Three Essays on The Economics of Crime Prevention
Marcelino Guerra
5/18/23
The Areninhas Project
Fortaleza is the fifth largest Brazilian city and is the state capital of Ceará, located in the Northeast region of the country
The city has 120.6 sq mi of territorial area and population around 2.7 million - similar to Houston and Chicago
- From 2010 to 2020, Fortaleza’s population grew by 10%
Census 2010 divides the city into 3,043 census tracts - a.k.a neighborhoods - and 3,020 are populated
- An average tract has 812 people in 0.105 sq km
In July 2014, The City of Fortaleza began an ongoing urban renewal project called “Areninhas”
The intervention consists of synthetic football turf, a playground and an outdoor gym. The surroundings are improved with further infrastructure development
The project targets vulnerable communities and aims to provide an amenity that promotes physical activity and stimulates a sense of community
In 8 years, 103 football arenas were built in Fortaleza and more than 160 in the rest of the State
There are two types of equipment - Areninhas type 1 and 2. They differ by the size of the turf field and presence of locker rooms and bench
Each equipment have three employees. Two guards (one in the morning, one in the evening) control access to the fields, and one janitor cleans and organizes the place.
According to the City Hall, maintenance costs (salaries, energy and water) range from R$ 120,000 to 125,000/year.
Currently, 102 neighborhoods have an equipment - around 94,000 people covered
City Hall and State Government invest R$ 24.87 millions/year in four social projects: Esporte em 3 tempos, Futpaz, Esporte superação, and Atleta Cidadão
Projects directly impacts around 32,000 children
Amateur football championships are held on weekends, dance/fitness classes occur regularly, and pick-up football games happen at daily basis
Data and Context
Fortaleza-CE commonly ranks as one of the most violent cities in the world, with homicide rates close to Cali-COL, St Louis-USA, New Orleans-USA, and Baltimore-USA
Between 2004 and 2015, homicide rates more than double. Many factors contributed to the violence escalation: from a police strike at the end of 2011 to the rise of crime syndicates
In 2019 the city managed to cut murder rates by more than 50%, compared to the previous year. However, in February 2020, the Military Police went on a general strike again, and the State went back to the top of the murder rates ranking
Detailed information about 12,081 murders within Fortaleza’s boundaries between January 1st, 2012 and November 30th, 2019
The 8,291 homicides that happened on the streets are considered in the analysis
5,560 murders occurred between 8:00 am and 10:00 pm
Murders are aggregated to months and at the census tract level. The outcome is homicide rates per 1,000 people
Note: complete addresses were converted to coordinates (lat and lon) using Google’s API
Research Design - Murders
I conjecture tracts treated until July 2019 and after January 2020 are similar in observables and unobservables, and this control group’s choice produces an apples-to-apples comparison
Potential control areas within 650 meters of a treated neighborhood are excluded (2), as well as fields built between September and December 2019 (11)
Different comparison groups are also built by matching similar never treated neighborhoods using neighborhoods and boroughs’ demographics available in Census 2010
In total, there are 31 treated and 55 control neighborhoods
Total population in treated areas is 28,365 and in control tracts is 51,017
This sample covers 3.2% of the city’s population and around 6% of land area
To account for endogeneity concerns and evaluate the impact of the Areninhas Project, I combine the difference-in-differences design with a comparison group based on areas treated after the sample final year and/or matched places.
The causal effect of the neighborhood intervention is estimated using the approach proposed by Wooldridge (2021):
\[\text{Murder Rate}_{im}=g_{m}+c_{i}+\beta_{q}(w_{it}.fq_{t})+...+\beta_{T}(w_{it}.fT_{t})+\varepsilon_{im}\text{ } \text{ } \text{ }\text{ }\text{ } (1)\]
where \(g_{m}\) is the treated at month \(m\) cohort fixed effects (month-year), and \(c_{i}\) refers to census tract fixed effects. The main idea of ETWFE is that consistent estimations for ATT’s can be obtained by allowing for full cohort and timing heterogeneity.
ETWFE splitting the sample: Time-of-Day, Day-of-Week, Age, Gender, and Criminal record heterogeneity
To estimate gender, day-of-week, age and criminal record-specific effects, I consider the following:
\[\text{Murder Rate}^{j}_{im}=g_{m}+c_{i}+\beta_{q}(w_{it}.fq_{t})+...+\beta_{T}(w_{it}.fT_{t})+\varepsilon_{im}\text{ } \text{ } \text{ }\text{ }\text{ } (2)\]
where \(\text{Murder Rate}^{j}_{im}\) is the outcome measured in tract \(i\) for deceased individuals of gender/of age/with criminal-record/at day-of-week or time-of-day \(j\) in month \(m\).
Results Crime I
ETWFE point estimate suggests that the urban renewal policy caused a decrease in murder rates decreased around 2/3 in treated neighborhoods
S & A (2021) confidence intervals rules out reductions smaller than 26%
The figure shows ETWFE and S&A (2021) dynamic treatment effects. Months are binned to years, and endpoints are also binned at -4 (or less) and 2 (or more). Coefficients are normalized to event time -1
Day-of-week columns show a larger effect during weekdays, and time-of-day analysis points to a bigger effect during the afternoon (from 12 pm to 5:59 pm) - when most social projects happen
Pupils that study during the afternoon and commute from home to school around 1:00 pm and from school to home between 5:00-5:30 pm would benefit the most
One advantage of ETWFE over other estimators is that it can also be applied using methods other than linear regression, including count models
Poisson estimates show that treatment effects accumulate over time. During the first 24 months, murder occurrences decreases are 19 and 39%, respectively, but point estimates are noisy. > 36 months estimate represents a 63% decrease in murder counts
Results Crime II
Temporal displacement is estimated using equation (1) but with murders happening during closed hours - from 22:00 pm to 7:59 am
Point estimate suggests an increase around 35% in homicide rates, but it is not statistically significant at usual levels (p-val=.49)
Neighborhoods are built around treated census tracts using the equipment’s distance to the other tracts’ centroids. The homicide data covers the opening hours
Point estimates are not statistically significant, and there is no evidence of crime moving around corners
The table shows the results using different sample choices
Point estimates are similar to the ones using the preferred sample and show that results are not driven by neither treated areas at the begging nor the end of the program
Online Appendix contains results from different control group choices using matching. Results are fairly similar across different samples and point to average murder rates reductions between 57% and 67%
Cost-Benefit Analysis
Back-of-envelop calculation shows that the 31 treated areas are having 18.86 fewer homicides every year. Extrapolating the effects to all 102 areas covered by 103 football fields, there are 62.68 fewer murders/year due to this policy. I follow Pereira, Almeida and Oliveira (2020), who attach R$ 1.119 million for a blue-collar Brazilian men
Amenity value is based on private turf field rentals at an hourly rate of R$ 100, considering 56 hours/week available for practice
Building costs vary from R$ .24 to R$ 1.7 million depending on the type of the equipment, and maintenance costs are composed by salaries, water and energy. There are four social projects connected to areninhas with an estimated cost of R$ 24.87 million/year
Appendix
Time line of Events
Crime information from November \(29^{th}\), 2011 to January \(10^{th}\), 2012
Comprehends 22 treated SAs (Ceará), and 50 SAs as control - 26 in Pernambuco, 24 in Alagoas
6 days of strike, around 70%-90% decrease in the number of Police officers on the streets
![]()
Crime information from January \(1^{st}\), 2012 to February \(18^{th}\), 2012
52 treated SAs in Bahia, and 50 SAs as control (Pernambuco and Alagoas)
12 days of strike, around 33%-50% decrease in the number of Police officers on the streets
![]()
Identification and Estimation I
DiD Estimates
Murder increased by 123.7% and 78.38%, while robbery grew by 202.3% and 43.59% in Ceará and Bahia, respectively
There is no reaction of motor theft, but motor robbery increased 5-fold and 96.5% in Ceará and Bahia, respectively
Murder and robbery elasticities range from [-1.56, -1.37] and [-2.24,-0.87]
QDiD Estimates
There are more cops allocated in the capital and metro area
Crime concentrates in few security areas
Treatment effects are much higher at upper percentiles
Takeaways
Violent crimes have a huge cost to society and this paper shows the essential role of deterrent effect on crime when police patrolling levels are low
- In particular, criminals switch to violent methods when \(p_{j}\) was very low
Time dynamics show perceived risk of arrest matched reality as time passed
Large elasticities associated to fewer cops suggest non-linear effects of police presence on crime
Daily monetary damages range from R$ 12.23 (US$ 6.58) to R$ 16.16 (US$ 8.69) million
Appendix
Data and Context
Police Blitzes
3,423 blitzes identified in 548 cells between 2012 and 2013
Crime data
Violent crimes (homicide and robbery) at street level
Allocation of Blitzes I
A blitz interrupts the flow of vehicles and people through a physical, visual, and audible warning. Police officers proceed with checks and inspections in selected targets
This sudden increase in policing (5-10 policemen) in a street segment usually lasts between 3 and 6 hours
The Bureau of Police Operations met every Friday to decide where to allocate blitzes in the following week. That decision was mainly based on the past spatial distribution to create residual deterrence
Allocation of Blitzes II
Identification and Estimation
As mentioned before, it is challenging to disentangle the effect of police on crime. I deal with the simultaneity using police crackdowns that essentially tried to surprise drivers in space and time and do not depend on shocks to crime levels. The reduced-form model that captures the direct effect of the increase in policing on crime:
\[ ln(\lambda_{idt}) = \delta Blitz_{idt}+\theta Blitz^{2}_{idt} +\rho WBlitz_{idt}+ \Gamma lag(Blitz_{idt}, dt-j) +c_{ipw}+ \alpha_{d} + \varepsilon_{idt} \\\]
The use of Poisson distribution is justified because the dependent variable \(\text{Crime}_{idt}\) is the number of crime occurrences (count variable) measured at cell \(i\) and day-time of day \(dt\). \(Blitz_{idt}\) refers to the number of hours spent by police officers in cell \(i\) at day-time of day \(dt\), \(WBlitz_{idt}\) is the weighted average of blitzes hours in the surrounding area of hexagon \(i\) at day-time of day \(dt\), and \(lag(Blitz_{idt}, dt-j)\) represents the hours of policing work at cell \(i\) at day-time of day \(dt-j\)
Conditional on daily and cell-by-day-of–week-by-time-of-day fixed effects, the allocation of blitzes in a small area \((0.126 km^{2})\) at a given hour of day is treated as good as random and used to identify the causal effect of local police interventions on crime. We rule out simultaneity exploiting the design of the intervention
The main neighborhood specification considers a spatial weight matrix \((W)\) that is based on the inverse distance of the cells’ centroids. Distance cutoffs set the “areas of influence” of the police intervention
The weights of the \(W\) matrix are defined as \(w_{ij}=\cfrac{1}{d_{ij}}\), where \(d_{ij}\) is the distance between census tracts \(i\) and \(j\)
To capture residual deterrence and temporal displacement, the reduced-form model incorporates part of the history of blitzes with temporal lags of hours spent by police in the past days (distributed lag model)
Results
Blitzes have a meaningful and statistically significant direct effect on violent crime occurrences, and an additional hour spent by this policing assignment causes an average decrease of 21% on daily crime counts in cells at a given 6-hour period of day. There are diminishing returns of public safety to hours spent by the police in a single area
We do not find spatial displacement/diffusion of benefits
There is a small temporal displacement in the next 6 hours after an intervention, but residual deterrence cancels it in the next 2-3 days
Mechanisms
To partially disentangle these mechanisms, we explore the information about the number of motor vehicles and motorcycles seized during each intervention
If incapacitation is the main channel through which blitzes reduce violent crime occurrences, we would see larger effects for blitzes that seize more vehicles, which is not the case
