Parental divorce is not uniformly disruptive to

children’s educational attainment

Jennie E. Branda, Ravaris Mooreb, Xi Songc, and Yu Xied,1
a
Department of Sociology, University of California, Los Angeles, CA 90095-1551; bDepartment of Sociology, Loyola Marymount University, Los Angeles, CA
90045; cDepartment of Sociology, University of Chicago, Chicago, IL 60637-1767; and dDepartment of Sociology, Princeton University, Princeton, NJ 08544

Contributed by Yu Xie, February 14, 2019 (sent for review August 1, 2018; reviewed by Michael Hout and Florencia Torche)

Children whose parents divorce tend to have worse educational families with divorced parents systematically differ from two-parent
outcomes than children whose parents stay married. However, not families in both observed and unobserved characteristics. They
all children respond identically to their parents divorcing. We focus have adopted a range of methods to address concerns over selection
on how the impact of parental divorce on children’s education into divorce, adding credibility to key findings regarding the negative
varies by how likely or unlikely divorce was for those parents. impact of parental divorce on children’s attainment (1). This prior
We find a significant negative effect of parental divorce on edu- work, however, has not considered how the effects of parental divorce
cational attainment, particularly college attendance and comple- vary according to the selection into, or the likelihood of, divorce.
tion, among children whose parents were unlikely to divorce. In this study, we consider how the effects of parental divorce
Families expecting marital stability, unprepared for disruption,
on children’s education differ across families who differ with
may experience considerable adjustment difficulties when divorce
respect to their propensity of divorce. The propensity of divorce
occurs, leading to negative outcomes for children. By contrast, we
find no effect of parental divorce among children whose parents
is a parsimonious measure of the likelihood that parental divorce
were likely to divorce. Children of high-risk marriages, who face occurs based on socioeconomic, psychosocial, and family con-
many social disadvantages over childhood irrespective of parental ditions. We suggest that marital disruption is not uniformly dis-
marital status, may anticipate or otherwise accommodate to the ruptive across families. For some, such events elicit little social-
dissolution of their parents’ marriage. Our results suggest that psychological and behavioral response. For others, such events
family disruption does not uniformly disrupt children’s attainment. are unexpected shocks and necessitate considerable adjustment.
The degree of disruption varies by the likelihood and corre-
|
divorce educational attainment | heterogeneity | causal analysis | sponding expectation that such events occur. Through examining
propensity score heterogeneity in the effects of divorce by the observed likelihood
that children experience a parental divorce, we shed light on

P arental divorce is, on average, associated with unfavorable

outcomes among children, including their ability to complete
high school and to attend and complete college (1, 2). However,
subpopulations of children for whom the causal effects of divorce
may be considerable, modest, or absent. We carefully attend to
the assumptions and methods necessary to estimate causal effects
is divorce uniformly negative for children’s attainment and well- across these subpopulations of interest, and offer alternative in-
being? We contend that families differ in their expectation of terpretations of our results based upon observed and unobserved
and ability to adjust and respond to marital disruption. Families selection into divorce. Our results suggest that a simple distinc-
expecting marital stability, unprepared for disruption, may ex- tion between children with divorced and nondivorced parents
perience considerable adjustment difficulties when divorce oc-
curs, leading to worse outcomes for children; by contrast, divorce Significance
among families who have come to expect disadvantage and in-
stability may not incur the same negative consequences.
While parental divorce is generally associated with unfavor-
Studies of parental divorce indicate that children’s response to
able outcomes for children, it does not follow that every di-
divorce varies by socioeconomic characteristics and family well-
vorce is equally bad for the children it affected. We find that
being. Children of more educated parents and white children parental divorce lowers the educational attainment of children
experience greater effects of parental divorce than children of who have a low likelihood of their parents’ divorcing. For these
less educated parents and racial and ethnic minority children (2– children, divorce is an unexpected shock to an otherwise-
6). Additionally, children of married parents with high levels of privileged childhood. However, we find no impact of parents’
conflict are no better off, and in fact may fare worse in some divorcing on the education of children who have a high likeli-
respects, than children of single parents (5–8). We argue that a hood of a divorce occurring. Disadvantaged children of high-
shared latent factor underlying observed interactions between risk marriages may anticipate or otherwise accommodate to
family socioeconomic status and well-being and parental divorce the dissolution of their parents’ marriage. Social discourse and
can be captured by the likelihood, or propensity, of divorce oc- policy aimed at promoting marital stability among disadvan-
curring. That is, the negative effects of parental divorce are taged families, for whom unfortunate events are common, are
greater among white children and children of more educated misguided.
parents because they are less accustomed to negative disruptive
events and disadvantaged circumstances than racial and ethnic Author contributions: J.E.B. and Y.X. designed research; J.E.B., R.M., X.S., and Y.X. per-
minority children and children of less educated parents (9–11). formed research; J.E.B., R.M., and X.S. analyzed data; and J.E.B., R.M., X.S., and Y.X. wrote
the paper.
More advantaged children are also unlikely to be embedded in a
social network in which family instability is more likely to occur Reviewers: M.H., New York University; and F.T., Stanford University.

and less stigmatized (12–14). By contrast, parental divorce may The authors declare no conflict of interest.
not further impede the educational attainment of children who Published under the PNAS license.
have grown accustomed to adverse events in their lives via al- 1
To whom correspondence should be addressed. Email: [email protected].
ready high levels of socioeconomic instability and family conflict. This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
Scholars studying the effects of parental divorce on children have 1073/pnas.1813049116/-/DCSupplemental.
primarily relied on observational data, while acknowledging that Published online March 26, 2019.

7266–7271 | PNAS | April 9, 2019 | vol. 116 | no. 15 www.pnas.org/cgi/doi/10.1073/pnas.1813049116

oversimplifies how parental divorce impacts children’s educational
attainment.
Results
Predicted Likelihood of Parental Divorce. With observational data,
the key to our identification strategy is the ignorability assump-
tion, that is, the assumption that parental divorce is uncorrelated
with unobserved factors that affect children’s outcomes (15). To
guard against potential selection bias and improve confidence in
the ignorability assumption, we condition the analyses on an
extensive set of observed characteristics using linked data from
the National Longitudinal Survey of Youth (NLSY) and the
National Longitudinal Survey’s Child–Mother file (NLSCM),
including maternal family background, socioeconomic, maternal
cognitive and psychosocial, and family formation and well-being
factors. We observe significant differences by parental divorce
status for most of the indicators we observe, suggesting greater Fig. 1. Effects of parental divorce on children’s educational attainment:
socioeconomic disadvantage and lower family well-being among Unadjusted and adjusted for the propensity score. Notes: Sample restricted
to children who were at least 18 y old in 2012 and whose parents were
children whose parents divorce than among those whose parents
married at the time of their birth. Parental divorce is measured as divorce
stay married (SI Appendix, Table S1). that occurred when children were 0–17 y old. Estimates are based on linear
We model the probability that a child experiences a parental probability models. Adjusted models control for propensity of parental di-
divorce over the course of childhood (age 0–17) as a function of vorce and child age in 2012 (estimates not shown). Propensity scores were
the predivorce covariates (SI Appendix, Table S2). As results estimated by a logit regression model of parental divorce on the set of

SOCIAL SCIENCES
from models predicting parental divorce are seldom presented in predivorce covariates. Analytic sample (n = 5,176) is further restricted to age
prior work on divorce effects on children, the literature has no 19 and above for college attendance (n = 4,982), and age 23 and above for
firmly established criteria by which to determine the strength of college completion (n = 3,901).
the prediction model. Our model incorporates a rich set of
theoretically informed covariates based on the literature on the
determinants of divorce. From SI Appendix, Table S2, we observe propensity for parental divorce at the median, we predict that
that mothers who themselves were raised in large families with among children whose parents stay married, about 81%
fathers present are less likely to divorce all else equal. Mother’s complete high school, 56% attend college, and 23% complete
self-esteem is negatively associated and depressive symptoms are college, while among children whose parents divorced, about
positively associated with the odds of divorce. High cognitive 78% complete high school, 50% attend college, and 17%
ability and higher academic achievement among mothers are complete college.
positively associated with divorce, all else equal; we note, nev-
Estimated Heterogeneous Effects of Parental Divorce. We next as-
ertheless, that descriptive statistics suggest that high-propensity
children have mothers with lower cognitive ability and achieve- sess whether the effects of divorce vary with the propensity for
ment (SI Appendix, Table S3). Education and household income parental divorce. There are important substantive payoffs to
generally reduce the odds of divorce, while mothers’ employ- understanding effect heterogeneity by the propensity for treat-
ment, especially employment at a private company without ment, whether or not the ignorability, or selection on observ-
flexible hours, increases odds of divorce. Family formation fac- ables, assumption holds true. The propensity score provides a
tors influence the likelihood of divorce, with women adopting parsimonious measure of an extensive set of observed covariates
more traditional family values and attitudes (e.g., delayed sexual that indicate the likelihood of divorce. Effect variability by the
debut and no prior marriages) less likely to divorce. Arguing propensity score lends itself to interpretations based on both
about chores is positively associated with divorce, while arguing observed and unobserved selection (16). That is, if the igno-
about money is negatively associated with divorce. Parents who rability assumption does not hold, we can interpret effect vari-
differ in their educational attainment and who are of different ability by the propensity score as resulting, at least partially, from
races are more likely to dissolve marriages. However, those unobserved selectivity (17, 18).
raised in different religions are less likely to divorce, perhaps We present local polynomial matching–smoothing heteroge-
reflecting strong selection into cross-religion marriages. In gen- neity results in Fig. 2. The x axis represents the continuous
eral, the likelihood of divorce increases as socioeconomic status propensity score and the y axis represents observed differences in
and psychosocial and family well-being decreases. (i) high school completion, (ii) college attendance, and (iii)
college completion between children whose parents did and did
Estimated Effects of Parental Divorce. We present linear probability not divorce. We observe a sizable negative effect of divorce on
model estimates of the effects of parental divorce on children’s educational attainment, particularly college attendance and
educational attainment in Fig. 1. Measures of children’s educa- completion, among children who had a low likelihood of expe-
tional attainment include high school completion by age 18, riencing a parental divorce, an effect that declines (i.e., becomes
college attendance by age 19, and college completion by age 23. less negative) as the propensity increases. The effect nears zero,
Unadjusted estimates suggest that divorce is associated with an or becomes positive, for children with a high propensity for pa-
8% lower probability of children’s high school completion, a rental divorce. The pattern in effects is curvilinear for high
12% lower probability of college attendance, and an 11% lower school completion (with little difference between children whose
probability of college completion. The magnitudes of the coef- parents had low and moderate likelihoods of divorce), and nearly
ficients are reduced when estimates are adjusted for the pro- linear for college attendance and completion (steeper for college
pensity of parental divorce, but retain significance. We observe completion than attendance). In each case, the general trend
that, net of the propensity for parental divorce, divorce is asso- indicates a reduction in the negative effect of parental divorce on
ciated with a 4% lower probability of children’s high school children’s education as the propensity for divorce increases.
completion, a 7% lower probability of college attendance, and We present heterogeneous effect estimates by propensity
a 7% lower probability of college completion. Holding the score strata in Fig. 3. Given the shape of response functions, and

Brand et al. PNAS | April 9, 2019 | vol. 116 | no. 15 | 7267

 effects of parental divorce on college attainment using samples
-.25 -.2 -.15 -.1 -.05 0 .05 .1
of more disadvantaged families (e.g., Fragile Families and Child
Wellbeing Study), who bear the most similarity to stratum 3, than
those we observe here.
Treatment effect

We underscore that we are comparing the effects of pa-

rental divorce on children’s educational outcomes across
strata, not children’s levels of educational attainment. Chil-
dren whose parents are unlikely to divorce have advantaged
family background characteristics and attain higher levels of
education. Educational outcomes differ far more by the pro-
0 .2 .4 .6 .8 1
pensity to divorce, as a summary proxy for family socioeco-
Propensity for divorce nomic well-being, than by parental divorce status. As a result,
High school completion College attendance
low-propensity children with divorced parents outperform
College completion high-propensity children with married parents. For example,
about 54% of children whose parents have a low propensity of
divorce but in fact divorce attend college, while about 43% of
Fig. 2. Matching–smoothing (MS) heterogeneous effects of parental di-
vorce on children’s educational attainment for (i) high school completion,
children whose parents have a high propensity of divorce but
(ii) college attendance, and (iii) college completion. Notes: Sample restricted remain married attend college (SI Appendix, Table S3).
to children who were at least 18 y old in 2012 and whose parents were In these analyses, we present simple results pertaining to the
married at the time of their birth. Parental divorce is measured as divorce heterogeneous effects of parental divorce on children’s outcomes
that occurred when children were 0–17 y old. Propensity scores were esti- as a function of the estimated propensity of divorce under
mated by a logit regression model of parental divorce on the set of pre- ignorability. They are informative descriptive results in their own
divorce covariates. Analytic sample (n = 5,176) is further restricted to age 19 right (20). If ignorability is true, we may interpret the pattern in
and above for college attendance (n = 4,982), and age 23 and above for
college completion (n = 3,901).
the effect of divorce as a function of the likelihood, or pro-
pensity, of disruption. However, if ignorability does not hold,
such that we have heterogeneous responses to latent determi-
to preserve cases at the tails of the propensity distribution where nants of divorce, the same results are still interpretable because
selection bias is most likely to occur, we construct three pro- they indicate variation in effects of parental divorce by the latent
pensity score strata. Families in which divorce is most likely unobserved parental resistance to divorce, a consideration
(stratum 3) have the most disadvantaged socioeconomic and missed in a critique of this approach (21). That is, we assume that
family well-being attributes (SI Appendix, Table S3). As we see lower observed propensity for divorce is associated with lower
from SI Appendix, Table S3, the estimated propensity score re- unobserved resistance to divorce, with lower resistance meaning
mains unbalanced according to the normalized differences in that parents choose divorce despite potential negative effects for
means (19). Given the coarseness of the strata, we did not expect children’s well-being.
balance. We adjust for the propensity score in all our models.
Very few individual covariates contain significant differences Sensitivity Analyses. In these analyses, we invoked the ignorability
across strata; if we further adjust for selected covariates with assumption. Whether this assumption is reasonable is a sub-
significant differences, our results remain substantively simi- stantive rather than a methodological issue, which depends upon
lar to those we present here, differing by no more than 1–2 the quality of the exogenous covariates in capturing potential
percentage points.
As we expect, given the matching–smoothing results in Fig. 2,
we find no significant effects for children who have a high pro-
pensity for parental divorce (stratum 3). We find significant ef-
fects for children who have a low propensity and midpropensity
for parental divorce (i.e., strata 1 and 2), with the largest effects
observed among children with the lowest propensity (although
estimated relatively imprecisely). Among children with a low
propensity for parental divorce, we observe a 6% lower level of
high school completion (81% predicted value among children of
divorced parents relative to 86% among children of nondivorced
parents, with the propensity held at the median), a 12% lower
level of college attendance (54% relative to 66%), and a 15%
lower level of college completion (21% relative to 36%). Among
children with a moderate propensity of parental divorce, we
observe a 4% lower level of high school completion, and a 7%
lower level of college attendance and completion. High school
completion point estimates are in fact similar for children across
the propensity for parental divorce, although imprecise for high- Fig. 3. Heterogeneous effects of parental divorce on children’s educational
propensity children, while college attendance and completion attainment: by propensity score strata. Notes: Sample restricted to children
rates markedly differ. who were at least 18 y old in 2012 and whose parents were married at the
We find larger effects for children who have a low propensity time of their birth. Parental divorce is measured as divorce that occurred
when children were 0–17 y old. Estimates are based on linear probability
for divorce than for the full sample (reported in Fig. 1), a con-
models. Adjusted models control for propensity of parental divorce and
sequence of overlooking cross-strata heterogeneity. Typically children’s age in 2012 (estimates not shown). Propensity scores were esti-
reported average effects under an assumption of effect homo- mated by a logit regression model of parental divorce on the set of pre-
geneity are weighted toward high-propensity children and ob- divorce covariates. Analytic sample (n = 5,176) is further restricted to age 19
scure larger effects for low-propensity children. This pattern of and above for college attendance (n = 4,982), and age 23 and above for
effect heterogeneity may help explain results suggesting smaller college completion (n = 3,901).

7268 | www.pnas.org/cgi/doi/10.1073/pnas.1813049116 Brand et al.

A B higher ðλ ≤ − 10Þ. Results for college attainment are less sensi-
tive to confounding. For college attendance, effects for children
with a low propensity are reduced to nonsignificance when γ ≥ 30
and λ ≤ − 20; for college completion, effects remain significant
for every value we consider. Effects on college are quite robust
for children with a moderate propensity as well, reaching non-
significance when γ ≥ 30 and λ ≤ − 10 or γ ≥ 20 and λ ≤ − 20. As
we note above, values greater than 10% for γ or λ for any particular
confounder are unlikely, lending confidence in our main findings of
the effects of divorce on college attainment for children with a low-
to-moderate propensity.

C D Discussion
Children whose parents divorce tend to have lower levels of
educational attainment than children whose parents stay mar-
ried. With careful attention to the assumptions needed to esti-
mate effects, we assess whether the impact of parental divorce
varies across families with varying likelihoods of divorce. Our
approach yields comprehensible and noteworthy results. Effects
of parental divorce on children’s educational attainment vary
inversely with the likelihood of divorce. We find significant ef-
fects of divorce on children’s educational success among those
with a low-to-moderate likelihood of parental divorce. For them,

SOCIAL SCIENCES
educational attainment rates are generally high, yet significantly
Fig. 4. Depiction of treatment effect heterogeneity by the propensity for differ depending upon whether or not their parents divorce,
parental divorce (P) and unobserved resistance to divorce (U) for (A) all units particularly for college attendance and completion. Parental di-
under the ignorability assumption, (B) treated units under the ignorability
vorce may trigger an acute sense of deprivation among these
assumption, (C) all units, and (D) treated units. Notes: A darker color indi-
cates a larger treatment effect. Data from ref. 22.
relatively advantaged children, whose peers tend to be likewise
advantaged and for whom family instability is uncommon and
comes as a shock. Conversely, we find no significant effect of
selection bias. We recognize that even with a rich set of pre- divorce on children’s education among those who have a high
treatment covariates, potential confounders remain (e.g., un- likelihood of parental divorce. Educational attainment rates
observed paternal characteristics). We address the possibility of among children whose parents have a high probability of divorce
unobserved confounding for heterogeneous effects with sensi- are relatively low, and these rates are roughly the same whether
tivity analyses. We assess a range of values of bias that may be or not parents divorce. Families prone to disruption have high
produced by an unobserved confounder (SI Appendix, Table S4). levels of socioeconomic hardship and/or a context in which
The effect reaches nonsignificance when the unobserved con- family shocks and economic distress are normative. That is, for
founder has a strong effect on children’s education (γ) and/or a these children, parental divorce is but one of many disadvan-
large difference between children of divorced and those of taged socioeconomic and family events faced during childhood,
nondivorced parents (λ). Suppose, for example, that father’s full- rendering the effects of any particular event less disruptive and
less severe.
time employment status, unobserved in our data, enhances levels
Divorce is a highly selective process; we cannot plausibly ac-
of education and is lower among fathers who get divorced (20).
count for all of the factors that influence both parents’ likelihood
When λ equals −10%, we assume that the prevalence of fathers
of divorce and children’s educational outcomes. One key ad-
having been full-time employed is 10% lower in the divorced
vantage and primary motivation for our focus on treatment effect
group than in the nondivorced group; when γ equals 10%, we heterogeneity by the propensity score is the heightened recog-
assume that children whose fathers are full-time employed have nition of potential violations of the assumption that we ade-
a 10% advantage in completing high school (or attending or quately adjust for all potential confounding factors. A researcher
completing college) over children whose fathers are not full-time can begin with such an assumption to carry out meaningful
employed (all else being held equal). Most of the family back- analyses without necessarily committing to the validity of the
ground, socioeconomic, psychosocial, and family formation and assumption (19–21). Indeed, even when unobserved selectivity is
well-being indicators differ by less than 5%; a few exceptions present, it is informative to understand variation in effects along
include maternal college completion (i.e., differs by 11%) and the propensity score (22). Our analyses yield an important pat-
parental arguing about money (i.e., differs by 12%) among those tern of effect heterogeneity by the estimated propensity of pa-
with a low propensity for divorce. We would thus not expect rental divorce based on observed covariates. If we accept the
many unobserved factors represented by λ to exceed 10%. assumption that we have accounted for all confounding factors,
To assess whether unobserved differences of the assumed the results suggest larger effects among children with a lower
magnitude would render estimated effects to nonsignificance, we likelihood of parental divorce. If we do not accept this assump-
let the values of λ range from −20 to −10% and the values of γ tion, we can nevertheless interpret the findings to reflect dif-
range from −40 to 40%. The sensitivity results when λ and γ are ferential unobserved selectivity of parental divorce: our results
negative are the same as those when γ and λ are positive, so there then reveal an association between lower resistance to divorce
is no loss of information by not including a positive range for λ. and larger effects of divorce. That is, given an observed low
We observe that the effect of divorce on high school completion likelihood of divorce, a divorce nonetheless can occur when
for children with a low and moderate propensity for parental unobserved characteristics render some parents less resistant
divorce is reduced to nonsignificance when unobserved differ- to divorce than others with similar observed characteristics.
ences between divorce and nondivorce are 10% or higher Lending confidence to our substantive interpretation, sensi-
ðγ ≥ 10Þ and the prevalence of the unobserved factor differs be- tivity analyses indicate that our main empirical findings are
tween children whose parents do and do not divorce is 10% or highly robust to confounding.

Brand et al. PNAS | April 9, 2019 | vol. 116 | no. 15 | 7269

 Children do not respond uniformly to family disruption. A outcomes for any individual, the individual-level causal effect as defined in
question of sociological importance is whether variations in their Eq. 1 is unidentifiable. Researchers are constrained to estimate the average
treatment effect (ATE), defined here as the overall average difference in
responses can be detected with characteristics that predict pa-
outcomes between children whose parents did and did not divorce:
rental divorce. This paper set out to answer this research ques-
tion and has yielded a clear answer. We describe important ATE = EðYð1Þ − Yð0ÞÞ. [2]
variation in the disruptive effect of parental divorce by the pre-
dicted likelihood of divorce based on observed characteristics, To address concerns of selection bias, our analytical approach begins with the
ranging from significant effects among children whose parents estimation of the propensity for parental divorce (P) based on observed
covariates (X):
are unlikely to divorce to no effects among children whose
parents are likely to divorce. While the effect of divorce is P = PðDi = 1jXi Þ. [3]
seemingly greatest among more advantaged children who may
not anticipate disruption, this is not to say that we should shift Under the ignorability assumption, conditioning on the propensity score is
attention away from children who expect disadvantage. It is as sufficient as conditioning on the full array of covariates X for the esti-
mation of treatment effects (24, 25). Departing from most previous research
telling that the educational attainment among these children is
on parental divorce effects on children, our approach necessitates that we
unaffected by parental divorce. Social discourse and policy explicitly model parental divorce as a first step. Average treatment effects
aimed at promoting martial stability among disadvantaged fam- conditional on the observed propensity for parental divorce take the
ilies, without attending to socioeconomic and family conditions following form:
in which adverse events are expected, are misguided.
ATEp = EðYð1Þ − Yð0ÞjP = pÞ. [4]
Materials and Methods
We estimate a series of linear probability models of the effects of parental
Data. The NLSY is a nationally representative sample of 12,686 respondents
divorce on children’s high school completion, college attendance, and col-
who were 14–22 y old when first surveyed in 1979. These individuals were
lege completion as follows:
interviewed annually through 1994 and biennially thereafter. In 1986, the
National Longitudinal Survey began a separate survey of the children of Yi = α + β 1 D i + β 2 P i + « i . [5]
NLSY women, the NLSCM. Data have been collected every 2 y since 1986. As
of 2012, the 6,283 women of the NLSY were 47–54 y old and had given birth Given our primary concern with the marginal effects of parental divorce on
to about 11,500 children. We link data on women from the NLSY with data education, the LPM is an appropriate modeling specification (26).
on children from the NLSCM (n = 11,512 children and n = 4,931 mothers) and
treat children as our units of analysis. We constructed measures of whether Estimating Heterogeneous Treatment Effects. We adopt two approaches for
and when a child (0–17 y old) experienced a parental divorce using NLSCM- estimating effect heterogeneity under the ignorability assumption. First, we
provided month and year of birth for children and NLSY-provided marriage use a matching–smoothing method consisting of the following steps (20, 21):
start and end dates for parents. We restrict our sample to 8,319 children who (i) estimate propensity scores for all units; (ii) match treated units to un-
were born to 3,940 married mothers. This restriction focuses on a relatively treated units with a matching algorithm; (iii) plot the observed difference in
homogenous population of children who were all at risk for a parental di- a pair between a treated unit and an untreated unit against a continuous
vorce from the time of birth. We further restricted the sample to those who representation of the propensity score; and (iv) use a local polynomial model
were at least 18 y old by 2012 (n = 7,258 children). Over a third of the sample to smooth the variation in matched differences to obtain the pattern of
(n = 2,420 children) experienced a parental divorce over the course of treatment effect heterogeneity as a function of the propensity score. Sec-
childhood. Our sample is further restricted to full data on educational at- ond, if there appears to be effect heterogeneity that could be sufficiently
tainment for models of high school completion (n = 5,176), and to age 19 captured by discrete strata of the estimated propensity score, we estimate
and above for models of college attendance (n = 4,982), and age 23 and stratum-specific effects. The number of strata we construct depends upon
above for models of college completion (n = 3,901). the shape of the nonparametric response function. Using this approach, we
Covariates used to construct the propensity of parental divorce are de- define the stratum-specific conditional average treatment effect as follows:
scribed in SI Appendix, Table S1. Missing values for the covariates were
imputed based on predivorce characteristics. Allowing our treatment to ATEs,p = EðYð1Þ − Yð0ÞjS = s, P = pÞ, [6]
occur anytime between a child’s birth and age 17 limits our pretreatment
where S = {1, 2, . . . s} indicates the stratum of the estimated propensity score.
covariates to those at the time of the child’s birth, which does not allow for
We estimate linear probability models of the form described in Eq. 5 sepa-
the adjustment of time-varying confounders. Still, as the dissolution pro-
rately by propensity score strata. We do not highlight other important axes
cess is likely to begin well before any formal separation is observed (23),
of variation in effects beyond the likelihood of parental divorce, such as
too much precision in the window of observation may lead to conditioning
gender, race, or timing of divorce, although we do so in other work and
on endogenous variables that amplify bias in estimating the effects of an
observe consistent patterns (27). We contend that focusing on the in-
impending divorce.
teraction of parental divorce with the estimated likelihood of divorce ad-
The average age at the time of parental divorce is roughly 6–7 y across the vances the existing literature on marital disruption in a theoretically sugges-
propensity for divorce. The narrow gap in the age of children at the time of tive way, as elucidated above.
divorce across strata allows us to eliminate the possibility that the timing of
As there is no a priori reason why the ignorability assumption holds true,
divorce in children’s lives drives differences in estimated effects by the we also acknowledge the presence of unobserved factors that affect the
propensity of divorce. We note, however, that the estimated propensity of likelihood of divorce. Of these unobserved factors, some are systematic,
divorce is not entirely uncorrelated with the hazard rate of divorce. The reflecting parents’ unwillingness, or resistance, to divorce. We denote the
duration of marriages is shorter among those with a high propensity for unobserved resistance to divorce as U. For example, parents’ resistance to
divorce because these parents have a shorter gap between marriage and divorce may be partly affected by their concern that children’s future out-
birth of the child (SI Appendix, Table S3). Nevertheless, the difference in comes will be negatively affected by a disruption. We describe the latent
marriage duration among divorced parents across the propensity for divorce
divorce function D* ð · Þ as follows:
is, somewhat surprisingly, small.
D* = P − U, [7]
Estimating Treatment Effects. For focal child i, the treatment effect (TE) of
parental divorce is defined as the difference between the two potential where P is the propensity of divorce based on observed covariates (26, 27),
outcomes in the treated (i.e., divorced parents) and untreated (i.e., non- and U is resistance to divorce, distributed between 0 and 1. Parents divorce
divorced parents) states (D = 1, 0): when D* ð · Þ exceeds 0:


TEi = Yi ð1Þ − Yi ð0Þ. [1] 1 if D* ≥ 0,
D= [8]
0 otherwise.
That is, we ask whether a child whose parents divorced had different out-
comes than he or she otherwise would have had if his or her parents had not In this model, we allow for the presence of U that affects not only divorce but
divorced. Given the impossibility of observing both treated and untreated also children’s attainment subsequent to divorce. In general, the treatment

7270 | www.pnas.org/cgi/doi/10.1073/pnas.1813049116 Brand et al.

effect varies both by P and by U. The ignorability assumption for estimating effects are sensitive to unobserved confounding covariates by quantifying
heterogeneous treatment effects refers to the special case where the treat- how the results obtained under the ignorability assumption would change if
ment effect varies by P and not by U, that is, U being ignorable. we relaxed the assumption. A standard approach is the calculation of a bias
We propose a simple model for parental divorce that incorporates un- factor (29, 30). The sensitivity of the estimated effects to unobserved
observed response heterogeneity (24, 28). We define the effect of parental treatment–outcome confounding can be assessed by subtracting the bias
divorce on children’s outcomes to be a function of both Pand U, that is, the factor from the point estimate and confidence interval of the treatment
marginal treatment effect (MTE): effect obtained under ignorability. The bias term is equal to the product of
two (stratum-specific) parameters:
MTE = EðYð1Þ − Yð0ÞjP = p, U = uÞ. [9]
BS = γ S λS , [12]
In Fig. 4, we depict alternative ways in which we can interpret treatment
effect heterogeneity. The darker shaded regions indicate a larger treatment where
effect magnitude (i.e., larger negative effects of parental divorce on child-
ren’s outcomes). In Fig. 4 A and B, we assume ignorability but allow for di- γ S = EðYjU = 1, D = d, S = s, P = pÞ − EðYjU = 0, D = d, S = s, P = pÞ [13]
vorce effect heterogeneity by P. Under this assumption, Eq. 9 is reduced to
the following: and

ATEp = TTp = EðYð1Þ − Yð0ÞjP = pÞ, [10] λS = PrðU = 1jD = 1, S = s, P = pÞ − PrðU = 1jD = 0, S = s, P = pÞ. [14]

That is, γ is the mean difference in children’s education associated with a

where ATEp refers to the average treatment effect and TTp refers to the
unit change in an unobserved binary confounder, U, and λ is the mean
treatment effect on the treated, both assumed to vary by the propensity for
difference in the unobserved confounder between the children of divorced
divorce P but not by the unobserved resistance to divorce U.
and nondivorced parents, both conditional on the estimated propensity for
In Fig. 4 C and D, we consider the general case of Eq. 9 and allow MTE to
divorce and propensity strata.
be a function of both P and U. Fig. 4 A and C depicts effects for all units. In
this case, we cannot estimate ATEp. However, if we have treatment het-
erogeneity bias but not baseline heterogeneity bias, the conventional Data Sharing. All data used here are publicly available at the Bureau of Labor
method for estimating heterogeneous treatment effects by the propensity Statistics National Longitudinal Surveys (https://www.bls.gov/nls/nlsy79.htm).

SOCIAL SCIENCES
score under ignorability still yields valid average treatment effects for the Code for data analysis is archived on Open Science Framework (https://osf.io/
treated (20, 24). That is, we are able to identify TTp as follows: dvgwu/).

Zp ACKNOWLEDGMENTS. We thank Xiang Zhou for valuable consultation and

1 Elizabeth Thomson for useful comments on a prior version of this manu-
TTp = EðYð1Þ − Yð0ÞjP = p, D = 1Þ = MTEðP = p, U = uÞdu, [11]
p script. We benefited from facilities and resources provided by the California
0
Center for Population Research at University of California, Los Angeles, and
Office of Population Research at Princeton University, which receive core
the integration of MTE over U = u given the propensity score P = p. Note that the
support (Grants P2C-HD041022 and P2C-HD047879) from the Eunice Kennedy
integral of [11] is systematically correlated with the propensity score P: lower
Shriver National Institute of Child Health and Human Development. National
observed propensity for divorce is associated with lower unobserved resistance Institutes of Health Grant R01 HD07460301A1 provided financial support for
to divorce, and thus more density in the integral for lower levels of U. In Fig. 4D, this research. Versions of this article were presented at the Center for Re-
we illustrate this relationship among treated units, that is, the subpopulation for search on Inequality and the Life Course at Yale University, Population Stud-
whichP > U. We note the high correlation between P and U among treated units ies Center at the University of Michigan, Department of Sociology at
in Fig. 4D: At low values of P, the estimated effect includes proportionally more Stanford University, Population Studies Center at the University of Pennsyl-
children whose parents have low values of resistance U; at high values of P, the vania, Office of Population Research at Princeton University, Center for Pop-
estimated effect includes more variation with respect to U, and thus propor- ulation and Development Studies at Harvard University, Center for
Demographic and Social Analysis at the University of California, Irvine, In-
tionally more children whose parents have high values of U.
ternational Sociological Association Research Committee on Social Stratifi-
cation and Mobility (RC28) 2017 Spring Meeting, and Population Association
Conducting Sensitivity Analyses. Sensitivity analyses provide a general of America 2017 Annual Meeting. The ideas expressed herein are those of
framework for investigating the extent to which the estimated treatment the authors.

1. McLanahan S, Tach L, Schneider D (2013) The causal effects of father absence. Annu 15. Heckman J (2005) The scientific model of causality. Sociol Methodol 35:1–98.
Rev Sociol 39:399–427. 16. Xie Y (2013) Population heterogeneity and causal inference. Proc Natl Acad Sci USA
2. Bernardi F, Boertien D (2016) Understanding heterogeneity in the effects of parental 110:6262–6268.
separation on educational attainment in Britain: Do children from lower educational 17. Brand J, Simon Thomas J (2013) Causal effect heterogeneity. Handbook of Causal
backgrounds have less to lose? Eur Sociol Rev 32:807–819. Analysis for Social Research, ed Morgan SL (Springer, New York), pp 189–214.
3. Bernardi F, Radl J (2014) Parental separation, social origin, and educational attain- 18. Xie Y, Brand JE, Jann B (2012) Estimating heterogeneous treatment effects with
ment: The long-term consequences of divorce for children. Demogr Res 30:1653–1680. observational data. Sociol Methodol 42:314–347.
4. Martin MA (2012) Family structure and the intergenerational transmission of edu- 19. Imbens GW, Rubin DB (2015) Causal Inference for Statistics, Social, and Biomedical
cational advantage. Soc Sci Res 41:33–47. Science (Cambridge Univ Press, Cambridge, UK).
5. Jaffee SR, Moffitt TE, Caspi A, Taylor A (2003) Life with (or without) father: The 20. Killewald A (2016) Money, work, and marital stability: Assessing change in the gen-
benefits of living with two biological parents depend on the father’s antisocial be-
dered determinants of divorce. Am Sociol Rev 81:696–719.
havior. Child Dev 74:109–126.
21. Breen R, Choi S, Holm A (2015) Heterogeneous causal effects and sample selection
6. Musick K, Meier A (2010) Are both parents always better than one? Parental conflict
bias. Sociol Sci 2:351–369.
and young adult wellbeing. Soc Sci Res 39:814–830.
22. Zhou X, Xie Y (December 13, 2018) Marginal treatment effects from a propensity
7. Amato P (2000) The consequences of divorce for adults and children. J Marriage Fam
score perspective. J Polit Econ, 10.1086/702172.
62:1269–1287.
23. Furstenberg F, Kiernan K (2001) Delayed parental divorce: How much do children
8. Thomson E, McLanahan SS (2012) Reflections on “Family structure and child well-
benefit? J Marriage Fam 63:446–457.
being: Economic resources vs. parental socialization.” Soc Forces 91:45–53.
24. Rosenbaum P, Rubin D (1983) The central role of the propensity score in observational
9. Amato PR (2001) Children of divorce in the 1990s: An update of the Amato and Keith
(1991) meta-analysis. J Fam Psychol 15:355–370. studies for causal effects. Biometrika 70:41–55.
10. Bumpass L, Lu HH (2000) Trends in cohabitation and implications for children s family 25. Rosenbaum P, Rubin D (1984) Reducing bias in observational studies using sub-
contexts in the United States. Popul Stud (Camb) 54:29–41. classification on the propensity score. J Am Stat Assoc 79:516–524.
11. McLoyd V, Cauce A, Takeuchi D, Wilson L (2000) Marital processes and parental socializa- 26. Angrist J, Pischke J-S (2009) Mostly Harmless Econometrics (Princeton Univ Press, Princeton).
tion in families of color: A decade review of research. J Marriage Fam 62:1070–1093. 27. Brand JE, Moore R, Song X, Xie Y (2019) Why does parental divorce lower children’s
12. Brand JE, Thomas JS (2014) Job displacement among single mothers: Effects on educational attainment? A causal mediation analysis. Sociological Science, in press.
children’s outcomes in young adulthood. AJS 119:955–1001. 28. Heckman J, Urzua S, Vytlacil D (2006) Understanding instrumental variables in models
13. Cherlin A (2004) The deinstitutionalization of American marriage. J Marriage Fam 66:848–861. with essential heterogeneity. Rev Econ Stat 88:389–432.
14. McDermott R, Fowler J, Christakis N (2013) Breaking up is hard to do, unless everyone 29. Gangl M (2013) Partial identification and sensitivity analysis. Handbook of Causal
else is doing it too: Social network effects on divorce in a longitudinal sample. Soc Analysis for Social Research, ed Morgan SL (Springer, New York), pp 377–402.
Forces 92:491–519. 30. VanderWeele T (2015) Explanations in Causal Inference (Oxford Univ Press, Oxford).

Brand et al. PNAS | April 9, 2019 | vol. 116 | no. 15 | 7271