision
roughly
of current
dependent
the central crime
of city,
PUMS
remains
city before
at the sample state,
the first column, income
indicators
marital
status
(high, middle,
of the household
crime
for those
estimates
households
and low),
has a larger effect
residing
on those
400,000
exclusively
Doing so limits observations,
in Tables
or
1980,
change
in the per capita
moving
marginal
6 and 7.
With
array of househo/d-/eve/
where
and across SMSAS. within 22
the SMSA
the
cities leaves
effects
when
the full set
the exception
of
controls:
the age, sex, race, years of education, the household
For
reported
8a and 8b include
multinominal Iogit models
within
models
and zero if the household
in Tables
a wide
8a and 8b.
in one of the large central
used in Tables
and whether
with
moving
to stay or leave.
Probit and Iogit yield similar
also include
head,
7
we focus
from linear probability
the five-year
that were
the regressions
Table
from cities is due to
are presented
means. zo All specifications
20 We have also experimented effects
analysis
of crime.
and region controls
level.
sample.
As before,
the effect
even
into the city.
analysis
approximately
in that city through
1980.
into a
to guide our specification
in the following
about whether
mobility
we present
analysis
net migration
is equal to one if a household
in 1975
rate captures
evaluated
city residents
move
to a manageable
all of the crime-related Therefore,
or not to move in the estimation,
actually
city-level
burdens
of the individual-level
variable
in our sample
of the total households
of the preceding
of interpretation,
of whether
in the sample
of large cities in 1975,
of the 1980
The results simplicity
all households
of residents.
to residents
15%
the decision
the computational almost
outflows
the sample
estimating
and
head is a homeowner,
in
that allow for differential These
as would
results
suggest
be expected.
that
the armed
forces
in 1975,
error term
for households
Failing to account roughly
twenty
the crime variable
standard report
the corresponding
Table
8.
city.
to Table
city residents from moving Table
6, there with
to the suburbs. into the SMSA,
suburban
The second
effects,
estimate
from the comparable
changes
and population
purposes,
in the bottom
that suburban
Table
in the fact that the
1.61
residents
panel of
of
leaving
offsetting
the
those
effects
here. on
the outflow
outside
The net migration
8, which
we
crime rates matter
by reducing
is to discourage
of the impact
of
the SMSA
regressions
in
looks only at the decision
of city crime in column
value from the aggregated
regressions account
effect
gain in moving
for the full sample
with
in
only the former.
the individual-level
cannot
controls
are
all of the variation
regressions
city population
whereas
Since the aggregated
ratio-bias
that
For comparison
city population.
panel of the table.
that
error estimates
crime rates having two
is to increase
to stay or go, reflects
The household-level indistinguishable
is some evidence
errors.
as reflected
crime is associated
depressing
both of these
on city residents
only aggregate
reported
The first effect
6 capture
estimation,
in the
the standard
very little information
from the city-level
8a includes
is consistent
city population.
is actually
for correlation
Because
7 and 8 are not very different.
Each additional
In contrast
We account
leads to standard
to individual-level
coefficient
1 of Table
This difference
correlations
is at the city level, there
errors in Tables
households.
in 1975.
too small on the crime coefficients.
aggregates
Column
college
in a given city so as not to understate
for within-city
times
from city-level
or attends
regression
do not, the similarity for the observed
changes.
23
1 of Table
regression
in the bottom
potentially
suffer
of the two
sets of estimates
negative
relationship
from ratio-bias,
between
8 is
but
suggests
crime
rate
Column
2 of Table
8a is identical
household-level
controls
these
significantly
variables
improve
This increases
Ievel aggregates,
where
Having
for city-level
black,
have children,
stay in central
there
the R2 of the regression,
our level of confidence
was concern characteristics
Members
of the military
and those
likely to leave the city.
The only surprise
for these
the rich are somewhat
income.
High income
muted
compared
income
groups.
non-blacks. interesting blacks, factors, sample
6.
difference
columns
blacks
is associated there
into households
with
effect
with
is no such effect. and without
more likely to leave central
families
children
with
children,
are somewhat
whereas
the opposite
college
are far more
to household
but the differences
is the effect
in central
cities,
Households
cities in response
between across
races.
controlling
with
is true of households
without
children
children.
and
The most non-
for other
of Table
to rising crime.
across
Among
8b split the are
High income
more likely to leave cities than low income
24
are
blacks
of income.
The final two columns
Sect ion IV: co nclusions
controlled
in cities.
according
of crime is observed
children.
male,
is that having
8b divide the sample
remaining
using city-
tend to carry similar coefficients
and non-blacks
of crime
are more likely to
more likely to remain
of Table
substantially with
attending
to crime,
while
was incomplete,
attainment
the coefficients,
are more sensitive
impact
heads who are young,
8a divide the sample
little differential between
blacks,
among
The other covariates
The first two
As before,
With
of Table
households
to Table
high income
household educational
columns
the estimated
that the set of controls
of
to note that
in the earlier estimates
or have lower
other factors,
that a full complement
It is interesting
are single,
cities,
The last three
1 except
are added to the specification.
is little changed.
controlled
to column
families
This paper examines
the connection
decline.
Rising city crime rates appear
reported
city crime
There
is associated
is some evidence
Instrumenting
than OLS.
increased
out-migration
income poor.
households
migrants
to play a causal approximately
Almost
of criminal
justice
all of the impact
with
are much
urban flight,
rather than a decrease
children
role in city depopulation.
a one person
system
decline
in new arrivals. to changes
the SMSA
city
Each
in city residents.
tend to keep people
severity
yields slightly
Migration
in cities.
larger is due to
decisions
of high-
in crime rates than those
are also more responsive
more likely to stay within
and central
of crime on falling city population
are much more responsive
Households
crime,
that rising crime rates in the suburbs
using measures
estimates
with
between
to crime. than those
Crime-related leaving
of the out-
the city for
other reasons. Increases
in crime rates and the subsequent
of costs on those residents
is the direct
the quartile double
costs of increased
reported
crimes
for cities in the bottom
and Miller,
Cohen,
incorporating
out-of-pocket
and lost life,
Valuing
1993
who remain
of cities with the greatest
from 0.065
unchanged (1 988)
city residents
has victim
21 The total life. If the value falls to $6,100,
costs,
estimated
The most obvious
increase
to O. 127;
(0.069
(1 993)
physical
in 1973
victim
associated
the typical with
25
years,
crime roughly essentially
in 1993).
Cohen
costs per crime,
damages,
lost productivity,
crime reported
it,21 The difference
to be $1,0
cost to city
rates were
to 0.074
and psychological
million,
life is assumed
crime
estimate
a number
Over the last twenty
in crime saw per capita
costs of crime are sensitive
of a human
from the city impose
crime victimization.
quartile
a lost life at $2.7
costs of $10,200
behind.
per capita
and Rossman
exodus
to the value
in large cities
in crime
assigned
million dollars,
in
increases
to a human
the cost per crime
between
top and bottom
Since the estimated reported
In addition
to the direct
in property
elasticity
of housing
long run (1 O year) employment
with
elasticity
of the city population) low quartile
is capitalized
into housing
-- an additional
capitalized
concentrations
crime-related
large,
category,
percent
were
(Smith
1978,
if the direct
22 Declining
mobility
in central
In 1975, 50 percent
high income.
6, the increases
victimization
$16,000
of poverty
income
additional
values
values
causing
(1 -5 year)
in population
intermediate
between
decline
between
in median
1978,
two
and a
and
top and bottom
will also decrease Thaler
for housing,
further
levels of approximately declines
into a $5,500
imposes
find the short-run
an elasticity
differential
Property
crimes
those
quartile housing
cities
(7
values
in
if the disamenity
and Hellman
costs of crime computed
above
two
of crime
and Naroff are fully
.22
Because
particularly
Assuming
translates
cities.
per year.
only those
migration
the demand
to employment
and taking
migration
crime-related
and Katz (1 992)
respect
per person
sets of cities includes
decreases
close to zero.
the crime-related
high versus
1979)
value
have similar effects
estimates, percent
costs of crime,
Blanchard
$600
in costs may be even greater.
city population
values.
into almost
the two
the true difference
Decreased
costs on cities.
cities translate
crime gap between
to the police,
declines
quartile
roughly were
is concentrated cities,
30 percent classified
values
cost to city residents, cited
of central
to be
fell into our low
and the remaining
would
by income
of the direct
20
group in Table
have caused
of the disamenity
are a reflection
26
city residents
income
cities by 1993
above.
do not appear
of net-migration
due to capitalization but rather
the rich, crime will increase
the effects
as middle
Using our estimates
in crime in the top quartile
housing
although
among
the fraction
do not entail costs of
an
of high income rise to 31.6 magnitude
residents
percent,
to fall to 18,2
While
it is unlikely
could be substantively
households
imposes
extent
peer effects
that
Benabou
1993,
negative
Cutler
percent
and the fraction
that the real economic
large, it is possible
externalities
are important and Glaeser
on those
to human
1995),
of low income effects
that the decline who remain
capital
the exodus
to
of a change
of this
in the number
of rich
in the central
development
residents
city.
To the
(Case and Katz
of skilled residents
1991,
hurts those
who
stay. Fiscal stress
on city government
mobility
patterns.
services
such as police protection,
hospitals, spending
In particular, on police,
crime difference per capita private
Low-income
health
23 These
top and bottom
quartile
Public education
represented
estimates
public transportation,
U.S. cities,
cities translates
costs are greater educational
from city finance
and city increases
in future
we estimate
that the
into an additional
$30
for low-income
programs.24
Combining
data for the years
1976,
the
1981,
either the current crime rate, the five-year lagged change in crime, or both. We include same covariates as in Section II of the paper. When entered separately, the coefficient
the on
per capita
is 696
same regression, individually,
is 648
(SE= 225)
(SE= 262).
When
and the coefficient
significant,
on five-year
both the level and changes
both enter positively.
statistically
in police expenditure
who
on
per capita
lead change
children,
by
per capita
crimes
the five-year
with
public
The poor are also less likely to be covered
in special
are obtained
We regress
and differential
need for locally-provided
from the 59 largest
in police expenditure.23
insurance.
and 1991.
will have greater
public education,
Using panel-data
are disproportionately
1986,
residents
high and rising crime rates are associated
between
annually
is likely to be linked to rising crime
The two
coefficients
Full regression
results
lagged
changes
in crime are included
are jointly, are available
in crime in the
but not from the authors
on request. 24 Crime
also carries
with
generally not borne by cities. If there are increasing governments,
declining
it large costs for courts, returns
populations
jails, and, prisons,
to scale in goods and services lead to inefficiencies 27
in producing
but these
provided
by city
such goods.
are
estimated
declines
expenditure,
in housing
for a fixed
between
top and bottom
property
tax revenue.25
about
14 percent
Wilson
1983),
Empirical
property
quartile
In comparison,
returns
levels of police
of differential 40%
of the mean
in Massachusetts
after the passage
rates
city’s
anticipated
of Proposition
fell by 50 percent
crime
losing
2 X (Ladd and
in California
after
13.
however,
constant
into roughly
municipalities
tax revenues
of Proposition
tax rate, the impact
cities translates
of total tax revenue
evidence,
by roughly
effective
and property
implementation
value due to crime with the increased
suggests
that the production
to scale (Bergstrom
of these
and Goodman
goods is characterized
1973,
Gonzalez
and Mehay
1987), 25 Crime income within
taxes,
is also likely to reduce on a per capita
the SMSA
Consequently, proportionately increase overstate
suggests
other sources
basis,
that
many
of city revenues,
such as sales taxes
The fact that most of the crime-related will continue
to work
in the central
movers
city.
the tax losses through these other channels is likely to be smaller than via property taxes. Also, to the extent that state governments
aid to cities experiencing
fiscal stress
the true costs borne by city residents. 28
(Yinger
and Ladd 1989),
this figure
and
remain
may
Table
1: Summary Mean
Variable
392986
City Population City Crime Suburban
Per Capita Crime
Statistics
for Panel Data,
Std. Dev. 727,406
Minimum 94,320
1976-1993 Maximum 7,530,493
.090
.025
.037
.207
.047
,015
.013
.125
.066
.025
.015
.208
18,521
2,916
11,028
29,159
.229
.174
.004
.836
.272
.025
,197
.375
Per Capita SMSA
Unemployment
Rate State
Real Income
Per
Capita O\OBlack State
Age Distribution:
0-17
years
18-24
years
.121
,013
.085
.154
25-44
years
.300
.029
.233
.364
45-64
years
.190
.013
.139
.231
.117
.020
.075
.186
.022
.012
.004
.087
.019
.011
.001
.082
.004
.032
-.169
.265
.0006
.0084
-,0471
.0540
-.0001
.0056
-.0280
.1377
.0012
.0030
-.0132
.0283
.0011
.0036
-.0301
.0213
65 years
and up
State Prison Commitments/ Reported State
Crime
Prison Releases/
Reported
Crime
A In(City
Population)
A (City Crime
Per
Capita) A (Suburban
Crime
Per
Capita) A Prison Commitment Rate A Prison Release
Rate
Notes U.S.
to Table central
occasional
1: Sample
cities with missing
and release
is comprised
population
data,
number
data are defined
of annual observations
greater
than
100,000
of observations
at the state
is equal to 2,165.
level and are changes
All other data are defined
at the city level unless otherwise
available
annually
YO high school graduates,
0/0 black,
which
decennial
census
except
are available years.
only in census
years.
from
1976-1993
in population
for 137
in 1975.
Due to
Prison commitment
in rates per reported
noted.
All variables
YO home-owners, % black is linearly
crime.
are
YO manufacturing, interpolated
and
between
Table
2: OLS Estimates
of the Relationship
Variable A (Per Capita
(1)
(2)
-1.039
-1.035
City Crime)
(.076)
A (Per Capita Suburban Per Capita Crime
City Population
(3)
(.075)
-.992
-1.024
-1.123
(.0786)
(.078)
(.082)
(.086)
.525
.424
.448
.499
.440
(.104)
(.105)
(.108)
(.121)
(.124)
City
----
----
----
----
Rate
Per Capita Os (-1) (-1 )
YO aged 0-17
(-1)
18-24
YO aged 25-44
YO aged 45-64
‘AA state f-l)-,
----
----
(-1 )
(-1)
% aged
(6)
.539
Unemployment
YO black
(5)
(.104)
Crime
lncome*l
(4)
Rate
-1.135
(-1 )
Suburban
and Crime
Crime)
Per Capita
State
Between
(-1)
(-1 )
(-1 )
population
.002
-.058
-.052
(.032)
(.066)
(.069)
-.046
.082
,148
(.056)
(.127)
(,134)
-.045
-.139
-.228
-.046
-.146
-.232
(.032)
(.050)
(.058)
(.032)
(.050)
(.062)
.11
.43
.38
.24
.48
.64
(.57)
(1.28)
(2.48)
(.58)
(1.30)
(2.54)
-.034
.028
-.019
-.033
.039
-.014
(,004)
(.039)
(.042)
(.005)
(.039)
(.043)
-.031
-.594
-.193
-.046
-.574
-.107
(.085)
(.177)
(.280)
(.086)
(.180)
(.283)
.568
-.668
-.285
.504
-.723
-.416
(.127)
(.286)
(.396)
(.133)
(.295)
(.401)
-.045
-.952
-.704
-.086
-.949
-.689
(.076)
(.260)
(.362)
(.078)
(.267)
(.369)
.207
-.601
-.073
.167
-,631
.023
(.151)
(.242)
(.410)
(.154)
(.257)
(.418)
.543
.634
(.096)
.356
.560
.644
.334
(.097)
(.129)
(.194)
(.124)
(.191) . .
No
Yes
Yes
No
Yes
Yes
No
No
Yes
No
No
Yes
.310
.390
.490
.315
.393
.495
City Fixed Effects? Region-year interactions R-squared Notes:
Dependent
observations 100,000
from
variable
is Aln(city
1976-1993
in population
for 137
in 1975.
population). U.S.
Number
Sample
central
of observations
Year and region dummies
are included
in all columns
interactions
All variables
are available
linearly
are present.
interpolated
between
about the level of geographic heteroskedasticity
decennial
census
is comprised
cities with
is equal to 2,165
greater
than
in all columns.
except those where region-year annually, except for YO black, which is See Table 1 for further information years.
disaggregation of the covariates. Standard errors in parentheses.
by city size.
of annual
population
Estimates
control
for
Table
3: First-stage
and Reduced-form Releases
on Crime
A per capita
(1)
Variable
Crime
of State
Prison Commitments
and
Rates and Citv Populations
city crime
(3)
0/0 A city population
(4)
(5)
(6)
-.405
-.393
-.283
.457
,317
.274
Per
(.085)
(.087)
(.108)
(.300)
.295)
(.368)
-.192
-.173
-.027
.988
.945
.999
Per
(.095)
(.097)
(.117)
(.336)
.330)
(.404)
.413
.407
.282
-,689
.742
-.703
(.076)
(.078)
(.090)
(.270)
(.264)
(.312)
A Prison Commitments
(2)
Estimates
(-1)
A Prison Commitments Crime
(-2)
A Prison Releases Per Crime
(-1 )
A Prison Releases Per Crime
(-2)
.303
.300
.157
-.464
-.544
-.857
(.014)
(.076)
(.091)
(.262)
(.256)
(.311)
Unemployment
-.003
-.008
-.052
-.041
-.129
-.187
Rate*
(.009)
(.014)
(,017)
(.032)
(.048)
(.061)
.23
.66
.06
-.14
.34
-1.95
(.16)
(.36)
(.07)
(.55)
( 1.20)
(2.36)
100
(-1)
State Per Capita Income*IOG (-1] ‘A black (-1)
YO
aged 0-17
YO aged
18-24
YO aged 25-44
YO aged 45-64
(-1)
(-1 )
(-1 )
(-1 )
.003
.010
.001
-.035
.034
-.005
(.001)
(.011)
(.012)
(.004)
(.037)
(.039)
-.546
-.284
(.171)
(.267)
.007
.009
-.044
-.025
(.023)
(.051)
(.076)
(.083)
-.057
-.119
-.019
.601
-.577
-.278
(.036)
(.084)
(.111)
(.126)
(.283)
(.390)
-.038
-.073
-.091
.013
-.742
-.391
(.021)
(.077)
(.101)
(.074)
(.258)
(.353)
-.016
-.020
-.360
.236
-.486
.519
(.042)
(.070)
(.110)
(.147)
(.239)
(.386)
-,052
-.084
-.023
.616
.718
.349
(.027)
(.036)
(.055)
(.031)
(,123)
(.189)
City Fixed Effects?
No
Yes
Yes
No
Yes
Yes
Region-Year
No
No
Yes
No
No
Yes