Changes in Crime Rates During COVID-19 Pandemic

Introduction and Overview

Study Context and Authors

This study, titled "Changes in Crime Rates during the COVID-19 Pandemic," examines the impact of the COVID-19 pandemic and associated societal events on crime rates. It was authored by Mikaela Meyer, Ahmed Hassafy, Gina Lewis, Prasun Shrestha, Amelia M. Haviland, and Daniel S. Nagin, and published in Statistics and Public Policy (2022, Vol. 9, No. 1, pp. 97–109). The article received date was June 2021 and accepted in April 2022, with online publication on June 2, 2022.

Research Questions and Scope

The research estimates changes in the rates of five FBI Part 1 crimes during two key periods in 2020: the spring COVID-19 pandemic lockdown and the period following the killing of George Floyd, extending through December 2020. The analysis uses weekly crime rate data from 28 of the 70 largest cities in the United States, spanning January 2018 to December 2020. The primary aim is to meticulously document changes or the absence of changes in crime rates, while accounting for seasonal trends, rather than to ascertain the underlying specific causes which could include social unrest, economic conditions, or altered police behavior.

Abstract Summary of Key Findings

• Homicide rates were elevated throughout 2020, even prior to the March lockdowns.

• Auto thefts saw significant increases during the summer and the remainder of 2020.

• Robbery and larceny experienced significant declines across all three post-pandemic periods.

• Burglary rates showed point estimates suggesting a decline for all four periods of 2020, but only the pre-pandemic period's decline was statistically significant.

• An openness index, developed to measure the degree of city openness during lockdowns, revealed a positive and significant association with larceny and robbery rates. This implies that lockdown restrictions effectively reduced these specific offense rates, whereas no detectable association was found for the other three crime types.

• While opportunity theory offers a plausible partial explanation for some findings, no single overarching crime theory can fully account for all the observed results.

Keywords

COVID-19; Crime trends; Pandemic lockdowns; Part 1 crimes.

Comparison to Existing Literature and Study Contributions

This study contributes to a nascent body of academic work analyzing crime rates during the pandemic by addressing several gaps found in earlier analyses:

• Extended Data Coverage: This research analyzes data through December 2020 from 28 cities, a longer and broader scope compared to many early analyses that usually ended by summer 2020 or focused on fewer cities. This allows for a more comprehensive examination of seasonal trends and the persistence of crime rate changes.

• Systematic Change Focus: Unlike studies that focused on identifying crime spikes in individual cities, this research aimed to detect evidence of systematic change across a combined analysis of multiple cities.

• City Representativeness: To the authors' knowledge, this is the first study to compare annual crime rates between the 28 included cities and all 70 largest cities in the U.S., demonstrating similar distributions of crime rates in 2018.

• Detailed Lockdown Measures: The study uniquely assembled data on the precise timing and strictness of city-level lockdowns over time, allowing for the creation and inclusion of an 'openness index' in the analysis.

• Comprehensive Period Coverage: It focuses specifically on the impact of initial lockdowns, the subsequent relaxation of restrictions, and the period following George Floyd's killing through the entire year of 2020, using panel regression models to account for seasonality and year-to-year trends.

• Post-George Floyd Analysis: This is one of the few articles to analyze crime trends across multiple cities following the killing of George Floyd, retaining the full weekly panel structure of the data, in contrast to some studies that averaged crime rates across cities into a single time series.

Previous studies cited include:

• Piquero et al. (2020): Analyzed a spike in domestic violence in Dallas during the initial two weeks of lockdown.

• Campedelli, Aziani, and Favarin (2021): Found mixed associations with lockdowns and crime rates in Los Angeles in March 2020.

• Mohler et al. (2020): Examined daily police calls-for-service in Los Angeles and Indianapolis, yielding mixed results.

• Ashby (2020): Used SARIMA models to assess changes in specific crime classes in 16 large U.S. cities up to May 10th, finding no evidence for serious assault changes and city-specific variations.

• Abrams (2020): Analyzed data from 23 U.S. cities, reporting no statistically significant increase in homicides in the four weeks post-lockdown, but declines in drug, violent, and most property crimes.

Data and Methodology

Data Sources

• Outcome Variables: Weekly crime counts for specific crime types, sourced from publicly available crime incident reports of 28 of the 70 largest U.S. cities. These 28 cities provided daily or weekly data in an accessible format for 2018, 2019, and 2020. Examples include cities like Nashville, Louisville, Houston, and Seattle, which published NIBRS data.

• Crime Types Analyzed: Five of the seven FBI UCR Part 1 crimes:

  • Homicide

  • Robbery

  • Burglary

  • Larceny

  • Auto Theft

• Excluded Crime Types: Due to data quality concerns:

  • Rape: While 26 cities provided data, inconsistencies in classification across cities led to its exclusion.

  • Aggravated Assault: Aggregated yearly totals from weekly data in 2018 were, on average, 20\% lower than UCR reported totals, indicating inconsistencies in classifying "aggravated" versus less serious assaults. Results for aggravated assault are included in supplementary materials.

• Population Data: U.S. Census Bureau Subcounty Resident Population Estimates from 2010 to 2019 were used to calculate rates per capita.

• City Representativeness: It was confirmed that the distributions of annual crime rates in 2018 for the 28 cities in the study were similar to those for all 70 most populous U.S. cities (refer to Figures A.1–A.5 in supplementary materials).

• Lockdown and Reopening Data: Information on lockdown dates and the timing/strictness of reopening phases was collected from media sources (CNN 2020; The New York Times 2020) and state/county health departments. This data covered 27 of 28 cities for lockdown dates and 23 of 28 cities for reopening phases. Data for Milwaukee (lockdown date and reopening phases), Mesa, Raleigh, St. Paul, and Tulsa (reopening phases) was incomplete.

Lockdown Measures

Two distinct measures were created to quantify the lockdown stage:

1. Time Since Lockdown: This measure tracked the amount of time (in two-week intervals) that had passed since a city's initial lockdown began. These intervals were specific to each city's lockdown timeline.

2. Openness Index: An integer scale from 0 (least open) to 14 (most open). A value of 14 was assigned to all weeks prior to a city's COVID-related lockdown. Post-lockdown values were determined by the operational status of seven economic sectors: "Entertainment," "Food and Drink," "Industries," "Outdoor and Recreation," "Personal Care," "Places of Worship," and "Retail." Each sector contributed 0 (not operating), 1 (some industries not full capacity), or 2 (all industries full capacity) to the index. The variation of this index across cities and over time is visualized in Figure A.6 and Figure 1 (supplementary materials).

Statistical Models

The study employed two sets of Poisson regression models, with the city-week pair as the unit of analysis. All models included an offset term for city population, city fixed effects, and year and month fixed effects. Standard errors were clustered at the city-level. Week units were seven days long, composing approximate month units of four or five full weeks.

Response Variable

For all models, the response variable was the city-wide weekly crime count, denoted as y_{cwr}, where c represents the city, w is the week of the year, and r is the year.

Model of Crime Rates by Openness Index

This model examined the relationship between crime rates and city openness levels during the lockdown period, concluding in Week 21 of 2020 (the week of George Floyd's killing). The model had 2695 city-week observations for homicide and 2818 for other crime types. The equation is:
E[y{cwr}] = ext{exp}(1 imes ext{Pop}c + ext{offset} + ext{log}( ext{Pop}c) + ext{fixedeffects} + ext{opennessindex}{cwr})
In the provided text, the fixed effects were broken out as follows:
E[y{cwr}] = ext{exp}( ext{log}( ext{Pop}c) + ext{intercept} + ext{Pop}c + ext{alpha} + ext{gamma}r + ext{theta}{m(w)} + ext{zeta}c + eta imes ext{openness index}_{cwr})

Where:

• ext{Pop}_c: Offset term for the population size of each city.

• ext{alpha}: Intercept reflecting log crime rates per week in January 2018 for the omitted city (Atlanta).

• ext{gamma}_r: Year fixed effects (2018 set to zero).

• ext{theta}_{m(w)}: Month fixed effects (January set to zero).

• ext{zeta}_c: City fixed effects (22 parameters).

• eta: Coefficient for the openness index.

The coefficient of interest, exp(eta), represents the average multiplicative change in weekly crime count associated with a one-unit increase in the openness index. Polynomial specifications for the openness index were tested but showed no clear evidence of nonlinear associations.

Model of Crime Rates by Weeks into Lockdown

This model explored how crime rates changed as each city's lockdown progressed, also ending in Week 21 of 2020. The variable time since lockdown (t{cwr}) was defined as: t{cwr} = I[r=2020] imes (w - (LWc - 1))^+ where LWc is the lockdown week for city c, and (x)^+ indicates x if positive and zero otherwise. This variable ranged from 0 to 11 and was then discretized into biweekly periods (T{cwr} = 0, 2, 4, 6, 8, 10). The regression equation was: E[y{cwr}] = ext{exp}( ext{log}( ext{Pop}c) + ext{alpha} + ext{gamma}r + ext{theta}{m(w)} + ext{zeta}c + eta{T{cwr}})

Where:

• eta{T{cwr}}: Six parameters of interest associated with the biweekly time-since-lockdown periods.

• exp(eta{T{cwr}}) indicates the average multiplicative change in weekly crime count for each biweekly period, relative to the start of 2020, after accounting for seasonal trends.

The model included 3187 city-week observations for homicide and 3310 for other crime types.

Model of Seasonal Crime Rates in 2020

This comprehensive model investigated how 2020 seasonal crime trends differed from 2018 and 2019 across four specific periods:

1. Pre-Pandemic: January and February 2020.

2. Pandemic Lockdown: March through May 2020.

3. Summer Protests: June through August 2020.

4. End of Year: September through December 2020.

There were 4120 homicide city-week observations and 4273 for other crime types. The model equation was:
E[y{cwr}] = ext{exp}( ext{log}( ext{Pop}c) + ext{alpha} + ext{gamma}r + ext{theta}{m(w)} + ext{zeta}c + eta1 I[r=2020]I[w imes 8] + eta2 I[r=2020]I[8 < w imes 21] + eta3 I[r=2020]I[21 < w imes 34] + eta_4 I[r=2020]I[34 < w])
Where:

• eta1, eta2, eta3, eta4: Parameters representing the four distinct periods of 2020.

• exp(eta) on the incident rate ratio scale indicates the average multiplicative change in weekly crime count for each 2020 period, relative to the same weeks in prior years.

Post-estimation tests (two-sided t-tests) were conducted to compare these coefficients, specifically whether later 2020 periods differed significantly from the pre-pandemic period.

Results: Detailed Findings

Description of Weekly Crime Data and Openness Index

• Population Variation: The average population across the analyzed cities exceeded 1.2 million, with a substantial standard deviation of over 1.6 million, highlighting significant variation in city sizes.

• Crime Counts and Rates (2018–2019 Average):

  • Larceny: Highest average weekly count (528.278 incidents) and rate (50.805 incidents per 100,000 people).

  • Homicide: Lowest average weekly count (2.792 incidents) and rate (0.326 incidents per 100,000 people).

• Variability: The variation in crime counts between cities was considerably larger than the average week-to-week variation within cities; this pattern also held, to a lesser extent, for crime rates.

• Openness Index Visualization: Figure 1 illustrates the variation in the openness index from Week 10 to Week 21 across 23 cities, showing diverse paths to reopening (e.g., Chicago remained closed past Week 21, Dallas had multiple reopening steps, Boston reopened minimally).

Degree of Openness and Time into Lockdown Analyses

Openness Index Findings (Equation (2))

• Larceny and Robbery: Showed substantial, positive, and statistically significant associations with the openness index (p < 0.001). This means as cities became more open, weekly rates for larceny and robbery increased, implying that lockdown restrictions effectively reduced these offenses. For example, in Week 10 (index = 14), the robbery rate was predicted to be 33\% higher than if the index was 0 (Weeks 14-16). By Week 21 (average index = 6.4), robbery rates were predicted to be 14\% higher relative to Weeks 14-16.

• Homicide, Auto Theft, and Burglary: Weekly changes in the openness index were not significantly associated with changes in their weekly offense rates.

Speculation on Nuances of Null Findings

Possible reasons for the lack of a significant association for homicide, auto theft, and burglary with the openness index include:

1. Enforcement and Compliance: The actual degree and success of local law enforcement efforts and public compliance with lockdown measures might mediate local policy effects, and data on compliance is difficult to obtain.

2. Statistical Power: Limited variability in weekly changes of the openness index over time and across cities could reduce statistical power.

3. Sensitivity to Change: Offenders of homicide, auto theft, and burglary might not be sensitive to modest changes in openness but only respond to more drastic shifts between open and mostly closed states.

Biweekly Lockdown Period Findings (Equation (3))

• Robbery and Larceny: Demonstrated significant reductions across most or all biweekly periods into the lockdown (Figures 3 and 4). Larceny rates dropped more quickly at the lockdown's start, while robbery rates took approximately a month to reach their lowest point.

• Homicide: Displayed a more complex pattern (Figure 5). At the onset of lockdowns, homicide rates were approximately 20\% lower than usual seasonal levels (p < 0.05), with this decrease possibly persisting for another two weeks. Following this, rates appeared to rebound to near usual seasonal levels for about four weeks. In the 9^{th} and 10^{th} weeks after lockdown, homicide rates were approximately 15\% higher than usual seasonal levels, though with wide confidence intervals.

• Auto Theft and Burglary: Most biweekly period estimates were not statistically significant and remained close to typical seasonal levels (Figures 6 and 7). However, abrupt and significant increases were observed for both crime types in the final 9^{th}-10^{th} biweekly period (p < 0.05 for auto theft and p < 0.01 for burglary). These late increases might be linked to the early response following George Floyd's murder (the second week of this period) or a threshold effect of increasing city openness, among other factors.

Overall, strong evidence suggests larceny and robbery rates decreased during lockdowns and were sensitive to lockdown severity and timing, increasing with greater openness. Homicide rates showed a short-term reduction, then fluctuation. Burglary and auto theft rates remained near typical levels until the very end of the lockdown period, with abrupt increases.

Analyses for All of 2020 (Seasonal Crime Rates)

These analyses, based on Equation (4), compared 2020 crime trends across four defined periods (Pre-Pandemic, Lockdown, Summer Protests, End of Year) to seasonal patterns in 2018 and 2019. Significant and often substantial changes were found, with mixed directions and magnitudes.

Homicide Trends (Figure 8)

• Sustained Increase: Homicide rates in 2020 were significantly higher throughout the year compared to prior years (p < 0.05 for all periods).

• Pre-Pandemic Increase: In January and February 2020, homicide rates increased by nearly 25\% relative to 2018 and 2019 levels.

• Later Period Exacerbation: Increases were even larger in the summer protest and end-of-year periods, reaching approximately 50\% (p < 0.001 for both). While these point estimates are larger, the statistical difference between the pre-pandemic estimate and the later period estimates was not significant at the p < 0.05 level (but was significant at p < 0.10). This finding challenges explanations that solely attribute the homicide surge to events following George Floyd's killing, suggesting the increase preceded these events.

• Lockdown to Summer Protest Period: A substantial and strongly significant increase (approximately 25\% higher, p < 0.001) was observed in homicide rates from the lockdown period to the summer protest period.

• Aggravated Assaults (Supplementary): Aggravated assault rates showed a partial mirroring of homicide trends, increasing significantly in the summer protest and end-of-year periods (p < 0.01), but showing no significant changes in the first two periods.

Auto Theft Trends (Figure 9)

• Early 2020 Stability: Auto theft rates in the pre-pandemic and lockdown periods of 2020 were similar to those in 2018 and 2019.

• Significant Later Increases: In the summer protest and end-of-year periods, auto theft rates significantly increased by over 20\% compared to the same months in prior years (p < 0.001 for both).

Burglary Trends (Figure 10)

• General Decline: Point estimates for burglary rates were negative across all four periods, suggesting a general decline.

• Statistical Significance: Only the pre-pandemic estimate was statistically significant (p < 0.05). The lockdown period estimate was significant at the p < 0.10 level. This suggests burglary rates may have experienced a modest 5-10\% decline in 2020, potentially for reasons unrelated to the pandemic.

Robbery Trends (Figure 11)

• Pre-Pandemic Stability: Robbery rates remained virtually unchanged in the pre-pandemic period compared to 2018-2019 levels.

• Sharp Pandemic Declines: Robbery rates declined sharply across all three pandemic periods (lockdown, summer protests, and end of year), with all estimates significant at p < 0.01. The declines were substantial, 14\% or more, with the most significant decrease (20\%) occurring during the lockdown period, coinciding with the lowest pedestrian traffic.

Larceny Trends (Figure 12)

• Pre-Pandemic Increase: Larceny rates in the pre-pandemic period were significantly up relative to 2018-2019 levels (p < 0.05). This makes the subsequent declines even more pronounced.

• Significant Pandemic Declines: Similar to robbery, larceny rates declined by 17\% or more during the three pandemic periods (p < 0.01 for all), relative to prior levels.

Discussion and Conclusions

Summary of Overall Crime Rate Changes

The study found substantial and often mixed changes in crime rates from March through December 2020, compared to the same periods in 2018 and 2019.

• Homicide: Significant increases were observed throughout 2020, including the pre-pandemic period.

• Auto Theft: Significant increases occurred in the latter half of 2020 (summer protest and end-of-year periods), also significant when compared to the 2020 pre-pandemic period.

• Robbery and Larceny: Experienced sustained declines throughout the pandemic periods.

• Burglary: Showed a consistent pattern of negative point estimates across all pandemic periods, though only the pre-pandemic period's decline was statistically significant at the 5\% level, with the lockdown period estimate being significant at the 10\% level.

Impact of Lockdowns

• Robbery and Larceny: Strong evidence shows that the severity of lockdowns reduced these offense rates, which were sensitive to lockdown strictness and timing, with rates increasing as cities became more open.

• Homicide, Auto Theft, and Burglary: The openness index did not show a direct impact of lockdown severity on these crime types.

• Biweekly Analysis: Further supported significant declines for larceny and robbery during lockdowns, with suggestive short-term fluctuations for other crime types.

Notably, crime types that typically move in tandem (e.g., homicide and robbery) diverged during the pandemic, with homicide and auto thefts increasing while larceny and robbery declined. This complex pattern presents a challenge for future research.

Theoretical Implications (Opportunity Theory)

The authors note that while their aim was to document changes rather than ascertain causes, opportunity theory offers a post hoc and partial explanation for some of the observed patterns.

Crimes Explained by Opportunity

• Robbery decline: Possibly due to reduced pedestrian traffic, especially at night, limiting potential targets.

• Larceny decline: Attributable to business closures, reduced foot traffic in retail areas, and possibly increased security measures (e.g., for mask enforcement).

• Auto theft increase: May have resulted from more vehicles left idle on streets coupled with fewer people (including police presence) outside, increasing opportunities for surreptitious theft.

Crimes Not Easily Explained

• Homicide: Opportunity theory struggles to explain the homicide increase. While increased time at home and substance use might contribute to domestic homicides, media reports don't support this. Additionally, reduced social gatherings in bars and less loitering by young men should theoretically reduce non-domestic homicides, contrary to the observed increases.

• Burglary: More time spent at home by residents would reduce opportunities for domestic burglaries (consistent with observed declines). However, closed businesses might increase opportunities for non-residential burglaries. These opposing forces might have balanced out, leading to no systematically detectable change, though this reasoning is presented cautiously as a