Scholarly Research and Expert Articles

1. Statistics

1. Post-Lockdown ‘Household Harm’ – Covid-19 and Beyond

Every day, there are almost 20 people a minute that are physically abused by someone close to them. That comes to about 10 million women and men a year that are victims of domestic violence.

Domestic violence incidents reported in the United States increased by 8.1 percent after COVID-19 lockdown orders, according to a recent study published by the National Commission on Covid-19 and Criminal Justice (NCCCJ). According to experts, this increased amount is merely a “floor” and not representative of the actual dramatic increase in household abuse and violence. For example, the Phoenix Police Department saw a 140% increase in DV-related homicides in the first half of 2020, and according to international sources, DV Hotline emergency calls, have substantially increased worldwide. This is despite many DV organizations and law enforcement agencies’ activities and resources being scaled back due to pandemic-related and other issues.

Domestic Violence Statistics

National Statistics

• Every 9 seconds in the US a woman is assaulted or beaten.
• On average, nearly 20 people per minute are physically abused by an intimate partner in the United States. During one year, this equates to more than 10 million women and men.
• 1 in 3 women and 1 in 4 men have been victims of (some form of) physical violence by an intimate partner within their lifetime.
• 1 in 5 women and 1 in 7 men have been victims of severe physical violence by an intimate partner in their lifetime.
• 1 in 7 women and 1 in 18 men have been stalked by an intimate partner during their lifetime to the point in which they felt very fearful or believed that they or someone close to them would be harmed or killed.
• On a typical day, there are more than 20,000 phone calls placed to domestic violence hotlines nationwide.
• The presence of a gun in a domestic violence situation increases the risk of homicide by 500%.
• Intimate partner violence accounts for 15% of all violent crimes.
• Women between the ages of 18- 24 are most commonly abused by an intimate partner.
• 19% of domestic violence involves a weapon.
• Domestic victimization is correlated with a higher rate of depression and suicidal behavior.
• Only 34% of people who are injured by intimate partners receive medical care for their injuries.

Intimate Partner Physical Abuse

• More than 10 million Americans are victims of physical violence annually.
• 20 people are victims of physical violence every minute in the United States.
• 1 in 3 women and 1 in 4 men is a victim of some form of physical violence by an intimate partner during their lifetimes.
• 76% of intimate partner physical violence victims are female; 24% are male.
• 1 in 7 women and 1 in 18 men are severely injured by intimate partners in their lifetimes.
• Domestic violence accounts for 15% of all violent crimes in the United States.
• Domestic violence is most common among women aged 18- 24 and 25-34.
• A majority of physical abuse is committed by dating partners rather than spouses.
• More than 75% of women aged 18-49 who are abused were previously abused by the same perpetrator.
• Slightly more than half of intimate partner physical violence is reported to law enforcement.

Children and Domestic Violence

• 1 in 15 children are exposed to intimate partner violence each year, and 90% of these children are eyewitnesses to this violence.
  (see: https://www.thehelpsavefoundation.org/the_help_save_foundation/contact-us/)

2. The Truth About Domestic Abuse and Domestic Violence Statistics (and the lack thereof):

Some reasons for the lack of more accurate and meaningful statistics are:

• Few cases of verbal, emotional, financial, and sexual abuse are in the count, yet they must be brought into the public eye and stopped before becoming a more permanent danger and nightmarish way of life.
• Victims have fewer places to contact since many organizations that once could have helped them have closed or are working remotely with fewer resources.
• Visible signs of abuse are not as apparent during remote learning sessions as they are in person, so third-party reports are scarce, although many teachers have reported incidents they saw in the background.
• Police are less available and preoccupied with public safety from rising street gangs, subway crimes, and other more visible felonies. Unfortunately, toxic behavior, coercive control, household abuse, and domestic violence cases are also serious threats to public safety, but they must reach a dangerous threshold before law enforcement is notified and can act.

What can be done to obtain better data?

Getting Male Participation

Note that women account for 99.2% of intimate partner violence victims, and what we see missing is the support of males who care about this subject and are willing to act. One exception is “Men Stopping Violence (MSV) a national training organization dedicated to educating men to create safer communities for women and girls. As such, MSV offers diverse training programs to individuals, families, non-profit, government and, corporate groups.” Also, note that 80% of DV cases involve minorities, yet because these groups are fighting for social justice and equality under the law, self-examination, and cultural improvement, even of such critical importance, could understandably receive a lower priority.

Another international organization focused on men against DV. Gender equality is a topic of a human rights group at the UN, see https://engagingmen.futureswithoutviolence.org/

Caged Birds Sing Inc. was co-founded by a man. As the spokesperson for their first PR campaign, A Stuart Kaplan, is dedicated to soliciting both female and male volunteers and professionals, emphasizing the need for males to step forward in this female-driven arena. He states: “If we are causing most of the problems, we males better jump in to help fix them.”

CBSI will also be gathering data from global sources and conducting surveys, to penetrate the current walls of silence that inhibit the production of more accurate and meaningful statistics.

One significant attempt to raise the Lack of Data issue is found in the following article excerpt:

“Data collected in surveys of nearly 400 adults for 10 weeks beginning in April 2020 suggest that more services and communication are needed so that even front-line health and food bank workers, for example, rather than only social workers, doctors, and therapists can spot the signs and ask clients questions about potential intimate partner violence. They could then help lead victims to resources, said Clare Cannon, assistant professor of social and environmental justice in the Department of Human Ecology and the lead author of the study:  “COVID-19, Intimate Partner Violence and Communication Ecologies” See: an informative article by Karen Nikos Rose dated 02.24.2021 – https://www.ucdavis.edu/curiosity/news/covid-19-isolation-linked-increased-domestic-violence-researchers-suggest

This form of communicating with victims, on a much larger sampling basis, can also help the dearth of meaningful Post-Lockdown ‘Household Harm ‘statistics, and the information that can be extrapolated from them.

When domestic violence stats show up on paper, the results are more than what you might expect.

Domestic violence most often happens behind closed doors, making it appear that it happens less often than what it actually does.

This evidence points to not only a growing awareness to help battle this issue but more of a concern on taking care of these victims.

Too many of these victims are left without any help that they desperately need because many states don’t have the funding or the resources to account for all of these cases.

Certain states don’t take domestic violence as seriously as they should, and victims are left without any real resolution.

Here’s a closer look at the nation’s domestic violence statistics and then a state-by-state statistics that just might shock you.

3. Domestic Violence Statistics, US

(see: https://domesticviolence.org/statistics/)

America has some pretty alarming statistics concerning intimate partner domestic violence. Here are some of those facts. Every day, there are almost 20 people a minute that are physically abused by someone close to them.

That comes to about 10 million women and men a year that are victims of domestic violence.

As many as 1 in 4 women and 1 in 9 men face severe domestic violence from an intimate partner, resulting in injury, stress disorders, contracting sexually transmitted diseases and many other devastating results.

On an average day, there are more than 20,000 phone calls that are made to domestic violence hotlines across the nation.

Domestic violence concerning intimate partner accounts for 15% of all violent crimes.

When there’s the presence of a gun in the home, the risk of homicide is increased by over 500% where domestic violence is taking place.

1 in 7 women and 1 in 18 men at some point in their lives, will feel threatened or fearful that they may be harmed or killed while being stalked by a past intimate partner.

As few as 34% of victims ever receive medical treatment for their injuries caused by their intimate partner.

State-by-State Statistics on Domestic Violence

It’s important to see some of the state facts in regard to domestic violence. Some of these facts may actually surprise you!

Alabama

As many as 31% of women and 26.9% of men in the state of Alabama experience domestic violence that’s been caused by an intimate partner.

Intimate partner violence accounted for 16% of all domestic violence crimes.

In 15% of those cases, a firearm was used one way or another.

24 of those domestic violence crimes resulted in homicide in 2013.

In 2013, there were 2,872 domestic violence aggravated assaults and 32, 587 domestic violence simple assaults that took place.

Alaska

There are studies that have shown that just about 50% of English-speaking adult women in the state of Alaska have faced intimate partner or sexual violence at one point or another.

These numbers get even uglier amongst American Indian/ Alaska Native women, were around 84% experience some form of domestic violence at some point in their lifetime.

Alaska leads the nation with the highest homicide rate of female victims killed by a male perpetrator.

Arizona

As many as 36.5% of women and 27.1% of men experience some form of domestic violence from an intimate partner.

The state of Arizona only recognizes 1st and 2nd-time offenders of domestic violence is only flagged. It’s not until the 3rd offense that they are charged with domestic violence. The most common sentence for this was probation.

In 2014, there were 109 deaths that were related to domestic violence.

Back in 2012, Arizona ranked 8th nationally for femicides per capita.

Arkansas

37.3% of women and 35.6% of men living in the state of Arkansas will experience intimate partner domestic violence at one time or another.

In just one day, 453 domestic violence victims were being served by shelters and other programs in the state of Arkansas.

As many as 18.6% of Arkansas women will be stalked by another individual at some point in their lifetime.

California

Every 56 minutes, another forcible rape is happening in the state of California.

In 2007, California experienced 174, 649 domestic violence cases, where a lot of them went unreported. Nearly 40% of those situations involved a deadly weapon.

When guns were found to be in the home that was experiencing domestic violence, a startling 64.5% of the victims claimed the gun was used against them.

In just a single day, as many as 5,800 women and children were being served in domestic violence shelters scattered all throughout the state.

Colorado

In 2014, 25 Colorado residents were killed by an intimate partner and nearly 70% of those cases a gun was found to be the weapon of choice.

Back in 2014, there were 904 victims being served at domestic violence programs, but nearly 163 had to be turned away because there weren’t enough resources to accommodate them.

Around 325,000 Colorado residents have been the victim of stalking.

Connecticut

In the state of Connecticut, just about 32.9% of women and 33.9% of men have experienced intimate partner domestic violence at some point.

An astonishing 1 in 3 criminal court cases has found to be situations of family violence.

In just one year, Connecticut’s relief programs served 46,750 victims of domestic violence.

Delaware

34.9% of women and 36.7% of men in the state of Delaware will experience domestic violence in some form at one point during their lifetime.

27,014 incidents of domestic violence were reported back in 2012.

Nearly everyday, domestic violence services have to turn away an average of 30 victims due to lack of resources.

Between the years 2003 and 2012, 74 women became victims of homicide. 25 of them were domestic violence-related.

Florida

There were as many as 108,030 domestic violence incidents that took place during 2013. Many cases remained unreported.

Back in 2013, 170 residents that were living in the state of Florida were killed by domestic violence.

From 2006 to 2012, guns were used 56% of the time concerning domestic violence homicide.

Georgia

35.1% women and 39.9% of men living in the state of Georgia, experience some form of domestic violence at some point during their lifetime.

During 2014, there were 117 Georgia residents that lost their lives due to domestic violence. In 29% of these situations, their children witnessed the murder of their parents.

There are 27 counties in the state of Georgia that have no domestic violence services and another 26 counties that are extremely limited.

Hawaii

35.7% of women and 21.8% of men living in Hawaii, experience domestic violence at some point during their lifetime.

An average of 1 in 7 Hawaiian women are raped at some point.

As many as 41% of all domestic violence programs are understaffed or don’t receive enough funding, or both.

Idaho

In 2012, less than 2% of all intimate partner domestic violence incidents were reported to police in the state of Idaho, and 0% concerning sexual abuse.

Between 2003 and 2012, nearly 69% of all domestic violence homicides used a firearm.

The Idaho court system dismissed around 37% of misdemeanor domestic violence charges and right around 31% of felony domestic violence charges in 2013.

Illinois

37.7% of women and 25.7% of men living in the state of Illinois have experienced some type of domestic violence at one point or another.

There were 65,800 reported intimate partner violence incidents that took place in 2014. Several others went unreported.

Between 2013 and 2014, the state of Illinois experienced 84 deaths that were due to domestic violence. 15 of them happened to be children.

Indiana

40.4% of women and 26.8% of men that live in the state of Indiana will experience domestic violence at some point during their lifetime.

Well over half of all the domestic violence homicides in the state of Indiana were committed with a gun.

On any given day, there are about 1,807 victims/survivors being served by domestic violence programs and nearly 182 of them did not have their needs met due to lack of resources.

Iowa

31.3% of women and 19.6% of men living in the state of Iowa have been the victims of domestic violence at one time or another.

There were 6,628 domestic violence cases that were reported to the police back in 2011. Many others did not get reported.

There have been researchers that have estimated that around 55,340 people living in the state of Iowa, experienced sexual violence in 2009.

Kansas

29% of women and 23% of men living in the state of Kansas experience domestic violence at one time or another.

There were 23,508 domestic violence incidents that were reported to the police back in 2013. Not every agency reported their numbers to the state, so this number is much lower than the actual incidents.

Between 2003 and 2012, over half of all Kansas domestic violence homicides were found to have had a firearm involved.

Kentucky

37.5% of Kentucky women and 31% of men living in the state of Kentucky have experienced domestic violence in one form or another, at some point during their lifetime.

As many as 1 in 9 women living in the state of Kentucky have been the victim of forcible rape.

Between 2013 and 2014, there were programs that provided bed-nights for 119,718 victims. An eye-opening 45% of those individuals were children.

Louisiana

33.4% of women and 28.4% of men that live in the state of Louisiana have experienced some form of domestic violence at one time or another.

81% of all female homicides that are due to domestic violence are found to be done by a partner or an ex-partner.

Louisiana ranked 4th in the nation back in 2010 for female homicide. Around two-thirds of those cases involved the use of firearms.

Maine

36.6% of women and 26.7% of men that live in the state of Maine have dealt with domestic violence at one time or another.

There were 5,487 domestic violence incidents that were reported to the police back in 2013. Only 17.4% of the time were any arrests made.

There were 439 victims that were served at one of Maine’s eight domestic violence relief programs on a particular day. 55 had to be turned away due to lack of resources.

Maryland

42.1% of women and 27.2% of men living in the state of Maryland will experience some form of domestic abuse throughout their lifetime.

In 2012, 50 people lost their lives due to domestic violence in the state of Maryland.

On one particular day, there were as many as 1,085 victims being served at domestic violence programs. 160 had to be turned away because they simply didn’t have the resources to help them.

Massachusetts

An astonishing 50% of women and around 25% of men will at some point be subject to non-rape sexual assault.

Back in 2008, Massachusetts declared domestic violence to be a public health emergency.

There were as many as 266 victims that were murdered because of intimate partner violence, and 74 of the perpetrators were killed also.

Michigan

41.8% of women and 23% of Michigan men will deal with domestic violence during their lifetime.

There were 103,331 reported domestic violence incidents in 2009. Several others were left unreported.

Just about 18.2% of Michigan women have been stalked by a partner or ex-partner at some point in their lifetime.

Minnesota

33.7% of women and 23.5% of men living in Minnesota have experienced some form of domestic violence at one time or another.

There were an estimated 80% of domestic violence victims back in 2002, that did not report the violence to the police.

1 in 3 women that were homeless in Minnesota, were homeless because of domestic violence.

Mississippi

40.1% of women and 25.8% of men living in the state of Mississippi have experienced domestic violence of some form at least once in their lifetime.

During a random phone survey time period in 2014, there were 181 calls to a local or state violence hotline in a single day. Just about 7 hotline calls were coming in every hour.

Mississippi has dropped from the 5th highest female fatalities in the country to 34th in the nation, especially because of the Mississippi Attorney General’s Office Domestic Violence Unit that was put in place.

Missouri

36.1% of women and 40.4% of men living in Missouri have experienced domestic violence in some form at one time or another.

1 in 7 women has been sexually abused in the state of Missouri.

In 2012, there were as many as 1,504 rapes or attempted rapes that were reported.

Montana

39.2% of women and 32.6% of men living in Montana have dealt with domestic violence at least once during their lifetime.

On a single day in 2014, there were 238 domestic violence victims being served, where 122 of those were children and 116 were adults.

These are facts leading up to 2020, what facts would you like to see on this list?

Recent Posts

• Signs of Domestic Abuse
• Domestic Violence Awareness Month
• Why Domestic Violence Victims Don’t Leave
• What is Sexual Coercion?

2. Conclusion From a 2022 National Institute of Health Paper

Domestic Violence

Authors: Martin R. Huecker; Kevin C. King; Gary A. Jordan; William Smock.

For the complete paper, Go to: ncbi.nlm.nih.gov/books/NBK499891/#

Prognosis

Without proper social service and mental health intervention, all forms of abuse can be recurrent and escalating problems, and the prognosis for recovery is poor. Without treatment, domestic and family violence usually recurs and escalates in both frequency and severity.[3][22][23]

• Of those injured by domestic violence, over 75% continue to experience abuse.
• Over half of battered women who attempt suicide will try again; often they are successful with the second attempt.

In children, the potential for poor outcomes is particularly high as abuse inflicts lifelong effects. In addition to dealing with the sequelae of physical injury, the mental consequences may be catastrophic. Studies indicate a significant association between child sexual abuse and increased risk of psychiatric disorders in later life. The potential for the cycle of violence to continued from childhood is very high.

Children raised in families of sexual abuse may develop:

• Attention deficit hyperactivity disorder (ADHD)
• Conduct disorder
• Depression
• Bipolar disorder
• Panic disorder
• Sleep disorders
• Suicide attempts
• Post-traumatic stress disorder (PTSD)

Health Outcomes

There are multiple known and suspected negative health outcomes of family and domestic violence. There are long-term consequences to broken bones, traumatic brain injuries, and internal injuries.

Patients may also develop multiple comorbidities such as:

• Asthma
• Insomnia
• Fibromyalgia
• Headaches
• High blood pressure
• Chronic pain
• Gastrointestinal disorders
• Gynecologic disorders
• Depression
• Panic attacks
• PTSD

Pearls and Other Issues

Screening: Tools

• The American Academy of Pediatricians has free guides for the history, physical, diagnostic testing, documentation, treatment, and legal issues in cases of suspected child abuse.
• The Center for Disease Control and Prevention (CDC) provides several scales assessing family relationships, including child abuse risks.
• The physical examination is still the most significant diagnostic tool to detect abuse. A child or adult with suspected abuse should be undressed, and a comprehensive physical exam should be performed. The skin should be examined for bruises, bites, burns, and injuries in different stages of healing. Examine for retinal hemorrhages, subdural hemorrhages, tympanic membrane rupture, soft tissue swelling, oral bruising, fractured teeth, and organ injury.

Screening: Recommendations

• Evaluate for organic conditions and medications that mimic abuse.
• Evaluate patients and caregivers separately
• Clinicians should regularly screen for family and domestic violence and elder abuse
• The Elder Abuse Suspicion Index can be used to assess for elder abuse
• Screen for cognitive impairment before screening for abuse in the elderly
• Pattern injury is more suspicious

Risks

• Failure to report child abuse is illegal in most states.
• Failure to report intimate partner and elder abuse is illegal in many states.

Legal

It is important to be aware of federal and state statutes governing domestic and family abuse. Remember that reporting domestic and family violence to law enforcement does not obviate detailed documentation in the medical record.

• Battering is a crime, and the patient should be made aware that help is available. If the patient wants legal help, the local police should be called.
• In some jurisdictions, domestic violence reporting is mandated. The legal obligation to report abuse should be explained to the patient.
• The patient should be informed how local authorities typically respond to such reports and provide follow-up procedures. Address the risk of reprisal, need for shelter, and possibly an emergency protective order (available in every state and the District of Columbia).
• If there is a possibility the patient’s safety will be jeopardized, the clinician should work with the patient and authorities to best protect the patient while meeting legal reporting obligations.
• The clinical role in managing an abused patient goes beyond obeying the laws that mandate reporting; there is a primary obligation to protect the life of the patient.
• The clinician must help mitigate the potential harm that results from reporting, provide appropriate ongoing care, and preserve the safety of the patient.
• If the patient desires, and it is acceptable to the police, a health professional should remain during the interview.
• The medical record should reflect the incident as described by the patient and any physical exam findings. Include the date and time the report was taken and the officer’s name and badge number.

National Statutes

Federal Child Abuse Prevention and Treatment Act (CAPTA)

Each state has specific child abuse statutes. Federal legislation provides guidelines for defining acts that constitute child abuse. The guidelines suggest that child abuse includes an act or failure recent act that presents an imminent risk of serious harm. This includes any recent act or failure to act on the part of a parent or caretaker that results in death, physical or emotional harm, sexual abuse, or exploitation.

Elder Justice Act

The Elder Justice Act provides strategies to decrease the likelihood of elder abuse, neglect, and exploitation. The Act utilizes three significant approaches:

Patient Safety and Abuse Act

The Violence Against Woman Act makes it a federal crime to cross state lines to stalk, harass, or physically injure a partner; or enter or leave the country violating a protective order. It is a violation to possess a firearm or ammunition while subject to a protective order or if convicted of a qualifying crime of domestic violence.

Enhancing Healthcare Team Outcomes

Domestic violence may be difficult to uncover when the victim is frightened, especially when he or she presents to an emergency department or healthcare practitioner’s office. The key is to establish an assessment protocol and maintain an awareness of the possibility that domestic and family violence may be the cause of the patient’s signs and symptoms.

Over 80% of victims of domestic and family violence seek care in a hospital; others may seek care in health professional offices, including dentists, therapists, and other medical offices. Routine screening should be conducted by all healthcare practitioners including nurses, physicians, physician assistants, dentists, nurse practitioners, and pharmacists. Interprofessional coordination of screening is a critical component of protecting victims and minimizing negative health outcomes. Health professional team interventions reduce the incidence of morbidity and mortality associated with domestic violence. Documentation is vital and a legal obligation.

• Healthcare professionals including the nurse should document all findings and recommendations in the medical record, including statements made denying abuse
• If domestic violence is admitted, documentation should include the history, physical examination findings, laboratory and radiographic finds, any interventions, and the referrals made.
• If there are significant findings that can be recorded, pictures should be included.
• The medical record may become a court document; be objective and accurate.
• Healthcare professionals should provide a follow-up appointment.
• Reassurance that additional assistance is available at any time is critical to protect the patient from harm and break the cycle of abuse.
• Involve the social worker early
• Do not discharge the patient until a safe haven has been established.

Resources

National

The following agencies provide national assistance for victims of domestic and family violence:

• Centers for Disease Control and Prevention (800-CDC-INFO (232-4636)/TTY: 888-232-6348
• Childhelp: National Child Abuse Hotline: (800-4-A-CHILD (2-24453))
• The coalition of Labor Union Women (cluw.org): 202-466-4615
• Corporate Alliance to End Partner Violence: 309-664-0667
• Employers Against Domestic Violence: 508-894-6322
• Futures without Violence: 415-678-5500/TTY 800-595-4889
• Love Is Respect: National Teen Dating Abuse Helpline: 866-331-9474 /TTY: 866-331-8453
• National Center on Domestic and Sexual Violence
• National Center on Elder Abuse
• National Coalition Against Domestic Violence (www.ncadv.org)
• National Network to End Domestic Violence: 202-543-5566
• National Organization for Victim Assistance
• National Resource Center on Domestic Violence: 800-537-2238 
• National Sexual Violence Resource Center: 717-909-0710

References

Sapkota D, Baird K, Saito A, Anderson D. Interventions for reducing and/or controlling domestic violence among pregnant women in low- and middle-income countries: a systematic review. Syst Rev. 2019 Apr 02;8(1):79.

Klein LB, Chesworth BR, Howland-Myers JR, Rizo CF, Macy RJ. Housing Interventions for Intimate Partner Violence Survivors: A Systematic Review. Trauma Violence Abuse. 2021 Apr;22(2):249-264. 

Marie-Mitchell A, Kostolansky R. A Systematic Review of Trials to Improve Child Outcomes Associated With Adverse Childhood Experiences. Am J Prev Med. 2019 May;56(5):756-764. 

Lewis NV, Dowrick A, Sohal A, Feder G, Griffiths C. Implementation of the Identification and Referral to Improve Safety programme for patients with experience of domestic violence and abuse: A theory-based mixed-method process evaluation. Health Soc Care Community. 2019 Jul;27(4):e298-e312.

Sarkar R, Ozanne-Smith J, Bassed R. Systematic Review of the Patterns of Orofacial Injuries in Physically Abused Children and Adolescents. Trauma Violence Abuse. 2021 Jan;22(1):136-146. 

Gao S, Assink M, Liu T, Chan KL, Ip P. Associations Between Rejection Sensitivity, Aggression, and Victimization: A Meta-Analytic Review. Trauma Violence Abuse. 2021 Jan;22(1):125-135.

Zeppegno P, Gramaglia C, di Marco S, Guerriero C, Consol C, Loreti L, Martelli M, Marangon D, Carli V, Sarchiapone M. Intimate Partner Homicide Suicide: a Mini-Review of the Literature (2012-2018). Curr Psychiatry Rep. 2019 Feb 21;21(3):13.

Hackenberg EAM, Sallinen V, Handolin L, Koljonen V. Victims of Severe Intimate Partner Violence Are Left Without Advocacy Intervention in Primary Care Emergency Rooms: A Prospective Observational Study. J Interpers Violence. 2021 Aug;36(15-16):7832-7854. 

Gottlieb A, Mahabir M. The Effect of Multiple Types of Intimate Partner Violence on Maternal Criminal Justice Involvement. J Interpers Violence. 2021 Jul;36(13-14):6797-6820. 

Roscoe LA, Schenck DP. Victim of Abuse, or Bully? The Case of the 800-Pound Man. Narrat Inq Bioeth. 2018;8(3):261-271. 

Wahi A, Zaleski KL, Lampe J, Bevan P, Koski A. The Lived Experience of Child Marriage in the United States. Soc Work Public Health. 2019;34(3):201-213. 

Harland KK, Peek-Asa C, Saftlas AF. Intimate Partner Violence and Controlling Behaviors Experienced by Emergency Department Patients: Differences by Sexual Orientation and Gender Identification. J Interpers Violence. 2021 Jun;36(11-12):NP6125-NP6143.

Jiang Y, DeBare D, Colomer I, Wesley J, Seaberry J, Viner-Brown S. Characteristics of Victims and Suspects in Domestic Violence-Related Homicide – Rhode Island Violent Death Reporting System, 2004-2015. R I Med J (2013). 2018 Dec 03;101(10):58-61.

Stone LB, Amole MC, Cyranowski JM, Swartz HA. History of childhood emotional abuse predicts lower resting-state high-frequency heart rate variability in depressed women. Psychiatry Res. 2018 Nov;269:681-687. [PMC free article] [PubMed]15.

Hansen JB, Killough EF, Moffatt ME, Knapp JF. Retinal Hemorrhages: Abusive Head Trauma or Not? Pediatr Emerg Care. 2018 Sep;34(9):665-670. [PubMed]16.

Skott S. Disaggregating Violence: Understanding the Decline. J Interpers Violence. 2021 Aug;36(15-16):7670-7694. [PubMed]17.

Lifflander AL. Hard Times and Hard Stops. JAMA. 2019 Mar 05;321(9):837-838. [PubMed]18.

Saywitz KJ, Wells CR, Larson RP, Hobbs SD. Effects of Interviewer Support on Children’s Memory and Suggestibility: Systematic Review and Meta-Analyses of Experimental Research. Trauma Violence Abuse. 2019 Jan;20(1):22-39. [PubMed]19.

Jordan KS, Steelman SH, Leary M, Varela-Gonzalez L, Lassiter SL, Montminy L, Bellow EF. Pediatric Sexual Abuse: An Interprofessional Approach to Optimizing Emergency Care. J Forensic Nurs. 2019 Jan/Mar;15(1):18-25. [PubMed]20.

US Preventive Services Task Force. Curry SJ, Krist AH, Owens DK, Barry MJ, Caughey AB, Davidson KW, Doubeni CA, Epling JW, Grossman DC, Kemper AR, Kubik M, Landefeld CS, Mangione CM, Silverstein M, Simon MA, Tseng CW, Wong JB. Interventions to Prevent Child Maltreatment: US Preventive Services Task Force Recommendation Statement. JAMA. 2018 Nov 27;320(20):2122-2128. 

Wardle D, Finnerty F. Work of the BASHH Sexual Violence Special Interest Group. Sex Transm Infect. 2018 Dec;94(8):552. 

Hawcroft C, Hughes R, Shaheen A, Usta J, Elkadi H, Dalton T, Ginwalla K, Feder G. Prevalence and health outcomes of domestic violence amongst clinical populations in Arab countries: a systematic review and meta-analysis. BMC Public Health. 2019 Mar 18;19(1):315. 

Berhanie E, Gebregziabher D, Berihu H, Gerezgiher A, Kidane G. Intimate partner violence during pregnancy and adverse birth outcomes: a case-control study. Reprod Health. 2019 Feb 25;16(1):22.

© 2022, StatPearls Publishing LLC.

This book is distributed under the terms of the Creative Commons Attribution 4.0 International License, which permits use, duplication, adaptation, distribution, and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, a link is provided to the Creative Commons license, and any changes made are indicated.

Bookshelf ID: NBK499891PMID: 29763066

3. New Method to Gather Domestic Violence Statistics: New Internet Search Algorithm Shows up to 8x More Lockdown-Related Incidents than London Police Reports

“Our findings show the London lockdown led to a gradual increase in the DV-related crimes recorded by the MPS and the effect of the lockdown remained positive until mid-June. The impact was nevertheless modest, with about 10–15 extra DV crimes per day relative to a normal average of over 200 crimes per day. In sharp contrast, although exhibiting a similar lockdown timing structure, we find a 40% increase (at peak) in our search-based DV index, seven to eight times larger than the increase in police-recorded crimes and much closer to the increase reported by the UK’s National Domestic Abuse Helpline in relation to helpline calls and contacts.

The broader lesson from our analysis is that it cautions against relying solely on police-recorded crimes or calls-for-service to assess the scale of the DV problem during crises like COVID-19. In such assessments, the use of complementary data sources is important, as it would allow researchers to move towards demarcating the lower and upper bounds of likely impacts on DV. One promising avenue in this respect is to engage with organizations supporting DV victims, encouraging the collection and provision of systematic data from DV helplines. Our algorithm for measuring temporal variation in DV incidence using internet search activity provides another viable strategy to complement assessments based on police records. Although our analysis by no means provides a definite answer to how to best construct a real-time indicator of DV, it can hopefully serve as a starting point that can be extended and further validated.”

Abstract

In contrast to widespread concerns that COVID-19 lockdowns have substantially increased the incidence of domestic violence, research based on police-recorded crimes or calls-for-service has typically found small and often even negligible effects. One explanation for this discrepancy is that lockdowns have left victims of domestic violence trapped in-home with their perpetrators, limiting their ability to safely report incidents to the police. To overcome this measurement problem, we propose a model-based algorithm for measuring temporal variation in domestic violence incidence using internet search activity and make precise the conditions under which this measure yields less biased estimates of domestic violence problem during periods of crisis than commonly used police-recorded crime measures. Analysing the COVID-19 lockdown in Greater London, we find a 40% increase in our internet search-based domestic violence index at the peak occurring 3–6 weeks into the lockdown, -seven to eight times larger than the increase in police-recorded crimes and much closer to the increase in helpline calls reported by victim support charities. Applying the same methodology to Los Angeles, we find strikingly similar results. We conclude that evidence based solely on police-recorded domestic violence incidents cannot reliably inform us about the scale of the domestic violence problem during crises like COVID-19.

1 INTRODUCTION

During the COVID-19 pandemic, there has been a major discrepancy between crisis-induced surges in domestic violence as perceived by practitioners in the field and the effects reported in empirical studies based on police records of domestic violence incidents. Reports from women’s support charities, domestic abuse helplines and frontline workers in countries such as Australia, France, the United Kingdom, the United States and China raised significant concerns from an early stage of the crisis, suggesting increases in domestic violence help seeking following the implementation of self-isolation and quarantine measures of anywhere between 25% and 80% (see, e.g. Allen-Ebrahimian, 2020; Human Rights Watch, 2020; UN Women, 2020; Wagers, 2020). Yet, in stark contrast to these alarming numbers, recent empirical studies exploiting police records of domestic violence incidents have found either relatively modest or no increases in family violence following lockdowns and self-isolation (Campedelli et al., 2021; Ivandic et al., 2021; Leslie & Wilson, 2020; McCrary & Sanga, 2020; Mohler et al., 2020; Payne & Morgan, 2020; Piquero et al., 2020). Indeed, a recent systematic review and meta-analysis (Piquero et al., 2021) combined 37 estimates from a total of 18 studies of the change in domestic violence (henceforth, DV) incidence during the early stages of the pandemic. The majority of the studies included in the review were based on police records, while a few of the included studies used data from health records and hotline registers. Eight of the included estimates showed a decrease in DV incidence while 29 showed an increase (with an overall average of about 8%), thus highlighting the contradictory nature of evidence in the literature.

Against this background, this paper has two objectives: (1) to highlight the potential limitations and biases in using police data to quantify the scale of the domestic violence problem during crisis, and (2) to propose an algorithmic methodology for measuring short- to medium-term temporal variation in domestic violence incidence based on internet search data.

From a policy perspective, there is an urgent need to quantify the impact of crises like COVID-19 pandemic on DV: at times where governments face unprecedented demands on limited resources, optimal policy responses to support victims of DV can only be implemented if the scale of the problem is known. However, the quantification of the prevalence of DV is difficult at the best of times due to data limitations, and the pandemic has exacerbated this difficulty in various ways. Victimization surveys have, under normal circumstances, become an accepted way of estimating prevalence rates for DV. However, these surveys are neither available in real time nor do they provide temporally granular enough information to adequately analyse the consequences of the COVID-19 crisis. By contrast, police records of DV incidents are often available at daily frequencies and even in real time, and in many cases contain fine-level information on location. We present evidence, based on daily counts of DV-related crimes recorded by the London Metropolitan Police Service (MPS) and a simple regression accounting for an overall trend, seasonality, day-of-week and weather effects, that the London lockdown in the spring of 2020 brought about an increase in recorded DV crimes of around 5%–7% (at peak) compared to levels before the pandemic.

However, the vast majority of victims of DV do not report these crimes to the police (see, e.g. Podaná, 2010; UN Women, 2020) and, importantly, there is every reason to believe that reporting behaviour itself would have been significantly affected by quarantines and self-isolation. For example, recent evidence presented by Campbell et al. (2020) shows that among DV victims who decide to contact the police for help, a large portion report waiting for the perpetrator to leave the scene before calling 911. The pandemic and associated lockdown measures implemented by many governments conceivably left victims of DV trapped in-home with their perpetrators, limiting their opportunity to safely report incidents to the police (Campbell, 2020; Kofman & Garfin, 2020). Thus, any analysis of police-recorded DV incidents runs the risk of underestimating the DV problem during crisis and lockdown. Help seeking behaviours other than police contact, although also having become more difficult, are likely to have been less affected by self-isolation and quarantine measures, as they generally allow for more anonymity and carry less consequences for both victim and perpetrator.

To complement available data sources, we propose a simple model that (i) delivers an algorithmic methodology for measuring temporal variation in DV incidence based on DV-related internet search activity, and (ii) makes precise the conditions under which this measure provides us with a less biased estimation of the DV problem during the crisis than traditional, police-recorded crime measures. Our approach uses pre-crisis data—in our case over 5 years—to relate daily internet search activity for DV-related terms to daily police-recorded DV incidents (both observed). The intuition for the approach is that both reflect the same underlying (unobserved) temporal variation in DV incidence, leading to a positive correlation that is stronger for the most relevant/least noisy internet search terms. Our algorithmic design further accounts for differential trends, seasonality and searches occurring on days contiguous to the underlying incident. More critically, it allows us to use estimated signal-to-noise ratios to create a composite measure of DV-related search activity, which we interpret as a search-based DV index. There are two conditions under which this measure yields estimates of the DV problem during this pandemic that are less downwardly biased than those based on police-recorded crime data: quarantine and self-isolation measures have made help seeking generally more difficult for DV victims, and has hampered help seeking through police relatively more than via the internet.

We present four results. First, after using the pre-2020 data to train our algorithm, we use the first 75 days of 2020 to test the validity of DV index. We find that our search-based index exhibits positive correlation and simple co-movement with the recorded DV crimes. Further verification of validity of our approach is offered by the fact that our search-based DV index and the police-recorded DV crimes exhibit similar relationships to a key time-varying exogenous factor: weather (see, e.g. Butke & Sheridan, 2010). Reassuringly, we find that higher temperatures are not only significant predictors of DV-incidents recorded by the London MPS but are also highly correlated with our search-based DV index. Second, analysing the London lockdown of spring 2020, we observe a closely aligned timing of increases in DV incidents recorded by the London MPS and increases in our composite DV index: while the lockdown had no immediate impact, a significant effect emerged somewhere between 3 and 6 weeks into the lockdown. Third, in level terms, however, we find a 40% increase (at peak) in our search-based DV index, -seven to eight times larger than the increase in police-recorded crimes but only about half the size of the increase noted for helpline contacts. Fourth, replicating our results for London using daily police and internet search data for the city of Los Angeles, California, we obtain qualitatively remarkably similar results.

Research use of Google data has expanded rapidly in the last decade. There are three broad circumstances—all relevant here—where such data have proven particularly useful. First, where there are issues in relation to obtaining accurate/truthful reporting. For instance, Stephens-Davidowitz (2014) used Google search data on terms involving racially charged language and showed that the local racially charged search rate is a robust negative predictor of Obama’s vote share. Second, where it is possible to identify search terms that are specific to the topic of interest, and when the act of searching itself may be closely related to it. For instance, noting that job search activity is notoriously difficult to measure, Baker and Fradkin (2017) develop a job search index based on Google searches for both broader terms such as ‘jobs’ and specific terms such as ‘Dallas jobs’ or ‘tech jobs’. Third, when there is a need to track developments in real time. An early contribution showing the potential for ‘nowcasting’ across multiple settings using Google search data was Choi and Varian (2020). More recently, Ferrara and Simoni (2020) show that Google data can be provide information for nowcasting GPD until official macroeconomic information becomes available.

2 SETTING AND STATISTICAL FRAMEWORK

Starting mid-March 2020 the UK government implemented a string of measures to limit the spread of the coronavirus. On March 16, the Prime Minister announced that everyone should begin social distancing. Later the same week, schools, theatres, nightclubs, cinemas, gyms and leisure centres were ordered to close. Finally, on the evening of March 23, a stay-at-home order effective immediately was announced and all non-essential shops and services were ordered to close. The police were granted powers to issue fines and send people home. The impact of the policy measures on people’s movement was strong. A sharp drop in mobility followed after social distancing was announced, and after the announcement of the full stay-at-home order, mobility was down to 10%–20% of the pre-lockdown level (Batty et al., 2021).1 The easing of the lockdown was gradual from mid-May. Nevertheless, mobility remained below 50% of pre-lockdown levels through to the end of June. Our focus will be on this initial lockdown with our sample period running through to 22 June 2020. Hence we will focus on this initial crisis response as an event study and our framework presented below will correspondingly be cast in terms of a pre-lockdown and a lockdown regime.

After the initial lockdown, there have been two further national lockdowns in October of 2020 and in January of 2021 (mixed in with regionally implemented policy measures for most of 2020). However, the impact on mobility of each of these subsequent lockdown was substantially lower than the initial spring 2020 lockdown (see previous footnote for reference).

We now set out a simple framework that gives rise to an algorithmic methodology for measuring temporal variation in DV incidence based police-reported DV crimes and DV-related internet search data.

2.1 Setup

Let t ∈ {1, …, T} denote time, where the unit of time is a day. Lockdown occurs at some time t0 and continues to the end of the sample period. Hence the overall sample period is split into two regimes, R  ∈  {0, 1}, with Rt=0 (pre-lockdown) if t<t0 and Rt=1 (lockdown) if t≥t0. Let nt denote the number of DV incidents/victims at time t. Although not directly observed, nt will have some distribution, and the concern is that this will have changed with the lockdown. Hence let fR(n) be the probability mass function for n in regime R.

A given victim of abuse i at time t, may seek help through alternative routes. Let pit∈{0,1} indicate whether she contacts the police, leading to a recorded DV crime. Similarly, let yit∈{0,1} denote whether she seeks support via an internet search. Note that the two help seeking responses are not mutually exclusive: for any given victim, either none, either, or both help seeking actions may occur. We will, however, in the following assume that pit and yit are statistically independent. A reason why this may not be realistic is that victims may use Google searches to find information about how to contact the police. This motivates one of our robustness checks (see Section 4.4) in which we remove all search terms directly related to police. For DV, repeat victimization is also common whereby the same individual i may seek help on multiple occasions through alternative channels. This could potentially cause correlation between pit and yit′ within an individual across different points in time, and hence between Pt and Yt′. As our algorithm will focus on same (or contiguous) day correlations, any potential autocorrelation stemming from repeat victimization will not influence our analysis.

In the data we observe the daily count of incidents recorded by the police. This is, we observe Pt≡∑i=1ntpit. Similarly, assume for now that we also observe the daily search intensity Yt≡∑i=1ntyit. One of the issues below will be the construction of the measure Yt.

2.2 Help seeking behaviour across regimes

Each help seeking behaviour is guided by the net benefit to victim i from taking that action, which may be regime specific. Hence let VkR denote the net systematic (or ‘common’) benefit to a victim from taking action k  ∈  {p, y} in regime R  ∈  {0, 1}.

Additionally, a given victim i obtains an individual-specific additive random utility εi1k

 from taking action k and an additive random utility εi0k

 of not taking action k which are assumed to be i.i.d. extreme value distributed across individuals and actions. Under these assumptions the probability of any given victim i in regime R taking action k will take the standard logit form,

πkR=Prkit=1|R=expVkR1+expVkR,fork∈{p,y}andR∈{0,1}.(1)

As specified, the action kit∈{0,1}

 is independent across i for any given t. Moreover, as the probability πkR

 of taking this action is common across i it follows that both count variables Pt

 and Yt

 are, given nt

, both binomially distributed and independent of each other.

A key issue is that the perceived individual benefit from seeking help, and hence help seeking behaviour as represented by πkR, may differ across the regimes, making it challenging to infer the change in DV incidence from help seeking data. For instance, only if Vp1=Vp0, and hence πp1=πp0, will the observed proportional change in Pt, accurately reflect the proportional change in the DV incidence level. A similar argument of course applies to help seeking via the internet. In general, we cannot a priori assume that Vk1=Vk0 for either action. Under the weaker assumption that the lockdown measures made help seeking generally more difficult for victims, whereby ΔVk≡Vk1−Vk0≤0 for both actions k = {p, y}, the observed proportional change in either action serves as a lower bound for the underlying proportional change in abuse incidence. Moreover, if help seeking via the police was discouraged relatively more, ΔVp<ΔVy, then the proportional change in help seeking via the internet provides a less downwardly biased estimator of the change in DV incidence.

A potential threat to the assumption ΔVk≤0 for both actions would be ‘substitutability’: if the lockdown decreased the perceived benefit to contacting the police, this could potentially have shifted help seeking onto alternative routes.

Further related concerns include the possibilities that the lockdown meant that individuals had more time available for doing internet searches and that, as a result of people spending more time at home, DV incidence became more visible to neighbours and leading to more third-party reporting and searches. We will return to discuss these potential caveats below.

2.3 Relating internet searches to police reports

Daily counts of police-recorded DV crimes and daily search activity will be correlated as both reflect the same underlying (unobserved) temporal variation in DV incidence. To see this, consider the covariance between Pt

 and Yt

 within either given regime Rt∈{0,1}

. Using that Pt

 and Yt

 are, conditional on nt

, both binomially distributed and independent, and using the law of iterated expectations, it is easily shown that,

CovPt,Yt|Rt=πpRtπyRtVarnt|Rt>0.(2)

Intuitively, Pt

 and Yt

 are positively correlated as both tend to be large on days when DV incidence nt

 is large. In our empirical application, Pt

 and Yt

 will both be used in index form. As this merely re-scales each by a multiplicative constant, the statistical properties are preserved.

In practice, we observe daily search intensities Yjt (in index form) for a set J of DV-related search terms. Hence, in order to create a single composite measure Yt we need to apportion relative weight across the various terms. To do so, we will use pre-lockdown data and draw on (2). This equation can be taken to apply for each term j  ∈  J, whereby the relative covariances of the various Yjt’s with Pt will indicate the relative frequency with which victims use the J terms: using πjy0 to denote the pre-lockdown propensity for a victim to search on term j  ∈  J it follows from Equation (2) that for two alternative terms, j and j′, Cov(Pt,Yj′t|Rt=0)/Cov(Pt,Yjt|Rt=0)=πj′y0/πjy0. Note that conditioning on pre-lockdown in central here: if data were pooled across regimes, the relative covariance Cov(Pt,Yjt)/Cov(Pt,Yj′t) would only correspond to the relative search frequency if πjyR and πj′yR remained constant not only in relative, but also in absolute terms. For this reason, we will use pre-lockdown data to construct our composite DV index.

However, measured search intensities can be expected to contain a fair amount of noise, for example, due to random searches by non-victims. Hence consider the regression specification,

Yjt=αj+λjPt+vjt,forj∈J,(3)

where vjt

 represents noise. The ordinary least squares estimator of λj

 is of course λ^j=Cov^(Pt,Yjt)/Var^(Pt)

. Applying this on pre-lockdown data will allow us to identify search terms that are more commonly used by victims—as indicated by their relative values of λ^j

—and that contain relatively less noise. We will use this approach to construct our measure Yt

. In particular, we will use pre-lockdown data to estimate (a version of) Equation (3) for each j  ∈  J, and terms with an estimated positive correlation, λ^j>0

, will be given a weight in the composite index that corresponds to its signal-to-noise ratio.

2.4 Algorithm

The exact algorithm used in constructing the composite index Yt accounts for two further complications. First, to account for the possibility that police reports and internet searches have different growth over time, seasonality etc., rather than directly relating Yjt to Pt, we relate the unexpected component of Yjt to the corresponding unexpected component of Pt after removing year-, month- and day-of-the-week effects. Second, while victims can be expected to contact the police at the time of a DV incident, on-line help seeking may be distributed around the time of the event, either in the days following the event or, if tensions are building in advance, in the days before. To account for this, we relate the unexpected component of Pt to the unexpected components of Yj,t+s for a set of days around t.

To implement our algorithm, we use data on daily counts of DV-related crimes, Pt

, recorded by the London MPS, from 1 April 2015 to 22 June, 2020, and presented in detail in the following section. With the lockdown occurring on March 23 (after an initial announcement a week earlier, see above), this effectively gives us five pre-crisis years and three lockdown months. In order to test the validity of our search-based DV index, we ‘train’ our algorithm on data up to the end of 2019, and use the first 75 days of 2020 as ‘testing’ period. As for internet search data, we select a set J potentially DV-relevant search terms, also presented below. For each search term j  ∈  J, we used Google Trends to generate a daily search index Yjt

, spanning our full sample period. Using these remaining terms, we apply the following algorithm.

1. We regress Pt, on year-, month- and day-of-the-week dummies using pre-2020 data, and obtain the residual, denoted ε^t. These represent the unexpected daily variation in DV crimes.
2. We correspondingly regress each search term intensity Yjt, j ∈ J, on year-, month- and day-of-the-week dummies and obtain the residuals, denoted ϵ^jt. These represent the unexpected daily variation in the search intensity for term j.
3. Still using pre-2020 data, we relate ε^t to ϵ^j,t+s for a set of ±K days around t by estimating ϵ^j,t+s=αjs+λjsε^t+ωj,t+s for each j ∈ J and s ∈ {−K, …, +K}, and we compute ( j, s)-specific signal-to-noise ratios, denoted σjs=(λ^js)2Var^(ε^t)/[(λ^js)2Var^(ε^t)+Var^(ωj,t+s)].
4. Using the estimated signal-to-noise ratios as weights we construct a composite index, Yt=∑j∈J∑sσjsYj,t+s, from the individual search terms for the full sample period.

The final daily composite index Yt

 is therefore a weighted average of the original J search indices, along with their leads and lags. In our leading case, we use a window of ±3 days which has the advantage that the index value on any given day t will reflect searches done on all days of the week. In our leading case, we thus estimate 23 × 7 = 161 signal-to-noise ratios and just over two-thirds (110) of these was positive and hence used in construction of the composite index.

We re-scale Yt to have a mean of 100 over the ‘training period’ (1 April 2015 to 31 December 2019). For ease of comparison, we also re-scale Pt to have a mean of 100 over the same period.

3 Data

The main data sources used for the current analysis are thus daily data on DV crimes recorded by the London MPS and a data on DV-related Google searches.

MPS Domestic Violence Crime Data

Data on the daily count of DV-related crimes recorded by the MPS were obtained by a Freedom of Information request. As noted above, our data covers the period 1 April 2015, to 22 June 2020. The data exhibit some general time patterns. Figure 1 shows the average daily count of DV-related crimes by year, month and day of the week. Panel A shows that the daily average has increased from about 205 in 2015 to 245 in 2019, which corresponds to an average annual increase of 4.5%. The data for 2020 cover only the time up to 22 June and, of course, incorporates the first lockdown period. The steady growth over time makes simple comparisons—for instance comparing a given week to the corresponding week a year before—somewhat problematic. Panel B shows a strong seasonal pattern, with reported DV incidence being lower in the first and fourth quarter and higher between late spring and end of summer. Panel C shows a strong day-of-the-week pattern, with incidence being about 10% higher on weekends than during weekdays. Finally, panel D shows the daily counts from 1 February 2020 to the end of the sample period.

3.1 Google search data

Daily data for a set J of 35 DV-related search terms were obtained. The initial selection of terms was made to be deliberately broad in order to subsequently narrow the set down by studying their variation and covariation with the DV-crime data. Based on listing of terms commonly used in relation to DV (NCDSV, 2017), terms were selected to cover three broad categories:

1. Terms that relate to general help seeking from helplines and charities.
2. Terms that describe abusive relationships and forms of abuse
3. Terms that relate to police and legal protection.

A complication when using daily Google search data is that daily data are only available for search windows up to 9 months. In the online Appendix, we describe how daily data series are generated for our full sample period. Google Trends provides search intensities in index form with values between 0 and 100 where a value of 0 is given for terms/days with low search volume. Twelve of the 35 terms had zeros on majority of days and thus low variation; these terms eliminated, reducing our set to 23 terms. Table 1 lists all terms used and which terms had ‘high/low’ variation.TABLE 1. Selection of search terms

Search term	Daily variation	Relative weight	Search term	Daily variation	Relative weight
Group 1: Seeking support	Group 2: Searching on abuse
Abuse help	High	0.023	Abusive partner	High	0.200
Abuse helpline	High	1.268	Abusive relationship	High	2.884
Abuse support	High	0.307	Threat of violence	Low	–
Refuge	High	1.294	Partner violence	Low	–
Women’s refuge	High	0.280	Domestic violence	High	4.207
Refuge helpline	Low	–	Domestic abuse	High	3.317
Refuge centre	Low	–	Emotional abuse	High	1.184
London refuge	Low	–	Psychological abuse	High	1.625
Violence refuge	Low	–	Controlling relationship	High	0.735
Shelter	How	0.715	Coercive control	High	0.250
London abuse	High	0.057	Group 3: Police/legal protection
Women’s aid	High	0.635	Domestic violence protection	Low	–
Victim support	High	0.042	Report domestic abuse	Low	–
National domestic violence helpline	Low	–	Abuse police	High	0.718
Domestic abuse charity	Low	–	Abuse protection	High	0.302
Domestic violence support	High	0.269	Reporting abuse	High	0.303
Domestic violence help	Low	–	Domestic violence police	High	0.809
			Domestic violence law	High	1.575
			Domestic violence charges	Low	–

• Notes: The tables lists the Google search terms used in the construction of the composite DV-search intensity index. The daily variation for a given search term is classified as ‘Low’ (‘High’) if it contains zeros on more (less) than half of all days. For terms with high variation, the table reports the relative weight place on that term, averaged over the ±K days used in the construction of the composite index.

Turning to the covariation with the DV-crime data and the implementation of our algorithm, our main specification includes three leads/lags, that is s in the range ±3. The (relative) weight placed on term j and day s in the construction of the composite index was σ˜js=σjs/[∑j′∈J0∑sσjs′]. For each of the 23 terms with high daily variation, the table shows its relative weight, averaged over days s. The search terms that, based on their covariation with recorded DV crimes, got the highest weight in our internet search-based DV index include ‘abuse helpline’, ‘domestic violence’, ‘domestic abuse’, ‘abusive relationship’, ‘emotional abuse’, ‘psychological abuse’ and ‘domestic violence law’.

It should be noted that while we use DV-crime data from the London MPS, the Google Trends data are for England. There are two reasons why our methods can be expected to be robust to this geographical discrepancy. First, the MPS is by far the largest territorial police force in England, covering over 8 million people, or about 15% of the entire population of England. Second, while our algorithm relates the unexpected daily variation in Google searches to the unexpected daily variation in DV crimes, these ‘unexpected’ components are in relation to the year, month and day-of-the-week that are controlled for. In fact, many of the days with high levels of DV crimes are highly predictable and include, for instance, all New Year’s Days, many bank holiday weekends etc. and these are, of course, common across the whole of England. Hence, one way to view the algorithm is that it uses the crime data to statistically identify high-risk days and then identifies search terms that spike on nearby days.

The lockdown that is the subject of the current study was of course national. Nevertheless, when comparing the DV crimes recorded by the MPS and DV-related Google search intensities for England, the caveat remains that the impact of the lockdown may have been different in London compared to the rest of the country. For instance, London being an urban area naturally has a high population density which may affect both DV incidence and its reporting (Peek-Asa et al., 2011). Related to the lockdown (but not DV), it has been shown by Sun et al. (2021) that even within London, local crime rates—for instance for robbery, burglary and theft—in March through May 2020, had some association with the local COVID-19 infection rate. Similarly, Campedelli et al. (2020) found that COVID-19 containment policies in Chicago impacted crime in different ways across local areas.

Weather Data

In our analysis below we will further account for weather as a factor affecting DV incidence. We use data on daily average temperature (in ∘C) and rainfall (in mm) from the London Heathrow weather station covering the full sample period, obtained from the National Climatic Data Centre. As noted, April and May of this year were unusually warm and dry. Panel A of Figure 2 shows the daily average temperature (in ∘C) with the horizontal red lines indicating the average temperature by month over the past 5 years. The second half of May was also unusually warm. Panel B shows rainfall per day, indicating that the key period from the beginning of the lockdown through to early June saw barely any rainfall at all.

4 RESULTS

The main aim of this section is to compare how the effect of lockdown on the two different measures of DV incidence, the police-reported crimes and our search-based DV index. But we will start by first present the results from a validation exercise and some descriptive analysis of how the measures compare. Later on we will also present a set of robustness checks and the results from a corresponding analysis using data from Los Angeles.

4.1 Validation

Since we used pre-2020 data to train our algorithm we can use pre-lockdown 2020 data for checking the validity of our index as a measurement of temporal variation in DV incidence. In particular, we will use the first 75 days of 2020, from 1 January to March 15, as our testing period over which we can check that our search-based DV index has predictive power for recorded DV crimes.

In Figure 3 we plot the residualized daily series for the DV index and the DV crimes, in each case having removed year-, month- and day-of-the-week fixed effects to account for potentially different trends and seasonality (see above). Panel (A) plots the two series for the training period from 1 April 2015 to 31 December 2019. The figure visually highlights co-movement between the two series, with a simple Pearson correlation coefficient of 0.27. The figure highlights key spikes in the DV-crime series around each New Year. Panel (B) plots the continuation of the same residualized series for the testing period of 1 January to 15 March 2020. Simple visual inspection suggests that the two series remain highly correlated. Indeed, the correlation is 0.23 (when leaving out New Year’s Day), which is only marginally lower than for the training period.

In the Online Appendix (see Table A.1), we provide further evidence on the synchrony of the two series within the testing period by regressing the daily DV crimes on the DV index. We present regressions both in levels and first-difference form and specifications that include a lead and a lag of the DV index. The results are particularly strong in the first-difference form, highlighting that day-to-day changes in the DV index is associated with same day changes in DV crimes.

4.2 Descriptive evidence

In Figure 4 we provide some first descriptive evidence of how the search-based DV index and the recorded DV crimes compare over the lockdown. In Panel A we first show the continuation, into the lockdown, of the series plotted in Figure 3 above, that is, the residualized DV index and DV crimes. This shows that, while the two series followed each other closely over the testing period, this feature breaks down after mid-March of 2020.

In order to show this divergence more clearly we collapse the daily counts of DV-related crimes recorded by the London MPS to the weekly level, and plot average daily DV crimes between 1 February and 22 June 2020 (Panel B). The figure suggests that the London lockdown was associated with a steady increase in DV crimes starting after 1 April and continuing through to the end of May, with a peak increase in slightly below 20% compared to pre-lockdown levels.

In the same panel we contrast the MPS data with our search-based DV index and with data on helpline calls and contacts obtained by the UK’s National Domestic Abuse Helpline. Compared to police-recorded DV crimes, the increase in the search-based DV index after lockdown measures were implemented was substantially larger and sharper. Indeed, after an initial drop for the 2 weeks starting 16 March (the day of the first announcement about social distancing) and 23 March (the start of the official stay-at-home order), the search-based DV index strongly increased early in April, peaking at around 35% above pre-lockdown levels throughout the entire month. Strikingly, the post-lockdown increase in our search-based DV index closely follows the increase reported by the UK’s National Domestic Abuse Helpline in relation to helpline contacts and calls. However, whereas the search-based DV index increased by roughly 35% at peak, helpline contacts increased in the order of roughly 60% compared to levels before the London lockdown.

The evidence presented here is, however, purely descriptive. It is well understood that intimate partner violence exhibits seasonal variation, with DV incidents more likely to occur during the summer months, starting in May (see, e.g., Campbell et al., 2020). Relatedly, empirical research investigating the relationship between weather and crime shows that temperature is positively correlated with aggressive behaviour, especially domestic violence (see, e.g. Butke & Sheridan, 2010; Sanz-Barbero et al., 2018). Thus, in assessing the impact of the pandemic and associated lockdown measures, it is important to account for time and meteorological effects. This eliminates one of our data sources—information on helpline contacts—from any further analysis, since it was made available to us for a very limited time span only (10 February to 4 May 2020) and at lower temporal granularity.

4.3 Regression analysis

To assess the impact of the London lockdown, we estimate a regression that accounts for an overall trend, seasonality and day-of-the-week effects. Hence our model for outcome Dt∈{Pt,Yt}

 is given by:

Dt=α+βy+γm+δd+ζxt+ft−t0It≥t0+εt,t=1,…T,(4)

where βy

, γm

 and δd

 are year-, month- and day-of-the-week fixed effects, controlling for a trend, seasonality and weekly cycles respectively. Moreover, and as mentioned above, one factor that may have played a role was the weather. Hot weather is a well-documented factor that increases the DV incidence, and London saw a particularly warm and dry April in 2020. To account for this, we use data on daily average temperature (in ∘C

) and rainfall (in mm) in London over the sample period, and xt

 thus includes controls for temperature and rainfall. Turning to the lockdown, It≥t0

 is a dummy for t being within the lockdown period, and f(t−t0)

 is a flexible, but continuous, function of lockdown duration. Note that f(0) is not restricted to be zero. Hence it allows for an immediate lockdown effect. Our baseline specification for f(·) is a quadratic function, possibly with a distinct effect at weekends,

f(τ)=ϕ0+ϕ1τ+ϕ2τ2+ϕ3Iwkend,(5)

where Iwkend

 is a weekend (Saturday/Sunday) indicator.

The first three columns of Table 2 present estimates of (4) using daily MPS counts of DV crimes as dependent variable. In column (i), we estimate a basic version of (4), ignoring weather and separate weekend effects. The estimates suggest, if anything, a negative immediate effect. However, a positive effect emerged over the following weeks and peaked after about 50 days of lockdown (=−ϕ1/(2ϕ2)), aligning well with the visual impression from Panel A of Figure 4. The finding that the impact of the lockdown grew with duration naturally accords with General Strain Theory (Agnew, 1992) which would suggest that reduced freedom of mobility, increased uncertainty, financial and emotional pressure could lead to the build-up of negative emotions, leading to a gradually increasing impact on DV incidence. The estimated coefficients imply an increase in DV crimes of around 5% at the peak compared to pre-lockdown levels. In column (ii), we add controls for weather, confirming a strong effect of temperature: a one degree Celsius increase in the daily (average) temperature is associated with a 0.8% increase in DV crimes per day. Rainfall is estimated to have a negative, but less precisely estimated, impact. However, the prolonged period of above-average temperature and dry weather observed for April and May only accounts for a small part of the rise in reported DV crimes during the lockdown period. Finally, in column (iii), we allow for the lockdown to have a differential effect on weekends. The negative effect here indicates that recorded DV crimes during the lockdown had a much smaller weekend–weekday difference than pre-crisis. The estimated coefficient translates into about 15 DV crimes per day, implying that the weekend–weekday difference during the lockdown was only about half of pre-crisis difference (see Figure 1). However, our coefficients of interest are barely affected by the inclusion of weekend effects.TABLE 2. The effect of the london lockdown on domestic violence

	Police-recorded DV incidents	Search-based DV index
	(i)	(ii)	(iii)	(iv)	(v)	(vi)
Lockdown (ϕ0)	−7.778**	−6.411*	−4.710	−7.417	−6.724	−6.745
	(3.804)	(3.695)	(3.622)	(7.667)	(7.590)	(7.721)
Days of Lockdown (ϕ1)	0.435**	0.371**	0.380**	2.182***	2.159***	2.159***
	(0.170)	(0.162)	(0.158)	(0.323)	(0.322)	(0.322)
Days Sq. (ϕ2)	−0.00405**	−0.00329**	−0.00333**	−0.0238***	−0.0235***	−0.0235***
	(0.00165)	(0.00158)	(0.00158)	(0.00318)	(0.00318)	(0.00318)
Temperature (∘C)		0.843***	0.845***		0.293***	0.292***
		(0.0685)	(0.0680)		(0.103)	(0.103)
Precipitation (mm)	−3.063**	−2.989**		2.076	2.075
		(1.525)	(1.522)		(1.660)	(1.662)
Weekend × Lockdown (ϕ3)		−7.138***				0.0878
			(2.293)			(4.263)
Observations	1,910	1,910	1,910	1,905	1,905	1,905

• Notes: Standard errors in parentheses
• *p < 0.10, **p < 0.05, ***p < 0.01
• The dependent variable in specifications (i) to (iii) is the daily count of DV-related crimes recorded by the London MPS between 1 April 2015 and 22 June 2020 in index form (100  =  average daily count over the period 1 April 2015 to 31 December 2019). The dependent variable in specifications (iv) to (vi) is a composite index of DV-related search intensity at daily frequency (100  =  average daily intensity over the period 1 April 2015 to 31 December 2019). All regressions include year-, month- and day-of-the-week fixed effects.

In the last three columns of Table 2, we re-estimate (4) with our internet search-based DV index as dependent variable. The estimates in column (iv) suggest that, while the timing of the increase in DV search intensity is more or less identical to that of DV crimes (both peaking roughly after 50 days of lockdown), the magnitude of the increase is about seven to eight times larger with an estimated increase of about 35%–40% at the peak. In column (v), we further control for weather. Here we find, reassuringly, that higher temperatures generate more DV-related searches according to our composite index: a one degree Celsius increase in the daily temperature is associated with close to a 0.3% increase in DV-related searches. The estimated effect of rainfall is, in contrast, highly imprecise. The lower estimates and precision is natural given that the weather measurements are local to London whereas the search data are for the whole of England. In column (vi), we find no indication of any differential impact of the lockdown on DV-related searches on weekends versus weekdays.

In order to avoid influence of the parametric form, we next replace the function f(t−t0) with a set of dummies for 2-week periods relative to the time of lockdown, starting with the 2 weeks leading up to the formal lockdown.

The results are presented Figure 5. Panel (A) shows that there was no significant effect on the number of DV-related crimes recorded by the MPS early on in the lockdown. There was, however, a significant increase in recorded DV crimes from the end of April through into early June (weeks 5–10). Nevertheless, the estimated effects are generally quite small—an increase of 5%–7%, at peak, compared to the pre-lockdown average. Panel (B) shows the corresponding estimates for the search-based DV index. The two figures again exhibit similar timing, suggesting that the first few weeks of the lockdown remained relatively quiet. However, after that, our index shows a sharp increase approaching mid-April. At this stage, DV-related searches were about 40% higher than their pre-lockdown average. Over the following 2 months, searches gradually fell back down towards pre-lockdown levels, but remained significantly above the pre-lockdown level.

Relating back to the conceptual framework from Section 2.1, if we were to assume that the increase in the search-based DV index accurately reflects the impact of the lockdown on DV incidence, whereas the lower increase in DV-related crimes reflects a reduced reporting rate by victims, then we can make a back-of-the-envelope estimation of the ‘missing’ number of recorded DV crimes over the sample period. Specifically, under this assumption and using that the average difference between the estimated effects in Figure 5 is 22% and the average number of recorded DV crimes per day over the baseline period is 221, this suggests that there were on average 48 missing recorded DV crimes per day. Hence the model predicts that the MPS would have recorded a further 4,700 DV crimes over the sample period had the rate of reporting to the police itself not been lowered by the lockdown.

4.4 Robustness checks

We have carried out a number of sensitivity checks in relation to our estimates for the search-based DV index which we describe in Table A.2 in the online Appendix.

As a first set of checks, we have verified that the results are robust to the choice of window used in the construction of Yt: while our main specification (reiterated in column 1 of Table A.2) uses ±3 days, columns 2–4 shows that the results are robust to using ±2 and ±1, and also to using only three lags and no leads. In each of these alternative specifications, the key pattern on the estimated coefficients remains: the point estimates of the immediate effect of the lockdown ϕ0 are all negative but not significant, while the estimated effects of the lockdown duration, ϕ1 and ϕ2, remain positive and negative respectively and within ±20% relative to our main estimates and always highly statistically significant.

Our second set of checks concerned the choice of terms and we present three alternative specifications (columns 5–7). First, as the potential impact of the lockdown on domestic violence received a substantial amount of media attention during the spring lockdown, we checked that our results are robust to excluding the generic term ‘domestic violence’ from the composite measure as it was commonly used in media and hence a natural ‘focal term’ for anyone doing internet searches to follow-up on these news stories.2

Second, as noted above, the key relationship (2) between help seeking via the police Pt and internet searches Yt that formed the basis for our algorithm was derived under the assumption of independence between pit and yit. This assumption could easily be violated if victims of abuse use internet searches for information on potential police protection. Hence we present a specification where we exclude all terms involving the word ‘police’ from our composite measure. Third, one concern could be that searches on some of the included terms (see Table 1) could be related other forms of non-domestic abuse. Hence we present a specification where we exclude all terms that are not specifically DV related. All three reductions of the set of search terms have only a minor impact on the estimated coefficients.

The third and final set of specification checks presented in Table A.2 (columns 8–9) test for robustness with respect to key dates. As noted above, bank holidays are days with comparably high levels of DV incidence. Hence we present the results from a specification that includes dummy variables to control for day t being one of the eight annual bank holidays (of which there were four within the lockdown days included in the estimation). Also, internet search activity on DV-related terms may be affected by efforts to raise awareness etc., most notably through DV campaigns. The launch of DV campaigns are commonly timed to coincide with two key days of the year: the International Women’s Day on 8 March and the International Day for the Elimination of Violence Against Women on 25 November (Agüero, 2019).3 Hence in the final robustness check we include controls to day t being within a week of either of these two key dates. The results from these final two specifications are largely indistinguishable from our main specification.

4.5 Results for Los Angeles, California

We next repeat the analysis for the city of Los Angeles, California. To that end, we combine three data sources for the period 1 April 2015 through 31 July 2020: (i) daily counts of DV-related calls for police service recorded by the Los Angeles Police Department (LAPD); (ii) daily search intensities in the State of California for a set 37 DV-related search terms from Google Trends; and (iii) information on average daily temperature and rainfall in LA from the National Centers for Environmental Information (NCEI). Further details about the Los Angeles data, including sources, trends and search-term selection, are provided in the online Appendix. In implementing our algorithm, we find that the search terms that get the highest weight in the search-based DV index for LA are ‘domestic violence hotline’, ‘abusive husband’, ‘reporting abuse’, ‘abuse support’ and ‘domestic violence charges’.

Figure 6 shows the LA counterpart of Figure 5. The evidence is qualitatively remarkably similar to the London case, despite institutional heterogeneities, different approaches in dealing with the pandemic, and differences in the relationship between the police force and the general public. Panel (A) shows that the LA lockdown led to a gradual increase in DV-related calls for police service, peaking after 7–8 weeks at roughly 15% above pre-lockdown levels, followed by a gradual decline. Our results here are similar to Miller et al. (2020) Focusing specifically on Los Angeles, they find an increase in DV-related 911 calls (and in calls to a DV hotline) at the initial lockdown, but a decrease in recorded DV crimes. Turning to Panel (B), we observe that the increase in our search-based DV index had a similar timing structure, but, whereas the increase in recorded DV calls was about 15% at peak, the increase in the DV index was around 30%. There is also a second significant spike in our search-based DV index after 15–16 weeks of lockdown.

5 DISCUSSION

Many types of crises—be it disease outbreaks like the current pandemic, severe economic downturns, or natural disasters—carry the risk of increasing DV (Anastario et al., 2009; Anderberg et al., 2016; Bermudez et al., 2019; Onyango et al., 2019). Effective policy responses require up-to-date reliable data on the scale of the problem. However, conventional data sources have severe limitations in this respect. Police-recorded DV is a particular case in point: victims of DV frequently do not report their abuser to the police, and COVID-19 and its associated restrictions has made reporting an abusive partner even more difficult.

Researchers before us have highlighted the limitations of police-recorded crime data, such as calls-for-service or reported crimes, to act as a proxy for the actual incidence of crime (see, e.g. Carr & Doleac, 2018; Pepper et al., 2010). For example, in a study on juvenile curfews and gun violence, Carr and Doleac (2018) argue that policy interventions aimed at reducing gun-involved crime also affect reporting rates, and exploit ShotSpotter data as a proxy in place of unobserved gun crime incidence.

Other sources of data specifically related to DV incidence include helpline data. Indeed, the data on calls and contacts from the UK’s National Domestic Abuse Helpline presented above showed an increase in the order of 60% in the first few weeks after the London lockdown compared to pre-pandemic levels. However, these data, too, are imperfect as it is not gathered sufficiently systematically over time to allow for a finer analysis. Indeed, Leslie and Wilson (2020) show that failing to account for seasonal trends results in over-estimating the effect of the COVID-19 crisis on DV by almost 50%. Similarly, data on homicides suggest a compositional shift towards domestic cases during the crisis. In 2020, the MPS launched 126 murder investigations while the corresponding number in 2019 was 150.4. However, out of those 22 (17.5%) were domestic killings in 2020 compared to only 16 (10.7%) in 2019. The numbers are, however (fortunately), too low for any conclusive statistical analysis.

In this paper we have adopted and implemented a framework for generating a DV index based on DV-related internet search activity. This has clear advantages compared to police and helpline calls data. However, it also shares some caveats. First, it is difficult to disentangle help-seeking by victims from that of third-party individuals. Indeed, non-victims may have increased their DV-related search intensity either due following upl on the increased media attention devoted to DV incidence during the lockdown or due to being more likely to overhear DV incidents among neighbours. The literature on DV incidence during the lockdown has made some attempts to disentangle the issue (Ivandic et al., 2021; Leslie & Wilson, 2020) but one recent major review (Piquero et al., ) remained silent on it. The same caveat applies to the current paper. We do show in our robustness analysis that our findings are robust to eliminating the most frequent and generic search term ‘domestic violence’. If this term would be relatively more used by non-victims, this would suggest that our results are not driven by third-party individuals.

A second caveat is that a lockdown means that individuals have more time available to spend on the internet. Indeed, there is evidence, for instance from the American Time Use Survey, that suggests that individuals spent more time playing games and using computers for leisure purposes.5 What speaks against the possibility that our results might simply reflect larger time availability is the finding that searches did not increase proportionately more during weekdays when time availability would likely have changed the most.

Our findings show the London lockdown led to a gradual increase in the DV-related crimes recorded by the MPS and the effect of the lockdown remained positive until mid-June. The impact was nevertheless modest, with about 10–15 extra DV crimes per day relative to a normal average of over 200 crimes per day. In sharp contrast, although exhibiting a similar lockdown timing structure, we find a 40% increase (at peak) in our search-based DV index, -seven to eight times larger than the increase in police-recorded crimes and much closer to the increase reported by the UK’s National Domestic Abuse Helpline in relation to helpline calls and contacts.

The broader lesson from our analysis is that it cautions against relying solely on police-recorded crimes or calls-for-service to assess the scale of the DV problem during crises like COVID-19. In such assessments, the use of complementary data sources is important, as it would allow researchers to move towards demarcating the lower and upper bounds of likely impacts on DV. One promising avenue in this respect is to engage with organizations supporting DV victims, encouraging the collection and provision of systematic data from DV helplines. Our algorithm for measuring temporal variation in DV incidence using internet search activity provides another viable strategy to complement assessments based on police records. Although our analysis by no means provides a definite answer to how to best construct a real-time indicator of DV, it can hopefully serve as a starting point that can be extended and further validated.

ACKNOWLEDGEMENT

We are grateful to Ruhi Pindolia for excellent research assistance, to the information managers at the London Metropolitan Police Service for help and support, to Arnaud Chevalier for constructive comments, and to Sofia Amaral, Timo Hener, Andreas Kotsadam and Ana Tur-Prats for insightful discussions.