Phyton Replication Exercise
Description
6 attachmentsSlide 1 of 6attachment_1attachment_1attachment_2attachment_2attachment_3attachment_3attachment_4attachment_4attachment_5attachment_5attachment_6attachment_6.slider-slide > img { width: 100%; display: block; }
.slider-slide > img:focus { margin: auto; }
Unformatted Attachment Preview
Econ 191
Replication Exercise
Due November 10th
Replicating a paper is a great way to learn how empirical work is practiced. In this exercise,
we ask you to read a paper published in economics, and then we will guide you in replicating
the main results. You may use whatever software you like, although programming hints will
be provided for Stata only. Many of the commands that we ask you to run in Stata are also
available in Statas dropdown menu. You can use it, but make sure to write each command
in a do-file as we have shown in class.
Please hand in, via bCourses, three files: (i) your log file, (ii) your do-file, (iii)
and a concise write up of your answers to the questions below.
You are welcome to work in groups of up to 3 but you must each upload your own submission
to bcourses and write the names of people you worked with at the top of your write up.
1
Overview of Paper
Lets begin by motivating and reviewing the research design employed in the paper. Around
2-4 sentences is sufficient for each of the questions in this section. Precision and brevity will
be rewarded!
1. Imagine you observe cross-sectional data across a city for a given month. In particular,
for each district in the city, you observe the average number of police officers on duty
each day (police_officers), the total number of car thefts (car_thefts) and a set of
district-level characteristics (X1 , X2 , …, XK ). You run the following OLS regression of
car thefts on the number of police officers, controlling for district characteristics.
1
car_theftsd = ? + ?police_officersd +
K
X
?k Xk,d + ?d
k=1
Does the estimated ?b capture the causal effect of police presence on car thefts? Why or
why not? If ?b is a biased estimate of the true causal effect, in which direction do you
think it is biased?
2. Summarize the research design/identification strategy employed in Di Tella & Schargrodsky (2004). What variation in police presence does their estimator use and how
does this uncover the causal effect of police presence on car thefts?
3. What is the key identifying assumption in their research design1 ? Describe what this
identifying assumption means in this particular setting and evaluate its plausibility.
2
Setup
1. Download the dataset MonthlyPanel2 from Bcourses.
2. Open Stata, set up the directory, and create a new do file to keep track of your steps.
3. Open a log file called replication.txt and load the data MonthlyPanel2.dta.
4. Describe the data to understand the different variables.
3
Data Creation
1. Create a categorical variable named category that takes the value of 1 if there is a
Jewish institution on the block (variable distanci equals 0); the value of 2 if location is
one block from the nearest institution (variable distanci equals 1); the value of 3 if the
location is 2 blocks away; and the value of 4 if it is more than two blocks away from the
nearest Jewish institution.
2. Drop the observations for which mes takes the value of 72.
1
2
Hint: The key identifying assumption for all difference-in-differences designs.
The authors use mes 72 and 73 as different periods from the second half of July, but for simplicity we
aggregated those into mes 73 for the variable totrob2
2
2
3. Generate a variable month that takes the same values as mes, but takes the value of
mes+1 if mes is above 7.
4. Replace the variable month with the value of 8 if the value of month is 74. (Hint: Use
the tab command to make sure your month variable goes from 4 to 13).
5. Use the label define and label values commands to label the months (4 should correspond
to April, .., 7 to July(1-17), 8 to July(18-31), 9 to August, … 13 to December). (Hint:
Use the label define and label values commands.) 3
6. Save the database as DataClean.dta.
Descriptives4
4
1. Construct a graph that shows the evolution of the average number of car thefts over
the months by the distance category you constructed.
To do that, it is useful to first collapse the data using the mean of the car thefts
(totrob2 ) by month and category.
Then, use the twoway and connected commands to make plot the evolution. Label
the categories according to the distance, name the y-axis Average number of car
thefts, the x-axis Months and title the graph Evolution of Average car thefts by
category.5
Export the graph and include it in your write up.
2. Replicate columns (A)-(D) of Table 2 of the paper.
Use the database DataClean.dta.
Use the estpost and tabstat commands to store the mean and standard deviation
of the total number of car thefts by month (totrob2 ), for each one of the categories
depending on the distance to the nearest Jewish institution (The variable you
constructed). 6
3
Refer to https : //stats.idre.ucla.edu/stata/modules/labeling ? data/ for help
After each item, there is a suggested way of constructing tables/figures. If If you use a software different
from Stata, or prefer using different commands, no need to follow the steps, just replicate the graph and table
and answer Q4.3
5
Hint: We did a very similar think in the Rep. Exercise Tutorial.
6
i.e: estpost tabstat totrob2 if category==x , statistics(mean sd) by(month) nototal
est store ax
4
3
After storing, use the esttab command to export the file using a rtf format, naming
the columns No Policy in place. Depending on your version of stata, your code
should look like:
1) esttab a1 a2 a3 a4 using table2.rtf, cells(Mean(fmt(5)) ///
SD(par fmt(3))) collabels(“No Policy in place”) replace label
2) esttab a1 a2 a3 a4 using table2.rtf, main(mean) aux(sd) ///
collabels(“No Policy in place”) nostar replace label
Compute the number of blocks and paste it on the table after opening it with
Word. (Hint: you can use the summarize command of the block id variable by
category restricting to one particular month, or the unique command for the block
id variable restricting to a particular category and month).7
Change the name of the columns, and make a title and a footnote to make it more
similar to the table on the paper. Include the table in your write up.
3. Looking at the months before the terrorist attack, what can you conclude about the car
thefts on the blocks closer/further from the Jewish institutions? Are they comparable?
What can you say about the difference after the attack? Explain.
5
Diff-in-Diff
1. Replicate columns (A)-(C) of Table 3 of the paper.8
Drop observations from July(18-31) (i.e. those for which month == 8)
Create a post-treatment time indicator. Namely, generate a variable called post
that equals 1 if and only if month > 8.
Label the variables you just constructed according to the names on Table 3 of the
paper.
Consider the specification in the first column of Table 3 9 .
7
i.e. unique observ if month==4 & category==y
Again, we present a suggested way to replicate the table, but feel free to use your own software/code.
9
See p191 of paper.
8
4
Car Theft
it
= ?0 Same Block Police
it
+ Mt + Fi + ?it
The code for running and outputting this regression looks like this:
gen same_block = (distanci==0)
gen same_block_police = same_block*post
areg totrob same_block_police i.month, absorb(observ) robust
outreg2 using table3.doc, keep(*_block_police) replace lab nocons
Use this code block and write the rest of the code to replicate columns (A)-(C) of
Table 3, which you can export to a .doc. Include the table in your write up.
2. In column (A) of Table 3 we see that the coefficient on Same-Block Police is -0.07752.
What is the interpretation for this?
3. Describe what we learn from Table 4. Why do you think the authors include this table
in their paper?
5
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/4981049
Do Police Reduce Crime? Estimates Using the Allocation of Police Forces After
a Terrorist Attack
Article in American Economic Review · March 2004
DOI: 10.1257/000282804322970733 · Source: RePEc
CITATIONS
READS
454
2,029
2 authors:
Rafael Di Tella
Ernesto Schargrodsky
Harvard University
Universidad Torcuato di Tella
101 PUBLICATIONS 13,353 CITATIONS
55 PUBLICATIONS 3,085 CITATIONS
SEE PROFILE
All content following this page was uploaded by Rafael Di Tella on 23 May 2014.
The user has requested enhancement of the downloaded file.
SEE PROFILE
Do Police Reduce Crime? Estimates Using the Allocation of
Police Forces After a Terrorist Attack
By RAFAEL DI TELLA
AND
ERNESTO SCHARGRODSKY*
An important challenge in the crime literature is to isolate causal effects of police
on crime. Following a terrorist attack on the main Jewish center in Buenos Aires,
Argentina, in July 1994, all Jewish institutions received police protection. Thus, this
hideous event induced a geographical allocation of police forces that can be
presumed exogenous in a crime regression. Using data on the location of car thefts
before and after the attack, we nd a large deterrent effect of observable police on
crime. The effect is local, with no appreciable impact outside the narrow area in
which the police are deployed. (JEL K42)
Classical criminology assumes that criminals
are rational beings who weigh the costs and
bene ts of their actions. Gary Becker (1968)
produced the rst fully edged theory of crime
based on rational behavior. His research led to
an upsurge of interest in the economics of criminal behavior [see, for example, Isaac Ehrlich
(1973), Ann Witte (1980), Ehrlich and George
Brower (1987), James Andreoni (1991), Richard Freeman (1996), Steven Levitt (1997),
Pablo Fajnzylber et al. (2000), inter alia]. One
of the central predictions of Beckers theory is
that crime will decrease when police presence
increases. A basic problem with this prediction
is that it has largely failed to nd empirical
support. In a survey of the literature, Samuel
Cameron (1988) reports that in 18 out of 22
papers surveyed researchers found either a positive effect of police presence on crime or no
relationship between these variables. More recent surveys by Thomas Marvell and Carlisle
Moody (1996) and John Eck and Edward Maguire (2000) reach similar conclusions.
There is, however, a serious endogeneity
problem with these studies that arises from the
simultaneous determination of crime and police
presence (see Franklin Fisher and Daniel Nagin,
1978). It is likely that the government of a city
in which the crime rate increases will hire more
police of cers. Areas beset by high crime will
thus end up with more police of cers than areas
with low crime rates, introducing a positive bias
in the police coef cient in a crime regression. A
central challenge in the crime literature has been
to break this endogeneity in order to identify
causal effects of police on crime.
Two recent papers use a time-series strategy
to address this problem. Using data for the
United States, Marvell and Moody (1996) nd
Granger-causation between crime and police
running in both directions. In a similar vein,
Hope Corman and H. Naci Mocan (2000) exploit high-frequency data for New York City to
show that increases in the number of police
of cers cause a reduction in one out of ve
crime categories (speci cally, burglary). Monthly
data are used because hiring and training delays
in the response of the police authority to an
increase in crime will mitigate simultaneity bias
present in low-frequency data. In order to
* Di Tella: Harvard Business School, Boston, MA 02163
(e-mail: rditella@hbs.edu); Schargrodsky: Universidad Torcuato Di Tella, Min?ones 2177, (C1428ATG) Buenos Aires,
Argentina (e-mail: eschargr@utdt.edu). We thank a coeditor, an extremely constructive referee, Jushan Bai, Sebastian Galiani, Erzo Luttmer, Robert MacCulloch, Sam
Peltzman, Andrea Rotnitzky, several key informants, and
seminar participants at the 2001 AEA New Orleans meetings, the University of CaliforniaBerkeley, Stanford University, Econometric Society, LACEA, UTDT, UdeSA,
Getulio Vargas, AAEP, and UNLP for helpful suggestions.
The second author thanks SCID at Stanford University for
their hospitality. Matias Cattaneo, Luciana Esquerro, and
Magali Junowicz provided excellent research assistance.
The database and computer programs used in this paper are
available at http://www.people.hbs.edu/rditella and www.
utdt.edu/ eschargr.
115
116
THE AMERICAN ECONOMIC REVIEW
validly address the simultaneity concern, these
identi cation strategies depend, crucially, on
the assumption that the authorities are unable to
forecast crime- ghting needs.1
Levitt (1997) develops a different approach
using instrumental variables to break simultaneity. He documents the presence of an electoral
cycle in police hiring and uses the timing of
gubernatorial and mayoral elections to instrument for police presence in a panel of 59 large
U.S. cities from 19701992. Using two-stage
least-squares (2SLS) techniques, Levitt nds a
negative and signi cant effect of police on violent crime. The pattern across individual crime
categories is surprising, with murder exhibiting
the largest (and the only signi cant) coef cient,
and with very imprecise estimates for the categories in which the rational model is presumed
to be more relevant (e.g., property crimes). Still,
the validity of the instrument might be questioned. The timing of elections may affect crime
by way of channels other than the number of
police of cers on the street. Levitt avoids some
of these concerns by controlling for the unemployment rate and public spending, although
police effort and crime reporting (as well as
police hiring) may also respond to the timing of
elections, particularly if the police are the target
of political manipulation. Similarly, the behavior of judges and prosecutors may be affected
by elections, something that could logically reduce criminal activity during such times.2
A more severe concern raised by Justin McCrary (2002) is that Levitts 2SLS estimates
suffer from a computational error (see also Levitts reply, 2002). When the mistake is corrected
the replication results show no effect of police
on crime at standard signi cance levels. The
1
Criminologists often emphasize the bene ts of anticipating crime patterns. David Bayley (1998), for example,
states The key assumption behind smarter law enforcement
is that crime is not evenly scattered through time and space.
Police are not faced with meeting all crime threats everywhere all the time. Instead, each form of crime displays a
particular pattern which, if understood, provides opportunities for law enforcement (Bayley, 1998, p. 174). On the
allocation of police resources to protect high crime areas,
often called hot spots, see Lawrence Sherman et al.
(1989) and Sherman and David Weisburd (1995).
2
On the incentives faced by members of the judiciary
see, for example, Richard Posner (1993).
MARCH 2004
state of the evidence leads Levitt (2002) to
wonder: If electoral cycles can provide no
more than suggestive evidence of a causal impact of police on crime, are there other identi cation strategies that can do better?
In this paper we present a different approach
to estimate the causal effect of police on crime.
On July 18, 1994 terrorists exploded a bomb
that destroyed the Asociacion Mutual Israelita
Argentina (A.M.I.A.), the main Jewish center in
Argentina. Eighty- ve people died and more
than 300 were wounded in the attack. One week
later the federal government assigned police
protection to every Jewish and Muslim building
in the country. Because the geographical distribution of these institutions can be presumed to
be exogenous in a crime regression, this hideous
event constitutes a natural experiment whereby
the simultaneous determination of crime and
police presence can be broken.3
We collected information on the number of
motor vehicle thefts per block in three neighborhoods in Buenos Aires before and after the
terrorist attack. The information covers the
nine-month period beginning April 1 and ending December 31, 1994. We also collected information on the location of each Jewish
institution in these neighborhoods. We then estimated the effect of police presence on car
theft. Our difference-in-differences estimates
show that blocks that receive police protection
experience signi cantly fewer car thefts than
the rest of the neighborhoods. The effect is
large. Relative to the control group, car thefts
fall by 75 percent in the blocks in which the
protected institutions are situated. However, the
effect is extremely local. We nd no evidence
that police presence in a given block reduces car
theft one or two blocks away from the protected
buildings.
There has been considerable interest in identifying the mechanisms by which police presence reduces crime. Is it that police presence
makes criminal activity less attractive (deterrence), or is it that police of cers apprehend
criminals leaving fewer of them around to com3
On natural and randomized experiments, see the discussions in Robert LaLonde (1986), Joshua Angrist (1990),
Angrist and Alan Krueger (1991), Daniel Hamermesh
(1999), and Bruce Sacerdote (2001).
VOL. 94 NO. 1
DI TELLA AND SCHARGRODSKY: DO POLICE REDUCE CRIME?
mit crimes (incapacitation)? Being based on
changes in crime levels in particular locations
(i.e., the protected blocks) our results are unlikely to re ect changes in the numbers of incarcerated criminals, which should affect all
neighborhood blocks, not just those containing
Jewish institutions.4 Thus, we believe that our
estimates are most appropriately interpreted as
the causal deterrent effect of police staf ng on
car theft. However, it is still possible that car
thefts were displaced in a way that we are
unable to measure, in which case the effect of
policing may be smaller than our estimates
suggest.
The rest of the paper is organized as follows.
In Section I we describe our data. In Section II
we discuss the empirical strategy and present
our results. Section III concludes.
I. Data Description
On July 18, 1994 a terrorist attack destroyed
the main Jewish center (A.M.I.A.) in Buenos
Aires, Argentina.5 Seven days later, on July 25,
the federal government decided to provide 24hour police protection to more than 270 Jewish
and Muslim institutions (including synagogues,
mosques, clubs, cemeteries, and schools) in Argentina. Muslim institutions were protected for
fear of potential retaliations after the Islamic
organization, Hezbollah, claimed responsibility
for the attack. Nearly ten years after the attack
this protection is still provided.
A signi cant proportion of the protected
buildings are Jewish institutions within Buenos
4
Daniel Kessler and Levitt (1999) use Californias sentence enhancement laws for a selected group of crimes to
distinguish between incapacitation and deterrence. See also
Levitt (1998). Articles studying responses to increases in
detection probabilities include Avner Bar-Ilan and Sacerdote (2001) on red light violations, and Robert McCormick
and Robert Tollison (1984), on fouls committed by basketball players.
5
This was the second terrorist attack in the city of
Buenos Aires. The Israeli embassy had been destroyed on
March 17, 1992. In the months immediately following this
rst attack, the most prominent Jewish centers, including
A.M.I.A., had been given more attention by of cers on
patrol. But surveillance was not generalized and declined
gradually. Information on these attacks can be found in
www.atentado-amia.com.ar, www.daia.org.ar, and www.
bnaibrith.org.
117
Aires proper.6 Although providing this surveillance required the distraction of a nonnegligible
proportion of the police forces protecting the
areas in which these buildings are located, the
police forces made a serious effort to maintain
previous levels of police presence in the rest of
these neighborhoods. Government of cials
worried that compromising police protection
throughout the neighborhoods might generate in
the residents ill feelings towards the Jewish
community.7 Because the personnel commitment could not be met with the normal number
of police assigned to these neighborhoods, the
increased police presence was achieved with
of cers reassigned from administrative tasks at
the Central Police Department, the Communications Division, and the Mounted Police.8
The data analyzed in this paper are from three
noncontiguous, Buenos Aires neighborhoods
that collectively represent about 3.2 percent of
the citys area and account for 6.9 percent of its
population. One police station is located in each
neighborhood.9 The neighborhoods were selected on the basis of three criteria: they were
the areas with the largest numbers of Jewish
institutions in the city;10 signi cant portions of
the neighborhoods were not close to a protected
institution (more than 50 percent of blocks are
more than two blocks removed from a protected
6
Approximately 85 percent of the Jewish population of
the country lives in Buenos Aires and its suburbs.
7
Institutional information for this paper was gathered
through a series of interviews with key informants, including the Secretary of Security (third level of authority in the
federal government, behind the president and ministers), the
Chief of the Federal Police, and the Minister of the Interior
during the period under consideration as well as a former
federal judge, a former federal prosecutor, and the director
of a nongovernmental organization devoted to protecting
civil rights.
8
For example, more than one-third of approximately 200
police of cers stationed in Once, one of the neighborhoods
with the highest density of Jewish institutions, had to be
reassigned to protection duties. The personnel necessary to
maintain the previous level of police presence in the rest of
the neighborhood was pulled from outside of this police
station.
9
There are 53 police stations in Buenos Aires. Adrian
Pelacchi (2000) provides an in-depth discussion of the institutional features of crime and the police force in Argentina.
10
There are no Muslim institutions in the neighborhoods
considered in our study.
118
THE AMERICAN ECONOMIC REVIEW
MARCH 2004
FIGURE 1. TIMELINE OF EVENTS
institution), providing a control group for our
study; and three was the maximum number of
police stations for which we were able to convince police authorities to provide us data.11
There are a total of 876 blocks in these three
neighborhoods. The block constitutes the unit of
observation for our study.12
We obtained all the information available to
the police (with the exception of the victims
name) about each auto theft in these neighborhoods for the nine-month period starting April
1, 1994 and ending December 31, 1994. Figure
1 presents a timeline of the events in our study.
April 1 to July 17 constitutes the period before
the terrorist attack. The interim period of July
18 to July 31 includes a rst week during which
surveillance had not yet been introduced and a
second week during which police began to implement the protection policy. By the end of the
last week of July police protection was fully
functioning and known to the public. Finally,
August 1 to December 31 covers the period of
police protection.
Although victims tendency to underreport
often results in of cial records underestimating
crime levels, this is a minor problem for car
thefts in Buenos Aires for two reasons. First,
police intervention is required to activate car
insurance against theft, a type of insurance carried by most car owners in Buenos Aires (89
percent according to the of cial victimization
survey, Ministerio de Justicia, 2000). Second,
because criminals often use stolen cars in the
commission of other crimes, victims who report
car thefts to police forestall confusion about
their involvement in such crimes. The victimization survey cited above reports that 87 percent of Buenos Aires car thefts are reported to
the police, compared to only 29 percent for all
types of crime. A further advantage of auto theft
data is that this category of crime is expected to be
more sensitive to police presence.13 Most robberies occur after a brief period of surveillance of the
intended victim. Criminals concentrating their attention on mobile victims might miss the presence
of police. A parked car, on the other hand, gives
criminals time to gather information on areas in
which they intend to commit crimes.
Car theft information obtained from the police includes the address at which the stolen
vehicle was parked, make and year of the vehicle, day and time of the report, and whether the
robbery was violent. During the period of analysis 794 nonarmed car thefts were reported in
these neighborhoods.14 Although they normally
occur in the middle of blocks, car thefts in many
cases are reported at corners so as to facilitate
victims verbal descriptions of crime locations
at the time they le police reports. We assigned
one-quarter of each car theft reported at a corner
to each of the intersections four blocks.15
13
11
The police stations daily records, which register auto
thefts on the same pages as reports of every other type of
crime or incident, are not available to the public. The Chief
of the Federal Police had to issue a special authorization
instructing police station personnel to transcribe the data for
us.
12
We consider a block as the segment of a street between two corners. With few exceptions, Buenos Aires is a
perfect grid city, with streets crossing perpendicularly at
corners. Each block is about 100 meters (110 yards) long.
Ninety-four percent of Buenos Aires car robberies
occur in the street (Ministerio de Justicia, 2000).
14
We exclude a small number (63) of armed robberies
reported during this period as well as 86 misreports that
correspond to nonexisting or incomplete addresses or to car
thefts that occurred outside of our sample neighborhoods
(i.e., that were reported to the wrong police station).
15
This procedure assigns some fractions of thefts to
blocks outside the boundaries of the neighborhoods under
study, which reduces the total number of car thefts from 794
to 778.75.
VOL. 94 NO. 1
DI TELLA AND SCHARGRODSKY: DO POLICE REDUCE CRIME?
TABLE 1DEMOGRAPHIC CHARACTERISTICS
Demographic
characteristics
Home ownership rate
Overcrowding rate
Poverty rate
Education of household
head
Number of household
members
Female population
Unemployment rate
Age
Number of census tracts
OF
CONTROL
AND
119
TREATMENT AREAS
Census tracts without
Jewish institutions
(A)
Census tracts with
Jewish institutions
(B)
0.696
(0.008)
0.014
(0.001)
0.042
(0.003)
11.653
(0.147)
2.719
(0.023)
0.556
(0.001)
0.053
(0.001)
38.005
(0.128)
53
0.663
(0.017)
0.017
(0.002)
0.052
(0.008)
11.052
(0.300)
2.685
(0.054)
0.552
(0.003)
0.059
(0.003)
37.690
(0.223)
14
Difference
(C) 5 (A) 2 (B)
0.032
(0.019)
20.002
(0.003)
20.010
(0.009)
0.600
(0.335)
0.034
(0.059)
0.003
(0.003)
20.005
(0.003)
0.315
(0.258)
Notes: Columns (A) and (B) present the mean of each variable for census tracts without and
with Jewish institutions in our sample. Column (C) presents the differences of means.
Standard deviations are in parentheses. Home ownership rate is the percentage of owneroccupied houses. Overcrowding rate is the percentage of households with more than three
people per room. Poverty rate is the percentage of households with at least one unmet basic
need (overcrowding; four or more members per working member and household head
with low educational attainment; poor quality housing; school-age children not attending
school; or no fecal evacuation system). Education of the household head is the average
educational attainment of the household head in number of years. Female population is
the percentage of women in the total population. Unemployment rate is the rate of
unemployment for the population of age 14 or higher. Age is the average age of the
population.
Source: 1991 Population Census.
The completed data set included information
on the geography of these neighborhoods, in
particular, the precise location of each Jewish
institution. There are 45 protected institutions in
this part of the city. Thirty-seven of them are
within these neighborhoods, while the rest are
near the boundaries (but less than three blocks
away).16 The geographical distribution of
blocks, institutions, and car thefts is summarized in Table A1 in the Appendix.
Using information from the 1991 census, Table 1 compares socioeconomic characteristics
potentially related to crime victimization and
car ownership across areas without and with
Jewish institutions. The lowest level of aggre-
16
None of the protected institutions in our sample is
located at a corner.
gation for which census information is available
in Buenos Aires is census tracts (fracciones
censales), which cover approximately eight to
ten contiguous hectares. Tests of means reveal
no statistical differences between census tracts
that contain and do not contain Jewish institutions along the following dimensions: home
ownership rate, percentage of overcrowded
households, percentage of poor households,
number of household members, percentage of
women, employment rate, and age. The only
dimension along which these census tracts differed was years of education of the household
head: 11.65 and 11.05, respectively, for tracts
without and with Jewish institutions. We interpret these results as evidence that the surveillance policy was randomly assigned across
socioeconomic characteristics. Table A2 in the
Appendix compares demographics and car theft
120
THE AMERICAN ECONOMIC REVIEW
MARCH 2004
TABLE 2MONTHLY EVOLUTION OF CAR THEFT
Month
April
May
June
July (117)
July (1831)
August
September
October
November
December
Number of
blocks
More than two
blocks from
nearest Jewish
institution
(A)
Jewish
institution on
the block
(B)
One block
from nearest
Jewish
institution
(C)
Two blocks
from nearest
Jewish
institution
(D)
0.09955
(0.248)
0.10840
(0.235)
0.07853
(0.196)
0.03926
(0.145)
0.03926
(0.146)
0.11836
(0.287)
0.10176
(0.256)
0.12112
(0.267)
0.09623
(0.240)
0.10176
(0.268)
452
0.12162
(0.361)
0.08783
(0.205)
0.12837
(0.286)
0.02027
(0.069)
0.02702
(0.078)
0.04729
(0.175)
0.01351
(0.057)
0.06081
(0.215)
0.02702
(0.078)
0.02702
(0.078)
37
0.12111
(0.287)
0.07763
(0.181)
0.07763
(0.215)
0.05900
(0.210)
0.07298
(0.217)
0.06677
(0.219)
0.09006
(0.276)
0.09782
(0.260)
0.11024
(0.288)
0.11645
(0.278)
161
0.12278
(0.297)
0.09734
(0.259)
0.06969
(0.186)
0.03097
(0.141)
0.06858
(0.238)
0.12721
(0.304)
0.09845
(0.248)
0.08849
(0.236)
0.10176
(0.217)
0.10619
(0.225)
226
Difference
(E) 5
(B) 2 (A)
Difference
(F) 5
(C) 2 (A)
Difference
(G) 5
(D) 2 (A)
0.02206
(0.060)
20.02056
(0.035)
0.04983
(0.047)
20.01899
(0.013)
20.01224
(0.014)
20.07106
(0.031)
20.08825
(0.015)
20.06031
(0.037)
20.06921
(0.017)
20.07474
(0.018)
0.02156
(0.025)
20.03076
(0.018)
20.00090
(0.019)
0.01973
(0.017)
0.03371
(0.018)
20.05159
(0.021)
20.01170
(0.024)
20.02330
(0.024)
0.01400
(0.025)
0.01468
(0.025)
0.02323
(0.022)
20.01106
(0.020)
20.00884
(0.015)
20.00829
(0.011)
0.02931
(0.017)
0.00884
(0.024)
20.00331
(0.020)
20.03263
(0.020)
0.00553
(0.018)
0.00442
(0.019)
Notes: The rst four columns present the mean and standard deviation (in parentheses) of the number of car thefts for each
type of block per month. The average number of car thefts for July can be obtained by summing the subperiods. The last three
columns present the differences of means of columns (B), (C), and (D) relative to column (A), with standard deviations in
parentheses.
rates for the neighborhoods under study relative
to the whole city.
A key dimension in our empirical exercise is
the distance of each block in our sample to the
nearest Jewish institution, whether or not the
building is within our neighborhoods. We distinguish among blocks that contain a Jewish
institution, blocks that are contiguous in any
direction to a block containing a Jewish institution, and blocks that are two blocks away in any
direction from a block containing a Jewish institution. We then compare these with blocks
that are more than two blocks away from a
block containing a Jewish institution.
Table 2 presents means (and standard deviations) of auto thefts for each month for each
type of block. The bottom row tallies the number of blocks of each type. For the month of
July we consider, separately, the period before
and after the terrorist attack. For the post-July
period, the table shows that, relative to the control group (i.e., blocks more than two blocks
away from the nearest Jewis
