Labour Economics 54 (2018) 152–171

Contents lists available at ScienceDirect

Labour Economics
journal homepage: www.elsevier.com/locate/labeco

Parent–child interactions and child outcomes: Evidence from randomized

intervention
Jun Hyung Kim a,∗, Wolfgang Schulz b, Tanja Zimmermann c, Kurt Hahlweg b
a
Department of Economics, University of Chicago, 1126 E. 59th st. Chicago, IL 60637, United States
b
Institute for Psychology, Braunschweig University of Technology, Germany
c
Center For Psychological Medicine, Hannover Medical School, Germany

a r t i c l e i n f o a b s t r a c t

JEL classification: Parent-child interactions are determined endogenously by child behavior, making identification of causal effects
D13 challenging. We overcome this endogeneity by analyzing a randomized, universal parent-training intervention on
J13 parents of preschool children. Evaluation of adolescent outcomes 10 years after the program suggests improve-
J18
ments to externalizing behaviors and wellbeing of children in the intervention group, mediated by changes to
J24
parenting during early childhood. These outcomes are not explained adequately by extant models of parent-child
Keywords: interactions, and so we explore alternative explanations. We show that benefits of early childhood interventions
Early childhood intervention extend beyond low-socioeconomic households.
Randomized controlled trial
Parental investment
Socioemotional skill
Discipline
Child development

1. Introduction tablishes a causal effect of parent-child interactions on outcomes during

This study reports empirical evidence of a causal effect of parent-
child interactions during preschool years on adolescent children’s be-
havioral outcomes. We focus on two aspects of parent-child interac-
tions—parents’ positive engagement with a child and parental disci-
pline. Psychologists have long recognized that parents’ harsh discipline
and poor monitoring lead to negative child outcomes,1 but economists
have only recently begun to study the role of parent-child interactions in
the human capital development of children. Empirical investigations to
date suggest that harshness and warmth during parenting predict adoles-
cents’ behavioral outcomes, and explain a portion of the variation in sib-
lings’ income (Björklund et al., 2010; Dooley and Stewart, 2007; Fiorini
and Keane, 2014). However, causal effects of parent-child interactions
remain a challenging question due to the endogeneity problem. For ex-
ample, a child’s misbehavior might be caused by parental punishments,
or parental punishments might be a response to a child’s misbehavior.
Experimental variation of a parent’s response might be a solution, but
most interventions that target parenting do not have follow-ups beyond
a few years after an intervention (Sanders et al., 2014). No evidence es-
∗
Corresponding author.
E-mail addresses: junhyung@uchicago.edu (J.H. Kim), wolfgang.schulz@tu-bs.de (W. Schulz), Zimmermann.Tanja@mh-hannover.de (T. Zimmermann),
k.hahlweg@tu-bs.de (K. Hahlweg).
1
See Dishion and McMahon (1998); Patterson (1986); Patterson and Dish-
ion (1985) and Patterson et al. (1984).
adolescence and beyond.
We study randomized parenting intervention for the parents of
preschool children in Germany.2 The intervention is a level 4 implemen-
tation of the Triple P Positive Parenting Program that targets parents of
children aged 2.6 to 6, a critical period for shaping long-term outcomes
of children (Cunha and Heckman, 2007). Parents are trained how to
engage with and effectively discipline their children while maintaining
the quality of the parent-child relationship. Since children do not partic-
ipate directly in the training, the effect of the intervention is channeled
through changes to a parent’s behaviors. Other than a language require-
ment,3 participants are not screened based on socioeconomic status or
child development risk factors, and thus findings from the intervention
are unlikely confined to children in low-socioeconomic households.
2
Psychologists distinguish prevention and treatment trials. Treatment trials
target subjects diagnosed with problems of interest, and prevention trials do
not. Since we assess the role of parenting styles generally, we consider only
prevention trials.
3
All parents must be fluent in German.

https://doi.org/10.1016/j.labeco.2018.08.003
Received 3 August 2017; Received in revised form 27 July 2018; Accepted 13 August 2018
Available online 21 August 2018
0927-5371/© 2018 Elsevier B.V. All rights reserved.
J.H. Kim et al.
The randomized intervention is based on the model of coercive dis-
cipline (Dishion and Patterson, 2015; Granic and Patterson, 2006; Pat-
terson, 1982; Patterson et al., 1992); a parent has limited control over
responses to a child’s misbehaviors. Failure to properly reinforce pro-
social behaviors and discipline anti-social behaviors leads to low aca-
demic and social skill during childhood, and delinquency and low self-
esteem during adulthood. The model suggests that improving a par-
ent’s ability to respond to a child’s misbehaviors leads to better child
outcomes. Analysis of the intervention suggests that improvements to
parental discipline and positive engagement during early childhood im-
prove behavioral outcomes and mental wellbeing during early adoles-
cence. Assessed 10 years after the intervention at ages 12 to 16, children
in the intervention group exhibited lower externalizing behaviors by 0.2
to 0.3 standard deviations and higher subjective wellbeing by 0.1 to 0.2
standard deviations. These long-term effects are mediated by changes to
parental discipline and positive engagement, which remain persistent at
least 4 years after intervention. We do not find an effect on children’s in-
ternalizing behaviors and social relationships. Although the intervention
had no overall effect on reducing internalizing behaviors, an increase in
positive engagement mediated a reduction in internalizing behaviors.
We build on evaluations of follow-up data collected 1 to 4 years
after an intervention (Hahlweg et al., 2010; Heinrichs et al., 2006;
2009), which suggest that intervention reduces harsh discipline and in-
creases positive engagement of the mother. The father’s parenting also
improves, but in smaller magnitudes. There were modest reductions in
children’s externalizing and internalizing behaviors throughout the pe-
riod. The current study contributes to these analyses in several ways.
First, we examine adolescent outcomes. Second, wellbeing and social re-
lationships are analyzed, in addition to externalizing and internalizing
behaviors. Third, linear mediation analysis is used to examine the role
of parenting behaviors in improving child outcomes. Fourth, the pres-
ence of non-compliers are assessed using two-stage least squares and
selection-bias tests from Black et al. (2017). Finally, inverse-probability
weighting is used to account for attrition.
This study contributes to literature on child human capital develop-
ment by providing evidence that parent-child interactions play a causal
role in behavioral skill development. We also show that parent-child in-
teractions can be improved by the intervention we study. Well-known,
effective early childhood interventions such as Perry Preschool Study,
Carolina Abecedarian Project, and Nurse-Family Partnerships involve
both home-visit and parenting components (Elango et al., 2015). Cur-
rent findings suggest that parent-child interactions are mediation chan-
nels in these programs. We show that the effectiveness of early child-
hood intervention is not necessarily limited to low-SES households. We
argue that extant models of parent-child interactions in economics are
inadequate at explaining the link between improved discipline and child
outcomes. We propose an alternative framework that accords with psy-
chological theory and empirical findings.
One limitation of this study is that the experiment is based on cluster-
randomization with 17 clusters, where 11 clusters are in the treatment
group and 6 in the control group. Conventional parametric tests based
on an asymptotic distribution, such as parametric t-test, are inappro-
priate. Our statistical inference is based on a bootstrap t-test using the
wild-cluster bootstrap procedure (Cameron and Miller, 2015), which al-
lows precise estimation of p-values, even with a small number of clus-
ters. We use the 10% significance level during tests. Combined with a
small number of clusters, our inferences have limited statistical power
to detect meaningful effects. We therefore caution that findings should
be interpreted as suggestive, as Benjamin et al. (2018) suggest.
2. Parent–child interaction models in psychology and economics
We review extant models of parent-child interactions during child-
hood from psychology and economics, including implications to this
study. Psychologists Gerald Patterson and others developed a model in
which coercive and inconsistent parenting during childhood leads to
153
Labour Economics 54 (2018) 152–171
delinquency in adolescence (Dishion and Patterson, 2015; Granic and
Patterson, 2006; Patterson, 1982; Patterson et al., 1992). Consider a
child who refuses to comply with a parent’s effort to modify the child’s
behaviors. In families in which children develop antisocial behaviors,
parent-child interactions are characterized by two patterns. In the first,
a parent uses coercive discipline strategies such as scolding and threat-
ening, and the child responds with hostility. The interaction escalates,
resulting in feelings of anger and contempt for both. In the second, the
parent-child interaction is characterized by lax and permissive parent-
ing; a parent responds to a child’s defiance by either giving up on disci-
pline or promising the child additional rewards for compliance. In both
cases, the parent fails to instill reasonable compliance to social norms in
the child. Combined with poor parental monitoring and the training in
coercive exchanges in the household, the child develops low academic
and social skills, and is more likely to engage in delinquency and re-
port low self-esteem during adolescence. Improving a parent’s ability
to reinforce pro-social behaviors and discipline anti-social behaviors re-
duces delinquency and improve self-esteem during adolescence through
changes to childhood parent-child interactions.
The idea that parent-child interactions affect the human capital de-
velopment of a child is not new to economics. A model from Baumrind4
has been reinterpreted as a model of a child’s preference formation and
restriction of the child’s occupation choice (Doepke and Zilibotti, 2017).
This parenting style model, however, does not correspond to the coer-
cive discipline that this paper assesses. A parent can be authoritarian
(i.e., strict and demanding) and simultaneously either consistent or in-
consistent in carrying out rewards and punishments. The intervention
that we study involves changing the way discipline is implemented. It is
unknown whether this change implies a change in parenting style, and
if so, in which direction.
Some researchers use principal-agent frameworks to describe parent-
child interactions. Weinberg (2001) proposes a model in which a parent
chooses between monetary incentive and corporal punishment to moti-
vate a child’s behaviors. In the model, paternalistic parents with binding
budget constraints choose corporal punishment, which is assumed non-
monetary. The parenting intervention we study does not involve finan-
cial transfers, but learning effective disciplinary strategies can be inter-
preted as lowering the cost of implementing non-harsh discipline rela-
tive to harsh discipline. The model, however, does not address whether
one incentive is more effective than another. Akabayashi (2006) and
Cosconati (2009) connect child outcomes and parents’ incentive use
through child effort and a child’s preference formation. A parent in
Akabayashi’s model transfers incentives to a child in response to the
child’s effort choices, measured using noisy signals. One implication
of the model is that maltreatment equilibrium might develop, during
which a parent repeatedly punishes a child by transferring too little in-
centive after initially overestimating the child’s effort. Improvements
to the parent’s signal-extraction (i.e., monitoring) technology, perhaps
induced by parent training, lower the probability of repeated punish-
ment in the model. The parent in Cosconati’s model imposes limits
on a child’s leisure activities to facilitate human capital development.5
Cosconati structurally estimates the model, demonstrating that the ef-
4 Baumrind (1966, 1968), and Maccoby and Martin (1983) presents 4 (ini-
tially 3) categories of parenting styles—authoritative, authoritarian, permissive,
and uninvolved. The authoritative parenting style is high on both control and
warmth dimensions of parenting, and associates with better behavioral and aca-
demic outcomes than other styles do. Authoritarian parenting is high on control
but low in warmth. Permissive parenting is high on warmth but low in control,
and uninvolved parenting is low on both dimensions and associates with the
worst child outcomes. Empirical findings, however, point to large heterogeneity
in the effect of parenting style across ethnic and socioeconomic groups, espe-
cially regarding the relative effect of authoritative and authoritarian parenting
styles. See Spera (2005) for more details.
5
Cosconati connects the severity of a parent’s control on a child’s leisure time
using a Baumrind parenting style model. However, we argue it is more appro-
J.H. Kim et al.
fect of parental control is heterogeneous in the sample; strict limits are
counterproductive to development of more academically advanced chil-
dren.
Although these models are interesting, they are not helpful to under-
standing and interpreting empirical analyses in this paper, which focus
on parenting intervention based on a coercive parenting model. It is thus
worthwhile to describe how insights from psychology can be understood
in the context of an economic model. We therefore sketch a simple model
of parent-child interactions, emphasizing distinctions from other models
that allow us to assess coercive parenting in an economic framework.
2.1. A model of parenting skill
We describe parent-child interactions using parenting skills as an
essential parameter that distinguishes coercive and non-coercive par-
enting. Consider a child and one parent who interact for a single period
during early childhood. The outcome of interest is the child’s adolescent
non-delinquency Y; the child can be either delinquent (𝑌 = 0) or not
delinquent (𝑌 = 1) during adolescence. Outcome 𝑌 = 1 is realized with
positive probability 𝑝 = 𝑒 × 𝐴, where e ∈ {0, 1} indicates whether the
child exerts effort during the first period toward building academic and
social skills. 𝑒 = 1 if the child exerted effort (i.e., acted pro-socially) and
𝑒 = 0 if the child did not exert effort (i.e., acted anti-socially). A ∈ (0, 1)
captures household, school, and neighborhood environments that influ-
ence adolescent outcome realization. If the child chooses 𝑒 = 1, outcome
𝑌 = 1 is realized with probability A. Otherwise 𝑌 = 0 with certainty. The
child is assumed myopic and does not internalize consequences of child-
hood effort choices realized during adolescence. The child pays effort
cost v for choosing 𝑒 = 1, which can be compensated for by the parent
to promote pro-social behaviors.
The distinction of the current model from other principal-agent mod-
els is that the parent does not have full control over how compensation
is realized. The parent announces parenting policy b ∈ {0, 1}, but the
utility transfer the child receives is a Bernoulli random variable B with
probability 𝑞 (𝑒, 𝑏) ≡ 𝑃 𝑟(𝐵 = 1|𝑒, 𝑏), where q increases in e and b. For sim-
plicity, assume 𝑞 (𝑒, 0) = 𝑞 (0, 𝑏) = 0 so nothing is transferred to the child
if the child chooses no effort or the parent does not announce 𝑏 = 1. The
child observes the parent’s announcement b and solves
max −𝑣 ⋅ 𝑒 + 𝐵 ⋅ 𝑞 (𝑒, 𝑏)
𝑒∈{0,1}
Specifically, if the parent announces 𝑏 = 1, the child’s choice becomes
max {−𝑣 + 𝐵 × 𝑞 (1, 1), 0}
Parenting skills determine a parent’s ability to respond to a child’s be-
havior by compensating for the child’s effort when the child chooses
𝑒 = 1. If inconsistent or over-reactive, the parent does not always suc-
ceed with rewarding pro-social behaviors or disciplining anti-social be-
haviors appropriately, even if that was the intention. These considera-
tions imply q(1, 1) < 1, and q(1, 1) is low when a parent uses coercive
parenting. The probability that a parent responds as intended to a child’s
behavior, q(1, 1), is interpreted as parenting skill.
The parent’s problem is announcing parenting b to the child, such
that
max −𝑐 ⋅ 𝑏 + 𝑒∗ (𝑏) × 𝐴
𝑏∈{0,1}
where e∗ (b) is the optimal response of the child.
We can now obtain the optimal solution of the child and parent. If
the parent announces 𝑏 = 1, the child chooses 𝑒 = 1 if q(1, 1) > v, and
chooses zero otherwise. The child chooses 𝑒 = 0 if the parent announces
𝑏 = 0. Given this solution of the child, the parent announces 𝑏 = 1 if q(1,
1) > v and −𝑐 + 𝐴 > 0 such that the incentive is sufficiently large to offset
the child’s effort cost, and environment A is favorable enough to justify
priate for our purposes to categorize it using other principal agent models due
to the similarities the models share.
154
Labour Economics 54 (2018) 152–171
the parent’s effort. The parent chooses 𝑏 = 0 if either −𝑐 + 𝐴 ≤ 0 or q(1,
1) ≤ v.
Based on this model, the observed discipline might have been an ap-
propriate response to a child’s behavior or an unjust response due to low
parenting skills. The effect of discipline on a child’s outcomes might be
positive for the former case but not for the latter. This implication is
also consistent with continued use of harsh parenting by some parents
even though they claim to know that their children’s outcomes will suf-
fer. The model implies that parents do not have sufficient skills to avoid
harsh parenting while trying to modify their children’s immediate be-
haviors. An appropriate intervention would be to increase parents’ skill
q(1, 1) so they can be in control of realized parenting. From the optimal
solution, increasing q(1, 1) increases e∗ , b∗ , and therefore the probability
of 𝑌 = 1.
3. Experimental design and background
3.1. Background
We evaluate an implementation of the level 4 Group Triple P Posi-
tive Parenting Program, a randomized intervention that provides educa-
tion and training to parents on parenting strategies (see (Sanders, 2012;
Sanders et al., 2014)). The program is available for levels 1 through 5,
where level 1 is an information campaign, and the intensity of involve-
ment increases at each level.6 Level 4, the implementation we study,
involves intensive parent training on-site in a group context, with prac-
tice sessions, homework, and instructor feedback.
The program draws on social learning models that highlight
the reciprocal and bidirectional nature of parent-child interactions
(Patterson, 1982). The model describes learning mechanisms, in which
coercive and dysfunctional parenting leads to future antisocial behav-
iors in children. Accordingly, the program teaches parents positive child
management skills as an alternative to coercive, inadequate, and in-
effective parenting practices. Emphasis is on a parent’s ability to use
child-initiated interactions as a context for incidental teaching, focusing
on a children’s core language competencies and impulse control (Hart
and Risley, 1975; 1995). The program also addresses risk factors linked
to adverse developmental outcomes in children, including poor parent
management practices, marital family conflicts, and parental distress
(Grych and Fincham, 1990).7
3.2. Comparison to other parenting interventions
A strength of this study is that it links parenting intervention in
early childhood to adolescent outcomes. Based on the most recent meta-
analysis of 101 Triple P programs from Sanders et al. (2014), the typical
follow-up period is 6 months after intervention, and no other Triple P
implementation administered follow-up measures beyond 1 year after
intervention. The follow-ups suggest that Triple P, on average, has a pos-
itive effect on child behaviors and parenting practices.8 They found that
greater effect sizes associated with more severe child problems at base-
line and using a targeted approach relative to a universal one, though
significant effect sizes were also found when using a universal approach.
Wilson et al. (2012) conducted a meta-analysis of Triple P inter-
ventions that consisted of 33 randomized controlled trials with child
6
Level 5 is program for children with special education needs.
7
These elements are combined with a population health perspective
(Biglan, 1995) that emphasizes the role of community support for parents. Par-
ents receive social and emotional support while receiving parental education,
and share difficulties of parenting with others. This aspect of the program is
achieved through media and the promotional strategy of Triple P.
8
The mean child age at the baseline during the studies included in this meta-
analysis was 5.85 years, but the average child age at baseline during the current
study was 4 years.
J.H. Kim et al.
behavior outcomes. Twenty-six of 33 studies used a waiting list con-
trol condition in which treatment was offered immediately after a post-
intervention assessment. The design precluded control group follow-ups
beyond the conclusion of the intervention. They point out that outcomes
for single parents tended to be non-significant, and the effect size for
mother-reported behavioral measures were sizable (0.61) and broadly
in line with those reported by other authors.
Lack of long-term follow-ups in other Triple P studies makes direct
comparisons between our study and extant literature difficult. Although
not directly comparable, Preparing for Life intervention in Ireland used
a level 4 Group Triple P and home visitations to improve the school
readiness of children, and were evaluated by Doyle et al. (2015, 2013),
and Doyle et al. (2016). The home visitation program begins in utero
and continues until the child is 5 years old, and the parenting course is
offered when the child is 2 to 3 years old. They find improvements to
parenting behaviors at 18 months of a child’s age. At 24 and 36 months,
they find improvements to a child’s health, particularly wheezing and
asthma. They also find that treatment lowers the risk of a child devel-
oping emotional and behavioral problems later in life. These effects are
stronger for boys.
Another partially comparable program is the Community Mothers
Programme (Johnson et al., 2000), which uses experienced volunteer
mothers in disadvantaged areas to support first-time parents with rear-
ing children up to 1 year of age. These community mothers visit par-
ticipants once a month, focusing on increasing the skills of parents by
“using a behavioral approach in which parents are encouraged to un-
dertake agreed tasks,” and by showing “the alternatives available to
parents in coping with various child-rearing problems.” Benefits were
found at both 1 year and 8 year follow-ups, when children were 8 years
old. At 8 year follow-ups, intervention-group parents were more likely
to monitor a child’s homework and TV watching, and to disagree to the
statement that corporal punishment is useful. Children visited libraries
more often, and children born after the intervention were more likely
to be immunized. Parents also reported greater self-esteem.
These studies suggest that a 10-year follow-up would
demonstrate positive influences on behavioral outcomes. Both
Sanders et al. (2014) and Wilson et al. (2012) report effect sizes
between 0.3 and 0.6 for short-term behavioral outcomes of children.
Johnson et al. (2000) shows that parenting-only interventions improve
both parents’ investment behaviors and parents’ perceptions of disci-
pline practices 7 years after an intervention, though these children are
about 6 years younger than the children in the current sample. Such
persistent changes in parents’ behaviors might be channels through
which long-term child effects are generated.
The Triple P program can be understood in the context of early child-
hood intervention literature since it addresses improvements to chil-
dren’s outcomes by altering early childhood environments. High-quality
early childhood interventions influence earnings, health, and criminal
behaviors of recipients (Elango et al., 2015). These findings are based
on interventions that feature both direct involvement with children and
home-visitation components that encourage parent participation. Thus,
it is difficult to distinguish direct effects of the program and indirect
effects through parents. The current study isolates the parent channel.
3.3. Experimental design
An intervention was implemented from 2001 to 2002 in Braun-
schweig, a small, urban city in Germany with a population of 250,000.
Program staff members presented the program to all 33 preschools in
the city, in which administrators of 23 preschools showed interest. Sev-
enteen of 23 preschools were selected randomly to participate due to
resource constraints. The intervention did not target at-risk or low-SES
households, unlike most other early childhood interventions. The inter-
vention excluded parents who could not understand German, so par-
ticipants were likely homogeneous regarding cultural backgrounds. All
parents received information about the program at the preschool and
155
Labour Economics 54 (2018) 152–171
decided whether to participate before randomization. To be eligible,
parents had to have a child aged 2.6 to 6 years old. Only one child
per family was targeted for the parenting intervention and data collec-
tion. Randomization was at the preschool level, with a 2/3 chance of a
preschool being in the intervention group. This experimental design en-
sures no control contamination between siblings and within preschools.
Of 915 eligible families in 17 preschools, 282 agreed to participate
in the program.9 After two families moved away, 280 families were in-
formed of their randomization status. The intervention group included
186 families in 11 preschools, and the control group included 94 fam-
ilies in 6 preschools. After randomization, 42 families assigned to the
intervention group refused to participate but agreed to respond to all
follow-up assessments.10
Although both parents were encouraged to participate, only about
6% of two-parent families did so. There was one single-father household.
The sample used in this study is larger than many other similar par-
enting interventions implemented before.11 However, we point out that
this study was limited in statistical power to detect small intervention
effects, primarily due to a cluster-randomization design with 17 clus-
ters. Cluster-randomization is popular among randomized-controlled tri-
als, and is sometimes the only option.12 The presence of a common el-
ement within each cluster diminishes the amount of independent in-
formation obtained from individual observations, lowering statistical
power.13 Smaller statistical power also implies that false positive rates
are higher (Benjamin et al., 2018). We do not rigorously calculate the
statistical power and false positive rates of the empirical model, but we
follow Benjamin et al. (2018) in reporting new statistical discoveries,
emphasizing that the evidence be interpreted as suggestive of causal
effects.
3.4. Program implementation
Interventions consisted of 4 weekly training sessions, 2 h each. Par-
ents were taught explanations for their children’s problematic behav-
iors, techniques to cope with the behaviors, and supportive strategies
for child development. These strategies were reinforced through role-
playing sessions. Subsequent to the group sessions, the parents had the
opportunity to hold 4 weekly individual telephone sessions, each 15–
20 min long, during which progress, questions, and difficulties that arose
regarding the Triple P Training could be discussed. These sessions stabi-
lized the implemented strategies and supported generalization to future
problems. Five licensed trainers led 28 groups of parents during the ses-
9
Thus, the experiment is based on parents who are self-motivated, and is not
nationally representative. The randomized design is not compromised since par-
ticipation choice is made before randomization. See Heinrichs et al. (2005) for
a study on reasons for non-participation. Privacy concerns were cited most of-
ten as the reason for opting out before randomization. Table 2 shows that the
differences in observable variables are small between the study’s sample and a
nationally representative sample.
10
Although non-compliers were in the intervention group, in Section 5.4, we
use their outcome measures to demonstrate that the effects of receiving the
intervention are robust to the presence of non-compliers.
11
This study is larger than all Triple P RCTs reviewed in Nowak and
Heinrichs (2008) except for one, and larger than all studies reviewed by
Wilson et al. (2012). Larger studies in Nowak and Heinrichs (2008) do not use
randomization.
12 Cluster-randomization reduces the risk of control-group contamination,
which might be significant in the present context; if some parents in the same
kindergarten were exposed to the treatment but others were not, the interven-
tion effect can easily spill over to control-group parents. It was also crucial to the
efficiency of program delivery; if participants were spread out over a large area,
they or the trainers would have to travel long distances for group sessions. Ei-
ther the program then becomes less cost-effective, or both program compliance
and survey retention suffers.
13
It also leads to underestimated standard errors when conventional paramet-
ric tests are used. We address this using a wild cluster bootstrap-t test.
J.H. Kim et al.
sions, usually at the participating preschools. Trainers were trained and
tested by the Triple P certification agency. During the program, trainers
received supervision weekly to maintain quality.
3.5. Interview procedure and measurement
Measurements were conducted 7 times—before intervention, imme-
diately after intervention, annually from 1 to 4 years after intervention,
and 10 years after intervention.14
3.5.1. Interview procedure
During the 10-year follow-ups, all families involved initially were
informed of the current project in writing. They were asked by phone
whether they were willing to participate. Data collection was conducted
using a combination of interviews and a written and electronically stan-
dardized survey. Interviews were conducted during home visits or at
the Technical University of Braunschweig. Parents (in 94% of cases, the
mothers) and children were interviewed concurrently but in separate
locations. The interviews were conducted by two interviewers, one for
the parent and one for the child (at least one of them with a Master’s
degree in Psychology, training to become a clinical psychologist [Ger-
man: Psychological psychotherapist]). Since some questions were sensi-
tive, children were questioned by an interviewer of the same gender. For
more sensitive questions, such as problematic behaviors, the respondent
used a tablet to answer electronically standardized surveys so that the
interviewer could not observe the answers.
The interview procedure caused minimum interference between par-
ents and children by ensuring they occurred simultaneously and at sep-
arate locations. In most cases, the mother responded to the parent ques-
tionnaire, but fathers responded when available. Literature that com-
pares self-reports of parents and children regarding problematic behav-
iors of a child (De Los Reyes and Kazdin, 2005; Stanger and Lewis, 1993;
Verhulst and Ende, 1992) commonly report positive but low correla-
tions between parent and child reports, suggesting that child self-reports
provide unique, important information on behaviors, especially among
older children.15
All interviewers were prepared for conversations with families
through intensive training, and were licensed to conduct diag-
nostic interviews or emotional disorders in children (Kinder-DIPS,
Schneider et al. (2009)). Participating families received compensation
of 80 € (110 USD, € 40 each for the adolescent and the parent). The
10-year follow-ups were conducted in March 2014. All procedures were
approved by the Human Subjects Protection Board of the German Asso-
ciation of Psychology.
3.5.2. Measurement
Data were collected primarily from mothers and children during the
10-year follow-ups. Cronbach’s alpha coefficients for internal consis-
tency appear in Table B.12. Parents’ discipline behaviors were mea-
sured using self-reports from a German version of the Parenting Scale
(Der Erziehungsfragebogen (EFB); Naumann et al. (2010), based on
Arnold et al. (1993).). The measure consists of 37 items designed to as-
sess laxness (i.e., inconsistency and ineffectiveness in discipline), over-
reactivity (i.e., harshness of discipline such as shouting and spank-
ing), and verbosity (e.g., long diatribes). They were measured using a
Likert-type scale, the anchors of which included “most likely” to “least
likely,” from 1 to 7 for responses regarding a child’s misbehaviors.
Table B.13 shows the parenting questionnaire used in the Triple P study,
14
Interview procedures for follow-ups until 4 years after an intervention are
described by Heinrichs et al. (2014).
15
More research is needed to determine the extent and direction of response
bias in child and parent self-reports. However, no evidence suggests favoring
one source over the other.
156
Labour Economics 54 (2018) 152–171
translated into English. Categorization of items followed the standard in
psychology literature.16
Positive engagement (Positive Parenting Questionnaire [PPQ]),
based on Strayhorn and Weidman (1988)., measured emotionally sup-
portive practices, such as praise and cuddling, in 13 items. Parents
rated the frequencies of each behavior during the past 2 months, from
‘never=0’ to ‘3=very often’. Child self-reported problematic behav-
iors at 10-year follow-ups were measured using the Youth Self Report
(Achenbach, 1991b), derived from the Child-Behavior Checklist. The
instrument consists of 112 items, designed for adolescents 11 to 18.
Each item is answered on a scale from zero (i.e., no symptom) to 2
(i.e., symptom is present most of the time or applies well), based on
a child’s experiences during the past 6 months. Sub-scales include in-
ternalizing disorders, externalizing disorders, attention problems, social
problems, and schizoid/compulsive abnormalities (i.e., thought prob-
lems). Internalizing disorders include withdrawal, somatic complaints,
and anxious/depressed sub-scales, and externalizing disorders include
delinquency and aggressive behaviors sub-scales. Wellbeing was mea-
sured using a German generic quality of life instrument for children
(i.e., KINDL, (Ravens-Sieberer and Bullinger, 1998)), which consists of
24 items answered using a 5-point Likert-type scale. Subscales include
physical wellbeing, emotional wellbeing, self-esteem, family, friends,
and school. We combined physical wellbeing, emotional wellbeing,
and self-esteem measures into a child wellbeing construct, and family,
friends, and school measures into a social relationship construct. All par-
enting and child outcome variables were constructed using sum scores,
and were normalized to a mean of zero and standard deviation of 1
within each follow-up. Sum scores restrict each item to equal weight,
and is a standard practice in psychological measurement.
Demographics measured at the baseline included household income,
mother’s education, and mother’s hours of work. Child’s IQ was mea-
sured at baseline using the Kauffman Assessment Battery for Children
(Kaufman and Kaufman, 1983). Child’s problematic behaviors were
measured using the Child Behavior Checklist (CBCL, Achenbach, 1991a)
at the baseline and during the first 4 follow-ups. Mother’s recalled par-
enting contained self-reports about the type of parenting that the mother
(i.e., trainee) received from her own parents. We used factor analysis to
derive a single factor for the parenting a mother received from her own
mother. Shown in Table B.15, punishment and controlling behaviors
had positive factor loadings, and warmth had negative loadings.
Although a standard measurement approach during psychological
research, self-reported measures are not free of self-report bias. The pri-
mary concern for this study is that mothers might be responding in a
socially desirable manner each time they respond, and during the next
4 years after intervention. We consider two extreme scenarios—either
a mother changed her behavior as reported or an unobserved third ele-
ment is causing the parent to lie in a way that is socially acceptable, im-
proving some of a child’s behavioral outcomes.17 In the latter case, the
intervention program was effective, but the proposed theoretical model
is no longer useful for interpretation. There is no way of removing con-
cerns of self-report bias in the study, but we assume measurements are
unbiased.18
16
We confirmed categorizations using exploratory factor analysis for each
follow-up.
17
The interview procedure for a child ensured that the adolescent felt the least
possible pressure from other family members or from the interviewer when an-
swering.
18 Discussed in Section 5.3, program-induced changes to parenting behaviors
correlated with child outcomes ten years after intervention. Mediation analy-
sis also suggests that parental positive engagement and a child’s internalizing
behaviors correlate, though no intervention effect was found for internalizing
behaviors. If responses on parenting were driven by self-report bias, parents
making biased reports were also parents whose children’s outcomes improved.
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table 1
Baseline Sample Characteristics.

Sample size Int’vn Mean Ctrl Mean p-value

Child’s age 278 4.03 4.13 0.43

Child’s gender(Female) 278 0.48 0.49 0.85
Number of siblings 278 1.04 1.19 0.31
Mother’s age 278 35.05 35.62 0.50
Mother’s work hours 278 2.78 3.52 0.16
Mother is HS grad 278 0.54 0.55 0.93
Single mother 278 0.05 0.11 0.23
Household income 271 4870.79 4327.96 0.17
CBCL Ext. 275 0.05 -0.10 0.31
CBCL Int. 275 0.08 -0.15 0.22
Mother’s harsh discipline 275 -0.05 0.11 0.21
Mother’s positive engagement 275 -0.02 0.05 0.53
Father’s harsh discipline 200 -0.03 0.08 0.61
Father’s positive engagement 200 -0.03 0.07 0.53
(IQ) KABC-mental 267 0.02 -0.05 0.74
(IQ) KABC-sequential 267 0.03 -0.06 0.58
(IQ) KABC-simultaneous 270 0.01 -0.05 0.78
(IQ) KABC-achievement 269 -0.01 0.00 0.94
Recalled parenting-own father 247 0.01 -0.01 0.89
Recalled parenting-own mother 267 0.00 -0.01 0.88

Note: The p-value is based on a two-tailed mean t-test between intervention and control groups using a wild cluster bootstrap-t test with 99,999 replications.
The Child Behavior Checklist (CBCL) measured a child’s problematic behaviors based on a parent’s report, in this case, a mother’s. Income is monthly household
income based on DEM from 2001 to 2002, when the exchange rate was approximately 1 DEM ≈ 0.54 USD. Recalled parenting measured parenting behaviors of a
responding mother (i.e., maternal grandparents of a respondent’s children). It was constructed by predicting factor scores from measures of warmth, control, and
punishment that a mother received. Higher scores mean less warmth, more control, and more punishment received by a mother.

3.6. Sample profile Table 2

Sample Comparison.
Attrition was 11% at the 10-year follow-up. We focus on parenting SOEP Triple P
behaviors of mothers since in most cases, only a mother attended the
Mean SD Mean SD
program. There was 1 single-father family, which was excluded from
analysis. The sample included 234 families in 17 preschools. Table 1 Mother’s age 35.18 4.94 35.18 4.99
Child is female 0.51 0.50 0.50 0.50
shows descriptive statistics of baseline characteristics. Pre-treatment
Household income 5090 2330 4685 1620
characteristics were balanced between intervention and control groups. Work hourr/day 2.51 3.05 3.06 3.17
The mean child age was 4, and child gender was split evenly. The av- Working full time 0.12 0.32 0.17 0.38
erage mother’s age was 35, and about half of the mothers graduated Working part time 0.41 0.49 0.39 0.49
from high school. Five percent of intervention and 10% of control group Single mother 0.17 0.38 0.07 0.26
Graduated high school 0.29 0.45 0.54 0.50
mothers were not married at the baseline. None of the t-tests rejected
Graduated college 0.16 0.37 0.41 0.49
the null of equal means between intervention and control groups, sug- Sample size 479 271
gesting the sample was balanced.
To assess the representativeness of the sample, Table 2 compares the Note: A SOEP sample was selected from wave S (2002) of sample E and F,
which are nationally representative and independently sampled from other
current sample with a sample of mothers in the nationally representative
samples in SOEP. Female respondents who are not immigrants and gave birth
German Socioeconomic Panel (SOEP) who would have been eligible for
between 1995 and 2000 were included in the sample, so they approximate
the program. The high school and college graduation rates, including the eligible population were Triple P available in their area. Net monthly
technical colleges, are much higher, and the share of single mothers is household income was converted to Deutsche Marks for 2002. High school
much lower in the current sample. Seventeen percent of mothers in the graduation equaled 1 if respondent’s last degree attained was upper sec-
sample work full-time, but only 12% in SOEP do. Therefore, although ondary or attended college. Full-time work status was defined as working
average monthly income is smaller in the sample, it is plausible that the 37 hours or more per week, and part-time status fewer than 37 hours. The
study sample is comprised of higher SES households. Triple P sample was collected from 2001 to 2002. High school graduation
Table B.11is a correlation table of all outcomes assessed during the was measured using “Abitur” status, which corresponds to high school in
study. The correlation between externalizing and internalizing behav- the United States. Full-time status was defined as working 7 hours or more
per day, and part-time status fewer than 7 hours. Household income was
iors was positive at 0.43. The correlation between wellbeing and social
measured in Deutsche Marks in the year it was answered, 2001 or 2002.
relationships was also positive at 0.60. Negative behaviors and quality
For both samples, college graduation equaled 1 if the respondent completed
of life (i.e., wellbeing and social relationships) correlated negatively, technical college or university.
with values between -0.64 and -0.38.

4. Econometric framework and outcome measures
Randomization was at the preschool level, so each preschool was
treated as a cluster. The standard model for evaluation of a randomized
experiment describes observed outcome 𝑌𝑖𝑘 of participant ik ∈ Ik by
𝑌𝑖𝑘 = 𝑍𝑘 𝑌𝑖𝑘 (1) + (1 − 𝑍𝑘 )𝑌𝑖𝑘 (0) (1)
where 𝑘 ∈ {1, … , 𝑁𝐾 } is an index for cluster and 𝐼𝑘 = {1𝑘 , … , 𝑁𝑘 } a set
of participants in cluster k. 𝑍𝑘 = 1 if cluster k was assigned to the inter-
vention group and 𝑍𝑘 = 0 otherwise. Intervention status was the same
for all participants in the same cluster. (𝑌𝑖𝑘 (0), 𝑌𝑖𝑘 (1)) are potential out-
comes for participant ik in cluster k. We test the null hypothesis of no
intervention effect, which is equivalent to counterfactual outcomes hav-
𝑑 𝑑
ing the same distribution: 𝑌𝑖𝑘 (0) = 𝑌𝑖𝑘 (1), where = denotes equality in
distribution. The intervention effect was estimated using ordinary least
squares:
𝑌𝑖𝑘 = 𝛼 + 𝛽𝑍𝑘 + 𝛾𝑋𝑖𝑘 + 𝜖𝑖𝑘 (2)

157
J.H. Kim et al.
The goal was to estimate 𝛽, the coefficient for assignment status Zk . 𝑋𝑖𝑘
is a vector of control variables. 𝜖𝑖𝑘 is an individual-specific error term
that might correlate within each cluster k, but is assumed to be inde-
pendent across k. To account for clustered error, confidence intervals
and p-values were calculated using the wild cluster bootstrap method
with 99,999 replications, which maintains cluster structure in each boot-
strap sample (see Cameron and Miller, 2015; Davidson and MacKinnon,
2010). p-values were calculated using bootstrapped t-statistics so that
the p-values are estimated more reliably than when Student’s t distribu-
tion is used.19 Wild cluster bootstrapping accounts for cluster random-
ization by adding the same wild residuals within each cluster, where
residuals are drawn from Webb’s 6-point distribution as Cameron and
Miller (2015) suggest. We are not aware of evidence suggesting that the
intervention injures children, so we conducted a one-tailed test at 10%
significance. For the remainder of the paper, all inferences are based on
wild cluster bootstrap t-tests with 99,999 replications.
Analyses explored gender differences in intervention effects. The lit-
erature suggests that gender differences exist during some interventions
(e.g., Heckman et al., 2013), but there is no theoretical foundation to
guide or interpret findings. We present these findings merely as empir-
ical discoveries.20
5. Results and discussion
We present an evaluation of the effect of Triple P interventions on
child behaviors measured 10 years after the intervention. Analyses adds
to evaluation of results from the first 4 years of follow-ups.21 Extant
analyses suggest that parents in the intervention group used less harsh
discipline and greater positive engagement during the 4 years after in-
tervention. These are reproduced in the next section. Child outcomes,
measured by maternal ratings of the Child Behavior Checklist (CBCL),
also improved. The effect size of the total CBCL score was 0.19 at the 4-
year follow-ups (Heinrichs et al., 2014).22 The current analysis assesses
10-year outcomes on child behaviors and wellbeing. Wellbeing was not
analyzed in previous evaluations. We also discuss findings in the context
of parent-child interaction models in psychology, and then in economic
models of parent-child interactions. Implications for these models were
consistent with mediation results. We analyze characteristics of non-
compliers, and estimate the effect on compliers using two-stage least
squares. Connection of empirical evidence and the economic model of
parent-child interactions, and analysis of non-compliers’ characteristics,
are unique to this paper.
5.1. Intent-to-treat
We present the intervention effect on parenting behaviors during
each of the follow-ups 1 through 4, and plot the effect in Figs. 1 and 2.
19
We do not gain statistical power using large bootstrap replications. Power
is limited by the sample’s size. The bootstrapping t-statistic allows more precise
inferences through asymptotic refinement.
20
Dishion and McMahon (1998) suggests that daughters are monitored more
than boys are. If so, daughters would benefit more from improvements to par-
enting.
21
See Hahlweg et al. (2010); Heinrichs et al. (2006, 2009); and
Heinrichs et al. (2014) For a study that assesses the role of varying incentives
to participation in a similar program, see Heinrichs and Jensen-Doss (2010).
22
Hahlweg et al. (2010) separates 2-year follow-up analyses for two-parent
(N=216) and single-parent families (N=59). For two-parent families, the effect
sizes on total CBCL score are 0.3, 0.19, and 0.32 for after intervention, 1-year
follow-up, and 2-year follow-up outcomes, respectively. For single-parent fam-
ilies, the effect sizes are -0.53, -0.71, and -0.57. The negative effects are due to
improvements in control group respondents. It is unclear how to interpret re-
sults from single-parents since the sample for this group was so small. We control
for single-parent status but do not conduct separate analyses for single-parent
households.
158
Labour Economics 54 (2018) 152–171
We estimate
𝑝𝑎𝑟𝑒𝑛𝑡𝑖𝑛𝑔𝑖𝑘 ,𝑡 = 𝛼𝑡 + 𝛽𝑡 𝑍𝑖𝑘 + 𝛾𝑡 𝑋𝑖𝑘 + 𝜖𝑖𝑘 ,𝑡 (3)
separately for parental discipline and positive engagement. Fig. 1 shows
that changes to parenting behaviors were persistent for at least 4 years
after intervention. Improvements to discipline behaviors were immedi-
ate after intervention, and an increase in positive engagement behav-
iors was significant 3 years after intervention. Analyzing components
of discipline separately in Fig. 2, verbosity, over-reactivity, and lax-
ness were lower in the intervention group, and the reduction in ver-
bosity was most consistent. Increases to positive engagement and re-
ductions to laxness suggest that improvements to discipline style did
not occur at the expense of overall discipline use. Fathers’ parenting be-
haviors also improved, though their response rate was much smaller.
Fig. B.6 and B.7 show parenting changes separately for boys and girls
sample.
One way to understand the persistence of intervention effects on
parenting behaviors is focusing on behavioral aspects of the program.
Triple P is Behavioral Parent Training (BPT), which provides training
on parenting skills through example and practice (Shaffer et al., 2001).
Part of the parenting skill involves breaking the negative reinforcement
trap—the parent tries to modify a child’s behaviors, the child resists,
and the parent either gives up, reinforcing the child’s resistance, or gets
tougher by yelling or spanking until the child complies, reinforcing the
parent’s harsh parenting. Reinforcement leads to the child’s problematic
behaviors, with harsh but ineffective parenting. Adjusting the parent’s
responses so the child’s behaviors can be modified without falling into
this trap is a skill learned through training, practice, and instructor feed-
back.23 An explanation for the persistence is that well-trained parents
create a different reinforcement mechanism in which better parenting
leads to better child behaviors, which allows the parenting skill to per-
sist through multiple channels (i.e., better parent-child relationships,
reduced parental stress, confidence in newly learned parenting meth-
ods, etc.). Long et al. (2017) compare Behavioral Parent Training (BPT)
and non-Behavioral Parent Training, in which BPT has a greater effect
on behavioral outcomes of parents and children. Extant literature does
not identify determinants of persistence on parenting.24
Table 3 shows intent-to-treat estimates in standard deviations in the
sample. The intervention effect on externalizing behaviors was -0.31,
significant at 10%. The effect on internalizing behaviors was positive,
contrary to expectations, but the magnitude was small and statistically
non-significant. The effect on quality of life was 0.161, and the effect on
child subjective wellbeing was 0.2, both significant at 10%. The effect
on a child’s social relationships was non-significant.
Examining effects separately for each gender, girls tended to benefit
more. For girls, the effect on externalizing behavior was -0.302, though
non-significant with 𝑝 = .115. The effect on wellbeing was greater at
0.3. For boys, effects were mostly smaller and none was significant.
The larger effect on externalizing behaviors for girls differed from
findings of targeted interventions, such as the Perry Preschool Pro-
gram and Abecedarian program, which found significant effects of early
childhood interventions for a male sub-sample for many outcomes.25
This finding accords with the idea that girls are monitored more, and
23
Parents are trained to correctly identify a child’s behavior, reinforce positive
behaviors, and provide short but consistent discipline. See Long et al. (2017) for
an introduction to parent training.
24
On the page 75 of Long et al. (2017): “Although the various reviews have
shown some support for child, parent, contextual, and program features as mod-
erators of response to parent training, most analyses are post-hoc and correla-
tional. Relatively little research has been conducted where these characteristics
have been studied as independent variables, which will be needed to establish
the validity of moderators.”
25
See Heckman et al. (2013) and García et al. (2017). Table 6 in
García et al. (2017) shows that the treatment effect of the Abecedarian program
is greater for males for a variety of outcomes.
J.H. Kim et al.
Fig. 1. Parenting Style by Mother and Father Note: Confidence intervals were calculated based on wild cluster bootstrap of intervention effect estimates with 99,999
replications. The Unit is standard deviations within each follow-up sample. Measures were collected through parent interviews, detailed by Heinrichs et al. (2014).
thus benefit more from the parenting program (see a related discus-
sion in Section 2). We cannot confirm whether girls were monitored
more in the sample. It is also possible that heterogeneity patterns
are affected not only by gender, but by household socioeconomics.
Havnes and Mogstad (2011) found that girls benefit more from univer-
sal childcare expansion in Norway.26 However, the policy in Havnes and
Mogstad (2011) is not directly comparable to the program we study. We
do not speculate further on the source of heterogeneity by gender, leav-
ing such investigations to future research.
Table 4 shows the robustness of estimates using various controls.
The first three columns show that when only demographic variables are
included, the effect on wellbeing is significant for both the pooled and
female samples. The middle three columns show results with only three
control variables of baseline skill and parenting measures. The effect on
externalizing behaviors is significant at 10% for the pooled sample. This
effect is now significant for boys, while not so in Table 3. The last three
columns show that none of the effects is significant when no controls are
included, though the magnitudes of the estimates are similar to those in
other models. We conclude that the estimates are consistent across mod-
26
Del Boca (2015) summarizes literature that suggests that untargeted child-
care expansion associates with better influences on girls.
159
Labour Economics 54 (2018) 152–171
els with various controls. The effects on externalizing behaviors are in
the range of −0.2 to −0.3 standard deviations, and the effects on wellbe-
ing are in the range of 0.1–0.3 standard deviations.27 The main results
in Table 3 include both demographic and baseline controls.
5.2. Connection to the theoretical model
Given current findings regarding the intervention and its outcomes,
we consider mechanisms that connect improved parenting and child out-
comes. We believe that the parents’ decision problem is the most plau-
sible channel for the observed effects, since the intervention interacted
only with the parents. Parents were not given material resources and the
time commitment was light, disqualifying channels through household
budget constraints. The intervention improved only some outcomes.
Thus it is implausible that results are driven entirely by response bias.
The spillover effect is disqualified since the study was randomized at the
preschool level and only one child was assessed per household. We can-
not reject the possibility that the intervention effect was driven by the
information provision, but we find it unlikely. Corporal punishment was
banned in Germany in 2000, and the government launched information
27
Table B.10presents model fit with only baseline control variables.
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table 3
Intent-To-Treat.

Variable name Ctrl mean Intervn effect Std. error p-Value R2

Pooled sample (N = 231)

Negative behavior total score -0.004 -0.120 0.157 0.276 0.105
Externalizing behavior 0.135 -0.310 0.151 0.095 0.111
Internalizing behavior -0.085 0.033 0.156 0.427 0.176
Quality of life total score -0.063 0.161 0.098 0.081 0.123
Wellbeing -0.090 0.200 0.110 0.069 0.156
Social relationship -0.005 0.067 0.076 0.199 0.074

Female sample (N = 111)

Negative behavior total score 0.119 -0.191 0.199 0.221 0.159
Externalizing behavior 0.045 -0.302 0.183 0.115 0.319
Internalizing behavior 0.260 -0.044 0.205 0.421 0.149
Quality of life total score -0.226 0.160 0.162 0.176 0.159
Wellbeing -0.323 0.301 0.151 0.043 0.175
Social relationship -0.048 -0.050 0.181 0.400 0.159

Male sample (N = 120)

Negative behavior total score -0.109 0.031 0.159 0.425 0.090
Externalizing behavior 0.212 -0.216 0.180 0.148 0.119
Internalizing behavior -0.380 0.175 0.138 0.130 0.138
Quality of life total score 0.077 0.092 0.142 0.272 0.104
Wellbeing 0.109 0.064 0.143 0.328 0.104
Social relationship 0.033 0.092 0.153 0.292 0.073

Notes: p-values are based on one-tailed t-tests, which are based on wild cluster bootstrap of t-statistics with 99,999 replications. p-values under 0.1 are in bold.
Negative behavior total score is a sum score of externalizing and internalizing behavior sub-scales based on Youth Self Report (YSR). Quality of life total score is a
sum score of Wellbeing and Social Relationship sub-scales, based on a German instrument for health-related quality of life for children (KINDL). These measures
are from 10-year follow-ups of the intervention, collected from 2012 to 2014. Control variables include demographics measured at the baseline, child behaviors,
and cognitive skills at the baseline, and parenting variables measured at the baseline and a mother’s recalled parenting received from her own mother.

Table 4
Intent-To-Treat with different control variables.

Variable name Effect p-Value R2 Effect p-Value R2 Effect p-Value R2

Pooled sample
Negative behavior total score -0.024 0.445 0.026 -0.070 0.346 0.030 0.010 0.523 0.000
Externalizing behavior -0.221 0.167 0.040 -0.290 0.081 0.050 -0.203 0.162 0.009
Internalizing behavior 0.091 0.719 0.084 0.094 0.712 0.016 0.127 0.797 0.003
Quality of life total score 0.130 0.153 0.027 0.085 0.261 0.031 0.088 0.242 0.002
Wellbeing 0.203 0.091 0.067 0.100 0.257 0.029 0.131 0.192 0.004
Social relationship 0.001 0.494 0.004 0.031 0.374 0.026 0.001 0.495 0.000
Demographic var. O X X
Baseline outcomes X O X
N 245 239 245

Female sample
Negative behavior total score -0.152 0.280 0.031 -0.127 0.287 0.058 -0.086 0.343 0.002
Externalizing behavior -0.235 0.166 0.049 -0.271 0.131 0.091 -0.202 0.164 0.009
Internalizing behavior -0.076 0.376 0.030 0.029 0.558 0.010 0.015 0.535 0.000
Quality of life total score 0.189 0.149 0.026 0.147 0.221 0.042 0.135 0.212 0.004
Wellbeing 0.348 0.034 0.073 0.213 0.141 0.042 0.206 0.142 0.008
Social relationship -0.047 0.600 0.031 0.021 0.456 0.046 0.005 0.487 0.000
Demographic var. O X X
Baseline outcomes X O X
N 115 113 115

Male sample
Negative behavior total score 0.086 0.715 0.032 -0.013 0.468 0.015 0.090 0.719 0.002
Externalizing behavior -0.203 0.198 0.037 -0.292 0.090 0.025 -0.200 0.213 0.009
Internalizing behavior 0.218 0.946 0.045 0.152 0.866 0.036 0.211 0.944 0.014
Quality of life total score 0.070 0.326 0.019 0.047 0.381 0.032 0.054 0.349 0.001
Wellbeing 0.092 0.304 0.027 0.037 0.408 0.028 0.074 0.325 0.001
Social relationship 0.012 0.472 0.011 0.029 0.437 0.024 0.000 0.500 0.000
Demographic var. O X X
Baseline outcomes X O X
N 130 126 130

Notes: p-values are based on one-tailed t-tests, which are based on wild cluster bootstrap of t-statistics with 99,999 replications. p-values under 0.1 are in bold.
Negative behavior total score is a sum score of externalizing and internalizing behavior sub-scales based on Youth Self Report (YSR). Quality of life total score is a
sum score of Wellbeing and Social Relationship sub-scales, based on a German instrument for health-related quality of life for children (KINDL). These measures
are from 10-year follow-ups of the intervention, collected from 2012 to 2014. Control variables include demographics measured at the baseline, child behaviors,
and cognitive skills at the baseline, and parenting variables measured at the baseline and a mother’s recalled parenting received from her own mother.

160
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Fig. 2. Harmful Discipline Sub-scales by Mother and Father Note: Confidence intervals were calculated based on wild cluster bootstrap of intervention effect
estimates with 99,999 replications. The unit is standard deviations within each follow-up sample. Measures were collected through parent interviews, detailed by
Heinrichs et al. (2014).

campaigns against such punishments (Bussmann, 2004). Both the inter-
vention and control groups would have been exposed similarly to infor-
mation about the negative effects of harsh parenting. The most likely
explanation is that the intervention improves parenting skills through
behavioral training, consistent with the model proposed in Section 2.1.
5.3. Mediation analysis
Figs. 3–5 show decompositions of intervention effects using linear
mediation analysis. The intervention was successful at changing a par-
ent’s behavior and a child’s outcomes, but we are able to argue that
the intervention effect is channeled through changes to a parent’s be-
haviors only if experimentally induced changes to a parent’s behaviors
correlate with experimentally induced changes to a child’s outcomes.
For each outcome, the intervention effect is decomposed into change in
discipline, change in positive engagement, and unobserved channels. By
unobserved channels we mean all experimentally induced changes not
captured by measured parenting changes.
Mediation analysis tests the validity of theoretical models that un-
derlie interpretation of data. The coercive parenting model suggests that
improvements to a parent’s reinforcement for prosocial behaviors and
discipline for antisocial behaviors lead to better behavioral outcomes
during adolescence. The economic model developed in Section 2.1 fo-
cuses on the role of improved discipline in a child’s behaviors. If these
models are valid, we should observe that experimentally induced im-
provements to parenting meaningfully explains intervention effects on
behavioral outcomes. Mediation analysis estimated a linear model:
The derivation of this equation is in Appendix A. Identification of the
model is based on results from Heckman et al. (2013) and Heckman and
Pinto (2015).
Each model controls for the same baseline characteristics as the ITT
model used to estimate Table 3. Fig. 3 shows mediation analysis esti-
mates in the pooled sample. Improvements to discipline and increases
to positive engagement both explain intervention effects. Lack of re-
sults for internalizing behaviors is due to the unmeasured channel, and
increases in positive engagement contributed to reductions of internal-
izing behaviors. Discipline and positive engagement both played roles
in reducing externalizing behaviors for girls. For boys, improvements to
discipline contributed to reductions in externalizing and internalizing
behaviors. Wellbeing for girls improved in the ITT estimate, but most
went through unmeasured channels. Increases to positive engagement
explains wellbeing improvements for boys, though the overall effect was
non-significant.
Mediation results are consistent with implications of the parent-child
interaction model in this paper. Harsh discipline measures how inconsis-
tent, over-reactive, and verbose a parent is when disciplining a child’s
behavior. These are characteristics of inept parenting that fail to dis-
courage anti-social behaviors, leading to low academic and social skills.
Positive engagement can be either interpreted as direct parental invest-
ment or positive reinforcement of pro-social behaviors. Consistent with
the economic model in Section 2.1, improvements to parental discipline
contribute to both behavioral and wellbeing outcomes.
5.4. Accounting for non-compliance

[ ] [ ] Intent-to-treat estimates provide information on the effect of provid-

𝐸 𝑌1 − 𝑌0 = 𝐚1 𝐸 𝐷𝐼𝑆𝐶1 − 𝐷𝐼𝑆𝐶0
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
change in discipline style
[ ]
+ 𝐚2 𝐸 𝑆 𝑈 𝑃 𝑃1 − 𝑆 𝑈 𝑃 𝑃0
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
increase in positive engagement
( )
+ 𝜏1 − 𝜏0 + 𝐛1 − 𝐛0 𝑋𝐾 ′
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
other factors
ing an intervention, but is not the effect of participating in the program
itself due to the presence of non-compliance in the intervention group.
This section investigates the nature of non-compliance in the sample,
(4) assessing whether findings are robust to non-compliance. Table 5 shows
161
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Fig. 3. Mediation analysis for pooled sample Notes: p-values are based on right-side, one-tailed t-tests, which were calculated from wild cluster bootstrap of t-statistics
with 99,999 replications. The stars in the figure indicate statistical significance of each component against the null hypothesis of zero contribution. See Table B.14 for
p-values from mediation analysis.

Table 5
Baseline sample characteristics of compliers and non-compliers.

Compliers Noncompliers Control Two-sided p-values

(1) (2) (3) (1)-(2) (2)-(3)

Sample size 144 41 93
Child age 4.03 4.00 4.13 0.90 0.46
Child female 0.47 0.54 0.49 0.52 0.67
# siblings 1.08 0.88 1.19 0.16 0.12
Mother age 35.47 33.59 35.62 0.09 0.04
Mother work hour 2.53 3.63 3.52 0.10 0.85
Mother HS 0.56 0.49 0.55 0.59 0.64
Single mother 0.04 0.07 0.11 0.51 0.55
Net income 5040.15 4304.88 4327.96 0.06 0.93
CBCL (ext) 0.13 -0.22 -0.10 0.07 0.44
CBCL (int) 0.09 0.05 -0.15 0.87 0.41
Harsh discipline (mother) -0.05 -0.03 0.11 0.94 0.42
Positive engagement (mother) -0.09 0.19 0.05 0.10 0.29
Harsh discipline (father) -0.04 0.03 0.08 0.75 0.76
Positive engagement (father) 0.01 -0.27 0.07 0.20 0.21
IQ (KABC ment.) 0.09 -0.24 -0.05 0.09 0.43
IQ (KABC seq.) 0.04 0.01 -0.06 0.92 0.78
IQ (KABC simu.) 0.15 -0.46 -0.05 0.01 0.12
IQ (KABC achiv.) 0.07 -0.28 0.00 0.13 0.35
Recalled parenting (father) -0.04 0.18 -0.01 0.27 0.48
Recalled parenting (mother) -0.09 0.35 -0.01 0.02 0.06

Note: p-values are based on mean t-tests between respective groups using bootstrap-t tests with 99,999 replications. Income was monthly household income based
on DEM from 2001 to 2002, when the exchange rate was approximately 1 DEM ≈ 0.54 USD. Recalled parenting behaviors measured parenting behaviors of the
parents of a responding mother (i.e., maternal grandparents of a respondent’s children).

baseline characteristics of compliers, non-compliers, and those in the
control group. Since the compliance decision was made after random-
ization, there is evidence of self-selection. Mean mother age was 35.47
for compliers, higher than 33.59 for non-compliers. Average household
income was also higher for compliers. These mean differences are sig-
nificant at 10%. Average household income was higher for compliers,
but baseline externalizing behaviors were also higher than for non-
compliers. Baseline child IQ was higher for compliers on two of the
four IQ sub-scales. Based on the measure of recalled parenting, complier
mothers reported receiving more warmth, less control, and less pun-
ishment from their own mothers than non-compliers did. Non-complier
mothers received less warmth, more control, and more punishment than
both compliers and the control group did. Compliers tended to be moth-
ers whose children engaged in greater externalizing behaviors, and who
received more warm parenting from their own parents.
Differences are also observable in outcomes. Table 6 shows that, on
average, the children of non-complier mothers reported greater adoles-
cent externalizing behaviors than those of complier mothers, but less
than children of control group mothers. Children of non-complier moth-
ers reported greater internalizing behaviors, lower wellbeing, and lower
social relationships than children in other groups. Although the children
of non-complying mothers engaged in less problematic behaviors at the

162
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table 7
Non-compliers versus control group.

(N=109) 𝛽 est. p-value Std.Error

Negative behavior total score -.057 0.399 0.138

Externalizing behavior -.289 .049 0.139
Internalizing behavior 0.110 0.300 0.191
Quality of life total score -.022 0.460 .186
Wellbeing 0.00 0.499 0.229
Social relationship -.057 0.342 .138

Fig. 4. Mediation analysis for female sample Notes: p-values are based on right-
side, one-tailed t-tests, which were calculated from wild cluster bootstrap of
t-statistics with 99,999 replications. The stars in the figure indicate statistical
significance of each component against the null hypothesis of zero contribution.
See Table B.14 for p-values from mediation analysis.
Notes: p-values are based on one-tailed t-tests, which are based on the wild
cluster bootstrap of t-statistics with 99,999 replications. p-values under 0.1
are in bold. Negative behavior total score is the sum of externalizing and in-
ternalizing behavior sub-scales, based on the Youth Self Report (YSR). Qual-
ity of life total score is the sum of wellbeing and social relationships sub-
scales, based on a German instrument for health-related quality of life for
children (i.e., KINDL). These measures are from 10-year follow-ups of inter-
vention, collected from 2012 to 2014. Control variables include demograph-
ics measured at the baseline, child behavior and cognitive skill measures at
baseline, and parenting variables measured at baseline, and a mother’s par-
enting received from her own mother.
Table 8
Two-stage least squares.
Variable name TSLS estimate p-value

Pooled sample (N=231) First stage F: 298.514

Negative behavior total score -0.170 0.267
Externalizing behavior -0.430 0.066
Internalizing behavior 0.050 0.421
Quality of life total score 0.220 0.082
Wellbeing 0.280 0.069
Social relationship 0.090 0.200

Female sample (N=111) First stage F: 98.833

Negative behavior total score -0.280 0.213
Externalizing behavior -0.440 0.110
Internalizing behavior -0.060 0.419
Quality of life total score 0.230 0.180
Wellbeing 0.440 0.046
Social relationship -0.070 0.391

Male sample (N=120) First stage F: 201.872

Negative behavior total score 0.040 0.422
Externalizing behavior -0.280 0.135
Internalizing behavior 0.230 0.118
Quality of life total score 0.120 0.274
Wellbeing 0.080 0.333
Social relationship 0.120 0.287

Fig. 5. Mediation Analysis for Male Sample Notes: p-values are based on right-
side, one-tailed t-tests, which were calculated from wild cluster bootstrap of
t-statistics with 99,999 replications. The stars in the figure indicate statistical
significance of each component against the null hypothesis of zero contribution.
See Table B.14 for p-values from mediation analysis.
Table 6
Mean outcome of complier group, non-complier group, and control group.
Compliers Non-compliers Ctrl group
Notes: p-values are based on one-tailed t-tests, which are based on the wild
cluster bootstrap of t-statistics with 99,999 replications. p-values under 0.1
are in bold. Negative behavior total score is the sum of externalizing and in-
ternalizing behavior sub-scales, based on the Youth Self Report (YSR). Qual-
ity of life total score is the sum of wellbeing and social relationships sub-
scales, based on a German instrument for health-related quality of life for
children (i.e., KINDL). These measures are from 10-year follow-ups of inter-
vention, collected from 2012 to 2014. Control variables include demograph-
ics measured at the baseline, child behavior and cognitive skill measures at
baseline, and parenting variables measured at baseline, and a mother’s par-
enting received from her own mother.

Negative behavior total score -0.026 0.109 -0.004

Externalizing behavior -0.085 -0.012 0.135 baseline, 10 years later, they were overtaken by the children of complier
Internalizing behavior 0.017 0.124 -0.085 mothers in terms of mean externalizing behaviors.
Quality of life total score 0.085 -0.171 -0.063
Wellbeing 0.092 -0.129 -0.090
We additionally use a framework from Black et al. (2017) to in-
Social relationship 0.053 -0.191 -0.005 vestigate the role of unobserved variables on non-compliers. We esti-
𝑗
mate 𝑌𝑑=0 = 𝛼𝑑=0 + 𝛽𝑑=0 𝑍𝑘 + 𝛾𝑑=0 𝑋𝑖𝑘 ,𝑑=0 + 𝜖𝑖𝑘 ,𝑑=0 for 𝑑 = 0, where d is
Notes: Negative behavior total score is the sum of externalizing and inter-
an indicator of participating in the intervention. We therefore estimate
nalizing behaviors sub-scales, based on the Youth Self Report (YSR). Quality
Eq. (2) but only for the control and non-compliance groups using as-
of life total score is the sum of wellbeing and social relationship sub-scales,
based on a German instrument for health-related quality of life for children
signment status as a treatment variable. Table 7 shows that children in
(i.e., KINDL). These measures are from the 10-year follow-ups, collected the non-compliance group reported less externalizing behaviors, greater
from 2012 to 2014. internalizing behaviors, and about the same wellbeing as children in the
control group. Coefficient estimates for 𝛽 were smaller than 0.2, and the

163
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table 9
Intent-To-Treat (IPW).

Variable name Ctrl mean Intervn effect Standard error p-value

Pooled sample (N=224)

Negative behavior total score -0.004 -0.103 0.154 0.293
Externalizing behavior 0.135 -0.308 0.152 0.094
Internalizing behavior -0.085 0.070 0.145 0.330
Quality of life total score -0.063 0.126 0.101 0.128
Wellbeing -0.090 0.156 0.104 0.094
Social relationship -0.005 0.048 0.093 0.315

Female sample (N=111)

Negative behavior total score 0.119 -0.183 0.210 0.234
Externalizing behavior 0.045 -0.307 0.195 0.126
Internalizing behavior 0.260 -0.025 0.197 0.449
Quality of life total score -0.226 0.125 0.184 0.255
Wellbeing -0.323 0.261 0.168 0.075
Social relationship -0.048 -0.069 0.218 0.386

Male sample (N = 113)

Negative behavior total score -0.109 0.053 0.165 0.379
Externalizing behavior 0.212 -0.208 0.183 0.158
Internalizing behavior -0.380 0.205 0.141 0.097
Quality of life total score 0.077 0.069 0.136 0.306
Wellbeing 0.109 0.046 0.139 0.369
Social relationship 0.033 0.063 0.142 0.330

Notes: p-values are based on one-tailed t-tests, which are based on the wild cluster bootstrap of t-statistics with 99,999 replications. p-values under 0.1 are in bold.
Negative behavior total score is the sum of externalizing and internalizing behavior sub-scales, based on the Youth Self Report (YSR). Quality of life total score is
the sum of wellbeing and social relationship sub-scales, based on a German instrument for health-related quality of life for children (i.e., KINDL). These measures
are from the 10-year follow-ups of intervention, collected from 2012 to 2014. Control variables included demographics measured at baseline, child behavior and
cognitive skill measures at baseline, and parenting variables measured at baseline, and a mother’s parenting received from her own mother.

null hypothesis of 𝛽 = 0 was not rejected for all outcomes other than ex-
ternalizing behaviors. For externalizing behaviors, the estimate of 𝛽 was
-0.289. Since none of these groups received intervention, we interpret
these estimates to mean that parents in the non-compliance group had
unobserved information, suggesting that their children would engage in
less externalizing behaviors in the future in comparison to those in the
control group.28
We discuss the two-stage least squares estimates, in which random-
ized assignment status was used as an instrument for program partici-
pation. We interpret these estimates as the effect of the intervention on
compliers. As expected from descriptive statistics, Table 8 shows that
the two-stage least squares estimates were greater than intent-to-treat
estimates. The effect on externalizing behaviors was -0.43, and the effect
on child wellbeing was 0.28. Corresponding ITT estimates were −0.31
and 0.2, respectively. For girls, the effect on externalizing behaviors was
similar to the effect in the pooled sample, and the effect on wellbeing
was greater at 0.44. The effect on externalizing behaviors was no longer
significant, with 𝑝 = .11. For boys, all effects were greater than for those
in the pooled sample, but none were significant and all were smaller
than effects for the female sub-sample.
the missing status of individuals at 10-year follow-ups using available
covariates.30 The model was estimated separately for intervention and
control groups to account for the possibility that the pattern of attrition
was different between the two.31
Table 9 repeats the intent-to-treat analyses in Section 5.1 with IPW
applied. Overall, estimates were qualitatively similar to the main anal-
yses. The effect on externalizing behaviors remained robust to IPW, and
the magnitude changed little in both the pooled sample and gender sub-
samples. The effect on child wellbeing also remained robust, though its
magnitude declined from 0.2 to 0.156 in the pooled sample.
5.6. The father’s role
Although fathers’ parenting improved, only 6% of fathers in two-
parent households participated directly in the program. It is unclear
whether improvements were due to the spillover effect from the mother,
responses to changes to a child’s behaviors, or another channel. Me-
diation analyses on parenting variables from both parents had non-
significant and inconsistent results. Investigating the role of other adults
in the household, including the father, is promising for future research.

5.5. Accounting for attrition and non-response

Attrition was low at 11% 10 years after intervention, but attrition
and item non-response might still introduce bias in intervention effect
estimates. We test the robustness of the main analyses using inverse
probability weighting (IPW). Assuming that missing data are indepen-
dent of the unobserved variables conditional on observable variables,29
the probability of missing in the sample can be used to correct for attri-
tion bias. The probability is constructed as a logit model that predicts
28
None of these mothers received Triple P interventions, but it is unknown
whether they sought alternative programs. We therefore use only one part of
the test that Black et al. (2017) describe since we cannot compare compliers
with always-takers in the control group.
29
This assumption is essential to IPW being able to reduce attrition bias, but
unfortunately, there is no way to test it.
6. Conclusion
This study reports evidence of the effect of parent-child interactions
during early childhood on behavioral outcomes during early adoles-
cence. The evidence derives from evaluating a randomized interven-
tion that targets parents of preschool children and provides education
and training on non-harsh discipline methods and positive engagement.
The intervention reduced externalizing behaviors and improved well-
being of children 10 years after the intervention, during a child’s early
30
Covariates include the number of a child’s siblings, single mother status,
household income at baseline, mother’s age, mother’s high school graduation,
mother’s hours of work, and whether a pregnancy was planned. The logit model
includes all of 278 households.
31
See Campbell et al. (2014) for detailed procedures.

164
J.H. Kim et al. Labour Economics 54 (2018) 152–171

adolescence. We investigate heterogeneous effects by gender, account of Human Development at the University of Chicago and the Institute
for non-compliance in the intervention group, and adjust for attrition. for Psychology at Braunschweig University of Technology for supporting
The effects were greater for girls and remain robust when accounting this study.
for non-compliance and attrition. Mediation analyses suggest that ex-
perimentally induced changes in parenting are channels through which Appendix A. Derivation of linear mediation model
interventions improved child outcomes.
This study suggests that good discipline and positive engagement We suppress cluster index k. The discussion here follows that found
during childhood reduce adolescent externalizing behaviors, which is in Heckman and Pinto (2015). Consider an observed child’s outcome
often used as a measure of socio-emotional skill. This study is first to 𝑌 = 𝐷𝑌1 + (1 − 𝐷)𝑌0 (A.1)
provide causal evidence on the effects of parent-child interactions dur-
where D is a binary variable indicating assignment to the intervention
ing childhood on adolescent outcomes. We interpret findings in light of
group, 𝐷 = 1 represents the intervention group assignment, and 𝐷 =
theoretical models in psychology and economics, and highlight the role
0 otherwise. Subscript d ∈ {0, 1} is used to represent variables when
of parenting skills during parent-child interactions. For early childhood
assignment status is fixed at d. The linear system is defined as:
intervention literature, we show that light-touch parenting interventions
improve behavioral outcomes observed as late as early adolescence, 𝑌𝑑 = 𝜅𝑑 + 𝐚𝑑 𝐏𝑑 + 𝐛𝑑 𝑋 + 𝜖̃𝑑
even in households that are not at-risk. These results support short, di- where 𝜅 d is an intercept and ad and bd are parameter vectors. X are pre-
rect parental interventions, or interventions in combination with other program variables that do not respond to the program. However, their
programs. effects on the outcome can be influenced by the program, and hence
We are cautious to claim external validity of a randomized trial in a the subscript on bd . 𝜖̃𝑑 is a zero-mean error term assumed to be inde-
population with disparate characteristics and institutional backgrounds
since the intervention consisted of German-speaking parents and chil-
pendent of all other regressors. ( Pd is a vector
) of parental inputs that
respond to the program: 𝐏𝑑 = 𝐏𝑗𝑑 ∶ 𝑗 ∈ 𝐽 , where J is an index set of
dren in the German city of Braunschweig. Evidence is based on cluster-
randomization design, with 11 clusters in the treatment and 6 clusters parental inputs. We denote Jm to be an index set of parental inputs that
in control. Statistical power was lower and false positive rates were are measured.
higher than in many other studies in economics literature. Thus, we Although the treatment program measured various parental behav-
present our study as the first of many potential future studies that will iors, many aspects of parental inputs were plausibly not observed and
deepen understanding of causal relationships between parent-child in- measured, some of which might respond to an intervention program.
teractions and child outcomes. We therefore suggest directions for fu- We decompose the parental input ad Pd into measured and unmeasured
ture research. Short-term benefits of parenting interventions have been inputs:
∑ 𝑗 𝑗 ∑ 𝑗 𝑗
well-established, but longer effects have not. There should be more 𝑌 𝑑 = 𝜅𝑑 + 𝐚𝑑 𝐏𝑑 + 𝐚𝑑 𝐏𝑑 + 𝐛𝑑 𝑋 + 𝜖̃𝑑
follow-ups to extant studies. Models from Patterson and Dishion, and 𝑗∈𝐽𝑚 𝑗∈𝐽 ⧵𝐽𝑚
from social interaction literature in economics, point to roles of peers in ⏟⏞⏞⏟⏞⏞⏟ ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
measured parental inputs unmeasured parental inputs
delinquent behaviors of adolescents (see Agostinelli, 2018 for a recent ∑
study). Research should assess whether parenting affects the friendship = 𝜏𝑑 + 𝐚𝑗𝑑 𝐏𝑗𝑑 + 𝐛𝑑 𝑋 + 𝜖𝑑 (A.2)
networks of adolescents. Other adults in the household and a child’s 𝑗∈𝐽𝑚

siblings might interact with parenting and parent-child interactions in ∑ ( )

where 𝜏𝑑 = 𝜅𝑑 + 𝑗∈𝐽 ∖𝐽𝑚 𝐚𝑗𝑑 𝐸 𝐏𝑗𝑑 and 𝜖𝑑 = 𝜖̃𝑑 +
generating child outcomes. Recent studies (Lee et al. 2013, Long et al. ( ( ))
∑ 𝑗 𝑗 𝑗
2017) suggest that parents with specific genotypes are more likely to use 𝑗∈𝐽 ∖𝐽𝑚 𝐚𝑑 𝐏𝑑 − 𝐸 𝐏𝑑 is a mean-zero error term.
harsh parenting when exposed to economic adversity. Collecting genetic 𝑑𝑖𝑠𝑡 𝑑𝑖𝑠𝑡
Without loss of generality, assume 𝜖̃1 = 𝜖̃0 where = means equality
information would allow researchers to investigate the heterogeneous
in distribution.
effect of parenting. Finally, it would be useful to identify variables that
The goal is to decompose the treatment effect into changes in mea-
affect the effectiveness and persistence of parent training.
sured and unmeasured parental inputs. Consider decomposition:
Funding [ ]
[ ] ∑ 𝑗 𝑗 𝑗 𝑗 ( )
𝐸 𝑌1 − 𝑌0 = 𝜏1 − 𝜏0 + 𝐸 𝐚1 𝐏1 − 𝐚0 𝐏0 + 𝐛1 − 𝐛0 𝑋
This intervention was funded by the Deutsche Forschungsgemein- 𝑗∈𝐽𝑚
[ ]
schaft (DFG; German Research Foundation), HA 1400/14-1-5. ∑ ( )
= 𝜏1 − 𝜏0 + 𝐸 𝐚𝑗1 𝐏𝑗1 − 𝐚𝑗1 𝐏𝑗0 + 𝐚𝑗1 𝐏𝑗0 − 𝐚𝑗0 𝐏𝑗0 + 𝐛1 − 𝐛0 𝑋
Conflict of interest 𝑗∈𝐽𝑚
[ ]
∑ ( ) ( ) ( )
The authors declare no competing interests. = 𝜏1 − 𝜏0 + 𝐸 𝐚𝑗1 𝐏𝑗1 − 𝐏𝑗0 + 𝐏𝑗0 𝐚𝑗1 − 𝐚𝑗0 + 𝐛1 − 𝐛0 𝑋
𝑗∈𝐽𝑚
Authors’ contributions (A.3)

KH, WS and TZ contributed to the design of the study, acquisition

of data, and supervision of data collection. JHK contributed to analysis
and interpretation of data in the manuscript. KH, WS, and JHK con-
tributed to manuscript preparation. All authors read and approved the
final manuscript.
Acknowledgments
We thank the editor and the two anonymous reviewers who pro-
vided many valuable comments. This paper benefited from discussions
with James Heckman, Stéphane Bonhomme and Alessandra Voena. Mi-
lan Wiedemann, Sylvi Kuperman, Tanya Rajan, Yu Kyung Koh, and Ling-
wei Wu provided helpful comments. We thank the Center for Economics
If measured and unmeasured inputs are independent in the outcome
equation without treatment, a0 can be estimated consistently. If they are
not independent, unmeasured inputs appear in the error term. The error
term of the regression equation will correlate with the regressors and ad
will be inconsistent. We then have plim𝐚̂ 1 ≠ plim𝐚̂ 0 even if 𝐚1 = 𝐚0 .
It is unobvious how to test this assumption. We nonetheless point out
that both the program and measurement system of parenting behaviors
were designed based on the same theoretical framework. Measurements
were chosen to capture parenting behaviors that were predicted to re-
spond to the program and affect childhood and adolescent outcomes.
We assume that parenting captured all inputs that correlate among one
another due to careful design of the study.

165
J.H. Kim et al. Labour Economics 54 (2018) 152–171

We test and do not reject hypotheses that plim𝐚̂ 1 ≠ plim𝐚̂ 0 . We reject Appendix B. Appendix tables and figures
plim̂𝐛1 ≠ plim̂𝐛0 for a few regressors in a few outcomes. The magnitudes
of these differences were very small.
With 𝐚1 = 𝐚0 , (A.3) can be written: Table B10
[ ] Model fit with control variables.
[ ] ∑ 𝑗( 𝑗 ) ( )
𝐸 𝑌1 − 𝑌0 = 𝜏1 − 𝜏0 + 𝐸 𝐚1 𝐏1 − 𝐏𝑗0 + 𝐛1 − 𝐛0 𝑋 (A.4) Pooled sample
𝑗∈𝐽𝑚
Variable name R2 R2 R2
(A.2) can be written: Negative behavior total score 0.102 0.025 0.028
∑ 𝑗 𝑗 Externalizing behavior 0.091 0.030 0.033
𝑌𝑑 = 𝜏𝑑 + 𝐚𝑑 𝐏𝑑 + 𝐛𝑑 𝑋 + 𝜖𝑑 , 𝑑 ∈ {1, 0} Internalizing behavior 0.175 0.082 0.014
𝑗∈𝐽𝑚 Quality of life total score 0.117 0.024 0.030
Wellbeing 0.147 0.059 0.027
and (A.1) can be written Social relationship 0.073 0.004 0.026
Demographic var. O O X
𝑌 = 𝐷𝑌1 + (1 − 𝐷)𝑌0 Baseline outcomes O X O
( ) ∑
= 𝜏0 + 𝜏1 − 𝜏0 𝐷 + 𝐚𝑗 𝐏𝑗 N 231 245 239
𝑗∈𝐽𝑚 Female sample
∑ ∑ ( ) R2 R2 R2
+ 𝐛𝑘 𝑋 𝑘 + 𝐛𝑘1 𝐷𝑋 𝑘 + 𝐛𝑘0 (1 − 𝐷)𝑋 𝑘 Variable name
Negative behavior total score 0.239 0.027 0.055
𝑘∈𝐾 ∖𝐾 ′ 𝑘∈𝐾 ′
Externalizing behavior 0.300 0.037 0.075
+𝜖 (A.5) Internalizing behavior 0.149 0.030 0.010
( ) Quality of life total score 0.154 0.020 0.038
where 𝐏𝑗 = 𝐷𝐏𝑗1
+ (1 − 𝐷)𝐏𝑗0 , observed parental input. 𝜏1 − 𝜏0 is contri- Wellbeing 0.160 0.052 0.034
∑ ( 𝑘 𝑘
)
bution from unmeasured parental input. 𝑘∈𝐾 ′ 𝐛1 𝐷𝑋 + 𝐛0 (1 − 𝐷)𝑋 𝑘
𝑘 Social relationship 0.159 0.030 0.046
Demographic var. O O X
accounts for linear parameters of Xk where 𝐛𝑘1 ≠ 𝐛𝑘0 . (A.5) can be es- Baseline outcomes O X O
timated directly using observed data, which can be used to construct N 111 115 113
(A.4). The final decomposition is:
Male sample
[ ] [ ] [ ] Variable name R2 R2 R2
𝐸 𝑌1 − 𝑌0 = 𝐚 𝐸 DIS𝐶1 − DIS𝐶0 +
1
𝐚 𝐸 SUP𝑃1 − SUP𝑃0
2
Negative behavior total score 0.089 0.031 0.015
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ Externalizing behavior 0.110 0.028 0.008
change in discipline style increase in positive engagement
( ) Internalizing behavior 0.127 0.031 0.029
+ 𝜏1 − 𝜏0 + 𝐛1 − 𝐛0 𝑋𝐾 ′ (A.6) Quality of life total score 0.102 0.018 0.031
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ Wellbeing 0.103 0.025 0.028
other factors Social relationship 0.071 0.011 0.024
Demographic var. O O X
Baseline outcomes O X O
N 120 130 126

166
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Fig. B6. Parenting style by mother, female sample Note: Confidence intervals were calculated based on wild cluster bootstrap of intervention effect esti-
mates with 99,999 replications. The unit is standard deviations within each follow-up sample. Measures were collected through parent interviews, detailed by
Heinrichs et al. (2014).

Table B11
Cross-correlation table of outcomes at 10-year follow-up.

Variables 1. 2. 3. 4. 5. 6.

1. (YSR) Total score 1.0000

2. (YSR) Externalizing behavior 0.7923 1.0000
(0.0000)
3. (YSR) Internalizing behavior 0.8420 0.4346 1.0000
(0.0000) (0.0000)
4. (KINDL) Total score -0.6467 -0.4168 -0.6286 1.0000
(0.0000) (0.0000) (0.0000)
5. (KINDL) Wellbeing -0.5847 -0.3557 -0.6257 0.9098 1.0000
(0.0000) (0.0000) (0.0000) (0.0000)
6. (KINDL) Social relationship -0.5731 -0.3886 -0.4952 0.8774 0.6015 1.0000
(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

Note: 𝑁 = 245. p-values are in the parentheses.

167
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Fig. B7. Parenting style by mother, male sample Note: Confidence intervals were calculated based on wild cluster bootstrap of intervention effect estimates
with 99,999 replications. The unit is standard deviations within each follow-up sample. Measures were collected through parent interviews, detailed by
Heinrichs et al. (2014).

Table B12
Internal Consistency of Survey Measures.

Name What does it measure Cronbach’s alpha Reported by

Parenting scale parent’s dysfunctional discipline 0.81 mother

Positive Parenting Questionnaire (PPQ) parent’s supportive behavior 0.85 mother
KINDL (physical) child’s physical wellbeing 0.79 child
KINDL (mental) child’s mental wellbeing 0.66 child
KINDL (self-esteem) child’s self-esteem 0.79 child
CBCL (externalizing behavior) child’s externalizing behavior (baseline) 0.9 mother
CBCL (internalizing behavior) child’s internalizing behavior (baseline) 0.86 mother
Youth Self Report (externalizing behavior) child’s externalizing behavior (10yr f.u.) 0.84 child
Youth Self Report (internalizing behavior) child’s internalizing behavior (10yr f.u.) 0.89 child
Youth Self Report (compulsive disorder) child’s compulsive disorder (10yr f.u.) 0.63 child

Note: Detailed discussions of each measure appear in Section 3.5.2.

168
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table B13
Parenting questionnaire (EFB).

Question Response scale: 1 through 7

Verbosity
2 Before I respond to a problem I warn or admonish my child several times
I warn or admonish it only once
4 If I forbid something to my child I’m talking about this very little
I’m talking about this very much
6 If my child is naughty or behaving inappropriately I get him into a long discussion about it
I let myself to no long discussions
9 If my child is naughty or behaving inappropriately I teach and urge him in detail
I tell him clearly and concisely what is important
11 If my child does not respond to ”no”, ... I speak with him, trying to convince him
I take some other action
27 If I had to do something, because my child has misbehaved, ... I’ll stick with what I have said, ignoring his protests
I stick to it and talk to my child about the fact that he is not to complain
Over-reactivity
3 If I’m excited or stressed, I am ... petty than usual
not petty than usual
5 If my child annoys me and gives me no peace I can ignore well
I can not ignore
10 If my child is naughty or behaving inappropriately I scream loudly to my child
I speak quietly with my child
14 After there is a problem with child I remain angry
Things return to normal quickly
17 When there is a problem with my child Things get out of control and I do things I wouldn’t want
Things don’t get out of control
18 If my child is naughty or behaving inappropriately, ... I beat/slap/hit him
I do not beat/slap/heat him
22 If my child is naughty or behaving inappropriately, ... I react on it without letting me get out of the socket
I am so angry and frustrated that my child knows it
23 If my child is naughty or behaving inappropriately, ... I do not use swear words
I use curse and swear words
26 If my child does something that I do not want ... I never say vulgar or hurtful things
I say vulgar or hurtful things
29 If my child does something that I do not want ... I do not insult my child
I insult my child
30 When my child misbehaves, I do not beat/slap/heat him
I beat/slap/hit him
31 When I punish my child ... I do not swear
I swear loudly
Laxness
1 When my child misbehaves, I do something right away
I do something later
7 I threaten with things where I’m sure I would do it
that I know that I will not do
8 I am a mother / father, who sets limits on what is allowed for my child
allows my child to do what he wants
12 When I want my child to stop doing something I tell him to stop in a resolute tone
I ask him to stop in a nice tone
15 If we are not at home, I treat my child as I would do at home
I let my child get away with more than at home
16 If my child does something I do not like ... I do something about it every time
I let it pass easily
19 If my child does not do what I asked it, ... I often let it go or do it myself
I take some other action
20 If I say a reasonable threat or warning ... I do not put them into practice
I put them into practice
21 If saying no doesn’t work... I promise my child a reward, so that he behaves
I take some other action
24 If I have my child forbidden something ... I let my child do it nevertheless
I stick to what I said
28 When my child is angry, because I have forbidden him something ... I take that ban back
I’ll stick to what I said
Other negative behavior
13 If my child is not in my area ... I do not know what he’s doing
I know what he’s doing
32 If my child is naughty or behaving inappropriately, ... so I threaten to punish him
I do not threaten him with a penalty
33 If my child has misbehaved, ... I talk quite normally with my child
I do not talk at length with my child
34 If my child misbehaves despite my warning, I threaten that I will leave him alone
I try not to make the threat
35 If my child is naughty or behaving inappropriately, ... I leave the room angry
I do not leave the room angry
36 If my child does not respond to a ban, I talk to him and try to persuade him
I take some other action
37 If I had to do something because my child misbehaved, I stick to what I said, ignoring his protests
I tell the child not to complain

169
J.H. Kim et al. Labour Economics 54 (2018) 152–171

Table B14
Mediation analysis p-values.

Sub-sample Outcome Discipline_P Positive_P Unmeasured_P

Pooled sample YSR_Total 0.122 0.168 0.355

Pooled sample YSR_EXT 0.059 0.399 0.121
Pooled sample YSR_INT 0.274 0.041 0.326
Pooled sample KIN_Total 0.127 0.167 0.059
Pooled sample KIN_WB 0.139 0.148 0.071
Pooled sample KIN_SK 0.209 0.218 0.284
Female sample YSR_Total 0.261 0.028 0.238
Female sample YSR_EXT 0.034 0.024 0.211
Female sample YSR_INT 0.462 0.014 0.438
Female sample KIN_Total 0.235 0.281 0.141
Female sample KIN_WB 0.236 0.270 0.054
Female sample KIN_SK 0.320 0.294 0.434
Male sample YSR_Total 0.113 0.339 0.370
Male sample YSR_EXT 0.229 0.398 0.206
Male sample YSR_INT 0.124 0.437 0.091
Male sample KIN_Total 0.133 0.132 0.412
Male sample KIN_WB 0.098 0.088 0.464
Male sample KIN_SK 0.183 0.326 0.444

Note: YSR_Total: youth self report total score. YSR_EXT: (YSR) Externalizing behavior. YSR_INT:
(YSR) Internalizing behavior. KIN_Total: KINDL Total Score. KIN_WB: (KINDL) wellbeing score.
KIN_SK: (KINDL) Social relationship score.

Table B15
Factor analysis of parenting measures that a mother received from mother’s own parents.

Factor Uniqueness

m_punish_v 0.7562696 0.4280563

m_warmth_v -.5977642 0.642678
m_control_v 0.4182435 0.8250723
m_punish_m 0.7424803 4,487,231
m_warmth_m -.622869 6,120,342
m_control_m 0.4184856 8,248,698

Note: v-subscript refers to parenting that a mother received from her own father and m-subscript
refers to parenting a mother received from her own mother.

170
J.H. Kim et al.
References
Achenbach, T.M., 1991. Manual for the child behavior checklist/4-18 and 1991 profile.
Department of Psychiatry, University of Vermont Burlington, VT.
Achenbach, T.M., 1991. Manual for the youth self-report and 1991 profile. Burlington:
University of Vermont, Department of Psychiatry.
Agostinelli, F., 2018. Investing in Children’s Skills: An Equilibrium Analysis of Social In-
teractions and Parental Investments. Unpublished manuscript
Akabayashi, H., 2006. An equilibrium model of child maltreatment. J. Econ. Dyn. Control
30 (6), 993–1025.
Arnold, D.S., O’Leary, S.G., Wolff, L.S., Acker, M.M., 1993. The parenting scale: a measure
of dysfunctional parenting in discipline situations.. Psychol. Assess. 5 (2), 137–144.
Baumrind, D., 1966. Effects of authoritative parental control on child behavior. Child Dev.
37 (4), 887–907.
Baumrind, D., 1968. Authoritarian vs. authoritative parental control. Adolescence 3 (11),
255.
Benjamin, D.J., Berger, J.O., Johannesson, M., Nosek, B.A., Wagenmakers, E.-J., Berk, R.,
Bollen, K.A., Brembs, B., Brown, L., Camerer, C., et al., 2018. Redefine statistical
significance. Nat. Hum. Behav. 2 (1), 6.
Biglan, A., 1995. Translating what we know about the context of antisocial behavior into
a lower prevalence of such behavior. J. Appl. Behav. Anal. 28 (4), 479–492.
Björklund, A., Lindahl, L., Lindquist, M.J., 2010. What more than parental income, educa-
tion and occupation? an exploration of what swedish siblings get from their parents.
B.E. J. Econ. Anal. Policy 10 (1).
Black, D., Joo, J., LaLonde, R., Smith, J. A., Taylor, E., 2017. Simple Tests for Selection:
Learning More from Instrumental Variables. CESifo Working Paper Series No. 6392.
Bussmann, K.-D., 2004. Evaluating the subtle impact of a ban on corporal punishment of
children in germany. Child Abuse Rev. 13 (5), 292–311.
Cameron, A.C., Miller, D.L., 2015. A practitioner’s guide to cluster-robust inference. J.
Human Resour. 50 (2), 317–372.
Campbell, F., Conti, G., Heckman, J.J., Moon, S.H., Pinto, R., Pungello, E., Pan, Y., 2014.
Early childhood investments substantially boost adult health. Science 343 (6178),
1478–1485.
Cosconati, M., 2009. Parenting Style and the Development of Human Capital in Children.
University of Pennsylvania Ph.D. thesis.
Cunha, F., Heckman, J., 2007. The technology of skill formation. Am. Econ. Rev. 97 (c),
31–47.
Davidson, R., MacKinnon, J.G., 2010. Wild bootstrap tests for IV regression. J. Bus. Econ.
Stat. 28 (1), 128–144.
De Los Reyes, A., Kazdin, A.E., 2005. Informant discrepancies in the assessment of child-
hood psychopathology: a critical review, theoretical framework, and recommenda-
tions for further study.. Psychol. Bull. 131 (4), 483.
Del Boca, D., 2015. Child care arrangements and labor supply. IDB Working Paper Series
No. 569.
Dishion, T.J., McMahon, R.J., 1998. Parental monitoring and the prevention of child and
adolescent problem behavior: a conceptual and empirical formulation. Clin. Child
Fam. Psychol. Rev. 1 (1), 61–75.
Dishion, T.J., Patterson, G.R., 2015. Developmental psychopathology, 2, 3, pp. 503–541.
Doepke, M., Zilibotti, F., 2017. Parenting with style: altruism and paternalism in intergen-
erational preference transmission. Econometrica 85 (5), 1331–1371.
Dooley, M., Stewart, J., 2007. Family income, parenting style and child behavioural-Emo-
tional outcomes. Health Econ. 16 (2), 145–162.
Doyle, O., Fitzpatrick, N., Lovett, J., Rawdon, C., 2015. Early intervention and child phys-
ical health: evidence from a dublin-based randomized controlled trial. Econ. Human
Biol. 19, 224–245.
Doyle, O., Harmon, C., Heckman, J.J., Logue, C., Moon, S., 2013. Measuring Investment in
Human Capital Formation: An Experimental Analysis of Early Life Outcomes. NBER
Working Paper No. 19316.
Doyle, O., Harmon, C., Heckman, J.J., Logue, C., Moon, S.H., 2016. Early skill forma-
tion and the efficiency of parental investment: a randomized controlled trial of home
visiting. Labour Econ..
Elango, S., García, J.L., Heckman, J.J., Hojman, A., 2015. Early childhood education.
NBER Working Paper No. 21766.
Fiorini, M., Keane, M.P., 2014. How the allocation of Children’s time affects cognitive and
non-Cognitive development. J. Labor Econ. 32 (4), 787–836.
García, J.L., Heckman, J.J., Ziff, A.L., 2017. Gender differences in the benefits of an influ-
ential early childhood program. NBER Working Paper No. 23412.
Granic, I., Patterson, G.R., 2006. Toward a comprehensive model of antisocial develop-
ment: A Dynamic systems approach. Psychol. Rev. 113 (1), 101.
Grych, J.H., Fincham, F.D., 1990. Marital conflict and children’s adjustment: a cognitive–
contextual framework.. Psychol. Bull. 108 (2), 267.
Hahlweg, K., Heinrichs, N., Kuschel, A., Bertram, H., Naumann, S., 2010. Long-term out-
come of a randomized controlled universal prevention trial through a positive parent-
ing program: is it worth the effort? Child Adolesc. Psychiatry Ment. Health 4, 14.
Hart, B., Risley, T.R., 1975. Incidental teaching of language in the preschool. J. Appl.
Behav. Anal. 8 (4), 411–420.
Hart, B., Risley, T.R., 1995. Meaningful Differences in the Everyday Experience of Young
American Children.. Paul H Brookes Publishing.
Havnes, T., Mogstad, M., 2011. No child left behind: subsidized child care and children’s
long-run outcomes. Am. Econ. J. 3 (2), 97–129.
171
Labour Economics 54 (2018) 152–171
Heckman, J.J., Pinto, R., 2015. Econometric mediation analyses: identifying the sources
of treatment effects from experimentally estimated production technologies with un-
measured and mismeasured inputs. Econ. Rev. 34 (1–2), 6–31.
Heckman, J.J., Pinto, R., Savelyev, P.A., 2013. Understanding the mechanisms through
which an influential early childhood program boosted adult outcomes. Am. Econ.
Rev. 103 (6), 2052–2086.
Heinrichs, N., Bertram, H., Kuschel, A., Hahlweg, K., 2005. Parent recruitment and re-
tention in a universal prevention program for child behavior and emotional prob-
lems: barriers to research and program participation. Prevention Sci. 6 (4), 275–286.
doi:10.1007/s11121-005-0006-1.
Heinrichs, N., Hahlweg, K., Bertram, H., Kuschel, A., Naumann, S., Harstick, S., 2006. Die
langfristige wirksamkeit eines elterntrainings zur universellen prävention kindlicher
verhaltensstörungen: ergebnisse aus sicht der mütter und väter [the long-term effec-
tiveness of a parenting program for universal prevention of child behavior problem.
Zeitschrift für Klinische Psychologie und Psychotherapie 35, 82–96.
Heinrichs, N., Hahlweg, K., Naumann, S., Kuschel, A., Bertram, H., Ständer, D., 2009.
Universelle prävention kindlicher verhaltensstörungen mit hilfe einer elternzentri-
erten maßnahme: ergebnisse drei jahre nach teilnahme [Universal prevention of child
behavior problems with a parent training: Results three years after participation].
Zeitschrift für Klinische Psychologie und Psychotherapie 38, 79–88.
Heinrichs, N., Jensen-Doss, A., 2010. The effects of incentives on families’ long-term out-
come in a parenting program. J. Clin. Child Adolescent Psychol. 39 (5), 705–712.
Heinrichs, N., Kliem, S., Hahlweg, K., 2014. Four-year follow-up of a randomized con-
trolled trial of triple p group for parent and child outcomes. Prevention Sci. 15 (2),
233–245.
Johnson, Z., Molloy, B., Scallan, E., Fitzpatrick, P., Rooney, B., Keegan, T., Byrne, P.,
2000. Community mothers programme-seven year follow-up of a randomized con-
trolled trial of non-professional intervention in parenting. J. Public Health 22 (3),
337–342.
Kaufman, A.S., Kaufman, N.L., 1983. Kaufman assessment battery for children. Am. Guid-
ance Serv..
Lee, D., Brooks-Gunn, J., McLanahan, Sara S., Notterman, D., Hahlweg, K., Garfinkel, I.,
2013. The Great Recession, genetic sensitivity, and maternal harsh parenting. Proc
Natl Acad Sci U S A 110 (34), 13780–13784. doi:10.1073/pnas.1312398110.
Long, N., Edwards, M.C., Bellando, J., 2017. Parent training interventions. In: Handbook
of Childhood Psychopathology and Developmental Disabilities Treatment. Springer,
pp. 63–86.
Maccoby, E.E., Martin, J.A., 1983. Socialization in the context of the family: Paren-
t-child interaction. In: Mussen, P.H.S.e., Heatherington, E.M.V.e. (Eds.), Handbook of
Child Psychology: Vol. 4. Socialization, Personality, and Social Development. Wiley,
pp. 1–101.
Naumann, S., Bertram, H., Kuschel, A., Heinrichs, N., Hahlweg, K., Döpfner, M., 2010. Der
erziehungsfragebogen (efb). Diagnostica.
Nowak, C., Heinrichs, N., 2008. A comprehensive meta-analysis of triple p-positive par-
enting program using hierarchical linear modeling: effectiveness and moderating vari-
ables. Clin. Child Fam. Psychol. Rev. 11 (3), 114.
Patterson, G.R., 1982. Coercive Family Process, 3. Castalia.
Patterson, G.R., 1986. Performance models for antisocial boys.. Am. Psychol. 41 (4), 432.
Patterson, G.R., Dishion, T.J., 1985. Contributions of families and peers to delinquency.
Criminology 23 (1), 63–79.
Patterson, G.R., Dishion, T.J., Bank, L., 1984. Family interaction: a process model of de-
viancy training. Aggress. Behav. 10 (3), 253–267.
Patterson, G.R., Reid, J.B., Dishion, T.J., 1992. Antisocial Boys, 4. Castalia.
Ravens-Sieberer, U., Bullinger, M., 1998. Assessing health-related quality of life in chron-
ically ill children with the german kindl: first psychometric and content analytical
results. Qual. Life Res. 7 (5), 399–407.
Sanders, M.R., 2012. Development, evaluation, and multinational dissemination of the
triple P-Positive parenting program. Annu. Rev. Clin. Psychol. 8 (1), 345–379.
Sanders, M.R., Kirby, J.N., Tellegen, C.L., Day, J.J., 2014. The triple P-Positive parenting
program: a systematic review and meta-analysis of a multi-level system of parenting
support. Clin. Psychol. Rev. 34 (4), 337–357.
Schneider, S., Unnewehr, S., Margraf, J., 2009. Diagnostisches interview psychischer
störungen im kindes-und jugendalter (k-dips).
Shaffer, A., Kotchick, B.A., Dorsey, S., Forehand, R., 2001. The past, present, and future of
behavioral parent training: interventions for child and adolescent problem behavior..
Behav. Anal. Today 2 (2), 91.
Spera, C., 2005. A review of the relationship among parenting practices, parenting styles,
and adolescent school achievement. Educ. Psychol. Rev. 17 (2), 125–146.
Stanger, C., Lewis, M., 1993. Agreement among parents, teachers, and children on inter-
nalizing and externalizing behavior problems. J. Clin. Child Psychol. 22 (1), 107–116.
Strayhorn, J.M., Weidman, C.S., 1988. A parent practices scale and its relation to parent
and child mental health. J. Am. Acad. Child Adolesc. Psychiatry 27 (5), 613–618.
Verhulst, F.C., Ende, J., 1992. Agreement between parents’ reports and adolescents’ self-
-reports of problem behavior. J. Child Psychol. Psychiatry 33 (6), 1011–1023.
Weinberg, B.a., 2001. An incentive model of the effect of parental income on children.
Journal of Political Economy 109 (2), 266–280. doi:10.1086/319556.
Wilson, P., Rush, R., Hussey, S., Puckering, C., Sim, F., Allely, C.S., Doku, P., Mc-
Connachie, A., Gillberg, C., 2012. How evidence-based is an ‘evidence-based parenting
program’? a PRISMA systematic review and meta-analysis of triple p. BMC Med. 10
(1), 130.