A global analysis of the impact of COVID-19 stay-at-home restrictions on crime | Nature Human Behaviour

The impact of stay-at-home restrictions on crime

The COVID-19 pandemic and subsequent restrictions represent a series of ‘natural experiments’ in which population-wide changes affected routines, social interactions and the use of public space. An interrupted time series (ITS) design can then be used to assess the impact of the treatment while accounting for pre-COVID-19 crime trends18. ITS analyses provide an estimate of changes in levels of crime following an ‘interruption’ in the time series, while accounting for potential confounders such as long-term trends, autocorrelation and other time-varying confounders19. In an ITS analysis, the assumption is that, without the intervention (that is, COVID-19 restrictions), there would be no change in the pre-intervention trend18.

The dependent variable in our analyses is police-recorded daily reported crime incidents for six major crime categories: assault, theft, burglary, robbery, vehicle theft and homicide. To ensure that the crime categories were as comparable as possible across contexts, we utilized definitions from the International Classification of Crime for Statistical Purposes20 for reference when collecting and aggregating crime data from each site (Supplementary Table 3). Not all crime categories were available for each city, and in some contexts certain crimes are not treated as separate categories. For example, in Seoul, burglary is not considered separately from robbery, and motor vehicle theft is not distinguished from theft. To ensure that the crime categories are as comparable as possible, we excluded these combined outcomes from the analyses (Supplementary Tables 4–10).

The treatment variable for the current analyses is a dummy variable defined by the date on which stay-at-home restrictions or recommendations were first implemented in each city, state/province or country (Supplementary Table 11). The effects of stay-at-home restrictions are modelled as a step function, whereby 0 represents the time period before and (if applicable) after the implementation of stay-at-home restrictions, and 1 represents the time period during stay-at-home restrictions. In this way, the step function estimates the extent to which the restrictions had an immediate effect on crime during the intervention period.

Given the count nature of our data, and the variation in frequency of daily crime incidents across cities (ranging from 0 to >500 daily incidents), in the present analyses we estimated Poisson generalized linear models using a logit-link function. This flexible approach provides an estimate of the level change in crime incidents after the implementation of stay-at-home restrictions. All tests are two-tailed, and models adjust for seasonality, autocorrelation and potential outliers. In addition, we included average daily temperature in Celsius as a covariate to account for potential fluctuations in crime due to higher temperatures21.

As an initial step, we conducted a series of descriptive analyses to evaluate the changes in crime before and after the implementation of COVID-19 stay-at-home restrictions. First, we calculated the average number of crimes in each city and category before and after the implementation of restrictions (Supplementary Table 15). Second, we plotted the 7-day moving average of daily crime counts for each city and crime. The moving average trend was indexed to equal 100 at the date on which the first stay-at-home restrictions were implemented. In this way, we can compare the direction of the trend immediately before and following restrictions across cities with different levels of crime. The mean trend for each type of crime was plotted alongside each city’s trend (Fig. 1). Supplementary Figs. 1–6 present the moving average trends broken down by city and type of crime. The full (non-smoothed) time series plots for each city and type of crime can be found in Supplementary Figs. 7–25.

Fig. 1: Seven-day moving average time series plots of daily numbers of crimes.

a, Assault (n = 23). b, Burglary (n = 20). c, Robbery (n = 24). d, Theft (n = 16). e, Vehicle theft (n = 20). f, Homicide (n = 25). Each time series is indexed to equal 100 on the day the first stay-at-home restrictions were implemented. The blue line indicates the average trend across all cities with available data. Zero time is the date on which stay-at-home restrictions were implemented.

Full size image

The descriptive results suggest that stay-at-home restrictions are associated with declines in all types of crime, with the exception of homicide. In Barcelona, for example, police-recorded thefts declined from an average of 385.2 to 38.1 per day (Supplementary Table 15). However, there still appears to be substantial variation across crime categories and cities in the size and direction of crime trends following the implementation of restrictions. Over time, the mean trend begins to return to pre-treatment levels of crime.

Next, we estimated the size of the level change in daily crimes that can be attributed to stay-at-home restrictions using ITS analysis. The analyses of trends for six categories of police-recorded daily reported crime incidents across 27 cities result in over 100 estimates of effect. To summarize this information, we used meta-analytical techniques to estimate the grand mean effect of stay-at-home restrictions for each type of crime (Table 1). The estimates of effect, expressed as the incidence rate ratio (IRR) with 95% confidence interval, are presented in Figs. 2 and 3 for violent and property crimes, respectively. The high number of hypotheses tested increases the possibility that we may detect a significant result due to chance. We therefore urge caution in interpreting the results for individual cities. The breakdown of effect sizes and summary effects by city are available in Supplementary Table 16. Across our sample, crime declined overall by 37% following the implementation of stay-at-home restrictions.

Table 1 Summary effect sizes from meta-analyses using cities with any available crime categories

Full size table

Fig. 2: IRR and 95% CI of stay-at-home restrictions on daily number of violent crimes.

a, Assault (n = 23). b, Robbery (n = 24). c, Homicide (n = 21). Overall summary effects estimated using random-effects meta-analytic techniques. ES, effect size. SaH (days), number of days under stay-at-home restrictions from the beginning of 2020 until the end of the respective time series from May to September 2020. Full results by city and crime can be found in Supplementary Table 17.

Full size image

Fig. 3: IRR and 95% CI of stay-at-home restrictions on daily number of property crimes.

a, Burglary (n = 20). b, Theft (n = 16). c, Vehicle theft (n = 20). Overall summary effects estimated using random-effects meta-analytic techniques. ES, effect size. SaH (days), number of days under stay-at-home restrictions from the beginning of 2020 until the end of the respective time series from May to September 2020. Full results by city and crime can be found in Supplementary Table 17.

Full size image

For assault, the summary effect suggests that the implementation of stay-at-home restrictions was associated with a 35% reduction in daily assaults (Fig. 2a). The I2 value of 98.4% suggests substantial heterogeneity in the effect sizes across cities and crime outcomes (Table 1). Similarly, effect sizes for robbery vary across cities, however no cities experienced a statistically significant increase in the number of daily robbery incidents following restrictions. The average size of the level change following restrictions was 46%.

The results for homicide suggest that overall there was a marginal decline in the number of daily homicides following the implementation of stay-at-home restrictions (14%, Fig. 2c). However, only three cities (Lima, Cali and Rio de Janeiro) saw a statistically significant decline in homicides. The I2 statistic (54.6%, Table 1) indicates relatively less heterogeneity in effects compared with other crime outcomes.

The distribution of effect sizes for burglary ranges from an 84% decline (Lima) to a 38% increase (San Francisco) in the number of daily incidents. The summary effect is relatively smaller compared with assault, where on average burglaries fell by 28% following the implementation of restrictions.

All cities with available data on theft experienced a significant drop in the number of daily incidents, however the I2 statistic (99.2%) still indicates high levels of heterogeneity between cities. Even cities with less restrictive, more voluntary stay-at-home recommendations (for example, Malmö and Stockholm) experienced marginal declines in the number of daily thefts. The results for vehicle theft also suggest heterogeneity in effects across cities, with 8 out of 18 cities experiencing no statistically significant change in the number of incidents following restrictions. The mean drop in vehicle thefts across cities was 39%.

Stringency of restrictions and size of decline

The next step is to evaluate why we find such substantial heterogeneity in effect sizes across cities. Heterogeneity in effect sizes can be attributed to, for example, variations in the characteristics of local or national policies. We estimate the extent to which variations in effect sizes are associated with the stringency of stay-at-home restrictions and wider COVID-19-related containment policies. To measure stringency, we drew from the Oxford Government Response Tracker documentation and coding of COVID-19 policy responses22. The stringency of stay-at-home restrictions is measured on a scale from 0 (no measures) to 3 (do not leave the house with minimal exceptions).

For the current analyses, we took the average of the stay-at-home scores between the first day of implementation to either the lifting of restrictions or the end of the time series, whichever came first (Supplementary Table 11). Using mixed-effects meta-regression techniques, we are able to assess whether more severe restrictions on routine activities are associated with greater declines in daily crimes (that is, larger negative effect sizes).

The results in Table 2 suggest that more stringent stay-at-home restrictions were associated with significantly more negative effect sizes for burglary, robbery, theft and vehicle theft. In essence, this suggests that more severe restrictions on ‘non-essential’ movement and activities were associated with significantly larger declines in crime. While the coefficients are negative for assault, the association was not significant at the conventional threshold of 0.05. However, inspection of the scatterplots suggests that Barcelona may be an outlier (Supplementary Fig. 26). When Barcelona is excluded from meta-regression analyses, more stringent stay-at-home restrictions are negatively associated with effect sizes for assault (Supplementary Table 19). The adjusted R2 values for burglary and robbery show that the stringency of stay-at-home restrictions accounts for about 35% of the variation in effect sizes across cities.

Table 2 Meta-regression results for the stringency of stay-at-home restrictions and overall stringency index on the size of the effect of stay-at-home orders on police-recorded crime

Full size table

Additional analyses

As an additional set of analyses, we evaluated the possibility that other COVID-19-related policy responses may account for variations in the size of the effect instead of, or in addition to, stay-at-home restrictions. For example, based on strain perspectives, we may expect smaller declines in cities and contexts where there is less economic support for those affected by unemployment or financial strain due to the pandemic. This would be because individuals experiencing strain are motivated to cope by seeking out alternative, possibly illegal income opportunities. In addition, since stay-at-home restrictions were often implemented alongside a wide range of policies that regulated leisure activities, routines and opportunities, we also examined the relationship between the overall stringency index and effect sizes for each type of crime.

The results show that more severe restrictions on school opening, working from home, public events, private gatherings and internal movement are not significantly related to the size of effects (Supplementary Table 21), with one exception: More stringent reductions or closures of public transportation are associated with more negative effect sizes for robbery and vehicle theft only. More economic support was not associated with the variation in effect sizes for any type of crime. The results for the overall stringency index were generally in line with the main findings for stay-at-home restrictions, whereby more stringent combinations of containment policies were associated with greater declines in burglary, robbery, vehicle theft and theft. However, comparing the model fit (adjusted R2 values) suggests that accounting for the overall combined policy response does not substantially improve the model fit.

Further, while the stringency indices and sub-indices provide systematic and comparable measures of COVID-19 containment policies across countries, they do not provide a measure of actual behavioural changes. We therefore conducted additional analyses to assess the relationship between changes in mobility indices as measured by the Google COVID Community Mobility Reports23,24, and effect sizes for each type of crime. Bivariate correlations between mobility measures and stringency measures suggest that more stringent stay-at-home restrictions are associated with greater declines in visits to commercial locations and parks, as well as increases in users remaining in their residences (Supplementary Table 14). The results using mobility indices are generally in line with the results using the stringency index measures, whereby cities that saw greater declines in the use of public space saw larger declines in crime, with the exception of homicide (Supplementary Table 22).