J Fam Econ Iss
DOI 10.1007/s10834-014-9435-y
ORIGINAL PAPER
Disruption in Parental Co-habitation and its Effects
on Short-Term, Medium-Term, and Long-Term Outcomes
of Adolescents
Andrew Hussey • Debjani Kanjilal
Anil Nathan
•
Ó Springer Science+Business Media New York 2014
Abstract In this paper the relationships between a disruption in parental co-habitation and various categories of
adolescent outcomes over multiple time horizons are
explored. Using data from the National Longitudinal Study
of Adolescent Health (Add Health), we estimated the
effects of a change from living with both parents to just
one, on academic and employment outcomes, the likelihood to indulge in risky behaviors, mental health outcomes
and body mass index measures, from less than 1 year to
over 14 years after the change. Propensity score matching
methods were used to control for individual characteristics
and pre-existing differences in the family environment that
may increase the chances of separation, and the results are
compared to those obtained using ordinary least squares or
probit methods. Results showed evidence of adverse effects
of living with one parent in the short term, medium term
and long term. Adolescents living with one parent had
lower academic achievement in all term lengths, poor
mental health in the short to medium term, and were more
likely to engage in risky behaviors in the medium to long
term.
A. Hussey
Department of Economics, University of Memphis,
423 Fogelman Administration Building, Memphis,
TN 38152, USA
D. Kanjilal (&)
Department of Business & Economics, Elizabeth City State
University, Campus Box 781, 1704 Weeksville Rd,
Elizabeth City, NC 27909, USA
e-mail:
[email protected]
A. Nathan
Department of Economics and Accounting, College of the Holy
Cross, 1 College St, Worcester, MA 01610, USA
Keywords Disruption in parental co-habitation
Propensity-score matching Add health
JEL Classification
I21 J11 J12
Introduction
Living in a single-parent family is largely perceived to have
negative effects on children’s well-being, and several earlier
academic studies have confirmed this perception (Amato and
Booth 1997; Keith and Finlay 1988; Manski et al. 1992;
Mayer 1997). Be it due to divorce, separation, or the death of
a parent, the transition from a dual parent household to a
single parent household may be accompanied by children’s
decreased academic performance or attainment, and may
alter the eventual marital and divorce probabilities of the
children (Cid and Stokes 2013; Corak 2001; Ermisch and
Francesconi 2001; Frisco et al. 2007; Keith and Finlay 1988).
Moreover, children from these families may have increased
likelihood of engaging in potentially risky behaviors, such as
drinking, smoking, drug use, or sex, and may suffer from
significant stress, depression, or other psychological problems (Amato 2000; Amato and Keith 1991; Brown 2006;
Dronkers 1999; Schramm 2005).
Adverse effects may arise through a number of channels.
For example, single parents might invest less time or
income in their children, or provide improper guidance or
role modeling (Downey 1994; Entwisle and Alexander
1996; McLanahan and Sandefur 1994; Myers and Myers
2014). Living in a single-parent family can also reduce
economic and social resources available to children (Bane
and Jargowsky 1988; McLanahan and Sandefur 1994).
Disruption of cohabitation can also have an effect on parents’ stress and health (Kohn and Averett 2014), which
123
J Fam Econ Iss
may in turn affect parenting practices. On the other hand, if
in an intact family there is friction among parents, then a
divorce may not adversely affect children’s outcomes, or
might even prove to be good for the child (Brown 2006;
Dronkers 1999). Indeed, some studies have found no causal
effect of living in a single-parent family on children’s
outcomes (Entwisle and Alexander 1996; Lang and
Zagorsky 2001; Marsh 1990; Sanz-De-Galdeano and Vuri
2007).
Thus, despite a large and growing body of research in the
area, the literature investigating the effects of parental
divorce or separation on children’s outcomes remains
inconclusive, or varied in its findings. This is likely due to a
few reasons. First, the outcomes investigated have often been
narrow and varied across studies. Second, many studies used
local or non-nationally representative data. Third, studies
have differed in their time frame of analysis or age of the
children when parental separation occurs.1 Finally, since
single-parent families are likely to be different from twoparent families, and in ways that may also independently
affect children’s outcomes, the effect of living with a singleparent is likely to be confounded with these factors. Correspondingly, studies in the literature have differed methodologically, in particular with regard to how they handle this
non-random assignment of changes in family structure.
In this paper, we analyzed the impact of an adolescent’s
transition from living with two parents to living with one
parent on several key academic, behavioral and healthrelated outcomes often associated with adolescents’ future
wellbeing. The academic and economic outcomes we
considered include GPA, high school graduation, college
attendance, college degree, employment and income. The
behavior related outcomes included in our study are the
propensities to engage in the potentially risky behaviors of
drinking, smoking, drug use and sex. Finally, we also
considered the health-related outcomes of BMI and mental
health indicators. Since it has been suggested that time
since the disruption of parental co-habitation is an important determinant of its effects on children (Chase-Lansdale
et al.1995; Cherlin et al. 1995), we separated our analysis
into short term (one year or less after parental separation),
medium term (about 7 years after) and longer term (about
14 years after) effects on adolescent outcomes.
We used data from the National Longitudinal Survey of
Adolescent Health (Add Health), a large national survey of
adolescents, to explore these relationships. The richness of
the data allowed for the inclusion of detailed control
variables not available to many prior researchers. Because
1
For example, Cid and Stokes (2011) looked at the role of parental
separation on school drop-out behavior, while Couch and Lillard
(1997) investigated much longer-term intergenerational correlation in
adult earnings as a function of parental divorce.
123
we wished to isolate causal effects of living with one parent
on children’s outcomes, we required changes in parental
co-habitation over time. Thus, we relied on reports in each
survey wave of whether or not adolescents were living with
both of their parents. An individual that reported living
with only one parent after initially reporting living with
both served as an indicator of parental separation. We thus
looked at changes in adolescent behavior due to a change in
co-habitation.2 .
The main estimation method used in this paper was the
propensity score matching technique, which is widely used
in statistics and medicine. Although this method is
becoming more common in economic applications
involving treatment effects, its use has been minimal in the
literature on the effects of parental divorce and separation.3
The method assists in creating an appropriate control group
for the treatment group of interest. In our case, we compared two groups who were otherwise observably similar,
but one group receives a treatment (living with single
parent) and the other group does not (remains living with
both parents). Under certain assumptions, the difference in
the outcomes of the two groups can be attributed solely to
the treatment. In addition to multiple methods of matching
on the basis of propensity score, we compared our findings
to those using ordinary least squares (OLS) and probit
regressions. A similar methodology was used by Frisco
et al. (2007) in their estimation of the influence of parents’
union dissolution on changes in adolescents’ mathematics
course work gains, overall GPA, and course failure rates.
They found that academic achievement of children
declined as a result of parental union dissolution. However,
their study was limited to only the short-term academic
performances of children. Our work can be seen as an
extension of their analysis, with our addition of several
categories of outcomes and multiple observations at different time periods following the dissolution of parental cohabitation.
Our results suggest that adolescents who experienced
parental separation had lower GPAs, poorer mental health,
and increased participation in risky activities in the shortterm. They also had a lower likelihood of graduating from
high school on time, lower likelihood of obtaining a college degree, and were more likely to smoke in the medium
and long-terms. These results imply the advocacy of
2
Tumin et al. (2014) found that the vast majority of all marital
separations end in divorce. Further, they found that separations
leading to divorce are relatively short —8 to 9 months. Using samples
from the National Longitudinal Survey of Youth, 1979, and the
National Surveys of Family Growth, they found that between 2 and
4 % of marital disruptions among women are caused by widowhood.
3
For example, see Basu et al. (2008), Dehejia and Wahba (2002),
Heckman et al. (1998), Imbens (2000) and Rosenbaum and Rubin
(1983).
J Fam Econ Iss
programs to either keep parents together or to support
families that are disrupted.
Empirical Methods
In analyzing the effects of change in parental co-habitation
on adolescents’ outcomes, we considered two groups, one
whose parents separated between the first two survey
waves, Waves I and II (the treatment group), and the other
whose parents did not (the control group). The main difficulty of attempting to determine causal relationships in
this context is that adolescents whose parents separate are
likely to differ from those whose families remain intact,
and in ways that are correlated with subsequent academic,
behavioral and health outcomes that can be observed in the
data. Our fundamental solution to this problem was to
follow that which is used by much of the literature, which
was to control for selection on observables. In this case, we
used the richness of the Add Health data to control for a
diverse array of initial characteristics of the individual,
family, school, and community. We then effectively created a counterfactual for individuals in the treatment group
using individuals from the control group who were most
similar in terms of these covariates. Specifically, each
observation in the treatment group was matched with one
or more observations in the control group. Under certain
assumptions, the average difference in outcomes can then
be attributed to the disruption in parental co-habitation
(Rosenbaum and Rubin 1983).
More concretely, we can define:
Y1: outcome of an adolescent living with one parent
(exposed to the treatment)
Y0: outcome of an adolescent living with both parents
(not exposed to the treatment)
D: indicator of a change in parental co-habitation (the
treatment)
X: set of covariates.
Matching requires the assumption that all relevant differences between the two groups will be captured by the set
of covariates. That is, ðY0 ; Y1 Þ?DjXj:
Children living with one parent were matched with a
control group of children living with both parents, with
whom the distribution of the covariates is as close as
possible to the group with the treatment. Propensity score
matching provides a natural method for weighting each of
the covariates, thus avoiding the problem of finding an
exact match for the treatment group. While finding an exact
match would severely limit the number of possible covariates to be matched on, propensity score allows matching
on a large number of covariates by collapsing the relevant
information into a single index, or ‘‘propensity score.’’ The
propensity score (PS) is defined as the probability of
receiving the treatment conditional on the set of covariates.
Thus, PSðXÞ ¼ PðD ¼ 1jXÞ:.
In practice, we estimated the propensity score using a
logit model including a large number of covariates. Using
the obtained estimates, for each individual we predicted the
probability of undergoing a parental separation (conditional on the observables), regardless of whether or not
individuals actually experienced this disruption in parental
co-habitation. We used only pre-disruption (Wave I) variables to predict the likelihood of disruption, since later
variables may themselves be affected by this change in cohabitation. Individuals in the treatment group were then
matched with individuals in the control group on the basis
of this estimated propensity score.
The first assumption thus became: ðY 0 ; Y1 ?Dj
PSðXÞ:This assumption merely states that parental separation effectively happens randomly to adolescents, conditional on the rich set of covariates as embodied in the
propensity score. Another necessary assumption is that the
propensity score was confined to be between zero and one,
which necessarily resulted from our logit specification.Under these assumptions, we estimated, in the words
of the treatment effect literature, the average treatment
effect on the treated (ATT), corresponding to the average
effect of living with one parent on adolescents in such an
environment:
(
)
X
1 X
ATT ¼
Yi
wði; jÞYj
ND i2D
J2C
i
where:
Yi: outcome of an adolescent living with one parent
Yj: outcome of an adolescent living with both parents
ND: number of adolescents living with one parent
Cj: set of matched control adolescents
w (i, j): the weight function
There are several possible different weighting methods
(Imbens 2000). These include one-to-one nearest neighbor
matching, k-nearest neighbors matching (k [ 1), kernel
density matching and local linear regression matching. We
used the simple average k-nearest neighbors matching
method, specifically using two nearest neighbors. This
approach created the counterfactual outcome as a simple
average over the outcomes of the nearest neighbors:
^ 0i jPðXi Þ; Di ¼ 0Þ ¼ 1
EðY
2
2
X
Y0j
j¼1
fDj2A2
We also used a kernel density matching method, where all
adolescents in the counterfactual group (living with both
parents) are considered for matching and the closer
observations (in terms of their propensity score) are given
more weight and the further observations are given lower
123
J Fam Econ Iss
Table 1 Descriptive statistics of treatment and control groups
(i)
Treatment Sample
Mean (Std. dev.)
(ii)
Control (unmatched)
Mean (Std. dev.)
(iii)
Control (matched)
Mean
Difference: (i) and (ii)
p value
Difference: (i) and (iii)
p value
15.708
15.447
15.742
0.000
0.709
(1.43)
(1.412)
Males
0.418
(0.493)
0.457
(0.498)
0.446
0.113
0.403
Black
0.196
0.138
0.197
0.001
0.967
(0.397)
(0.345)
0.484
0.992
0.876
0.106
0.694
0.286
0.075
0.535
0.507
1.356
0.000
0.887
5.141
0.001
0.668
Age
American Indian
0.046
0.046
(0.021)
(0.210)
Asian
0.085
0.080
(0.280)
(0.272)
Others
0.088
0.079
(0.283)
(0.271)
1.367
1.756
(1.190)
(1.342)
No. of siblings
Mother’s education
5.209
5.617
(2.388)
(2.402)
Mother works full time
0.757
0.756
0.761
0.962
0.907
Father works full time
(0.428)
0.907
(0.429)
0.934
0.914
0.030
0.727
(0.290)
(0.247)
4.476
0.002
0.532
3.894
0.000
0.955
1.697
0.000
0.616
0.160
0.037
0.721
1.320
0.010
0.755
16.556
0.364
0.695
Closeness to mother
Closeness to father
Skipped school
Grade repeated
Attention problem
TV hours
4.442
4.558
(0.845)
(0.738)
3.888
4.305
(1.190)
(0.922)
1.933
1.075
(7.352)
(4.703)
0.169
0.134
(0.375)
(0.341)
1.343
1.212
(1.133)
(1.025)
16.147
15.500
(14.953)
(14.483)
Play computer games
2.691
2.716
3.028
0.935
0.513
Hang with friends
(6.108)
2.013
(6.355)
1.966
1.991
0.331
0.744
(1.022)
(0.979)
0.490
0.112
0.370
0.615
0.000
0.822
0.349
0.090
0.978
0.642
0.000
0.476
0.084
0.093
0.421
0.331
0.094
0.301
Play sports
Felt depressed
Felt fearful
Felt sad
Private school
Urban
123
0.46
0.499
(0.498)
(0.500)
0.627
0.447
(0.805)
(0.699)
0.348
0.302
(0.588)
(0.545)
0.676
0.523
(0.723)
(0.644)
0.07
0.094
(0.256)
(0.292)
0.299
0.263
J Fam Econ Iss
Table 1 continued
(i)
Treatment Sample
Mean (Std. dev.)
(ii)
Control (unmatched)
Mean (Std. dev.)
(iii)
Control (matched)
Mean
Difference: (i) and (ii)
p value
Difference: (i) and (iii)
p value
0.208
0.529
0.838
0.245
0.603
0.969
0.396
0.036
0.839
0.068
0.001
0.381
(0.458)
(0.440)
0.202
0.190
(0.402)
(0.392)
0.246
0.235
(0.431)
(0.424)
0.389
0.340
(0.488)
(0.474)
North east
0.083
0.138
(0.277)
(0.345)
Caring neighbor
0.707
(0.455)
0.761
(0.426)
0.709
0.010
0.942
0.900
0.913
0.889
0.362
0.588
(0.299)
(0.281)
3.940
0.089
0.786
2.774
0.000
0.240
101.19
0.000
0.242
0.227
0.000
0.969
0.430
0.001
0.815
Rural
West
South
Feel safe in neighborhood
Happy in neighborhood
GPA
PVT scores
3.958
4.038
(0.974)
(0.955)
2.708
2.888
(0.853)
(0.825)
100.061
102.700
(14.279)
(14.165)
0.229
0.147
(0.420)
(0.355)
0.422
0.347
(0.494)
(0.476)
Drugs
0.244
0.202
0.242
0.033
0.938
Sex
(0.430)
0.385
(0.401)
0.237
0.381
0.000
0.892
(0.487)
(0.425)
454
4961
Smoke
Alcohol
N
All variables obtained from Wave I of Add Health. Sample restricted to households with cohabiting parents in Wave I. Treatment group includes
individuals who transitioned from living with both parents to living with one parent between Waves I and II. Control group includes individuals
for whom parental co-habitation remained intact between Waves I and II
weight. Specifically, we use a biweight (quartic) distribution kernel smoothing function to match treated observations with the untreated sample, using a bandwidth of
0.6.
For purposes of comparison of results, in addition to
using these matching estimators, we initially carried out
ordinary least squares and probit regressions. We ran OLS
regressions for all outcomes that are continuous and probit
regressions for all binary outcomes, where the marginal
effects were reported. However the advantages of using
matching estimators over OLS and probit regressions in the
case of a binary treatment where there is selection into the
treatment were twofold. First, any misspecification problems using OLS and probit were diminished because the
functional forms in constructing counterfactuals are
flexible. Secondly, the distribution of propensity scores for
the treatment and control group needed to have a common
support. This allowed for treatment and control groups to
be essentially the same in all characteristics except for
treatment itself. The comparison between the treatment and
control groups was more direct than with OLS or probit, as
it replicated an experimental design.
Data
Add Health
The data used in this paper came from the National Longitudinal Survey of Adolescent Health (Add Health) that
123
J Fam Econ Iss
Table 2 Wave II ordinary least squares and probit marginal effects estimates
Live with one parent
Age
Males
Black
No. of siblings
GPA
Smoke
Alcohol
Drugs
Sex
BMI
MHI
Depressed
-0.134***
0.032
-0.022
0.038
0.052
0.039
1.12**
0.032*
(0.008)
(0.022)
(0.030)
(0.025)
(0.031)
(0.292)
(0.385)
(0.013)
0.008
-0.011**
0.034***
-0.010*
0.097***
-0.574***
0.146*
0.003
(0.002)
(0.009)
(0.004)
(0.005)
(0.004)
(0.005)
(0.052)
(0.069)
-0.257***
-0.021*
-0.022
0.005
-0.039**
0.586***
-1.349***
-0.013*
(0.023)
(0.010)
(0.015)
(0.011)
(0.015)
(0.143)
(0.191)
(0.006)
-0.177***
-0.111***
-0.148***
-0.013
0.199***
0.654***
-0.122
-0.001
(0.038)
(0.008)
(0.019)
(0.017)
(0.024)
(0.212)
(0.284)
(0.008)
-0.011
-0.004
0.012*
-0.003
-0.018***
-0.092
0.161*
0.000
(0.009)
(0.004)
(0.005)
(0.004)
(0.005)
(0.052)
(0.068)
(0.002)
Mother’s education
0.024***
(0.006)
-0.001
(0.002)
0.006
(0.003)
0.006*
(0.003)
-0.013***
(0.003)
-0.076*
(0.031)
-0.098*
(0.042)
-0.001
(0.001)
Mother works full time
-0.019
0.002
0.016
-0.001
0.007
-0.019
-0.291
-0.005
(0.028)
(0.011)
(0.016)
(0.013)
(0.016)
(0.158)
(0.212)
(0.007)
Father works full time
Closeness to mother
Closeness to father
Skipped school
Grade repeated
Attention problem
TV hours
Play computer games
Hang with friends
Play sports
-0.027
0.001
0.017
0.005
-0.015
-0.591*
-0.712*
-0.012
(0.048)
(0.019)
(0.028)
(0.021)
(0.028)
(0.272)
(0.358)
(0.011)
-0.002
-0.007
-0.016
-0.009
-0.008
0.115
-0.377**
-0.006
(0.020)
(0.007)
(0.011)
(0.008)
(0.011)
(0.108)
(0.145)
(0.004)
-0.005
-0.008
-0.027**
-0.020**
-0.020*
0.316***
-0.861***
-0.009**
(0.017)
(0.006)
(0.010)
(0.007)
(0.009)
(0.091)
(0.121)
(0.003)
-0.008
0.001
-0.000
0.002*
0.004**
0.041**
0.055**
0.001*
(0.004)
(0.001)
(0.001)
(0.001)
(0.002)
(0.016)
(0.020)
(0.000)
-0.198***
0.033*
-0.032
0.012
0.020
-0.0009
0.296
0.003
(0.040)
(0.016)
(0.021)
(0.017)
(0.021)
(0.208)
(0.282)
(0.008)
-0.069***
0.005
0.032***
0.021***
-0.006
-0.109
0.931***
0.009***
(0.013)
(0.005)
(0.007)
(0.006)
(0.007)
(0.071)
(0.094)
(0.002)
-0.002*
(0.001)
0.000
(0.000)
0.000
(0.001)
-0.000
(0.000)
-0.001*
(0.000)
0.019***
(0.005)
0.001
(0.007)
0.000
(0.000)
-0.001
-0.001
-0.000
0.001
0.002*
0.005
0.003
-0.001
(0.002)
(0.001)
(0.001)
(0.001)
(0.001)
(0.012)
(0.015)
(0.000)
--0.004
0.026***
0.049***
0.033***
0.025***
-0.158*
-0.112
-0.000
(0.012)
(0.005)
(0.007)
(0.006)
(0.007)
(0.069)
(0.093)
(0.003)
0.077***
-0.025*
0.044**
0.021
0.048***
-0.137
-0.149
-0.004
(0.024)
(0.010)
(0.014)
(0.011)
(0.014)
(0.137)
(0.185)
(0.006)
Felt depressed
0.111
0.008
-0.055
0.031
0.115**
-0.252
(0.070)
(0.025)
(0.036)
(0.031)
(0.042)
(0.375)
Felt fearful
0.037
-0.042**
-0.068**
-0.028
-0.014
0.050
0.555
0.011
(0.047)
(0.013)
(0.025)
(0.019)
(0.025)
(0.249)
(0.336)
(0.010)
Private school
0.100*
-0.029*
-0.049*
0.014
-0.053*
-0.040
-0.056
-0.007
(0.043)
(0.016)
(0.024)
(0.020)
(0.024)
(0.233)
(0.319)
(0.009)
-0.043***
-0.006
-0.031***
-0.042***
-0.340***
-0.552***
-0.007*
0.09*
(0.006)
-0.000
(0.010)
0.003***
(0.007)
0.001
(0.010)
-0.002**
(0.092)
0.011*
(0.124)
-0.060***
(0.004)
-0.001*
(0.000)
(0.000)
(0.000)
GPA
PVT scores
Smoke
Alcohol
123
(0.001)
(0.000)
(0.000)
(0.005)
(0.007)
-0.209***
0.104***
0.138***
0.181***
-0.392
0.513
0.002
(0.042)
(0.024)
(0.020)
(0.025)
(0.224)
(0.303)
(0.008)
-0.054
0.106***
0.212***
0.192***
0.130
0.579**
-0.000
(0.030)
(0.013)
(0.014)
(0.017)
(0.116)
(0.225)
(0.007)
J Fam Econ Iss
Table 2 continued
GPA
Drugs
Smoke
Alcohol
Drugs
Sex
BMI
MHI
Depressed
-0.163***
0.134***
0.223***
0.172***
-0.266
0.113
0.005
(0.038)
(0.017)
(0.021)
(0.021)
(0.202)
(0.275)
(0.008)
Sex
0.033
(0.027)
0.081***
(0.015)
0.068***
(0.020)
0.077***
(0.015)
0.123
(0.182)
0.223
(0.246)
0.017
(0.08)
N
3952
4663
5288
5262
4112
5285
5285
5273
Standard errors shown in parentheses. OLS regression was performed in the cases of GPA, BMI and mental health (depressed and sad). For all
other dependent variables, probit regressions were performed. Only coefficients with statistical significance in at least one regression are included
in the table. Also included in each regression were controls for racial background, regional school dummies, and neighborhood variables
Significance levels: p \ 0.10. * p \ 0.05. ** p \ 0.01. *** p \ 0.001
was conducted by the Carolina Population Center of the
University of North Carolina at Chapel Hill. Conducted in
four waves, Add Health is a longitudinal survey of a
nationally representative sample of young adults who were
in grades 7–12 during 1994–1995. A series of in-home
interviews were conducted in 1994–95 for Wave I and
were followed up in 1996 for Wave II and 2001–2002 for
Wave III. Wave IV data collection was completed in 2008
with the full set of variables recently released.
Add Health is one of the largest longitudinal surveys of
adolescents and has a wide range of information about their
personal characteristics, health, risky behaviors, daily
activities, families, friends, romantic partners, peer groups,
schools, neighborhoods, and communities. Because of its
comprehensiveness, Add Health is now being used by
many researchers in a variety of social and behavioral
science disciplines for analyses of adolescents’ academic,
social, physical, psychological and economic phenomena.
The richness of the data and its longitudinal nature allowed
us to have a relatively complete and dynamic look at the
lives of these children as they transition into adulthood.
The data also allowed us to consider a wide variety of
individual outcomes, and to include a relatively large
number of control variables in our analysis.
Dependent Variables
We considered academic variables (high school graduation,
grade point average, standardized test scores, college
attendance, and college degree), employment variables
(whether employed in Wave III and personal incomes in
Waves III and IV), behavioral variables (smoking, drinking, drug use and sex), and health outcomes (BMI in
Waves II and III and obesity in Wave IV, and mental health
in Waves II and III) as outcomes potentially affected by the
change in parental co-habitation. An indicator variable for
high school graduation was created primarily from high
school transcript information linked to the survey data. To
maximize sample size, we supplemented transcript information with students’ self-reports of their graduation
status. We considered only on-time high school graduation
(4 years or less) for two reasons. First, this allowed us to
avoid the inclusion of GEDs, which may or may not be
comparable to a regular high school diploma. More
importantly, it allowed us to include the youngest cohort
surveyed, initially in 7th grade, which, by the time of the
third survey wave, was old enough to have graduated from
high school on time.
Overall high school grade point average was also
obtained from student transcripts. In consideration of shortterm effects on grades, we followed Sabia (2007) in calculating Wave II grade point average. We used the average
response of the student’s self-reported grades in both the
most recent math class and the most recent English/language arts class. Giving equal weight to math and science
classes, a 4 was assigned for each reported ‘‘A’’ grade, 3
for each ‘‘B’’ grade, 2 for each ‘‘C’’ grade and 0.5 for a
‘‘D’’ or lower grade. Wave I GPA was constructed in a
similar fashion for use as a control variable for initial
academic inclination. For Wave IV academic outcomes, we
considered two reports, namely if individuals ever attended
college and if individuals have a college degree conditional
on having attended college.
A unique feature of the Add Health data is that each
respondent was administered the ‘‘Add Health Picture
Vocabulary Test’’ (PVT), an abbreviated version of the
Peabody Picture Vocabulary Test—Revised. We included
the individuals’ scores on this test in Wave III as an
alternative, standardized measure of academic/intellectual
development.
In both Waves III and IV, respondents were asked to
report their annual personal income before taxes in
2001/2002 and 2006/2007 income years, respectively. In
addition, in Wave III they were also asked about their
employment status at that point of time. Information from
these responses provided the employment variables we
include in our analysis.
The Add Health data contains detailed individual
behavioral information. Based on student self-reports, we
created dummy variables corresponding to whether the
123
J Fam Econ Iss
Table 3 Wave III ordinary least squares and probit marginal effects estimates of academic and employment outcomes
GPA
Live with one parent
Age
Males
Black
No. of siblings
Mother’s education
Mother works full time
Father works full time
Closeness to mother
Closeness to father
Skipped school
Grade repeated
Attention problem
PVT scores
Graduated on time
Employed
Log income
-0.141**
-0.111
-0.056***
0.005
-0.240
(0.046)
(0.451)
(0.011)
(0.015)
(0.144)
0.054***
0.718***
0.030***
0.041***
0.295***
(0.008)
(0.114)
(0.003)
(0.004)
(0.042)
-0.294***
1.481***
-0.026***
0.031*
0.637***
(0.021)
(0.374)
(0.008)
(0.012)
(0.118)
-0.292***
-7.014***
0.022*
-0.107***
-0.240
(0.032)
(0.549)
(0.010)
(0.020)
(0.174)
-0.001
-0.783***
-0.004
0.000
0.004
(0.008)
(0.132)
(0.003)
(0.004)
(0.043)
0.043***
(0.005)
0.984***
(0.080)
0.009***
(0.002)
-0.017***
(0.003)
-0.060*
(0.026)
0.007
0.703
0.022*
0.030*
0.067
(0.024)
(0.416)
(0.009)
(0.014)
(0.132)
0.060
-0.051
0.031*
0.073**
-0.024
(0.041)
(0.703)
(0.016)
(0.026)
(0.227)
-0.021
-0.277
-0.003
0.014
0.077
(0.017)
(0.283)
(0.006)
(0.010)
(0.091)
-0.018
-1.24***
0.004
-0.007
-0.051
(0.014)
(0.240)
(0.005)
(0.008)
(0.077)
-0.021***
-0.084*
-0.003***
-0.001
-0.007
(0.003)
(0.034)
(0.001)
(0.001)
(0.011)
-0.370***
-5.044***
-0.115***
-0.031
-0.314
(0.033)
(0.547)
(0.015)
(0.020)
(0.180)
-0.058***
0.801***
0.001
-0.000
-0.026
(0.011)
(0.187)
(0.004)
(0.006)
(0.059)
TV hours
-0.001
(0.001)
0.002
(0.013)
-0.000
(0.000)
0.000
(0.000)
-0.001
(0.004)
Play computer games
-0.003
-0.051
-0.001*
-0.000
0.013
(0.002)
(0.030)
(0.000)
(0.001)
(0.010)
-0.029**
-0.465*
-0.003
0.002
-0.036
(0.010)
(0.182)
(0.004)
(0.006)
(0.059)
0.101***
-0.295
0.038***
0.017
-0.004
(0.020)
(0.361)
(0.008)
(0.012)
(0.113)
0.119*
2.06*
-0.028
-0.068
-0.450
Hang with friends
Play sports
Felt depressed
(0.058)
(0.976)
(0.022)
(0.036)
(0.323)
Felt fearful
0.051
-1.93**
0.002
-0.024
0.118
(0.039)
(0.654)
(0.012)
(0.023)
(0.210)
Private school
0.056
3.844***
0.034**
-0.034
-0.114
(0.028)
GPA
PVT scores
(0.523)
(0.012)
(0.021)
(0.190)
2.807***
0.059***
-0.030***
-0.074
(0.239)
(0.005)
0.002***
(0.008)
0.000
(0.077)
0.012**
0.014***
(0.000)
Smoke
Alcohol
123
(0.000)
(0.000)
(0.004)
-0.249***
-0.410
-0.050***
-0.052*
0.254
(0.036)
(0.585)
(0.014)
(0.021)
(0.186)
-0.025
1.576***
-0.005
0.018
0.076
(0.025)
(0.439)
(0.009)
(0.015)
(0.137)
J Fam Econ Iss
Table 3 continued
GPA
PVT scores
Graduated on time
Employed
Log income
-0.191***
-0.892
-0.017
-0.002
-0.129
(0.032)
(0.533)
(0.011)
(0.018)
(0.166)
Sex
-0.164***
(0.029)
-1.472**
(0.476)
-0.040***
(0.011)
0.008
(0.016)
0.134
(0.150)
N
3952
6005
6164
5871
4238
Drugs
Standard errors shown in parentheses. OLS regression was performed in the cases of GPA, PVT scores and log of Income. For all other
dependent variables, probit regressions were performed. Only coefficients with statistical significance in at least one regression are included in
the table. Also included in each regression were controls for racial background, regional school dummies, and neighborhood variables
Significance Levels: p \ 0.10. * p \ 0.05. ** p \ 0.01. *** p \ 0.001
individual had: (1) smoked every day for 30 days, (2)
consumed alcohol when not with parents or other adults in
the family, (3) used drugs (marijuana or cocaine) in their
life, and (4) engaged in sexual activities ever in their life.
These behavioral indicators from Waves II and III constituted our next set of dependent variables. For Wave IV
behavioral indicators, we had information only on whether
an individual had smoked every day for 30 days.
Our final set of outcome variables was respondents’
body mass index (BMI) and obesity and indicators of
mental health. Waves II and III have self-report of individuals’ weight and height. We calculated BMI using the
formula: (weight in pounds times 4.88) divided by (square
of height in feet). In Wave IV, individuals were asked
whether they considered themselves obese or not and we
used this self-report as the outcome variable for the final
wave.4
We also looked at self-reports of respondents’ mental
health. In particular, we relied on nineteen of the twenty
questions from the Center of Epidemiological Studies
Depression (CES-D) scale. Adolescents were asked to
report how often in the past week they experienced each of
the nineteen symptoms, and received a score between 0 and
3 for each question. A score of 0 indicated ‘‘never or
rarely,’’ 1 indicates ‘‘sometimes,’’ 2 indicates ‘‘a lot of the
time,’’ and 3 indicates ‘‘most of the time or all of the time.’’
The scoring of the positive items was reverse coded. The
CES-D scale, constructed by adding up the values from
each of the nineteen questions (thus ranging between 0 and
4
The correlation between the BMI’s calculated from self-reported
weights and heights in Wave III is 0.78. Note that these waves are
5 years apart, so there may be some changes in both height and
weight across these years. Also, using a rough guideline of defining
obesity as having a BMI C30, binary variables on obesity were
created for Waves II and III. The correlation between Waves III and
IV is 0.6113. The correlation between Waves II and IV is 0.3625.
Note that Waves III and IV are 7 years apart and Waves II and IV are
12 years apart.
57), has been widely used in past research to examine
depression (Fletcher 2009, 2008; Roberts et al. 1991;
Radloff 1977). Thus, a high CES-D score indicates poor
mental well-being and we represented that as Mental
Health Index (MHI). Furthermore, following Roberts et al.
1991, we created a binary variable ‘‘Depressed’’ using cutoff points from the CES-D scoring. For males, this variable
equals 1 if the score was 22 and above and zero otherwise.
For females, a cut-off point of 24 was used. This information was included in our analysis for Wave II.5
Disruption in Parental Co-habitation
Because changes in family structure aid in identification of
causal effects, we desired to observe individuals both
before and after a change in parental co-habitation occured.
In each wave, adolescents were asked about the presence of
various household members for up to twenty such members. This lead us to create a binary variable which takes a
value of one if both mother and father were present in the
household in Wave I but only one was reported present in
Wave II (approximately one year later).6 Since we wish to
look at the effects of the change in parental co-habitation,
we drop from the analysis adolescents who reported living
with only one parent during Wave I. Unfortunately, a
weakness of our study is that we are unsure if the disruption in parental co-habitation was due to divorce, separation, death, or other reasons, and our results need to be
taken in such context. While Wave I of the survey has a
parent questionnaire where parents were asked about their
marital status at that point of time, later waves do not
include this information. However, our variable indicating
5
All of relevant data to construct the ‘‘Depressed’’ variable is not
available in Waves III and IV.
6
While constructing this variable, apart from the presence of both
biological parents, we also included the presence of two foster parents
or adoptive parents.
123
J Fam Econ Iss
Table 4 Wave III ordinary least squares and probit marginal effects estimates of other outcomes
Smoke
Live with one parent
Alcohol
Drugs
Sex
BMI
MHI
-0.044
0.085*
-0.016
0.023
0.015
0.058
(0.034)
(0.024)
(0.025)
(0.020)
(0.400)
(0.225)
Age
-0.055***
0.006
-0.047***
0.002
0.432***
-0.071
(0.006)
(0.004)
(0.004)
(0.003)
(0.071)
(0.040)
Males
0.056***
0.013
0.091***
-0.023*
0.245
-0.546***
Black
(0.016)
(0.012)
(0.012)
(0.001)
(0.195)
(0.112)
-0.241***
-0.134***
-0.000
0.023
0.489
-0.012
(0.017)
(0.020)
(0.018)
(0.014)
(0.290)
(0.166)
No. of siblings
-0.002
-0.009*
-0.003
0.001
-0.119
0.031
(0.005)
(0.004)
(0.004)
(0.004)
(0.071)
(0.040)
Mother’s education
0.001
(0.004)
0.004
(0.003)
0.009***
(0.003)
-0.003
(0.002)
-0.105*
(0.043)
-0.012
(0.024)
Mother works full time
Father works full time
Closeness to mother
0.009
0.020
0.010
0.001
-0.018
-0.102
(0.018)
(0.013)
(0.013)
(0.011)
(0.217)
(0.124)
-0.145
-0.027
0.047*
0.014
-0.026
-0.507
(0.031)
(0.023)
(0.022)
(0.017)
(0.372)
(0.210)
0.005
-0.006
-0.015
-0.022*
0.188
-0.136
(0.012)
(0.010)
(0.010)
(0.009)
(0.148)
(0.085)
-0.007
-0.007
-0.034
-0.007
0.507***
-0.325***
(0.010)
(0.008)
(0.008)
(0.007)
(0.124)
(0.071)
0.004*
0.000
0.003*
0.000
0.077***
0.012
(0.002)
(0.001)
(0.001)
(0.002)
(0.021)
(0.011)
0.056*
-0.025
0.015
-0.010
-0.039
0.576***
(0.024)
(0.018)
(0.018)
(0.016)
(0.285)
(0.165)
Attention problem
0.010
0.008
0.016**
-0.001
-0.236*
0.308***
(0.007)
(0.006)
(0.006)
(0.005)
(0.098)
(0.055)
TV hours
0.000
(0.001)
0.001
(0.000)
-0.000
(0.000)
0.000
(0.000)
0.023***
(0.007)
0.006
(0.004)
Play computer games
-0.001
-0.001
-0.000
-0.001
0.030
0.011
(0.001)
(0.001)
(0.001)
(0.001)
(0.015)
(0.009)
-0.051
Closeness to father
Skipped school
Grade repeated
Hang with friends
Play sports
Felt depressed
0.031***
0.017**
0.025***
0.021***
-0.261**
(0.008)
(0.006)
(0.0106)
(0.005)
(0.095)
(0.054)
-0.027
0.053***
0.011
0.038***
0.150
-0.162
(0.108)
(0.016)
(0.022)
(0.012)
(0.010)
(0.187)
0.006
-0.024
0.067
0.007
0.057
(0.043)
(0.033)
(0.036)
(0.028)
(0.044)
-0.068*
-0.055*
-0.021
-0.012
0.343
0.358
(0.027)
(0.022)
(0.020)
(0.017)
(0.340)
(0.197)
-0.013
-0.057**
0.028
-0.017
-0.232
-0.511**
(0.027)
(0.022)
(0.021)
(0.017)
(0.318)
(0.187)
GPA
-0.062***
0.001
-0.018*
-0.033***
-0.514***
-0.115
PVT scores
(0.010)
0.000
(0.007)
0.003***
(0.008)
0.001**
(0.007)
-0.001*
(0.125)
0.011
(0.072)
-0.021***
Felt fearful
Private school
(0.000)
Smoke
Alcohol
123
(0.000)
(0.000)
(0.000)
(0.007)
(0.004)
0.028
0.074***
0.045**
-0.457
0.097
(0.019)
(0.020)
(0.017)
(0.306)
(0.177)
0.150***
0.099***
0.094***
0.248
-0.098
(0.019)
(0.014)
(0.011)
(0.227)
(0.132)
J Fam Econ Iss
Table 4 continued
Smoke
Alcohol
0.173***
(0.023)
Sex
0.110***
(0.021)
0.031*
(0.015)
0.027
(0.016)
N
4663
5288
5262
Drugs
Drugs
Sex
BMI
MHI
0.058***
0.046**
-0.638*
0.072
(0.015)
(0.015)
(0.276)
(0.161)
0.424
(0.249)
-0.067
(0144)
4112
5285
5273
Standard errors shown in parentheses. OLS regression was performed in the cases of BMI and mental health (depressed and sad). For all other
dependent variables, probit regressions were performed. Only coefficients with statistical significance in at least one regression are included in
the table. Also included in each regression were controls for racial background, regional school dummies, and neighborhood variables
Significance Levels: p \ 0.10. * p \ 0.05. ** p \ 0.01. *** p \ 0.001
living with one parent in Wave I was moderately positively
correlated with the divorce variable obtained from the
parent questionnaire.7
Independent Variables
For our empirical strategy to be viable, we conditioned on
a variety of adolescent characteristics and behaviors at or
before the time of the first survey (prior to a change in
parental co-habitation). Because we wished to allow the
effect of a change in family structure to operate through a
number of potential channels, we avoided covariates
corresponding to later points in time, as subsequent
behaviors or characteristics could be influenced by the
prior change in family structure. In addition to Wave I
values for the outcome variables described above, we
included demographic variables, namely age at Wave I,
gender (with females as the omitted category), and racial
background (whites being the omitted category). We also
controlled for the number of siblings, mother’s education
level, whether mother and father work full time, and
closeness to mother and father. Adolescents were asked in
detail about their school activities; we included the
number of days they skipped school without excuse, and
indicators of whether they repeated grades and how much
difficulty they had paying attention in school. They were
also asked about their daily activities; we included the
number of hours they spent watching television, playing
video games, hanging out with friends, and whether they
played sports. We also controlled for the type of school
they went to (with public school being the omitted category), whether the school was in an urban, semi-urban or
7
The correlation between the parent answering as not divorced,
separated, or single in Wave I and student reporting that both parents
are together in Wave I was 0.594. However, the analysis is limited to
students whose parents separate between Waves I and II (in order to
develop a treatment group). Wave II does not have a parent
questionnaire, so a switch in status from the parents’ perspective
was not observed.
rural location (semi-urban is the omitted category), and
indicator variables corresponding to living in the West,
Midwest, South or Northeast regions of the United States
(Midwest being the omitted category). Finally, we controlled for neighborhood related variables, namely selfreports of whether they lived in a neighborhood where
people care, whether they felt safe, and how happy they
felt to be living in the neighborhood.8 .
Sample Selection
We drew our sample from the 20,745 adolescents who
were interviewed for the initial in-home survey. For outcomes occurring in Wave II and Wave III, we relied on the
14,738 and 15,197 responses to each of these surveys,
respectively. Individuals living with one parent at the time
of the initial survey were dropped from the sample. We
also dropped individuals who were initially aged 17 and
over, to avoid including non-minors who may have moved
out of their parent’s home by the time of Wave II. Finally,
conditioning on non-missing values for all of our included
covariates resulted in a sample size of 9,311.
Table 1 displays the sample means and standard deviations of each of the covariates for three different groups.
Group (i) is the treated group, comprised of individuals
whose parents experienced a change in co-habitation
between Waves I and II. Groups (ii) and (iii) are both made
up of individuals from intact families, where Group (ii)
contains the entire control sample and Group (iii) contains
only those individuals who were matched to the treated
group using the 2-nearest neighbors technique. Also
reported are the p values of the differences in means
between groups (i) and (ii) and between groups (i) and (iii).
Ideally, the covariate values of the treatment group and the
8
It is acknowledged that some of these variables may have
endogeneity and unobserved heterogeneity issues, which could bias
the coefficients on these independent variables. However, there
should be no major effect of this endogeneity on the variable of
interest which is parental separation.
123
J Fam Econ Iss
Table 5 Wave IV ordinary least squares and probit marginal effects estimates
Ever attended college
Has college degree
Log Income
-0.064***
-0.063***
-0.049
0.061*
-0.024
(0.015)
(0.018)
(0.028)
(0.026)
(0.032)
0.031***
0.051***
0.080***
-0.035***
0.021***
(0.004)
(0.006)
(0.008)
(0.004)
(0.006)
-0.072***
-0.079***
0.290***
0.043***
0.031
(0.012)
(0.015)
(0.022)
(0.012)
(0.016)
0.052***
0.089***
-0.016
-0.086***
0.065**
(0.015)
(0.023)
(0.033)
(0.012)
(0.025)
No. of siblings
-0.010*
-0.012*
0.005
0.001
-0.012*
(0.004)
(0.006)
(0.008)
(0.004)
(0.006)
Mother’s education
0.035***
(0.003)
0.049***
(0.003)
0.012*
(0.005)
0.000
(0.003)
-0.016***
(0.004)
Live with one parent
Age
Males
Black
Mother works full time
Father works full time
Closeness to mother
Closeness to father
Skipped school
Grade repeated
Smoke
Obesity
0.005
0.024
0.069**
-0.005
-0.003
(0.013)
(0.017)
(0.025)
(0.013)
(0.018)
0.041
0.058*
0.068
-0.044
-0.016
(0.023)
(0.030)
(0.043)
(0.023)
(0.031)
-0.006
0.002
0.035*
0.002
0.024
(0.009)
(0.012)
(0.017)
(0.008)
(0.012)
0.003
0.005
-0.023
0.010
0.017
(0.007)
(0.010)
(0.015)
(0.007)
(0.010)
-0.002
-0.005*
-0.001
0.003**
0.003
(0.001)
(0.002)
(0.002)
(0.001)
(0.002)
-0.140***
-0.158***
-0.199***
0.063***
-0.026
(0.019)
(0.021)
(0.034)
(0.019)
(0.023)
Attention problem
-0.000
0.001
-0.001
-0.001
-0.018*
(0.006)
(0.008)
(0.011)
(0.006)
(0.008)
TV hours
0.000
(0.00)
-0.001
(0.001)
-0.002*
(0.001)
0.000
(0.000)
0.001**
(0.000)
Play computer games
-0.001
-0.005***
-0.002
0.000
0.003*
(0.001)
(0.001)
(0.002)
(0.001)
(0.001)
Hang with friends
Play sports
0.002
-0.012
0.030**
0.023***
-0.019**
(0.006)
(0.007)
(0.011)
(0.005)
(0.008)
0.022
0.060***
0.046**
0.051*
-0.024*
(0.011)
(0.015)
(0.022)
(0.011)
(0.016)
(continued)
Felt depressed
0.002
-0.032
0.106
0.012
0.001
(0.029)
(0.042)
(0.062)
(0.030)
(0.002)
0.002
-0.051
-0.023
-0.015
0.022
(0.020)
(0.027)
(0.040)
(0.018)
(0.028)
0.057**
0.081**
0.016
-0.021
-0.048
(0.021)
(0.026)
(0.036)
(0.019)
(0.026)
GPA
0.112***
(0.007)
0.210***
(0.010)
0.108***
(0.015)
-0.054***
(0.007)
-0.051***
(0.010)
PVT scores
0.005***
0.006***
0.002**
0.000
0.001
(0.000)
(0.001)
(0.001)
(0.000)
(0.001)
Smoke
-0.081***
-0.117***
-0.074*
(0.020)
(0.023)
(0.035)
Alcohol
0.004
0.033
0.030
Felt fearful
Private school
123
0.004
(0.025)
0.029*
-0.002
J Fam Econ Iss
Table 5 continued
Ever attended college
Has college degree
Log Income
Smoke
Obesity
(0.019)
(0.014)
(0.018)
(0.026)
(0.013)
Drugs
0.003
-0.020
0.019
0.063***
-0.056**
Sex
(0.016)
-0.063***
(0.022)
-0.097***
(0.032)
-0.011
(0.017)
0.087***
(0.022)
0.11
(0.016)
(0.019)
(0.029)
(0.016)
(0.021)
N
6199
6199
4238
5415
4112
Standard errors shown in parentheses. OLS regression was performed in the case of log of Income. For all other dependent variables, probit
regressions were performed. Only coefficients with statistical significance in at least one regression are included in the table. Also included in
each regression were controls for racial background, regional school dummies, and neighborhood variables
Significance Levels:
p \ 0.10. * p \ 0.05. ** p \ 0.01. *** p \ 0.001
control group should be as close as possible. For each
covariate the difference in means between the treated and
control groups decreased significantly for the matched
sample as compared to the unmatched sample. The p values suggest that there are no statistically significant differences in means of the covariates between the treatment
group and the matched control group. Correspondingly,
there was a reduction in the standardized bias for nearly all
of the covariates in the matched sample in comparison to
the unmatched sample.
Results
Estimated coefficients from OLS and the marginal effects
of probit regressions on short-term (Wave II) outcomes are
presented in Table 2. Each column represents a separate
regression corresponding to a specific outcome variable. As
seen in the table, a disruption in parental co-habitation had
a statistically significant negative effect on immediate
outcomes in terms of scholastic performance (GPA) and
mental health (MHI and Depressed), and sexual behavior.
The coefficients on several of the covariates were statistically significant. In general, males were found to have
higher BMI while females were more likely to be depressed and sad. An individual who reported repeating a grade
tended to have a lower GPA. We also found that mother’s
education had a positive effect on GPA and the closer one
feels to the father, the less likely that he/she indulged in
risky behavior and also less likely to feel depressed. While
sports participation was associated with higher GPA, it also
increased the likelihood of consuming alcohol and drugs as
well as engaging in sexual activities. Indulgence in risky
behaviors also significantly increased with more time spent
with friends.
When we allowed for more time between parental separation and the outcome measurement, living with one
parent was found to be more influential on a larger number
of outcomes. Our medium term (Wave III, occurring about
7 years after the original survey) OLS and probit results are
presented in Tables 3 and 4. The estimates of education
and employment outcomes are presented in Table 3
whereas those of risky behavior and health outcomes are
presented in Table 4. We found that there was a significant
negative effect of parental co-habitation disruption on
academic outcomes, including both GPA and timely
graduation from high school. Living with one parent was
associated with a decrease in overall high school GPA of
about 0.14 points and a 5.6 % lower likelihood of graduating within 4 years. Also, wages significantly decreased
and smoking behavior significantly increased as well.
Other covariates in Wave III regressions (Tables 3, 4)
impacted outcomes as predicted.9All academic and
employment outcomes increased significantly with
mother’s education. Sports participation was associated
with an increase in both GPA and the likelihood of on-time
high school graduation, but also an increased likelihood of
consuming alcohol and having sex, similar to what we
found for Wave II
The longer term (Wave IV) OLS and marginal effects of
the probit regressions are presented in Table 5. Adolescents living in a single-parent household between Waves I
and II were 6.4 % less likely to have attended college and
6.3 % less likely to have earned a college degree in the
long run. They were also more likely to be earning less
compared to adolescents who did not experience a change
in family structure between Waves I and II, although this
effect was only marginally significant. In the long run, they
were also 6.1 % more likely to engage in smoking.
9
In the interest of conserving space, we do not report the coefficients
of all the covariates in the OLS and probit regressions, especially the
ones whose coefficients were not statistically significant in any of the
regressions. However, a full list of coefficients of all covariates is
available upon request.
123
J Fam Econ Iss
Table 6 Propensity Score Matching Estimates of Parental Separation on multiple outcomes of adolescents
Wave II
2-NN
ATT
Wave III
K-D
ATT
Wave IV
2-NN
ATT
K-D
ATT
2-NN
ATT
N
K-D
ATT
Treated
Untreated
Academics
GPA
-0.202**
-0.153**
-0.159*
-0.171**
189
3763
(0.071)
(0.059)
(0.076)
0.412
(0.063)
0.049
1123
4882
(0.667)
(0.541)
-0.052***
-0.055***
1171
4993
(0.015)
(0.013)
1215
4984
1215
4984
1112
4759
751
3487
266
4396
297
4990
299
4963
300
4973
223
3889
300
4985
300
4985
PVT scores
Graduated on time
Ever attended college
Has college degree
-0.032
-0.050***
(0.018)
(0.015)
-0.025
-0.046**
(0.019)
(0.016)
Employment
Employed
Log income
0.009
0.009
(0.018)
(0.015)
-0.224
-0.175
-0.044
-0.047
(0.182)
(0.151)
(0.030)
(0.026)
0.086*
(0.038)
0.081*
(0.032)
0.069*
(0.032)
0.067*
(0.027)
Risky behavior
Smokea
0.009
(0.033)
Alcohola
Drugs
Sex
0.046
(0.027)
-0.007
-0.008
-0.037
-0.024
(0.025)
(0.029)
(0.031)
(0.026)
0.060
0.049
0.038
0.024
(0.031)
(0.027)
(0.030)
(0.025)
0.065
0.069*
0.048
0.033
(0.036)
(0.030)
(0.025)
(0.020)
0.394
0.115
0.508
0.164
-0.004
-0.013
(0.378)
(0.322)
(0.546)
(0.460)
(0.040)
(0.033)
0.925
1.505***
0.215
0.091
(0.580)
(0.470)
(0.314)
(0.252)
0.035
0.050**
(0.022)
(0.019)
Health and mental health
BMI/obesity
Mental health
Depressed
ATT corresponds to the estimate of the average treatment on the treated. Standard errors shown in parentheses. 2-NN indicates a 2-nearest
neighbors weighting method and K-D indicates use of a kernel density weighting method (with a biweight distribution kernel smoothing function
using a bandwidth of 0.6). Untreated sample sizes correspond to use of K-D
Significance Levels:
a
p \ 0.10. * p \ 0.05. ** p \ 0.01. *** p \ 0.001
indicatesone observation in the treatment effects calculation was dropped due to not having propensity scores with a common support
Mother’s education and sports participation had a positive
significant effect on educational achievement and income.
Also, adolescents were more likely to engage in smoking
in the long term with more time spent with friends in
Wave I.
In order to investigate changes in academic achievement, risky behaviors, and mental health, we controlled for
risky behavior, test scores, grades, and mental health in
123
Wave I.10Furthermore, academic achievement was significantly lowered the more one engaged in risky behavior
like smoking, drinking, drug use and sex in Wave I.
10
We did not use lagged dependent variables due to attenuation bias
(see Maddala and Rao 1973). However, the results using the lagged
dependent variables are available upon request.
J Fam Econ Iss
We now turn to our results from propensity score
matching. As described above, the propensity scores were
estimated using a logit model, and then non-parametric
estimation was carried out on all outcome variables using
both the two-nearest neighbors and kernel density
approaches.11 The results of the second stage are reported
in Table 6. The effect on GPA using these two methods
was a little bit stronger in both Waves II (ranging between
0.153 and 0.202 GPA points lost, as compared to 0.134
points lost using OLS) and III (ranging between 0.159 and
0.171 GPA points lost, as compared to 0.141 points lost
using OLS). There was a similar issue with the attending
college and getting a college degree variables as well,
where the propensity score matching estimates were a little
stronger (although not by much). On time graduation rates
had similar propensity score matching estimates (between a
5.2 and 5.5 % decreased likelihood) as probit estimates.
Effects on income were non-significant using propensity
score matching (this effect was weakly significant using the
OLS estimate).
Regarding behavioral outcomes using the propensity
score matching methods, the effects of separation on
smoking behavior were virtually the same in Waves III and
IV as before. The propensity score estimates of the effects
on sexual behavior (between 6.5 and 6.9 % increased
likelihood in Wave II and between a 3.2 and 5 % increased
likelihood in Wave III) were a little stronger using propensity score matching. Also, the effects on mental health
and depression were slightly stronger after using propensity
score matching.
Conclusion
The effects of living with one parent on outcomes of
children have been widely debated in the literature. An
understanding of the potential effects of separation or
divorce on short and long term outcomes of children is
important for the design of new divorce laws that might
combat the incidence of divorce and thus its adverse effects
on children. While some researchers have concluded that
living with a single parent does not have an adverse effect
on children, others have found significant negative effects.
Our paper underscores the importance of time frame in
analyzing this issue; we found evidence of not only adverse
effects of change in family structure in the immediate term
(approximately one year or less following the change), but
also significant medium and longer term (about 3–5 years,
11
We also checked the robustness of these two methods using onenearest and three-nearest neighbors matching, along with normal and
Epanechnikov kernels, and these results were similar. These results
are available upon request.
and later) effects on the academic achievement of children
and their propensity to engage in smoking.
We used data from four waves of the National Longitudinal Survey of Adolescent Health, a large national data
set involving students initially in junior high and high
school. The data allowed us to control for an unusually rich
set of covariates, including demographics, parents’ characteristics, home and school environment, school characteristics, daily activities, and neighborhood variables.
Unlike the majority of prior research, we included a wide
range of outcomes which may be related to future wellbeing, including academic performance, performance on a
standardized test, timely graduation from high school,
college attendance, and college degree, propensity to
engage in risky behaviors like drinking, smoking, drug use
and sex, and BMI/obesity and mental health outcomes.
Furthermore, to our knowledge, our paper is one of the
very few in the literature to use propensity score matching
to uncover the effects of change in family structure on
children’s outcomes. We used both two-nearest neighbor
and kernel density weighting methods, and compared our
results to those from OLS and probit regressions.
We found evidence that parental separation negatively
impacts academic outcomes (GPA, graduating on time,
college attendance, and obtaining a degree) behavioral
outcomes (smoking), and mental health. Furthermore,
many of these effects became a little stronger after using
propensity score matching, which suggests that these
effects have been slightly underestimated in the past literature when using more restrictive functional forms and
not using appropriate the comparison groups generated by
matching methods.
Our results underscore the importance of strong twoparent families for the well-being of children. Policies
could aim at providing incentives to maintain stable and
intact families. There could also be provision for economic
security of children and certain community resources like
support centers and more affordable education. Support
from schools is also an important predictor of their mental
well-being (Hussey et al. 2013). Since children who belong
to single-parent families are more likely to suffer from
emotional stress and lower academic performance, good
support systems and counseling services in schools should
be emphasized. Child support laws can also help to alleviate on-going disadvantages of children who undergo
parental separation (Allen et al. 2011; Hofferth and Pinzon
2011). Our results also underscore the importance of
addressing both short-term and longer-term implications of
parental separation. In addition to policies aimed at
immediate consequences of separation, such as lower
grades and increased likelihood of smoking, policies
affecting young adult outcomes, like college enrollment
and graduation, should also be considered.
123
J Fam Econ Iss
There are some limitations in our analysis. First, while
the propensity score matching approach can be seen as an
improvement over tradition multiple regression methods,
and while our data provides a rich set of control variables,
we cannot rule out the possibility of omitted variables
biasing our estimates, nor the possibility that one or more
of our included independent variables are endogenous.
Second, our sample only included changes in family
structure during a single year, when children were initially
between grades 7 and 10. To the degree that parental
separation may have a greater influence on younger aged
children, our results may underestimate the effects of
separation on children in the population as a whole.
Finally, it should be emphasized that our analysis remained
agnostic as to the cause for parental separation, as well as
the total length of time over which adolescents remain in a
single parent home. Our estimates likely reflect a mix of
the effects of parental divorce, death or other separation
(temporary or permanent). Given our findings of multiple
negative effects of separation, we view our results as being
likely lower bounds of the true effects of parental separation associated with marriage dissolution. However, our
analysis leaves room for further research investigating
particular causes for parental separation, and the heterogeneous effects they may have on multiple outcomes and
over multiple time horizons. Additional outcomes such as
grade level repetition and teenage pregnancy, as well as
multi-generational effects of parental separation, including
future marriage and fertility decisions of the child, should
also be addressed with the availability of appropriate data.
Acknowledgments The research uses data from Add Health, a
program project designed by J. Richard Udry, Peter S. Bearman, and
Kathleen Mullan Harris, and funded by a Grant P01-HD31921 from
the Eunice Kennedy Shriver National Institute of Child Health and
Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due Ronald Rindfuss and Barbara
Entwisle for assistance in the original design. Persons interested in
obtaining Data Files from Add Health should contact Add Health, The
University of North Carolina at Chapel Hill, Carolina Population
Center, 123 W. Franklin Street, Chapel Hill, NC 27516-2524 (
[email protected]). No direct support was received from grant P01HD31921 for this analysis.
References
Allen, B., Nunley, J., & Seals, A. (2011). The effect of joint-childcustody legislation on the child-support receipt of single
mothers. Journal of Family and Economic Issues, 32, 124–139.
doi:10.1007/s10834-010-9193-4.
Amato, P. R. (2000). The consequences of divorce for adults and
children. Journal of Marriage and Family, 62, 1269–1287.
doi:10.1111/j.1741-3737.2000.01269.x.
Amato, P. R., & Booth, A. (1997). A generation at risk: Growing up
in an era of family upheaval. Cambridge: Harvard University
Press.
123
Amato, P. R., & Keith, B. (1991). Consequences of parental divorce
on children’s well-being: A meta-analysis. Psychological Bulletin, 110, 26–46. Retrieved from http://www.jstor.org/stable/
353132. Accessed 24 Nov 2014.
Bane, M. J., & Jargowsky, P. (1988). The links between government
policy and family structure: what matters and what doesn’t. In A.
J. Cherlin (Ed.), The changing American family and public
policy (pp. 219–255). Washington DC: Urban Institute Press.
Basu, A., Polsky, D., & Manning, W. (2008). Use of propensity scores
in non-linear response models: the case for health care expenditures. NBER WP14086, National Bureau of Economic
Research. Retrieved from http://www.nber.org/papers/w14086.
pdf. Accessed 24 Nov 2014.
Brown, S. L. (2006). Family structure transitions and adolescent wellbeing. Demography, 43, 447–461. doi:10.1353/dem.2006.0021.
Chase-Lansdale, P. L., Cherlin, A.J., & Kiernan, K.E. (1995). The
long-term effects of parental divorce on the mental health of
young adults: A developmental perspective. Child Development,
66, 1614–1634. Retrieved from http://www.jstor.org/stable/
1131900. Accessed 24 Nov 2014.
Cherlin, A.J., Kiernan, K.E., & Chase-Lansdale, P.L. (1995). Parental
divorce in childhood and demographic outcomes in young
adulthood. Demography, 32, 229–318. Retrieved from http://
www.jstor.org/stable/2061682. Accessed 24 Nov 2014.
Cid, A., & Stokes, C. (2013). Family structure and children’s
education outcome: Evidence from Uruguay. Journal of Family
and Economic Issues, 34, 185–199. doi:10.1007/s10834-0129326-z.
Corak, M. (2001). Death and divorce: The long-term consequences of
parental loss on adolescents. Journal of Labor Economics, 19,
682–715. Retrieved from http://www.jstor.org/stable/10.1086/
322078. Accessed 24 Nov 2014.
Dehejia, R. H., & Wahba, S. (2002). Propensity score-matching for
non-experimental causal studies. The Review of Economics and
Statistics, 84, 151–161. doi:10.1162/003465302317331982.
Downey, D. B. (1994). The school performance of children from
single-mother and single-father families: Economic or interpersonal deprivation? Journal of Family Issues, 15, 129–147.
doi:10.1177/019251394015001006.
Dronkers, J. (1999). The effects of parental conflicts and divorce on
the well-being of pupils in Dutch secondary education. European
Sociological Review, 15, 195–212. Retrieved from http://apps.
eui.eu/Personal/Dronkers/articles/europeansociologicalre
view1999.pdf. Accessed 24 Nov 2014.
Entwisle, D. R., & Alexander, K.L. (1996). Family type and
children’s growth in reading and math over the primary grades.
Journal of Marriage and Family, 58, 734–746. Retrieved
from http://www.jstor.org/stable/353500. Accessed 24 Nov
2014.
Ermisch, J. F., & Francesconi, M. (2001). Family structure and
children’s achievements. Journal of Population Economics, 14,
249–270. doi:10.1007/s001480000028.
Fletcher, J. M. (2008). Adolescent depression: Diagnosis, treatment,
and educational attainment. Health Economics, 17, 1215–1235.
doi:10.1016/j.socscimed.2008.12.005.
Fletcher, J. M. (2009). Childhood mistreatment and adolescent and
young adult depression. Social Science and Medicine, 68,
799–806. doi:10.1002/hec.1319.
Frisco, M. L., Muller, C., & Frank, K. (2007). Parents’ union
dissolution and adolescents’ school performance: Comparing
methodological approaches. Journal of Marriage and Family,
69, 721–741. doi:10.1111/j.1741-3737.2007.00402.x.
Heckman, J. J., Ichimura, H., & Todd, P. (1998). Matching as an
econometric evaluation estimator. The Review of Economic
Studies, 65, 261–294. Retrieved from http://www.jstor.org/
stable/2971733. Accessed 24 Nov 2014.
J Fam Econ Iss
Hofferth, S., & Pinzon, A. (2011). Do nonresidential fathers’ financial
support and contact improve children’s health? Journal of
Family and Economic Issues, 32, 280–295. doi:10.1007/s10834010-9237-9.
Hussey, A.J., Kanjilal, D., & Okunade, A.A. (2013). Effects of peers
and social environment on adolescent psychological well-being.
International Journal of Business and Social Research, 3, 69–83.
Retrieved from http://thejournalofbusiness.org/index.php/site/arti
cle/view/75/74. Accessed 24 Nov 2014.
Imbens, G. W. (2000). The role of the propensity score in estimating
dose-response functions. Biometrika, 87, 706–710. doi:10.1093/
biomet/87.3.706.
Keith, V.M., & Finlay, B. (1988). The impact of parental divorce on
children’s educational attainment, marital timing and likelihood
of divorce. Journal of Marriage and the Family, 50, 797–809.
Retrieved from http://www.jstor.org/stable/352648. Accessed 24
Nov 2014.
Kohn, J., & Averett, S. (2014). Can’t we just live together? New
evidence on the effect of relationship status on health. Journal of
Family and Economic Issues, 35, 295–312. doi:10.1007/s10834013-9371-2.
Lang, K., & Zagorsky, J.L. (2001). Does growing up with a parent
absent really hurt? The Journal of Human Resources, 36,
253–273. Retrieved from http://www.jstor.org/stable/3069659.
Accessed 24 Nov 2014.
Maddala, G.S., & Rao, A.S. (1973). Tests for serial correlation in
regression models with lagged dependent variables and serially
correlated errors. Econometrica, 41, 761–774. Retrieved
from http://www.jstor.org/stable/1914095. Accessed 24 Nov
2014.
Manski, C.F., Sandefur, G.D., McLanahan, S., & Powers, D. (1992).
Alternative estimates of the effect of family structure during
adolescence on high school graduation. Journal of the American
Statistical Association, 87, 25–37. Retrieved from http://www.
jstor.org/stable/2290448. Accessed 24 Nov 2014.
Marsh, H. W. (1990). Two-parent, stepparent, and single-parent
families: Changes in achievement, attitudes, and behaviors
during the last two years of high school. Journal of Educational
Psychology, 82, 327–340. doi:10.1037/0022-0663.82.2.327.
Mayer, S. (1997). What money can’t buy: Family income and
children’s life chances. Cambridge: Harvard University Press.
McLanahan, S., & Sandefur, G. D. (1994). Growing up with a single
parent: What hurts, what helps. Cambridge: Harvard University
Press.
Myers, S., & Myers, C. (2014). Family structure and school-based
parental involvement: A family resource perspective. Journal of
Family & Economic Issues. doi:10.1007/s10834-014-9409-0.
Radloff, L. (1977). The CES-D scale: A self report depression scale
for research in the general population. Applied Psychological
Measurement, 1, 385–401. doi:10.1177/014662167700100306.
Roberts, R., Lewinsohn, P., & Seeley, J. (1991). Screening for
adolescent depression: A comparison of depression scales.
Journal of the American Academy for Child Adolescent Psychiatry, 30, 58–66. doi:10.1097/00004583-199101000-00009.
Rosenbaum, P. R., & Rubin, D.B. (1983). The central role of the
propensity score in observational studies for causal effects.
Biometrika, 70, 41–45. Retrieved from http://www.jstor.org/
stable/2335942. Accessed 24 Nov 2014.
Sabia, J. (2007). The effect of body weight on adolescent academic
performance. Southern Economic Journal, 73, 871–900.
Retrieved from http://www.jstor.org/stable/20111933. Accessed
24 Nov 2014.
Sanz-De-Galdeano, A., & Vuri, D. (2007). Parental divorce and
students’ performance: Evidence from longitudinal data. Oxford
Bulletin of Economics and Statistics, 69, 321–338. doi:10.1111/j.
1468-0084.2006.00199.x.
Schramm, D. (2005). Individual and social costs of divorce in Utah.
Journal of Family and Economic Issues, 27, 133–151. doi:10.
1007/s10834-005-9005-4.
Tumin, D., Han, S., & Qian, Z. (2014). Meanings and measures of
marital separation. Working paper. Retrieved from http://
paa2014.princeton.edu/papers/141524. Accessed 24 Nov 2014.
Andrew Hussey is an associate professor at University of Memphis
that specializes in labor economics, the economics of education, and
applied econometrics. Dr. Hussey’s areas of research include
investigating the effects of higher education on labor market
outcomes, explaining income inequality and race and gender differentials in labor market and education outcomes, and adolescent
obesity. Dr. Hussey received his Ph.D. from Duke University in 2006.
Debjani Kanjilal is an assistant professor at Elizabeth City State
University where she teaches multiple classes in Economics and
Business Administration. Her primary research interest is in the area
of health and labor economics, specifically adolescent behavior and
outcomes. She is also engaged in inter-disciplinary research in the
area of Management and Information Systems. Dr. Kanjilal received
her Ph.D. in economics from the University of Memphis in 2010.
Anil Nathan is an assistant professor at Holy Cross that specializes in
the economics of education and health economics, along with
implementing matching methods to control for selection on observables. Dr. Nathan has publications examining end of life care for
dementia patients along with analyzing survival probabilities of NFL
football players in relation to the rest of the general population. Dr.
Nathan received his Ph.D. in economics from Duke University in
2008.
123