Exploring the Effects of COVID-19 Containment Policies on Crime: An Empirical Analysis of the Short-term Aftermath in Los Angeles
EExploring the Effects of COVID-19 Containment Policies onCrime: An Empirical Analysis of the Short-term Aftermath inLos Angeles
Forthcoming at
American Journal of Criminal Justice
Gian Maria Campedelli † , Alberto Aziani , Serena Favarin , † Corresponding Author: [email protected] University of Trento — Department of Sociology and Social Research, Trento (Italy) Università Cattolica del Sacro Cuore — Department of Political and Social Sciences, Milan (Italy) Transcrime - Joint Research Centre on Transnational Crime, Milan (Italy)1 st October 2020
Abstract
This work investigates whether and how COVID-19 containment policies had an immediate impacton crime trends in Los Angeles. The analysis is conducted using Bayesian structural time-series andfocuses on nine crime categories and on the overall crime count, daily monitored from January 1 st th th to March 16 th and from March 4 th to March 28 th Keywords:
Coronavirus; Bayesian Modelling; Social Distancing; Urban Crime; Causal Impact; Time-Series; Routine Activ-ity Theory; Crime Pattern Theory a r X i v : . [ s t a t . O T ] O c t ntroduction In the first months of 2020, California was one of the first States to be affected by the spread ofa new virus belonging to the coronavirus family, named Sars-CoV-2. On March 4 th , six cases ofCOVID-19 were confirmed in Los Angeles County rising the total number of cases for the countyup to seven. Following this, the Los Angeles County Board of Supervisors and the Department ofPublic Health declared a health emergency. From that moment on, the Los Angeles population hadbeen invited to adopt simple social distancing strategies that limit their exposure to others—e.g.,remaining home when sick—and to prepare for the possibility of more significant social distanc-ing requirements. The institutional intervention started to become stronger on March 16 th , withthe prohibition of all events comprising fifty or more attendees. On March 19th, the CaliforniaDepartment of Public Health further reinforced the containment strategy by ordering all individ-uals living in the State to stay at home (County of Los Angeles, 2020).Distancing measures simultaneously affect the daily routines and the social interactions ofmillions of people. Daily commuters are forced to spend their days at home; household membersshare the same living spaces throughout the entire day; people can connect to their peers onlytelematically. Yet, the impact of the lock-down policies, and of the virus itself, outreach the alter-ation of people’s everyday routines and social relations. In the medium term, the losses due to theslowdown of the economic system may transform into higher unemployment and destitution formany. The scale of social distancing and lock-down policies adopted to mitigate the deadly conse-quences of the COVID-19 constitutes an unprecedented instrument to investigate contemporarysocieties and the short-term changes in crime trends.Despite the rapidly growing attention of scientists to the consequences on crime of the COVID-19 public health emergency, examinations of the impact on different types of crimes of milderand stricter policies remain underdeveloped. Moreover, our understanding of how well certaintheoretical frameworks provide support in explaining immediate changes in the occurrence ofdifferent crimes in response to the pandemic is limited. This work attempts to address these gapsby investigating the extent to which measures taken to contain COVID-19 impact on nine crimecategories, and on the overall crime count, in the city of Los Angeles in the immediate aftermath oftheir promulgation. Specifically, we concentrate on a set of primarily instrumental crimes, namelyburglary, theft, shoplifting, robbery, and vehicle theft and on more ‘expressive’ offenses, namelybattery (simple assault), intimate partner assault, assault with a deadly weapon, and homicide. We rely on the categorization provided by Cohn and Rotton (2003), who state that that “Expressive crime in-volves [...] violence that is not directed at the acquisition of anything tangible or designed to accomplish anythingspecific other than the violent outcome itself. Assaults, disorders, and domestic violence are examples of expressivecrime. Instrumental crime [...] involves behavior that has a specific tangible goal, such as the acquisition of property.Predatory crimes, such as theft, burglary, and robbery, are examples of instrumental crime” (Cohn and Rotton, 2003,p. 252). Some authors have argued that also assault and domestic violence–and within it, intimate partner assault–aregoal-oriented because, through the commission of these crimes, offenders seek to gain control over another person orassert their identity (Tedeschi and Felson, 1994). We agree that all crimes can have a certain extent of rationality. Still, th to March 16 th , the secondone including all data points from March 4 th to March 28 th ), the study dynamically comparesthe effects of milder containment policies prompted during the first two weeks of March withstricter measures put in place in the second half of the month. This design provides an informa-tive framework in which to assess the evolution of the effects as a result of policy tightening in theimmediate aftermath. Section 3 presents the statistical outcomes for the selected crime categoriesand the overall number of crimes in the city of Los Angeles. Finally, section 4 and section 5 reviewthe most important results of the analyses, focusing on their theoretical interpretations and themost important policy implications entailed by the study. as argued by Cohn and Rotton (2003), some crimes are more instrumental than others. As an example, “[a]lthoughrobbery is more commonly classified as a violent crime, the violence involved is usually subservient and instrumentalto the goal of taking another person’s property” (Cohn and Rotton, 2003, p. 359). Research Background
Late modern history has been marked by the outbreak of several pandemics. In 1918, the so-called “Spanish Flu” infected about 27% of the world’s population for an almost three-year period,with death estimates ranging from 17 to 50 million globally (Taubenberger and Morens, 2006).In 1957, the “Asian Flu” led to a total of 1.1 million deaths worldwide. In 1968, the A/H3N2influenza shocked the world, causing about 1 million deaths in total. In 2002, the SARS emergedin China. During the period of infection, between November 2002 and July 2003, there were 8,098reported cases of COVID-19 and 774 deaths. In 2009 the “Swine Flu” emerged from the UnitedStates and spread quickly around the world with an estimate of about 61 million cases in theUnited States alone. These public health emergencies modified and influenced many componentsof human society. Researchers investigated these changes from different standpoints. The interestof criminologists in pandemics is, instead, more recent as it mainly emerged in reaction to thespread of COVID-19.Apart from a few studies focusing on the relation between SARS and suicides in Hong Kong(Chan et al., 2006; Cheung et al., 2008), research on crime and deviance is emerging only nowin response to the COVID-19 pandemic. Ashby (2020) used seasonal auto-regressive integratedmoving average models to analyze trends in serious assaults (in public places and residences), bur-glaries (residential and non-residential), vehicle thefts in 16 large American cities. The analysesindicated no statistically significant changes in serious assaults. In some cities, Ashby (2020) ob-served significant reductions in residential burglaries, but only minor changes in non-residentialburglaries. Theft of motor vehicles also decreased in certain cities, while results are mixed whenconsidering thefts from motor vehicles. Mohler et al. (2020) analyzed the counts of calls for ser-vice in Indianapolis and Los Angeles. The authors performed regressions in which they includedan indicator for treatment–i.e., the period after the introduction of shelter in place orders–testingfor differences in means of calls for six crimes in the period January 2 to March 16, 2020, whichacted as baseline scenario. The results indicate a significant decrease in burglaries, robberies, andvandalism in Los Angeles and a significant increase in calls for service for domestic violence inboth cities. Domestic violence is the focus also of the studies by Piquero et al. (2020) and by Leslieand Wilson (2020). Piquero et al. (2020) identified a statistically significant increase in domesticviolence in the first two weeks after the lockdown; yet, they also observed a subsequent decrease.Focusing on a sample of 15 metropolitan areas, Leslie and Wilson (2020) found that social distanc-ing measures were associated with a 10% increase in domestic violence service calls and observedthat the increase might be actually higher due to underreporting.Despite the importance of the topic, and the growing number of available studies, scientistshave not yet answered many questions related to the possible effects of quarantine, social dis-tancing, and self-isolation on crime and deviant behaviors. In particular, on the one hand, the4hort-term effect on crime of social-distancing policies characterized by different intensities ondifferent types of crimes, from more instrumental ones–e.g., burglary–to more expressive and se-rious ones–e.g., homicide–is still to be determined. Few studies presented some results for LosAngeles (i.e., Ashby (2020); Mohler et al. (2020)), but they considered partially different crimes,partially different time frames, used different data (i.e., calls for service), and exploited differentstatistical methods (i.e., SARIMA model, regressions) compared to the analysis we are presentinghere (i.e., nine categories of recorded crimes analyzed using Bayesian structural time-series mod-els). The triangulation and comparison of different results are fundamental, given the novelty ofthe topic. At the same time, this analysis is the occasion to (indirectly) investigate the explanatorypower of theories of crime as routine activity (Cohen and Felson, 1979), crime pattern theories(Brantingham and Brantingham, 1984) and general strain theory (Agnew, 1992) in the aftermathof such a major change in social interactions like the one due to the introduction of COVID-19containment measures.
Routine activity (Cohen and Felson, 1979) and crime pattern theories (Brantingham and Branting-ham, 1984) stress that how the characteristics and interactions of individual-level activities com-mand the spatial and temporal distribution of offending and victimization. On assuming thesenotions, members of a community can be modeled as potential offenders, potential victims, andpotential guardians who move and interact in a socio-geographical space. Starting from thesepremises, routine activity theory postulates that offenders and victims–or targets–usually meetduring everyday non-criminal activities (Brantingham and Brantingham, 1984). Behavioral deci-sions then determine how the various agents react to each other’s presence and actions. Crimeoccurs in the context of the everyday routines as the three factors mentioned above converge inspace and time: a motivated offender, a victim or potential target, and the absence of a capableguardian (Brantingham and Brantingham, 1984).Today, many criminological studies, especially those on ‘volume’ and urban crimes, rely onideas emerging from theories that focus on situations and opportunities as triggers of crime (Co-hen and Felson, 1979; Brantingham and Brantingham, 1984, 1995; Clarke, 1995, 2009; Wortleyet al., 2008). Because of the strong attention that these theories give to ordinary interaction in ge-ographical and social space, we also rely on them to formulate our hypotheses on the short-termimpact of the COVID-19-related social distancing measures. On the other hand, general straintheory postulates that stress generator factors like limited freedom of movement, strict physicaland social isolation, in addition to economic uncertainty and concerns, may push youths, andpeople in general, to be more prone to commit crimes. These factors may introduce new nega-tive stimuli while simultaneously removing positive ones thus generating negative feelings suchas disappointment, depression, fear, and anger (Agnew, 1992).Overall, public measures intended to contain the spread of the virus cause people to spend5ore time at home and lose the density of social interactions. Accordingly, crime opportunitiesand places where crimes occur are likely to change from past observations and experiences. Thesechanges can be particularly significant in the immediate aftermath of the health emergency. Undermild policies–i.e., from March 4 th to March 16 th –we expect a contraction in most urban crimes asthe density of targets reduces in many areas of the city (Angel, 1968). We hypothesize crime reduc-tion will be strong in crime generators areas, by which we mean “particular areas to which largenumbers of people are attracted for reasons unrelated to any particular level of criminal motiva-tion they might have or to any particular crime they might end up committing” (Brantingham andBrantingham, 1995, p. 7). In Los Angeles, typical examples include the Hollywood entertainmentdistrict, the financial district with its high concentration of offices, and the famous Staple Center.Also flows of people to some crime attractor areas will be reduced because people have to avoidconcentrating in bars, nightclubs, or shopping malls, but also in high-intensity drug traffickingand prostitution areas.To various extents, we foresee a reduction in the number of batteries, assaults with deadlyweapons, homicides, robberies, burglaries, shoplifting, thefts, and stolen vehicles as a consequenceof a reduced interaction of people in the urban environment. Based on routine activity and crimepattern theories, we hypothesize that shoplifting diminishes the most. The reduction in the num-ber of open shops and the limitations on the number of entrances–only a certain amount of peopleare allowed to be simultaneously in the shops depending on the premises’ square meters–reducethe opportunities for crime by simultaneously increasing the guardianship and reducing the ex-posure of targets to potential offenders.Routine activity (Cohen and Felson, 1979) and crime pattern theories (Brantingham and Brant-ingham, 1984) suggest that stricter social distancing policies–i.e., the period from March 16 th toMarch 28 th –should have a stronger impact on the crimes considered than mild policies. This isdue to the further reduction in movement and social interaction induced by the reinforcementof social distancing measures. At the same time, the prolonged stay-at-home order may triggerincreases in types of crimes in the medium- and eventually long-term as effects of the increasedstress to which people are exposed. While according to the general strain theory all types of crimemay be influenced by this dynamic (Agnew, 2001), in line with the findings of Schoepfer and Pi-quero (2006), expressive crimes are more likely to be directly influenced by the increased strainthat people experience. Finally, differently from the other crimes considered, intimate partnerassaults are likely to increase. Patriarchy and gender inequality are often considered to be theroot causes of intimate partner violence; yet, situational determinants are also recognized as in-fluencing this form of crime (Wilkinson and Hamerschlag, 2005). As a consequence of the spreadof COVID-19, couples–including dysfunctional ones–spend more time together in their homeswith a reduced presence of possible informal guardians like relatives and acquaintances, two fac-tors that may lead to an increase in violence outbreaks (Hayes, 2018). At the same time, the straincaused by the pandemic makes people more likely to respond with anger to confrontations andto be less concerned about hurting others thus possibly boosting violent crimes (Agnew, 2001;6roidy, 2001), both in the period under mild policies and even more in that under stricter ones. Evaluating the causal link and impact of certain policies is a crucial aspect of research and prac-tice. Criminologists working on different topics have long attempted to assess the extent to whichpublic interventions aimed at reducing crime are actually effective in fulfilling their mission. Thestandard for assessing the causal impact of a certain intervention is represented by RandomizedControlled Trials (RCT) (Rubin, 1974). However, this research design is often unfeasible due to is-sues related to financial costs, ethics or practical obstacles–e.g., complex regulatory requirements.While Pearl (2009) demonstrates that post-facto observational studies cannot provide evidence ofcausal inference due to the potential presence of confounding factors, several quasi-experimentalalternatives have been proposed to overcome the difficulty of running RCT in certain scientificfields. Among these methods are interrupted time series. Interrupted time series have gainedpopularity in sociology and criminology, and they have been applied to several different researchproblems (Biglan et al., 2000; Humphreys et al., 2013; Pridemore et al., 2013, 2014; Humphreyset al., 2017). This method makes it possible to assess the effect of a certain policy by analyzingthe change in the level and slope of the time series after an intervention has been applied, com-pared to the structure of the temporal dynamic before the intervention. More recently, scientistshave developed a framework for evaluating the causal influence of a certain intervention relyingon Bayesian statistics. Following this later evolution, this work investigates the effect of socialdistancing and related measures in the attempt to contain COVID-19 on criminal trends in LosAngeles using Bayesian structural time-series (BSTS) models (Brodersen et al., 2015). Specifically,we apply a method relying on diffusion regression state-space which predicts a counterfactualtrend in a synthetic control that would have occurred in a virtual counterfactual scenario with nointervention–thus, in a scenario where no containment policies are promulgated. This approachallows us to quantify the short-term impact and statistical significance of the containment policieson our variable of interest, namely the number of crimes over time. BSTS are state-space modelsspecifically defined by two equations. The first, i.e. the observation equation, being: y t = Z T t α t + ε t (1)where y t is a scalar observation, Z t is the d -dimensional output vector and ε t ∼ N (0 , σ t ) and ε t isa scalar observation error with noise variance σ t . The observation equation connects the observeddata y t to a latent d -dimensional state vector α t . The second, equation, instead is the state equation ,which reads: α t +1 = T t α t + R t η t (2)7here T t is a d × d transition matrix, R is a d × q control matrix, η t is a q -dimensional systemserror with a q × q state diffusion matrix Q t such that η t ∼ N t (0 , Q t ) . This second equationspecifically governs the dynamic change of the state vector α t through time. The inferential di-mension in the model comprises three components. First, draws of the model parameters θ andthe state vector α (given y n , i.e. the observed data in the training period) are simulated. Second,the model uses posterior simulations to simulate from p (˜ y n +1: m | y n ) , which is the posteriorpredictive distribution, with ˜ y n +1: m as the counterfactual time series and y n as the observedtime series before the intervention. Third, using posterior predictive samples, the model computethe posterior distribution of the point-wise impact y t − ˜ y t for each time unit t .The Bayesian framework in which the model is embedded allows flexibility and inferentialpower, enabling the method to effectively estimate the cumulative difference between the actualdata and a counterfactual scenario. The proposed modeling architecture, through the comparisonbetween a univariate and a multivariate model (which includes two covariates, i.e., daily temper-ature and presence of holidays) and the exploitation of a long timeframe, controls the risk of ex-cluding relevant patterns that may not be specifically related to the pandemic and avoid the risk ofignoring long-term dynamics, a pitfall that would lead to biased estimates. Moreover, the weeklycomponent embedded in the estimation technique preserves the inherent seasonal componentoften exhibited by criminal activity.Concerning the implementation part, we have relied on the CausalImpact package availablein R. Each model has been fit by only considering a single target time-series mapping the trend ofcrime categories in the time window under consideration at a time. The package does not allowto simultaneously model multiple target variables and, having included a crime series as the targetone and all the others as covariates would have contradicted the requirement to only consider co-variates that are not influenced by the interventions. Each model included a seasonal component:characterizing weekly seasonality (number of seasons set equal to 7 with season duration equal to1, i.e., one day). This allowed to account for the well-known seasonal oscillations of crime overthe days of a week (with temperature as a covariate being able to control higher-level seasonalcomponents). , We performed our analyses by including weekly seasonal terms to account forweekly variations in crime trends. By doing so, it was possible to consider also possible seasonalvariations that do not directly depend on temperature-related mechanisms, but instead relate toroutine activities of people and places during social occurrences (e.g. school closure) (Andresenand Malleson, 2013). Furthermore, to obtain accurate estimates for each model a total of 1,000Markov Chain Monte Carlo (MCMC) samples are drawn. Finally, the prior of each model, ex-pressed in terms standard deviation of the Gaussian random walk at the local level, has been keptequal to 0.1 as suggested by the authors of the package in absence of ground truth. The value repre-sents a good compromise between a high standard deviation that would assume that the variationsin the signal are all explained by the intervention and a very small standard deviation that imposes8hat such variations are instead solely due to high noise in the data. Exploiting this analytical strategy, we separately analyze the period up to March 16 th to test the ef-fect of milder containment policies (the first post-intervention time window goes from March 4 th to March 16 th ) and up to March 28 th to test the effect of the introduction of stricter containmentpolicies (the second post-intervention time window goes from March 4 th to March 28 th ). Thesetwo post-intervention time-windows enable us to obtain a more comprehensive and inherentlydynamic description of the reality. In fact, given the progressive tightening of imposed restric-tions, avoiding a comparative analysis of the evolution of the effects over different weeks wouldsomewhat oversimplify and overly aggregate our statistical estimates. Conversely, the possibilityto observe the evolution of the effects helps in assessing the strength of a trend, anticipating thelikely developments and better explaining its behavior by framing it in our theories of reference.Figure 1 presents the data on people’s mobility (driving, transit and walking) in the city of Los An-geles from January 1 st th when stricter policies were implemented. Our interest isto understand changes in the crime patterns in the immediate aftermath of the health emergencyto assess and analyze shot-term dynamics. Indeed, in the medium- and long-run other factorsmight play a more central role in explaining the changes in crime trends (e.g., unemployment,social riots).
To conduct our analysis, we first drew upon the (Los Angeles Open Data portal). Unlike priorstudies that analyzed the counts of calls for service in Los Angeles (e.g., Mohler et al. (2020)), weused data on crime reported in the city. Exploiting the website API, we accessed two differentdatasets on crimes reported by the Los Angeles Police Department (LAPD). The first one com-prised all crime incidents reported from 2017 to 2019. The second one referred to all the crimesreported from January 1 st to March 30th 2020. The two datasets contain detailed daily observa-tions regarding each reported crime with information on the type of offense–organized in 140 The R code used to perform the analyses here presented is publicly available in a dedicated repository at Link. The two post-intervention time windows considered are not mutually exclusive. The first one includes the daysbetween the 4 th of March to the 16 th of March and the second one considers the days between the 4 th of March to the28 th of March. The second time window includes the first one. Our interest is to map the evolution of the policies andtheir influence on crime trends given that policy effects in this context are cumulative rather than mutually exclusive.The focus of our analysis is not to compare two different time period of post intervention in order to understand, forexample, if mild policies are better than harsh policies, but to assess the effect of those policies on crime over time. The mobility trends in Los Angeles are also confirmed by other studies that used other data to analyze trends incommuting and general mobility (e.g., data provided by Cuebiq) (Klein et al., 2020; Ruiz-Euler et al., 2020). st ataset First Day Considered N of Observations Table 1: Number of Observations and Starting Point per DatasetFor the purpose of the present work, we only relied on the reported date of crime occurrenceand on the offense categories as meaningful sources of information to analyze city-wide criminaltrends. A brief description of the three datasets (1. the one comprising crimes from 2017 to 2019;2. the one with offenses up to March 2020; 3. the merged one used in the models) is providedin Table 1. We extracted the crime categories of interest and we then grouped observations bydaily counts, obtaining separated time series for each crime category. While our models focus onthe post-intervention period from March 4 th to March 28 th , we lastly accessed data on April 7th,thus ensuring that a larger number of crime reports that were not possibly included in the datasetin their first week after occurrence were actually imputed by the LAPD soon thereafter. Figure2 displays the number of observations per crime category in the period from January 1 st th ss au l t w i t h D . W . B a tt e r y B u r g l a r y I n t i m . A ss l t. R obbe r y S hop li ft. T he ft H o m i c i de s S t o l en V eh i c l e N ( / / − / / ) Figure 2: Number of Observations per Crime Categoryidays were not changed because of the pandemic. The method adopted generally recommendsusing more than two controls to evaluate the effect of an intervention on the response time series.However, we can assume that no other daily predictor of crime has been strongly influenced bythe containment policies during the period analyzed. Differently from other studies on the topic,the use of time series spanning 39 months was made to reduce the potential biases arising fromthe exclusion of hidden trend dynamics, preserving the seasonality and long-term dependenciesof crime. Table 2 reports the main descriptive statistics for the time series that are part of theanalysis.
This section presents the results on the causal impact of the policy interventions starting on March4 th per each offense type and for the overall number of crimes. For each crime, the same analyticalstructure is provided. We ran two different models. The first one was a univariate model that onlyconsiders the time series of interest without controls. The second model integrates two covariatesto control for spurious effects and unobservable dynamics. At the same time, the two models wereperformed on two time-windows; the first capturing the impact of mild policies (from March 4 th toMarch 16 th ), the second including also the first weeks in which stricter policies entered into force(from March 4 th to March 28 th ). Figure 3 shows the time series of the crimes considered; while,12 ariable Min 1 st Q Median Mean St. Dev. 3rd Q Max
All Crimes 254.00 570.00 628.00 619.00 69.80 669.00 796.00Assault (with Deadly Weapon) 6.00 24.00 29.00 29.44 7.53 34.00 61.00Battery 15.00 46.00 52.00 52.15 9.51 58.00 93.00Burglary 13.00 31.00 38.00 38.70 9.96 45.00 93.00Homicide 0.00 0.00 0.00 0.72 0.00 1.00 5.00Intimate Partner Assault 10.00 36.00 41.00 41.60 8.84 47.00 78.00Robbery 7.00 20.00 24.00 24.23 5.85 28.00 48.00Shoplifting 2.00 15.00 18.00 18.00 4.92 21.00 33.00Theft 19.00 52.00 62.50 61.72 13.3 61.72 71.00Stolen Vehicle 19.00 40.00 46.50 46.74 9.39 52.00 88.00Holiday 0.00 0.00 0.00 0.02 0.16 0.00 1.00Max Temperature 52.00 69.00 76.00 75.84 9.20 82.00 108.00
Table 2: Descriptive Statistics of the Considered Time Seriesto facilitate the reading of the results and summarize the statistical outcomes of the models, Table3 presents all the statistical results (full statistical outcomes are available in the SupplementaryMaterials, from Table A1 to A10). A ss au l t w / D W B a tt e r y B u r g l a r y I n t i m a t e A g . R obbe r y S hop li ft i ng 2060100 T he ft S t o l en V eh i c l e 012345 H o m i c i de O v e r a ll Figure 3: Time Series of Considered Crimes13 rime Type First post interventiontime window(March 4 th – March 16 th ) Second post interventiontime window(March 4 th – March 28 th ) Univariate With Covariates Univariate With Covariates
Assaults D.W. -2.98%[-19%, 13%] -1.5%[-18%, 13%] -11%**[-23%, 2.8%] -6.3% (6%)[-18%, 5.5%]Battery (Simple Assault) -0.6%[-12%, 11%] 0.78%[-9.2%, 11%] -11%**[-21%, -0.99%] -7.6%**[-16%, 0.39%]Burglary 0.89%[-14%, 15%] -0.58%[-14%, 11%] -4.8%[-15%, 5.5%] -7.3%*[-17%, 3.3%]Intimate Partner Assault -4%[-16%, 6.4%] -2.5%[-13%, 8.6%] -0.28%[-11%, -11%] 3.3%[-5.6%, -12%]Robbery -24%***[-38%, -8.5%] -23%***[-38%, -8.7%] -21%***[-33%, -9.3%] -19%***[-30%, -8.7%]Shoplifting -14%***[-30%, 2.4%] -15%***[-30%, 0.34%] -31%***[-42%, -20%] -32%****[-43%, -21%]Theft -9.1%**[-19%, 0.57%] -9.6%**[-19%, -1%] -24%***[-31%, -17%] -25%***[-31%, -18%]Stolen Vehicles 1%[-9.4%, 11%] 0.06%[-10%, 9.9%] 1.5%[-6.5%, 9.6%] -0.12%[-7.4%, 7.5%]Homicides -15%[-88%, 57%] -10%[-84%, 59%] -28%[-79%, 25%] -24%[-76%, 31%]Overall Crimes -5.6%***[-10%, -1.5%] -5.4%**[-9.5%, -1%] -15%***[-18%, -11%] -14%***[-17%, -11%]
Table 3: Model Results - Relative Cumulative Effect per Each Crime (95% C.I. Between Parenthe-ses)
Assault with Deadly Weapons
Concerning assaults with deadly weapons, both the univariate and the multivariate models, forboth time windows, i.e. up to March 16 th and up to March 28 th , do not show any statistically sig-nificant effect of the containment policies. Although all the models report negative coefficients,indicating a reduction in the trend in absolute and relative terms, the posterior probability of acausal effect is, respectively, 64% and 52% for the first time window and 94% and 85% for the secondtime window. The results provided by the models referring to the second time window–especiallyfor the univariate one–are much closer to a statistically significant outcome, suggesting a differ-ent effect when stringent policies are introduced. Consequently, it is possible that this negativeeffect will show statistically significant results in the case of prolonged or even more stringentcontainment measures.Table A1 reports the outcomes of the analysis. Assault with Deadly Weapons corresponds to crimes labelled as Assault with Deadly Weapon, Aggravated Assaultby the LAPD. According to the California Penal Code 245(a)(1), an assault with a deadly weapon occurs when anindividual wrongfully attacks a victim with an object that can seriously injure or even inflict death. attery and Simple Assault The occurrence of batteries is affected by strict policies, while it is not so by loose ones. Upto March 16 th , containment policies do not seem to have resulted in a reduction of battery andsimple assaults. Conversely, if we consider the introduction of more stringent measures of socialdistancing, it is possible to observe a statistically significant reduction of this crime. The modelsconsidering crimes recorded until March 28 th show a relative effect of -11.0% in the univariatemodel and of -7.6% in the multivariate one. The posterior probability of a causal effect is, respec-tively, 98% and 96% indicating statistically significant outcomes for both models (full results areavailable in Table A2). Burglaries
For what concerns burglaries, the policies to contain the spread of COVID-19 have not producedany significant effect in the first four weeks from their introduction. Although we expected aslight decrease in their overall occurrence, given the extensive agreement over the fact that bur-glars prefer to target unoccupied homes (Shover, 1991; Mustaine, 1997; Tseloni et al., 2002) andthe consequent increased guardianship enforced by people staying at home due to the pandemic,statistical outcomes do not corroborate our hypothesis. In the first time window, the effects areminimal and the results of the two models diverge in their directions. The univariate model showsan increase of 0.89% with respect to the predicted value in the absence of an intervention, whilethe multivariate displays an effect equal to -0.58%, with posterior probabilities respectively being56% and 50%. The models performed up to March 28 th show a non-significant reduction of bur-glaries in the city with a higher posterior probability for both the univariate and the multivariatemodels–respectively, 80% and 91%–compared to the other time-frame analyzed. Full results canbe found in Table A3. Although
Battery and Simple Assault are coded as a single type of offense one in the original database under thecrime category BATTERY - SIMPLE ASSAULT, California assault law, disciplined by Penal Code 240PC, provides twodistinct definitions for battery and assault. A simple assault is the attempt to use force or violence against someoneelse, while battery is the actual use of force or violence against one or more individuals. In California, according to the Penal Code, burglary is defined as the act of entering any structure, room, orlocked vehicle with the intention to commit a theft or a felony. Furthermore, an individual can be considered guilty ofburglary even if the intended crime, once entered, is never been committed. Our initial hypothesis is that containmentpolicies should have a clear impact on this crime category as people encouraged or forced to stay in their housesincrease guardianship, thus reducing the chance for an individual to enter a property without being noticed. ntimate Partner Assault COVID-19-containment policies, we hypothesize, could cause intimate partner assaults to in-crease as a consequence of individuals spending longer time at home in a potentially stressfulsituation. Instead, results suggest that the policies adopted have not prompted any immediatesignificant change in intimate partner assaults. The models considering the days from March 4 th to March 16 th as period of intervention show non-significant negative effects (-4.0% and -2.5%).The univariate model considering the entire period identifies a small non-significant negative ef-fect (-0.28%). Finally, the multivariate model indicates an increase in intimate assaults due to thepolicies (+3.3%); but this increase is not statistically significant. The explanation for the absenceof a clear and significant signal in the post-intervention period may be connected to the fact thatdynamics of this crime are complex and the forced cohabitation of partners is not an immediatetrigger of violence within the household. While in the case of shoplifting, for example, the clo-sure of shops has a direct impact on the thefts within the shops themselves, in the case of intimatepartner violence the increase may be delayed as tensions intensify. Robbery
Robbery shows a significant change in the post-intervention period. This applies to both theunivariate and the multivariate cases for both the temporal windows selected. In the univariatecase, the relative effect is estimated as a reduction of 24% in robberies, with a cumulative totalof 202 criminal events against a predicted 266 in a non-intervention scenario in the first timewindow. In the second time window, the estimated effect is a reduction of 21%, with a cumulativetotal of 439 robberies against a predicted 533 in a non-intervention scenario (See Table A5 in theSupplementary Materials for the Absolute Effect). In the multivariate case, the effects are slightlyreduced in both the time frames; being respectively, -23% and -19%. The magnitude of the effectsis thus high. The posterior probability is 99.7% and 99.8%, respectively, for the univariate andmultivariate models up to March 16 th and 99.7% and 99.8% for the models up to March 28 th . To analyze intimate partner assaults, we have combined both simple and aggravated assaults–Intimate Partner -Aggravated Assault” and “Intimate Partner - Simple Assault— in the original database by LAPD. These two offensesfall within the broader set of domestic violence crimes, which includes other forms of within-family violence–e.g.,parents being violent against their children. The California Penal Code defines an intimate partner as a current orformer spouse, a fiancé, a co-parent of a child, a person with whom the perpetrator had a dating relationship or aperson with whom the perpetrator lives. Section 211 of the California Penal Code defines robbery as the act of taking personal property from someoneelse, against the targeted victim’s will, using force or fear. Robberies are classified as felonies. Data on robberiescorrespond to
Robbery in the original database compiled by the LAPD. hoplifting The statistical outcomes of the models referring to shoplifting indicate a significant reductionafter the introduction of the state of the emergency in Los Angeles. The results hold for all models.In the first time window analyzed, 189 shoplifting cases were registered, significantly fewer thanthe cumulative numbers of occurrences predicted by the virtual scenarios with no intervention:220 for the univariate and 223 in the multivariate cases (Table A6 in the Supplementary Materials).In the models up to March 28 th , against cumulative predicted values of 462 and 471 in a scenariowithout intervention, 320 shoplifts were recorded by the LAPD. This indicates an estimated rel-ative reduction of 14% in the univariate case, and a 15% reduction in the multivariate one in theperiod with mild policies. With the introduction of stricter policies, the relative reduction is 31%in the univariate case, and 32% in the multivariate one. The probability of the effects being a causalconsequence of the policies is 95.6% and 97.2% for the first time window and 99.7% and 99.9% forthe second time window, providing strong statistical evidence. Theft
In line with what we found for robbery and shoplifting, thefts also appear to be affected by theapplication of the COVID-19-related containment measures. All models show a statistically sig-nificant reduction of thefts. In the first analyzed window, the models report a 9.1% decrease in theunivariate case (significant at the 96.7% level) and a 9.6% reduction in the multivariate one (signifi-cant at the 98.3%). In the second window considered (up to March 28 th ), the models estimate a 21%decrease in the univariate case (significant at the 99.7%) and a 25% reduction in the multivariatecase (significant at the 99.9%). According to the California Penal Code 459.5, shoplifting is the offense of entering a commercial establishmentduring business regular hours, with the intent of committing a theft crime worth $950 or less, regardless of the actualcompletion of the theft. This crime is identified as
Shoplifting-Petty Theft ($950 & Under) in the LAPD’s database. Inthe present work, we have also considered shoplifting grand-thefts, related to attempted thefts of property worthmore than 950$–i.e.,
Shoplifting-Grand Theft ($950.01 & Over) in the original dataset. For the purpose of this study, we have combined in the general category “Theft” both petty thefts–i.e.,
TheftPlain-Petty ($950 & Under) –and grand thefts–i.e.,
Theft-Grand ($950.01 & Over) Excpt, Guns, Fowl, Livestk, Prod . TheCalifornia Penal Code defines petty theft as the act of stealing–or wrongfully taking–an object belonging to someoneelse when the value of the property is equal to $950 or less. Grand theft, instead, is a more serious offense and pertainsto acts in which the property has a value higher than 950$. tolen Vehicles The policies have not had any significant effect on vehicle thefts. Indeed, the relative effects arenot only not significant, but also small for all models: 1.0% and 1.5% in the univariate cases, 0.06%and -0.12% in the multivariate ones. There are different possible (complementary) explanationsfor this finding. On the one hand, while at home, a car owner may be an ineffective guardian ofher/his own car. On the other hand, because the theft of cars is often related to their immediateuse (Cherbonneau and Wright, 2011), the slowdown of productive and social activities–includingother crimes–may mean that offenders have less need to steal a vehicle. The combination of thesedynamics, which are pushing in opposite directions, might explain the absence of a clear impactof social distancing policies on stolen vehicles in Los Angeles.
Homicides
Overall, COVID-19 containment policies do not show any statistically significant effect on homi-cides in the short aftermath of their deployment. In neither of the two selected temporal win-dows do the models detect a sufficient statistically strong variation in the trend. Nonetheless,compared to the first batch of data–up to March 16 th –the models also considering the strict poli-cies point in a two-fold direction. First, the relative effect has increased, ranging from -15% to -28%according to the univariate model and from -10% to -24% according to the multivariate specifi-cation. Second, the posterior tail-area probability p has decreased, thus increasing the posteriorprobability of a causal effect, which ranged from 67% and 63% in the first batch of results, to 86%and 80%. Overall Crimes
Finally, we considered all reported crimes in the two selected temporal windows under analysis;this aggregated variable also accounts for all those offenses that go beyond the categories previ-ously considered. Figure 4 displays the evolution of the post intervention relative marginal effectin the two considered time windows, for all crimes and model types. In particular, the graph rep-resents the percent changes in crime between our synthetic control, which we set at zero, and theactual registered crimes. Our empirical results indicate a significant decrease in the overall crime Stolen vehicles refers to two different offenses, namely grand theft auto and the unlawful taking or driving of avehicle. The main difference between the two offenses pertains to the duration of the crime itself. If, for instance, aperson steals a car with the intent to keep it, this is often considered grand theft auto. Conversely, if the offender aimsat using the car for a ride–or, in any case, for a short timeframe–the act is usually considered as the unlawful takingor driving of a vehicle. The original database compiled by the LAPD provides a single crime category:
Vehicle - Stolen . The category
Homicides comprises those crimes which are disciplined by California’s Homicide Laws and cor-responds to the Criminal Homicide crime category in the original database. A person committing a homicide canbe prosecuted in several ways depending on the characteristics of the action. Among these are first-degree murder,second-degree murder, capital murder, voluntary manslaughter, involuntary manslaughter and vehicular manslaugh-ter. Notably, homicides are the most serious offenses among those considered in this study–and the least common.
Assaultw. DW Battery Burglary IntimateP. Agg. Robbery Shoplifting Theft StolenVeh. Homicide Overall U n i v a r i a t e M u l t i v a r i a t e Model Type P o s t − I n t e r v en t i on R e l a t i v e M a r g i na l E ff e c t ( % ) Significance <95% >95%
Figure 4: Graphical Summary of Model Results (with 95% Confidence Intervals)When focusing on all reported crimes, the results are in line with those seen for robbery,shoplifting, theft, and battery. Nonetheless, the cumulated count of these crimes cannot explainon its own the overall decrease given the numerosity of other crimes. As such, these findings sug-gest that, potentially, most crimes have diminished in the considered short-term period leadingto a general reduction of crime in the entire city of Los Angeles. In turn, this result indicates that,besides the categories chosen for this study, further work is needed to disentangle mechanismsrelated to single offenses or crime categories. 19
Discussion & Conclusion
In line with our hypotheses, the statistical results show that robberies, thefts, and shoplifting hada statistically significant reduction already in the post-intervention period up to March 16 th , whenonly mild policies were applied. Robbery recorded the single largest decrease (-23%/-24% depend-ing on the statistical specifications). Mohler et al. (2020), applying a different statistical strategyon different data and concentrating on a partially different time frame, also observed a signif-icant reduction in robberies in Los Angeles in the immediate aftermath of the introduction ofsocial-distancing measures. After robberies, shoplifting (-14%/-15%) and thefts (-9.1%/-9.6%) arethe crime categories for which the differences between the observed data and the virtual scenarioare the strongest. Overall crimes also significantly declined (-5.4%/-5.6%). Theft, robbery andshoplifting are offenses involving a direct act targeting a property, thus making it likely that com-mon underlying principles govern the similar decreasing trends of these crimes. These reductionscan be explained in light of the reduction of social interactions as people avoid public spaces andspend more time at home as well as by closure of public places and the presence of quotas on thenumber of entries into shops and malls. In turn, this leads to a reduction in criminal opportu-nities and to an increase in guardianship. Coherently with the ideas emerging from routine ac-tivity theory (Cohen and Felson, 1979) and crime pattern theory (Brantingham and Brantingham,1984), most crime reductions are larger after the introduction of stricter containment policies,which further reduce criminal opportunities while increasing informal guardianship. The mod-els considering the entire period up to March 28 th show a significant reduction for shoplifting(-31%/-32%), thefts (-24%/-25%), robberies (-21%/-19%), and for overall crimes (-15%/-14%). Inaddition, battery (-11%/7.6%) also started to show a significant decrease after the adoption of strictdistancing policies.Contrarily, the models did not detect any significant change in the trends of homicides, as-saults with a deadly weapon, intimate partner assaults, but also of burglaries and stolen vehicles.Previous studies have listed several stressors specifically emerging during quarantines comprisingthe duration of the quarantine itself, the fear of infection, frustration and boredom, and inade-quate supply of basic commodities, services, and information (Brooks et al., 2020). In addition,families under financial and psychological stress as a result of the pandemic, increased their al-cohol use at home (Colbert et al., 2020). This might also increase the occurrence of violent andexpressive crimes since there is a strong relationship between the use and abuse of alcohol andviolence (Parker and Auerhahn, 1998). In line with these evidence on the negative psychologicaloutcomes of quarantine, our statistical results suggest that besides reduction of opportunities andchange in social life, increased strain is likely to have a leading role in explaining crime dynamicsimmediately after the introduction of social distancing measures. In this perspective, the lack ofsignificant variations in the trends of the aforementioned crimes might be interpreted as the resultof a combined effect of a partial reduction in situational and opportunistic triggers of crime, onthe one hand, and worsening of the balance between positive and negative psychological stimuli20n the other.In light of this, intimate partner assault warrants some further considerations. Firstly, intimatepartner assault is a crime which is strongly related to strain dynamics (Eriksson and Mazerolle,2013; Piquero et al., 2020). Secondly, crime factors pertaining to routine activity theory and crimepattern theory and factors pertaining to general strain theory all suggest an increase in intimatepartner assaults. It is not so for other considered crimes, with the partial exception of vehicletheft. Nonetheless, we do not observe any statistically significant increase in the count of intimatepartner assaults. By contrast, Mohler et al. (2020), in their study, saw significant increases in do-mestic violence calls for service in Los Angeles. In this regard, to be noted is that many of suchcalls concern domestic disturbances without a physical assault (MacDonald et al., 2003; Mohleret al., 2020).More in general, our analysis suggests that containment policies in the immediate aftermathhave had a stronger impact on more instrumental crimes compared to crimes for which expres-sive motivations are more relevant. Similarly, more serious offenses appear to be less influenced bysocial-distancing measures. The combination of motivation and seriousness identifies two macroclasses of crimes, which may help the interpretation of the effects of containment policies. Moreinstrumental/less serious crimes are strongly influenced by the reduction of social interactionsand the sudden modification of everyday habits. While routine activity theory and crime patterntheory seem to be a powerful tool to interpret changes in this first class of criminal behaviors, theyare less effective in capturing the dynamics of more expressive/more serious crimes whose inter-pretation needs to be integrated with a broader set of factors (Hayward, 2007). Strong motivation,low self-control and criminal opportunities are concurrent causes in the explanation of crimecommitment in the aftermath of the policy introduction, but the combination of these causes isnot equal for all crime types (Longshore and Turner, 1998). It appears that strong motivationand low self-control are more likely to play a central role in more serious crimes, whereas crim-inal opportunities do so in triggering less serious offenses. In this regard, the strain theory ofcrime (Agnew, 1992) may act as complementary interpretative framework to the rational choiceapproach, explaining the emerging trends also in light of the sudden amplification of stressorsresulting from COVID-19 containment measures. Future studies may investigate these dynamicmore in depth by exploiting a panel approach and by identifying variables capable of representingthe different theories we exploited in our interpretative framework.Finally, burglaries and vehicle thefts appear to fall outside this reasoning and possibly nullifyit. Yet, as previously said, these are crime categories which unify different actual criminal dynam-ics. Burglaries can be residential or non-residential; separating the two may shed light on distinctopposite processes, as residential burglaries are expected to decrease, while non-residential onesto increase as a byproduct of different dynamics in guardianship. In support of this, Ashby (2020),who could separate residential and non-residential burglaries in Los Angeles, observes–at leastduring some weeks–significant changes with respect to both residential and non-residential bur-glaries. It is thus important to mention again that our results capture short-term dynamics within21n open time-frame that will likely extend over several weeks ahead, thus potentially leading tonew patterns. Furthermore, we may have witnessed, especially in relation to the period up toMarch 16 th , a mixed dynamic combining both the effect of the mild policies and the fear of conta-gion. Fear of contagion has probably played a role in reshaping people’s individual and collectivebehavior (Brooks et al., 2020), even before the second half of March when stronger measures wereadopted. This process, directly impacting on the density and frequency of social contacts, haspotentially fostered the effect of the policy recommendations on crime.In addition, social distance restrictions may have also influenced the reporting rate of criminaloffenses. People avoid spending time outside their homes, and this may reduce their willingnessor ability to go to the police to report a crime. In the case of Los Angeles, problems related toa reduction of crime reporting are mitigated by the possibility to report a crime using an onlineform for specific crimes (e.g., theft or theft from vehicles) or calling a dedicated number. More-over, while underreporting issues may affect the count of all the crimes considered, the crimes forwhich we do not observe any significant change–with the important exception of intimate part-ner violence–are the ones that are more likely to be reported–i.e., homicide, assaults with deadlyweapons, stolen vehicle, and burglary. These factors help to reduce possible biases due to increasesin underreporting.By contrast, the dynamics related to the lack of–or impossibility to–reporting may especiallyinfluence the count of intimate partner assaults. Indeed, the cohabitation of victim and offendermay make it difficult for the victim to report the offense to the police. Nevertheless, since neigh-bors are likewise at home, they are also better able to exert partial guardianship on the episodes ofintimate partner violence that are happening in their surroundings and call the police. In this re-gard, studies have demonstrated that neighbors can play a role in reporting abuses (Paquin, 1994).In addition, scholars have shown that informal actors (e.g., family, friends, and neighbors) are of-ten approached by the victims of intimate partner violence because they are proximal and maybe able to intervene before, during, and after the violent event (McCart et al., 2010; Wee et al.,2016). Related to this, the police attention to urban crimes may have diminished due to the needto enforce social distancing measures, thus partially countering the crime mitigating effect drivenby looser social interactions. However, the reduction of police capabilities is likely to have only aonly marginal effect on intimate partner assaults, which are serious crimes that often takes placein a private environment.Additional caution should be adopted in interpreting the results obtained. As the extant crim-inological literature indicates, crime clusters in time and space, meaning that its spatio-temporaldistribution is not random (Doran and Lees, 2005; Grubesic and Mack, 2008; Mohler et al., 2011;Weisburd, 2015). Therefore, there are very few reasons to think that this patterned nature willentirely change because of containment measures. Future work will hence need to investigatethe potential heterogeneity of policy effects across different areas of the same city, also analyzingthe potential correlates of diverging trends in terms environmental and socio-economic factors.Furthermore, according to anecdotal evidence, police were asked to limit or stop making low-22evel arrests to manage jail crowding and narrow down the spread of the virus in prisons. Thesechanges in police practice could have influenced crime trends and we could expect a downwardbias for the estimates of less serious crimes (e.g., shoplifting). The same downward bias cannotbe foreseen for homicide or intimate partner assaults that are not low-level crimes, but seriousoffences instead. An opposite trend that could have affected crime rates in the opposite directionin the period under analysis is the early release of prisoners in order to contrast the spread ofcoronavirus within prisons. The early release of prisoners could potentially lead to an increase incrime rates due to recidivist behaviors of former inmates. The two mechanisms described herego in opposite directions (decrease vs increase in crime trends), therefore they potentially tend tooverall balance themselves. This work has outlined how COVID-19 containment policies have influenced criminal trendsin the city of Los Angeles already in the first weeks after their introduction. Policy implicationsemerge from these analyses. First, studying criminal dynamics in this anomalous time should helpin creating protective measures for the most vulnerable subjects influenced by these changes, in-cluding homeless people, women and children. Homeless people, for instance, could experience anincrease in the likelihood of becoming victims of crimes, as a consequence of lower guardianshipand social control. Furthermore, as people are likely to suffer more stress in this period thus beingmore prone to commit certain crimes, including violent ones, it might be worth empowering re-mote contact points, facilitating distance reporting to make it easier for victims, who cannot freelymove, to connect with the police. Distance reporting, for instance, could include text messagingservices to reduce the risk of being stopped or heard by the abusive member of a household.Finally, as crime takes new forms and dynamics, law enforcement agencies will be requiredto modify or re-define resource allocation for the new priorities. The task, especially in highlypopulated and heterogeneous cities like Los Angeles, can be herculean. In fact, while many po-lice departments and other institutions in the United States have tailored their actions based onpredictive policing software, these models can suddenly become of little help. Depending on theextent to which these policies will force crime to change–beyond mere temporal trends–predictivemodels built on millions of past observations may no longer be informative. This situation thusurges alternative predictive tools that can take into account disruptions of social life as the trig-gers of new criminal risks, prompting data-driven strategies to re-assess criminal patterns andcountering strategies.
Conflict of Interest
The authors declare no conflict of interest. 23 unding
The authors did not receive any funding for the present work.
Acknowledgments
The authors wish to thank Jay Aronson, Laura Dugan, Marco Dugato, Gary LaFree, Clarissa Man-ning, Cecilia Meneghini, Riccardo Milani, James Prieger and two anonymous referees for theirprecious feedback and comments on previous versions of this manuscript.24 eferences
County of Los Angeles, Public Health Issues Order to Prohibit Group Eventsand Gatherings, 2020. URL: https://lacounty.gov/covid19-news/public-health-issues-order-to-prohibit-group-events/ , library Catalog: la-county.gov Section: COVID-19 Newsroom.E. G. Cohn, J. Rotton, Even criminals take a holiday: Instrumental and expressivecrimes on major and minor holidays, Journal of Criminal Justice 31 (2003) 351–360.URL: .doi: .J. T. Tedeschi, R. B. Felson, Violence, aggression, and coercive actions., American Psycholog-ical Association, Washington, 1994. URL: http://content.apa.org/books/10160-000 .doi: .J. K. Taubenberger, D. M. Morens, 1918 Influenza: the Mother of All Pandemics -Volume 12, Number 1—January 2006 - Emerging Infectious Diseases journal - CDC(2006). URL: . doi: .S. M. S. Chan, F. K. H. Chiu, C. W. L. Lam, P. Y. V. Leung, Y. Conwell, Elderly suicide and the 2003SARS epidemic in Hong Kong, International Journal of Geriatric Psychiatry 21 (2006) 113–118.URL: http://doi.wiley.com/10.1002/gps.1432 . doi: .Y. T. Cheung, P. H. Chau, P. S. F. Yip, A revisit on older adults suicides and Severe Acute RespiratorySyndrome (SARS) epidemic in Hong Kong, International Journal of Geriatric Psychiatry 23(2008) 1231–1238. URL: http://doi.wiley.com/10.1002/gps.2056 . doi: .M. P. J. Ashby, Initial evidence on the relationship between the coronavirus pandemic andcrime in the United States, Crime Science 9 (2020) 6. URL: https://doi.org/10.1186/s40163-020-00117-6 . doi: .G. O. Mohler, A. L. Bertozzi, J. Carter, M. B. Short, D. Sledge, G. E. Tita, C. D.Uchida, P. J. Brantingham, Impact of social distancing during COVID-19 pandemic oncrime in Los Angeles and Indianapolis, Journal of Criminal Justice 68 (2020) 101692.URL: .doi: .A. R. Piquero, J. R. Riddell, S. A. Bishopp, C. Narvey, J. A. Reid, N. L. Piquero, Staying Home,Staying Safe? A Short-Term Analysis of COVID-19 on Dallas Domestic Violence, Amer-25can Journal of Criminal Justice 45 (2020) 601–635. URL: https://doi.org/10.1007/s12103-020-09531-7 . doi: .E. Leslie, R. Wilson, Sheltering in Place and Domestic Violence: Evidence from Calls for Ser-vice during COVID-19, SSRN Scholarly Paper ID 3600646, Social Science Research Net-work, Rochester, NY, 2020. URL: https://papers.ssrn.com/abstract=3600646 . doi: .L. E. Cohen, M. Felson, Social Change and Crime Rate Trends: A Routine Activity Approach,American Sociological Review 44 (1979) 588–608. URL: . doi: , publisher: [American Sociological Association, Sage Publi-cations, Inc.].P. J. Brantingham, P. L. Brantingham, Patterns in crime, Macmillan, New York, 1984.R. Agnew, Foundation for a General Strain Theory of Crime and Delinquency*, Crimi-nology 30 (1992) 47–88. URL: https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1745-9125.1992.tb01093.x . doi: , _eprint:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1745-9125.1992.tb01093.x.P. L. Brantingham, P. J. Brantingham, Criminality of place, European Journal on Criminal Policyand Research 3 (1995) 5–26. URL: https://doi.org/10.1007/BF02242925 . doi: .R. V. Clarke, Situational Crime Prevention, Crime and Justice 19 (1995) 91–150. URL: . doi: , pub-lisher: The University of Chicago Press.R. V. Clarke, Situational Crime Prevention: Theoretical Background and Current Practice, in:M. D. Krohn, A. J. Lizotte, G. P. Hall (Eds.), Handbook on Crime and Deviance, SpringerNew York, New York, NY, 2009, pp. 259–276. URL: http://link.springer.com/10.1007/978-1-4419-0245-0_14 . doi: , series Title: Handbooksof Sociology and Social Research.R. K. Wortley, L. G. Mazerolle, S. Rombouts, Environmental Criminology and Crime Analysis,Routledge, 2008. Google-Books-ID: 8U8ObTP81QsC.R. Agnew, Building on the Foundation of General Strain Theory: Specifying the Types of StrainMost Likely to Lead to Crime and Delinquency:, Journal of Research in Crime and Delinquency(2001). URL: https://journals.sagepub.com/doi/10.1177/0022427801038004001 .doi: , publisher: Sage Publications.26. Schoepfer, A. R. Piquero, Self-Control, Moral Beliefs, and Criminal Activity, Deviant Behav-ior 27 (2006) 51–71. URL: https://doi.org/10.1080/016396290968326 . doi: , publisher: Routledge _eprint: https://doi.org/10.1080/016396290968326.D. L. Wilkinson, S. J. Hamerschlag, Situational determinants in intimate partner violence, Ag-gression and Violent Behavior 10 (2005) 333–361. doi: , place:Netherlands Publisher: Elsevier Science.L. M. Broidy, A Test of General Strain Theory*, Criminology 39 (2001) 9–36.URL: https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1745-9125.2001.tb00915.x . doi: , _eprint:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1745-9125.2001.tb00915.x.D. B. Rubin, Estimating causal effects of treatments in randomized and nonrandomized stud-ies., Journal of Educational Psychology 66 (1974) 688–701. URL: http://content.apa.org/journals/edu/66/5/688 . doi: .J. Pearl, Causal inference in statistics: An overview, Statistics Surveys 3 (2009) 96–146. URL: https://projecteuclid.org/euclid.ssu/1255440554 . doi: , pub-lisher: The author, under a Creative Commons Attribution License.A. Biglan, D. Ary, A. C. Wagenaar, The Value of Interrupted Time-Series Experiments for Com-munity Intervention Research, Prevention Science 1 (2000) 31–49. URL: https://doi.org/10.1023/A:1010024016308 . doi: .D. K. Humphreys, M. P. Eisner, D. J. Wiebe, Evaluating the Impact of Flexible AlcoholTrading Hours on Violence: An Interrupted Time Series Analysis, PLoS ONE 8 (2013)e55581. URL: https://dx.plos.org/10.1371/journal.pone.0055581 . doi: .W. A. Pridemore, M. B. Chamlin, E. Andreev, Reduction in Male Suicide Mortality Followingthe 2006 Russian Alcohol Policy: An Interrupted Time Series Analysis, American Journal ofPublic Health 103 (2013) 2021–2026. URL: https://ajph.aphapublications.org/doi/abs/10.2105/AJPH.2013.301405 . doi: , publisher: AmericanPublic Health Association.W. A. Pridemore, M. B. Chamlin, M. T. Kaylen, E. Andreev, The Effects of the 2006 Russian Al-cohol Policy on Alcohol-Related Mortality: An Interrupted Time Series Analysis, Alcoholism:Clinical and Experimental Research 38 (2014) 257–266. URL: http://doi.wiley.com/10.1111/acer.12256 . doi: .D. K. Humphreys, A. Gasparrini, D. J. Wiebe, Evaluating the Impact of Florida’s “StandYour Ground” Self-defense Law on Homicide and Suicide by Firearm: An Interrupted27ime Series Study, JAMA Internal Medicine 177 (2017) 44. URL: http://archinte.jamanetwork.com/article.aspx?doi=10.1001/jamainternmed.2016.6811 .doi: .K. H. Brodersen, F. Gallusser, J. Koehler, N. Remy, S. L. Scott, Inferring causal impactusing Bayesian structural time-series models, The Annals of Applied Statistics 9 (2015)247–274. URL: https://projecteuclid.org/euclid.aoas/1430226092 . doi: , publisher: Institute of Mathematical Statistics.M. A. Andresen, N. Malleson, Crime seasonality and its variations across space, Applied Geog-raphy 43 (2013) 25–35. URL: . doi: .B. Klein, T. LaRocky, S. McCabey, L. Torresy, F. Privitera, B. Lake, M. U. Kraemer, J. S. Brown-stein, D. Lazer, T. Eliassi-Rad, Assessing changes in commuting and individual mobility in majormetropolitan areas in the United States during the COVID-19 outbreak, preprint, 2020.A. Ruiz-Euler, F. Privitera, D. Giuffrida, B. Lake, I. Zara, Mobility Patterns and Income Distributionin Times of Crisis: U.S. Urban Centers During the COVID-19 Pandemic, preprint, 2020. URL: https://ssrn.com/abstract=3572324 .E. G. Cohn, Weather and Crime, The British Journal of Criminology 30 (1990) 51–64. URL: https://academic.oup.com/bjc/article/30/1/51/523870 . doi: , publisher: Oxford Academic.S. Field, The Effect of Temperature on Crime, The British Journal of Criminology 32 (1992) 340–351. URL: https://academic.oup.com/bjc/article/32/3/340/319313 . doi: , publisher: Oxford Academic.S. Towers, S. Chen, A. Malik, D. Ebert, Factors influencing temporal patterns in crime in alarge American city: A predictive analytics perspective, PLoS ONE 13 (2018). URL: . doi: .N. Shover, Burglary, Crime and Justice 14 (1991) 73–113. URL: .E. E. Mustaine, Victimization risks and routine activities: A theoretical examination using agender-specific and domain-specific model, American Journal of Criminal Justice 22 (1997)41. URL: https://link.springer.com/article/10.1007/BF02887340 .A. Tseloni, D. R. Osborn, A. Trickett, K. Pease, Modelling Property Crime Using the BritishCrime Survey: What Have We Learnt?, British Journal of Criminology 42 (2002) 109–128. URL: odelling-property-crime-using-the-british-crime-survey-what-have-we-learnt(59a91227-09c6-4474-b0da-9259a4f001ac).html . doi: .M. Cherbonneau, R. Wright, Auto Theft, in: M. H. Tonry (Ed.), The Oxford Handbook ofCrime and Public Policy, Oxford University Press, 2011, pp. 191–222. Google-Books-ID: ZcYV-DAAAQBAJ.S. K. Brooks, R. K. Webster, L. E. Smith, L. Woodland, S. Wessely, N. Greenberg, G. J. Rubin, Thepsychological impact of quarantine and how to reduce it: rapid review of the evidence, TheLancet 395 (2020) 912–920. URL: https://linkinghub.elsevier.com/retrieve/pii/S0140673620304608 . doi: .S. Colbert, C. Wilkinson, L. Thornton, R. Richmond, COVID-19 and alcohol in Australia: Industrychanges and public health impacts, Drug and Alcohol Review (2020). URL: . doi: .R. N. Parker, K. Auerhahn, Alcohol, Drugs, and Violence, Annual Review of Sociology24 (1998) 291–311. URL: https://doi.org/10.1146/annurev.soc.24.1.291 . doi: , _eprint: https://doi.org/10.1146/annurev.soc.24.1.291.L. Eriksson, P. Mazerolle, A general strain theory of intimate partner homicide, Aggression andViolent Behavior 18 (2013) 462–470. URL: . doi: .J. M. MacDonald, P. W. Manz, G. P. Alpert, R. G. Dunham, Police use of force: examining the rela-tionship between calls for service and the balance of police force and suspect resistance, Journalof Criminal Justice 31 (2003) 119–127. URL: . doi: .K. Hayward, Situational Crime Prevention and its Discontents: Rational ChoiceTheory versus the ‘Culture of Now’, Social Policy & Administration 41 (2007)232–250. URL: https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-9515.2007.00550.x . doi: , _eprint:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1467-9515.2007.00550.x.D. Longshore, S. Turner, Self-Control and Criminal Opportunity: Cross-Sectional Test of theGeneral Theory of Crime, Criminal Justice and Behavior 25 (1998) 81–98. URL: https://doi.org/10.1177/0093854898025001005 . doi: , pub-lisher: SAGE Publications Inc.G. W. Paquin, A Statewide Survey of Reactions to Neighbors’ Domestic Violence, Jour-nal of Interpersonal Violence 9 (1994) 493–502. URL: https://doi.org/10.1177/ . doi: , publisher: SAGE PublicationsInc.M. R. McCart, D. W. Smith, G. K. Sawyer, Help seeking among victims of crime:A review of the empirical literature, Journal of Traumatic Stress 23 (2010) 198–206. URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/jts.20509 . doi: , _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/jts.20509.S. Wee, M.-J. Todd, M. Oshiro, E. Greene, V. Frye, Modifiers of Neighbors’ Bystander In-tervention in Intimate Partner Violence: A Concept Mapping Study, Violence and Gen-der 3 (2016) 55–63. URL: .doi: , publisher: Mary Ann Liebert, Inc., publishers.B. J. Doran, B. G. Lees, Investigating the Spatiotemporal Links Between Disor-der, Crime, and the Fear of Crime, The Professional Geographer 57 (2005) 1–12.URL: . doi: http://link.springer.com/10.1007/s10940-008-9047-5 . doi: .G. O. Mohler, M. B. Short, P. J. Brantingham, F. P. Schoenberg, G. E. Tita, Self-Exciting Point Pro-cess Modeling of Crime, Journal of the American Statistical Association 106 (2011) 100–108.URL: . doi: .D. Weisburd, The Law of Crime Concentration and the Criminology of Place, Criminology53 (2015) 133–157. URL: http://doi.wiley.com/10.1111/1745-9125.12070 . doi: . 30 upplementary Materials Daily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 26 317 26 317Prediction (S.D.) 27 (2.2) 327 (26.9) 27 (2.2) 327 (27.1)95% C.I. [23,32] [275,379] [23,31] [274, 374]Absolute Effect (S.D.) -0.79 (2.2) -9.50 (26.9) -0.41 (2.3) -4.89 (27.1)95% C.I. [-5.1, 3.5] [-61.6, 42.2] [-4.7, 3.6] [-56.9, 43.1]Relative Effect (S.D.) -2.98% (8.2%) -2.98% (8.2%) -1.5% (8.4%) -1.5% (8.4%)95% C.I. [-19%, 13%] [-19%, 13%] [-18%, 13%] [-18%, 13%]Post. tail-area prob. p: 0.36409 0.48134Post. prob. causal effect: 64% 52%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 24 604 24 604Prediction (S.D.) 27 (1.8) 675 (45.7) 26 (1.5) 644 (38.4)95% C.I. [23, 30] [585, 760] [23, 29] [568, 718]Absolute Effect (S.D.) -2.9 (1.8) -71.3 (45.7) -1.6 (1.5) -40.4 (38.4)95% C.I. [-6.2, 0.76] [-155.6, 19.00] [-4.6, 1.4] [-114.2, 35.6]Relative Effect (S.D.) -11% (6.8%) -11% (6.8%) -6.3% (6%) -6.3% (6%)95% C.I. [-23%, 2.8%] [-23%, 2.8%] [-18%, 5.5%] [-18%, 5.5%]Post. tail-area prob. p: 0.0625 0.14801Post. prob. causal effect: 94% 85%
Table A1: Causal Impact Analysis - Assaults with Deadly Weapons31 aily crime counts up to March 16 th Univariate With Cov.Average Cumulative Average Cumulative
Actual 50 597 50 597Prediction (S.D.) 50 (2.9) 601 (34.4) 49 (2.7) 592 (32.5)95% C.I. [44, 56] [533, 668] [44, 54] [530, 652]Absolute Effect (S.D.) -0.3 (2.9) -3.6 (34.4) 0.39 (2.7) 4.62 (32.5)95% C.I. [-5.9, 5.3] [-71.2, 64.1] [-4.5, 5.6] [-54.5, 66.7]Relative Effect (S.D.) -0.6% (5.7%) -0.6% (5.7%) 0.78% (5.5%) 0.78% (5.5%)95% C.I. [-12%, 11%] [-12%, 11%] [-9.2%, 11%] [-9.2%, 11%]Post. tail-area prob. p: 0.41895 0.40672Post. prob. causal effect: 58% 59%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 45 1116 45 1116Prediction (S.D.) 50 (2.5) 1256 (63.1) 48 (2) 1207 (50)95% C.I. [45, 55] [1128, 1380] [44, 52] [1111, 1307]Absolute Effect (S.D.) -5.6 (2.5) -139.7 (63.1) -3.7 (2) -91.3 (50)95% C.I. [-11, -0.5] [-264, -12.4] [-7.6, 0.19] [-191.1, 4.74]Relative Effect (S.D.) -11% (5%) -11% (5%) -7.6% (4.1%) -7.6% (4.1%)95% C.I. [-21%, -0.99%] [-21%, -0.99%] [-16%, 0.39%] [-16%, 0.39%]Post. tail-area prob. p: 0.02486 0.03534Post. prob. causal effect: 98% 96%
Table A2: Causal Impact Analysis - Battery and Simple Assault32 aily crime counts up to March 16 th Univariate With Cov.Average Cumulative Average Cumulative
Actual 34 406 34 406Prediction (S.D.) 34 (2.4) 402 (29.2) 34 (2.5) 408 (29.6)95% C.I. [29, 38] [345, 461] [30, 39] [359, 465]Absolute Effect (S.D.) 0.3 (2.4) 3.6 (29.2) -0.2 (2.5) -2.4 (29.6)95% C.I. [-4.6, 5.1] [-54.9, 60.8] [-4.9, 3.9] [-59.2, 46.6]Relative Effect (S.D.) 0.89% (7.2%) 0.89% (7.2%) -0.58% (7.2%) -0.58% (7.2%)95% C.I. [-14%, 15%] [-14%, 15%] [-14%, 11%] [-14%, 11%]Post. tail-area prob. p: 0.44173 0.49627Post. prob. causal effect: 56% 50%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 33 816 33 816Prediction (S.D.) 34 (1.9) 858 (46.3) 35 (1.8) 881 (46.1)95% C.I. [31, 38] [769, 948] [31, 39] [787, 966]Absolute Effect (S.D.) -1.7 (1.9) -41.6 (46.3) -2.6 (1.8) -64.5 (46.1)95% C.I. [-5.3, 1.9] [-131.7, 47.1] [-6, 1.2] [-150, 29.3]Relative Effect (S.D.) -4.8% (5.4%) -4.8% (5.4%) -7.3% (5.2%) -7.3% (5.2%)95% C.I. [-15%, 5.5%] [-15%, 5.5%] [-17%, 3.3%] [-17%, 3.3%]Post. tail-area prob. p: 0.19832 0.09328Post. prob. causal effect: 80% 91%
Table A3: Causal Impact Analysis - Burglary33 aily crime counts up to March 16 th Univariate With Cov.Average Cumulative Average Cumulative
Actual 38 454 38 454Prediction (S.D.) 39 (2.3) 473 (27.5) 39 (2.1) 466 (25.3)95% C.I. [35, 44] [424, 529] [34, 43] [414, 512]Absolute Effect (S.D.) -1.6 (2.3) -18.8 (27.5) -0.96 (2.1) -11.52 (25.3)95% C.I. [-6.3, 2.5] [-75.4, 30.3] [-4.9, 3.4] [-58.4, 40.2]Relative Effect (S.D.) -4% (5.8%) -4% (5.8%) -2.5% (5.4%) -2.5% (5.4%)95% C.I. [-16%, 6.4%] [-16%, 6.4%] [-13%, 8.6%] [-13%, 8.6%]Post. tail-area prob. p: 0.24535 0.37313Post. prob. causal effect: 75% 63%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 39 969 39 969Prediction (S.D.) 39 (2.2) 972 (56.0) 38 (1.7) 938 (41.7)95% C.I. [34, 43] [860, 1071] [34, 41] [855, 1021]Absolute Effect (S.D.) -0.11 (2.2) -2.68 (56.0) 1.2 (1.7) 30.9 (41.7)95% C.I. [-4.1, 4.4] [-102.3, 109.3] [-2.1, 4.5] [-52.1, 113.7]Relative Effect (S.D.) -0.28% (5.8%) -0.28% (5.8%) 3.3% (4.4%) 3.3% (4.4%)95% C.I. [-11%, 11%] [-11%, 11%] [-5.6%, 12%] [-5.6%, 12%]Post. tail-area prob. p: 0.48324 0.22015Post. prob. causal effect: 52% 78%
Table A4: Causal Impact Analysis - Intimate Partner Assault34 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 17 202 17 202Prediction (S.D.) 22 (1.7) 266 (20.5) 22 (1.6) 262 (19.5)95% C.I. [19, 25] [225, 304] [19, 25] [225, 302]Absolute Effect (S.D.) -5.3 (1.7) -63.6 (20.5) -5 (1.6) -60 (19.5)95% C.I. [-8.5, -1.9] [-101.8, -22.6] [-8.3, -1.9] [-100.1, -22.8]Relative Effect (S.D.) -24% (7.7%) -24% (7.7%) -23% (7.5%) -23% (7.5%)95% C.I. [-38%, -8.5%] [-38%, -8.5%] [-38%, -8.7%] [-38%, -8.7%]Post. tail-area prob. p: 0.00333 0.00208Post. prob. causal effect: 99.67% 99.79%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 18 439 18 439Prediction (S.D.) 22 (1.3) 553 (32.8) 22 (1.2) 545 (30.4)95% C.I. [20, 25] [490, 619] [19, 24] [486, 604]Absolute Effect (S.D.) -4.6 (1.3) -114.2 (32.8) -4.3 (1.2) -106.3 (30.4)95% C.I. [-7.2, -2.1] [-180.5, -51.3] [-6.6, -1.9] [-165.4, -47.4]Relative Effect (S.D.) -21% (5.9%) -21% (5.9%) -19% (5.6%) -19% (5.6%)95% C.I. [-33%, -9.3%] [-33%, -9.3%] [-30%, -8.7%] [-30%, -8.7%]Post. tail-area prob. p: 0.00298 0.00218Post. prob. causal effect: 99.70% 99.78%
Table A5: Causal Impact Analysis - Robbery35 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 16 189 16 189Prediction (S.D.) 18 (1.5) 220 (18.2) 19 (1.5) 223 (17.7)95% C.I. [15, 21] [184, 255] [16, 21] [188, 256]Absolute Effect (S.D.) -2.6 (1.5) -31.1 (18.2) -2.8 (1.5) -34.1 (17.7)95% C.I. [-5.5, 0.44] [-66.3, 5.30] [-5.6, 0.064] [-67.4, 0.766]Relative Effect (S.D.) -14% (8.3%) -14% (8.3%) -15% (7.9%) -15% (7.9%)95% C.I. [-30%, 2.4%] [-30%, 2.4%] [-30%, 0.34%] [-30%, 0.34%]Post. tail-area prob. p: 0.04353 0.02808Post. prob. causal effect: 95.64% 97.19%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 13 320 13 320Prediction (S.D.) 18 (1) 462 (26) 19 (1) 471 (26)95% C.I. [16, 21] [412, 513] [17, 21] [421, 521]Absolute Effect (S.D.) -5.7 (1) -142.0 (26) -6.1 (1) -151.5 (26)95% C.I. [-7.7, -3.7] [-193.4, -92.0] [-8, -4.1] [-201, -101.3]Relative Effect (S.D.) -31% (5.6%) -31% (5.6%) -32% (5.5%) -32% (5.5%)95% C.I. [-42%, -20%] [-42%, -20%] [-43%, -21%] [-43%, -21%]Post. tail-area prob. p: 0.00348 0.001Post. prob. causal effect: 99.65% 99.90%
Table A6: Causal Impact Analysis - Shoplifting36 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 55 662 55 662Prediction (S.D.) 61 (2.9) 728 (35.3) 61 (2.8) 732 (33.2)95% C.I. [55, 66] [658, 797] [56, 67] [670, 801]Absolute Effect (S.D.) -5.5 (2.9) -66.1 (35.3) -5.8 (2.8) -70.0 (33.2)95% C.I. [-11, 0.34] [-135, 4.13] [-12, -0.64] [-139, -7.67]Relative Effect (S.D.) -9.1% (4.8%) -9.1% (4.8%) -9.6% (4.5%) -9.6% (4.5%)95% C.I. [-19%, 0.57%] [-19%, 0.57%] [-19%, -1%] [-19%, -1%]Posterior tail-area probability p: 0.0333 0.01663Posterior prob. of a causal effect: 96.67% 98.33%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 47 1175 47 1175Prediction (S.D.) 62 (2.2) 1548 (56.2) 62 (2.1) 1557 (53.4)95% C.I. [58, 66] [1444, 1651] [58, 66] [1453, 1658]Absolute Effect (S.D.) -15 (2.2) -373 (56.2) -15 (2.1) -382 (53.4)95% C.I. [-19, -11] [-476, -269] [-19, -11] [-483, -278]Relative Effect (S.D.) -24% (3.6%) -24% (3.6%) -25% (3.4%) -25% (3.4%)95% C.I. [-31%, -17%] [-31%, -17%] [-31%, -18%] [-31%, -18%]Post. tail-area prob. p: 0.0035 0.00109Post. prob. causal effect: 99.65% 99.89%
Table A7: Causal Impact Analysis - Thefts37 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 45 536 45 536Prediction (S.D.) 44 (2.4) 530 (28.6) 45 (2.4) 536 (28.8)95% C.I. [40, 49] [477, 586] [40, 49] [483, 591]Absolute Effect (S.D.) 0.46 (2.4) 5.55 (28.6) 0.029 (2.4) 0.343 (28.8)95% C.I. [-4.2, 4.9] [-49.8, 58.8] [-4.6, 4.4] [-55.1, 52.8]Relative Effect (S.D.) 1% (5.4%) 1% (5.4%) 0.064% (5.4%) 0.064% (5.4%)95% C.I. [-9.4%, 11%] [-9.4%, 11%] [-10%, 9.9%] [-10%, 9.9%]Post. tail-area prob. p: 0.41646 0.48507Post. prob. causal effect: 58% 51%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 46 1140 46 1140Prediction (S.D.) 45 (2) 1123 (49) 46 (1.7) 1141 (43.3)95% C.I. [41, 49] [1032, 1214] [42, 49] [1054, 1225]Absolute Effect (S.D.) 0.67 (2) 16.84 (49) -0.055 (1.7) -1.369 (43.3)95% C.I. [-2.9, 4.3] [-73.6, 107.8] [-3.4, 3.4] [-85.0, 86.2]Relative Effect (S.D.) 1.5% (4.4%) 1.5% (4.4%) -0.12% (3.8%) -0.12% (3.8%)95% C.I. [-6.5%, 9.6%] [-6.5%, 9.6%] [-7.4%, 7.5%] [-7.4%, 7.5%]Post. tail-area prob. p: 0.37151 0.49896Post. prob. causal effect: 63% 50%
Table A8: Causal Impact Analysis - Stolen Vehicles38 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 0.62 8.00 0.62 8.00Prediction (S.D.) 0.72 (0.26) 9.38 (3.36) 0.69 (0.25) 8.94 (3.28)95% C.I. [0.21, 1.3] [2.69, 16.3] [0.21, 1.2] [2.73, 15.5]Absolute Effect (S.D.) -0.11 (0.26) -1.38 (3.36) -0.072 (0.25) -0.935 (3.28)95% C.I. [-0.64, 0.41] [-8.28, 5.31] [-0.58, 0.41] [-7.48, 5.27]Relative Effect (S.D.) -15% (36%) -15% (36%) -10% (37%) -10% (37%)95% C.I. [-88%, 57%] [-88%, 57%] [-84%, 59%] [-84%, 59%]Post. tail-area prob. p: 0.326 0.37437Post. prob. causal effect: 67% 63%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 0.52 13.00 0.52 13.00Prediction (S.D.) 0.72 (0.19) 18.11 (4.83) 0.69 (0.19) 17.14 (4.84)95% C.I. [0.34, 1.1] [8.39, 27.3] [0.31, 1] [7.71, 26]Absolute Effect (S.D.) -0.2 (0.19) -5.1 (4.83) -0.17 (0.19) -4.14 (4.84)95% C.I. [-0.57, 0.18] [-14.34, 4.61] [-0.52, 0.21] [-13.04, 5.29]Relative Effect (S.D.) -28% (27%) -28% (27%) -24% (28%) -24% (28%)95% C.I. [-79%, 25%] [-79%, 25%] [-76%, 31%] [-76%, 31%]Post. tail-area prob. p: 0.13814 0.2012Post. prob. causal effect: 86% 80%
Table A9: Causal Impact Analysis - Homicides39 aily crime counts up to March 16 th Univariate With Cov.Avg. Cum. Avg. Cum.
Actual 558 6700 558 6700Prediction (S.D.) 592 (14) 7098 (165) 590 (13) 7086 (152)95% C.I. [567, 620] [6805, 7441] [564, 615] [6772, 7376]Absolute Effect (S.D.) -33 (14) -398 (165) -32 (13) -386 (152)95% C.I. [-62, -8.8] [-741, -105.1] [-56, -6] [-676, -72]Relative Effect (S.D.) -5.6% (2.3%) -5.6% (2.3%) -5.4% (2.1%) -5.4% (2.1%)95% C.I. [-10%, -1.5%] [-10%, -1.5%] [-9.5%, -1%] [-9.5%, -1%]Post. tail-area prob. p: 0.00555 0.01493Post. prob. causal effect: 99.44% 98.51%
Daily crime counts up to March 28 th Univariate With Cov.Avg. Cum. Avg. Cum.