Examining the Role of Mandatory Arrest Policies on Child Maltreatment

Examining the Role of Mandatory Arrest Policies on Child

Maltreatment

Introduction and Motivation

Domestic violence in the United States accounts for an estimated 1,200 deaths and two

million injuries to women each year (Black & Breiding, 2008). Early public attitude towards

domestic violence was based on the belief that abuse was best handled within the family and,

while injunctions were available to married women, criminal penalties against the spouse were

negligible (Bourg & Stock, 1994). Recently, the feminist movement, various grassroots

organizations and the research community have significantly shifted the way the criminal justice

system treats and prosecutes perpetrators of domestic assaults. Several alternatives for appropriate

enforcement in response to intimate partner violence have been researched, proposed and adopted

as policy. These alternatives vary from doing nothing to more formal criminal sanctions such as

arrests, restraining orders and coerced treatments (Fagan, 1996). For example, the Violence

Against Women Act of 1994 (VAWA) was the first U.S. federal legislation to acknowledge

domestic violence as a crime (NNEDV Safety Net Project and The Confidentiality Institute,

2015). It created new legal remedies and penalties (e.g. pro arrest policies such as mandatory and

preferred arrest); authorized extensive grants to states, NGOs and federal agencies for various

anti-violence programs; and created a civil rights remedy (Silk & Hurwitz, p. 6).

Much of the research following the emergence of VAWA has focused on the effect of pro-

arrest policies on recidivism rates (Hirschel et al., 2007). However, since not all pro-arrest

policies are the same, policy effectiveness can be influenced by factors such as how the policy is

written and implemented and the degree of support for and monitoring of such policies within the

community (Browne & Williams, 1989). Many advocates view the actual arrest of batterers as

the ultimate goal of mandatory arrest legislation (Fagan, 1996). Interestingly, to date, proarrest

policy research has been slow to examine the net-widening effect of domestic violence arrest

practices (Hirschel et al., 2007). Initially, domestic violence statutes only applied to violence

between couples. The definition today encompasses a wide range of relationships–

current/former spouse and cohabitant, child in common, dating relations and those related by

marriage or blood (National Institute of Justice, 2008).

A great deal of research has focused on the deterrence of sanctions in the context of

intimate partner violence, but few have examined the impact of sanctions on child abuse. Thus, the

goal of this paper is to contribute to the understanding of this complex phenomenon by exploring

the impact of a deterrent, such as mandatory arrest provision, on child maltreatment.

Child abuse is not a new phenomenon; however, it was not until “the battered child

syndrome” was published in 1962 (Kempe, Silverman, Steele, Droegemueller & Silver) that

researchers began to explore the etiology of child abuse. Studies have shown that children who are

abused display signs of negative externalizing behavior, have an increased risk of aggressive

behavior, experience anxiety and depression, and may perpetuate the cycle of violence by

increasing the probability that the child grows up to be a perpetrator or victim of domestic

violence (Summers, 2006; Kernic et al., 2003; Lichter &McClosky, 2004).

The magnitude of this problem is surprising. In 2008, the Child Protective Services

received 3.3 million reports of children being neglected or abused, out of which 772,000 children

were determined to be victims of abuse (Fang et al., 2012). According to the National Child Abuse

and Neglect Database System, 1,640 children died from abuse and neglect in 2012. This translates

to a rate of 2.20 children per 100,000 children in the general population and an average of four

children dying every day from abuse or neglect. The average lifetime cost of nonfatal child

maltreatment in 2008 was $210,012 and the average lifetime cost per death was $1,272,900 (Fang

et al., 2012). Clearly, the impact of child maltreatment generates a significant economic burden for

not only the victims but also, by extension, for society. Thus, child violence is a social issue of

concern for individuals and policy makers alike.

The political potency of issues relating to child and other familial violence has garnered a

certain urgency to the development of new laws and sanctions. Thus, this study seeks to answer

the following research question: Do aggressive sanctions such as mandatory arrest and mandatory

prosecution, serve as effective deterrents to child violence?

In this paper, I develop and present a simple model of violence between children and their

abusers. This is a Stackelberg type model in which the abuser (parent/guardian) maximizes

expected utility subject to the stochastic reaction function of the victim (child). The comparative

statics yield clear predictions about the impact of sanctions and deterrents on child maltreatment.

The predictions are intuitive. As the cost and probability of external intervention rise, the

probability of violence falls.

I follow this model with an empirical estimation of the impact of aggressive sanctions,

such as mandatory arrest, on child victimization. I find a statistically significant and negative

effect of these laws on child fatality rates (controlling for state and year fixed effects). However,

while examining the impact of these laws on child maltreatment (controlling for state trends), I

find a significant but positive relationship. Although this positive relationship contradicts my

model, it could be a consequence of an increase in the reporting of maltreatment and/or reprisal

rate by abusers.

The contribution of my paper is two-fold. First, I provide a theoretical framework that

models child violence, in which I allow for a strategic interaction between the child and his

abuser. In recent years, researchers have developed game theoretic models with respect to

marriage and intimate partner violence (Manser & Brown, 1980; Lundberg & Pollak, 1996; Bloch

& Rao, 2002; Tauchen & Witte, 1991); little attention, however, has been paid to analyzing child

violence. While no model can capture all the complex factors that affect child violence, my goal

is to construct a model that generates testable hypotheses and stimulates further theoretical

research, thus allowing for a better understanding of how sanctions influence child victimization.

My second contribution is that I attempt to bridge the gap between the intimate partner and

child abuse literature by empirically testing the effect of domestic violence statutes on child

maltreatment. Literature on the co-occurrence of child abuse and spousal abuse has been

expanding as more researchers are evaluating the impact of spousal abuse on children. As

previously stated, none have evaluated the impact of sanctions on child maltreatment. Policies and

laws have an important role in the identification and cessation of child maltreatment. Thus, in

order for academics to participate in this important debate, it is necessary to evaluate which

policies are effective at exposing and reducing family violence.

The rest of this paper is organized as follows: Section 1.2 maps the mandatory arrest and

child maltreatment research that is relevant to my paper. Section 1.3 presents a model of child

violence and examines the theoretical effect of interventions on child violence. Section 1.4

presents descriptive evidence. Sections 1.5 and 1.6 present the empirical estimates of the impact

of mandatory arrest policies on reported maltreatment rates and child fatality rates. Section 1.7

concludes.

1.2 Literature Review

Until the 1970’s, the criminal justice system tended to overlook domestic violence cases –

domestic abuse was treated as a private affair and most police agencies discouraged their officers

from arresting perpetrators of domestic violence (Bourg & Stock, 1994).

Mandatory arrest laws, according to Iyengar (2009), emerged largely in response to two

events. First, a court decision in Connecticut found the local police accountable for failure to

adequately respond to a domestic violence incident.

Second, the Minneapolis Domestic Violence Experiment was conducted. It randomly

assigned a police response to domestic violence calls. Researchers found that arrest was the most

effective method police used to reduce domestic violence. The other methods – counseling and

separating the individuals – were considerably less effective at deterring future arrests (Sherman &

Berk, 1984). This study was instrumental in spearheading research and policy evaluation for

mandatory arrest laws. The results of this experiment were used by the US Department of Justice,

academics, legislators, and criminal justice spokespersons to justify and support mandatory arrest

policies (Mignon & Holmes, 1995). For example, the Violence Against Women Act was signed

into law in 1994 as a federal response to violent crimes against women. Since then, according to

the National Institute of Justice (NIJ), 15 states have adopted mandatory arrest policies within the

past two decades.

After Sherman and Berk’s 1984 study, several other regional experiments financed by

NIJ (commonly known as the Spousal Assault Replication Program (SARP)) were conducted to

test the deterrent effect of arrest policies on domestic violence. All the studies paint a complex

picture of the effectiveness of arrest as a deterrent. In 1992, Sherman, Schmidt and others

conducted another controlled experiment using data from the Milwaukee Police Department.

Their study revealed that while arrest deters repeat domestic violence incidences in the short run,

there was no evidence of an overall long term deterrent effect. Moreover, they found a decrease

in violence with groups who were employed, married, or white, but violence increased with arrest

groups who were unemployed, unmarried, high school drop-outs, or African American. Similarly,

Berk et al. (1992) used data from Colorado Springs and found that arrest did have a deterrent

effect on employed batterers but not on unemployed batterers. In contrast, Dunford’s

(1992) study on Omaha, and Hirschel & Hutchison’s (1992) study on Charlotte found that arrest

was no more effective an intervention than mediation or separation. And finally, Pate & Hamilton

(1992) using data from Metro-Dade found that arrest only marginally affected recidivism after six

months.

The findings from these studies are merely snapshots and are limited in that they often

include a single jurisdiction or small sample sizes. Overall, the studies failed to replicate the

Minneapolis findings, and the authors reported inconsistencies in the direction and impact of

arrest in domestic violence cases. Furthermore, when a finding of no effect is reported the readers

have no formal way to assess whether the failure to find an effect was due to the absence of an

effect or to the likelihood that the research design would not find an effect if it did exist (Garner

et al., 1995).

It is argued that mandatory arrest laws can result in a more consistent and punitive

response to domestic violence; it is one form of state intervention that reduces the incidence of

abuse and sends the appropriate societal message to abusers (Forell, 2013). Furthermore, it is

hypothesized that those households residing in states with aggressive legislation tend to have a

lower probability of domestic violence (Dugan, 2003).

In sharp contrast to the hypothesis that mandatory arrest laws may reduce violence and

recidivism rates, Iyengar’s research found evidence to suggest that at the state level mandatory

arrest laws increased the risk of intimate partner homicides.

Using the FBI’s Supplementary Homicide Reports from 1976-2003, Iyengar (2009) tested

to see if mandatory arrest laws affected the levels of domestic violence. A difference-indifference

analysis revealed that intimate partner homicides increased by about 60% in states with

mandatory arrest laws, but familial homicides declined in response to mandatory arrests. In

contrast, there were no significant effects of arrest laws on homicides. Furthermore, the author

suggested that mandating arrest might not deter abusers from killing their victims if it decreased

reporting by victims or increased reprisal by abusers.

Iyengar’s study suggests that mandatory arrest laws may deter reporting, thereby

nullifying the potential deterrence intended by the required arrest. In fact, given current penalties,

the low probability of arrest and a high probability that prosecutors decline such cases, mandatory

arrests laws may also have a negligible effect on future arrests and convictions (Sloan et al.,

2013).

Most of the studies have been beset with methodological concerns that lead to strikingly

different estimated effects of arrest laws on intimate partner abuse. Thus, researchers and

policymakers who are unaware of the differences between studies may make generalizations about

mandatory arrest laws that cannot actually be supported (Zeoli et al., 2011).

While the focus of interest in the domestic violence literature is understanding intimate

partner abuse, more recent inquiry has sought to explore children’s exposure to domestic

violence. Previously children were thought of as being tangential and disconnected to the

violence between their parents, and commonly labeled “silent witnesses” (McIntosh, 2003). More

recent qualitative research, however, has disputed this opinion as researchers have established the

interconnectedness between men’s abuse of women and child abuse (Connolly et al., 2006;

Cunningham & Baker, 2004; Edleson, 1999; Guille, 2004; Hester et al., 2000). McDonald and

colleagues (2006) estimated domestic violence exposure at 15.5 million U.S. youth by calculating

the number of children in two-parent homes where violence had occurred.

Children exposed to domestic violence may also be direct victims of maltreatment. A

number of reviews have examined the co-occurrence of documented child maltreatment in

families where adult domestic violence is also occurring (Hartley, 2002; Appel & Holden, 1998;

Edleson, 1999; McGuigan & Pratt, 2001; Bowker et al., 1988; Shipman et al., 1999; Stark &

Flitcraft, 1988). Appel & Holden (1998) found that there exists a 41% median co-occurrence of

child maltreatment and adult domestic violence in families studied, while a majority of the studies

found a 30% to 60% overlap (Edleson, 1999; McGuigan & Pratt, 2001).

An obvious question confronting researchers and policymakers is: what types of policy

initiatives would promote a child’s wellbeing? In the early 1960s, with the support and

encouragement of the federal government, U.S. states began enacting laws mandating the

reporting of child abuse to government authorities. Some states mandated universal reporting

while others targeted certain professionals (non-universal reporting). Palusci and Vandervort

(2014) reviewed 213 counties in the 18 states that had universal reporting statutes. They found

states with these statutes had higher confirmed report rates for neglect, but none for other

maltreatment types. While universal reporting has been touted as increasing identification of

children who suffer from child abuse and neglect, Palusci and Vandervort (2014) and Mathews

and Bross (2008) suggest that additional reports made may not necessarily imply that more

maltreated children would be found, especially for more serious cases such as physical and

sexual abuse. Shpiegel and colleagues (2013) argue that the lack of sufficient evidence coupled

with broad statutes, which may be ambiguous, can directly or indirectly affect the substantiation

of reports. Thus, when state statutes do not provide comprehensive and understandable guidelines

of what constitutes harm, substantiation decisions become increasingly problematic.

To date, there has been little detailed research conducted on the impact of policies on the

actual maltreatment of children. Certainly, a contextual understanding of the importance of such

provisions on child violence and an empirical analysis that distinguishes child maltreatment cases

from other domestic violence cases is lacking. As such, examining the effect of statutes and

policies that are meant to act as deterrents to child violence remains an issue to be further

explored.

1.3 Theoretical Model

In a seminal contribution, Becker (1968) showed that the most efficient way to deter a

crime is to impose the severest possible penalty (to maintain adequate deterrence) and economize

on the cost of enforcement (Dhami & Nowaihi, 2011). If we applied that proposition to the domain

of family violence, would harsher mandates and policies effectively reduce the probability of

violence?

A more interesting question is do most abusers intend to abuse, or is it an irrational act? If

the abuser has an impulse-control problem, why would a harsh deterrent such as mandatory arrest

or no-drop policy work? One can argue that deterrence works even for those with control

problems because it brings to light the wrongness of the act. Consequentially, systems of morality

and social pressure act as a deterrent. Moral intolerance could also impact legislation. State

legislators that declare the moral impropriety of child abuse may enact stricter sanctions, acting in

concert with their constituents.

The following model is based on the assumption that there is a rational component to

family violence. In this section, I present a simple model of child violence that draws on the

previous work of Tauchen, Witte and Long (1985) who model intimate partner abuse. Like the

authors, I view violence as an instrument for controlling the victim’s behavior, z.

In a game theoretic context, I assume the dominant decision maker to be the parent, who

maximizes his or her expected utility subject to the stochastic reaction function of the child. The

parent imposes a behavior standard (𝑧) on the child and threatens a level of violence (v) if the rules

are not obeyed. Randomness is an essential feature in which violence occurs for instrumental

purposes (Tauchen et al., 1985). The random term ( ) reflects random changes in the child’s

behavior or the parent’s perception of the child’s behavior, or even the child’s misinterpretation of

𝑧  . If not included, the parent would set behavior standard (𝑧) and the child would either obey or

disobey with certainty.

In the remainder of this section I derive: 1) the child’s choice problem; 2) the parent’s

choice problem; and 3) the resulting equilibrium.

1.3.1 The Child’s Choice Problem

To keep this model simple, I assume the child’s utility to be a function of behavior and

violence, with z being the child’s only choice variable. The child evaluates his utility function

before knowing the state of the world in which he will be in.

Formally the child’s utility function is:

𝑈𝑐 (𝑧 + , 𝑣) (1)

𝑈𝑣𝑐 < 0

Uzc<0 beyond some z

The parent sets the desired behavior, 𝑧  and inflicts a level of violence (𝑣) if the child fails to

meet the standard. Note, this model only accounts for violence that is potentially criminal, 𝑣𝑙,

where the CPS would deem as physical abuse or neglect. There are two types of parents that are

not modeled here. First, parents who choose to never be criminally violent or who would never

resort to violence, 𝑣 < 𝑣𝑙. Second, I do not consider parents who are likely to threaten violence

before a policy intervention, but not after. The assumption being that these parents will have the

same response as the parents for whom a threat is still being made after external sanctions or

interventions.

Formally, the parent threatens to inflict violence that would be potentially criminal if:

(2)

I assume the child has a perceived probability density function, 𝑔𝑐( ), for the random

variable that measures the child’s misinterpretation of 𝑧.  I also assume the distribution function

has the following properties: 1) has a zero mean; 2) it is unimodal: where its first derivative is

positive for < 0 and negative for > 0; and 3) it is twice continuously differentiable.

Thus, the probability of violence (π) is:

(3)

For interest, I assume the case for a problem child, wherein the child’s behavior (𝑧 + ) may fall

short of the required behavior (𝑧) , 𝑧 + < 𝑧 .

Then the expected utility of the child is:

E𝑈𝑐 = π𝑈𝑐(𝑧 + , 𝑣) + (1 − π)𝑈𝑐(𝑧 + , 0) (4)

Maximizing EUc by the choice of z yields the following first order necessary condition for the

child’s optimum behavior

(5)

The nature of the problem implies . The child must choose a level of behavior

before he realizes the state of the world, i.e. violence or no violence. Thus, the immediate

disutility he gets from improving his behavior is not conditioned on whether there will be

violence in the future.

(6)

The child weighs the cost of his behavior in terms of his reduced utility versus the benefit of

his behavior which is the reduced probability of violence. For at-risk children so that

equation (6) implies,

(7)

Behavior that the child enjoys or that is not costly i.e. Ucz ≥ 0, never results in violence as

the child obeys willingly or without cost. Thus, U c < 0 focuses on the at-risk children.

Totally differentiating equation (6) with respect to and yields

This yields the comparative static result,

(9)

As the parent raises the behavior standard, the child responds by increasing his optimal

behavior.

The child has a perceived probability density function, 𝑔𝑐( ), for the random variable .

Thus, if  = 0, and if < 0, then we can eliminate children who enjoy behaving or for when the cost

of behaving is zero ( . For such children, the cost of obeying the standard set by the parent

would always be less than the cost of violence incurred by disobeying. Such a child will never fall

below the set standard, guaranteeing no violence. This is not a problem child.

Also, we can eliminate children who are always trying to please. If , then setting a

higher 𝑧  would increase the child’s behavior by more than the set standard. This would result in

the parent increasing the behavior standards, until the child was perfect. Children like this respond

to the threat of violence and by exceeding any increase in 𝑧,  eliminating the probability of

violence Thus, our child of interest is one for which . Such a child naturally responds to

an increase in the behavioral standard, but only partially.

Next, I model the parent’s expectations concerning the child’s reaction to the rules.

1.3.2 The Parent’s Choice Problem

The parent’s utility is a function of the behavior of the child and the cost of sanctions

(fines, imprisonment) given outside intervention, 𝐶𝑃. The cost of outside intervention enters into

the parent’s utility function as it imposes costs such as legal fees, loss of potential income from

being arrested, and disapproval and loss of support from the community

As with the child, I assume the parent has a perceived probability density function, 𝑔𝑝( ),

for the random variable in the child’s behavior. Let π(v) be the probability of violence if the

parent perceives that the child has fallen short of the behavior standard.

(10)

Then, assuming external intervention occurs only when there is violence, the

parent's utility function is:

(11)

Let 𝜌 be the probability of external intervention when there is violence. As noted

previously, violence reflects abuse and neglect that would be legally penalized.

Formally, the parent’s expected utility can be expressed as follows:

𝐸𝑈𝑃 = 𝜌𝜋(𝑣)𝑈𝑝[𝑧 (𝑧) + , 𝐶𝑃] + (1 − 𝜌)𝜋(𝑣)𝑈𝑝[𝑧 (𝑧) + , 0]

+ (1 − 𝜋(𝑣))𝑈𝑝[𝑧 (𝑧) + , 0] (12)

Maximizing 𝐸𝑈𝑃 by the choice of 𝑧  yields the following first order necessary condition for the

parent’s choice problem:

d𝐸𝑈𝑃 𝑝

𝑃

d𝑈𝑃(𝑧, 𝐶) 𝑑𝑧 𝑝

𝑃

𝑝 ,

d󰑧 𝑑𝑧

(13)

The nature of the problem implies . The parent derives immediate utility from

the child’s behavior that is not conditional on external intervention. Then,

(14)

This implies,

(15)

That is, the parent gets utility when the child responds positively to the behavior standard.

Totally differentiating with respect to 𝐶𝑃 and 𝑧  yields

d𝐸𝑈𝑃 d𝐸𝑈𝑃

[𝜌𝑔𝑝( ) (𝑑𝑑𝑈𝐶𝑃𝑝 𝑃

)] 𝑑𝐶

2 𝑃 𝑃 2 d󰑧2 𝑑𝑧 d󰑧 𝑑𝑧2

(16)

Thus,

This leads to an important implication; the parent lowers the behavior standard (or raises the

violence threshold) as the cost of sanctions increase7.

Totally differentiating equation (13) with respect to and :

(11)

This implies,

(12)

Another important implication is that the parent lowers the behavior standard when the

probability of detection or intervention increases. Thus, the parent takes the cost of sanctions and

the probability of external intervention into consideration when setting the behavior standard for

the child.

is negative by the concavity of

. Thus, By the same logic,

1.3.3 Comparative Statics

And,

(14)

The two comparative static results, equations (20) and (21), are highly intuitive; the

probability of external intervention and the costs imposed on the individual by sanctions,

presumably could promote the reduction, or cessation of violent behavior. It would be less likely

for the parent, or any abuser, to engage in violence if he or she perceived the cost of sanctions to

be more certain or more severe. Correspondingly, the theory predicts that implementing sanctions,

such as mandatory or pro-arrest policies – where arrest is certain given the reporting of abuse,

should reduce the true incidence of child maltreatment.

This model can also be extended to allow for violence to be treated as a continuous

variable and for variations in responses across victims. For example, as children get older, they

may become more active and focused in trying to prevent or intervene in the abuse. Older children

are more likely to strike back, leave or report an abusive parent. Thus, one would expect to see the

probability of violence fall with children who are older and less dependent on the parent or the

abuser. This would add another dimension and depth to the model – the abuser responds to the

age of the victim, reducing violence as the child ages.

The relation between sanctions and behavior is complex, and hence it comes as no surprise

that the literature on domestic violence documents mixed findings on the impact of sanctions and

deterrence. The ultimate goal of this model is to stimulate further theoretical research that focuses

on child abuse and prevention, thus enhancing an understanding of how sanctions influence child

victimization.

A theoretical model alone, however, is not enough to determine the relative magnitude

and deterrent effect of sanctions and provisions. Thus, testing the effectiveness and evaluating the

legislative impact of policies that try to curb child violence remains an empirical question.

1.4 Data and Descriptive Evidence

Deterrence theory asserts that the perception of swift, severe, and certain legal sanctions

contributes to an abstention from or reduction of illegal behavior (Heckert & Gondolf, 2000). The

comparative static result of my theoretical model follows in line with this theory, in that, as the

probability of external intervention and cost rise, violence should fall. A mandatory arrest

provision is one such intervention – arrest is certain conditional on reporting. As previously

discussed, a great deal of research has been conducted with regard to these legal sanctions on

intimate partner abuse; however, none have examined the deterrent effects of these provisions on

child abuse. Thus, the primary focus of this research is to test the effectiveness of such sanctions

on child maltreatment rates.

State statistics about child maltreatment are derived from the data collected by the

National Child Abuse and Neglect Data System (NCANDS). The Children’s Bureau analyzes and

publishes the data in an annual report that is available for download on its website. I use these

reports in conjunction with the data provided through Cornell University – the National Data

Archive on Child Abuse and Neglect (NDACAN) to construct a panel dataset from the years

1990 to 2010 for all 50 states (exclusive of District of Columbia).This is the first study to

evaluate multi-state and multi-year comparisons of child maltreatment across the span of 20

years.

NCANDS was established in response to the Child Abuse Prevention and Treatment Act of

1988. As part of the act, it collects data aggregated at the state level through an annual survey. The

survey asks each state to report the number of children who were the subjects of abuse or neglect;

the number of child victims of maltreatment by age, sex and race; the number of reports and

investigations of child abuse and neglect; the number of child fatalities; reporting of abuse to

Child Protective Services (CPS); and other statistics (NDACAN). I calculate maltreatment and

fatality rates using census estimates for state population with respect to age.

The Children’s Bureau defines child fatality as children who have died due to abuse or

neglect, and victims of maltreatment as children who have experienced or who were at risk of

experiencing abuse or neglect. Perpetrator is defined as a parent or caretaker who has maltreated

a child.

Police officers have three tiers of decision making power: full discretion (discretionary

arrest laws); discretion with the state indicating a preference for arrest (preferred arrest laws); and

little to no discretion (mandatory arrest laws). In determining the classification as to which states

have these laws, I use the classification scheme from a study conducted by Zeoli et al. (2011).

Table 1.1 lists the legislative date for each state that passed mandatory and recommended arrest

policies.

I control for state-year factors that are potentially associated with the incidence of family

violence which may in turn affect the child fatality and maltreatment rate. Economic measures

such as median household income and unemployment rates were taken from the US Census

Bureau. Additionally, income-inequality measures such as theil and gini, and human capital index

measures such as high school and college attainment rate, were taken from a state-level panel

data constructed by Frank, Mark (2009). I also control for state crime levels for violent and

nonviolent crimes as reported by FBI’s Uniform Crime Reports, and I account for social controls

such as divorce rates (Wolfers, 2006; CDC).

Other controls include, unemployment rate for males, male-female employment ratio (US

Census Statistical Abstracts); execution (Donohue and Wolfers, 2005); population density (US

Census Statistical Abstracts); and share of prisoners to state population which may be indicative of

police behavior and crime enforcement levels in a given state (US Census Statistical Abstracts).

Table 1.2 gives a summary of the data sources I use in my analysis.

1.5 Empirical Strategy

1.5.1 Parallel Trends and Policy Exogeneity

Prior to the 1980s, the statutory structure for handling domestic violence cases could

charitably be described as a benevolent neglect of a family problem (Buzawa & Buzawa, 1996).

Since the late seventies, statutory changes have sought to mainly alter the official reaction and

response to domestic violence. As noted earlier, there had been a recurrent exercise of discretion

by the criminal justice system to avoid arresting and prosecuting domestic violence offenders.

Current legislation mandating arrests have resulted from the interplay of pressure from feminist

groups, concerned legislators, and professionals in the criminal justice system (Buzawa &

Buzawa, 1996). However, according to the authors, this pressure was not due to the level of

violence, but rather the perceived government treatment of offenders. They also argued that

alternative reforms such as mediation were dismissed as inappropriate or sexist as American

society became more conservative and punitive towards domestic violence offenders.

Stark (1993) posits that the most important reason for passing mandatory arrest laws was to

control police behavior; reducing the level of violence was only of distant concern after the desire

to avoid liability from inaction. For example, after the case of Thurman v City of Torrington,

threats of future lawsuits served as a motivation for municipalities to protect themselves from

liability, creating the desire to monitor and regulate police intervention in domestic violence

incidents (Stark, 1993). This argument is further substantiated by Iyengar (2009), who suggests

that the timing of arrest law passages is tied to the publication of the MDVE results, the promotion

of these results by the Justice department in subsequent years, and finally federal funding of these

policies after 1994. The preceding arguments collectively suggest that the motivation for most of

the mandatory arrest laws does not appear to have been tied to the level of domestic or family

violence.

Another underlying assumption here is that no other event, beside the implementation of

the mandatory arrest policy alters the temporal path of either the treated or control groups. My

data allows me to test and relax this identifying assumption as I look for graphical evidence of

whether the two treatment groups diverged before the passage of the laws. In Figure 1.1, the

dotted line represents the mean of the outcome variable for the two states (Utah and Rhode Island)

that passed the mandatory arrest provision in 2000, with the solid line representing states with no

such provision. As observed in the graph, the trends for child maltreatment prior to 2000 are very

similar across the two groups, lending credibility to the empirical strategy in the next section.

1.5.2 Empirical Model

Using a difference in difference framework, I exploit the variation across states and time

in the implementation of mandatory arrest policies. This allows me to identify the causal effect of

these deterring policies under the assumptions of parallel trends and policy exogeneity.

My two main dependent variables are reported incidences of child maltreatment and child

fatality rates. Since the nature of the database allows for some reporting effect, I expect the

reporting of maltreatment to increase in states with mandatory arrest laws. However, following the

prediction of my comparative static results in Section 1.3.3, I expect the true incidence of

maltreatment to fall. I test this prediction by investigating the impact of mandatory arrest laws on

child fatality rates. The underlying premise of this approach is that, with or without intervention,

child fatality is always reported to the police; it is therefore immune to the reporting effect.

The explanatory variables of interest are states with mandatory arrest and recommended

arrest laws. Since states with mandatory arrest statutes allow for little to no discretion by the

police when making arrests, I predict a larger impact on child maltreatment than states with no

such provisions.

My choice of controls is motivated by previous domestic violence research. I consider state

and demographic controls that could affect familial violence and the decision to report. For

example, schooling, employment, and income may alter the perpetrator’s risk of offending by

increasing the opportunity cost of engaging in violence. As a result, I expect a negative correlation

between these variables and child maltreatment. I also consider the judicial environment that may

help capture the effects of deterrence caused by incarceration and additional policing. Finally, to

allow for policies that are evolving over time, address time shocks and control for heterogeneity, I

include state fixed effects, year fixed effects and state specific linear time trends.

Comparing child maltreatment rates before and after the passage of mandatory arrest laws,

I estimate the impact of these laws on maltreatment rates per 1000 children. I use the following

empirical specification:

𝑦𝑆𝑇 = 𝛽𝑜 + 𝛽1 𝑀𝑎𝑛𝑑𝐴𝑟𝑟𝑆𝑇 + 𝛽2 𝑋𝑆𝑇 + 𝛾𝑇 + 𝜃𝑆 + 𝜑𝑆𝑇 + 𝜖𝑆𝑇 (15)

For each state S in year T, 𝑦𝑆𝑇 is the maltreatment rate per 1000 children; MandArr is an

indicator equal to 1 if a state has the mandatory arrest statute in effect at year T, and 0 otherwise;

𝛾𝑇, 𝜃𝑆 and 𝜑𝑆𝑇 are year fixed effects, state fixed effects and state-specific time trends respectively;

𝑋𝑆𝑇 is a vector of control variables. The coefficient of interest is 𝛽1 which measures the effect of

the mandatory arrest law provision on the child maltreatment rate.

Column (2) of Table 1.4 reports the coefficients from this regression.

Using the same specification as (22), I also estimate the regression where the regressor of

interest is the presence of recommended arrest policies. The recommended arrest effect variable is

defined as 1 in states that have recommended arrest laws in effect at time T. Column (4) of Table

1.4 reports the coefficients from this regression.

1.6 Estimation Results

1.6.1 Main Results

Table 1.4 presents the estimates of the impact of mandatory arrest laws on reported child

maltreatment rates based on the difference in difference framework. Each column represents a

separate regression. I find a positive and highly statistically significant relationship between states

that have implemented mandatory arrest laws and reported child maltreatment rates. The results

suggest that mandatory arrest laws are responsible for an additional 3.33 children being reported

as maltreated per 1000 children. The effect is relatively large in magnitude, suggesting that

mandatory arrest policies result in an approximate 23.6% increase in reported child maltreatment

rates. This may seem surprising; however there are two explanations for the results. The likely

explanation is that reporting of maltreatment increased in states with mandatory arrest laws;

alternatively, recidivism may have increased in these states. Increased reporting lowers the

probability of a child being abused. On the other hand, abusers may blame the victim for being

arrested, thus penalizing the victims with repeat or escalating violence once the abuser is released.

To verify the reporting hypothesis, I use data from NDACAN to examine if states with

mandatory arrest laws indeed see an increase in the reporting of child maltreatment to the CPS.

Column (1) of Table 1.6 reports the results of these estimates. I find a positive and statistically

significant relationship, suggesting that states with mandatory arrest provisions see an

approximate 16.8% increase in the reporting of child maltreatment rates. In contrast, I find a

negative and statistically insignificant coefficient for states with recommended arrest laws. This

supports the theory that in mandatory arrest policy states, the police and the CPS are recording

more of the violence that citizens are reporting to them. Figures 1.3 and 1.4, show the trend in

reporting rates to the CPS by law enforcement personnel.

Although child maltreatment is the most comprehensive measure of violence against

children, it may often go unreported. Since maltreatment mostly occurs in the privacy of the home,

it may be impossible to know what really happened as important facts can be either concealed or

go undiscovered. Child fatality (death due to maltreatment) can serve as a useful proxy and

reliable measure of violent crimes towards children. Since the death of a child is much harder to

conceal, it is highly unlikely that fatalities would go unreported. Accordingly, I estimate the

impact of mandatory arrest laws on child fatality rates.

Column (1) of Table 1.5 reports the results using only state and year fixed effects as

controls. Estimates indicate that mandatory arrest law states see a decrease in child fatalities by

0.38 children per 100,000 children. This corresponds to a 21% reduction in child fatality rates in

states with mandatory arrest laws compared to states with no such policies. The results are

statistically significant at the 5 percent level. However, as column (2) shows, the addition of a

state-specific time trend to the model lessens the effect of these laws and makes the estimates

indistinguishable from zero (from -0.38 and 5 percent significance to -0.051 and insignificant). It

is important to note that the model specified in column (1) underscores the magnitude and

economic significance of mandatory arrest statutes’ impact on child fatality rates. Thus, using

fatality as a proxy for child maltreatment yields valuable information about the true victimization

rate of children.

My second hypothesis assumes that the increase in maltreatment rates may be due to an

increase in the recurrence of child maltreatment. Figure 1.5 looks at the trend in recurrence within

six months of offending from the years 2000 to 2010. One important characteristic can be seen.

There seems to be a general reduction in the risk of recidivism over time. However, this may not

be the result of any particular intervention. In fact, the high rate of offending observed

immediately after the original event is a result of the high-risk people recidivating quickly,

leaving the remaining sample in the risk set (medium and low risk individuals) to recidivate at

different points in time (Kurlychek et al., 2012).

There is also a lack of consensus in the domestic violence literature as to the effectiveness

of these laws on recidivism rates. In fact, most research indicates that arrest may have an effect

in delaying or reducing intimate partner recidivism. Combining data from all five replication

(SARP) studies, Maxwell and colleagues (2001) concluded that arrest only slightly reduced

recidivism. Additionally, Hilton and colleagues (2007) find that by arresting higher-risk

perpetrators a small beneficial effect of arrest, possibly in delaying recidivism, can be realized. In

sum, to reduce recidivism, the best available evidence suggests that a police response,

particularly one that results in an arrest, is the most effective offender-focused solution (Maxwell

& Robinson, 2014).

Although it is likely that both the reporting of abuse and recidivism are operating, I find no

strong evidence to suggest that states with mandatory arrest provisions see an increase in the risk

of recurrence. Collectively, I view these findings as compelling evidence to support the hypothesis

that states with mandatory arrest laws not only see an increase in the reporting of abuse but also a

decrease in the true incidence of maltreatment.

Having established that mandatory arrest laws have a significant impact on overall child

maltreatment, I now estimate the effect of these laws stratified by age of the victims. One

extension of my theoretical model is that as children get older, they may become more active in

trying to prevent or intervene in the abuse. Abused children entering school age also have a

higher probability of being detected by concerned outsiders, thereby reducing the physical contact

between the abuser and the child. When I stratify by age (Table 1.7), I find that the impact of

mandatory arrest laws is considerably stronger for older victims. The estimates however are

insignificant, regardless of age. These findings are consistent with the hypothesis that as children

get older, their abuse has a higher probability of being detected and reported.

1.6.2 Robustness Checks

Table (1.4) columns 7-9 estimate the sensitivity of the results to an alternate specification.

Since the maltreatment rate data is intrinsically a count of child victims subjected to abuse within

a discrete time period, I use the negative binomial likelihood function as a robustness check to

estimate the original specification11. The corresponding regression estimates and patterns of

statistical significance are similar across both models. Evidence from the negative binomial

regression serves as a check on the linear results rather than an alternative to the difference in

difference estimation.

I now conduct other robustness checks to analyze the plausibility of the identifying

assumption. First, it may be that the adoption of mandatory arrest policies coincides with the

adoption of other laws that address family violence, in which case my estimate could suffer from

omitted variable bias. To address this, I run a negative binomial regression and I include controls

for other state statutes such as ‘hard’ no-drop prosecution and universal reporting laws. Table 1.9

reports the estimates of these coefficients. Of these additional controls, none have a statistically

significant impact on child maltreatment. More importantly, the inclusion of these other policies

does not reduce the statistical significance and positive effect of mandatory arrest laws on child

maltreatment rates.

Second, there may be a concern that changes in maltreatment rates preceded the

mandatory arrest policies. To address this, I create a placebo indicator which pretends that the

treatment takes place one year earlier. Accordingly, I run a difference in difference and negative

binomial regression. Table 1.10 reports the results from the estimation. All estimated coefficients

are insignificant and close to zero. Thus, the placebo treatment does not influence changes in

child maltreatment rates. This specification adds plausibility to the assumption that the

maltreatment rates of states with and without mandatory arrest provisions follow a similar trend

before the treatment.

Finally, there may be an interaction effect between arrest and certain demographic

characteristics of the perpetrator such as the male and female unemployment rate, income,

divorce and education. The estimates in Table 1.11 include these potential mechanisms, the policy

change (states with mandatory arrest laws), and the interaction between the two terms. In the

presence of unemployment, mandatory arrest seems to be a promising treatment and seems to

deter all unemployed suspects. This result is of interest because unemployment by many is

thought to increase spousal violence (Berk et al., 1992). Columns 5, 6 and 7 estimate the

interactions of education, income and divorce. While all three estimates have negative signs, none

are statistically significant at conventional levels.

1.7 Conclusion

Motivated by the cyclicality of abusive relationships in a domestic setting, this paper

investigates the causal effect of policy interventions on children who are abused. The theory is

premised on a simple hypothesis of deterrence, predicting that sanctions that are certain and severe

contribute to the overall decline in observed child maltreatment. One example of such a sanction is

a mandatory arrest provision, i.e. a policy that makes arrests certain conditional on reporting.

Using data from the National Child Abuse and Neglect Data System from 1990 through 2010, I

empirically test the impact of these laws on child maltreatment rates. I find that reported

maltreatment rates increased in states with mandatory arrest laws. Evidence from the

OLS estimates for child fatality rates demonstrates that the increase in reported maltreatment is

not due to recidivism but, in fact, more people reporting child abuse to the CPS and law

enforcement. The most important result that emerges from the data, however, is that while

reported abuse increases in states with mandatory arrest laws, the true incidence of maltreatment

actually falls.

It is also important to note that variations across counties in a state, along with their actual

procedural implementation may differ and inadvertently bias my results. Additional studies of

what happens within and among states over time will help determine the true impact of mandated

arrest laws on child maltreatment. Also, once data with more heterogeneous groups of victims

and longer follow-up periods become available, I suggest further investigation into the impact of

interventions on the recurrence of child abuse.

This, and other future research, documenting the impact of various laws and their impact

on abused children can help legislators design customized policies to meet the needs of victims.

More importantly, continued cooperation between society and the criminal justice system to help

identify and prosecute perpetrators is needed to address the pervasive problem of child

victimization.

Figure 1.1 Parallel Trends

Notes: Figures plot group-specific yearly averages for maltreatment rate.

Passed Policy in 2000 states includes Utah and Rhode Island. “No Policy” states are states which had no mandatory

arrest policies in place by 2010.

Figure 1.2 Trends in Child Maltreatment Rates, States with Mandatory Arrest Laws

Notes and Sources: Data is from the Child Maltreatment Reports and the National Data Archive for Child Abuse and

Neglect (NDACAN). The Dashed line marks the timing of the mandatory arrest provision. As of 2010, 15 states have

implemented mandatory arrest policies.

Figure 1.3 Report Source to CPS: Law Enforcement Personnel (percent)

For the solid line, the “zero” marker on the X-axis indicates the year in which the state passed the mandatory arrest

policy. The non-solid lines, the “zero” marker on the X-axis indicates the year in which the state passed the

recommended arrest policy. See Table 1.1 for state-specific legislative dates.

Figure 1.4 General Trend in the Reporting of Child Abuse by Law Enforcement

Figure 1.5 General Trend in Percent Recurrence of Abuse within 6 months (1999-2010)

Note: Figure 1.4 shows a general trend in the reporting of child abuse from 1990-2010. Figure 1.5 shows the general

trend in the recurrence of abuse from 1999-2010. The solid line indicates states which have passed mandatory arrest

laws. The non-solid lines represent reporting rates of states that have discretionary and recommended arrest policies.

See Table 1.1 for state-specific legislative dates.

Figure 1.6 Victimization Rates per 1,000 Children, by Age Cohort

For the solid line, the “zero” marker on the X-axis indicates the year in which the state passed the mandatory arrest

policy. The non-solid lines, the “zero” marker on the X-axis indicates the year in which the state passed the

recommended arrest policy. See Table 1.1 for state-specific legislative dates.

Figure 1.7 Victimization Rates per 1,000 Children, by Gender

For the solid line, the “zero” marker on the X-axis indicates the year in which the state passed the mandatory

arrest policy. The non-solid lines, the “zero” marker on the X-axis indicates the year in which the state passed the

recommended arrest policy. See Table 1.1 for state-specific legislative dates.

Figure 1.8 Perpetrators by Relationship to Victims (2010) & Victims by Age Cohort (2010)

Source: Child Maltreatment Report (2010)

Figure 1.9 Reported Maltreatment by Type (2010)

Source: Child Maltreatment Report (2010)

Table 1.1 States with Mandatory Arrest, Recommended Arrest and Discretionary Arrest Policies

State

Policy

State

Policy

Discretionary

Mandatory 1996

Mandatory 1989

Recommended 1991

Discretionary

Mandatory 1991

Recommended 1993

Discretionary

Mandatory 1994

Recommended 1994

Mandatory 1987

Discretionary

Mandatory 1991

Discretionary

Recommended 1994

Discretionary

Mandatory 2001

Discretionary

Mandatory 2000

Discretionary

Recommended 2002

Discretionary

Mandatory 1998

Mandatory 1990

Discretionary

Recommended 2000

Discretionary

Mandatory 2000

Discretionary

Mandatory 1995

Mandatory 2002

Discretionary

Mandatory 1999

Discretionary

Mandatory 1996

Discretionary

Recommended 1995

Recommended 1989

Discretionary

Source: Zeoli et al. (2009)

Table 1.2 Summary of Data Sources

Variables

Definitions

Sources and years

Dependent Variables

Child maltreatment

Children who have experienced

or who were at risk of

experiencing abuse or neglect.

NDACAN (1990-1999)

Children’s Bureau (2000-2010)

Child fatality rate

Children who have died due to

abuse or neglect

Reporting of maltreatment

A report source is defined as the

category or role of the person

who notified a CPS agency of the

alleged child maltreatment

Explanatory variables of Interest

Classification of states with mandatory

arrest and recommended arrest

provisions

Mandatory arrest states are states

which require an arrest

conditional on a report of

domestic violence.

Recommended arrest states are

states where officers are

instructed but not required to

make a warrantless arrest.

Zeoli et al. 2011 “Mandatory,

Preferred, or Discretionary: How

the classification of Domestic

violence warrantless arrest laws

impact their estimated effects on

intimate partner homicide”

Family and State Environment

Employment population ratio

Male unemployment rate

Unemployment rate

Median Household income

Poverty rate

Population density per square mile

(Proxy for urban rate)

(Total population/ Land area)

U.S. Census Bureau - Statistical

Abstracts Series,

Bureau of Labor Statistics

Divorce rate

Wolfers, Justin. 2006. Did

Unilateral Divorce Raise

Divorce Rates? A Reconciliation

and New Results (1990 – 2000)

CDC divorce rates (2000-2010)

Fraction of child population that is white

Fraction of child population that is black

College attainment rate

High school attainment rate

Theil index

Information about the victim’s race

U.S Census Bureau -Current

Population Surveys

Human Capital Index Measures

Income-Inequality Index

Frank, Mark. W. 2009 "Income

Inequality, Human Capital, and

Income Growth: Evidence from

a State-Level VAR Analysis."

State Judicial Environment

Death penalty & execution rate

Wolfers, Justin. 2006. Uses and

Incarceration rates

Prisoner to population ratio

Abuses of Empirical Evidence

in the Death Penalty Debate

U.S. Census Bureau - Statistical

Abstracts Series

Crime rate

Crime to population ratio

FBI Uniform Crime Report

(1990-2010)

Variable

Obs

Mean

Std. Dev.

Min

Max

Maltreatment rate per 1000 children

968

14.15387

8.22681

0.688966

82.62695

Victims ages 0-3

960

15.55099

8.962154

1.072743

113.9011

Victims ages 4-11

960

12.48899

7.530598

1.158864

95.88476

Victims ages 12-15

960

11.36321

6.924545

1.169128

88.93894

Reporting rate per 1,000 persons

953

6.92981

2.827277

1.777521

29.07375

Fatality rate per 100,000 children

963

1.770668

1.079786

7.863395

State-year controls:

Divorce rate

865

4.384362

1.208081

10.8

Median household income (2012)

1050

52819.17

8417.866

32018

77506

High school attainment rate

1050

0.543162

0.04365

0.385799

0.637845

College attainment rate

1050

0.160779

0.040332

0.07519

0.342751

Employment population ratio

850

63.70447

4.400458

49.1

73.3

Unemployment rate, males

850

5.650471

2.125332

15.8

Execution rate

950

0.037048

0.096375

0.961847

Prisoner to population ratio

1050

0.003491

0.001509

0.000444

0.008856

Table

1.3

Summary Statistics

Total crime to population ratio

1050

0.009219

0.004609

0.001307

0.024887

Population density per sq. mi.

1050

180.3238

245.2472

0.96

1186.41

Fraction of child pop. that is black

1023

0.128014

0.114996

0.003415

0.469574

Fraction of child pop. that is white

1023

0.699382

0.190164

0.127249

1.023237

Theil index

1050

0.756171

0.173633

0.443031

1.487246

Table 1.4 Effect of Mandatory and Recommended Arrest Laws on Child Maltreatment Rates

Dependent variable mean 14.154

Mandatory arrest law

effect

1.558

(0.969)

3.334**

(1.566)

0.1689** (0

.0755)

Recommended arrest law

effect

2.8043

(2.855)

5.0243

(4.483)

0.2049

(0.1797)

Mand or Recom

arrest law effect

2.019*

(1.089)

3.961**

(1.761)

0.217**

(0.093)

Estimation method

OLS

Neg. Bin.

Neg. Bin

Controls for other violent

crimes

State-year economic &

social controls

State demographic

controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

588

549

547

588

549

547

R-squared

0.697

0.774

0.704

0.772

0.712

0.784

Notes:

The dependent variable for each column is the column title per 1000 children. Robust standard errors are reported in parentheses. Coefficients that are significant

at the .01, .05 .1 percent levels are marked with ***, **, *.

Mandatory arrest states are states which require an arrest conditional on a report of domestic violence. Recommended arrest states are states where officers are

instructed but not required to make a warrantless arrest when a domestic violence offense is reported.

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Child Maltreatment Rate per 1000 children

Mand or Recom arr states are states with either mandatory arrest or recommended arrest statutes.

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants. Indexed crimes included in the violent crime

variable are murder, robbery, assault, and rape.

State economic control variables include the variables: male unemployment rate, employment-population ratio, and state median household income (BLS and

US Statistical Abstracts), college and high school attainment rate, and the theil income inequality index (Frank, 2009), population density per square mile (U.S

Statistical Abstracts).

State social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty (number of state executions by year).

State demographic controls are based on the March Current Population Survey and include variables for the fraction of the child population that is black and

white.

Table 1.5 Effect of Mandatory and Recommended Arrest Laws on Child Fatality Rates

Child fatality rate per 100,000 children

Dependent mean

1.771

Mandatory arrest

-0.380**

-0.0508

-0.216**

(0.180)

(0.236)

(0.0977)

Recommended arr.

-0.183

-0.248

-0.114

(0.192)

(0.252)

(0.112)

Mand or Recom arr.

-0.339**

-0.121

-0.196**

(0.144)

(0.180)

(0.08)

Estimation method

OLS

Neg. Bin.

Neg. Bin

Neg. Bin.

Controls for other violent

crimes

State-year economic & social

controls

State fixed effects

Year fixed effects

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

State-specific time trend

Observations 538 538 538 538 538 538 538 538 538

R-squared 0.451 0.575 0.447 0.576 0.451 0.576 - - -

Notes:

The dependent variable for each column is the child fatality rate per 100,000 children. Robust standard errors are reported in parentheses. Coefficients that are

significant at the .01, .05 .1 percent levels are marked with ***, **, *.

Mandatory arrest states are states which require an arrest conditional on a report of domestic violence. Recommended arrest states are states where officers are

instructed but not required to make a warrantless arrest when a domestic violence offense is reported.

Mand or Recom arr states are states with either mandatory arrest or recommended arrest statutes.

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants. Indexed crimes included in the violent crime

variable are murder, robbery, assault, and rape.

State economic control variables include the variables: male unemployment rate, employment-population ratio, and state median household income (BLS and

US Statistical Abstracts), college and high school attainment rate, and the theil income inequality index (Frank, 2009), population density per square mile (U.S

Statistical Abstracts).

State social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty (number of state executions by year).

State demographic controls are based on the March Current Population Survey and include variables for the fraction of the child population that is black and

white.

1.6 Effect of Mandatory and Recommended Arrest Laws on the Reporting of Child Abuse

Reporting of child abuse per 1000 persons

Dependent variable mean

6.93

Mandatory arrest law effect

1.167**

0.151**

(0.586)

(0.0615)

Recommended arrest law effect

-0.602

-0.116

(0.546)

(0.0748)

Estimation method

OLS

Neg. Bin.

All controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

527

R-squared

0.834

0.832

Notes:

Each column represents a separate regression. The dependent variable for each column is the Reporting of child

abuse rate per 1000 persons. Robust standard errors are reported in parentheses. Coefficients that are significant at

the .01, .05 .1 percent levels are marked with ***, **, *.

All regressions include State FE, Year FE, State economic and social policy controls, Crime controls, and a

statespecific time trend.

All controls include:

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants.

Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

State economic control variables include the variables: male unemployment rate, employment-population ratio, and

state median household income (BLS and US Statistical Abstracts), college and high school attainment rate, and the

theil income inequality index (Frank, 2009), population density per square mile (U.S Statistical Abstracts). State

social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty

(number of state executions by year).

(1)

(2)

(3)

(5)

Table

State demographic controls are based on the March Current Population Survey and include variables for the

fraction of the child population that is black and white.

Tables 1.7 Effect of Mandatory Arrest Laws on Child Maltreatment Rates (Stratified by Age)

Maltreatment rate/1000

children by age

(1)

(2)

(3)

(4)

(6)

(8)

Ages 0-3

Ages4-11

Ages12-15

Ages 0-3

Ages4-11

Ages12-15

Dependent variable mean

15.55

12.48

11.36

Mandatory arrest effect

1.710

2.549

2.288

0.0585

0.0952

0.0821

(1.951)

(1.740)

(1.590)

(0.0889)

(0.0851)

(0.0837)

Estimation method

OLS

Neg. Bin

Observations

527

R-squared

0.807

0.786

0.778

Notes:

The dependent variable for each column is the column title per 1000 children. Robust standard errors are reported in

parentheses. Coefficients that are significant at the .01, .05 .1 percent levels are marked with ***, **, *.

Mandatory arrest states are states which require an arrest conditional on a report of domestic violence.

Each column represents a separate regression. All regressions include state by year controls, state fixed effect, year

fixed effect, and a state specific time trend

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants.

Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

State economic control variables include the variables: male unemployment rate, employment-population ratio, and

state median household income (BLS and US Statistical Abstracts), college and high school attainment rate, and the

theil income inequality index (Frank, 2009), population density per square mile (U.S Statistical Abstracts). State

social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty

(number of state executions by year).

State demographic controls are based on the March Current Population Survey and include variables for the

fraction of the child population that is black and white.

1.8 Effect of Recommended Arrest Laws on Child Maltreatment Rates (Stratified by Age)

Maltreatment rate/1000

children

(1)

(2)

(3)

(4)

(6)

(8)

Ages 0-3

Ages4-11

Ages12-15

Ages 0-3

Ages4-11

Ages12-15

Dependent variable mean

15.55

12.48

11.36

Recommended arrest effect

0.194

1.083

0.966

0.0277

0.0823

0.0788

(1.145)

(1.172)

(1.087)

(0.0701)

(0.0837)

(0.0868)

Estimation method

OLS

Neg. Bin

Neg. Bin.

Neg. Bin

Observations

527

R-squared

0.806

0.784

0.777

Coefficients that are significant at the .01, .05 .1 percent levels are marked with ***, **, *.

Recommended arrest states are states where officers are instructed but not required to make a warrantless arrest

when a domestic violence offense is reported.

The dependent variable for each column is the column title per 1000 children. Robust standard errors are reported in

parentheses.

Each column represents a separate regression. All regressions include state by year controls, state fixed effect, year

fixed effect, and a state specific time trend

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants.

Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

State economic control variables include the variables: male unemployment rate, employment-population ratio, and

state median household income (BLS and US Statistical Abstracts), college and high school attainment rate, and the

theil income inequality index (Frank, 2009), population density per square mile (U.S Statistical Abstracts). State

social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty

(number of state executions by year).

State demographic controls are based on the March Current Population Survey and include variables for the

fraction of the child population that is black and white.

Table

1.9 Robustness Check: Effect of Mandatory Arrest, Hard No Drop Prosecution Policies, And

Mandatory Reporting Of Child Abuse for All Persons on Child Maltreatment Rate

Dependent variable mean

14.154

Mandatory arrest law

0 .1703 **

(0 .076)

Hard no-drop prosecution policy

-0.0695

(0 .071)

Mandatory reporting law

2.625

(60.057)

Estimation method

Negative Binomial

All controls

Observations

588

R-squared

Notes:

The dependent variable for each column is the Child maltreatment rate per 1000 children. Robust standard errors are

reported in parentheses. Coefficients that are significant at the .01, .05 .1 percent levels are marked with ***, **, *.

Mandatory arrest states are states which require an arrest conditional on a report of domestic violence.

Hard no drop policy: is one in which the state will push forward a prosecution using all means available.

Mandatory reporting laws: any person who suspects child abuse or neglect is required to report.

In addition to hard no-drop policy and mandatory reporting by all persons laws.

Each column represents a separate regression. The regressions include all state by year controls in addition to, state

fixed effect, year fixed effect, and a state specific time trend.

1.10 Robustness Check: Effect of Mandatory Arrest Effect on Child Maltreatment Rate

Using Placebo Indicator

(1)

Child maltreatment rate per1000 children

Placebo Mandatory arrest effect

.05617

(0 .0687)

Estimation method

Negative Binomial

Other state controls

Child Maltreatment Rate per 1000 children

Table

Observations

547

Notes:

Placebo Mandatory arrest: A Placebo indicator which pretends that the treatment (policy) is implemented one year

earlier.

Each column represents a separate regression. The regressions include all state by year controls in addition to, state

fixed effect, year fixed effect, and a state specific time trend.

Table

1.11 Robustness Check: Interaction Effects between Mandatory Arrest and Demographic

Characteristics

(1)

(2) (3) (4) (5)

(6)

(7)

Child maltreatment rate per 1000 children

Dependent mean

14.154

Mandatory arrest

0.16797**

0.459***

0.3642**

0.4086**

0.2166

0.2698

0.3197

(0 .07547)

(0.1165)

(0.1526)

(0.164)

(0.2968)

(0.5457)

(0.2723)

MandXunemp

- 0.059***

(0.0211)

mandXunempm

-0.0346

(0.0232)

mandXunempf

-0.05146

(0.0356)

mandXcollege

-.32376

(1.902)

mandXHHincome

-1.70e-06

(9.70e-06)

mandXdivorce

-0.03468

(0.0605)

Observations

588

549

Notes:

Each column represents a separate regression using the negative binomial estimation method. Column (1) gives the

estimates of the baseline regression.

The interaction terms include demographic characteristics of the perpetrator and mandatory arrest laws:

unemployment rate (unemp), male and female unemployment rates (unempm & unempf), education (college),

median household income (HHincome), and divorce.

All regressions include State FE, Year FE, State economic and social policy controls, Crime controls, and a

statespecific time trend. See notes from table (1) All controls include:

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants.

Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

Table

State economic control variables include the variables: male unemployment rate, employment-population ratio, and

state median household income (BLS and US Statistical Abstracts), college and high school attainment rate, and the

theil income inequality index (Frank, 2009), population density per square mile (U.S Statistical Abstracts). State

social policy controls are taken from Wolfers (2006) and include divorce rates, and indicators for the death penalty

(number of state executions by year).

State demographic controls are based on the March Current Population Survey and include variables for the

fraction of the child population that is black and white.

1.12 Effect of Arrest Policies on Reported Child Maltreatment Rates per 1,000 children

4 5 6 7

14.154

Child maltreatment rate per 1,000 children

Dependent mean

Mandatory arr.

1.558

3.334**

(0.969)

(1.566)

Recommended arr.

2.8043

5.0243

(2.855)

(4.483)

Mand/Recom arr.

2.019*

3.961**

(1.089)

(1.761)

Recom vs Disc arr.

4.172

5.112

(2.98)

(4.99)

Mand vs Disc arr.17

1.206

3.431**

(1.080)

(1.589)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

588

549

547

377

461

R-squared

0.697

0.77

0.70

0.77

0.71

0.78

0.70

0.79

0.74

0.81

Table

Table 1.13 Effect of Arrest Policies on Child Fatality Rates per 100,000 Children

3 4 5 6 7

Child fatality rate per 100,000 children

Dependent mean

1.771

Mandatory arr.

-0.380**

-0.0508

(0.18)

(0.236)

Recommended arr.

-0.183

-0.248

(0.192)

(0.252)

Mand/Recom arr.

-0.339**

-0.121

(0.144)

(0.18)

Recom vs Disc arr.

-0.229

-0.255

(0.193)

(0 .262)

Mand vs Disc arr.22

-0.392**

-0.094

(0.185)

(0.248)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

538

376

451

R-squared

0.451

0.575

0.447

0.576

0.451

0.576

0.502

0.61

0.43

0.56

Chapter 2: The Effect of Medical Marijuana Laws on Child Maltreatment:

Evidence from State Panel Data, 1995-2014

2.1 Introduction and Motivation

It has been argued that marijuana use is part of a post-modern consumer culture which

crosses class, gender, race, age and geographic boundaries (Simpson, 2003). Marijuana is also

the most widely used illicit drug in the United States. In 2014, 27 million people aged 12 or older

used an illicit drug in the past 30 days – these estimates are driven primarily by marijuana use

with 22.2 million Americans reporting the use of marijuana in the past 30 days (SAMHSA,

2014). Public opinion on marijuana has shifted considerably in the last 40 years, with the

majority of Americans (53%) now favoring legalization (PEW Research, 2015). Furthermore, a

2014 Gallup Poll found that 76 percent of Americans favored no jail time for those convicted of

minor marijuana possession (PEW Research, 2014). This broader social acceptance of the drug

has been reflected by policies being implemented at the state level that have allowed for the use

of marijuana to be decriminalized, legalized, or approved for medicinal purposes in 23 states plus

the District of Columbia.

How does the legalization of medical marijuana affect child maltreatment?

We might expect that children living with substance-abuse caregivers may experience a

greater risk for maltreatment. In fact, the 2010 Fourth National Incidence Study found that illegal

drug use was a factor in 9.5% of cases of physical abuse and about 12.5% of all neglect cases

(Sedlak et al., 2010). Parents who have substance use problems are more likely to contribute to

severe family dysfunction, be physically abusive, and commit child neglect than those without

diagnosed substance abuse problems (Ammerman et al., 1999; Appleyard et al., 2011).

Caregiver substance misuse has also been documented as a predictor of severity in child

maltreatment cases (Sprang, Clark & Bass, 2005; Staton-Tindall, Sprang & Straussner, 2016).

Some studies suggest that cannabis may act as a gateway drug that encourages other forms of

illicit drug use such as cocaine and heroin or alcohol use (see Jeffery DeSimone, 1998; Hall &

Lynskey, 2009; Wen et al., 2015). While little attention has been paid to marijuana use and

family violence, the link between illicit drug abuse, alcohol abuse and child maltreatment has

been well documented. For example, Famularo and colleagues (1992) find specific associations

between alcohol use and physical maltreatment, and cocaine abuse and sexual maltreatment.

Considering the implications of a gateway effect, marijuana use could indirectly elevate the risk

of child maltreatment. The gateway effect is one of the principal reasons cited in defense of laws

prohibiting the use or possession of marijuana (Morral, McCaffrey & Paddock, 2002). Despite a

number of scientific studies disputing this claim (see DeSimone, 1998, Tarter et al. 2016), the

debate over the most appropriate policy has been generally polarized due to differing positions

on the drug’s harm.

Though clinical trials have demonstrated the benefits of cannabis in alleviating chronic

and neuropathic pain, other scientific studies have indicated significant physical and

psychotropic side-effects of the drug (Leung, 2011). Regular marijuana use has been linked to

adverse health outcomes, including mental slowness, short-term memory loss, impaired reaction

times, and accentuation of anxiety and depression (Crean, Crane & Mason, 2011; Cellucci,

Jarchow & Hedt, 2004). Chronic use of marijuana in the long run increases the risk for a number

of psychosocial outcomes including diminished relationship quality, lower satisfaction with life,

and greater need for economic assistance (Dubowitz et al., 2015; Volkow et al., 2014). These

effects can often lead to an unstable and chaotic environment for children, in which case basic

needs such as nutrition, supervision and medical care may go unmet (Staton-Tindall et al., 2016).

Similarly, parents with depression and anxiety disorders are less likely to prevent injury and harm

to their children and more likely to exhibit stress or aggravation during parenting (Chung et.al.,

2005).However, mental health problems cannot be causally connected to involvement in drug or

marijuana use, even if it can be illustrated that its (ab)use may exacerbate pre-existing psychiatric

disorders (Simpson, 2003; Crome, 1999).

So far, studies examining the link between marijuana use and psychosocial disorders have

not addressed the nature of the following relationship: does marijuana use lead to such disorders

or do issues such as anxiety and depression lead to the (over-) use of marijuana? In fact, one

qualitative study found that parents who used marijuana reported that the drug improved their

parenting by allowing them to relax and manage difficult emotions relating to parenting –

thereby preventing them from yelling at or hitting their children (Thurstone et al., 2013). It must

be noted that these results are preliminary in nature and must be interpreted with caution as they

only include data from 11 parents in five focus groups. Additionally, conflicting results among

most of the studies seem to be a result of differences in the degree of exposure, individual

sensitivity, and drug potency (e.g., CBD/THC ratio) (Niesink & Laar, 2013). Thus, research

focused on the benefits and consequences of marijuana use merits further investigation.

There is a growing body of empirical research examining the link between illegal drug

consumption and intimate partner violence. However, due in part to differences in research

design, the empirical support for this notion is rather mixed. While similar efforts exist with

respect to marijuana use and child abuse (see Friesthler, Gruenewald & Wolf, 2014), no single

analysis has assessed the overall impact of medical marijuana laws (MML) on child

maltreatment rates across the United States.

Several studies have found that legalizing medical marijuana and decriminalizing its use

leads to greater access and use of the drug (see Cerda et al., 2012; Anderson, Hansen & Rees,

2015; Freisthler et al., 2013). For example, at the state level, Pacula et al. (2013) conclude that

states which allowed medical marijuana distribution through dispensaries or home cultivation

had higher levels of past month marijuana use than states with no such laws. Additionally,

legalization reduces the need for judicial and correctional spending on marijuana related offenses

and facilitates the reallocation of police resources toward other violent crimes such as domestic

violence. Thus this paper seeks to study the link between marijuana use and violence aimed at

children, with the main purpose of examining the role that changes in marijuana legislation may

play in the incidence of child abuse.

This paper begins by extending the current MML-crime literature by providing a

comprehensive evaluation of the impact of MMLs implemented at the state level on reported

child victimization rates. I show that specific dimensions of medical marijuana regulation

differentially influence the magnitude of reported incidences of child abuse, a finding which

sheds new light on the current literature. More specifically, using fixed effects analysis applied to

data from the Child Maltreatment Reports (1995-2014) and the FBI’s Uniform Crime Reports

(UCR), I show that states that allow for home cultivation in addition to decriminalizing its use

see an increase in the magnitude of reported incidences of child maltreatment rates This, of

course, does not mean that marijuana legislation caused an increase in maltreatment rate. As

noted in the first chapter, two factors influence reported incidences of abuse: actual maltreatment

and the proportion of maltreated cases that are reported. Establishing a distinction between the

two definitions can help understand the true impact of these laws.

There is an apparent gap between the enforcement of child safety laws and marijuana

statutes. For example, there are no formal guidelines instructing child welfare professionals on

how to handle cases where marijuana use has been recommended by a physician. Furthermore,

reports of abuse and neglect that come to the attention of Child Protective Services (CPS) do not

differentiate between specific substances used; thus very little is known about which specific

drugs may be more likely to result in maladaptive parenting behaviors (Freisthler et al., 2015).

Indeed, marijuana is still considered as an illicit drug by many professionals, and anecdotal

evidence exists where CPS workers have removed children or denied custody because of the

parents’ legal use of marijuana. This speculation of child endangerment due to marijuana use can

lead to an increase in the reporting of child maltreatment cases.

To test the reporting hypothesis, I use an alternative proxy for maltreatment rates that is

less likely to be biased by reporting: rates of child fatality from abuse and neglect. Child

fatalities must always be reported, and using an extreme form of an incident is a common

strategy among economists studying crime (see Iyengar, 2011; Levitt, 1998). I find that states

that allow for home cultivation in addition to decriminalizing its use see a decrease in child

fatality rates. This is obviously an imperfect proxy for overall maltreatment, but the fact that

there is a consistent decline in fatality rates in states with marijuana regulation is evidence that

states with MML do see a reduction in actual maltreatment.

Given that there is limited research on the relationship between marijuana consumption

and child maltreatment, estimating the impact of various marijuana laws remains crucial. This

paper improves on the existing literature in that it is the first to analyze the impact of drug

regulatory variables on child victimization in a nationally representative, state-level panel dataset

spanning 19 years.

The remainder of the paper is organized as follows. Section 2.2 provides a brief historical

background of MMLs in the U.S. Section 2.3 summarizes the limited research examining the

impact of MML on crime in general and, more specifically, family violence. Sections 2.4 and 2.5

discuss data sources and methodology. I present the results from my analyses of the impact of

these laws on child victimization rates in Section 2.6. Finally, Section 2.7 concludes.

2.2 Background

2.2.1 Cannabis the Drug

Cannabis is largely derived from the female plant of cannabis sativa, with the two main

active ingredients being delta-9-tetrahydrocannabinol (THC) and Cannabidiol (CBD).THC

accounts for both the physical and psychotropic effects of cannabis, and hence is also the most

widely studied. The mechanisms by which CBD exerts its effect are not precisely known, and by

itself has almost no effect on normal physiological processes (Niesink & Laar, 2013). Not much

is known of the safety and side effects of CBD either. Few studies have described the effects of

CBD for therapeutic applications in clinical trials (Bergamaschi et al., 2011). While there is

evidence from controlled trials that cannabinoids are effective in relieving nausea, alleviating

severe pain, and improving appetite in people with HIV and cancer-related illnesses

(Bergamaschi et al., 2011), chronic cannabis use is also associated with psychiatric toxicity and

long-term psychiatric conditions (Reece, 2009). However, to date, there is no conclusive

evidence to support the relationship between chronic cannabis use and the occurrence of

psychosis. In fact, very few studies that have been published distinguish between the types of

cannabis used, and none have given an indication of the THC/CBD ratio (Niesink & Laar, 2013).

2.2.2 A Brief History of Medical Marijuana Laws (MMLs) in the United States

America’s connection with cannabis dates back to the early 1600s. The cultivation of

cannabis (hemp) was the primary reason for America’s colonization; it was produced initially by

Jamestown settlers who were ordered by King James I in 1619 to grow 100 plants specifically

for export (Deitch, 2003). Hemp cultivation remained a prominent industry until the mid-1800s,

and throughout this period, the plant was commonly used by physicians to treat a broad spectrum

of ailments (Anderson, Hansen & Daniel, 2013; Pacula et al., 2002). From 1850 to 1942,

marijuana was listed in the United States Pharmacopeia and National Formulary; the official list

of recognized medicinal drugs (Anderson, Hansen & Daniel, 2013). However, in 1937, the

Marihuana Tax Act –which did not criminalize marijuana but did impose prohibitive taxes on its

use – was passed after research indicated a link between marijuana use and crime (Deitch, 2003).

Since then, several other laws were signed, including the Boggs Act (1952), the Narcotic Control

Act (1956) and the Federal Controlled Substances Act (1970), which effectively discontinued the

use of marijuana for medicinal purposes and ultimately criminalized it at the federal level (Blitz,

1992; Deitch, 2003).

The Controlled Substances Act classified marijuana as a Schedule I drug with high

potential for abuse and no currently accepted medical uses in treatment. In 1973, Oregon

became the first state to decriminalize cannabis – whereby possession of one ounce or less was

treated as a misdemeanor with no jail time. By 1978, Nebraska became the eleventh state to pass

the decriminalization legislation. During the Reagan administration, however, Congress passed

several anti-drug legislation bills, which effectively ended the wave of states decriminalizing the

possession of marijuana.

In 1996, California became the first state to legalize medical marijuana by passing the

Compassionate Use Act (California Proposition 215). It removed criminal penalties for using,

possessing and cultivating marijuana for medical purposes. The law provided immunity from

criminal prosecution or sanction to physicians who recommended or prescribed marijuana to

their patients. Despite federal restrictions, since 1996, 23 states have adopted medical marijuana

laws, instituting their own specific restrictions for use, cultivation, possession limits, and

allowance of dispensaries. While some states did allow doctors to prescribe marijuana before

1996, it had no practical effect since it was against federal law for pharmacies to distribute the

drug (Anderson, Hansen & Daniel, 2013).

2.3 Literature Review

2.3.1 Marijuana and Crime

While there is so far a dearth of literature reporting on the impact of medical marijuana

laws on child maltreatment rates, considerable research has been conducted on the broader

subject of the impact of marijuana on negative life outcomes, criminal proclivity, and violent

behavior.

As marijuana is a psychoactive substance, behavioral changes following its use are

common and expected. The nature of some of these behavioral changes is subject to widespread

academic debate, particularly when hostile or violent behavior is in focus. Some studies find that

short-term use of marijuana may actually inhibit aggressive behavior, while long-term use can

alter the nervous system in a way that augments violent proclivities (National Research Council,

1993). However, other studies have found socio-emotional deficits in marijuana users, including

patterns of interpersonal withdrawal, hostility, and diminished interpersonal skills (Platt et al.,

2010; Clopton et al., 1979; Roser et al., 2012). For example, Ansell et al. (2014) found hostile

and impulsive behaviors as well as perceptions of hostility in others to be associated with

marijuana use among subjects with no reported substance dependence. These findings were

supported by Theunissen et al. (2012) who found that infrequent users of marijuana experience

stronger effects on attention and inhibition following marijuana use relative to chronic users.

However, further research is needed to examine whether these associations are causal since

increases in interpersonal hostility could act an acute stressor that motivates the use of marijuana.

It is well documented that there is an association between illegal drugs and crime, but

knowledge of this association alone may not be helpful in guiding policy designed to reduce

violent crime (Markowitz, 2005). One might expect that marijuana use causes acceleration in

criminal involvement; however, the empirical support for this notion is rather mixed. Bennet et

al. (2008) conducted a systematic review of the literature and a meta-analysis of the strength of

the relationship between types of drug used and criminal behavior. Results of the review of the

ten studies found that marijuana users were 1.5 times more likely than non-marijuana users to

commit a criminal offense. However, the authors found the relationship to be weaker than hard

drugs such as heroin, crack, and cocaine.

To the extent that marijuana legalization may increase a broader acceptance and use of

the drug, these policies could then help identify the potential causal relationships between

marijuana access, use, and societal costs or benefits. Indeed, several studies have concluded that

there is a positive association between MMLs, decriminalization, and consumption, and these

policies could potentially shift its availability to adolescents. For example, Thurstone, Lieberman

and Schmiege (2011) examined the prevalence of medical marijuana use among 80 adolescents

in a substance treatment program in Denver, Colorado. Their study found that a large portion of

the participants (48%) reported acquiring marijuana from someone with a medical marijuana ID

card, concluding that persons with such ID cards may provide a source of supply for teenagers.

Researchers have also tried to make use of the variation in MMLs to tease out the causal

relationships between state policies and crime rates. Potentially, MMLs can lead to lower rates of

crime and violence – prospective customers would prefer legal outlets because of reduced risk of

arrest and the state regulatory systems could then design safeguards against potential criminal

involvement (Shepard & Blackley, 2016).

While earlier studies may have suggested a positive link between

decriminalization/MMLs and higher incidences of rape, robberies and assault, there has been

little evidence providing a systematic involvement of criminal behavior in states that have passed

MMLs. In fact, evidence from recent studies show that MMLs may be associated with lower

crime rates. For example, using an ecological cross-sectional design, Kepple and Friesthler

(2012) studied the density of medical marijuana dispensaries and property/violent crime rates in

Sacramento, California. Their study found no link between marijuana dispensaries and crime;

concluding that the safety measures used by dispensaries (eg., doormen, and video cameras) may

have had countervailing effects on criminal activity. Morris et al. (2014), generalized these

results by using state level panel data from 1990-2006 to estimate the association between MML

and property/crime rates. Findings from the research showed that the legislation preceded the

decline in homicide and assault, suggesting a drop in violent crime. Morris et al. (2014) and

Shepard & Blackley (2016) reached similar conclusions, concluding that the passage of medical

marijuana laws did not precipitate an increase in criminal behavior, and furthermore possibly

reduced it.

2.3.2 Marijuana and Interpersonal Violence

The majority of the studies that have examined the co-existence of substance abuse and

interpersonal violence have focused on alcohol, without including other commonly used

substances such as marijuana. Of recent concern within the study of associations between

substance use and violence is intimate partner violence, or IPV. So far research findings on the

association between marijuana use and IPV have been inconsistent. Using data from 96 studies,

Moore et al. (2008) conducted a meta-analytic review to quantitatively evaluate the relationship

between specific drug use and intimate partner aggression. Their results suggest that the

psychopharmacological effects of the drug produces increased aggression between intimate

partners. Some studies have found that marijuana use is highly correlated with psychological

abuse (Bennett et al. 2008), while others have linked more severe forms of IPV and IPV

recurrence to marijuana use (Wofordt et al., 1994; Chermack et al. 2001). A major limitation of

the previous studies is that most of them have been cross-sectional. It is thus important to test

whether marijuana use is predictive of subsequent IPV (Smith et al., 2014). Reingle, et.al. (2012)

used longitudinal data from the National Longitudinal Study of Adolescent Health to examine

the association between IPV and marijuana use. The authors found that consistent marijuana use,

independent of alcohol use, was a strong predictor of intimate partner aggression for both victims

and perpetrators.

Contrary to the previous literature some studies have suggested that marijuana use may

be inversely associated with IPV. For example, a nine-year longitudinal study examining a

community sample of newly married couples found that after controlling for important covariates

(e.g., anti-social behavior, alcohol use), frequent use of marijuana generally predicted less

frequent partner aggression over the first nine years of marriage (Smith et al., 2014).

Additionally, Stuart et al. (2013) found that women were less likely to perpetrate physical

aggression on days in which they had used marijuana relative to non-use days. There are also

some studies suggesting a weak causal link between marijuana and IPV. The analyses of

FalsStewart et al. (2003) indicated that the consumption of opiate-based drugs and cannabis were

not associated with an increase in the likelihood of male-to-female partner aggression at any

level of severity. Testa et al. (2003) reached similar conclusions; albeit being the most common

drug used by the survey respondents. The authors reported that within ongoing relationships hard

drug use (cocaine, heroin), but not marijuana use alone, predicted severe IPV and recurrence of

IPV. In addition to the lack of association between marijuana use and partner aggression, the

authors suggest that marijuana use may help suppress aggressive behavior.

The literature regarding the effects of marijuana use on child abuse and neglect have been

limited. Using survey data from respondents living in 50 mid-size cities in California, Freisthler

et al. (2015) used linear mixed effects multilevel modeling to assess the impact of marijuana use

on abusive/neglectful parenting. They found a significant and positive association between the

density of medical marijuana dispensaries and frequency of child physical abuse by parents.

Their findings suggest that parents who are current users of marijuana engage in physical abuse

more frequently, and may also have higher aggressive tendencies than their counterparts who do

not use marijuana. However, the authors found little evidence to suggest that past year marijuana

use related to supervisory neglect, and in fact found there was a negative relationship between

marijuana use and physical neglect.

In summary, the current literature provides mixed and inconclusive evidence about the

marijuana and domestic violence nexus, and is uncertain about the effects of MMLs on child

victimization. Since the majority of the survey studies have been correlational or cross-sectional,

the generalizability of the results may be limited due to over-reliance on self-reported measures,

lack of information regarding severity and nature of the offenses, and response biases (e.g. social

desirability bias) (Moore & Stuart, 2005). Freisthler et al. (2015) note that their reliance on data

gathered through a list-assisted telephone sample of only landlines, likely underestimates the

abuse and neglect rates among populations with no landlines. Thus the causal mechanisms of the

relationship between marijuana use and child maltreatment remain unknown. Furthermore, due

to social desirability bias, some parents may not disclose if they are abusive or neglectful

parents, and may even report their abusive practices at lower rates than would be accurate. These

inconsistent conclusions continue to fuel the debate about the marijuana-violence relationship.

Indeed, a better understanding of how changes in marijuana legislation may affect child

maltreatment at the population level is needed.

2.3.3 Extension of the Literature Review

The potential externality e ect of marijuana legalization on crime and illegal drug

consumption has been of primary concern in the current drug policy debate. Studies have shown

that legalizing marijuana is associated with an increase in marijuana consumption among all

ages. Consequently, the welfare implications will depend largely on whether marijuana use itself

generates negative or positive externalities to children, and on the extent to which marijuana

serves as a gateway to harder illicit drugs and to substance abuse.

A major societal concern about marijuana intoxication is the psychological and physical

effects which may directly affect the well-being of children and other non-users of cannabis. For

example, marijuana is known to impair motor skills, trigger psychiatric illnesses including mood

disorder and latent schizophrenia, and cause short term memory loss and temporal distortions

(Platt et al., 2010; Roser et al., 2012). These effects can increase the potential risk of parental

neglect and abuse. On the other hand, preliminary clinical research supports the potential

medicinal value of marijuana (see Walsh, Nelson & Mahmoud, 2003). Positive impacts on

parenting are likely to result if parents used the drug under medical supervision to relieve chronic

pain, anxiety, seizures, and other illnesses. Thus, depending on the degree to which these positive

and negative effects are experienced across populations on average, marijuana legalization could

either increase or decrease the risk of maladaptive parenting.

There is also a possibility of an indirect link to violence. A large body of research has

established a positive causal connection between alcohol abuse, illicit drug use (e.g. cocaine and

heroin) and domestic violence. A 1998 study by the National Center on Addiction and Substance

Abuse found that children whose parents abused illicit drugs and alcohol were three to four times

more likely to be severely maltreated than children of parents who were not substance abusers

(Reid Macchetto & Foster, 1998).

There has also been substantial research on whether marijuana use is likely to precede the

use of harder illicit drugs and other addictive substances such as alcohol. The findings thus far

have varied. For instance, a longitudinal study found that among adults with no history of

alcohol abuse, those who reported marijuana use during the first wave of the survey were more

likely to develop an alcohol use disorder within three years relative to those who reported no

marijuana use (Pacek, Martins & Crum, 2012). Wen et al. (2015) also found a positive

association between MMLs and frequency of binge drinking for adults over 20 years of age.

Additionally, a meta-analytic review by Merill et al. (1994) found that cocaine use was 17 times

more likely in adults that used marijuana as children. Other studies have also found a significant

correlation between marijuana use and illicit drug dependence (see Fergusson, Boden

& Horwood, 2006).While these findings appear to support the gateway hypothesis, authors

Morral, McCaffrey & Paddock (2002) suggest that factors such as familial relations, social

environment, and genetic predisposition to illicit drugs may be more reliable predictors of future

drug consumption. Further, using data from the 1993-2010 Behavioral Risk Factor Surveillance

System, Anderson et al. (2013) found that MMLs may significantly reduce the probability of past

month alcohol use and frequency of binge drinking. Clinical studies have also suggested that

smoking marijuana may prevent the development of tolerance to opiates (see Cichewicz and

Welch, 2003) and that MMLs are associated with a significant reduction in prescription

opioidrelated mortality (Bachhuber et al., 2014). The current debate provides little evidence to

support or refute the suggestion that marijuana use and MMLs are causally linked to the

subsequent abuse of alcohol and licit-illicit drugs.

Some studies suggest that marijuana may not only be a gateway to harder illicit drugs but

also to crime and criminal behavior. According to the research conducted by Evans (2013), the

probability of being arrested for a non-drug violent and income-producing crime is greater for

marijuana users than for non-users. If so, parental marijuana use may increase child

maltreatment. Indeed, some studies find that severe family dysfunction, such as parental

criminality, elevates the risk of child maltreatment (Juby & Farrington, 2001). Additionally,

Farrington (2010) suggests that poor parental supervision and parental criminality are the

strongest predictors of juvenile delinquency and anti-social behavior. If, however, criminality is

explained by marijuana’s illegality, rather than from the drug itself, legalization can break

marijuana’s link to violence. Thus, depending on which pathways are the strongest, marijuana

use and its medical availability can negatively or positively influence child welfare outcomes.

This highlights the need for rigorous empirical research in this area.

2.4 Data and Descriptive Evidence

According to the literature, MMLs should increase both the supply and demand for

marijuana, and thus increase the consumption of marijuana unambiguously (Anderson, Hansen

& Rees, 2013). Since marijuana is a psychoactive substance, it may influence a perpetrators’

perception of the expected costs or payoffs when supplying violence. If marijuana use does

increase a caregivers’ negligent or abusive behavior, I would expect legalization to lower the cost

of engaging in violence. Reducing the cost of violence is expected to raise the amount of

violence supplied; thus one possible outcome of legalization is a positive relationship between

marijuana use and child victimization. It is also quite possible that a caregiver, under the

influence of marijuana, may engage in certain types of maladaptive behavior, thus I separately

examine the relationship between MMLs and the most common types of maltreatment (neglect

and physical abuse).

To study the impact of medical marijuana laws and its different dimensions on child

victimization, I employ the use of three major datasets: Child Maltreatment Reports, National

Data Archive on Child Abuse and Neglect (NDACAN) and the FBI’s Uniform Crime Reports

(UCR). These data sources and variables have been described previously. Table 2.2 provides

definitions for my outcome measures. Table 2.3 presents descriptive statistics.

Data taken from the maltreatment reports and NDACAN capture the severity of child

victimization. The data assesses overall maltreatment rates, children who were victims of neglect

and physical abuse, victims by age, and deaths attributed to child maltreatment (fatality rates).

Additionally, publicly available data from the UCR provide information on crime and arrest rates

(e.g., offenses against the family, drug offenses, alcohol offenses) and help reinforce the findings

from the previous literatures’ marijuana-crime link.

In 2014 the CPS received 3.6 million referrals alleging child abuse and neglect, of which

more than 50 percent of the cases were investigated. 702,000 children were victims of abuse and

neglect (9.4 victims per 1,000 children) and an estimated 1,580 children died due to

maltreatment (2.3 per 100,000 children). According to the CPS there are two major risk factors

that may increase the likelihood of victimization – caregiver alcohol abuse and caregiver drug

abuse. While some states may have legalized marijuana, no formal guidelines exist on how

welfare workers should handle a caregivers’ recreational or even medicinal use of the drug.

In 2014, 27% of all child maltreatment cases were related to parental drug use (Child

Maltreatment Report, 2014). Since marijuana is still classified as a schedule I substance, child

welfare agents might not distinguish a parents’ use of marijuana from other illegal substances

such as heroin or cocaine even if it’s used for medical purposes. This could potentially increase

the number of at-risk victims being reported to the CPS – thereby increasing the reported

incidences of child maltreatment rates.

I also investigate an alternative proxy for maltreatment that is less likely to be affected by

the reporting bias: child fatality rates. Since institutions and authorities (law enforcement, state

vital statistics departments, medical examiners, hospitals, etc.) must report any deaths due to

maltreatment, and because such a report will be investigated by the CPS, this variable is unlikely

to face reporting bias. One limitation, however, is that it is an extreme outcome, and as such

could create noise in the proxy.

The primary independent variables of interest are states that have passed medical

marijuana laws (MMLs). To determine when a MML was passed within each state, I used the

research conducted by Pacula et al. (2015) and updated it with information from the official

legislative website of each US state, NCLS and NORML. Specifically, dichotomous indicators

are included for whether a state has the following: laws that allow for the medicinal use of

marijuana (MML); legal protection for patients to grow their own plants for medicinal purposes

(home cultivation laws); provisions for patients to use marijuana to mitigate chronic pain; and

decriminalization statutes in conjunction with MMLs.

All state laws allow patients to possess and use small quantities of marijuana without

being subjected to state criminal penalties. However there are variations within each state’s

MML –each have their own specific restrictions for possession limits, home cultivation and

qualifying conditions. For example, while only two states – California and Washington – allow

the use of marijuana to treat anorexia, the majority of states with MMLs include provisions for

conditions such as HIV-AIDS, cancer, cachexia, chronic pain, and other conditions approved by

the state health department. Possession and cultivation limits can also vary from one ounce and

six plants in Alaska to 2.5 ounces and 12 plants in Michigan (Hoffmann & Weber, 2010).

Currently, only 15 states allow for home cultivation of medical marijuana by patients.

Users in the states which have decriminalized possession may face a lower expected

penalty (Markowitz, 2005) and therefore a lower price of using marijuana. California’s

decriminalization statute (2010) provides an example of how the possession of small amounts

(less than one ounce) of marijuana constitutes a simple misdemeanor:

Except as authorized by law, every person who possesses not more than 28.5 grams of

marijuana, other than concentrated cannabis, is guilty of an infraction punishable by a fine of

not more than one hundred dollars.

States that have home cultivation and decriminalization laws greatly liberalize access for

patients and recreational users. It also provides a source of easily accessible marijuana for youth

recreational use and broadens the social approval of marijuana use (Pacula et al., 2015). If as

some research suggests, these laws lead to an increase in marijuana consumption and to an

increase in IPV, then one would expect that in states that have the most lenient form of the law

(i.e. home cultivation in conjunction with decriminalization) there would be an increased risk in

the frequency and severity of child maltreatment. Table (2.1) gives a summary of the 23 states

that have legalized marijuana for medical use and 17 states that have decriminalized its use.

Sociodemographic and economic characteristics may also play a role in determining a

perpetrators’ propensity toward supplying violence. These variables and their sources have been

described in the previous chapter. To aid in controlling for a variety of time-varying and

potentially confounding factors I include each state’s unemployment rate, male to female

employment ratio, percent population living below the poverty level, college and high school

attainment rate, median household income, population density, divorce rates, indicators of race,

percent of the state population incarcerated, and violent crime rates. These variables serve as

proxies for opportunities available to perpetrators. For example, higher unemployment and

poverty rates may correspond to higher stress and depression due to fewer opportunities being

available; parents then may be less likely to invest time and money in their children. In such

cases basic needs can be neglected. Conversely, states with a higher proportion of educated

people and higher income (real wages) have more well-paying employment opportunities, and

have a higher cost of engaging in violence. As a result, a negative relationship is expected for

education and household income.

Additional controls include the percent of the population between the ages of 15-24 and

25-45 (U.S. Census Statistical Abstracts); the number of police officers per 100,000 persons

(Census Bureau, UCR); female share of officers per 100,000 persons (Census Bureau, UCR);

arrest rates for types of offenses (UCR); and beer and alcohol consumption per capita (Beer

Institute). These variables account for other state-level changes that could separately explain

maltreatment rates. For example, arrest rates (enforcement) and the number of police officers

measure the effectiveness of a state’s efforts against crime in general. They are included in all

models to measure the risk of punishment for committing a crime. I expect that higher police

presence and higher arrest rates for family offenses would increase the cost of violence, thereby

reducing maltreatment. Miller & Segal (2014) found lower domestic violence escalation rates as

a result of an increase in reporting by female police officers. I expect a similar relationship with

respect to female share of officers and reported child maltreatment rates. Additionally, a number

of studies have found that alcohol is a significant contributory factor to child maltreatment,

linking alcohol consumption to reduced self-control, mental health issues, antisocial personality

characteristics, and thus a higher risk of physical abuse and neglect. Finally, to allow for

variation in MMLs, to address time shocks and control for heterogeneity I include state fixed

effects, year fixed effects and state specific linear time trends.

2.5 Empirical Strategy

2.5.1 Methodology

In light of the uncertainty of the effects of MMLs, this paper examines whether states that

have implemented these laws see a change in child victimization rates. Specifically, to estimate

these outcomes, I use the fixed effects model with Driscoll-Kraay standard errors to exploit the

within-state variation introduced by the passage of MMLs in 23 states plus D.C. over the 19 year

observation period.

One limitation of using the standard fixed effects model is that it does not account for

cross-sectional dependence. This cross correlation of errors could be due to omitted common

effects that may not be quantitatively measured, such as social norms or psychological behavior

patterns (Chudik & Pesaran, 2013). In order to test whether the residuals in my fixed effects

regression are spatially independent, I perform the Pesaran CD test, as recommended by Hoechle

(2007). The null hypothesis of the CD test states that spatial dependence is indeed present. Since

spatial dependence could lead to inconsistent coefficient estimates, I estimate the fixed effects

model that is robust to heteroscedasticity, autocorrelation, and very general forms of

crosssectional and temporal dependence.

To carry out the fixed effects analysis I estimate equation (1) where each of the child

offenses variable (i.e. child maltreatment rates, victims by age, fatality rates, arrest rates for

family offenses), is the dependent variable. Formally, my empirical specification may be

expressed as:

𝑦𝑆𝑇 = 𝛽𝑜 + 𝛽1 𝑀𝑀𝐿𝑆𝑇 + 𝛽2 𝑋𝑆𝑇 + 𝛾𝑇 + 𝜃𝑆 + 𝛿𝑆𝑇 + 𝜖𝑆𝑇 (1)

where for each state S, in year T, 𝑦𝑆𝑇 is the child offense rate outcome variable; the main

explanatory variable 𝑀𝑀𝐿𝑆𝑇 is a dichotomous indicator equal to 1 if a state implemented a

medical marijuana provision from year T forward, and 0 otherwise; 𝛾𝑇 and 𝜃𝑆 are year and state

fixed effects; 𝛿𝑆𝑇 is the state-specific time trend; and 𝑋𝑆𝑇 is a vector of control variables that

include sociodemographic, economic, crime and public policy indicators. The coefficient of

interest is 𝛽1 which measures the effect of the MML on child victimization rate.

Using the same specification as (1) I estimate four separate regressions where the

regressors of interest are states that have implemented home cultivation laws; provisions for

pain; and decriminalization laws in conjunction with MMLs and home cultivation laws (HCL).

HCL is an indicator set to 1 if the state provides legal protection for patients or caregivers to

grow marijuana for medicinal purposes, 0 otherwise. Decrim&HCL is a dichotomous indicator

equal to 1 if a state has implemented both home cultivation laws and decriminalization laws at

year T, 0 otherwise. Finally, MMLXDecrim is an indicator equal to 1 if a state has implemented

both MMLs and decriminalization laws at time T. As implied earlier, a state’s implementation of

MMLs is likely to either increase or decrease the likelihood of child victimization rates, thus the

impact of MMLs on child abuse is tested in each model.

2.6 Estimation Results

2.6.1 Main Results

Deterrence theory asserts that reducing the perceived severity of legal sanctions

associated with marijuana use will increase the demand for marijuana. However, the changes in

legislation (i.e. an increase in demand) could result in significant and negative spillover effects to

parents and their children, increasing the risk of child abuse and neglect.

Table 2.4 presents the impact of MMLs on reported maltreatment rates while controlling

for other time-varying explanatory variables. Each column reports the estimated effect of

statelevel marijuana legalization from a unique regression. In the first column, I present a

parsimonious specification that only includes state and year fixed effects. I find that the

legalization of medical marijuana is associated with a 13.2% increase in reported child

maltreatment rates. Adding state level controls in column (2) reduces the magnitude of the

estimated relationship and the significance falls from the 1% to 5% level; more importantly,

legalization is associated with a 11.4% increase in the reported incidences of child maltreatment.

I now extend the specification to include state-specific linear time trends to control for the

influence of unobserved factors at the state level that trend smoothly over time (e.g., citizen and

government sentiment toward marijuana use). Again, I find a statistically significant and positive

effect of MMLs on child maltreatment rates. Specifically, the estimates suggest that after the

passage of MMLs, states see an increase in reported incidences of child maltreatment by

1.30 per 1,000 children; this translates to a 9.8% increase in reported maltreatment rates.

Table 2.5 presents the estimates between MMLs and child maltreatment rates by age

group. I expect victimization rates to be higher for younger children since they are more

vulnerable to abuse and neglect than older children. Additionally, since marijuana is classified as

a schedule I substance at the federal level, it is more likely for parents to get reported and be

investigated for abuse and neglect if they use marijuana in the presence of younger children. I

find the estimates of reported incidences of abuse for younger children (ages of 0 and 3 years) to

be much larger in magnitude and are statistically significant compared to incidences of abuse for

older children. More specifically, for younger children between the ages of 0 to 3 (Table 2.5,

column 1), enforcement of MML is associated with a 16.5% increase in reported maltreatment,

and for children between ages 12 to 15 (column 4), a (statistically insignificant) 13.73% increase

in reported incidences of maltreatment.

Why does the implementation of MMLs increase child maltreatment?

If MMLs are associated with an overall increase in the incidence of reported

maltreatment rates, what could explain such an effect? There are two likely mechanisms through

which MML – legislation that aids patients with chronic health conditions – might affect child

maltreatment estimates. First, MMLs caused an increase in actual maltreatment. Second, it may

have increased the reporting rate of maltreatment. I employ two measures of child maltreatment

to attempt to distinguish between the mechanisms and correct for any potential reporting bias:

child fatality rates and arrest rates for family offenses.

First, I test whether reporting and arrest patterns for family offenses changed around the

implementation of the policy. Arrest data are frequently used in the crime literature as a measure

of crime and to account for changes in police reporting behavior. In column (2) of Table 2.6, I

present estimates of the impact of MMLs on arrest rates for offenses against the family, including

all the controls mentioned above. I find that states that adopt MMLs witness a (statistically

insignificant) 6.1% increase in arrest rates for family violence relative to states with no such

policies. However, as Dalbo & Aizer (2014) suggest, the estimated effect of arrest may not just

reflect changes in reporting but also changes in arrests conditional on reporting. For example,

police officers are more likely to arrest parents that use marijuana if the Department of Children

and Family Services and courts consistently rule that parental usage of medical marijuana places

the child at a substantial risk of harm. Indeed, until 2010, public opinion about legalizing

marijuana rarely shifted, with a majority believing the drug should be made illegal and usage of

the drug should be policed (PEW Research, 2014).

Next, I present regression estimates of the impact of MMLs on actual maltreatment.

Albeit a noisy proxy for maltreatment due to its low-frequency, child fatality rates can serve as

an appropriate proxy to measure an increase or decrease in actual maltreatment following the

implementation of state-level MMLs. The underlying premise of this approach is that child

fatalities will always be reported to the police and CPS, and as such it will be immune from any

reporting effect (Levitt, 1998).

The main results are shown in Table 2.7, column (3). Baseline estimates in column (1)

show that there is a negative and statistically significant relationship between MMLs and

changes in fatality rates. However, these estimates become smaller and insignificant after

controlling for socio-demographic factors and state-specific linear time trends (column 3). I find

that MMLs have a negative (-0.174) impact on child fatality rates; more specifically, the results

suggest that after the passage of the laws, states see a 9.45% reduction in child deaths due to

maltreatment. The lack of significance could be explained, in some part, due to noise in the child

fatality measure. While this finding does not provide evidence of a strong correlation between

MMLs and fatality, it does not necessarily negate the possibility that an economically significant

relationship exist. More importantly, the evidence suggests that there may indeed be a reporting

effect going on, and not an increase in actual maltreatment.

Tables 2.8 and 2.9 (column 2) provide additional evidence that MMLs may be associated

with an increase in the reporting of child maltreatment. Interestingly, the results show that there

is no significant positive relationship between the adoption of MMLs and rates for physical

abuse and neglect. Moreover, the estimates indicate there may be evidence of a drop in physical

abuse in states with medical marijuana policies. Specifically, I find that states with MMLs are

estimated to have 0.548 fewer children who are physically abused per thousand children relative

to states without MMLs, a reduction of 21.4 % when assessed against the sample mean. On the

other hand, I find a positive but statistically insignificant relationship between MMLs and

neglect, showing a 10.7% increase in the reported incidences of neglect. The pattern of results so

far is consistent with the reporting hypothesis: parents who use medical marijuana are more

likely to be subject to a child neglect inquiry since social workers may determine that marijuana

use would substantially impair a parent’s judgement and ability to care for their children’s basic

needs.

Figure 2.2 presents graphical evidence of the effect of MMLs on reported maltreatment

rates over time. The graph shows the means of yearly maltreatment rates before and after the

implementation date of MMLs, with 1 on the X-axis denoting the first full year of the law being

in effect. Prior to the implementation of MMLs, the maltreatment rates seem to be relatively

stable; however, after the first full year of the law being in effect, there is a sharp increase in the

reported incidences of maltreatment. After the fourth year, the treatment effect appears to be

decreasing over time, suggesting an initial reporting effect.

In summary, the estimates from the NCANDs and UCR data indicate a 10–13% increase

in reported child maltreatment rates after medical marijuana legalization. However, this positive

effect largely comes from the increase in the reporting and investigation of cases of child neglect.

More importantly, evidence from the child fatality estimates show that the actual incidence of

child maltreatment may be falling in states with MMLs.

2.6.2 Robustness Checks

Table 2.10 column (1) shows the estimates for pre-and post-legalization trends in child

maltreatment rates. I add controls for four years of MML policy leads and three years of policy

lags. In the years preceding the law, I find that reported maltreatment rates are negative and

stable, but statistically insignificant; suggesting no policy endogeneity, thus lending credibility to

the main estimates in Table 2.4. However, after the first full year of the law being in effect,

MMLs are associated with a significant increase in reported maltreatment rates. The estimates

for the reported maltreatment rates become even larger in magnitude, but are statistically

insignificant, during the third year of post-legislation. However, after five or more years, while

the estimates become negative, they continue to be statistically indistinguishable from zero.

It is somewhat surprising that the effect of MMLs does not grow over time; nevertheless

this pattern of results is consistent with Figure 2.2, showing the reported incidences of

maltreatment ramping up immediately after the legislation, and slowing down in the years after.

One potential reason for this could be due to the nature of the data – since NCANDs aggregates

the reports into a single yearly estimate, monthly growth over time may be missed. Additionally,

this phenomenon is consistent with the reporting hypothesis; the behavioral response seems to

follow immediately after the passage of the law. If, as anecdotal evidence suggests that opinions

change, whereby there is a greater social acceptance of marijuana by law enforcement and social

workers, especially for parents who use the drug for medicinal purposes, then I would not expect

to see growth over time.

Next, Table 2.4, Column (4) estimates the sensitivity of my results to an alternate

specification. Since maltreatment is intrinsically a count of victims within a discrete time period,

I use the negative binomial model as a specification check for my primary analyses. Table 2.4

presents coefficients on the maltreatment rate variable from the OLS fixed effects and negative

binomial specifications for completeness. The estimates confirm the results from my main

estimation – states with MMLs see a significant increase in the reported incidences of child

maltreatment, and a significant decline in fatality rates (Table 2.4, column 4). In addition, when I

use the coding preferred by Pacula et al. (2015) to obtain the effective dates of the laws, I find a

similar pattern of statistically significant results (Tables 2.15 & 2.16).

I now examine the impact of specific policy dimensions to capture the reporting effect

and the true maltreatment effect: that is, provisions that allow for home cultivation and

prescriptions for chronic pain. Since both provisions instrument for regulatory laxness, they are

more likely to increase social availability and access to the drug. As such, these provisions are

predicted to affect reporting behavior and consequently reported maltreatment rates through the

changing perceived risk associated with the enforcement of parental marijuana use. I thus, expect

parents and caregivers who grow marijuana, even if licensed, to be reported and investigated for

(risk of) child endangerment.

Column (2) from Tables 2.11 and 2.12 show the estimates of home cultivation laws

(HCLs) on child maltreatment and fatality rates. Overall, I find a positive and statistically

significant legislative effect on reported child maltreatment rates. Specifically, the results suggest

that states with HCLs are responsible for an additional 3.26 children being reported as maltreated

per 1000 children, translating to a 24.6% increase in reported maltreatment rates. Note that these

estimates are twice as large as the ones from MMLs. More importantly, when I estimate the

effect of HCLs on child fatality rates in Table 2.12, I find the magnitude of the coefficients to be

large and statistically significant, suggesting a 27% reduction in actual maltreatment rates. I find

similar and statistically significant results (Column 3, Tables 2.11 & 2.12) when I test the impact

of the provisions that allow the use of marijuana for chronic pain. Specifically, the

implementation of provisions that allow for chronic pain is associated with a 14% increase in

reported maltreatment rates, and a substantial 22.4% reduction in fatality rates.

I continue to explore the differential effects of state-specific medical marijuana

regulations by interacting MMLs with states that have decriminalized the possession of

marijuana. Tables 2.13 and 2.14 provide further evidence that the magnitudes of the interaction

terms are much larger in states that impose relatively lax restrictions than those with no such

policies. These findings are consistent with my previous estimates from Tables 2.11 and 2.12,

and build on the work by Pacula et al. (2015) who recognized the heterogeneity in the

implementation of state level marijuana regulations. Thus, the binary MML measure in Table 2.4

misses the heterogeneous effects and dynamics of these policies. Finally, these findings are

consistent with the interpretation that MMLs not only influence the reported incidences of

maltreatment, but they may also reduce the actual incidences of child maltreatment.

2.7 Conclusions

Recent research by Friestler and colleagues (2015) suggests that parental marijuana use is

related to higher incidences of physical abuse and neglect. However, to my knowledge, no

research has examined the relationship between state marijuana legislation and child

victimization rates. The central findings gleaned from this paper provide indirect evidence that

marijuana use, induced by increased access to medical marijuana, affects the reported incidences

of child maltreatment positively. Specifically, estimates from the fixed effects regression suggest

that after the passage of MMLs, states see a statistically significant (9.8%) increase in reported

maltreatment rates. Using Driskoll-Kraay standard errors, these results are robust to

heteroscedasticity, autocorrelation, and very general forms of cross-sectional dependence.

The findings from my main model raise an important follow-up question: does medical

marijuana legalization increase child maltreatment or child maltreatment reporting? I examine

one particular outcome of interest to proxy for the true incidence of maltreatment: child fatality

rates. I find a negative but statistically insignificant relationship between MMLs and child

fatality rates. However, as discussed by Pacula et al. (2015), MMLs vary greatly and can thus

generate heterogeneous effects. Indeed, I find the largest estimates when I look at specific

dimensions of MMLs, where the coefficients capture not only the reporting effect but also the

true effect on maltreatment. For example, states with provisions that allow for home cultivation

see a 24.6% increase in the reported incidences of maltreatment and, surprisingly, a statistically

significant 27.6% reduction in the actual maltreatment rate. Further, these findings run contrary

to the arguments suggesting a positive relationship between the legalization of medical marijuana

and violence.

Data limitations do not allow me to explore all of the other channels through which

MMLs may affect child outcomes – particularly pharmacological effects of the drug. However,

identifying one specific mechanism through which MMLs may affect maltreatment, such as child

fatality, does provide one piece of the puzzle. It is important to note that the negative relationship

between MMLs and child fatality rates does not necessarily imply a strict causal connection that

marijuana use reduces actual maltreatment. For instance, it is possible that marijuana regulation

reduces child fatality rates through its positive reporting effect. Even if growing or consuming

marijuana is legal, anecdotal evidence suggests that parental use of marijuana can be

controversial. However, with the passage of time, I expect attitudes and behaviors toward

parental medical marijuana use to be more tolerant and accepted. As such, it is unlikely that

reported maltreatment rates will continue to increase. Indeed, trend analyses provide further

evidence that child maltreatment may be decreasing over time. Clearly, distinguishing between

child maltreatment and reporting is a subject that warrants further attention. In sum, as the

narrative of medical and recreational marijuana legislation unfolds across the country, more

substantive research is needed to determine how marijuana use impacts child outcomes and

parenting.

Figure 2.1 Past-Month Use of Selected Illicit Drugs

Source: National Institute on Drug Abuse (2013)

Figure 2.2 Trends in Child Maltreatment Rates, by State MMLs

Notes and Sources: Data is from the Child Maltreatment Reports and the National Data Archive for Child Abuse and

Neglect (NDACAN), which provides prevalence of child maltreatment from 1995-2014. The Dashed line marks the

timing of the medical marijuana law. As of 2014, 23 states plus D.C. have implemented MMLs; the law provides

protection from criminal penalties for using marijuana for a medical purposes.

Table 2.1 MML Legislation Policies by State, 1996-2014

Provisions

State

Year Passed

Effective date

Pain

Home Cultivation

Marijuana Decriminalized

Possession Limit

Alaska

1998

1999

Yes

1975

1 oz/6 plants (3 mature, 3 immature)

Arizona

2010

2011

Yes

2011

2.5 oz/12 plants

California

1996

Yes

1976

8 oz/ 6 mature or 12 immature

Colorado

2000

2001

Yes

2011

2 oz/ 6 plants (3 mature, 3 immature)

Connecticut

2012

Not specified

District of Columbia

2010

2014

2 oz/Not specified

Delaware

2011

Yes

6 oz

Hawaii

2000

Yes

3 oz/ 7 plants (3 mature, 4 immature)

Illinois

2013

2014

2.5 oz

Maine

1999

Yes

1976

2.5 oz/6 plants

Maryland

2014

Yes

2011

Not specified

Massachusetts

2012

2013

Yes

2009

Not specified

Michigan

2008

Yes

2.5 oz/ 12 plants

Minnesotta

2014

1976

Not specified

Montana

2004

Yes

1 oz/ 4 plants (mature)

Nebraska

1977

Nevada

2001

Yes

2002

1 oz/ 7 plants (3 mature, 4 immature)

New Hampshire

2013

2 oz

New Jersey

2009

2010

Yes

2 oz/ Not specified

New Mexico

2007

Yes

6 oz/ 16 plants (4 mature, 12 immature)

New York

2014

1977

Not specified

North Carolina

1977

Ohio

1976

Oregon

1998

Yes

1973

24 oz/24 plants (6 mature, 18 immature)

Rhode Island

2006

Yes

2013

2.5 oz/ 12 plants

Vermont

2004

Yes

2 oz/ 9 plants (2 mature, 7 immature)

Washington

1998

Yes

2012

24 oz/15 plants

Notes and Sources: Own data collection. Referred from the following sources: Procon.org; NORML; Pacula et al. 2013

Table 2.2 Summary of Data Sources

Variables

Definitions

Sources and years

Dependent Variables

Child maltreatment

Children who have experienced or

who were at risk of experiencing

abuse or neglect.

NDACAN (1995-1999)

Children’s Bureau (2000-

2014)

Child fatality rate

Offenses against the family arrest rate

Children who have died due to abuse

or neglect

Family violence includes all types of

violent crime committed by an

offender who is related to the victim

either biologically or legally through

marriage or adoption.

Bureau of Justice Statistics

(1995-2014)

Explanatory variables of Interest

Medical Marijuana Laws

Decriminalization Laws

States that allow for the medical use

of marijuana.

Reduces penalties associated with the

use or possession of small amounts of

marijuana

NORML; State statutes;

Pacula et al. (2013);

ProCon.org

Family and State Environment

Female labor force participation rate

Unemployment rate

Median Household income

Poverty rate

Population density per square mile (Proxy

for urban rate)

(Total population/ Land area)

U.S. Census Bureau -

Statistical Abstracts Series,

Bureau of Labor Statistics

(1995-2014)

Divorce rate

Beer & Alcohol consumption per capita

Wolfers, Justin. 2006.

(1995 – 2000)

CDC divorce rates (2000-

2014)

Beer Institute (1195-2014)

Fraction of child population that is white

Fraction of child population that is black

Percent of the population between ages

15-24

Percent of the population between ages 25-

U.S Census Bureau

Current Population

Surveys (1995-2014)

College attainment rate

High school attainment rate

Human Capital Index Measures

Frank, Mark. W. (2009)

(1995-2014)

State Judicial Environment

Law enforcement Female

Officers

Law enforcement to population ratio

Female officers to population ratio

Bureau of Justice Statistics

U.S. Census Bureau -

Incarceration rates

Drug abuse arrest rates

Prisoner to population ratio

Statistical Abstracts Series

(1995-2014)

Crime rate

Crime to population ratio

FBI Uniform Crime Report

(1995-2014)

Table 2.3 Summary Statistics

Std.

Variable Obs Mean Dev. Min Max

Dependent Variables:

Child Maltreatment Rate per 1,000 children 964 13.2315 8.42456 0.70364 82.627

Child Fatality Rate per 100,000 children 957 1.84094 1.26986 0 16.775

Maltreatment by age: 0-3 957 15.0515 8.7571 0.67 113.901

Maltreatment by age: 4-7 957 12.4261 6.65694 1 103.869

Maltreatment by age: 8-11 957 10.1774 6.65694 1.05 88.9389

Maltreatment by age: 12-15 957 9.98481 6.69793 1.16913 88.9389

Physical Maltreatment Rate per 1,000 children 966 2.56123 2.33376 0 21.7933

Neglect Rate per 1,000 children 967 6.94018 5.44813 0 48.3264

Arrest Rates for Offenses Against the Family,

per 100,000 persons 956 39.094 35.4792 0.02083 230.472

Independent Variables:

Percent of the population: 15-24 1020 14.239 1.13257 10.8858 20.2158

Percent of the population: 25-44 1020 28.0525 2.54601 22.882 36.8265

Beer consumption per Capita 1020 1.24542 0.20897 0.67 1.91

Alcohol Consumption per Capita 1020 2.3604 0.52117 1.2 4.7

Arrest Rates for Drug Abuse per 100,000

persons 986 390.072 177.734 4.5833 1105.24

Law Enforcement to Population Ratio 1018 0.00312 0.00096 0.00024 0.00933

Female Officers to Population Ratio 1018 0.00024 0.00022 1.6E-05 0.00177

Violent Crime to Population Ratio 1020 0.00438 0.00255 0.00067 0.02661

Prisoners to Population Ratio 1005 0.00413 0.00183 0.00085 0.01768

Poverty Rate 1020 13.1104 3.72273 4.5 26.4

Percent Black 985 13.8662 13.9825 0.44814 91.2646

Percent White 985 66.1245 20.1882 12.38 102.324

Population Density per sq. mile 1020 369.5 1319.51 1.06 10801.5

Labor force Participation Rate for Females 969 60.4909 4.44932 46.3 71.2

High School attainment Rate 1020 0.57703 0.05557 0.44878 0.74347

College Attainment Rate 1020 0.17783 0.045 0.08398 0.45932

Median Household Income 1020 53895.1 8260.62 35521 77506

Divorce Rate 919 3.98627 1.07179 1.5 10.4406 Unemployment Rate 1020 5.62978 1.96417 2.3 13.8

Table 2.4 Effects of MMLs on Child Maltreatment Rates

(1) (2) (3) (4)

Child Maltreatment Rate per 1,000 Children

Dependent variable mean

13.23146

MML = 1

1.758***

1.505**

1.303*

0.104*

(0.597)

(0.700)

(0.649)

(0.0583)

Estimation method

OLS

Neg. Bin

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

964

759

Within R-squared

.20

.25

.52

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variable for each column is the child maltreatment rate per1000 children. MML=1 if

a state implemented a medical marijuana provision. This table provides the coefficient estimates from the

regression model in (1) estimated by FE regression. Robust standard errors (in parentheses) are based on

Driscoll-Kraay spatial-autocorrelation and cluster-robust standard errors.

All controls include:

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000

inhabitants. Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

Other crime controls include, family and drug abuse arrest rates per 100,000 persons; law enforcement to

population ratio; female officers to population ratio; prisoner to population ratio

State economic control variables include the variables: unemployment rate, female labor force

participation rate, and state median household income (BLS and US Statistical Abstracts), college and high

school attainment rate, (Frank, 2009)

State socio- demographic controls are based on the March Current Population Survey and the U.S

Statistical Abstracts. They include variables for the percent of the child population that is black and white;

divorce rates, percent of the population that’s between the ages of 15-24 and 25-44; alcohol consumption

per capita; beer consumption per capita; population density per sq. mile.

Table 2.5 Effects of MMLs on Child Maltreatment Rates, by Age Cohort

(1)

(2)

(3)

(4)

Child maltreatment rate per 1,000

children, by age groups

0-3

4-7

8-11

12-15

Dependent variable mean

15.0515

12.426

10.177

9.985

MML = 1

2.448*

1.441

1.222

1.371

(1.395)

(1.200)

(0.965)

(0.973)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

742

Within R-squared

.422

.5062

.5146

.5408

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variable for each column is the child maltreatment rate per1000 children, by age cohort.

MML=1 if a state implemented a medical marijuana provision. This table provides the coefficient estimates from the

regression model in (1) estimated by FE regression. Robust standard errors (in parentheses) are based on

DriscollKraay spatial-autocorrelation and cluster-robust standard errors.

All controls include:

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000 inhabitants.

Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape. Other crime controls

include, family offenses and drug abuse arrest rates per 100,000 persons; law enforcement to population ratio; female

officers to population ratio; prisoner to population ratio

State economic control variables include the variables: unemployment rate, female labor force participation rate,

and state median household income (BLS and US Statistical Abstracts), college and high school attainment rate,

(Frank, 2009), population density per square mile (U.S Statistical Abstracts).

State socio- demographic controls are based on the March Current Population Survey and the U.S Statistical

Abstracts. They include variables for the percent of the child population that is black and white; divorce rates,

percent of the population that’s between the ages of 15-24 and 25-44; alcohol consumption per capita; beer

consumption per capita; population density per sq. mile.

Table 2.6 The Effect of MMLs on Arrest Rates for Family Offenses

(1) (2)

Arrest rates for family offenses per 100,000 persons

Dependent variable mean

39.09395

MML=1

0.493

2.397

(2.246)

(2.391)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

Within R-squared

784

.165

784

.6396

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variable for each column is the arrest rates for offenses against a family member,

per 100,000 persons. MML=1 if a state implemented a medical marijuana provision. This table provides

the coefficient estimates from the regression model in (1) estimated by FE regression. Robust standard

errors (in parentheses) are based on Driscoll-Kraay spatial-autocorrelation and cluster-robust standard

errors.

All regressions include state economic, socio-demographic policy and crime controls.

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000

inhabitants. Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

Other crime controls include, drug abuse arrest rates per 100,000 persons; law enforcement to population

ratio; female officers to population ratio; prisoner to population ratio

State economic control variables include the variables: unemployment rate, female labor force

participation rate, and state median household income (BLS and US Statistical Abstracts), college and

high school attainment rate, (Frank, 2009), population density per square mile (U.S Statistical Abstracts).

State socio- demographic controls are based on the March Current Population Survey and the U.S

Statistical Abstracts. They include variables for the percent of the child population that is black and white;

divorce rates, percent of the population that’s between the ages of 15-24 and 25-44; alcohol consumption

per capita; beer consumption per capita; population density per sq. mile.

Table 2.7 The Effect of MMLs on Child Fatality Rates

(1) (2) (3)

Child fatality rate per 100,000 children

(4)

Dependent variable mean

1.841

MML = 1

-0.298*

-0.093

-0.174

-.1396

(0.159)

(0.233)

(0.258)

(0.105)

Estimation method

OLS

Neg. Bin

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

957

794

within R-squared

.025

.147

.363

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variable for each column is the child fatality rate, per 100,000 children. MML=1 if a

state implemented a medical marijuana provision. This table provides the coefficient estimates from the

regression model in (1) estimated by FE regression. Robust standard errors (in parentheses) are based on

Driscoll-Kraay spatial-autocorrelation and cluster-robust standard errors.

All regressions include state economic, socio-demographic policy and crime controls.

Crime rate controls use FBI Uniform Crime reports for the number of violent crimes per 100,000

inhabitants. Indexed crimes included in the violent crime variable are murder, robbery, assault, and rape.

Other crime controls include, law enforcement to population ratio; female officers to population ratio;

prisoner to population ratio

State economic control variables include the variables: unemployment rate, female labor force participation

rate, and state median household income (BLS and US Statistical Abstracts), college and high school

attainment rate, (Frank, 2009), population density per square mile (U.S Statistical Abstracts).

State socio- demographic controls are based on the March Current Population Survey and the U.S

Statistical Abstracts. They include variables for the percent of the child population that is black and white;

divorce rates, percent of the population that’s between the ages of 15-24 and 25-44; alcohol consumption per

capita; beer consumption per capita; population density per sq. mile.

Table 2.8 The Effect of MMLs on Maltreatment Types: Physical Abuse

(1)

Physical abuse rate per 1,000 children

(2)

Dependent variable mean

2.561

MML=1

-0.229

-0.548

(0.309)

(0.532)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

761

within R-squared

.474

.581

Number of groups

Table 2.9 The Effect of MMLs on Maltreatment Types: Neglect

(1)

Child neglect rate per 1,000 children

(2)

Dependent variable mean

6.94

MML =1

0.340

0.746

(0.550)

(0.647)

All Controls

State fixed effects

Year fixed effects

State-specific time trend

Observations

762

within R-squared

.232

.482

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variables are child physical abuse and neglect rate, per 1,000 children. MML=1 if a state

implemented a medical marijuana provision. This table provides the coefficient estimates from the regression model

in (1) estimated by FE regression. Robust standard errors (in parentheses) are based on Driscoll-Kraay

spatialautocorrelation and cluster-robust standard errors. Crime rate controls include the number of violent crimes

per 100,000 inhabitants; family and drug abuse arrest rates per 100,000 persons; law enforcement to population ratio;

female officers to population ratio; prisoner to population ratio. State economic control variables include

unemployment rate, female labor force participation rate, and state median household income, college and high

school attainment rate, population density per square mile. State socio- demographic controls include the percent of

the child population that is black and white; divorce rates, percent of the population between the ages of 15-24 and

25-44; alcohol consumption per capita; beer consumption per capita; population density per sq. mile.

Table 2.10 Robustness of Estimates: The Effect of MMLs on Child Maltreatment Rates to

Control for Policy Leads and Lags

(1)

(2)

(3)

(4)

(5)

VARIABLES

All ages

0-3

4-7

8-11

12-15

3 Years Prior

-0.150

-0.893

-0.662

-0.905

-0.818

(0.981)

(0.838)

(0.812)

(0.851)

(0.861)

2 Years Prior

-0.322

-0.404

-0.156

-0.213

-0.140

(0.775)

(1.221)

(0.953)

(0.790)

1 Year Prior

-0.363

0.884

1.013

0.266

0.359

(0.994)

(1.330)

(1.148)

(0.856)

(0.802)

Year Effective

0.461

1.086

0.809

0.358

0.518

(0.684)

(1.109)

(0.910)

(0.765)

(0.731)

1 Year After

-0.00656

1.272

0.595

0.249

0.444

(0.629)

(1.266)

(0.996)

(0.727)

(0.710)

2 Years After

1.170

2.635*

1.827

1.068

1.304

(1.042)

(1.503)

(1.422)

(1.189)

(1.142)

3 Years After

3.102

6.360

5.070

3.924

4.290

(2.798)

(4.930)

(4.406)

(3.680)

(3.622)

4 Years After

0.973

4.354*

2.629

1.888

2.119

(1.339)

(2.279)

(2.068)

(1.675)

(1.693)

5+ Years After

-1.351

2.664

1.188

0.603

0.721

(1.508)

(2.440)

(2.227)

(1.807)

(1.738)

Observations

759

742

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variable for each column is the arrest rates for offenses against a family member, per

100,000 persons. MML=1 if a state implemented a medical marijuana provision. This table provides the

coefficient estimates from the regression model in (1) estimated by FE regression. Robust standard errors (in

parentheses) are based on Driscoll-Kraay spatial-autocorrelation and cluster-robust standard errors. All

regressions include state fixed effects, year fixed effects, state specific linear time trends. Crime rate controls

include the number of violent crimes per 100,000 inhabitants; family and drug abuse arrest rates per 100,000

persons; law enforcement to population ratio; female officers to population ratio; prisoner to population ratio.

State economic control variables include unemployment rate, female labor force participation rate, and state

median household income, college and high school attainment rate, population density per square mile. State

socio- demographic controls include the percent of the child population that is black and white; divorce rates,

percent of the population between the ages of 15-24 and 25-44; alcohol consumption per capita; beer

consumption per capita; population density per sq. mile.

Table 2.11 Heterogeneity in MMLs (HCL and MML-Pain): Child Maltreatment

(1) (2) (3)

Child maltreatment rate per 1,000 children

Dependent variable mean

13.231

MML=1

1.303*

(0.649)

HCL = 1

3.260**

(1.432)

MML-Pain = 1

1.854**

(0.839)

Observations

759

Within R-squared

.52

.526

.521

Number of groups

Table 2.12 Heterogeneity in MMLs (HCL and MML-Pain): Child Fatalities

(1) (2) (3)

Child fatality rate per 100,000 children

Dependent variable mean

1.841

MML =1

-0.174

(0.258)

HCL = 1

-0.508*

(0.293)

MML-Pain =1

-0.412*

(0.227)

Observations

794

Within R-squared

.363

.367

.365

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variables are child maltreatment rate per 1,000 children and child fatality rates per 100,000

children. MML=1 if a state implemented a medical marijuana provision. HCL =1 if a state allows for caregivers to

grow marijuana for medicinal purposes. MML-Pain =1 if a state has provisions that allow marijuana to be used for

chronic pain.

Table 2.13 Heterogeneity in MMLs: Child Maltreatment

(1) (2)

Child maltreatment rate per 1,000 children

(3)

Dependent variable mean

13.231

MML =1

1.303*

(0.649)

MML&Decrim =1

3.788**

(1.779)

Decrim&HCL = 1

4.308**

(2.000)

Observations

759

Within R-squared

.52

.527

.528

Number of groups

Table 2.14 Heterogeneity in MMLs: Child Fatalities

(1) (2)

Child fatality rate per 100,000 children

(3)

Dependent variable mean

1.841

MML =1

-0.174

(0.258)

MML&Decrim =1

-0.214

(0.429)

Decrim&HCL = 1

-0.323

(0.487)

Observations

794

Within R-squared

.363

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variables are child maltreatment rate per 1,000 children and child fatality rate per 100,000

children. MML=1 if a state has a medical marijuana provision. MML&Decrim =1 if a state has both MML and

decriminalization laws. Decrim&HCL =1 if a state has both home cultivation and decriminalization laws. All

regressions include state fixed effects, year fixed effects, state specific linear time trends. Crime rate controls

include the violent crime to population ratio; family and drug abuse arrest rates per 100,000 persons; law

enforcement to population ratio; female officers to population ratio; prisoner to population ratio. State economic

control variables include unemployment rate, female labor force participation rate, and state median household

income, college and high school attainment rate, population density per square mile. State socio- demographic

controls include the percent of the child population that is black and white; divorce rates, percent of the population

between the ages of 15-24 and 25-44; alcohol consumption per capita; beer consumption per capita; population

density per sq. mile. Robust standard errors (in parentheses) are based on Driscoll-Kraay spatial-autocorrelation and

cluster-robust standard errors.

Table 2.15 Robustness of Estimates with the Use of MML Effective Dates, by Age Cohort

(1)

(2)

(3)

(4)

(5)

Child maltreatment rates

per 1,000 children

All ages

0-3

4-7

8-11

12-15

Dependent variable mean

13.231

15.0515

12.426

10.177

9.985

MMLeffective=1

1.429**

3.021**

1.948*

1.630*

1.801*

(0.606)

(1.234)

(1.081)

(0.908)

(0.897)

Observations

759

742

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variables are child maltreatment rates, per 1,000 children and maltreatment by age groups.

MMLeffective=1 if a state’s medical marijuana provision became effective that year. This table provides the

coefficient estimates from the regression model in (1) estimated by FE regression. Robust standard errors (in

parentheses) are based on Driscoll-Kraay spatial-autocorrelation and cluster-robust standard errors.

All regressions include state fixed effects, year fixed effects, and state specific linear time trends Crime rate

controls include the number of violent crimes per 100,000 inhabitants; family and drug abuse arrest rates per

100,000 persons; law enforcement to population ratio; female officers to population ratio; prisoner to population

ratio. State economic control variables include unemployment rate, female labor force participation rate, and state

median household income, college and high school attainment rate, population density per square mile. State socio-

demographic controls include the percent of the child population that is black and white; divorce rates, percent of

the population between the ages of 15-24 and 25-44; alcohol consumption per capita; beer consumption per capita;

population density per sq. mile.

Table 2.16 Robustness of Estimates with the Use of MML Effective Dates, by Severity of Abuse

(1)

(2)

(3)

Severity of abuse

Physical abuse

rate/1,000 children

Neglect rate/1,000

children

Fatality

rate/100,000

children

Dependent variable mean

2.561

6.94

1.841

MMLeffective = 1

-0.489

1.057**

-0.083

(0.487)

(0.474)

(0.250)

Observations

761

762

794

Number of groups

*, **, and *** denote statistical significance at the 0.10, 0.05 and 0.01 levels, respectively.

Notes: The dependent variables are types of maltreatment (physical abuse and neglect), per 1,000 children and child

fatality rate per 100,000 children. MMLeffective=1 if a state’s medical marijuana provision became effective at year

T. This table provides the coefficient estimates from the regression model in (1) estimated by FE regression. Robust

standard errors (in parentheses) are based on Driscoll-Kraay spatial-autocorrelation and cluster-robust standard

errors.

All regressions include state fixed effects, year fixed effects, and state specific linear time trends Crime rate

controls include the number of violent crimes per 100,000 inhabitants; family and drug abuse arrest rates per

100,000 persons; law enforcement to population ratio; female officers to population ratio; prisoner to population

ratio. State economic control variables include unemployment rate, female labor force participation rate, and state

median household income, college and high school attainment rate, population density per square mile. State socio-

demographic controls include the percent of the child population that is black and white; divorce rates, percent of

the population between the ages of 15-24 and 25-44; alcohol consumption per capita; beer consumption per capita;

population density per sq. mile.

Students also viewed