Effects of COVID-19 shutdowns on domestic violence in US cities.

We empirically investigate the impact of COVID-19 shutdowns on domestic violence using incident-level data on both domestic-related calls for service and crime reports of domestic violence assaults from the 18 major US police departments for which both types of records are available. Although we confirm prior reports of an increase in domestic calls for service at the start of the pandemic, we find that the increase preceded mandatory shutdowns, and there was an incremental decline following the government imposition of restrictions. We also find no evidence that domestic violence crimes increased. Rather, police reports of domestic violence assaults declined significantly during the initial shutdown period. There was no significant change in intimate partner homicides during shutdown months and victimization survey reports of intimate partner violence were lower. Our results fail to support claims that shutdowns increased domestic violence and suggest caution before drawing inference or basing policy solely on data from calls to police.

Introduction

From the outset of the COVID-19 pandemic, news coverage and policymaking have prominently featured concerns that government-mandated restrictions on economic activity and personal mobility might increase domestic violence (DV).1 This attention to DV is well-motivated because of its high social and economic costs (Garcia-Moreno & Watts, 2011) and because stress, economic disruption and social isolation are established predictors of DV (Berg & Tertilt, 2012; Bright et al., 2020). Nevertheless, shutdowns were unprecedented, and they could reduce DV in some households by lowering exposure to DV triggers such as infidelity and alcohol consumption outside the home (Nemeth et al., 2012), limiting contact between non-cohabiting and former couples (Ivandic et al., 2020), and even strengthening some relationships (Sachser et al., 2021). Furthermore, increased public and private funding to support DV victims and survivors, together with increased media attention devoted to DV, around the time shutdowns were imposed (Bright et al., 2020) could have reduced repeated violence and escalation. Federal stimulus payments enacted in response to the pandemic also significantly lowered poverty rates, which may have reduced DV (Wheaton et al., 2021; Erten et al., 2022). As a result of these opposing factors, the effects of shutdowns on overall DV levels were theoretically ambiguous and likely to vary across populations.

Determining the overall impact of shutdowns on DV requires careful empirical analysis, but results needed to be produced and disseminated rapidly to contribute to ongoing debates about pandemic policy (Reingle Gonzalez et al., 2020). Because of this urgency, researchers from a variety of disciplines relied on readily available administrative data to assess DV incidence. In the US, the main source has been real-time data published by individual police departments that provide one or two key measures of DV: calls for service (911 calls or radio dispatches) and criminal incident reports. Despite the initial set of papers yielding mixed results, the claim that shutdowns increase DV incidence has been presented as an established fact in media coverage and in political debates about pandemic restrictions (e.g., Biggs, 2020).

This paper is motivated by the observation that the empirical studies finding increases in DV in US cities examine DV service calls as their exclusive (Leslie & Wilson, 2020; McCrary & Sanga, 2021; Nix & Richards, 2021) or primary (Hsu & Henke, 2021) outcome measure. While service call volume measures demand for police resources, it is a limited proxy for rates of specific crimes because “callers can be mistaken in what they report” (Ashby, 2020b; p. 1061) and “not all domestic violence calls are for activities that constitute crimes” (Klein, 2009; p. 1). Papers that examine DV crime rates are more likely to find decreases in DV, particularly when they account for seasonal variation using data from prior years (Abrams, 2021; Ashby, 2020a; Bullinger et al., 2021; Miller et al., 2020).2 However, because studies of the different police outcomes have differed in their geographic coverage, it is unclear if the divergence in estimates comes from systematic differences between the two types of police data or from geographic variation in the impact of shutdowns.

We address this important question by studying the 18 large, urban US police departments, serving over 14 million people, for which we were able to obtain incident-level data on both DV calls for service and DV assault crimes. We empirically estimate the impacts of shutdowns by comparing the differences in changes in each of our outcomes between the initial pandemic shutdown period in 2020 and the earlier part of the year and the changes between the same time periods in 2019. We find a decrease in DV assaults but an increase in DV calls during shutdowns. We also estimate models that account for the finding in the prior literature of an increase in DV calls during the period of voluntarily lower mobility that followed the nationwide emergency declaration but preceded mandated shutdowns (e.g., McCrary & Sanga, 2021). When we estimate models that also control for the pre-shutdown emergency period, we find both DV assault crimes and DV calls are lower during shutdowns, relative to the immediately preceding period. We also find no evidence that intimate partner homicides or reports of intimate partner violence in the National Crime Victimization Survey increased during shutdown months; suicides, which have been linked to DV (Stevenson & Wolfers, 2006), were lower. These results fail to provide empirical support for claims that DV increased because of pandemic shutdowns, and instead suggest that violence may have decreased.

Data and methods

Our sample includes the 18 large US police departments that provided incident-level, real-time data on both DV calls for service and DV assault crimes. We started our search with the full list of 107 local police departments that served a population of at least 250,000 people, according to the 2018 Law Enforcement Officers Killed and Assaulted (LEOKA) Data Collection (Kaplan, 2020). We checked city and county open data archives as well as departmental webpages to identify police departments from this list that published incident-level calls for service and crime data with DV information in real-time.3 Only 7 of them published this information online. To expand our sample, we identified 30 additional large departments that have published data on either DV calls or DV crimes in real-time since 2019 and submitted public records requests for data on the missing measure. This approach allowed us to identify departments that were most likely to maintain the relevant data and respond to our request. Of the 30 requests submitted, we received responses and usable data from 11 departments by July 14, 2021.4 The 18 police departments in our final sample, listed in Table 1 , include 2 of the 3 largest in the country, and collectively serve over 14 million people. Data sources, including date accessed, for all police data are listed in Table A1 and variables and keywords to identify DV are listed in Table A2.

Table 1

Sample of Municipal Police Departments

	Rank by Pop. Served	Population Served	Initial Shutdown	Initial Reopening
Los Angeles, CA	2	4,029,741	3/20	5/29
Chicago, IL	3	2,719,151	3/21	6/3
Fort Worth, TX	21	893,756	3/25	5/8
San Francisco, CA	23	889,282	3/17	9/1
Memphis, TN	38	652,226	3/24	5/6
Tucson, AZ	43	537,392	3/31	5/8
Mesa, AZ	48	504,873	3/31	5/8
Kansas City, MO	50	493,115	3/24	5/6
Virginia Beach, VA	58	451,001	3/30	5/15
Minneapolis, MN	60	428,261	3/28	6/1
New Orleans, LA	65	396,374	3/20	5/16
Chesterfield Co., VA	72	346,692	3/30	5/15
St. Paul, MN	82	309,756	3/28	6/1
St. Louis, MO	84	306,875	3/23	5/18
Cincinnati, OH	86	301,952	3/24	5/15
Orlando, FL	92	286,679	3/25	5/11
Durham, NC	96	273,759	3/26	6/1
Chandler, AZ	105	255,986	3/31	5/8
		14,076,871

Notes: This table lists the police departments included in the main estimation sample, which consists of all departments serving a population of 250,000 or more and providing real-time data on domestic-related calls for service and assault crimes. Sources and details can be found in the supplementary materials.

The initial shutdowns in these 18 cities started between March 17 and March 31, 2020.5 We focus on the impact of the initial shutdowns to avoid complications related to re-opening and repeated closures and therefore end the sample period on May 6, 2020, the earliest reopening date in our sample. This also allows us to compare our results to the existing literature that mainly investigates DV outcomes in the first few months of the pandemic. We separately analyze each of our two main outcomes, DV assault crimes and DV service calls. We use incident-level police data to calculate daily counts of DV calls and DV crimes and we compute rates using the LEOKA population served by the department (Kaplan, 2020). We focus on DV assaults because they are the most commonly reported DV crime category across police departments. Our measure of DV assaults is based on police criminal incident reports, and not on arrests or convictions.

The raw data clearly show opposing trends in these two outcomes during the pandemic, foreshadowing our main results. From January through mid-March, DV assault crimes in 2020 followed a similar seasonal pattern to those crimes in 2019 (Figure 1 , Panel A). After that, 2020 DV assaults decreased slightly relative to 2019 levels, as cities started to mandate shutdowns, leading to a sizable relative decline in April and early May. In contrast, DV service calls in 2020 tracked those in 2019 in January and February, but diverged in the month of March, when 2020 calls increased at a higher rate.

Fig. 1

Trends in DV assault crimes and service calls. This figure depicts trends for (A) police reports of DV assault crimes and (B) DV service calls to police. Daily trends were calculated as the 7-day moving average of daily records, aggregated across cities, per 100,000 total population served. The dashed vertical line on March 14 indicates the date after the nationwide emergency declaration and the solid vertical lines indicate the dates of city shutdowns. The trends indicate a relative decrease in DV assaults during the shutdowns compared to 2019 (A), and a relative increase in DV service calls (B).

In addition to showing the differential trends, Figure 1 also illustrates the disparity in rates between the two outcomes: DV calls are 4 times more frequent than DV assaults.6 This disparity highlights the fact that most DV calls do not lead to DV criminal incident reports, making it important not to rely on DV calls alone for tracking incidence. Furthermore, it would be inaccurate to assume that DV calls include all DV assault crimes, as not all DV assault crimes originate from such calls.7

The trends in Figure 1 also illustrate the key strategy underlying our empirical approach. To estimate the impact of shutdowns on our DV outcomes of interest, we need to compute a counterfactual for what DV levels would have been in the absence of shutdowns. We accomplish this by exploiting data from 2019 to account for seasonal variation in DV within the year and from the pre-shutdown months to account for inter-year variation in DV levels.8 Our models formally compare differences in DV outcomes during the pandemic shutdown period relative to the earlier months of 2020 to the differences between the same time periods in 2019.

Our first specification takes this form:

is the domestic violence outcome of interest, measured at the city-day level and scaled to city population. We include a vector of city and year fixed effects and account for seasonal and within-week variation with month and day of week fixed effects. The error term captures random city-day level independent shocks that affect outcomes. We report robust standard errors that allow for heteroskedasticity. Because larger cities are less subject to random fluctuations in their daily crime rates, we follow the usual practice and weight all regressions by city population.

In our first model, is an indicator variable that takes a value of 1 if a shutdown is effective in city i on day t. The coefficient captures the difference-in-differences estimate described above. All US cities were clearly affected by the pandemic and experienced shutdowns at around the same time. We therefore rely on 2019 to provide a “control” year that was unaffected by the pandemic and define the “pre” and “post” periods based on calendar date (month and day) within the year. While it is possible to compare locations with larger and smaller drops in voluntary mobility, it is not obvious that such measures capture meaningful variation in the severity of the pandemic (i.e., that places with smaller drops in mobility, possibly because they contain more essential workers, were somehow less affected by the stress, health impact, or other hardships caused by the pandemic)

We also estimate a second model that aims to match the prior literature focusing on the effect of the nationwide emergency declaration that preceded the mandatory shutdowns

In this model, the variable is replaced with a indicator that takes a value of 1 starting on March 14, 2020, the day after the nationwide emergency declaration. The coefficient is therefore a difference-in-differences estimate of the average change in outcomes between the period after the emergency declaration between 2020 and 2019 compared to the average change in outcomes that occurred between these years in the period between January 1 and March 13. We report results from this model for purposes of comparison but note that it is unable to address our policy question of interest, the impact of mandated shutdowns.

Finally, we report estimates from a model that includes both explanatory variables from the prior two models, which allows us to distinguish the effects of city-specific mandatory shutdowns from those attributable to the earlier nationwide emergency declaration:

In this model, represents the incremental change in DV during the shutdown period, in addition to the change caused by the voluntary reductions in mobility that followed the emergency declaration, which is estimated by .9 As in Model 2, these changes are between 2020 and 2019 in comparison to the average changes between these years that occurred between the beginning of the year and March 13.

Another parameter of interest is the average change in DV during the shutdown period between 2020 and 2019 relative to the period between January 1 and March 13 (i.e., the period before the emergency declaration in 2020) between the two years. That parameter is calculated by summing the and coefficients.

Results

We present our three main empirical findings in Table 2 . The first finding is that shutdowns are associated with a significant decrease in DV assault crimes across all models. While the period after the emergency declaration has lower DV assault crimes when combined with the shutdown period (Model 2), there was no measurable change during the pre-shutdown emergency period (Model 3). However, there was a large and significant drop in DV assaults during shutdowns, whether the comparison period is limited to the time before or after the emergency declaration (Model 3) or if it includes all pre-shutdown dates (Model 1). The magnitude of this drop is consistently around 0.19 per 100,000 population (p < 0.01), corresponding to 10.0% of the 2019 baseline.10

Table 2

Main Estimation Results

Panel A: Domestic Assault Crimes
	Using City Shutdowns	Using Emergency Declaration	Using Both
City Shutdowns (Shutdown Start - May 5)	-0.188*** [0.033]		-0.191*** [0.049]
Emergency Declaration (March 14 - May 5)		-0.142*** [0.030]	0.004 [0.045]
Shutdown Relative to pre-Emergency Declaration			-0.187*** [0.034]
Outcome variable 2019 mean	1.699

This table presents the results from estimating equations 1-3 in the paper using city-day level data, weighted by city population. Outcomes are rates of DV assault crimes (Panel A) or service calls (Panel B) per 100,000 population. Robust standard errors are shown in brackets. *** p < 0.01, ** p < 0.05, * p < 0.1.

Our second finding is an increase in the number of DV calls during the shutdown, relative to the start of the year. The increase is 0.178 per 100,000 population (p < 0.10) in Model 1, which includes the emergency period as part of the baseline. The emergency period from March 14 forward was itself associated with more DV calls: the estimate in Model 2 is a 0.394 increase (p < 0.01). When that period is excluded from the baseline in Model 3, the estimated increase in calls during the shutdown is 0.295 (p<0.001). The opposing direction of these first two findings indicates that the two measures are not interchangeable, as might have been imagined, and that pandemic shutdowns had differential effects on police measures of DV crimes and calls. By analyzing cities with police data on both DV call and crime, we reject the possibility that heterogeneity across city samples in the source of the conflicting results.11

The regression results from estimating Model 3 on DV calls data also show our third main finding: that DV calls to police increased after the emergency declaration but prior to the enforcement of mandatory shutdowns and should not be attributed to shutdowns themselves. The increase in daily DV calls during the pre-shutdown emergency period is 0.721 per 100,000 (p<0.01; 9.2% of baseline; Table 2).12 The incremental effect of the shutdown, relative to this elevated rate, is actually a decline of 0.426 DV calls per 100,000 (p<0.01). Despite this decline relative to the period immediately before shutdowns, DV calls were still elevated during shutdowns, compared to pre-pandemic period from January 1 to March 13 (0.295, p<0.01). This again confirms that calls for service and DV crimes show opposing trends during the shutdown period.

Our finding of significantly lower DV crime rates during shutdowns persists across multiple alternative sample definitions, including adding data from 2018 to expand the comparison group, omitting one city at a time from the sample (see Appendix B and Figures A1 and A2), as well as after excluding both Chicago and Los Angeles, the only two departments in our sample serving populations of over a million people. However, the increase in DV calls for service is less robust across samples. We found a significant relative decline in DV calls during shutdowns in Model 1 for the 16-city sample (Table 3 ). Estimates from that sample confirm the significant increase in DV calls following the national emergency declaration, as well as the significant relative decline following mandatory shutdowns, but the latter drop is sufficient to fully offset the prior increase.

Table 3

Excluding Chicago and Los Angeles

Panel A: Domestic Assault Crimes
	Using City Shutdowns		Using Emergency Declaration	Using Both
City Shutdowns (Shutdown Start - May 5)	-0.144*** [0.042]			-0.164*** [0.063]
Emergency Declaration (March 14 - May 5)			-0.091** [0.038]	0.024 [0.057]
Shutdown Relative to Pre-Emergency Declaration				-0.140*** [0.043]
Outcome variable 2019 mean	1.687

This table presents the results from estimating equations 1-3 in the paper using city-day level data, weighted by city population, and excluding Chicago and Los Angeles from the sample. Outcomes are rates of DV assault crimes (Panel A) or service calls (Panel B) per 100,000 population. Robust standard errors are shown in brackets. *** p < 0.01, ** p < 0.05, * p < 0.1.

We also confirmed the lack of city-specific pre-trends in outcomes in the 4 weeks leading up to the city shutdowns that are not accounted for by the initial emergency declaration. The only significant estimate in Table A5 is an increase in DV calls in the week before shutdowns in Model 1 (column 3), coinciding with the emergency declaration. This estimate becomes negative and insignificant after accounting for the emergency declaration in Model 3.

Table 4 reports separate estimates for the impact of shutdowns on simple DV assaults and on aggravated DV assaults (typically involving weapons or serious injuries), with significant declines present for both types of DV assaults. Consistent with simple DV assaults being far more common in the data, the magnitude of the estimated effect of shutdowns is much larger for simple DV assaults. Relative to the pre-emergency declaration period, there were 0.145 fewer simple DV assaults per 100,000 population per day during the initial shutdowns (Model 3, Panel A) and 0.042 fewer aggravated DV assaults (Model 3, Panel B). However, when scaled in proportion to the 2019 mean value of their outcome measures, the estimated reductions are more similar in size: 10.3% for simple DV assaults and 14.4% for aggravated DV assaults.

Table 4

Effects of Pandemic Shutdowns on Simple and Aggravated DV Assaults

Panel A: Domestic Simple Assault Crimes
	Using City Shutdowns	Using Emergency Declaration	Using Both
City Shutdowns (Shutdown Start - May 5)	-0.148*** [0.029]		-0.163*** [0.043]
Emergency Declaration (March 14 - May 5)		-0.107*** [0.027]	0.017 [0.039]
Shutdown Relative to pre-Emergency Declaration			-0.145*** [0.030]
Outcome variable 2019 mean	1.406

This table presents the results from estimating equations 1-3 in the paper using city-day level data, weighted by city population. Outcomes are rates of DV simple assault crimes (Panel A) or DV aggravated assault crimes (Panel B) per 100,000 population. Robust standard errors are shown in brackets. *** p < 0.01, ** p < 0.05, * p < 0.1.

Despite their value in providing a rapid view of the ongoing pandemic, it is important to note additional limitations of the real-time police data used in this analysis. Because these data rely on reporting by individual police departments, they are limited in scale and the findings may not generalize to other cities. Even their internal validity can be questioned because real-time crime data differ from official Justice Department reporting efforts in that there are no quality standards or requirements for data inclusion or coding. We attempted to validate the real-time crime data by comparing them to 2019 National Incident-Based Reporting System (NIBRS) data (Kaplan, 2021a) and found variation in match rates for DV crimes across cities (Figure A4). Quality concerns are even more significant for calls data because there are no federal data products related to police service calls.

The main findings of this paper call into question claims of conclusive empirical support from real-time police data that pandemic shutdowns increased DV. Although calls for service were higher, they increased before shutdowns were mandated. Furthermore, the increase in DV calls was clearly not matched by an increase in DV crimes, which declined substantially during shutdowns.

Interpretation

How can the opposite effects of shutdowns on DV calls and crimes be reconciled? We start with a simple framework in which we model the rate of DV calls to police as:where denotes the daily population share that experiences a DV incident. The rate at which DV incidents are reported to police is reflected in , the reporting rate for “true” DV incidents, which is defined as the average number of DV calls to police per DV incident. Although it is possible that some incidents produce multiple calls to police, the high share of DV crimes never reported to police suggests that is likely smaller than one. The second term in the equation captures DV calls that are not connected to corresponding DV criminal incidents. These could be other types of crimes, which are not domestic crimes, or incidents that are not crimes at all like verbal disputes that may involve intimate partners. A fraction of the population does not experience DV and produces “false” DV calls at a rate of . The rate of DV crimes recorded by police is:where , and are defined as in the equation for calls. The additional parameters are used to capture the fact that only a fraction of DV calls are recorded as criminal incidents. Because more information is available to police after the initial response, we expect that , and that crimes are more closely related to incidence than calls. The third term in the equation reflects the fact that some DV incidents that are not reported to police as DV calls are nonetheless reported as non-DV calls, at a rate of . A fraction of those calls will be recorded as DV incidents in the crime data.

Within this framework, how can we think about the opposite effects of shutdowns on DV calls and crimes? One possibility is that crimes decreased but reporting rates increased because of additional DV calls for non-criminal incidents or non-DV crimes . These calls could have increased during shutdowns from third-party reporters, such as neighbors, who had incomplete information about events in the home, and were more likely to be home themselves and possibly more aware of the issue because of increased media coverage of DV risks. Support for this mechanism is found in Greater London police data where the increase in DV calls during the pandemic was attributed to third-party callers (Ivandic et al., 2020).

Although we are not able to directly identify third-party callers in data from US departments, we do find indirect support for this mechanism in the 9 departments for which we can categorize DV calls based on severity.13 More severe DV calls indicate reported physical violence, while less severe calls indicate a domestic disturbance or verbal dispute. Panel A of Table 5 shows that our main estimation result of a decrease in DV assaults is present on this sample, including decreases for both simple and aggravated assaults. However, the new results in Panel B show that the increases in DV calls in both Models 1 and 3 were driven by an increase in less severe DV calls for verbal disputes. This supports an interpretation that some of the additional calls were related to non-criminal incidents, including from third-party callers with limited information. Such calls would correspond to an increase in in our model.

Table 5

Effects of Pandemic Shutdowns on Verbal and Physical DV Calls for Service

Panel A: Domestic Assault Crimes
	All		Simple Assault		Aggravated Assault
	(1)	(2)	(3)	(4)	(5)	(6)
City Shutdowns (Shutdown Start - May 5)	-0.198*** [0.039]	-0.218*** [0.061]	-0.157*** [0.034]	-0.174*** [0.052]	-0.042*** [0.013]	-0.044* [0.025]
Emergency Declaration (March 14 - May 5)		0.023 [0.057]		0.021 [0.048]		0.002 [0.024]
Shutdown Relative to pre-Emergency Declaration		-0.195*** [0.040]		-0.154*** [0.035]		-0.041*** [0.013]
Outcome variable 2019 mean	1.637		1.350		0.288

This table presents estimated effects of pandemic shutdowns from our two main models. Estimates from Model 1 are presented in odd-numbered columns, while those from Model 3 are in even-numbered columns. Regressions are estimated using city-day level data, weighted by city population, on a subsample of 9 cities for which we have information on the severity of DV 911 calls. The 9 cities in this subsample are: Chesterfield County, Chicago, Cincinnati, Kansas City, Los Angeles, Minneapolis, New Orleans, San Francisco, and St. Paul. Outcomes are rates of DV assault crimes (Panel A) or service calls (Panel B) per 100,000 population. Regressions include city, year, month, and day of week fixed effects. Robust standard errors are shown in brackets. *** p<0.01, ** p<0.05, * p<0.1.

One reason why third-party calls might have increased following the emergency declaration is that neighbors were more likely to be home and to call police about noise disturbances, whether from domestic disputes or from other causes. We examine this in Table 6 on our sample of 16 departments with this information.14 The first four columns replicate the main results on that sample, while the last two examine 911 calls for non-DV nuisance issues, including noise, general disturbances, and parties. We find that these calls were significantly elevated during shutdowns relative to periods before the shutdown (column 5; by 0.42 calls per 100,000 population) and before the emergency declaration (column 6; by 0.47 calls per 100,000 population, or 4.3% of the 2019 mean rate). This is consistent with the interpretation that the increase in third-party reporting of DV disputes comes in part from neighbors spending more time at home and being more likely to complain to police about noise. Because some, but not all, of these additional calls correspond to DV crimes, that effect would be captured by increases in both and in our model.

Table 6

Effects of Pandemic Shutdowns on Calls for Noise and General Disturbances

Sub-Sample of Cities with General Disturbance/Noise Calls
	Domestic Assault Crimes		Domestic Calls		Noise Calls
	(1)	(2)	(3)	(4)	(5)	(6)
City Shutdowns (Shutdown Start - May 5)	-0.189*** [0.034]	-0.194*** [0.050]	0.203** [0.098]	-0.368** [0.149]	0.419** [0.211]	0.168 [0.281]
Emergency Declaration (March 14 - May 5)		0.006 [0.046]		0.679*** [0.138]		0.298 [0.231]
Shutdown Relative to pre-Emergency Declaration		-0.188*** [0.035]		0.311*** [0.101]		0.466** [0.215]
Outcome variable 2019 mean	1.669		7.601		10.86
Observations	4,032	4,032	4,032	4,032	4,032	4,032

This table presents estimated effects of pandemic shutdowns from our two main models. Estimates from Model 1 are presented in odd-numbered columns, while those from Model 3 are in even-numbered columns. Regressions are estimated using city-day level data, weighted by city population, on the subsample of 16 cities (missing St. Louis and Tucson) for which we have information on the calls for noise complaints or general disturbances. Outcomes are rates of DV assault crimes, DV service calls, and non-DV noise or general nuisance calls per 100,000 population. Regressions include city, year, month, and day of week fixed effects. Robust standard errors are shown in brackets. *** p<0.01, ** p<0.05, * p<0.1.

Our data are also consistent with the possibility that increased reporting rates only partially explain the increase in DV calls to police in the period following the emergency declaration, and that some of the increase is from elevated rates of non-criminal verbal and emotional domestic abuse during that period.15

Given the relative timing of the effects on DV calls and crimes in US, it is also possible that some of the increased calls to police for domestic incidents before the beginning of the shutdown, in combination with the additional financial resources and public attention devoted to the issue of DV, had a deterrent effect, preventing escalation and lowering crime rates during the shutdown (Miller & Segal, 2019). This deterrence could apply to increased reporting of criminal DV incidents as well as non-criminal verbal disputes that might have escalated over time without police intervention.

An alternative possible reconciliation is that the additional DV calls reflected an increase in DV crimes (higher ), but that fewer crimes were recorded because of reductions in policing intensity for DV cases during the shutdown (lower values of one or more of the parameters , and ). This interpretation is favored in Bullinger et al. (2021), which characterizes the divergence between DV calls and recorded DV crimes in Chicago as reflecting “substantial underfiling [by police] of official incident reports for domestic crimes” (p. 267) and studies the ratio of reported DV crimes to DV calls for service as a measure of the extent to which police officers “avoid filing a domestic violence report” (p. 269).

The departments in our sample (including Chicago) all have written operational procedures for handling domestic disputes and designated personnel to address DV (US DOJ, 2007, 2013). Although police departments altered procedures to reduce officer exposure to and community spread of COVID-19, they have not relaxed recording requirements for DV, and in public statements assert that they continue to prioritize those cases (Police Executive Research Forum, 2020).16 The initial shutdown period also saw dramatic reductions in non-DV violent crimes (e.g., Miller et al., 2020; Abrams, 2021; Bullinger et al., 2021), freeing time and resources to address DV. Furthermore, the explanation that failures of police record-keeping is the source of the drop in crimes seems more likely for less serious crimes than for the assaults that we study. The significant decrease in aggravated assaults is informative on this point because those crimes are probably the least likely to be neglected in official reports from police responding to 911 calls. Furthermore, in Los Angeles, where crimes can be linked to arrest records, there is no evidence of less intensive policing of DV in the form of fewer arrests per crime during the initial shutdown (Miller et al., 2020).

Although we lack the data to directly examine the possibility that failures of police drive the observed crime reductions,17 we do observe an objective measure of police responsiveness to DV calls in 5 of our cities: the time between the 911 call and police arrival at the scene.18 Table 7 shows estimates from our two main models on this sub-sample. We confirm the decline in DV assaults in columns 1 and 2, but find no evidence that police were slower to respond to DV calls during the pandemic. Rather, columns 5 and 6 show significantly faster police response times following the emergency declaration and during the subsequent shutdowns. Relative to 2019 response times, police responded about 17% faster to DV calls during shutdowns.

Table 7

Effects of Pandemic Shutdowns on Police Response Time for DV Calls

Sub-Sample of Cities with Police Response Time
	Domestic Assault Crimes		Domestic Calls		Police Response Time for DV Calls (seconds)
	(1)	(2)	(3)	(4)	(5)	(6)
City Shutdowns (Shutdown Start - May 5)	-0.244*** [0.070]	-0.220** [0.099]	-0.026 [0.226]	-0.787** [0.317]	-116.562*** [22.738]	-60.939** [30.413]
Emergency Declaration (March 14 - May 5)		-0.030 [0.088]		0.951*** [0.271]		-69.561*** [26.211]
Shutdown Relative to pre-Emergency Declaration		-0.250*** [0.072]		0.164 [0.231]		-130.500*** [23.431]
Outcome variable 2019 mean	1.275		8.878		771.2
Observations	1,260	1,260	1,260	1,260	1,260	1,260

This table presents estimated effects of pandemic shutdowns from our two main models. Estimates from Model 1 are presented in odd-numbered columns, while those from Model 3 are in even-numbered columns. Regressions are estimated using city-day level data, weighted by city population, on a subsample of 5 cities for which we have information on police response times for DV calls. The 5 cities in this subsample are: Chandler, Cincinnati, Mesa, St. Louis, and Virginia Beach. Outcomes are rates of DV assault crimes (Column 1), DV service calls (Column 2) per 100,000 population, and average police response time in seconds for DV calls (Column 3). Regressions include city, year, month, and day of week fixed effects. Robust standard errors are shown in brackets. *** p<0.01, ** p<0.05, * p<0.1.

Evidence from federal data on deaths and crime victimization

Finally, to shed further light on the issue of police failing to respond fully to serious DV during shutdowns, we draw on additional federal data sources that were not available in real time.

We start by studying changes in the extreme outcome of homicide that is universally reported to police, thereby avoiding the interpretation challenge for other police data coming from the fact that DV reporting rates by victims and witnesses respond to external factors (Miller & Segal, 2019) and may have been affected by pandemic shutdowns. We examine newly released Supplementary Homicide Reports (SHR) from the Uniform Crime Reporting system (Kaplan, 2021b), with incident-level data that identifies the reporting police department, month of occurrence, and relationship between victim and offender. We compute difference-in-differences estimates for the impact of shutdowns using a simplified version of main estimation approach, by comparing the change in outcomes to the shutdown period in 2020 (April and May) from the pre-pandemic data from the start of 2020 (January and February) to the change of the same time periods in the prior year. Our sample includes 17 of our 18 cities because departments in Florida did not participate in the SHR in our time period.

Panel A of Figure 2 plots the IPH rates (per million population) for each of the four time periods and shows no evidence of a relative increase during shutdowns: the implied difference-in-differences impact of shutdowns is zero. Using monthly police department level data, we formally estimated the corresponding regression model comparing IPH rates in shutdown months (April and May in 2020) to January and February in 2020 and the same four months in 2019, with agency, year and month fixed effects, and confirm the economic and statistical insignificance of the estimate (<0.0001, s.e. of 0.342).19

Fig. 2

Effects of pandemic shutdowns on intimate partner homicide (IPH), suicide, and intimate partner violence (IPV) rates. This figure shows death and nonfatal crime rates from pre-pandemic and pandemic shutdown months in 2020 and from the same months in 2019. The IPH data are available at the monthly level for 17 of our 18 police departments (Orlando is missing). We show in (A) IPH rates per million population for January-February and April-May (omitting March, the month of the emergency declaration and start of shutdowns). Preliminary suicide data are at the state-quarter level, so we show in (B) the rates per 100,000 population for January-March and April-June for the 12 states in which our sample cities are located. Data on nonfatal IPV are nationwide and monthly, so (C) depicts rates per 100,000 population aged 12 or older for the same time periods as (A).

We also examined the more frequent outcome of suicide, which has been linked to DV in prior research (e.g., Stevenson & Wolfers, 2006). Our data on suicide rates are CDC estimates currently available at the state-quarter level (Ahmad & Cisewski, 2021). Panel B of Figure 2 shows the comparison between the first and second quarters of 2020 and 2019 for the 12 states containing any of our 18 cities. The relative change during the shutdown quarter is a reduction of 1.53 suicides per 100,000 population. We confirm the statistical significance of the drop in the corresponding regression model using state-by-quarter data, with state, year and quarter fixed effects, and weighting observations by population, resulting in a standard error of 0.433 (p-value = 0.001).

Finally, we examine data from the National Crime Victimization Survey (NCVS; United States. Bureau of Justice Statistics, 2021) to assess the possibility that nonfatal DV in the population, including crimes that were not reported police, increased significantly during shutdowns. We do not find that to be the case. Panel C of Figure 2 compares national rates of intimate partner violence (IPV) in January and February 2020 (pre-shutdown months) to those in April and May 2020 (shutdown months) and to the same two periods in 2019.20 The 2019 data reflect the usual seasonal pattern for IPV in the NCVS, where rates in April and May are typically higher than those at the start of the year. Rather than repeating this pattern, data from 2020 show a slight decline relative to the start of the year, leading to an implied reduction in IPV of 34 incidents per 100,000 population aged 12 and older. The corresponding regression model using Census region-by-month data, with region, year and month fixed effects yield an estimate of -17.9 incidents per 100,000 population (s.e. 11.2). When we separately examine crimes that are reported to police and those that are not, we find the decline is entirely attributable to a significant decline in IPV reported to police (-15.4, s.e. 7.4), which is consistent with the police data showing a decline in reported crimes.21

While it is still possible that shutdowns caused increases in DV that were not reported to police or to the survey, or caused changes in family dynamics that affected DV rates in subsequent post-shutdown months, none of the results in this analysis support a contemporaneous increase in DV during government-mandated shutdowns.

Conclusion

This paper uses incident-level police data on DV calls and crimes from major US cities that provided both measures to characterize the empirical evidence on the impact of pandemic more fully. Contrary to many media reports and claims by opponents of pandemic shutdowns (e.g., Biggs, 2020), the evidence presented here does not support an increase in DV rates during mandatory shutdowns. Instead, we find significant decreases in recorded DV assault incidents that we argue is unlikely to come from a lower propensity of police to file crime reports, in part because the decline is present for aggravated assaults. If police were failing to respond to these most serious crimes, we might expect to see increased fatalities related to DV during shutdowns. This is not what we find in data on intimate-partner homicides or suicides. We also find no evidence of an increase in nonfatal IPV in the population from national survey data on crime victimization.

The conclusion that recorded DV appears to have been lowered by shutdowns should not be interpreted as evidence that concerns regarding DV in the pandemic were unwarranted in the US. On the contrary, it is possible that increased federal funding for support services, as well as community and private sector efforts, contributed to raising awareness (and elevating DV calls to police during the initial emergency period) and improving support systems for victims and survivors. These measures may have contributed to lower DV assault rates and should therefore be considered during future pandemic shutdowns and also as ongoing policy efforts to reduce DV. This paper also illustrates the challenges faced by researchers who want to provide timely evidence to inform public policy related to DV. Enhanced real-time police data resources, with broad coverage across agencies and formal standards and requirements for data quality and elements, would be invaluable for future DV research and population health surveillance