Economics of Child Protection: Maltreatment, Foster Care, and Intimate Partner Violence

Abstract

Violence within families and child neglect are strikingly common: 700,000 children are found to be victims of abuse or neglect in the United States each year; over the course of childhood, 6% of children are placed in foster care, and 18% witness intimate partner violence. These children are at much higher risks of homelessness, criminal justice involvement, unemployment, and chronic health conditions compared to their neighbors. This article reviews the state of the economics literature on the causes and consequences of child maltreatment and intimate partner violence and calls for greater research into interventions aimed at improving child well-being.

1. INTRODUCTION

Family dysfunction, in the forms of child maltreatment and intimate partner violence (IPV), is strikingly common. In the United States, each year, 4 million cases of suspected child abuse and neglect are reported to Child Protective Services, 700,000 children are found to have been victims of abuse or neglect, and over 1,600 fatalities are attributed to abuse or neglect (U.S. Dep. Health Hum. Serv. 2017). A similar number of women in the United States are killed each year due to violence from an intimate partner, and surveys show that 37% of women in the United States experience IPV during their lifetime (Smith et al. 2017).

In addition to the direct welfare concerns for those who are abused or neglected, there are large externalities associated with such dysfunction. First, direct expenditures on Child Protective Services total over $30 billion per year (Rosinsky & Connelly 2016). More importantly, especially for policy makers who seek to target programs to at-risk youth, maltreated children are at particularly high risk of outcomes such as homelessness or imprisonment as young adults. For example, children who receive child protective services are 2–3 times more likely than other children to be incarcerated by age 21 (Berger et al. 2016, Cutuli et al. 2016, Lindquist & Santavirta 2014, Mersky & Topitzes 2010), are 1.6–2 times more likely to receive disability pensions (Vinnerljung et al. 2015), have substantially lower educational attainment (Boden et al. 2007, Fry et al. 2017, Mersky & Topitzes 2010), and are 2–4 times more likely to be victims of preventable death (Hjern et al. 2004). Fang et al. (2012) estimate that the lifetime cost of new abuse and neglect cases in the United States each year is between $100 billion and $500 billion.

One of the strongest risk factors predicting child maltreatment is the presence of IPV within the household (U.S. Dep. Health Hum. Serv. 2017). The US Department of Justice estimates that 18% of children have been exposed to IPV during their lifetimes, with 12% witnessing IPV directly, most often the child’s father or the mother’s partner victimizing the child’s mother (Hamby et al. 2011).¹ Dong et al. (2004) find that those who have experienced child abuse are three times more likely to have been exposed to IPV relative to those who have not. Others estimate that the increase in the risk of abuse is closer to 15 times higher if previously exposed to IPV (Osofsky 1999). Moreover, a stronger correlation between IPV and child abuse and neglect is observed in high-risk families (e.g., welfare recipients and shelter residents) (Appel & Holden 1998).

In terms of economic theory, there is considerable interest in modeling family decision making, with an emphasis on altruism and the ability of families to achieve Pareto optimality through within-family bargaining (Becker 1991, Browning & Chiappori 1998, Seiglie 2004). As Becker (1991) notes, child abuse may be viewed as an exception to the typical altruism found within families. As a result, economic models of crime appear to be more fruitful than models of altruism in the study of child abuse and neglect. With respect to IPV, we examine the evidence regarding the ability of within-family bargaining models, as well as models derived from behavioral economics, to explain this phenomenon.

The main aim of this review is to describe the state of the existing empirical work in economics on child maltreatment and the related subjects of foster care and IPV.² This includes a discussion of the hurdles to credible research designs in this area, as well as opportunities arising from new data sources and a growing interest in rigorous evaluation of new interventions.

The review is organized as follows. Section 2 discusses child abuse and neglect, and Section 3 describes the role of foster care in responding to such maltreatment. Section 4 considers IPV and its impact on child well-being, and Section 5 concludes.

2. CHILD ABUSE AND NEGLECT

2.1. Measurement and Risk Factors

This section describes how child abuse and neglect are measured to give a better sense of their prevalence and to highlight empirical concerns with common measures. As with other crime statistics, there are two ways child abuse and neglect are typically measured: official reports to child protection authorities and victimization surveys.

2.1.1. Official reports of maltreatment.

As noted in Section 1, child protection systems find that approximately 700,000 children are abused or neglected each year in the United States. Such findings rely on referrals from community members, including physicians, educators, neighbors, and family members. Investigations are then conducted and family courts decide on rehabilitation efforts.

There is wide geographic variation in the rate at which children are found to be maltreated. In 2015, the average annual maltreatment rate across states was 0.95%, with a standard deviation of 0.48%. Even for seemingly similar states, the rates can diverge. For example, Nebraska’s maltreatment rate is over 2.5 times the rate of neighboring Kansas.

The wide heterogeneity in official reporting rates suggests that there are different standards about when to report abuse and how evidence of abuse is adjudicated. This variation has implications when evaluating programs that aim to reduce child maltreatment. An intervention’s effects on official maltreatment rates could stem from a change in enforcement rather than a change in the underlying maltreatment of children. For example, a program aimed at providing family preservation services to prevent maltreatment may increase the monitoring of families as part of providing services. This additional attention could result in more reports even if abuse and neglect in these families decline relative to comparison families that are not monitored as closely.

2.1.2. Survey evidence.

In addition to official reports, victimization surveys provide data by asking respondents directly about maltreatment. Finkelhor et al. (2015) conduct a phone survey of children and find that 15% report maltreatment in a given year, with 38% of 14–17 year olds reporting having ever experienced abuse or neglect. Currie & Tekin (2012) employ data from the National Longitudinal Study of Adolescent Health (Add Health), which includes questions aimed at measuring abuse or neglect. For example, respondents were asked if they were left alone when they should have been supervised more than 10 times or if they were hit, kicked, or slapped by their parents or other adult caregivers more than 10 times. Of the respondents, 23% reported that they were maltreated in either of these ways, and 4.8% reported being victims of sexual abuse.

Such survey evidence suggests that at least some forms of child abuse and neglect are common. Akey policy question is when such conduct rises to a level that requires Child Protective Services to intervene—a question the answer to which varies across time and space, and one that is addressed in Section 3.

2.1.3. Linked administrative data.

A resource that is becoming increasingly available—and deserves more attention from empirical researchers—is linked administrative data, such as health and education data combined with data from child protection systems. For example, consider the intervention discussed in Section 2.1.1, aimed at reducing abuse and neglect through family preservation services. While official reports can be contaminated by increased attention brought by the program, and survey evidence for a particular program can be expensive to collect and relies on self-reports that may also be impacted by the program itself, linked administrative data offer a wide array of objective outcome measures. These include measures of health outcomes from Medicaid, schooling outcomes from state and district databases, educational attainment from financial aid applications, employment and earnings from Social Security or state unemployment insurance programs, welfare and disability receipt, and criminal justice involvement.

For example, Nordic countries have abundant linked administrative data, which are increasingly used in child welfare research (see, for example, Kloppen et al. 2015). In the United States, a prominent example of linked administrative data is the Illinois Integrated Database constructed by the Chapin Hall Center for Children (Goerge et al. 1994). The database links state and local administrative data sets going back to the 1980s, with high-quality child welfare data dating back to the early 1990s. Outcomes can be gathered from schooling data, Medicaid data, juvenile and adult corrections, and employment and earnings. Similarly, Cuyahoga County, Ohio, which contains the city of Cleveland, has linked child welfare data covering a similar time span (Coulton et al. 2016). The Children’s Data Network has recently linked such data from California and is being used to inform policy (Putnam-Hornstein et al. 2015). This growing tide of data resources will make it possible to evaluate programs by exploiting natural experiments and to reduce the cost and increase the feasibility of randomized trials in child protection interventions.

2.2. Risk Factors Predicting Abuse and Neglect

Administrative reports describe the risk factors that lead to abuse and neglect. Two that are often highlighted are drug and alcohol abuse and the presence of a caregiver with a history of IPV (Mowbray et al. 2017, U.S. Dep. Health Hum. Serv. 2017). Indeed, there is some evidence that efforts to curb alcohol consumption or methamphetamine use reduce maltreatment, as measured by child protection interventions such as foster care (Cunningham & Finlay 2013, Markowitz et al. 2014).

Exposure to IPV is a particularly strong risk factor for child maltreatment, as noted in Section 1 and investigated in more detail in Section 4. In a related study, Waldfogel et al. (2010) use data from the Fragile Families and Child Wellbeing Study, a longitudinal study of 5,000 children born between 1998 and 2000 with an emphasis on children born to unmarried parents. This study includes a number of health and maltreatment questions. Waldfogel et al. review the abuse and neglect research that uses these data and note that one of the more robust findings is that the presence in the household of a man who is not the child’s father elevates the risk of child maltreatment.

Parental criminal justice involvement is another risk factor for child maltreatment. For example, Berger et al. (2016) study administrative data in Wisconsin and find that, among all children with an official maltreatment report, 8% had a parent in state prison at some point during the 12 months following the report.

Increasingly, linked administrative data are being used to generate risk scores and decision support tools that can be used by child protection investigators. This introduces new capabilities as well as ethical questions involving privacy and the potential for discrimination to be encoded within risk-assessment algorithms (Cuccaro-Alamin et al. 2017). The development of these tools also provides a potential research opportunity: first, a test of whether the introduction of data- driven decision support tools affects investigator behavior and, second, a measure of whether the children in families investigated when these tools are available have better health and educational outcomes. Alternatively, the tools may generate discrete flags based on continuous risk scores, setting up the possibility of comparing cases that fall above and below the thresholds in a regression discontinuity design.

2.3. Consequences of Abuse and Neglect

Abuse and neglect are directly welfare relevant due to the harm they cause during childhood. However, as is the case for quality of life questions in health economics, it is difficult to measure changes in maltreatment, and with the measures that we do have—official reports to authorities, self-reports in victimization surveys, and administrative data on health outcomes—it is also difficult to estimate society’s willingness to pay for reductions in maltreatment.

Research has shown that those who are victims of abuse and neglect as children are substantially more likely to have costly outcomes such as health problems or criminal justice involvement. These estimates help quantify important externalities that can be part of a return-on-investment calculation for programs that prevent abuse and neglect.

Widom (1989) conducts the seminal study using relatively large survey evidence to estimate effects of childhood abuse and neglect on longer-term outcomes. The study compared abused or neglected children with children who were matched on observable characteristics. For young children prior to school entry, the characteristics include sex, age, race, and hospital of birth; for those of school age, the matching variables include sex, race, date of birth, class and elementary school, and neighborhood. Delinquency outcomes were noticeably higher among those who were abused and neglected. For example, 19% of maltreated males have a violent criminal record compared to 13.5% of the comparison group. Physical abuse was particularly related, leading to the conclusion that child abuse leads to a cycle of violence. In the end, the study is a call for more research on the mechanisms that explain why abuse and neglect lead to more violence to inform preventive measures.

Currie & Spatz Widom (2010) look at even longer-term effects with this matched sample. In 2003–2004, outcomes were available for a matched pair comparison among 358 subjects. Those who were victims of child maltreatment when they were young are found to have lower levels of employment (17 percentage points) and earn $9,000 less compared to $28,000 in earnings among the comparison group. This raw difference declines to $6,500 when demographic controls are included in the estimation.

Currie & Tekin (2012) study the effects of child maltreatment on crime outcomes using Add Health. They employ methods including rich controls, sibling comparisons, and twin comparisons and find that maltreatment is associated with a doubling of the likelihood that the child will commit crime. For example, they find an approximately 10-percentage-point increase in commission of a nondrug crime compared to a mean of 16%. The effects are particularly strong for experiences of sexual abuse and greater degrees of maltreatment.

Going forward, more research on the effects of abuse and neglect on health outcomes is needed. First, health measures offer more objective child safety assessments compared to official maltreatment reports. Second, there is a large medical literature devoted to the phenomena of adverse childhood events that documents the correlation between child abuse and neglect and poor health later in life (see, for example, Caspi et al. 2016, Danese et al. 2009, Felitti et al. 1998). Similarly, there is significant interest, in biology and economics, in the effects of prenatal health insults on long-term outcomes (Almond & Currie 2011). A better understanding of the long-term implications of abuse and neglect for chronic health conditions can inform policies that invest in prevention and amelioration of child maltreatment.

2.4. Interventions that Reduce Abuse or Neglect

There are few rigorous studies finding programs that reduce abuse and neglect (Levey et al. 2017). One exception is the Nurse Family Partnership. This is a postpartum care program where nurses make home visits to new mothers. Olds (2007) describes how intensive programs—with visits up to 2 years after birth and targeted to low-income mothers—have been evaluated using randomized controlled trials (RCTs). These programs have been shown to reduce child maltreatment and to improve cognitive development, as well as ameliorating longer-term problems including arrests and health behaviors such as youth smoking and drinking. Programs that are less intensive have been found to have smaller effects (Aos et al. 2004). Currently, a large-scale expansion of the Nurse Family Partnership in South Carolina is being studied with an RCT (Lantz et al. 2016).

Given that poverty is positively correlated with maltreatment, poverty amelioration programs should be evaluated on their ability to protect children. Paxson & Waldfogel (2002) use state-level panel data to study the relationship between families’ circumstances and the levels of child maltreatment reported and foster care placement. They find that several characteristics are associated with official reports of maltreatment, including the fractions of children with absent fathers, with two nonworking parents, with incomes below 75% of the poverty line, and with working mothers (holding income fixed). The results also show that decreases in state welfare benefits are associated with increases in child protection in the form of foster care placement. The results suggest that welfare programs can improve child safety.³

In addition to broader programs that have an effect on abuse and neglect, the most direct method to protect children is the foster care system. The next section considers this system in more detail.

3. CHILD PROTECTION AND FOSTER CARE

3.1. Measurement

One response to child abuse and neglect is the temporary removal of children from their families and their placement into foster care, typically a foster family, with the goal of family rehabilitation and reunification. Foster care placement is fairly common in developed countries. In the United States, Wildeman & Emanuel (2014) calculate that nearly 6% of children are placed in foster care during their childhood, using data from the 2000s. The rate of foster care placement varied by race and ethnicity, with over 15% of Native American children and over 11.5% of African American children being placed. Fallesen et al. (2014) find that the lifetime risk of foster care placement was similar in Denmark in 2000 but fell between 2000 and 2010.

Figure 1 shows that the number of children in foster care in the United States has varied widely over the past 50 years, with approximately 300,000 children in care in the 1960s and 1970s, increasing during the 1980s to 550,000 by the mid-1990s, with a subsequent decline to just over 400,000 by 2015. Each year, over 250,000 children enter foster care in the United States (U.S. Dep. Health Hum. Serv. 2016). While official statistics are not yet available for more recent years, it is expected that the number will have increased since 2015, in part due to the opioid epidemic.

Figure 1.

Number of children in foster care in the United States, 1962–2015. The gap in the data is due to the combination of two data sources, and there are no available data for that time period. Data are publicly available from the US House of Representatives Green Book, 1996, and the US Department of Health and Human Services Adoption and Foster Care Analysis and Reporting System.

The changes over the past 20 years mirror changes in crime rates more generally (Levitt 2004), which suggests that demand for foster care has declined since the mid-1990s.

In addition to the time series variation in placement, there is a striking degree of variation in placement across states in the United States. In terms of the number of children in foster care at a point in time, the mean is 6 per thousand, with a standard deviation of three per thousand. Alaska has the highest foster care rate, at 14 per thousand, while Virginia has the lowest, at 2.5 per thousand.

Figure 2 reports the fraction of children entering foster care in each state plotted against the fraction of children found to be maltreated. The first thing to note is the range of both axes. As noted in Section 2, the fraction of children found to be maltreated varies widely across states, as does foster care placement. Second, the relationship between the number of children found to have been maltreated and the fraction of children entering foster care is weak, with a slope of 0.08. For example, Michigan and Wisconsin have similar rates of foster care placement, but the official child maltreatment rate in Michigan is nearly four times higher than that of Wisconsin. Similarly, in New Hampshire, three out of every 1,000 children entered foster care in 2015 compared to eight per thousand in neighboring Vermont. This suggests that not only does the definition of maltreatment vary across places, but also the response to maltreatment varies. As with official maltreatment rates, changes in foster care entry rates can be difficult to interpret, as they may vary due to the prevalence of abuse and neglect or the policies that govern the rate of entry.

Figure 2.

Fraction of children entering foster care plotted against children maltreated in 2015. The slope is 0.08. Data taken from the US Children’s Bureau and the US Department of Health and Human Services Adoption and Foster Care Analysis and Reporting System.

The prevalence of foster care placement is the result of demand and supply forces. Child abuse and neglect by parents, coupled with child protection policies that determine the response to suspected abuse and neglect, affect the level of demand. Supply is provided largely by families willing to open their homes to foster children on a temporary basis, with an average length of stay of 1–2 years across the United States.

In addition to the level of foster care services, the supply responsiveness to the monthly payments made to foster parents can inform compensation policy. Doyle & Peters (2007) use a price theory model of the foster care market to estimate the effects of raising the monthly subsidy offered to foster parents on the supply of foster homes. Their contribution is the recognition that, in the late 1980s, there was a shortage of foster homes, and changes in subsidies within states were often tied to other welfare programs. In this setting, the model implies that the relationship between changes in subsidy levels and changes in the foster care population within states over time traces the foster care supply curve. The results suggest that, in states with shortages, an increase in the subsidy by 10% would increase the quantity supplied by 3%.

3.2. Causal Effects of Foster Care on Child Outcomes

The decision to place a child in foster care involves a trade-off between two goods: child protection and family preservation. A current policy concern receiving considerable attention is the potential benefit of diversion, where children at risk of foster care placement remain at home, often with family preservation services, although not always (Berrick & Hernandez 2016). Experiments with such interventions are just beginning to be contemplated and offer an opportunity for researchers. A better understanding of the causal effects of foster care placement on child outcomes is needed to inform such policies. This section describes the state of the evidence.

3.2.1. Selection on observables.

Few studies have considered the effect of removing children from home on child outcomes.⁴ Earlier reviews by the National Research Council and the Institute ofMedicine (Nat. Res. Counc. & Inst. Med. 1998) and by McDonald et al. (1997) find 10 studies from 1960 to 1996 that employ a comparison group of any kind. Most have small sample sizes and mixed evidence of correlations with child outcomes.⁵

Berger et al. (2009) use the National Survey of Child and Adolescent Well-Being, which includes 2,453 children who were investigated by Child Protective Services, 342 of whom were placed in foster care. Propensity score matching estimates suggest little effect of placement on child behavior problems or measures of cognitive skill over a 30-month time frame. While the estimates show robustness to different assumptions, they all rely on a selection-on-observables identifying assumption. While this is a prominent study using one of the largest child welfare survey data sets (Wildeman & Waldfogel 2014), the null results do not rule out large effects due to relatively imprecise estimates. For example, at the bounds of the confidence interval, externalizing behavior problems may increase by one-third of a standard deviation after foster care placement.

One of the largest studies of the effects of foster care placement on child outcomes is conducted by Jonson-Reid & Barth (2000), who link juvenile incarceration and child welfare administrative data in California for 160,000 children involved with the child protection system. They find little difference in incarceration rates for those who entered foster care compared to those who remained home. In-home services for abused children are associated with somewhat lower incarceration rates, especially for African American and Hispanic children.

3.2.2. Instrumental variable approach using randomly assigned investigators.

Doyle (2007b, 2008, 2013) uses an economics lens to study the effects of foster care on child outcomes. The model illustrates a moral hazard problem where child protection investigators face incentives that may lead to too many placements. If they fail to remove a child who is subsequently found to be abused or neglected, they will face punishment. If they emphasize child protection over family preservation, negative consequences of foster care placement for marginal cases are unlikely to be found until years later and are unlikely to be attributed to Child Protective Services. As a result, the theory suggests that Child Protective Services may overemphasize child protection at the expense of family preservation.

The main empirical strategy in these papers is an instrumental variables approach that exploits the fact that most cases are rotationally assigned to investigators within teams. This effectively randomly assigns cases to investigators. The investigator’s placement rate in all other cases is used to measure the tendency to recommend placement.

Using administrative data in Illinois for children suspected of being abused or neglected in the 1990s, Doyle shows that the investigator assigned to the case matters for placement (the first stage). Using this source of variation, Doyle finds that outcomes are much worse for children placed in foster care compared to those remaining at home. Teen motherhood and juvenile delinquency rates are found to be two times higher due to foster care placement, adult incarceration rates are three times higher, and employment and earnings are 40% lower. Meanwhile, emergency care—a proxy for child safety—is actually found to increase with foster care placement in the 90 days to 1 year after the investigation. There is at least one benefit to placement: Wellness visits are three times higher in foster care. This finding is reassuring given that all foster children are supposed to receive a wellness visit: The strategy finds effects where we know it should.

The estimated effects apply only to cases where the investigator assignment affects foster care placement: cases where investigators may disagree about the placement recommendation. From a policy perspective, this is a margin of interest: The results suggest that the system should be more family preservation oriented in the context of the study, Illinois in the 1990s, which at that time had one of the highest placement rates in the United States. The placement rate has fallen significantly since due to polices aimed at reducing the number of children in foster care.

Studies of more recent cases in Illinois and of cases in other state or national systems would be welcome comparisons. For example, systems differ in the quality of their foster homes, the quality of the foster care case managers who oversee the cases once the child is placed, and the placement rate, so that the marginal case will differ across places. If similar effects are found across a wide set of areas, then that would suggest that the one common feature—the act of removal—is likely responsible for the negative consequences, at least across the quality range of the systems studied. If the results differ across jurisdictions, it will be constructive to unpack why some systems perform better than others.

The only published study that has used a similar instrumental variable approach is that of Warburton et al. (2014). This study focuses on 16–18-year-old males and rich administrative data from Canada involving over 20,000 children. Foster care placement is found to lead to less criminal justice involvement for these older youths, although the estimate is imprecise, as it cannot rule out a 50% increase. The study also uses another identification strategy that exploits a change in child protection policy that sharply increased the number of placements in reaction to a child homicide. In an interrupted time series approach, the results suggest that foster care placement is associated with a very large reduction in high school graduation and a similarly large increase in welfare receipt as young adults. This sudden change in policy is common across jurisdictions in the United States and is a potentially fruitful source of placement variation for future research into the causal effects of foster care on child outcomes.

3.2.3. Effects of different types of foster care placements.

In addition to measuring the effects of foster care placement versus its alternatives, there is significant policy interest in measuring the effects of different types of placement. This remains a first-order research question. The two main types of placement that have been considered are kinship placement and institutionalization.

3.2.3.1. Kinship foster care.

Over the past 30 years, the largest shift in foster care policy has been the embrace of calling on relatives such as grandparents to provide foster homes. Placement with kin is thought to be less traumatic for children, and, as a result, most states look to find kinship placements prior to seeking nonrelatives. An empirical evaluation of the performance of relative foster homes compared to nonrelative ones is difficult because children with willing relatives may differ from those without.

Doyle (2007a) attempts to address the quality question by investigating a reform in Illinois that reduced the subsidy offered to relatives to provide foster care by 30%. Children placed in foster care following reports of abuse—a subset of entrants who were less likely to be affected by concurrent reforms—were 20% less likely to find a willing relative. This suggests that relatives do respond to financial incentives when it comes to deciding whether to provide foster care. Furthermore, Doyle investigates quality measures, including placement disruption and the likelihood that the child was taken to a physician for a wellness visit—a task expected of all foster parents for children entering foster care. These measures were stable or slightly improved following the reform. This suggests that the marginal kinship providers—those affected by the subsidy reduction—were of similar or slightly lower quality compared to the nonrelative caregivers enlisted to provide care in the relatives’ absence.⁶

3.2.3.2. Institutionalization versus family foster care.

There is also relatively little quantitative research showing that the dominant form of foster care—placing children with families—is superior to group homes or institutions such as orphanages. One of the more controversial projects in this regard is the Bucharest Early Intervention Project (Marshall 2014, Schoenmaker et al. 2014, Smyke et al. 2009, Zeanah et al. 2003), which randomized the placement of 136 infants into either usual care in the orphanage or placement with a foster family. IQ scores were substantially higher (by 0.6 standard deviations) for those placed with foster families at ages 3.5 and 4.5. Later, at age 8, scores were still higher (by 0.4 standard deviations) and marginally statistically significant. The Bucharest Early Intervention Project provides speculative evidence that earlier placement results in larger returns of family foster care in terms of cognitive ability.

3.2.4. Transition to adulthood.

Children who age out of foster care have received attention, in part because they can be tracked more easily and because they are a policy-relevant group given the support services provided as children transition from the foster care system to independent living. These children are at particularly high risk of poor outcomes such as criminal justice involvement and homelessness (Dworsky et al. 2013, Reilly 2003).

Given these outcomes, there is significant interest in aiding children as they transition out of foster care through what are known as independent living programs. Unfortunately, large- scale, quantitative assessments of these programs are generally lacking. A few programs have been evaluated, and the results suggest limited effectiveness (Greeson et al. 2015a,b). One exception is the Youth Villages Transitional Living Evaluation (Jacobs et al. 2015). This program includes support and counseling tailored to the needs of each youth and has been studied with a prospective RCT involving over 1,300 youths in Tennessee. One of the success stories appears to be in the area of housing, where homelessness was found to drop from 27% to 21% after 1 year for those randomized into the program compared to those in the control group (Jacobs et al. 2015). Meanwhile, the treatment group had 18% higher earnings compared to $3,488 for the control group. These improvements were not found to be statistically significant at 2 years, although homelessness was not assessed at that time (Skemer & Jacobs 2016).

There are many related topics of interest, from child poverty to child adoption, that could be the subject of another review. The next section addresses one of the largest predictors of child maltreatment and foster care entry: IPV.

4. INTIMATE PARTNER VIOLENCE AND CHILD WELL-BEING

In addition to abuse directed at children, exposure to IPV can also negatively affect children. In homes characterized by IPV, children are more likely to suffer stress, depression, and internalizing and externalizing behavioral problems (Herrenkohl et al. 2008, Sternberg et al. 2006, Wolfe et al. 2003). They are also more likely to express anger and depression (Carlson 2000). More recently, scientists have documented that children exposed to violence (this includes IPV, physical abuse, or bullying) exhibit telomere erosion between the ages of 5 and 10 (Shalev et al. 2013). Telomere length is considered an important physiological mechanism linking stress to cellular aging, disease, and mortality in humans. Evidence also suggests that boys who witness IPV are more likely to be abusive in adulthood (Roberts et al. 2010).

Why would children who have not suffered abuse themselves, but have witnessed or otherwise been exposed to IPV, suffer worse outcomes? Fathers who are violent toward their partners are less involved parents and exhibit more negative child-rearing practices (Guille 2004). But research has focused more often on the impact of IPV on mothers and their ability to parent. Several studies have documented the physical, emotional, and psychological toll of IPV borne by mothers. Women who have been abused are significantly more likely to report having asthma, diabetes, chronic pain, and activity limitations (Black et al. 2011). They are also three times more likely to report being in poor mental health. Other work has documented higher rates of post-traumatic stress disorder, lower self-esteem, and greater depression and anxiety (Golding 1999). As a result of these negative effects on physical and mental health, mothers who are the victims of IPV suffer deficits in terms of their parenting capacity (Stephens 1999) and the quality of the attachment to their children (Cleaver et al. 1999, Holden 2003, Levendosky et al. 2003).⁷

4.1. Causal Effects of Intimate Partner Violence on Child Outcomes

Estimating the causal impact of IPV on child outcomes is complicated by two main factors: the endogeneity of IPV and issues related to measurement. First, homes characterized by IPV are disproportionately disadvantaged as measured by race, ethnicity, and income.⁸ As such, the finding that these children suffer worse outcomes cannot necessarily be attributed to exposure to IPV and could simply reflect underlying differences in the level of disadvantage faced by these families, although researchers have made efforts to control for as many potential confounders as possible in their analyses. The second complication is difficulty measuring IPV based on survey data. Self-reports are noisy, and the relatively low prevalence of witnessing violence in any given year (6.6% report any exposure in the past year, and 2.6% report witnessing violence directly) would require very large data sets to study the impact of IPV on child outcomes. As noted above, large administrative data are more promising.

Aizer (2010) provides an example of a study that uses a large administrative data set to estimate the impact of reductions in IPV against pregnant mothers on newborn health. Exploiting exogenous variation in IPV from changes in local policies, she estimates significant reductions in IPV that result in significant improvements in birth weight, an important marker of newborn health that is strongly predictive of later life outcomes. The administrative data consist of the universe of births over a 10-year period in California linked with hospitalization data for the mothers that enable one to identify mothers admitted to the hospital for an assault while pregnant, independent of self-reports. Using health data (rather than crime data) to measure IPV offers the advantage of not having to rely on police reports, which are arguably more endogenous.

There is also evidence that IPV generates externalities outside the household. Baker & Hoekstra (2010) use a unique administrative data set linking child test scores and disciplinary infractions with measures of IPV among classmates’ parents. They document that, after the arrest of a child’s parent for IPV, the child’s classmates suffer worse test scores and behavior. The assumed mechanism is that the child’s behavior suffers as a result of the violence, generating externalities for their classmates. This study is an example of a creative use of linked administrative data that have, since the study was written, become even more prevalent.

Solutions to the problems of IPV with respect to child health and well-being consist of efforts either to ameliorate the negative effects of exposure on children ex post or to reduce or eliminate the prevalence of IPV. In this section, we discuss the existing evidence with respect to each option.

Researchers have examined the impact of interventions for children who have been exposed to IPV, but this research is largely outside of economics, and only a small number of the evaluations include valid comparison groups and follow-up assessments. Graham-Bermann & Hughes (2003) provide a review of what they refer to as exemplary studies—studies with a randomized design; outcomes based on well-measured and accepted outcomes related to child behavior; and the collection of information from mothers, children, and teachers, as well as some administrative records. In general, they conclude that these exemplary studies find that interventions that address the symptoms of the child and also provide parenting education and support are the most effective in terms of improving child internalizing and externalizing behavioral problems.

The finding that a program that targets mothers and children is especially effective is interesting and somewhat novel. Parenting programs that seek to modify parenting behavior and perceptions have generally been found to be ineffective, particularly when targeting disadvantaged families (Lundahl et al. 2006).

Perhaps the circumstances faced by women who have been the victims of IPV generate a different context in which the provision of parenting support can be more effective. This could be either because the mother is more receptive to the intervention or because moderating parenting practices at this sensitive moment has the potential to generate a greater impact (that is, the return is greater). Although the Heckman model (Cunha & Heckman 2007, Heckman 2006) of child investment and returns incorporates the concept of sensitive windows of investment, this concept is defined by time and the child’s place in the life cycle. In light of these results, one might consider expanding the model to include sensitive windows based on context, with future empirical research exploring the effectiveness of efforts to modify parenting behavior and perceptions at sensitive moments, such as after exposure to IPV. This work would ideally be based on larger samples, include families outside of the shelter system, and include long-term follow-up. The results might be generalizable to other populations and contexts as well, suggesting ways that we might design and deliver parenting programs to other vulnerable populations.

4.2. Interventions that Reduce Intimate Partner Violence

One way to reduce the harm of childhood exposure to IPV is to reduce this violence in the first place. Economics provides frameworks to consider this question in terms of household bargaining and as a part of the larger economics of crime.

4.2.1. A model of household bargaining.

Incidence of IPVhas fallen considerably over the past 20 years. According to the US Bureau of Justice Statistics, IPV against women has fallen from about 10 per 1,000 in 1993 to half that by 2008 (Catalano et al. 2009). Intimate partner homicide has declined by a similar amount.

We consider whether an economic analysis of IPV, which places violence within the context of a household bargaining model, can explain such dramatic declines. In such a model, men derive utility from abuse, and women derive disutility. The amount of violence in a relationship is a function of the relative costs and benefits of violence, as well as the outside option available to each member of the household. The outside option is defined by the utility enjoyed by each individual should the relationship dissolve (Farmer & Tiefenthaler 1997, Tauchen et al. 1991). Within this framework, multiple factors can influence the amount of violence by either (a) increasing the costs to men of perpetrating violence or (b) improving women’s outside options through social interventions and the labor market. Below, we discuss the empirical evidence regarding each possibility.

4.2.2. Increasing the costs of violence: the legal response to intimate partner violence.

Historically, states did not prosecute most crimes of IPV. For example, US law historically exempted husbands from allegations of rape, and police officers responding to reports of IPV were discouraged from arresting perpetrators; they were instead advised to give counsel. This began to change in the 1980s, when victims successfully sued police departments for failure to provide equal protection after repeated calls for assistance did not result in arrest (Worden & McLean 2010). Since then, the legal system has increasingly taken issues of IPV more seriously. Police departments, prosecutors, and judges have all adopted policies with the aim of increasing the likelihood and severity of punishment.

However, to date, there have been few evaluations of such policies that allow for causal inference. Some experiments have been conducted, but the research design is often compromised by treatment migration and treatment dilution. This is to be expected in social settings with such important consequences. An example of one such compromised study is the Minnesota Domestic Violence Experiment, initially conducted in the 1980s and evaluated by criminologists (Berk & Sherman 1988, Sherman & Berk 1984). In this study, police officers who responded to calls of IPV were randomly assigned to either arrest or merely counsel the perpetrators. Cases in which life- threatening injury seemed likely were excluded. In practice, nearly all of those perpetrators who were assigned to be arrested were, but many of those assigned to be counseled were in fact arrested. This is most likely because they were deemed by police officers to be sufficiently threatening that lack of arrest would significantly endanger the victim. Thus, the delivered treatment was ultimately endogenous, complicating causal inference. Angrist (2006) reanalyzes the data using instrumental variable techniques to uncover the treatment effect. Specifically, he uses treatment assignment as an instrument for treatment delivered. When he does this, he estimates significant negative effects of arrest on the probability of reoffense for IPV. This represents an example of how experiments in social settings can be improved with econometric techniques to generate causal estimates.

While this particular case was designed as an RCT, even in the absence of RCTs, the criminal justice setting often lends itself to analysis based on naturally occurring, plausibly exogenous variation: Police officers, prosecutors, defense attorneys, and judges are most often assigned randomly and can influence important outcomes. The main barrier to analysis, however, has been a lack of access to data. Fortunately, similar to the child welfare data linkages, great advances have lately been made in the area of criminal justice, not only making such data available to researchers but attempting to link such data with additional outcomes found outside the criminal justice arena. For example, the UChicago Urban Labs have been linking data in Chicago and New York to study innovative crime prevention strategies, some of which are described below. As local jurisdictions have increasingly sought more aggressive responses to IPV, more research is needed to better quantify the impact of such efforts in terms of recidivism and family outcomes more generally.

4.2.3. Improving women’s outside options.

Improving a woman’s outside option affects her threat or disagreement point, the point that represents the payoff should the negotiations break down and the union dissolve. As the outside option improves, so too does the payoff to dissolution, and the bargaining power the woman has when negotiating over the allocation of resources (including violence) within the household increases. The two main ways of improving women’s outside options involve improving labor market conditions for women and providing services to women that help terminate abusive relationships. It is important to note that, within the context of a household bargaining model, a woman’s actual earnings or use of services need not increase, only her potential earnings or use of services, to raise her threat point and thus bargaining power (Pollak 2005). Women can also enter the labor market or use services in an effort to signal their threat points but not necessarily dissolve the relationship, as discussed by Farmer & Tiefenthaler (1997). We discuss each possibility in turn.

4.2.3.1. Female labor market opportunities.

Numerous economists (and other social scientists) have considered the role of potential earnings in the labor market in affecting rates of IPV, driven in part by the strong secular increases in women’s wages and declines in IPV over the past 20–30 years. The two main barriers this research must address are data availability and identification. With respect to identification, structural methods, panel data, and instrumental variables techniques have all been used to identify the effects of women’s labor market conditions or outcomes on violence. Bowlus & Seitz (2006) use structural methods to estimate a negative impact of female employment on violence. Tauchen et al. (1991) and Farmer & Tiefenthaler (1997) use panel data to examine the impact of changes in a woman’s income over time on violence. Panel data enables one to overcome the potential for bias from time-invariant omitted variables, but the potential for reverse causality (that is, reductions in violence possibly resulting in greater employment and wages) remains.

Other studies do not use individual-level data linking wages and violence, but rather rely on administrative data linking local employment and earnings with rates of IPV. Examples include a study by Anderberg et al. (2016), who use data from the United Kingdom and variation in opportunities brought about by the Great Recession. In areas that witnessed larger increases in female unemployment, violence increased, in contrast to areas that witnessed larger increases in male unemployment, which saw declines in violence. Aizer (2010) uses panel data for local labor markets in California and links relative increases in demand for female labor with declines in IPV. Munyo & Rossi (2015) use plausibly exogenous changes in the real exchange rate in Uruguay that generate opposite effects on the average wages of men and women to estimate the impact of an increase in the wage gap on IPV. They find that IPV increases when the wage gap increases, consistent with a bargaining model.⁹

The above results notwithstanding, there is less uniformity regarding bargaining models of IPV when the data are from developing countries. For example, Erten & Keskin (2018) find that an increase in female education in Turkey increases female earnings but does not reduce IPV. Using data from Mexico, Bobonis et al. (2013) study the impact of a conditional cash transfer (CCT) program (allocated randomly) on IPV. They find that, although the receipt of CCTs reduces the probability of physical violence, it increases threats of violence. There are two potential means of explaining why these results differ from those described above. On the one hand, a CCT differs from an increase in potential labor market earnings. On the other hand, a necessary condition of a bargaining model is that the threat of termination of a relationship is viable. In many developing countries, the stigma associated with being unmarried may be particularly high, substantially lowering the likelihood of any threat to leave a marriage. Future research should consider the conditions under which a household bargaining model is useful in capturing the determinants of IPV.

4.2.3.2. Services for victims of intimate partner violence.

The passage of the National Violence Against Women Act (VAWA) in 1994 resulted in a major increase in the provision of services to victims of IPV. VAWA authorized a grant program administered through the Department of Justice’s Office of Violence Against Women. Services that have been provided under VAWA include victim assistance programs, transitional housing, shelter services, civil legal assistance, and court training and improvement grants. Most of these programs have not been evaluated rigorously. They do not lend themselves to RCTs, and treatment receipt is endogenous. Another likely barrier to evaluation is the fact that the anonymity of the recipients is of primary importance, limiting researchers’ access to such information.

The lack of research on the effectiveness of public provisions of such services may have contributed to past difficulties in reauthorizing VAWA (it was eventually reauthorized in 2013). The lack of a possibility for RCTs suggests significant opportunities for economists and other social scientists trained in using naturally occurring, quasi-experimental variation for causal inference. Given the difficulty of obtaining individual-level data, analyses that focus on exogenous variation in the timing and local placement of services for identifying variation are likely the most promising.

4.2.3.3. Reducing intimate partner violence: alternative models based on behavioral economics.

Alternative models in economics can also be used to understand the determinants of IPV. These models explicitly consider individual decision makers’ cognitive limitations and psychological biases. The former might include what Kahneman (2011) describes as a two-tiered model of cognition: The first level is fast, instinctive, and emotional, and the second level is slower, deliberative, and logical. The latter include imperfect optimization, bounded self-control, and nonstandard preferences (Congdon et al. 2011). In this section, we discuss whether behavioral models that incorporate cognitive limitations or psychological biases might explain IPV.

A two-tiered model of cognition can potentially be used to explain IPV and generate potential interventions. If men who abuse their partners are engaged in the first mode of thought—fast, instinctive, and emotional—then efforts to encourage them to engage more in the second mode of thought—slower, deliberative, and logical—might lead to a reduction in violence. This has yet to be tested. However, work by Heller et al. (2017) tests the potential for cognitive behavioral therapy (CBT) to influence violent behavior. The intervention stresses the importance of the slower, more deliberative means of thinking and provides subjects with tools to help the transition to slower thought. The authors find that, when the CBT was administered to at-risk youth, it reduced arrests for violent crimes. Whether this model of fast thinking accurately reflects the decision making of violent men and whether interventions involving CBT for the abusers can reduce future violence are important topics for future research.

Regarding psychological biases, there is evidence that bounded self-control may affect the incidence of IPV. This work, however, has focused on the bounded self-control of women: Women who have been abused and report their abuser to the police for prosecution at the time of offense often seek to drop the prosecution days or weeks later. In other words, these women plan to behave one way but often end up behaving otherwise, which is considered evidence of bounded self-control. This intrapersonal conflict can potentially respond to commitment devices.

Research has found that, in the context of IPV, women do seem to value a commitment device. Examining a policy that forces prosecutors to continue with the prosecution of IPV even if women state a preference for not doing so (a commitment device), Aizer & Dal Bo (2009) find that women are in fact more likely to report abuse when this policy is in place. This is consistent with a model of bounded self-control with a demand for commitment devices among women. It should be noted, however, that research on the impact of mandatory arrest for IPV indicates that this policy results in a decrease in reporting and an increase in intimate partner homicide. However, family violence (including abuse against children) declined as a result (Iyengar 2009). More work is needed to better assess whether bounded self-control (among either women or men) can explain IPV and whether efforts to encourage a commitment to behavior can reduce violence.

5. CONCLUSION

Child welfare is a multidisciplinary research area; how can economics make an impact? Economic theory is a useful way to frame problems and develop solutions. For example, the economics of crime emphasizes opportunity costs of would-be perpetrators as a predictor of maltreatment and suggests that antipoverty and employment programs can be successful in protecting children; household bargaining emphasizes that greater opportunities for mothers, whether or not they are realized, can empower them in ways that improve child well-being; moral hazard problems in child protection suggest a tendency toward an overuse of foster care; and behavioral economics can guide interventions that reduce the threat of violence within families.

Similarly, economics incorporates cost-benefit analyses as a natural discussion point when evaluating child welfare programs. The power of the economics approach lies in casting a wide net for those costs and benefits. For instance, a program designed to reduce foster care entry might compare the cost of the program with the savings from lower rates of placement, but economics emphasizes that this is not sufficient. We need to study a wide range of child outcomes to begin to understand whether a program has benefited the child and to gauge whether the costs of the program were indeed worth the expense.

This brings us to the recurring theme of this review, which is that the linked administrative data that are increasingly available will usher in a wave of research on this wider range of costs and benefits. These data include productivity measures from education, employment, and welfare agencies; externality measures from criminal justice; and health measures from state Medicaid programs or providers. These data can provide more objective measures of well-being compared to official reports of maltreatment, which may reflect policy changes rather than underlying abuse and neglect.

The uses of theory and of these new data sources provide the opportunity to employ econometric tools to answer these important questions. These tools include instrumental variables, which provide plausibly exogenous variation in the likelihood of treatment due to policies and practices of the child welfare system. Regression-discontinuity designs, for example, can be employed in settings where risk scores are used to flag at-risk cases to study the effectiveness of the tool and the programs the tool recommends.

Finally, economists should team with child welfare practitioners and researchers from other fields to implement RCTs when new interventions are being considered. In child protection, experimentation with different forms of foster care, including more informal placements with relatives coupled with family preservation support, is particularly needed and increasingly possible. Economists can push for experimentation with treatments that can reveal underlying behavioral responses—such as how families respond to financial incentives or how investigators respond to social incentives when performance measures are shared within teams—so that we can build on the lessons of these experiments when designing new policies.

The returns to programs that reduce violence within families and childhood neglect could be enormous considering the direct and indirect costs involved. Economic research on these programs should shed light on the underlying behavior of individuals implementing or exposed to them to inform policies that improve child well-being.