Introduction

The coronavirus (COVID-19) pandemic is affecting every country and just about every person in one form or another. Since the United States Centers for Disease Control (CDC) first learned about a cluster of cases in Wuhan, China on December 31, 2019, to the identification of the first U.S. case on January 21, 2020 from a man who returned back to Seattle after traveling to Wuhan, to the first set of cases hitting every state in the U.S., the virus has gained national attention. Policy and government officials, health professionals, scientists, as well as researchers have been at the forefront of a wide range of issues, questions, and responses. In this study, we focus on one issue that may be related to COVID-19: criminal behavior, specifically domestic violence.

On Sunday April 5, 2020, United Nations Secretary António Guterres addressed the issue of domestic violence within the context of international lockdowns associated with COVID-19 by noting: “We know lockdowns and quarantines are essential to suppressing COVID-19, but they can trap women with abusive partners... Over the past weeks, as the economic and social pressures and fear have grown, we have seen a horrifying surge in domestic violence.” The UN Secretary also called for nations to plan for the prevention and redressing of domestic violence as part of their overall response to COVID-19 to include efforts to keep domestic violence shelters open as essential services, setting up emergency warning systems in pharmacies and grocery stores, etc. (United Nations, 2020; United Nations Women, 2020).Footnote 1

There are ample theoretical reasons to envision both increases and decreases in crimes during the lockdowns. On the one hand, there could be an increase in crime because people are ordered to stay in their homes, resulting in more people spending more time confined with one another. This could increase the types of negative strains and negative emotional responses, such as anger, that are referred to by Agnew (1992) in General Strain Theory that could be indicative of a crime spike. In fact, on this score, Phumzile Mlambo-Ngcuka, executive director of UN Women, recently stated: “Confinement is fostering the tension and strain created by security, health, and money worries” (Bilyeau, 2020).Footnote 2 Alternatively, there could also be observed decreases in reports of specific types of crime, such as domestic violence, because some victims may be afraid to call the police or domestic violence shelters because their perpetrators are confined with them and can closely monitor their communications.

Routine activities theory may also be called upon here to offer predictions of increases or decreases in crime. Predicting more crime, routine activities may anticipate an increase in domestic violence cases because all three conditions specified by the theory would appear to be met: a motivated offender, a suitable target, and the absence of a guardian. There is also the potential for decreases in other crimes (such as robbery and burglary) because fewer people are out in bars and shopping malls and many more are staying at home, thereby reducing—or even eliminating—some of the elements needed for routine activities to successfully anticipate these types of criminal behavior.

In this study, we examine the extent to which a stay-at-home order in Dallas, Texas was associated with any increase in domestic violence defined as abuse or assault against a family member, household member (including previous household members), or a current or past dating partner. Before we present the results of our investigation, we offer a brief overview of the potential effects of lockdowns on two prominent forms of domestic violence – child maltreatment and partner abuse. This is followed by a summary of what is known from both media reports and empirical research on the relationship between lockdowns and crime. We then turn toward our analyses.

Background

Childhood maltreatment, including abuse and neglect, has been identified as a prevalent and severe public health issue that can lead to a myriad of negative outcomes across the life course. Prevalence rates of maltreatment vary dependent on the type of abuse or neglect. Self-reported incident rates demonstrate that supervision neglect (i.e., being left home alone as a child) is the most common form of maltreatment. According to The National Longitudinal Adolescent Health study (Hussey, Chang, & Kotch,, 2006), 41.5% of respondents indicated that they had been left home alone as a child. Physical assault was the second most common form of maltreatment (28.4%), followed by physical neglect (11.8%) and contact sexual abuse (4.5%). Within the entire country, approximately 37.4% of children under the age of 18 experience being the subject of a childhood maltreatment investigation at some point in their lifetime (Kim, Wildeman, Jonson-Reid, & Drake, 2017), with Black youth being most likely to be involved in such investigations (53%) and Asian youth least likely (10.2%). Socio-economic status (SES) has also been found to be related to neglect, with children belonging to lower SES families experiencing higher rates of neglect (Vanderminden et al., 2019). Additionally, it is estimated that 1 year of childhood maltreatment investigations costs the country approximately $585 billion, roughly 4% of the US gross domestic product in 2010 (Fang, Brown, Florence, & Mercy, 2012).

Childhood abuse and neglect also has long-lasting psychological effects on children. On this score, all forms of maltreatment (supervisory neglect, physical assault, physical neglect, and sexual assault) have been associated with increased trauma symptoms and suicidal thoughts in children aged 10–17, as well as an increased risk of underage drinking and illicit drug use (Vanderminden et al., 2019). Child abuse has also been linked to issues throughout adulthood, including a higher likelihood of mental health problems such as depression, and a decrease in intellectual and cognitive development (Anda et al., 2006; Lansford et al., 2002; Sousa et al., 2018). Individuals who have experienced childhood maltreatment also demonstrate a limited ability to deal with stress and an increase in antisocial behavior (Currie & Tekin, 2012; McCrory, De Brito, & Viding, 2011; Widom, Fisher, Nagin & Piquero, 2018). Additionally, the effects of maltreatment lead to worsened health across the lifespan and are linked to increased mortality in adulthood (Anda et al., 2006; Jonson-Reid, Chance, & Drake, 2007).

Risk factors for maltreatment have been extensively studied in the literature. Distinct demographic, socioeconomic, and familial factors have been found to be associated with childhood abuse and neglect. As aforementioned, children in lower SES families are more likely to be subjected to childhood maltreatment. At the county level, rates of neglect have been positively associated with rates of births by teens, percentage of births by unmarried mothers, drug-related offenses, and percentages of children receiving supplemental nutrition program (SNAP) benefits (Morris et al., 2019). More proximal factors that increase risk of childhood maltreatment include mother’s unhappiness, stress and reactivity, father’s drinking, and children’s problem behavior. Parents’ age has also been inversely related to maltreatment, with younger parents being more likely to abuse and neglect their children (Black, Heyman, & Slep, 2001).

The COVID-19 stay-at-home orders are creating major disruptions in daily life. The cessation of many face-to-face interactions due to the mandated responses to the pandemic has disrupted the lives of all Americans, especially the lives of families with children. These families are struggling to figure out how to meet the health, well-being, and educational needs of their children. Caregivers are coping with closures to schools and childcare facilities and they must now take on the daily demands of overseeing the educational needs of their children. Risk for child maltreatment may be elevated when families and caregivers are dealing with such stressful conditions. Caregiver adversities and strains have been repeatedly identified as key risk markers that increase the likelihood of child abuse and neglect (see Agnew, Rebellon, & Thaxton, 2000). Agnew et al. (2000) theorized that caregiver strain produces negative emotions in the caregiver, which are commonly expressed by caregivers withdrawing from or neglecting their nurturing responsibilities. Caregiver strains and family adversity, including lack of social support, economic hardship, unsupportive family structure and domestic violence may result in failed nurturing and child maltreatment (Agnew et al., 2000).

Most importantly, mandated responses to the COVID-19 pandemic have altered the availability of health and social service resources typically relied upon by vulnerable children and their families. Reduced social and educational support is particularly challenging for those families whose members have behavioral or mental health needs, medical needs, families with co-custody arrangements, or children in foster care or at risk for child maltreatment. School and childcare personnel function as guardians for children at risk for maltreatment and provide the highest percentage of referrals to child protection agencies (U.S. Department of Health & Human Services, 2020). With children mandated to stay home rather than attend school or go to childcare, many incidents of abuse may go undetected or unreported.

Beyond the effects of the stay-at-home order on child maltreatment, another important consideration is the amount of conflict between partners. Researchers have confirmed that varied forms of family violence often occur in tandem, with strong associations between partner abuse and child maltreatment (Herrenkohl, Sousa, Tajima, Herrenkohl, & Moylan, 2008).

Similar themes are evident within the context of family, domestic, and intimate partner violence. Due to the COVID-19 pandemic, levels of family adversity have increased in unprecedented ways. Many have lost their jobs and sources of income, while others are worried over future economic uncertainties. Times of financial stress and problems at work, possibly moderated by alcohol use, have been linked to perpetration of partner abuse by both men and women (Capaldi, Knoble, Shortt & Kim, 2012). As partner abuse can be driven by a need for power and control, when life spins out of control at work or financially, individuals find it difficult to cope with these pressures and may become abusive toward their family members.

Stay-at-home orders may have unintentionally compounded the threat for domestic violence victimization by trapping at-risk partners at home and disrupting access to social support and social service resources typically available to them. The mandated stay-at-home orders even mimic common forms of partner abuse such as forcing isolation from friends and family, preventing the victim from working or attending school, and generally controlling the victim’s associations, movements, and activities (National Domestic Violence Hotline, n.d.). With stay-at-home orders, those at risk for partner abuse are unable to seek safety or assistance from family, friends, or service providers.

Lockdowns & Crime

Much of the early investigations surrounding potential crime increases due to coronavirus have been published by the media, with activists and academics weighing in with their predictions. There are some initial indications and a few case studies surrounding the potential rise in crime associated with COVID-19 and the ensuing lockdown orders around the world (Wagers, 2020). In the state of Texas, the locale of our analysis, there have been stories about a potential rise in child abuse in Fort Worth (Solis & Martinez, 2020), reported increases in domestic violence, assault, and burglary in Houston (CW39, 2020), reported increases in domestic violence cases by the District Attorney of Montgomery County (Gonzalez, 2020), and increases in Google-based searches for help with domestic violence (Neuman, 2020; Townsend, 2020). On the other hand, preliminary data from Miami-Dade County (Florida) show that domestic violence arrests decreased throughout March (Ovalle & Rabin, 2020) and other data from Florida suggests that child abuse cases have decreased (McKinnon, 2020).Footnote 3 Relatedly, according to RAINN, the Rape, Abuse and Incest National Network, which manages the National Sexual Assault Hotline, during March there was a 22% increase in calls to the hotline (Kamenetz, 2020). For the first time in the history of the hotline, half of the calls came from minors. Of the minors who called the hotline, 67% reported their perpetrator was a family member and 79% reported that they were currently living with that perpetrator.

Analyses from the Associated Press (2020) providing an ‘early-look’ at several cities throughout the United States (and around the world) also show reductions in crime. Concurrently however, there have been increases in calls to mental health and suicide prevention hotlines during the time period of many stay-at-home orders (Goodman, 2020; Jackson, 2020).

At the same time, there has been little systematic examination of how the COVID-19 virus and the various stay-at-home orders have affected crime one way or the other from the academic and research community in large part due to data limitations. For the most part, available information tends to suggest little increase in crime throughout the United States (see Ashby, 2020; Shayegh & Malpede, 2020; Weichselbaum & Li, 2020; Campedelli et al., 2020). Using crime data from 17 large U.S. cities, Ashby (2020) used crime data in previous years in order to forecast the expected frequency of crime in the early part of 2020. His results showed that there was little change or even decreases in some crime types but the effects were not uniform across all the cities. In one Australian study, Payne, Morgan, and Piquero (2020) used monthly violent crime (common assault, serious assault, sexual offenses and breaches of domestic violence orders) data for Queensland, Australia in order to model the monthly offense rate between February 2014 and February 2020. (Most restrictions took place in the month of March throughout Australia, with a combination of federal and territory restrictions). They used the 2014–2020 data to forecast March 2020 data and then compared the forecast to the actual crime that was recorded in March 2020. Results showed that rates of recorded crime in March across the four violent crime categories did not differ from the forecast, and in most cases the actual March 2020 crime figures were lower than the forecasts generated.

Current Focus

Overall, there has been much more media reporting of potential short-term changes in crime due to COVID-19. In particular, the media has focused on the various stay-at-home policies that have been mandated. However, there have been few rigorous empirical analyses from the research community. Among those that have been conducted, the preliminary conclusion is that there is only limited support for the suggestion that COVID-19 and its associated containment measures has had an impact on violent crime and, in the more robust (but short-term) studies, there has been no strong evidence of an impact on domestic violence. Other studies around the world looking at domestic violence reveal mixed results, with some evidence of increases in some places but decreases in others. Nevertheless, most media reports and empirical studies have been short-term, not providing for an adequate time of observation after stay-at-home style orders have been implemented.

Our study was designed to overcome the ‘time-frame’ limitation. In particular, we use data from the Dallas Police Department to examine whether the stay-at-home order served as a ‘negative’ intervention, so to speak, that was associated with an increase in domestic violence. Our time frame includes more observed (daily) points of data than what has been found in the extant literature, thereby giving us greater confidence in our ability to detect both short-term and longer-term trends that occurred after the lockdown was instituted compared to a lengthy ‘pre-lockdown’ period.

Specifically, this study provides an initial, short-term examination of how COVID-19 and responses by the city and county of Dallas have affected one type of criminal activity, domestic violence. Using an array of modeling strategies, we examine how domestic violence crimes changed following the Dallas stay-at-home order on March 24, 2020.

Methods

Data

The city of Dallas, Texas is the site of our investigation. As of July 1, 2019, the city had a population of 1,300,000 individuals, making it one of the country’s ten largest cities (coming it at #9). The city is part of the Dallas-Fort Worth metroplex, which is the fourth largest metropolitan area in the United States. The city of Dallas and its surrounding counties is also home to a large number of Fortune 500 companies.

Public safety in the city of Dallas is overseen by the Dallas Police Department (DPD), which has over 3500 sworn personnel and over 500 civilian employees. Like in any other major city, Dallas has experienced its fair share of crime increases and decreases—depending on the year and crime type. Most recently, 2019 saw the city experience a large surge in the number of homicides. Yet during the first quarter of 2020, homicides were down significantly—as were many other violent crimes with the exception of aggravated assaults. Nevertheless, for most of the first quarter of 2020, most violent and non-violent crimes were holding steady with some slight fluctuations.

The DPD requires officers to enter all reports into the report management system (RMS). Any call or self-initiated activity (e.g., traffic stop) triggers the creation of an incident number. If, as a result of the incident, officers encounter a situation where a report must be completed, they will log into the DPD field-based reporting system and complete the report. Before submitting their report, officers are required to indicate if the incident is a domestic violence-related report by marking “yes” or “no” from a drop-down menu. After the report is completed, it is assessable in RMS and all data in the report are extractable. In fact, crime analysts pull data from RMS daily and create descriptive analyses for command staff members.

One type of report created by DPD crime analysts, and the focus of the current study, is the Family Violence Incident List.Footnote 4 Data recorded in the Family Violence Incident List include misdemeanor and felony domestic violence, child abuse, elderly abuse, and sexual assault for offenses. Misdemeanor domestic violence includes harassment, interference with 911 calls, unlawful restraint, deadly conduct, and arrest warrant arrests. Felony domestic violence includes aggravated assaults, violations of protective orders, and strangulation offenses. Child abuse includes offenses such as indecency with a child (persons <17 years old) and sexual assault of a child. Elderly abuse means a crime committed against someone who is 65 or older, while sexual assaults are those offenses committed against adults.Footnote 5

For this study, we rely upon domestic violence incident reports that occurred between January 1st, 2020 and April 27th, 2020, for a total of 118 days of incident counts. An index of daily counts of all domestic violence incidents in Dallas was created by summing together misdemeanor, felony, child, elderly, and sexual assault family violence incidents. In Table 1, descriptive statistics for total domestic violence and individual categories are displayed. The daily domestic violence incident count is the dependent variable in this study.

Table 1 Descriptive statistics
Full size table

Following the Dallas County order issued by Judge Clay Jenkins, the City of Dallas similarly adopted the ‘stay-at-home’ order on March 24th, 2020, with regulations taking effect on the same day. Appendix A includes all of the necessary details regarding the various lockdown orders in the City of Dallas, Dallas County, and the State of Texas. An indicator variable was created to represent the stay-at-home order intervention, with the 83 days before March 24th (January 1st to March 23rd) coded as zero and March 24th and afterwards (March 24th to April 27th) coded as one for the 35 day post-intervention period.

Analytic Plan

We take multiple approaches to test the effect of the stay-at-home order on domestic violence in the city of Dallas. First, we created a simple descriptive graph depicting domestic violence crime counts for the entirety of the data, as well as a 6-week window centered about the intervention date. Next, we conducted a trend analysis using Stata’s itsa command that estimates the effect of the stay-at-home order on our daily time series. This test estimates the linear trend in domestic violence before the stay-at-home order, detects any presence of change in domestic violence at the time the intervention began, and estimates the linear trend of domestic violence after the stay-at-home order. We also considered the descriptive graph and the potential for a second change in the trend.

A Dickey-Fuller testFootnote 6 for non-stationarity and a correlogram plot to determine if lagged values are autocorrelated with the domestic violence data series were used. Then, we made use of multiple OLS and Poisson regression models that include the lag terms that are correlated with the data series. Here, the OLS models provide an easy interpretation of the magnitude of the effect of the Dallas stay-at-home order – similar to a difference in means test – and the Poisson regression models are used because our time series consists of count data.

Lastly, we used an autoregressive integrated moving average (ARIMA) forecast model that uses only the data from before the stay-at-home order to predict future domestic violence incidents. Prediction intervals about the point estimates were also generated, allowing us to compare actual domestic violence to domestic violence that would have been predicted with data from before the stay-at-home order began. This was done in order to attempt to understand what crime analysts might have expected to see in domestic violence crime counts before the city’s stay-at-home order came into effect.Footnote 7

Results

The graphs in Fig. 1 provide a visual representation of the entire time series and a restricted 6-week window for domestic violence incidents in Dallas, TX, with the series appearing to follow a moving average pattern. To set up our data for trend analysis, we made use of the Statistical Analysis System (SAS) date format which presents dates as the number of days elapsed since January 1st, 1960 (StataCorp. 2019) (e.g., January 2nd, 1960 is 00001). Results of the domestic violence trend analysis are displayed in Table 2, with Fig. 2 graphically portraying the estimated linear models for observed domestic violence before and after the March 24th (SAS date 21,998) stay-at-home order. As can be seen in Table 2, the implementation of the stay-at-home order is not associated with a statistically significant increase in domestic violence incidents, and there is not enough evidence to suggest an upward trend in domestic violence incidents throughout the month after the stay-at-home order went into effect.

Fig. 1
figure 1

Domestic violence incidents in Dallas. Two way line plots of daily domestic violence incident counts in Dallas, TX, from January 1st to April 27th, 2020, and a centered look at the 3 weeks before and after the stay-at-home order. Domestic violence incident counts are the solid black line, and the dashed vertical line drawn at Mar 24 indicates the beginning of the city’s stay-at-home order regulations

Full size image
Table 2 Trend analysis results
Full size table
Fig. 2
figure 2

Trend analysis, March 24th (21998) breakpoint. Domestic violence trend analysis for before and after the stay-at-home order. Solid black lines represent the estimated trend and solid black points are observed data. The dashed vertical line is the intervention at March 24th, or 21,998 days after January 1st, 1960

Full size image

However, if we consider the possibility that there are two changes in the trend of domestic violence, which might be observed visually in Fig. 1 and Fig. 2, then we can estimate a second break in the trend on April 7th (SAS date 22,012), 2 weeks after the stay-at-home order went into effect. Figure 3 contains the visual representation of the two break point estimation in Table 2. According to the results of the two break point trend analysis, there is statistically significant evidence that the trend in domestic violence changed twice: it increased after March 24th and decreased after April 7th.

Fig. 3
figure 3

Trend analysis, March 24th (21998) and April 7th (22012) breakpoints. Domestic violence trend analysis for before and after the stay-at-home order. Solid black lines represent the estimated trend and solid black points are observed data. The dashed vertical lines are at March 24th and April 7th, or 21,998 and 22,012 days post January 1st, 1960

Full size image

OLS and Poisson Regression Results

Results of a Dickey-Fuller test in Table 3 suggest the domestic violence series is stationary, but the autocorrelation plot in Fig. 4 suggests the potential for some seasonality in the data, with the first, seventh, and fourteenth lag of domestic violence counts possibly exhibiting effects on future values. This presents the possibility that those past values in the series could be strong predictors of respective future values (e.g., domestic violence the day/week/2 weeks before any given day is a good predictor of domestic violence on that given day). As shown in eqs. 1 through 8, we take into account the effects of the lagged series in both the OLS and Poisson models. These models take the following form:

Table 3 Dickey-fuller test results
Full size table
Fig. 4
figure 4

Autocorrelation plot (correlogram). Autocorrelation plot of the lag terms with the data series. The shaded grey area is a 95% confidence interval band. If autocorrelation between a lag term and the data series extends above (positive) or below (negative) the confidence interval, there is a correlation between the lag term and the data series (ex: the first lag can help predict current data points)

Full size image

OLS

$$ \mathrm{Domestic}\ \mathrm{Violence}=\upalpha +\upbeta\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+\upvarepsilon $$
(1)
$$ \mathrm{Domestic}\ \mathrm{Violence}=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}1\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(2)
$$ \mathrm{Domestic}\ \mathrm{Violence}=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}7\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(3)
$$ \mathrm{Domestic}\ \mathrm{Violence}=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}14\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(4)

Poisson

$$ \mathit{\ln}\left(\mathrm{Domestic}\ \mathrm{Violence}\right)=\upalpha +\upbeta\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+\upvarepsilon $$
(5)
$$ \mathit{\ln}\left(\mathrm{Domestic}\ \mathrm{Violence}\right)=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}1\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(6)
$$ \mathit{\ln}\left(\mathrm{Domestic}\ \mathrm{Violence}\right)=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}7\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(7)
$$ \mathit{\ln}\left(\mathrm{Domestic}\ \mathrm{Violence}\right)=\upalpha +{\upbeta}_1\ \mathrm{x}\ \left(\mathrm{Stay}-\mathrm{at}-\mathrm{Home}\right)+{\upbeta}_2\ \mathrm{x}\ \mathrm{lag}14\left(\mathrm{Domestic}\ \mathrm{Violence}\right)+\upvarepsilon $$
(8)

The results of our regression analyses in Table 4 provide evidence that domestic violence increased in the days after the stay-at-home order went into effect. Model 1 can be interpreted similar to a difference in means test, with days after the stay-at-home intervention regulations were implemented having an estimated 3.4 more domestic violence incidents, increasing from 35.4 to 38.8. The stay-at-home intervention is statistically significant at the .10 level for Model 2 and 3, but not so in Model 4. Accounting for the effect of lagged values reduces the strength of the relationship between the stay-at-home order and the domestic violence series.

Table 4 OLS and poisson regression results
Full size table

Our Poisson regression models are not affected in the same way as the OLS models when modeling the lagged dependent variable. Across Models 5–8, the stay-at-home order remains statistically significant at the .05 level. Poisson regression coefficients can be interpreted as the expected change in the difference in the natural log of expected domestic violence counts (Bruin, 2006). If we take Model 5 as an example, we can work to estimate the expected percent increase and expected change in domestic violence counts after the stay-at-home order was issued by doing the following:

$$ {e}^{\mathrm{coefficient}}\hbox{--} 1=\mathrm{expected}\ \mathrm{percent}\ \mathrm{change} $$
(9)
$$ {e}^{0.092}\hbox{--} 1=\mathrm{expected}\ 9.6\%\mathrm{increase} $$
(10)
$$ {e}^{\mathrm{coefficient}}\ \mathrm{x}\ {e}^{\mathrm{constant}}=\mathrm{expected}\ \mathrm{daily}\ \mathrm{count}\ \mathrm{after}\ \mathrm{stay}-\mathrm{at}-\mathrm{home} $$
(11)
$$ {e}^{0.092}\ \mathrm{x}\ {e}^{3.566}=\mathrm{expected}\ 38.8\ \mathrm{domestic}\ \mathrm{violence}\ \mathrm{incidents} $$
(12)

Even when accounting for potential seasonality, the results of Models 6–8 suggest we should expect anywhere from a 7.1 to 8.4% increase in domestic violence counts in the days after the stay-at-home order was issued.

ARIMA Forecasting Model

Finally, we turn to the results of the ARIMA forecast model displayed in Fig. 5. After fitting an ARIMA model suggested by the auto.arima() function in RFootnote 8 to the data from the 83 days before stay-at-home order, we generated a predicted domestic violence count and two prediction intervals for the remaining 35 days. The 80 and 95% prediction intervals can be interpreted as containing the range of values of domestic violence counts we would expect to see 80 or 95% of the time based on the data prior to March 24th. We are most concerned with observed crime counts peaking above the upper bounds of the prediction intervals. Such increases essentially indicate a “spike” in crime counts as they go further beyond what we expect to see 95% of the time.

Fig. 5
figure 5

Domestic violence projection. Domestic violence ARIMA forecast for daily incident counts in Dallas, TX. Observed crime data is the solid black line, and the dashed black line represents the forecasted point estimate for domestic violence. The dark grey polygon is the 80% prediction interval and the light grey polygon is the 95% prediction interval. The dotted black line at March 24th represents the beginning of the stay-at-home intervention

Full size image

In the first week after the stay-at-home order went into effect, we see that the series rises above the 80% prediction interval for 3 days, peaks above the 95% interval 2 weeks afterwards, and once again rises above the upper bound of the 80% interval in the last week of our data. Overall 77% (27/35) observed days fall below the 80% interval, with only 1 day (2.8%) rising above the 95% prediction interval’s upper limit. This suggests one could have predicted this movement in crime before the stay-at-home order was put into place and that an upward trend after March 24th may have been part of an already upward moving average.Footnote 9

Conclusions

The coronavirus has wreaked havoc on the lives of persons worldwide. Numbers of infected cases and deaths change by the hour such that even listing a current number here would be immediately outdated. As scientists and doctors race for both treatments and vaccines, the ancillary effects of the COVID-19 epidemic are far-reaching. In this study, we were concerned with one of these additional outcomes, lockdowns or stay-at-home orders, and the extent to which such orders would result in an increase in domestic violence. We examined this issue within the context of an intervention time-series methodology using data from Dallas, Texas to compare domestic violence crimes for an 83-day period before the stay-at-home order was mandated to a 35-day period thereafter.

At first glance, we did not detect strong evidence of any change in the trend of domestic violence. However, upon further inspection of the time series, our results showed that there appeared to be an increase in domestic violence in the first 2 weeks after the stay-at-home order was implemented but then a decrease thereafter. Yet, some of that short-term spike seems to be associated with what appears to be an upward trend of domestic violence crimes that was already occurring prior to the stay-at-home order – leading us to cautiously interpret our regression models. While it may be true that domestic violence is higher (on average) after March 24 than before, it is difficult to directly and solely attribute the higher average count of incidents to the stay-at-home order. The latter finding of a short-term increase could be due to the fact that people were already voluntarily starting to stay at home in the days prior to the March 24 order because of CDC and public health officials warning of the dangers of this new virus and because of City, County, and State orders to limit social gatherings to less than 10 people. This interpretation also coincides with COVID-19 altering the routine activities of individuals. Let us explain a bit further.

On this point, Google created community mobility maps for counties and states in the U.S. and countries throughout the world that show the significant changes that the virus and its related consequences have had on the daily routines of people’s lives. Google uses local history (turned on) data to track how often and for how long people travel to different location types, compared with a baseline value (the median value for the same day of the week, in this case, in January and early February). Figure 6 plots the Google Mobility data for Dallas County from February 15, 2020 through April 11, 2020 (the last data available), superimposed with the stay-at-home order issued on March 24, 2020. Through April 11, the percentage change from baseline is staggering: retail/recreation down 47%, grocery/pharmacy down 14%, workplace down 42%, parks down 56%, and transit stations down 48%. The only exception to the downward trends, of course, is residential placement which increased 18% since the baseline but was increasing about a week prior as restaurants, bars, private clubs and other business started to suspend operations due to a Dallas County’s order on March 16 (see Appendix). As can be discerned, changes in person’s daily routines were already in motion well before the March 24 stay-at-home order.

Fig. 6
figure 6

Google mobility patterns, Dallas County, Dallas, Texas. February 29, 2020–April 11, 2020

Full size image

In short, and based on the projection models, we do not see, at least with the data we have, any lasting increase or sustained higher levels of domestic violence. There is some evidence to reject the null hypothesis that the lockdowns had no (increasing) effect on domestic violence, but not all of the results strongly support the idea that the stay-at-home intervention has domestic violence increasing/at sustained higher levels. This of course, does not indicate the stay-at-home order was the sole cause of that short-term spike, which as we noted earlier may have also been the continuation of an already increasing time series, perhaps due to the seriousness of warnings about COVID-19 from other locales. That increase could also have been associated with other things going on at the same time such as people working more from home, being furloughed, or laid-off from their employer. As previously noted, financial stress and problems at work have been linked to perpetration of partner abuse (Capaldi et al., 2012) and these stresses brought on by the COVID-10 pandemic could be related to the increase in domestic violence prior to the stay-at-home order.

We believe that our analyses offer an important, strong empirical baseline for which additional data from Dallas could be collected to assess the more long-term effects of the stay-at-home order with respect to domestic violence as well as other crime types. Yet, we are also mindful of the limitations of our work. First, while we had longer follow-up data than most studies to date, there is a need to consider additional data to see what happens when other restrictions are loosened or lifted altogether. Second, while we focused on crime counts, there is also a need to perform a deep dive into the data to examine if there are chronic abusers that were committing more of the offenses—as is the case with what has been learned from criminal career research (Piquero, Farrington, and Blumstein, 2003). Third, using a large city like Dallas for our case study may pose some generalizability concerns, especially with respect to the problem of domestic violence in rural communities. On this score, it is important to note that many rural communities in Texas did not experience any sheltering orders, but analysis of domestic violence across different types of communities, city, suburban, and rural is important going forward. Finally, we relied on official records and have surely missed out on other domestic incidents that did not come to the attention of the police department, which we suspect is a larger—but uncertain number.

Like crime, we will likely be living with COVID-19 for the rest of our lifetimes. It is incumbent on the community of scholars to continue to track its adverse effects on persons and their lives throughout the world.