Child Support and Educational Outcomes - Evidence from the
British Household Panel Survey
Ian Walker
University of Warwick and Institute for Fiscal Studies
and
Yu Zhu
University of Kent and Centre for the Economics of Education
Keywords: parental separation,
educational outcomes
JEL Code:
parental
incomes,
child
support,
D13, D31, J12, J13, J22
Abstract
It is widely accepted that Child support (CS) has played a positive role in reducing
child poverty among non-intact families. But little research has been done on the role of CS
on the real educational outcomes of the children concerned.
Using a sample of dependent children in non-intact families from the British
Household Panel Survey (BHPS), we find that CS received has an effect which is at least 10
times as large as that associated with variations in total household net income for two key
educational outcomes, namely school leaving at the age of 16 and attaining 5 or more good
GCSEs. This result is robust with respect to adding controls for characteristics of the child
and the custodial mother.
We then instrument average CS receipt using the retrospective information on the
mother’s fertility, relationship and unemployment, to investigate the causality of the
relationship. We find that the strong CS effect found in probit models carries through when
we allow CS, and other variables, to be endogenous. Moreover, applying the predicted CS,
from the first-stage coefficients from IV model for non-intact children, does not seem
to affect the educational outcomes for children living in intact families. This natural
“non-experiment” lends further support to the causal interpretation of the CS effect.
1
I. Introduction
The impact of parental separation on children has been a longstanding concern
of social scientists. While there has been a general consensus that parental separation
is associated with adverse outcomes for children (for surveys, see e.g. Amato and
Keith (1991), Haveman and Wolfe (1995) and Amato (2001)), the extent to which this
correlation is causal is far from clear. Causality becomes difficult to infer if there are
important omitted variables that are likely to be correlated with separation. Despite
the considerable evidence that income does affect child outcomes (see e.g. Dahl and
Lochner (2005), Plug and Vijverberg (2003, 2005), Chevalier et al. (2005) and
Francesconi et al. (2005)), many studies typically omit income while even fewer
studies have attempted to separate out the effects of one parent (mostly the father)
leaving on the outcomes for the children, from the effects of that parent's money
leaving.
Child support (CS) is the monetary transfer made by the non-custodial parent
(usually the father) to the custodial parent (usually the mother), for the care and
support of children of a relationship that has broken down. In most countries there is a
system of CS that both reduces the financial effects of separation on children and
raises the costs of separation to the non-custodial parent1.
It is widely believed that CS has played a positive role in reducing child
poverty among non-intact families (see e.g. Meyer and Hu (1999))2. Despite the
attention given to child poverty because of its apparent adverse effects on outcomes
for children, and to the role that CS plays in reducing child poverty, little research has
been done on the role of CS on the real educational outcomes of the children
concerned. This vacuum is particularly troublesome at a time when the system of CS
in the UK is about to be reformed in a way which seems likely to reduce CS
1
See Cancian et al (2003) for US evidence and González (2005) for evidence from across 16 countries.
2
CS not only affects child outcomes through a direct income effect, it may also cause indirect incentive
effects on maternal quality selection, fertility decisions, prenatal investments, and birth outcomes.
Walker and Zhu (2006) find that an increase in CS liabilities arising from the introduction of complex
CS rules in 1993 significantly reduced the risk of partnership dissolution for couples in the UK. Aizer
and McLanahan (2006) find that stricter state CS enforcement leads to fewer out-of-wedlock births and
an increase in both the average education of mothers and investment of prenatal care in the US.
2
payments3 unless compliance can be radically improved by moving to a system that
will entail a great deal more idiosyncratic variation in CS.
Notable exceptions are Knox (1996) and Argys et al. (1998), both of which are
based on the National Longitudinal Survey of Youth (NLSY). The sample in Knox
(1996) contains relatively young mothers aged between 23 and 31 and
disproportionately from lower socioeconomic status and minority families. Knox
shows that the positive effect of CS on achievement test scores for 5-8 year olds in
single-parent families is one order of magnitude higher than that from other income
sources. In contrast, while increases in overall family income improves the level of
cognitive stimulation in children’s home environments, CS does not appear to have
any additional effect over and above other forms of income. Moreover, the CS
coefficient in the instrumental variables model using state average CS levels and other
indicators of local economic environment remains statistically significant at the 10%
level. Knox also shows that CS levels predicted by economic conditions and state
average CS levels do not help explain variations in achievement test scores for
children living with both parents.
Argys et al. (1998) revisit the NLSY data a few years later, as the sample
becomes more representative of children who will ever be born to NLSY mothers.
Using this more representative sample, they find that CS coefficient is no longer
statistically significant under the Knox (1996) specification. Argys et al. show that it
is important to control for race and out-of-wedlock births and to distinguish between
cooperative and non-cooperative awards. The additional positive effect of CS on
children’s cognitive outcomes is only found for blacks in the divorced/separated
sample and for whites in the non-marital birth sample. Instrumenting CS and total
family income using local economic and demographic indicators as well as policy
variables relating to divorce, CS and welfare generosity of the state, the authors find
that the effects of CS persist, but only for blacks in the divorced sample.
Researchers have hypothesized the following mechanisms through which CS
may have beneficial effects on children over and above income from other sources
(Knox (1996), Aizer and McLanahan (2006)):
3
The system of CS in the UK was reviewed by David Henshaw for the DWP in July 2006. His
recommendations can be found at http://www.dwp.gov.uk/childsupport/henshaw_report.asp.
3
1) A “labelling effect”, i.e. CS may improve children’s outcomes more than other
sources of income, if mothers feel obliged to spend it directly on children. Del
Boca and Flinn (1994) find that coefficients associated with CS and alimony
income are significantly higher than those for other income in Engel curves for
expenditures on child-specific goods. Kooreman (2000) presents empirical
evidence that the marginal propensity to consume child clothing out of
exogenous child benefits is much larger than that out of other income sources
for households with one child in the Netherland. He interprets this finding as
evidence of the labelling effect of a child benefit system.
2) Improved family dynamics. CS may affect relationships between the noncustodial father and the custodial mother and between the absent father and the
child differently than income from other sources. The effect of increased
visitation, theoretically ambiguous a priori (e.g. Ermisch (2005), Del Boca and
Ribero (2001)). Both Knox (1996) and Argys et al. (1998) have allowed for
contact of the child with the absent father and found little evidence that it
matters.
3) Absence of any adversary effects that are often associated with other income
sources on child outcomes. It is possible that other types of income may bring
consequences for families that CS does not, i.e. stigma and work disincentives
associated with benefit income. For instance, maternal employment might
have negative effects through poor childcare (Gregg et al. (2005), Rhum
(2004)) while income from step-fathers might have separate negative effect.
4) A signal effect (reverse causality). Aughinbaugh (2001) finds that measures of
child achievement have a significant positive effect on both the possibility of
any CS being paid and the amount paid. She suggests that the results imply
that custodial mothers invest more in the child as a signal to the absent father
in order to secure future CS payments.
However, a plausible alternative to the above explanations is that the
relationship between CS receipts and child’s educational outcomes might reflect a
selection effect (unmeasured heterogeneity among families) rather than a causal
effect. For instance, payment of child support by an absent father might be correlated
with his commitment to the well-being of the child. From a policy point of view, it is
extremely important to be able to discriminate between these two explanations.
4
Our empirical work is based on the British Household Panel Survey (BHPS).
Unlike the NLSY which is overrepresented by younger mothers, our sample is
representative of all non-intact families. Using a sample of dependent children under
16 in non-intact families whose real educational outcomes are observable, we find
that CS received has an effect which is at least 10 times as large as that of total
household net income for two key educational outcomes, namely school leaving at the
age of 16 and attaining 5 or more GCSEs4. This result is robust with respect to adding
controls for characteristics of the child and the custodial mother5.
We then instrument current CS receipt using retrospective information on
mother’s fertility, relationship and employment before the birth of the child, to
investigate the causality of the relationship. We find that the strong CS effect found in
probit models carries through to the IV modelling. This strongly suggests that the
apparent simple correlation between CS and children’s educational outcomes is causal
rather than (entirely) driven by selection by some important unobservables that are
associated with CS compliance.
II. Data
The British Household Panel Survey (BHPS) is a nationally representative
sample of some 10,000 original sample members (OSMs) recruited in 1991. These
OSMs and all adult members of their families are interviewed in each following year.
The BHPS collects information on household organisation, housing, employment,
education, health and incomes in all waves. It is a particularly rich dataset with
lifetime histories of marriage, cohabitation, and fertility and employment transitions.
This allows us to match all children to their natural parents and establish the time of
departure of any absent parents, even for partnerships dissolved before the first wave.
We focus on school leaving and attainment of 5+ GCSE6s by the age of 16 for
all children in non-intact families in this paper. Figure 1 reports the two outcomes and
4
These are arguably the two most important measures of educational achievements in the UK: the
former captures the dropout rate at the minimum school leaving age while the latter is a key quality
indicator of the standard the pupil has achieved.
5
Moreover, allowing for endogeneity of separation arising from the 1993 CS reform and non-random
attrition of absent fathers makes little difference to the strong effect of CS received on both outcomes.
6
The General Certificate of Secondary Education (GCSE) is the principal means of assessing pupil
attainment at the end of compulsory secondary education since 1988. The GCSE examinations are set
according to nationally agreed criteria; with a quality assurance framework ensuring standards are
maintained
between
awarding
bodies
and
syllabuses
year
on
year
(http://www.dfes.gov.uk/qualifications/mainSection.cfm?sId=1).
5
CS receipts by family types, where non-intact families are further divided into two
groups depending on whether or not parents were living together when the child was
born (denoted as separated and single mothers in Table 1). A child is defined as living
in a non-intact family, if he or she does not live with both natural parents at any point
in time until he/she reaches the age of sixteen. Note that the way the family type is
defined means that children living in the same family can be classified into different
categories. It is apparent that the outcomes and CS receipts differ substantially across
intact and non-intact families, and between children whose parents separated and
children born out-of-wedlock within non-intact families7. For instance, the dropout
rates at 16 is below 18% for children living with both natural parents, as opposed to
28% for children whose parents have separated and a staggering 38% for children
living in non-intact families since birth. In contrast, the probability of attaining 5+
good GCSEs for children born-out-wedlock is only about half of that for children
living in intact families, with children in separated families in between.
There are 1540 distinct youths whose age 16 educational outcomes are
observable in the first fourteen waves of the BHPS, of which 414 (or 26.9%) live in
non-intact families. Table 1 reports summary statistics by family types which
confirms the importance of distinguishing between in-wedlock and out-of-wedlock
births. While over half of all separated mothers have received any CS, only less than
one-in-three single mothers do so. Conditional on receiving any CS, separated mother
received on average 51.9 pounds per week, almost twice as much as the 27.0 pounds
received by single mothers.
Table 1 reports child and mother characteristics which will be controlled for in
the regression analysis later. As might be expected, there is little difference in the
chance that the youth concerned is a boy or a twin across family types. On the other
hand, the differences in the average number of dependent children in the sample
period and birth order might reflect complex interactions of fertility and repartnership.
Recent studies suggest that family size and birth order are potentially important for
children’s educational attainment (see e.g. Booth and Kee (2005) and Lundberg
(2005)). Single mothers have significantly lower education than either separated or
intact mothers, while there is nothing to distinguish between the latter two groups.
7
The sample size does not allow us to further distinguish between (previously) married and cohabiting
couples, who are treated symmetrically for social security and child support purposes in the UK.
6
Table 1 also summarizes the retrospective information on the mother’s fertility,
relationship and employment history up to the point the child concerned was born and
her own experience of parental separation when she was sixteen. It turns out that nonintact mothers, and single mothers in particular, are more likely to start a relationship
and fertility at an early age, less likely to have any work experience and more likely to
have experience parental separation herself. Our identification strategy relies on these
lifetime event which predate the birth of the child.
0
.2
.4
.6
Figure 1: Educational Outcomes and CS receipts by Family Type
Intact
Father at birth
School Leaving at 16
Ever Received CS
No Father at birth
5+ Good GCSEs
Table 1: Summary Statistics by Family Types
Family Type
School leaving at 16 (%)
5+ Good GCSEs (%)
Proportion of time mother repartnered (%)
Ever Repartnered (%)
Ever received CS
CS Received conditional on receiving (£/week)
Total Net Income (£/week)
Child’s characteristics
Child being boy (%)
Number of kids in the family
Child being twin (%)
1st natural child of mother
2nd natural child of mother
3rd natural child of mother
4th or higher order natural child of mother
Intact
17.9
57.5
478.4
Separated
mothers
28.1
44.0
38.0
51.4
50.2
51.9
312.7
Single
mothers
37.9
29.9
36.9
54.0
31.0
27.0
258.2
50.7
1.9
2.3
43.4
37.6
14.0
5.0
53.5
1.9
1.8
43.4
37.3
14.4
4.9
55.2
2.4
3.4
78.2
11.5
6.9
3.5
7
Mother’s characteristics
Mother’s age left school
Mother non-white (%)
Retrospective information of the mother:
Mother’s age at 1st birth
Mother’s age at 1st relationship
Mother ever worked before birth of child
Mother not living with both parents when 16 (%)
Proportion of time child not living with dad (%)
Number of waves observed in sample
Obs
17.4
4.7
17.5
4.0
16.7
5.7
24.9
21.6
86.3
14.9
6.2
1126
23.0
20.2
84.1
19.0
42.9
5.4
327
20.5
17.6
60.9
36.8
100.0
6.8
87
Note: CS and income are in Jan 2006 prices.
III. Analysis and Results
1) Probit Models
We focus on the impact of CS on child educational outcomes in this paper.
Following Knox (1996) and Argys et al (1998), our main specification contains both
the mean CS receipts and mean total net family income in non-intact years during the
sample period8. Hence the coefficient of CS measures the extent to which CS has
larger effects than those from all income sources. However, unlike Argys et al (1998),
we do not control for pre-separation family income, as we only observe one third of
the separations taking place in the sample period for the non-intact families.
Table 2 presents the marginal effects for the probit models of both outcomes
for the sample of non-intact families. Model 1 is the simplest model specification
which only controls for repartnership, CS and total net income, all averaged over nonintact years in the sample period. While repartnerships are associated with increases
in family income, they are also likely to have an independent effect on child
outcomes. Model 2 adds such child characteristics as gender, number of dependent
children in the family, a twin indicator and dummies for birth order of the child.
Finally, Models 3 controls for two important characteristics of the mother: her age left
full-time education and an indicator for being non-white.
It is clear that of the three key variables of interest; only CS is significant in
explaining the differences in child educational outcomes. An increase in CS receipt
from zero to the mean level of £51.9 (conditional on receiving) will decrease dropout
rates by 16.5 percentage points and increase GCSE pass rates by 11.9 percentage
points respectively, large enough to offset the difference in educational achievements
8
Table 1 shows that this window of observation is 5.4 years for separated mothers and 6.8 years for
single mothers on average.
8
between intact and separated families. Moreover, the size of the CS effects is
remarkably robust with respect to the successive inclusion of child and mother
characteristics. If the observed correlation between CS and educational outcomes are
driven by a selection election, we would expect the size of the CS coefficients to
decrease in absolute value as we add more controls. Hence we interpret the robustness
of the CS effect in the simple probit models as evidence against the pure selection
story.
Table 2:
Marginal Effects (Robust standard errors9 are given in brackets)
Left school at 16 = 1
Mother
Repartnered
CS Receipt
(£100/week)
Total Net Income
(£100/week)
Child
Characteristics
Mother
characteristics
N
χ2(d.f.)
Log likelihood
Attained 5+ GCSEs = 1
Model 1
-0.007
(0.070)
-0.302
(0.100)
-0.015
(0.021)
No
Model 2
-0.052
(0.069)
-0.311
(0.094)
-0.006
(0.020)
Yes
Model 3
-0.053
(0.066)
-0.318
(0.096)
0.003
(0.018)
Yes
Model 1
-0.056
(0.081)
0.241
(0.096)
0.038
(0.021)
No
Model 2
-0.018
(0.080)
0.228
(0.105)
0.028
(0.020)
Yes
Model 3
-0.016
(0.079)
0.229
(0.103)
0.023
(0.020)
Yes
No
No
Yes
No
No
Yes
414
12.71
(3)
-243.85
405
41.91
(8)
-225.91
405
52.32
(10)
-216.47
414
14.20
(3)
-265.01
414
33.71
(9)
-250.25
414
38.34
(11)
-248.65
Notes: Child characteristics include gender, number of dependent children, dummy for twins, dummies for order
of births. Mother characteristics include age left full-time education and dummy for being non-white. Bold figures
indicate statistical significance at the 5% level.
2) Instrumental Variable Estimation
Child support estimates presented in Table 2 will suffer from the endogeneity
bias if the unobserved heterogeneity across households affect both CS receipts (and
payments) and children’s outcomes. Besides, measurement errors in CS and total net
income arising from misreporting and aggregating will also lead to biased results in
probit models. The standard approach to deal with endogeneity and errors-in-variables
is two-stage least squares.
9
Throughout the paper we report robust standard errors which account for possible hetroskedasticity
caused by the inclusion of siblings. Our sample size does not allow us to fully exploit the differences
between siblings (see e.g. Bjorklund and Sundstrom (2007)).
9
We allow total family income to be endogenous as well in our main
specification since CS is its component. We also treat proportion of time in
repartnerships as endogenous.
In the absence of convincing natural experiments we base our identification
strategy on lagged values. This is not entirely satisfactory and we therefore spend
some time establishing the validity of our choices. The set of instruments we use is
based on events that took place before the birth of the child, i.e. retrospective
information on mother’s fertility, relationship and employment. To maximize
efficiency, we also carry out statistical tests which checks for redundancy of
(combinations of) instruments. The preferred parsimonious specification includes the
following five excluded variables: an indicator for in-wedlock births; a quadratic in
the log of mother’s age at first birth; a quadratic in the log of mother’s age at first
relationship; an indicator for partnership still at risk by 1993 (i.e. Wave 3 of BHPS),
the year the CS reform was introduced10; and an indicator for mother ever worked
before the birth of the child.
Table 3A presents the second stage IV estimates for the main model
specification which controls for child and mother characteristics as outlined in Model
3 in Table 2. As in the probit models, neither repartnership nor total net income has a
significant effect on the two educational outcomes. On the other hand, the CS effect
found in the probit models remains statistically significant for both outcomes. While
the CS effect has increased substantially in size in comparison to the probit
specification, so has the standard error11. Indeed, given the large standard errors, one
can not reject the null hypothesis that the coefficients are the same across the two
different specifications. In interpreting these findings, more emphasis should be
placed on the direction and statistical significance rather than the exact size of the
coefficients for CS and total net income. It is also worth noting that the IV models
easily pass both the Anderson IV relevance test and the Hansen over-identification
test of all instruments at any conventional level of significance.
10
There was almost no CS paid pre 1993. The idea of including this dummy was to allow for
endogenous separation in response to the 1993 reform.
11
This is a common feature of the IV approach in general, given the trade-off between consistency and
efficiency. Moreover, least square estimate has a persistent bias towards zero when the regressor
concerned is measured with error. This is known as the attenuation effect (see e.g. Greene (2000)).
However, it could also be consistent with a Local Average Treatment Effect (LATE) or a credit
constraint story.
10
Table 3B presents the first stage results for the excluded variables used as
instruments. Note that the three endogenous variables are identified by different sets
of instruments which ensure that the predicted variables should retain some
independent variation in the second stage. Moreover, we can reject the joint
insignificance of the 7 IVs at the 1% level for all three endogenous variables.
Table 3A
IV Estimates Second Stages (robust s.e.’s in brackets), N=414
Outcomes
Mother
Repartnered
CS Receipt (£100/week)
Total Net Income (£100/week)
R-squared
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
-0.184
(0.236)
-0.567
(0.231)
0.015
(0.073)
0.323
5+ GCSEs
-0.035
(0.257)
0.687
(0.258)
0.030
(0.081)
0.382
17.363
(5)
0.004
LR
6.234
(4)
0.182
3.170
(4)
0.530
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
Table 3B
First Stage IV Results (robust s.e.’s in brackets), N=414
Endogenous
variables
Log mother’s age at 1st
birth
Square of log mother’s
age at 1st birth
In-wedlock birth
Log mother’s age at 1st
relationship
Square of log mother’s
age at 1st relationship
Partnership still at risk
by 1993
Mother Ever worked
before birth of child
Shea’s Partial R2
P-value
(joint
significance of IVs)
Repartnership
CS Receipt
(£100/week)
Total Net Income
(£100/week)
2.117
(4.304)
-0.366
(0.670)
0.119
(0.071)
0.057
(0.284)
0.000
(0.090)
-0.199
(0.060)
-0.178
(0.067)
0.073
0.0001
-2.134
(5.293)
0.447
(0.834)
0.095
(0.059)
-0.323
(0.230)
0.112
(0.073)
-0.125
(0.069)
-0.052
(0.061)
0.061
0.0002
38.113
(13.995)
-5.492
(2.198)
0.074
(0.241)
1.365
(1.068)
-0.415
(0.338)
-0.557
(0.252)
-0.264
(0.226)
0.047
0.0059
Note: Bold figures indicate statistical significance at the 5% level. Only excluded instruments are reported.
While CS is highly endogenous under all specifications, a formal test fails to
reject the exogeneity of repartnership and total net income. Tables A3A and A3B in
11
the Appendix present two-stage least squares estimates when repartnership and total
net income are treated as exogenous and redundant instruments are dropped as a result
of further tests. It is remarkable that the CS coefficients remain largely unchanged and
retain statistical significance. However, one should bear in mind that the exogeneity
of repartnership and total net income is only true in a statistical sense, i.e. one can not
reject the null that the covariance between the regressor and the error term is equal to
zero. Given the likelihood of measurement errors in these two variables, and the fact
that CS is a component of total net income, we would prefer the current main
specification which is less likely than the one in the appendix to be biased by
measurement error or endogeneity.
3) The “Two-parent” model: CS for Intact Children
So far we have only used the intact families in the descriptive analysis.
Following Knox (1996), we carry out an “natural experiment” by testing whether
predicted CS for children in intact families using the first-stage coefficients from the
IV model above help to explain the variations in educational achievements. If
predicted CS turns out to affect educational outcomes for intact children, then the
causal explanation found for non-intact children is called into question, as it might be
due to any remaining unobserved heterogeneity among families.
Table 4 presents the model estimates for the two-parent model. In contrast to
the results for non-intact families, the predicted CS has no effect at all on either
educational outcome. On the other hand, predicted total net income is statistically
significant for both outcomes12. This finding lends strong support to the causal
interpretation based on the non-intact family sample.
Table 4
Two-Parent Model Estimates (bootstrapped robust s.e.’s in brackets,
1000 repetitions), N=1126
Outcomes
Predicted Mother
Repartnered
Predicted CS Receipt
(£100/week)
Predicted Total Net Income
(£100/week)
Adj. R-squared
Leaving School at 16
0.141
(0.166)
-0.123
(0.086)
-0.129
(0.045)
0.107
5+ GCSEs
-0.391
(0.206)
0.108
(0.136)
0.235
(0.063)
0.382
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
12
It is also worth noting that the coefficient of the instrumented income is still smaller than the CS
effect for non-intact families, a result that would be consistent with a credit constraint explanation.
12
4) Sensitivity Analysis
One might be concerned with the use of an average measure of repartnership
and CS, which are averaged over the non-intact years in the sample period. Table 5
presents the second-stage IV estimates13 with binary measures for mother ever
repartnered and ever received CS during the sample period.
Table 5
IV Estimates Second Stages (robust s.e.’s in brackets), N=414
Outcomes
Mother Ever
Repartnered
Mother Ever Received CS
Total Net Income (£100/week)
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
-0.208
(0.332)
-0.853
(0.387)
0.119
(0.121)
5+ GCSEs
-0.040
(0.319)
0.824
(0.378)
-0.046
(0.116)
5.188
(5)
0.393
LR
3.497
(4)
0.478
4.257
(4)
0.373
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
It is striking that the CS effect remains positive for both outcomes, despite the
loss of efficiency through using dichotomous measures of repartnership and CS
receipt as indicated by the increases in standard errors and worse test statistic for the
Anderson IV relevance test14. The results support the hypothesis that improved family
dynamics associated with the receipts of CS (i.e. change at the extensive margin) are
more important than CS-induced changes in spending patterns (i.e. change at the
intensive margin) as far as children’s educational outcomes are concerned.
Table 6A and 6B present the second-stage IV estimates for non-intact families
with below- and above median family net incomes. For the poorest half of non-intact
families which is overrepresented by single parents and benefit recipients, the CS
effect is no longer statistically significant, although it still has the right sign and
similar magnitude. One possible explanation for this finding is that income has to be
above a critical threshold for CS to have a significant effect. The complex interaction
13
To save space, first stage IV estimates for the sensitivity analyses are not shown but will be available
from the authors upon request.
14
However, the test of joint significance of the excluded instruments can still be rejected at the level of
5% for all three endogenous variables.
13
between CS and social security benefits which withdraws out-of-work benefits pound
for pound might also be responsible. Besides, the measurement error problem might
also be more acute for benefit recipients.
In contrast, Table 6B clearly shows that our CS findings hold true for nonintact families with above median incomes, with very little change in the magnitude
of the effects.
Table 6A
IV Estimates Second Stages (robust s.e.’s in brackets), Below Median
Income, N=207
Outcomes
Mother
Repartnered
CS Receipt (£100/week)
Total Net Income (£100/week)
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
-0.195
(0.500)
-0.380
(0.444)
-0.198
(0.406)
5+ GCSEs
-0.155
(0.526)
0.418
(0.407)
0.564
(0.385)
7.756
(5)
0.170
LR
5.662
(4)
0.226
3.579
(4)
0.466
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
Table 6B
IV Estimates Second Stages (robust s.e.’s in brackets), Above Median
Income, N=207
Outcomes
Mother
Repartnered
CS Receipt (£100/week)
Total Net Income (£100/week)
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
-0.195
(0.232)
-0.527
(0.254)
-0.004
(0.090)
5+ GCSEs
0.087
(0.234)
0.731
(0.271)
-0.018
(0.088)
11.276
(5)
0.046
LR
3.412
(4)
0.491
4.628
(4)
0.328
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
14
Table 7
IV Estimates Second Stages (robust s.e.’s in brackets), Separated
Mothers Only, N=327
Outcomes
Mother
Repartnered
CS Receipt (£100/week)
Total Net Income (£100/week)
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
0.099
(0.255)
-0.296
(0.202)
-0.061
(0.064)
5+ GCSEs
-0.127
(0.283)
0.515
(0.264)
0.039
(0.077)
14.516
(5)
0.006
LR
6.417
(4)
0.093
4.164
(4)
0.244
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
Finally, Table 7 presents evidence for separated couples only. Although the
CS coefficient for school leaving at 16 dips below the 10% significance level, it
remains significant at just above the 5% significance level for attaining 5+ GCSEs.
This suggests that the causal effect we find earlier is not completely driven by the
inclusion of children born out-of-wedlock.
IV. Conclusion
Our findings suggest that CS payments have a beneficial effect on educational
outcomes that is well over and above income from other sources for children living in
non-intact families in the UK. This result is very robust with respect to the successive
addition of controls for child and mother characteristics, hence offers little support to
the selection explanation of the CS effect.
Moreover, our instrumental variable coefficients of CS based on retrospective
information on fertility, relationship and employment of the mother remain highly
significant. On the other hand, predicted CS using the first-stage coefficients from IV
model for non-intact children does not affect the educational outcomes for children
living in intact families. This natural experiment lends further support to the causal
explanation based on the instrumental variable model. We also carry out sensitivity
analysis to show that our results are robust with respect to a number of potential
pitfalls.
15
References
Aizer, A. and S. McLanahan (2006), “The Impact of Child Support Enforcement on
Fertility, Parental Investments, and Child Well-Being.” Journal of Human
Resources 41(1): 28–45.
Argys, L. M., Peters, H. E., Brooks-Gunn, J. and J. R. Smith (1998), “The Impact of
Child Support on Cognitive Outcomes of Young Children”, Demography,
153, pp. 159-173.
Amato, P.R. and B. Keith (1991), “Parental divorce and the well-being of children: A
meta-analysis”, Psychological Bulletin, 110, 26-47
Amato, P.R. (2001), “Children of divorce in the 1990s: an update of the Amato and
Keith (1991) meta-analysis”, Journal of Family Psychology, 153, 355-70.
Aughinbaugh, A. (2001), “Signals of Child Achievements as Determinants of Child
Support”, American Economic Review Papers and Proceedings, Vol 91, 2,
140-144.
Bjorklund, A. and M. Sundstrom (2007), “Parental separation and children’s
educational attainment: A siblings analysis on Swedish register data”,
Economica, forthcoming.
Booth, A.L. and H.J. Kee, 2005, “Birth Order Matters: The Effect of Family Size and
Birth Order on Educational Attainment”, IZA Discussion Paper 1713.
Cancian, M., D.R. Meyer, and H-O. Park (2003), “The Importance of Child Support
for Low-Income Families”, IRP Wisconsin mimeo.
Chevalier, A., C. Harmon, V. O’Sullivan and I. Walker (2005), “The Impact of
Parental Income and Education on the Schooling of their Children” IZA
Discussion Paper 1496.
Dahl, G. and L. Lochner (2005), “The Impact of Family Income on Child
Achievement”, NBER Working Paper 11279.
Del Boca, D. and C. Flinn (1994), "Expenditure Decisions of Divorced Mothers and
Income Composition," Journal of Human Resources Vol.29 pp. 742-761.
Del Boca, D. and R. Ribero (2001), “The Effect of Child-Support Policies on
Visitations and Transfers”, American Economic Review Papers and
Proceedings, Vol 91, 2, 130-134.
Ermisch, J. (2005), “New Families, Child Support and Divorced Fathers’ Contact
with their Children”, University of Essex, mimeo.
Francesconi, M., S.P.Jenkins and T. Siedler (2005), The Impact of Family Structure
during Childhood on Later Life, Anglo-German Foundation report.
González, L. (2005), “The Determinants of the Prevalence of Single Mothers: A
Cross-Country Analysis”, IZA DP 1667.
Gregg, P, Propper C., Washbrook E. and S. Burgess (2005), “The Effects of Early
Maternal Employment on Child Development in the UK”, Economic Journal,
Vol. 115, F48-80.
Greene, W., 2000, Econometric Analysis, 4th edition, Prentice Hall.
16
Haveman, R. and B. Wolfe (1995), “The determinants of children’s attainments: A
review of methods and findings”, Journal of Economic Literature, 33, 18291878.
Knox, V. W. 1996. "The Effects of Child Support Payments on Developmental
Outcomes for Children in Single-Mother Families." Journal of Human
Resources 31(4):816-840.
Kooreman, P. (2000), “The Labelling Effect of a Child Benefit System”, American
Economic Review, 90, 371-383.
Lundberg, S. (2005), “Men and islands: Dealing with the family in empirical labour
economics”, Labour Economics, 12, 591-612.
Meyer, D. R. and M-C. Hu (1999), “A Note on the Antipoverty Effectiveness of
Child Support among Mother-Only Families”, Journal of Human Resources,
34, 225-34.
Plug, E. and W. Vijverberg (2003), “Schooling, family background, and adoption: is it
nature or nurture?”, Journal of Political Economy, 111, 611-641.
Plug, E. and W. Vijverberg (2005), “Does family income matter for schooling
outcomes? Using adoptees as a natural experiment”, Economic Journal, 115,
879-906.
Rhum, C.J. (2004), “Parental employment and child cognitive development”, Journal
of Human Resources, 39, 155 92.
Walker, I. and Y. Zhu (2006), “Child Support and Partnership Dissolution”,
Economic Journal, 116, C93-C109.
17
Appendix:
Table A3A
IV Estimates Second Stages (robust s.e.’s in brackets), N=414
Outcomes
Mother
Repartnered
CS Receipt (£100/week)
Total Net Income (£100/week)
R-squared
Anderson canon corr
statistic Chi-sq (df)
P-value
Hansen J stat
Chi-sq (df)
P-value
Leaving School at 16
-0.183
(0.092)
-0.620
(0.229)
0.046
(0.027)
0.307
5+ GCSEs
0.140
(0.111)
0.801
(0.275)
-0.036
(0.034)
0.353
31.725
(5)
0.000
LR
6.092
(4)
0.192
2.893
(4)
0.576
Note: Other regressors include child and mother characteristics as outlined in Table 2. Bold figures indicate
statistical significance at the 5% level.
Table A3B
First Stage IV Results (robust s.e.’s in brackets), N=414
Endogenous variable
Mother Repartnered
Total Net Income (£100/week)
Log mother’s age at 1st birth
Square of log mother’s age at 1st birth
In-wedlock birth
Log mother’s age at 1st relationship
Square of log mother’s age at 1st relationship
Shea’s Partial R2
P-value (joint significance of IVs)
Repartnership
-0.253
(0.080)
0.089
(0.029)
-5.694
(4.724)
0.947
(0.750)
0.106
(0.057)
-0.406
(0.221)
0.141
(0.070)
0.074
0.0026
Note: Bold figures indicate statistical significance at the 5% level. All reported variables are excluded instruments
except for mother repartner and total net income
18