Bullock, Jessie; Pellegrino, Ana Paula
Article
How do Covid-19 stay-at-home restrictions affect
crime? Evidence from Rio de Janeiro, Brazil
EconomiA
Provided in Cooperation with:
The Brazilian Association of Postgraduate Programs in Economics (ANPEC), Rio de Janeiro
Suggested Citation: Bullock, Jessie; Pellegrino, Ana Paula (2021) : How do Covid-19 stay-at-home
restrictions affect crime? Evidence from Rio de Janeiro, Brazil, EconomiA, ISSN 1517-7580, Elsevier,
Amsterdam, Vol. 22, Iss. 3, pp. 147-163,
https://doi.org/10.1016/j.econ.2021.11.002
This Version is available at:
https://hdl.handle.net/10419/266980
Standard-Nutzungsbedingungen:
Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen
Zwecken und zum Privatgebrauch gespeichert und kopiert werden.
Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich
machen, vertreiben oder anderweitig nutzen.
Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten,
gelten abweichend von diesen Nutzungsbedingungen die in der dort
genannten Lizenz gewährten Nutzungsrechte.
Terms of use:
Documents in EconStor may be saved and copied for your personal
and scholarly purposes.
You are not to copy documents for public or commercial purposes, to
exhibit the documents publicly, to make them publicly available on the
internet, or to distribute or otherwise use the documents in public.
If the documents have been made available under an Open Content
Licence (especially Creative Commons Licences), you may exercise
further usage rights as specified in the indicated licence.
https://creativecommons.org/licenses/by-nc-nd/4.0/
H O S T E D B Y
Contents lists available at ScienceDirect
EconomiA
journal homepage: www.elsevier.com/locate/econa
How do Covid-19 stay-at-home restrictions affect crime?
Evidence from Rio de Janeiro, Brazil
☆☆,☆
Jessie Bullock
a,
⁎
,1
, Ana Paula Pellegrino
b,2
a
Department of Government, Harvard University, United States
b
Department of Government, Georgetown University, United States
article info
Article history:
Received 15 July 2021
Received in revised form 29 October 2021
Accepted 8 November 2021
Available online 1 January 2022
Keywords:
Crime
Violence
Extortion
Organized crime
Covid-19
abstract
How do changes in mobility impact crime? Using police precinct-level daily crime statistics
and shootings data from the state of Rio de Janeiro, Brazil, we estimate that extortion, theft,
and robberies decrease by at least 41.6% following COVID-19 mandated stay-at-home orders
and changes in mobility in March 2020. Conversely, we find no change in violent crimes,
despite fewer people being on the streets. To address the relationship between crime and
mobility, we use cellphone data and split the precincts into subgroups by pre-Covid-19-
related restrictions mobility quintiles. We estimate a similar average decrease in extortion
regardless of a precinct’s previous activity level, but find that the decrease in theft and
robberies is substantially higher for the more mobile precincts while it disappears for the
least mobile precincts. Using daily cellphone mobility data aggregated at the police precinct
level, we find that changes in mobility while the stay-at-home order is in place only have a
meaningful effect on robberies, which increase in likelihood when a precinct’s mobility
ranking is higher than the previous day. Together, these results suggest that the stay-at-home
order and associated decline in mobility strongly affected extortion and property crimes
while not interfering with the dynamics of violent crime. These findings support the hy-
pothesis that violent and property crime follow different dynamics, particularly where there
is a bigger impact of organized criminal groups.
© 2022 National Association of Postgraduate Centers in Economics, ANPEC. Production and
hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
https://doi.org/10.1016/j.econ.2021.11.002
1517-7580/© 2022 National Association of Postgraduate Centers in Economics, ANPEC. Production and hosting by Elsevier B.V. This is an open access article
under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
☆☆
Production and hosting by National Association of Postgraduate Centers in Economics, ANPEC.
☆
We thank the UNDP Policy Response Research Team and GRANDATA for providing access to the mobility data and for their inclusion in the ‘Exploring the
impact of COVID-19 and the policy response in LAC through mobility data’ project.” We thank the Harvard Political Economy of Development workshop for
helpful comments. All errors are our own.
]]
]]
]]]]]]
⁎
Corresponding author.
E-mail address: [email protected] (J. Bullock).
1
ORCID :0000-0003-2486-4439
2
ORCID :0000-0001-6058-0594
EconomiA 22 (2021) 147–163
1. Introduction
Immediately following COVID-19 mobility restrictions, commonly referred to as “lockdowns,”
3
crime changed around the
world as those who could stayed off streets and in their homes. Initial reports showed plummeting violent and non-violent
crime rates while domestic violence rose. Across the Global South and particularly in Latin America, international observers
predict that the pandemic strengthened the power of organized crime as they watched gangs provide social welfare for
community residents and, at times, enforced harsher mobility restrictions than the government (McDermott and Dudley, 2020).
In contexts where the government is wholly focused on managing the state’s response to the pandemic, it is possible that a
pandemic could be an opportunity for criminal groups.
We study the relationship between mobility and crime in the State of Rio de Janeiro, a quintessential case where criminal
groups and state wield overlapping influence. We first ask the question, how did stay-at-home restrictions impact both violent
and nonviolent crime? We study the impact of a mandatory stay-at-home order in Rio de Janeiro, Brazil on several types of
violent and nonviolent crime, as well as police killings. We then consider whether variation in quarantine compliance affected
violent and nonviolent crime differently.
We first compare crime levels in the 30 days before and after COVID-19 mobility restrictions were issued, exploiting detailed
daily police precinct-level data on shootings, lethal violence - including homicides and police killings – extortion, theft, and
robberies. We estimate this using an interrupted time series, also called a regression discontinuity in time (RDiT) design
(Hausman and Rapson, 2018). We then leverage granular data on mobility at the daily level, which we use to estimate variation
in levels of quarantine compliance, also at the police precinct level. We use this panel data to estimate the effect of changes in
mobility on crime levels once the stay-at-home order had already been enacted.
There are two main sets of results. First, regardless of model specification, bandwidth, or sample, we find that violent crime
did not change following the mandated restrictions, despite large reductions in mobility. This is inconsistent with many findings
from elsewhere around the world where violent crime plummeted following quarantine restrictions. It is also inconsistent with
early journalistic evidence, claiming that criminal groups halted violence in order to lock down and protect their communities
(Dalby, 2020; GloboNews, 2020).
If true, criminal groups merely halted increases in violence, as our results show there was no meaningful decrease in
shootings, homicides, or other forms of lethal violence. We find that police killings decreased slightly following the stay-at-
home order, but this effect dissipates over time and the correlation disappears when we consider mobility level of the precinct.
These results suggest that COVID-19 restrictions did not noticeably disrupt the dynamics of violence in Rio de Janeiro. We
interpret this as evidence that incentives and norms governing violent crime – predominately, thought not exclusively related to
organized crime – suffered little impact from mobility restrictions. They support the need for further investigation into the
relationship between mobility and incentives to engage in violent criminal behavior.
Second, we find that extortion and property crimes plummeted following mobility restrictions. In our preferred regression
discontinuity specification, we find that extortion decreased by 45.9%, theft by 69.4%, and robbery by 41.6% following the
mandate. This could be due to social distancing measures and fewer people on the streets (mechanically making it more
difficult to steal, rob, or extort from people), or could be due to criminal behavioral changes in the areas most likely for property
crimes to happen (which also happen to be the most mobile areas, or the areas at greatest risk for spreading COVID-19).
We take two strategies to address the possibility that the underlying mobility level was related to decreases in property
crime rates and extortion. We conduct a subgroup analysis using our preferred regression discontinuity specification, splitting
all police precincts in the state into quintiles that correspond to their pre-pandemic mobility level. The coefficient is similar in
magnitude across all five quintiles for extortion, indicating that extortion decreased following the stay-at-home order, but was
not very elastic to mobility levels.
On the other hand, there was no perceptible decrease in robberies and thefts among lower mobility police precincts, while
more mobile precincts registered decreases in robberies and thefts of a maximum of 38% and 37% in the upper quintiles,
respectively. Though this finding is partially indicative of the nature of the crimes, it suggests that the decreases in thefts and
robberies was mostly driven by the highest trafficked areas that locked down, whereas the decreases in extortion occurred
across all police precincts, from high to low mobility levels.
We then consider whether within-lockdown changes in mobility impacted crime levels, and if relative day-over-day changes
in mobility was related to crime. All correlations disappear except for that on robberies when our parameter of interest becomes
the relative change in precinct daily mobility levels. We find a positive but small relationship between robberies and increases
in mobility vis-a-vis the previous day.
Our paper is related to the broader literature on COVID-19 and the impact of mobility restrictions. First, it is closest to the
growing body of work on COVID-19, crime, and mobility (Ashby, 2020; Bullinger et al., 2021; Campedelli et al., 2020; Estevez-
Soto, 2020; Halford et al., 2020; Bullinger et al., 2021) in a variety of national and criminal contexts. Second, it speaks to others
that have studied the specific case of Rio de Janeiro’s response to COVID-19 and crime, including Bruce et al., 2021; Bullock,
3
The extent of “lockdowns” varied widely across and within countries. For a time series of within-country variation in the nature of lockdown restrictions in
this paper’s case, Brazil, see Brazil’s COVID-19 Policy Response from Oxford Blavatnik School’s COVID-19 Government Response Tracker. For a comparison with
measures adopted by other countries, see the comparative Stringency Index’s containment and closure policy indicators. We will refer to the mobility re-
strictions throughout the paper as lockdowns, keeping with local vocabulary.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
148
2022 and Monteiro (2020). These works consider in greater detail how criminal groups may have reacted to the pandemic. Our
findings are broadly consistent with those of Bruce et al., 2021), who argue that criminal groups known for extortion (in
Portuguese, milícias began forcing residents to reopen commercial establishments during the lockdown. If it is the case that
these groups suffered extortion rent losses, as our findings show, our stories are consistent with each other. Finally, we also
contribute to the literature on spatial dynamics of crime and its incentive structures (Weisburd et al., 2004; Tita and Griffiths,
2005; Bernasco and Block, 2009; Guerette and Bowers, 2009).
Section 2 introduces the context of crime in Rio de Janeiro and the government’s response to COVID-19. Section 3 introduces
our empirical strategies. We present results in Section 4, and conclude in Section 5.
2. Context
2.1. Crime in the state of Rio de Janeiro
A tragic feature of contemporary Rio de Janeiro is its high rate of crime and lethal violence. The homicide rate was 37.6 per
100,000 people in 2018 (Cerqueira et al., 2020), on par with the national rate for Honduras, Belize, or Venezuela, which rank
among the most murderous countries in the world. Property crimes are also high and have been rising in recent years, with
10,599 incidents of cargo robberies, 125,646 incidents of street robberies, and 54,366 incidents of vehicle robbery in 2017 alone
(Rolim, 2020). Lastly, the state of Rio de Janeiro ranks amongst the worst places in the world for police killings. Police kill
someone approximately every 10 hours oursin Rio de Janeiro (News, 2020) and are known for their brutal tactics that, in some
cases, provoke more violence (Magaloni et al., 2020).
Organized criminal groups play an important role in the spread of crime and violence throughout the state. In Rio de Janeiro,
there are several competing criminal groups that fight for territorial and market control – of both illegal and legal nature - as
well as control over voters and informal influence over police and elected officials. These frequent disputes have created a
constant state of turf wars, reputational disputes, or firefights with the police. Of the criminal groups present in Rio de Janeiro,
three major gangs concentrate their illegal drug trafficking and other criminal operations in informal settlements scattered
across the state. There are also several vigilante militia groups, who operate extensive protection rackets in the settlements they
dominate, extorting residents and local businesses for everything from weekly “security provision” to car parking spaces to gas,
cable, and electric utilities. While these groups originally formed to combat gangs and their illegal drug operations, they have
since entered this illegal market as well. Many violent crimes and extortion incidents are related to these groups, and nearly all
police killings are indirectly related to these groups (Magaloni et al., 2020). Even when an innocent bystander is assassinated by
the police, it usually happens during a targeted raid to arrest drug traffickers or when police are pursuing a member of a
criminal group (Andreoni et al., 2020). Specific groups are also involved in theft and robberies.
2.2. Rio de Janeiro’s response to COVID-19
The response to COVID-19 in Brazil was decentralized and poorly coordinated at the city and the state level.
4
Across the
country, state governors and mayors enacted a series of closure and COVID-19 containment measures that were soon to be
contested by the federal government – the disagreement ended up in the Supreme Court, that decided in favor of governors and
mayors. Rio de Janeiro State’s governor was one of the first in the country to close schools and suspend public events in early
March 2020. The city’s then mayor was initially reluctant to follow suit but ultimately also adhered to social distancing policies.
The timeline for COVID-19 mobility restrictions in Rio de Janeiro is as follows:
•
March 5: First Covid-19 case confirmed in the state, closure of state public schools.
•
March 13: Municipal
5
public school closure and state-mandated stay-at-home order.
•
March 27: Closure of all non-essential businesses by the Governor.
•
June 06: Governor relaxes Covid-19 non-essential business restrictions.
Only in early June, 12 weeks after the state-level stay-at-home order was enacted, did the governor begin to gradually relax
quarantine restrictions. Though there was variation in citizen compliance with the restrictions, overall trends show that mo-
bility decreased suddenly after the March 13 order, as shown in Fig. 1. Using GRANDATA mobility data (described below in
Section 3.2), we see that mobility plummeted in the city of Rio de Janeiro in mid-March, and gradually increased in the coming
weeks, though never quite to pre-pandemic levels. This pattern can be found throughout Brazil, regardless of specific dates
when public authorities issued orders. While the issuance and compliance to stay-at-home orders was irregular throughout the
country and often contradictory at the different levels of government, empirical data shows a clear change in mobility dynamics
following March 13, be it the result of personal restrictions or adherence to edicts.
4
See the Blavatnik School’s Brazil’s Covid-19 Policy Response project for more detailed information on measures and their variation between subnational
governments.
5
“Municipal” only refers to schools in the capital city of Rio de Janeiro. Other cities closed schools in different timelines.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
149
3. Empirical strategy
We use two strategies to estimate the impact of COVID-19 restrictions on crime. First, we estimate the initial effect of the
mobility restrictions and how quarantine changed crime. Second, we estimate how changes in mobility – once restrictions were
already put in place – may have been related to fluctuations in crime levels.
To begin, we estimate an interrupted time series model (also called a regression discontinuity in time) (Hausman and
Rapson, 2018) to estimate the effect of COVID-19 restrictions on crime and violence. These designs are growing increasingly
common in studies in the economics of crime field (Mummolo, 2017; Carr and Packham, 2020; Zoorob, 2020; Jassal, 2020; Ouss,
2020) and have been shown to be similar to experimental benchmarks (St. Clair et al., 2014). The specification is as follows, for
police precinct i and day t:
Crime Lockdown f days days Lockdown
u
( ) ( )
it t t t t d
i it
1 2
= + + + × +
+ +
(1)
where Y
it
represents the count of a particular crime or shooting registered in the precinct for a certain day (described in
greater detail below). The coefficient of interest is β
1
. Lockdown
t
is a dummy equal to one on after March 13 and f(days
t
)
represents linear, quadratic, and cubic functions that model time trends on either side of the treatment threshold in days, which
is the running variable. To account for the seasonality of crime, we include day of the week fixed effects (λ
d
), as well as police
precinct-level fixed effects (π
i
). Standard errors are clustered at the police precinct level.
To alleviate concerns about unequal compliance with quarantine restrictions or unequal risks of the spread of COVID-19
when comparing high-traffic areas to low-traffic areas, we conduct a subgroup analysis for each crime by mobility level (de-
scribed in greater detail in Section 3.2) of each police precinct. We subset the sample of police precincts into the five mobility
quintiles shown in Figs. 2 and 3 and run the same regression above to examine within-mobility level changes before and after
mobility restrictions.
Our second estimation strategy is focused on looking at day-to-day changes in crime. We are interested in how the day-to-
day changes in compliance with restrictions within a precinct affect crime and violence. We analyze how the day-to-day change
in mobility level of a police precinct is related to crime. We estimate the below panel fixed effects model:
Crime Mobility u
it
i t
d i it
1
, 1
= + + + +
(2)
In this model, we estimate regressions of the change in mobility from the prior day’s level (ΔMobility
i,t−1
) on each type of crime.
The coefficient of interest is again β
1
and the unit of analysis is the police precinct. We include day of the week fixed effects (λ
d
),
police precinct-level fixed effects (π
i
), and cluster standard errors at the police precinct level.
3.1. Crime and violence data
We measure the impact of quarantine restrictions on crime and violence using two primary sources of crime data. The first
contains official police reports, drawn from the Public Safety Institute (ISP) database. We obtained the daily police precinct-level
reports for all crimes in the state of Rio de Janeiro. There are 137 police precincts in the state. The second source is Fogo Cruzado,
a civic tech and data collection nonprofit that collects citizen-, media-, and government-reported data on shootings in the
greater Rio de Janeiro metropolitan area, verifies them, and publishes a georeferenced database of the shootings.
Fig. 1. Change in mobility following stay-at-home order. Note: This figure plots the daily average mobility level for the capital city of Rio de Janeiro vis-a-vis the
first day in the data set, March 1. A value of 0 implies that citywide out-of-home mobility was the same as March 1; a negative value means that out-of-home
mobility was lower. The stay-at-home order coincides with the large drop in early March. The minimum value occurred on March 19, one week after the
announcement of the stay-at-home order. Other municipalities in Rio de Janeiro state show similar trends. Source: GRANDATA.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
150
Our analysis considers six crimes or types of crimes taken from the ISP database. We first look at a set of three variables to
measure how lethal violence changed in response to the COVID-19 restrictions. For this, we analyze reported homicides, police
killings, and the ISP omnibus indicator of “lethal violence," which sums the daily totals of homicides, robbery followed by death,
and injury followed by death. We then look at three variables to measure how nonlethal violence and property crimes changed
in response to mobility restrictions. The first of these is a variable we constructed for extortion, summing the individual police
reports for both extortion and threats. The second variable we constructed is the total number of robberies, summing all of the
categories of robbery in the ISP database.
6
The third variable we constructed is the total number of thefts, summing all cate-
gories of theft in the ISP database.
7
We use the Fogo Cruzado data to analyze how shootings changed in response to mobility restrictions. The Fogo Cruzado
database captures a different quantity of interest than the ISP data, which all draws from police reports. Shootings are reported
by citizens, local leaders, or journalists to Fogo Cruzado, who then cross-check the data with verified sources (press, community
leaders, and their partners within law enforcement agencies) before eventually publishing and pushing them to their public
geocoded API (Perguntas Frequentes, 2019). Anyone that wants to report to Fogo Cruzado can do so anonymously on the cell
phone app, online, or via phone. Shootings are published once they are verified by the Fogo Cruzado team. Though there may be
some overlap in ISP and Fogo Cruzado data – especially if the shooting results in a death or injury, that medical staff are required
to report – some shootings will likely never be officially tallied, particularly if there is no subsequent incident to report to the
police. Yet shootings indicate that a situation could turn lethal, even if it has not done so already. For this analyses, we calculate
daily total shootings per police precinct so as to be comparable to the ISP data at the police precinct unit of analysis.
3.2. Mobility data
The second data source is related to mobility during COVID-19. We use mobility data taken from cell phone records to
construct estimates of how much out-of-home activity occurred before and after the stay-at-home restriction in each police
precinct. The cell phone mobility data was provided by GRANDATA, a private firm that calculated out-of-home estimates at
granular, sub-municipal levels.
8
Limitations of cellphone mobility data and comparison to other widely used datasets are re-
ported in the Appendix A.1.
Fig. 2. Average police precinct mobility quintiles. Note: This figure shows the distribution of average mobility level for precincts during the two-week period
before the pandemic began. As an example to aid in interpretation, the top 20% most active precincts, between the 80% and 100% dotted lines crossing the x-
axis, had average mobility levels of 90.5 or higher, calculated from the underlying hex-level mobility scores. The most active precincts were uniformly high
mobility, whereas the middling and lower mobile precincts included a mixture of high- and low- mobility hex cells. Source: GRANDATA and ISP.
6
These include cell phone robbery, bike robbery, cargo truck robbery, robbery of a business, car robbery, home robbery, robbery at an ATM or financial
institution, robbery while driving the victim to an ATM or financial institution, robbery of an ATM, bank robbery, robbery while on the street, and robbery while
in a bus.
7
These include bike theft, vehicle theft, cell phone theft, and theft while on the street or while on a bus.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
151
The smallest unit of analysis that GRANDATA reports is the H3 hex cell at the 8th level, (0.72 km
2
), which includes a few city
blocks in an urban setting. For each day, hex cells are ranked by percentile (0–100) of out-of-home mobility across the entire
state of Rio de Janeiro, where 0 indicates least mobile and 100 indicates most mobile. If a cell phone registered the same
location for more than eight consecutive hours, this was tagged as “home" and all home observations were dropped from the
dataset. Users with less than 10 events a day were also dropped from the dataset. Only out-of-home locations were used to
calculate mobility level. Hex cells with more out-of-home registries were ranked as higher, those with fewer, as lower. There are
16,745 city-day observation in the dataset for the state of Rio de Janeiro, spanning from March 1 to August 31st 2020.
We calculate the pre-pandemic mobility levels of the police precincts in the following way. First, we subset the data to just
March 1–12, the days before the stay-at-home order was issued. Second, for each hex cell, we calculate an average percentile
ranking. This is important because mobility may be quite different on a weekday versus weekend in some communities, and
one day’s percentile ranking may not reflect how active a few blocks are. Third, we categorize all hex cells for which police
precinct they are in or not. On average, there were 34 hexes per precinct. For hex cells that were on the boundary of two
precinct, we included them in both groups. Lastly, we calculated an average mobility level per precinct for the entire pre-period
by averaging the individual hex averages within each precinct.
Fig. 2 shows the distribution of pre-period averages per precinct. The first quintile with the lowest out-of-home mobility
levels included all precincts that had an average percentile ranking of 39 or lower. The second lowest quintile included all those
from 39 to 52, third from 52 to 68, fourth from 68 to 90, and the highest quintile included all from 90 to the maximum value of
100. Fig. 3 then shows the geographic distribution of the police precincts, colored by quintile. The red and orange colors
correspond to the most active police precincts, those in quintiles four and five, that registered the highest average number of
out-of-home pings per hex cell. Most of these correspond to the capital city (where police precincts are smaller in land mass
because they cover a denser population) and the city’s suburbs, although there is some within-city variation on which precincts
are the highest traffic. The yellow color mostly corresponds to precincts that cover medium-size cities, whereas the green and
blue correspond to smaller or rural cities further away from the capital.
We believe the pre-lockdown mobility levels can be useful for analysis because it captures a quantity of interest that
population density or population size does not. Not all people or cell phone users will live in a hex cell they pass through, but
Fig. 3. Geographic distribution of pre-lockdown mobility levels, by police precinct. Note: This figure shows all 137 police precincts in the state of Rio de Janeiro,
Brazil. Precincts are colored according to their average mobility level from March 1–12 before the stay-at-home order was mandated, where mobility is
measured as an average of the percentile rank of GRANDATA H3 hex cells within each precinct. Source: GRANDATA and ISP.
8
The authors were selected as part of an open call for research by the United Nations Development Program (UNDP), acting in partnership with GRANDATA
during the COVID-19 pandemic. For more information on the collaboration, see their joint press release.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
152
their cell phone will register a ping if they pass through an area they work, commute, run errands, etc. Some of these areas, such
as a downtown bus station or central business district, may not have a high density of residences but may be risky areas for
spreading COVID-19 due to the high level of activity in the area. The mobility data captures this trend, while residential po-
pulation data does not. The limitations of this data is in its format: by measuring all daily mobility levels by daily percentile rank
vis-a-vis the other hexes, rather than as a fixed value, it makes comparisons for each hex over time difficult. For this reason, we
averaged mobility levels over several days to leverage the data and create this pre-lockdown benchmark for each police pre-
cinct.
Fig. 4. Violent and property crimes after the stay-at-home restrictions. This figure represents the daily average per precinct for each type of crime reported by
the Institute for Public Safety (ISP), starting on January 1, 2020. The top series of plots shows all violent crimes, homicides, and police killings, while the bottom
panel shows extortions, thefts, and robberies. The lines are generated by an ordinarly least squares (OLS) regression without covariate adjustment on either side
of the March 13, 2020 date. The plots are aggregated at the precinct-day average across the entire state to aid in visualization, but in the following analyses the
unit of analysis is the police precinct-day, of which there are 137 observations for each day.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
153
4. Results
4.1. Effect of mobility restrictions on crime
Fig. 4 shows the our results graphically for six different types of crimes reported to the Public Safety Institute (ISP): the
omnibus lethal violence variable, homicides, police killings, extortion, theft, and robberies. Each plot shows the daily registries
for each type of crime across all police precincts in the state of Rio de Janeiro. The visual evidence for the crimes shown on the
top panel (lethal violence, homicides, police killings) is not indicative of large changes following the stay-at-home order. Both
lethal violence and homicides look like they may have increased slightly following mobility restrictions, though not at a level
that is statistically distinguishable from random noise. Police killings, also appear to have fallen, especially in the days im-
mediately following restrictions. Yet none of the violent crimes appear to have decreased at the magnitude or for the duration at
which those in the bottom panel have decreased. The bottom panel subplots representing extortion, thefts, and robberies
indicate a precipitous decline in property crimes committed following the March 13 restrictions.
Table 1 shows our core specification (Eq. (1)) in the top panel, regressing a variety of violent and non-violent crimes on the
interrupted time series estimator Lockdown. Our preferred model is a linear estimator that uses a bandwidth of 30 days before
and after the restrictions were announced (March 13), but we estimate models for a wider bandwidth at 60 days as well, shown
in the bottom panel. Given that the observations are at the granular day-police precinct level, we prefer the narrower sample
that just includes the days in the month before or after the restrictions began. Though there is no substantive difference
between the estimates in the 30-day and 60-day models in Table 1, we prefer the 30-day bandwidth to eliminate doubt that
changes in crime could be driven by something else.
9
All reported models include police precinct and day-of-week fixed effects.
We also report quadratic and cubic specifications of the functional form for both the 30- and 60-day bandwidths in Appendix
Tables A1 and A2, respectively.
The upper panel of Table 1 columns (1)–(3) shows that violent crime changes very little in the 30 days following COVID-19
restrictions. This is consistent with the lack of visual evidence in Fig. 4. Regardless of whether violent crimes were reported by
the police (homicides and other lethal crimes included in the “lethal violence" variable) or were monitored independently
(shootings are reported by CrossFire, independent of police), violent crime did not appear to change following quarantine
restrictions. Models with the 60-day bandwidth reflect the lack of change in violent crime post-lockdown, although there is
weak evidence that homicide and other lethal violence may have actually slightly increased in the 60 day bandwidth.
We take this as evidence supporting the view that violent conflict between criminal groups did not cease just because
restrictions were put in place. It is possible that criminal groups continued defending their territories or using force against
rivals, despite that stay-at-home order. Though there is no way to know from this data if reported homicide and lethal violence
are perpetrated by criminal groups, it is widely known that Rio de Janeiro’s criminal groups are the drivers of much of the city’s
lethal violence, whether they are fighting each other or fighting the police. For this reason, we are cautious about interpreting
the coefficient as a causal estimate of criminal violence, but we interpret it as suggestive evidence that restrictions alone did not
disrupt the dynamics of violence in a meaningful way.
Table 1
Effect of lockdown on crime and violence.
Dependent variable:
Shootings Violent Crimes Property Crimes
Lethal Homicides Police Extortion Theft Robbery
Violence Killings
(1) (2) (3) (4) (5) (6) (7)
30 Day Bandwidth
Lockdown -0.043 0.019 0.020 -0.025 ** -0.591 *** -0.682 *** -0.948 ***
(0.031) (0.018) (0.017) (0.011) (0.057) (0.082) (0.115)
Source Fogo ISP ISP ISP ISP ISP ISP
N 4453 6989 6989 6989 6989 6989 6989
R
2
0.226 0.094 0.088 0.080 0.300 0.243 0.646
60 Day Bandwidth
Lockdown -0.010 0.024 * 0.023 * -0.018 ** -0.614 *** -1.305 *** -1.457 ***
(0.021) (0.012) (0.012) (0.008) (0.040) (0.087) (0.090)
Source Fogo ISP ISP ISP ISP ISP ISP
N 8833 13,904 13,904 13,904 13,904 13,904 13,904
R
2
0.221 0.081 0.078 0.069 0.298 0.224 0.677
Note: All models estimate the effect of the stay-at-home restriction on daily precinct-level crimes from the Public Safety Institute (ISP) official crime statistics or
Fogo Cruzado’s shootings database. Models use either a 30- or 60-day bandwidth, a linear estimator, and control for police precinct and day of week. Clustered
standard errors at the police precinct level are shown in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.
9
In the pre-period, Rio de Janeiro’s summer, especially the Carnaval period, often registers higher rates of violent crime than the rest of the year. In the post-
period, several other important political events happened (a police killing that mobilized the Brazilian version of Black Lives Matter protests, impeachment
proceedings of the then-Governor of Rio de Janeiro that led to social unrest, etc.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
154
Police killings, shown in column (4), did appear to decrease slightly following the quarantine restrictions. When considering
just 30 days before and after the quarantine mandate began, our coefficient estimate of −0.025 for police killings translates into
3.4 fewer police killings per day in the entire state. The drop-off in police killings is especially acute around the cutoff point,
which is visible in Fig. 4. This decrease dissipates when considering a 60-day bandwidth. Though still statistically significant and
negative, the coefficient estimate for police killings of −0.018 translates to a decrease of 2.46 killings per day across the state.
There are a few possible explanations for why police killings decreased in Rio de Janeiro following mobility restrictions. First,
police themselves were infected early and at a high rate with Covid-19: around 8–10% of the police force contracted Covid-19
early in March and each precinct was operating with limited human resources (ISTOE, 2020). Reports show that one in three
police officers for the state of Rio de Janeiro had to be placed on sick leave at some point throughout 2020, and 65 officers died
due to Covid-19 (G1, 2021). The shortage of police officers may have affected their ability to conduct police raids, which require
a lot of manpower and are where most police killings occur (Magaloni et al., 2020; Andreoni et al., 2020). Second, police were
required to perform different tasks that were pandemic-related, such as limiting entry to public spaces, conducting traffic stops,
and monitoring occupancy at testing centers (ISTOE, 2020). The average officer may have spent less time doing crime control
(where they are more likely to use force) and more time doing public safety and enforcement duties. Finally, if there were fewer
people on the streets – including people committing acts of violence – then this may have lowered the likelihood that police
would encounter someone and use force.
10
Extortion, theft, and robbery all fell dramatically following the Covid-19 lockdown, as shown in columns (5)–(7). The
coefficient estimates translate into a decrease of 80.3 reported incidents of extortion, 92.8 fewer reported thefts, and 128.9
fewer robberies across the entire state. The estimates from the 60-day model show that the second month of quarantine
exacerbates these differences as extortion, theft, and robberies decrease even further. The coefficients for theft and robbery lose
their significance when modeling with a quadratic specification (Table A.1) and robbery loses significance when modeling with
a cubic specification (Table A.2). The decrease in registered extortion is robust to model specification or bandwidth.
One natural explanation for why extortion, theft, and robberies decreased so suddenly following COVID-19 restrictions is
that fewer people were out in public and businesses were closed during this time. When people began staying home, it de-
creased opportunities for thieves and robbers to steal cell phones, cars, or other valuables when away from home, and increased
the risk of breaking and entering a home because it was likely to be occupied. More complex robbery schemes – such as bank
robberies or cargo robberies – could be more exposed to police, in banks where they were not controlling adherence to COVID-
19 restrictions, or more visible, since there was less transit. Overall, there was a change in the opportunity structure for this type
of criminal behavior. Similarly, criminal specialists in extortion may have had a harder time collecting extortion payments when
non-essential businesses temporarily closed their doors or when people stopped using services. Journalists and other scholars
noted these as potential reasons for why extortion and property crime decreased during the pandemic (Monteiro, 2020; Bruce
et al., 2021). Finally, there is no reason to believe this decrease is due to changes in reporting by the population or processing
capacity on behalf of police. While under reporting is an issue in all these categories, thefts and robberies can be reported online
and police stations were kept open during lockdown. We expect changes in mobility to have no impact on incentives to report
crimes.
4.2. Effect of lockdown on crime, by pre-pandemic mobility level
We break down our results for each type of crime by mobility level to analyze differences in high activity areas compared to
low activity areas. For each police precinct, we use the pre-pandemic period (March 1–12) average mobility level for each
precinct, shown in Fig. 3. This is to account for within-crime variation in responses to the mobility restrictions. For instance, it is
possible that the most active of precincts (precincts that register the highest mobility levels pre-pandemic) locked down more
severely – and therefore changed daily life in a more noticeable way – than less active precincts, altering the likelihood of
different crimes being committed. In other words, we expect the change in mobility in a mostly non-essential commercial area
to be different than in police precincts of a more residential or rural makeup. If we expect a relationship between mobility and
the opportunity structure for violent and property crime, the impact of lockdown should vary between quintiles.
Table 2 breaks down the results for each type of crime by mobility quintile. The top panel shows the bottom 20% of precincts
that registered the fewest number of out-of-home cell phone pings in the pre-period, or what we call the “least mobile"
quintile. The bottom panel shows the top 20% of precincts that registered the highest density of out-of-home pings in the pre-
period, what we call the “most mobile" quintile.
The results for violent crime are consistent with those found in Table 1, providing more supporting evidence that violent
crime did not change following quarantine restrictions even when conditioning on previous mobility level of a precinct. Col-
umns (1)–(3) show that shootings, all incidents of lethal violence, and homicides did not change following COVID-19 restric-
tions regardless of mobility quintile. Fig. 5 illustrates this graphically. Panel A of Fig. 5 shows that the change in shootings by
quintile is indiscernible after lockdown for all precincts in the sample (The Fogo Cruzado database did not register any shootings
for the precincts in the lowest mobility quintile). This panel provides visual evidence that the near-zero coefficient on shootings
10
This explanation doesn't apply to police raids, where police seek out suspects to arrest and often execute them. Stray bullet victims are often victimized in
their homes.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
155
from Table 2 applies across mobility levels and isn't driven solely by one particular type of precinct. Panels B and C show that
the null effect on lethal violence and homicides is not driven by one particular type of precinct.
The (lack of) change across quintiles from the pre-pandemic period to post-lockdown period is constant, even though lethal
violence and homicides are slightly more frequent in the more mobile quintiles. Together, all three panels in Fig. 5 show that
COVID-19 restrictions and their impact on mobility – even in densely populated areas where the pandemic posed a greater
threat – did not change the incidence of violent crime. This can be interpreted as strong evidence that changes to mobility did
not alter the incentive or opportunity structure for violence.
In Column (4) of Table 2, we show our results for police killings by mobility quintile. These results add nuance to the findings
from Table 1, which suggested that police killings decreased slightly following COVID-19 restrictions. Though many of the
coefficients are also negative in Column (3), the only one that is statistically significant (and only at the p < 0.1 level) is that for
the police precincts in the 60–80% range of pre-pandemic mobility. According to Fig. 3, many of these precincts are not in the
capital, but are in the suburbs or surrounding cities in the capital. Fig. 6 shows that there is little to no change in police killings
after mobility restrictions in the bottom three quintiles (where police killings are rare in the first place), and the decreases in
police killings in the more mobile fourth and fifth quintiles appear modest and temporary. These results suggest that police
killings slightly decreased following the implementation of COVID-19 restrictions, but that these decreases were more no-
ticeable in the suburbs of the capital city.
Column (5) shows results for extortion and shows marked decreases in extortion rates across all mobility quintiles. The
coefficients range from a low of −0.449 to a high of −0.813, but all are highly statistically significant and visually noticeable in
Fig. 7. These results are similar in magnitude to the coefficients from Table 1, which indicates that the decrease in extortion was
of a similar magnitude across mobility quintiles. Even though extortion levels are lower in the least mobile quintile, they
showed an expressive decrease after restrictions, indicating that extortion decreased even in areas with lower population
density and lower out-of-home foot traffic. We believe this may be due to the commercial closings put in place with the
restrictions, making it more difficult for extortionists to demand rents from establishments that were temporarily closed. The
constant relationship across all quintiles indicates that relationship between extorted and the extorters does not depend on
mobility.
We then turn to the property crimes of theft and robbery, shown in Columns (6) and (7). There are two clear observations
from the regression coefficients by quintile in Table 2. First, the post-lockdown decrease in incidence of crime grows in
magnitude as the quintiles move from least active to most. In the least mobile quintile, the coefficient for both theft and robbery
is small and statistically insignificant. It is negative and significant for the 20–40% quintile, and increases in magnitude for each
subsequent quintile of activity level. Second, this is partially due to variation in pre-pandemic levels of theft and robbery in each
quintile. The higher mobility, more densely populated areas had higher baseline levels of theft and robbery to begin with, as
shown in Fig. 8. These two observations suggest that higher mobility areas were greater hotspots of theft and robbery before
COVID-19 restrictions, but that other, less mobile areas experienced significant decreases in theft or robbery relative to baseline
levels. For example, theft and robbery decreased by an average of 32% and 19% in the fifth quintile, respectively, while they
decreased by a larger 36% and 21% in the fourth or 37% and 38% in the third quintiles, respectively. The view that mobility
Table 2
Effect of lockdown on crime and violence, by mobility quintile.
Dependent variable:
Shootings Violent Crimes Property Crimes
Lethal Homicides Police Extortion Theft Robbery
Violence Killings
(1) (2) (3) (4) (5) (6) (7)
Quintile 1: Least Mobile
Lockdown – -0.004 -0.004 – -0.449 *** -0.070 -0.020
(0.028) (0.028) (0.112) (0.050) (0.031)
Quintile 2
Lockdown -0.006 0.037 0.041 -0.006 -0.549 *** -0.475 *** -0.348 ***
(0.064) (0.048) (0.047) (0.010) (0.121) (0.094) (0.121)
Quintile 3
Lockdown -0.049 0.010 0.011 0.001 -0.813 *** -0.384 *** -0.658 ***
(0.034) (0.034) (0.034) (0.012) (0.128) (0.101) (0.169)
Quintile 4
Lockdown -0.098 0.051 0.044 -0.062 * -0.749 *** -0.862 *** -1.097 ***
(0.065) (0.043) (0.043) (0.036) (0.151) (0.171) (0.273)
Quintile 5: Most Mobile
Lockdown 0.022 -0.022 -0.016 -0.041 -0.452 *** -1.245 *** -1.608 ***
(0.056) (0.036) (0.036) (0.034) (0.121) (0.289) (0.362)
Source Fogo ISP ISP ISP ISP ISP ISP
N 1565 1565 1565 1565 1565 1565 1565
Note: All models estimate the effect of the stay-at-home restriction on daily precinct-level crimes from the Public Safety Institute (ISP) official crime statistics or
Fogo Cruzado’s shootings database. Each panel estimates linear models for all six crimes reported (or shootings) within that mobility quintile using the 30-day
bandwidth. In Panel 1, there is no coefficient estimate for shootings or police killings because there were zero in both the pre- and post-lockdown period.
Clustered standard errors at the police precinct level are shown in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
156
(caption on next page)
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
157
Fig. 5. Violent crimes after the stay-at-home restrictions, by mobility quintile. Note: Each data point and curve shown in this figure is as described in Fig. 4. Each
quintile is a subset of the entire dataset corresponding to pre-lockdown mobility level. In the top panel, there were no shootings registered in the Fogo Cruzado
database before of after lockdown in the lowest mobility quintile precincts, therefore, only four plots are shown.
Fig. 6. Police killings after the stay-at-home restrictions, by mobility quintile. Note: Each data point and curve shown in this figure is as described in Fig. 4. Each
quintile is a subset of the entire dataset corresponding to pre-lockdown mobility level. The leftmost panel indicates that there were no police killings in any of
the precincts in the lowest mobility quintile.
Fig. 7. Extortion after the stay-at-home restrictions, by mobility quintile. Note: Each data point and curve shown in this figure is as described in Fig. 4. Each
quintile is a subset of the entire dataset corresponding to pre-lockdown mobility level.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
158
restrictions had a significant impact on theft and robberies across mobility levels (except for areas with lowest baseline activity
levels) is tenable, given these results. This indicates that mobility has a big impact on the incentive structure for property crime,
as documented in criminology literature on trajectories of crime (Weisburd et al., 2004; Bernasco and Block, 2009).
Fig. 8. Property crimes after the stay-at-home restrictions, by mobility quintile. Note: Each data point and curve shown in this figure is as described in Fig. 4.
Each quintile is a subset of the entire dataset corresponding to pre-lockdown mobility level.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
159
To summarize, in Tables 1 and 2, COVID-19 restrictions appear to be reliably correlated with decreases in extortion, theft,
and robberies. The effect of the stay-at-home order on extortion is not sensitive to the pre-pandemic activity level in a precinct,
and decreases across all precinct types. Decreases in theft and robbery are largest in higher-activity precincts. When breaking
down by activity level, we see that the decreases in theft and robbery post-lockdown are large in magnitude for all types of
precincts except those with the lowest activity levels. Violent crimes, however, do not appear to have reacted to the stay-at-
home order as strongly as nonviolent crimes. Across model specifications, bandwidths, and subsets by mobility level, we find no
evidence that shootings, lethal violence, or homicides changed following the implementation of quarantine measures. There is
slight evidence that police killings decreased, especially in the days immediately following restrictions, but these decreases
dissipate over time. When looking at the effect of these measures on police killings by mobility quintile, the relationship is weak
or non-existent, especially in the densely populated capital city where most of the police killings happen.
4.3. Relationship between changes in mobility and crime
We now turn to our panel results looking at how changes in mobility over time were related to changes in crime. We
estimate Eq. (2) in Table 3, restricting our sample to just looking at within-precinct variation in mobility and crime once COVID-
19 restrictions were put in place. The sample shown in this table begins on March 13, when the stay-at-home mandate was
issued, and ends 60 days after.
11
The coefficient β
1
for the change in mobility is reported for precinct-level shootings, as well as
the violent crimes and property crimes.
The way day-to-day mobility is calculated in our data is a relative measure, changes in ΔMobility do not just indicate changes
in mobility, they indicate changes in mobility relative to the surrounding precincts. For example, if a high traffic area that ranks
in the 95th percentile of activity in the days prior to the stay-at-home restrictions and dramatically reduces their mobility
following the lockdown order but at the same rate as all other areas, it is possible the area would still rank in the 95th
percentile, even though their mobility fell.
A change in relative mobility across time can measure of how responsive the incentive structure of a precinct is to such
change and how mobile the opportunity for crime – be it by changes to the trajectory of victims or of the perpetrator. In other
words, when it comes to mobility changes, we can expect some crimes to be more “sticky" than others. There are no significant
results for most reported crimes. Thus we may infer that violent crime and extortion are stickier in relation to territory in face of
changes in mobility.
The only noticeable relationship between a day-to-day change in mobility and crime is for robberies. An increase in a
precinct’s daily average mobility level by one percentile rank from the previous day corresponds to 0.01 more robberies. We
interpret this as evidence that higher trafficked areas facilitate robberies. Robbers and thieves have fewer targets when people
are not on the streets, or conversely, are under closer scrutiny when people are in their homes or businesses that are target
locations. This effect is most notable when mobility increases from one day to the next, bringing people onto the streets and
lowering the bar for a robber to commit a crime.
Yet we would expect a similar effect to appear for theft, which we do not see in the data. The main difference between the
two categories is that a robbery involves the use of violence. A closer analysis of the data could reveal if there is a relationship
between the level of violence employed in the robbery – the use of a handgun, for example, is a possible indicator of link to an
organized criminal group – and the stickyness of such crime. It may be the case that robberies that involve lower levels of
violence present the same relationship to relative changes in mobility as thefts. This merits further investigation.
Table 3
Effect of lockdown on crime and violence.
Dependent variable:
Shootings Violent Crimes Property Crimes
Lethal Homicides Police Extortion Theft Robbery
Violence Killings
(1) (2) (3) (4) (5) (6) (7)
ΔMobility 0.002 -0.001 -0.001 0.0005 0.0002 0.002 0.010 ***
(0.002) (0.001) (0.001) (0.001) (0.002) (0.002) (0.003)
Source Fogo ISP ISP ISP ISP ISP ISP
N 3965 8277 8277 8277 8277 8277 8277
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
11
We also extend the sample to the earliest date available before the stay-at-home mandate (March 1) and subset the sample even further to just 30 days after
the mandate. In either case, coefficients do not change in sign or substantive interpretation.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
160
5. Conclusion
In this paper, we used daily variation in mobility and crime to estimate how stay-at-home mandates and subsequent
changes in mobility affected crime. Our estimates focus on the state of Rio de Janeiro, which not only is a high-crime state in
Brazil, but is so because of the presence of multiple contesting criminal organizations.
We suggest that the presence of criminal organizations is why we saw the large decrease in extortion and property-related
crimes that are predicated on people being out of home or at work, yet observed no effect on violent crimes. Our estimates show
that extortion, theft, and robbery decreased sharply following mobility restriction orders, ranging from 41.6% to 69.4% de-
creases, yet violent crimes did not change and police killings only slightly fell, though that effect was temporary. The decreases
in extortion were nearly uniform across mobility levels, while the decreases in theft and robberies were more elastic to changes
in mobility. Only robberies appeared sensitive to day-over-day changes in mobility during the lockdown, increasing slightly
when a day’s mobility was higher than the prior.
These results tell a consistent story. Our findings suggest that the impact of mobility restrictions affected the number of
people on the street during commercial hours, the number of businesses open, and subsequently, the opportunities to rob, steal,
or extort for potential criminals. The way that potential victims of theft, robbery, or extortion move about the world was
restricted with the stay-at-home orders, and, therefore, these crimes decreased. Shootings, homicide, other lethal violence, and
police killings, however, are less random and are usually not predicated on the availability of extortion or theft targets or people
being “in the streets.” We show that this is particularly salient in the case of robberies, which are sensitive to day-over-day
relative change during the lockdown. With no evidence that violent crime is as sensitive to mobility and the availability of
potential targets, we have no reason to suggest that COVID-19 restrictions would have a large, negative impact on violent crime
rates.
These results could indicate that lethal violence is more latent to territorial and spatial dynamics of organized criminal
groups rather than to overall mobility levels in the population. Yet we would need further testing in order to confirm this
hypothesis. While current data reported by ISP nor Fogo Cruzado report gang or militia involvement, a possible test of this
theory is to compare the impact of COVID-19 restrictions on different types of crime across hexes under the control of different
groups. Indeed, we could expect violent confrontations to decrease if the pandemic triggered a ceasefire between the state’s
criminal groups or slowed the rate of turf wars, but anecdotal and journalistic evidence indicates that neither of these occurred.
Such findings will advance our understanding of how the presence of criminal groups changes the baseline level of violence in
territory, distinguishing them from other urban criminal dynamics.
Conflict of interest statement
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Acknowledgments
We thank the UNDP Policy Response Research Team and GRANDATA for providing access to the mobility data and for their
inclusion in the “Exploring the impact of COVID-19 and the policy response in LAC through mobility data” project. We thank the
Harvard Political Economy of Development workshop for helpful comments. All errors are our own.
Appendix A
A.1. Mobility data
Using cellphone data to measure mobility, especially amidst Covid-19 restrictions, has grown increasingly common. We
leverage a novel dataset, GRANDATA, that uses extremely disaggregated mobility data at the hexagon level, the smallest of
which is less than one square kilometer. To validate the GRANDATA measure, we compare it to one of the more well-known
datasets in Brazil that measures mobility at the municipality level. The comparison dataset, In Loco, uses a similar measurement
tactic to GRANDATA, based off of cell phone location pings. It also used pre-pandemic mobility data to estimate the home base
of each cellphone and from there calculated whenever the device left such area.
There are two important features that distinguish GRANDATA from the In Loco data. First, the dataset begins on February 1st
for In Loco, and only on March 1st for GRANDATA. Second, In Loco measures the level of “isolation", not a change in mobility.In
the GRANDATA dataset, a positive value indicates an increase in mobility relative to the reference date (March 1st) and a
negative value a decrease in mobility. The opposite is true for the In Loco data: a positive value indicates increases in isolation,
and vice versa. We inverted the In Loco index in Fig. A.1 to allow for easier comparison, using GRANDATA’s reported change in
mobility level at the municipality-level and comparing that with In Loco's municipality-level social isolation index. The direct
comparison shows that GRANDATA is more sensitive to increases in mobility in the city of Rio de Janeiro, both in the pre-
pandemic weeks and post-lockdown measures that lasted for months. Yet the overall trends during the critical lockdown
period, are quite similar, giving us confidence that the GRANDATA dataset is capturing underlying changes in mobility. The
difference between the two measures is related to how cellphone signals are interpreted.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
161
The limitations of using such datasets are well-known. First, they are aggregated indexes and the actual number of ob-
servations is unknown for each hex-day. The commercial nature of data-collection also means researchers are unable to as-
certain the continuity of methodology, since improvement rollouts are not necessarily reported. Additionally, each dataset has
their own limitation. In spite of these, there is no reason to believe in systematic bias in data collection when comparing cities in
Rio de Janeiro, where cellphone coverage and usage is more uniform among the population. In fact, the two datasets are closely
correlated. We were not able to access the sub-municipal level data from In Loco in order to run our analysis on this dataset.
A.2. Tables
See Table A.1.
See Table A.2.
Fig. A.1. Comparison of GRANDATA and In Loco mobility data in Rio de Janeiro. Note: This figure plots the mobility level for the city of Rio de Janeiro vis-a-vis
the March 1 in the GRANDATA and city-level In Loco datasets. A value of 0 implies that citywide out-of-home mobility was the same as March 1 (GRANDATA) or
February 1(In Loco); a negative value means that out-of-home mobility was lower than the reference date, a positive value means that it was higher. The stay-at-
home lockdown order coincides with the large increase in late mid to late March.
Table A.1
Effect of lockdown on crime and violence: quadratic specification.
Dependent variable:
Shootings Violent Crimes Property Crimes
Lethal Homicides Police Extortion Theft Robbery
Violence Killings
(1) (2) (3) (4) (5) (6) (7)
30 Day Bandwidth
Lockdown -0.152 *** 0.003 0.003 -0.062 *** -0.471 *** 0.106 -0.160
(0.048) (0.026) (0.026) (0.017) (0.088) (0.158) (0.186)
Source Fogo ISP ISP ISP ISP ISP ISP
N 4453 6989 6989 6989 6989 6989 6989
R
2
0.227 0.093 0.088 0.08 0.297 0.24 0.646
60 Day Bandwidth
Lockdown -0.060 * 0.018 0.018 -0.037 *** -0.535 *** -0.730 *** -0.916 ***
(0.033) (0.019) (0.018) (0.012) (0.061) (0.083) (0.121)
Source Fogo ISP ISP ISP ISP ISP ISP
N 8833 13,904 13,904 13,904 13,904 13,904 13,904
R
2
0.222 0.081 0.078 0.07 0.298 0.226 0.678
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
162
References
Andreoni, M., Londoño, E., Galdieri, D., 2020. Licença para matar’: por trás do ano recorde de homicídios cometidos pela polícia no Rio.The New York Times.
Ashby, M.P.J., 2020. Initial evidence on the relationship between the coronavirus pandemic and crime in the United States. Crime Sci. 9, 6.
Bernasco, W., Block, R., 2009. Where offenders choose to attack: a discrete choice model of robberies in Chicago. Criminology 47, 93–130.
Bruce, R., Cavgias, A., Meloni, L., 2021. Policy Enforcement in the Presence of Organized Crime: Evidence from Rio de Janeiro. SSRN Scholarly Paper ID 3678840.
Social Science Research Network.
Bullinger, L.R., Carr, J.B., Packham, A., 2021. COVID-19 and crime: effects of stay-at-home orders on domestic violence. Am. J. Health Econ.
Bullock, J., 2022. Limiting Aggressive Policing Tactics Can Reduce Police and Civilian Violence.
Campedelli, G.M., Aziani, A., Favarin, S., 2020. Exploring the Effects of COVID-19 Containment Policies on Crime: An Empirical Analysis of the Short-term
Aftermath in Los Angeles. Preprint. Open Science Framework.
Carr, J.B., Packham, A., 2020. SNAP schedules and domestic violence. J. Policy Anal. Manag. 〈https://onlinelibrary.wiley.com/doi/pdf/10.1002/pam.22235〉.
Cerqueira, D., Bueno, S., Alves, P.P., de Lima, R.S., 2020. Atlas de Violência. Technical Report. Instituto de Pesquisa Econômica Aplicada.
Dalby, C., 2020. Six Ways Coronavirus is Impacting Organized Crime in the Americas. Section: Analysis.
Estévez-Soto, P.R., 2020. Crime and COVID-19: Effect of Changes in Routine Activities in Mexico City. Technical Report. SocArXiv.
G1, 2021. Monitor da Violência: RJ é o estado com mais policiais mortos pela Covid no país c Rio de Janeiro c G1. G1.
GloboNews, 2020. Mesmo com quarentena, Brasil tem alta de 6% no número de assassinatos no 1 semestre.
Guerette, R.T., Bowers, K.J., 2009. Assessing the extent of crime displacement and diffusion of benefits: a review of situational crime prevention evaluations.
Criminology 47, 1331–1368.
Halford, E., Dixon, A., Farrell, G., Malleson, N., Tilley, N., 2020. Crime and coronavirus: social distancing, lockdown, and the mobility elasticity of crime. Crime Sci.
9, 11.
Hausman, C., Rapson, D.S., 2018. Regression discontinuity in time: considerations for empirical applications. Annu. Rev. Resour. Econ. 533–552.
ISTÓE, 2020. Covid-19 já matou pelo menos oito policiais no Rio – ISTOÉ Independente.
Jassal, N., 2020. Gender, law enforcement, and access to justice: evidence from all-women police stations in India. Am. Polit. Sci. Rev. 1035–1054.
Magaloni, B., Franco-Vivanco, E., Melo, V., 2020. Killing in the slums: social order, criminal governance, and police violence in Rio de Janeiro. Am. Polit. Sci. Rev.
552–572.
McDermott, J., Dudley, S., 2020. GameChangers 2020: Predictions for 2021 – Shaking Off the Virus. Section: Analysis.
Monteiro, J., 2020. O Comportamento dos Indicadores de Segurança Pública durante a Pandemia de Covid-19 no Rio de Janeiro. Working Paper. Ministério
Público. Rio de Janeiro.
Mummolo, J., 2017. Modern police tactics, police-citizen interactions, and the prospects for reform. J. Politics 1–15.
News, B., 2020. Rio Violence: Police Killings Reach Record High in 2019 – BBC News.
Ouss, A., 2020. Misaligned incentives and the scale of incarceration in the United States. J. Public Econ. 191, 104285.
Perguntas Frequentes, 2019.
Rolim, L., 2020. Segurança Pública em Números 2020. Technical Report. Instituto de Segurança Pública. Rio de Janeiro, Brazil.
St. Clair, T., Cook, T.D., Hallberg, K., 2014. Examining the internal validity and statistical precision of the comparative interrupted time series design by com-
parison with a randomized experiment. Am. J. Eval. 311–327.
Tita, G., Griffiths, E., 2005. Traveling to violence: the case for a mobility-based spatial typology of homicide. J. Res. Crime Delinquency 42, 275–308.
Weisburd, D., Bushway, S., Lum, C., Yang, S.M., 2004. Trajectories of crime at plces: a longitudinal study of street segments in the city of Seattle. Criminology 42,
283–322.
Zoorob, M., 2020. Do police brutality stories reduce 911 calls? Reassessing an important criminological finding. Am. Sociol. Rev. 176–183.
Table A.2
Effect of lockdown on crime and violence – cubic specification.
Dependent variable:
Shootings Violent Crimes Property Crimes
Lethal Homicides Police Extortion Theft Robbery
Violence Killings
(1) (2) (3) (4) (5) (6) (7)
30 Day Bandwidth
Lockdown -0.075 -0.031 -0.036 -0.022 -0.320 *** -0.666 *** 0.154
(0.065) (0.034) (0.033) (0.020) (0.115) (0.183) (0.250)
Source Fogo ISP ISP ISP ISP ISP ISP
N 4453 6989 6989 6989 6989 6989 6989
R
2
0.228 0.094 0.088 0.081 0.299 0.244 0.647
60 Day Bandwidth
Lockdown -0.098 ** -0.001 -0.001 -0.043 *** -0.506 *** 0.094 -0.284 *
(0.045) (0.024) (0.024) (0.015) (0.081) (0.119) (0.165)
Source Fogo ISP ISP ISP ISP ISP ISP
N 8833 13,904 13,904 13,904 13,904 13,904 13,904
R
2
0.222 0.081 0.078 0.07 0.298 0.229 0.679
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
J. Bullock and A.P. Pellegrino EconomiA 22 (2021) 147–163
163
