ArticlePDF Available

Abstract and Figures

We study the impact on children of increasing maternity leave benefits using a reform that increased paid and unpaid maternity leave in Norway in July 1977. Mothers giving birth before this date were eligible only for 12 weeks of unpaid leave, while those giving birth after were entitled to 4 months of paid leave and 12 months of unpaid leave. This increased time with the child led to a 2.7 percentage points decline in high school dropout and a 5% increase in wages at age 30. For mothers with low education we find a 5.2 percentage points decline in high school dropout and an 8% increase in wages at age 30. The effect is especially large for children of those mothers who, prior to the reform, would take very low levels of unpaid leave.
Content may be subject to copyright.
DISCUSSION PAPER SERIES
Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
A Flying Start? Maternity Leave Benefi ts and
Long Run Outcomes of Children
IZA DP No. 5793
June 2011
Pedro Carneiro
Katrine V. Løken
Kjell G. Salvanes
A Flying Start? Maternity Leave Benefits
and Long Run Outcomes of Children
Pedro Carneiro
University College London,
IFS, CeMMAP and IZA
Katrine V. Løken
University of Bergen
Kjell G. Salvanes
Norwegian School of Economics,
CEE, CESifo and IZA
Discussion Paper No. 5793
June 2011
IZA
P.O. Box 7240
53072 Bonn
Germany
Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: iza@iza.org
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 5793
June 2011
ABSTRACT
A Flying Start?
Maternity Leave Benefits and Long Run Outcomes of Children
*
We study the impact on children of increasing maternity leave benefits using a reform that
increased paid and unpaid maternity leave in Norway in July 1977. Mothers giving birth
before this date were eligible only for 12 weeks of unpaid leave, while those giving birth after
were entitled to 4 months of paid leave and 12 months of unpaid leave. This increased time
with the child led to a 2.7 percentage points decline in high school dropout and a 5%
increase in wages at age 30. For mothers with low education we find a 5.2 percentage points
decline in high school dropout and an 8% increase in wages at age 30. The effect is
especially large for children of those mothers who, prior to the reform, would take very low
levels of unpaid leave.
JEL Classification: J13, J18
Keywords: maternity leave, children’s outcomes
Corresponding author:
Kjell G. Salvanes
Department of Economics
Norwegian School of Economics
and Business Administration
Helleveien 30
N-5035 Bergen-Sandviken
Norway
E-mail: kjell.salvanes@nhh.no
*
This is a substantially revised and extended version of IZA DP No. 5362. We thank Gerard van den
Berg, Sandy Black, Richard Blundell, Christian Dustmann, Per-Anders Edin, Peter Fredriksson, Kjell-
Erik Lommerud, Uta Schoenberg, and seminar participants at Yale, University of Bergen, University of
Stavanger, University of Stockholm, Uppsala University, Tinbergen Institute, Norwegian School of
Economics, University College London, University of Texas, University of Rome, Universidade Nova
de Lisboa, Universidad Carlos III, University of Bologna, Colegio Alberto, ESSLE 2008, EEA 2008, and
SOLE 2009 for useful comments on the paper. Løken and Salvanes are thankful to the Research
Council of Norway for financial support. Carneiro gratefully acknowledges the financial support from
the Economic and Social Research Council for the ESRC Centre for Microdata Methods and Practice
(grant reference RES-589-28-0001), the support of the European Research Council through ERC-
2009-StG-240910-ROMETA and Orazio Attanasio’s ERC-2009 Advanced Grant 249612 “Exiting Long
Run Poverty: The Determinants of Asset Accumulation in Developing Countries”, and the hospitality of
the World Bank Research Group.
2
When it comes to paid maternity leave, the United States is in the postpartum dark ages.
One hundred and seventy-seven nations -- including Djibouti, Haiti and Afghanistan --
have laws on the books requiring that all women, and in some cases men, receive both
income and job-protected time off after the birth of a child. But here, the Family and
Medical Leave Act of 1993 provides only unpaid leave, and most working mothers don't
get to stay home with their newborns for the 12 weeks allowed by the law. Many aren't
covered by the FMLA; others can't afford to take unpaid time off. Some go back to work a
few weeks after giving birth, and some go back after mere days.
Sharon Lerner, Washington Post, June 13, 2010
Although the evidence on time use within families is limited and needs further study, the
increase in work from 1969 to 1996 has produced a reduction in the time available for
parents to spend with children. The increase in hours mothers spend in paid work,
combined with the shift toward single-parent families, resulted in families on average
experiencing a decrease of 22 hours a week (14 percent) in parental time available
outside of paid work that they could spend with their children.
Council of Economic Advisors (2009)
1. Introduction
There are huge disparities in maternity leave entitlements across different countries. On
one extreme, countries in Northern Europe (such as Sweden, Norway or Germany)
mandate very generous paid leave and long periods of job protection after birth. On the
other extreme there are a handful of countries such as the United States (US), which have
no paid leave mandate and offer little (if any) job protection (ILO, 1998).
These disparities were much smaller 30 to 40 years ago. In several countries, new
mothers had benefits similar to the ones currently in existence in the US, where the
federal mandate (which is adopted in almost all states) is 12 weeks unpaid leave for
women working in firms with 50 or more workers. One striking example, which is the
focus of our paper, is Norway. Prior to 1977, working mothers in Norway were entitled to
12 weeks unpaid leave, and to no paid leave. Currently the situation is very different: they
are entitled to a full year of paid leave and an additional year of job protection.
3
With the dramatic growth in female labor force participation, maternity leave
benefits have become more generous across the world. In the US, however, they have
remained fairly low, in spite of intense debate on this topic. A central question is whether
the absence of stronger maternity protection in the US is detrimental to child
development, or whether the high levels of benefits in Northern Europe are mainly
important for maternal health (and maternal welfare more generally), with little
consequence in children’s lives. This question is the focus of our paper.
Empirically, this is a notoriously difficult problem, as emphasized (for example)
by Bernal (2008) and Dustmann and Schönberg (2008) since mothers who spend more
time with their children after birth may have many unobservable attributes that affect
child development (or they use child care arrangements which are special in unobservable
dimensions). Furthermore, since additional time with children is generally associated
with less time at work and lower household income, it is difficult to isolate the two.
In our paper we address these empirical challenges, by studying a reform in
maternity leave benefits in Norway on long term outcomes of children, namely education
and earnings at age 30. The reform we analyze increased mandatory paid maternity leave
from 0 to 4 months and mandatory unpaid maternity leave from 3 to 12 months.
1
This new set of benefits applied to all eligible mothers having children after July
1
st
, 1977.
2
We estimate their long term impact on children using regression discontinuity,
comparing outcomes of children of eligible mothers born just after and just before this
particular date. We assess the importance of month of birth effects, and of any potential
1
This is equivalent to moving from the current level of maternity leave entitlements in the US to something
closer to Holland and several other countries in Southern and Central Europe.
2
Eligibility criteria, involving work requirements, are discussed below in detail. About 35% of women
giving birth in 1977 were ineligible for paid maternity leave benefits.
4
manipulation of the date of birth. We follow children to as late as 2007, when they are 30
years of age. We observe several long term outcomes, such as high school completion,
college attendance, and wages at age 30.
3
We begin with a very simple look at the data. Take individuals (and their mothers)
born only in two months of 1977: June (just before the reform was implemented) and July
(just after the reform). We can compare the outcomes of children in these two groups
(only for eligible mothers), by running a regression of the outcome of interest on an
indicator for being born in July. However, there may be differences in outcomes between
children born in these two months of 1977 for reasons unrelated to the reform, as
emphasized in a large literature on month of birth effects (Black, Devereux and Salvanes,
2008, present estimates for Norway). Therefore, it is important to use an earlier year,
prior to the implementation of the reform, as a comparison. We use data from 1975 to
estimate the difference in outcomes between children born in June and July prior to the
reform, and subtract it from the estimate of the effect of being born in July (vs. being
born in June) that we got from the 1977 data (a difference-in-differences estimator).
4
Table 1 presents estimates of the impact of the program using the single (first
column) and double differences (second column) estimators for a subset of the dependent
3
The appendix, Table A1, also shows IQ, height (males only), and teenage pregnancy (females only).
4
For the single difference we would run the following regression using data for children born in June and
July of 1977:
i
July
ii
uDY ++= *
βα
where
i
Y is the outcome of interest and
July
i
D is a dummy indicating whether an individual was born in
July.
β
measures the impact of benefiting of the reform on the outcome of interest, among children of
eligible mothers. For the difference in difference estimator, using data from children born in the months of
June of July of 1975 and 1977, we can run:
ii
July
i
July
iii
uDDDDY ++++=
19771977
***
βφγα
where
1977
i
D is a dummy indicating whether an individual was born in 1977. As before,
β
measures the
impact of benefiting of the reform on the outcome of interest, among children of eligible mothers.
5
variables we consider in the paper. Child outcomes are shown at the top: indicators for
whether a person is a high school dropout, whether she has ever attended college, and log
earnings at age 30. The results suggest that the reform had an impact on high school
dropout rates, and earnings at age 30, but not on college attendance, both in the single and
the double-difference specifications.
Then we examine two pre-birth maternal variables, which should not be affected
by the reform: years of education of the mother, and log annual income in 1975 (more
variables are shown in the appendix, Table A1). In both these dimensions, the set of
mothers giving birth in June of 1977 is similar to the set of mothers giving birth in July of
the same year (even when we use the differences in differences estimator).
Finally, we find no impact of the reform on maternal income right around the time
the mother gave birth (average log income in the year of birth and the year after b irth),
which means that the reform had no impact on unpaid leave.
5
We also look at maternal
labour supply and income 5 years after the birth of the child
6
, and see no statistically
significant effect of the reform on these variables, using both single (first column) or
double (second column) differences. This is why we argue that the main mechanism of
this reform was an increase in time with the child, with no short or long run consequences
on labour market outcomes.
In the rest of the paper we develop, expand and discuss these results in detail,
implementing a regression discontinuity estimator that uses information from children
5
Note that the small significant effect on income at year of birth in the first difference result is purely a
month effect, and a consequence of the fact that we only have annual (and not monthly) measures of
income for each mother. Mothers giving birth to the child later in the year have more months to work
before giving birth and therefore have a higher income during the year of birth. When controlling for this
problem using eligible mothers in 1975 there is no effect of the reform on income in the year of birth.
6
As opposed to more permanent effects of the reform on labour market outcomes of females, after
employers and mothers fully adjust their expectations and behaviours.
6
born in the other months of the year. The main patterns of table 1 survive a more
sophisticated estimation procedure. We will also examine a wider set of variables.
The literature on this topic is fairly wide, so we will not review it in detail. Good
reviews of the literature on maternal employment and child outcomes are available in
Blau and Currie (2006) and Bernal and Keane (2010). The Economic Journal featured a
recent symposium on this topic (Gregg and Waldfogel, 2005; Tanaka, 2005; Gregg,
Washbrook, Propper and Burgess, 2005). The literature is fairly inconclusive and plagued
with empirical problems, as these papers document. The Society for Research in Child
Development edited a recent volume on this topic (Brooks-Gunn, Han and Waldfogel,
2010) arguing that, at least for non-hispanic whites in the US, maternal employment in
the first year of life does not have particularly detrimental consequences on children
because its negative and positive aspects cancel each other out. But, as in most of the
literature, the authors caution against a causal interpretation of their estimates.
Recent papers directly examine maternity leave reforms. For the US, Rossin
(2011) studies the effect of the 1993 reform on children’s birth and infant health. She
finds support for some positive effects of the reform on children’s health outcomes. We
can also find three other empirical analyses of the effect of maternity leave reforms on
long term outcomes of children, using registry data with very large sample sizes for
Germany (Dustmann and Schönberg,, 2008), Denmark (Rasmussen, 2010), and Sweden
(Liu and Skans, 2010).
7
There are two important aspects of these papers relatively to the
literature described above: 1) they explore exogenous variation in maternity leave
resulting from legislative reforms to these benefits; 2) they are able to look at long run
7
We should also mention a set of recent papers studying Canadian reforms and focusing on short run
outcomes for children, by Baker and Milligan, (2008a, 2008b). These papers also find no significant effects
of the reform on children’s outcomes.
7
outcomes of children. Our data challenges the main result of these papers: that there is
little or no effect of maternity leave expansions on long run outcomes of children.
This is an important finding. We believe that there are two central aspects of our
study that distinguish it from the ones above and may explain our different results. First,
we consider a change in maternity leave entitlements when they were at a very low level,
similar to the US today. The papers we refer to consider expansions in maternity leave
from an already baseline level that is fairly generous. Even in Dustmann and Schönberg,
(2008), who study three different reforms in Germany, the earliest reform they consider is
an expansion from 2 to 6 months in paid maternity leave entitlements (the long term
outcomes considered in the study of that reform are wages at ages 24-26).
Second, we are able to look at education and labour market outcomes as late as
age 30. Other papers have examined earlier educational outcomes, or earlier labour
market outcomes. One problem with looking to early labour market outcomes is that
individuals’ careers may only stabilize much later.
8
In addition, our data lets us link mothers with their children which allows us to do
a rich analysis of impacts by subgroups of mothers; and it lets us construct good measures
of eligibility for the reform which is important since generally only a fraction of mothers
(those who are working a minimum amount of time) is eligible for these benefits.
9
8
In fact, we do not find any effect of the reform on earnings at ages 24 and 25.
9
One drawback of our data is that it does not contain direct measures of labour supply. This information is
not essential for estimating effects of the reform but it is useful to understand the mechanisms through
which it is operating. We do, however, observe total income in each year. There is no impact of the reform
on maternal income in 1977 and 1978. This means that the reform did not change the amount of unpaid
leave being taken by mothers giving birth after the reform. We do not consider the case that the reform had
no effect at all on leave taking behaviour, since it is highly unlikely. Below we present indirect evidence
suggesting that the new paid leave entitlement was fully taken-up by new mothers, and therefore the lack of
change in annual income is just a result of unchanged levels of unpaid leave. For example, when we
examine later reforms to maternity leave, for which we observe labour supply data, we see close to full
uptake of the new benefits. Therefore, we argue that the reform led to an increase in four extra months of
8
The paper proceeds as follows. Section 2 gives background information on
maternity leave legislation in Norway while Section 3 presents the empirical strategy.
Section 4 presents data and Sections 5 shows the results. Section 6 discusses (evidence
on) mechanisms by which the reform impacts child outcomes. Section 7 concludes.
2. Maternity Leave Reform and Institutional Background
2.1 Maternity Leave Reform
In 1956, maternity leave benefits became available to women in Norway through the
introduction of compulsory sickness insurance for all employees. Eligible mothers were
entitled to 12 weeks of essentially unpaid maternity leave. This is basically the same level
of benefits available for mothers in (nearly all states in) the US in 2011, provided that
they work in firms with 50 or more employees.
On July 1
st
, 1977, Norway saw the introduction of paid maternity leave and an
increase in unpaid leave, as illustrated in Figure 1.
10
With this reform, parents were given
the universal right to 18 weeks of paid leave with guaranteed job protection before and
after the birth of a child.
11
Maternity leave payments were equivalent to 18 weeks of pre-
leave actually taken by new mothers, without changing unpaid leave or maternal income. In addition, all of
the reforms to either paid or unpaid leave examined in the literature described above had important impacts
on the uptake of leave.
10
These changes were introduced together with a new law increasing workers rights (”Arbeidsmiljøloven”)
accepted June 3
rd
, 1977, by the Parliament and introduced July 1
st
, 1977 (see Prepositions, Ot.prp. nr. 71
and Innst.o. nr. 90). There were additional reforms after 1977. From 1987 and onwards the paid maternity
leave was extended almost yearly until 1993. From 1993 and up till now Norway has had the same paid
maternity leave of 42 weeks with 100% cover or 52 weeks with 80% cover. We have in this paper decided
to focus on the 1977 law for three reasons. It is a change in what we think is a critical period for the child,
for instance since breastfeeding is still an issue. It is easier to assess the first change in the law since the
latter reforms were anticipated to a larger degree. And, given that available adult data goes only up to 2007,
we have a much richer set of available outcomes for children born in 1977 than for those born later. We
leave the study of the other reforms for future work.
11
You could take a maximum of 12 weeks before the birth of the child; however most mothers worked
almost until day of birth as they wanted to save leave until after the child was born (Survey on fertility in
1977, Statistics Norway).
9
birth employment (i.e., 100% replacement rate). Of these 18 weeks, 6 could be taken by
the mother alone, while the rest could be shared between both parents. In practice, all
leave was almost exclusively taken by the mother (Rønsen and Sundström, 2002). In
addition, parents also got entitled to 1 year of unpaid job protection (on top of the 18 paid
and job-protected weeks of maternity leave).
Not all mothers were eligible to receive the new benefits, with eligibility
depending on their work and income history. Only women working at least 6 of the 10
months immediately prior to giving birth, and having more than 10000 NOK
12
of yearly
income, were eligible for leave and coverage.
Because of limitations in our data (we do not observe labour supply directly, and
we only have yearly income which includes wage income and benefits) we have to rely
on an imperfect measure of eligibility. In particular, we define eligible mothers as those
having at least 10000 NOK of salary in the calendar year before giving birth. Our use of
12 rather than 10 months of income to determine eligibility is likely to slightly overstate
the number of eligible mothers. We estimate that two thirds of all mothers giving birth in
Norway in 1977 were eligible for maternity leave benefits. We tried different alternative
definitions of eligibility, without significant changes in our empirical results.
Figure 2 shows the proportion of mothers who were eligible for maternity leave
entitlements from 1975–1979, by birth month of the child. Between 1975 and 1979 the
proportion of eligible mothers was always between 60% and 70%, and in 1977 it was
about 65%. Since we can only focus on eligible mothers in our analysis, this means that
our estimates ignore 35% of mothers and children giving birth in that year.
12
10000 NOK (USD 1725) refers to the lowest level of income providing pension points in the Norwegian
social security system in 1977.
10
In order to be able to identify the effects of the reform on children’s outcomes it is
crucial that mothers are not able to change their eligibility status immediately after the
reform is announced. Otherwise, the set of eligible mothers giving birth just before and
just after the reform would not be comparable. The maternity leave reform was
introduced during a big offensive from the sitting (very radical) parliament at the end of
its period. It is unlikely that it was expected since it came along with a lot of other
changes (unrelated to the maternity leave reform) and at the end of the legislative period.
The Government report became official on April 15
th
, 1977, and was approved on June
13
th,
1977
13
. This means that all women giving birth after the announcement of the law in
1977 were already pregnant when the law was introduced,
14
and because of the rule of
working 6 out of 10 months prior to giving birth, it was difficult for women to change
their status in the short term. We also checked national newspapers around 1976 and
1977 for news about the reform. We do not find any evidence that newspapers reported
anything on the reform before June 1977.
15
Therefore, it is plausible that eligibility status
is exogenous for mothers giving birth in 1977.
The 1970s in Norway was the decade of oil discovery, with increasing labour
force participation of women, and the implementation of several welfare reforms. We
have studied all possible laws and reforms occurring during that period that may have had
an impact on maternal and child outcomes. The only one we found was the abortion law
13
Propositions and regulations from the Government: Ot.prp nr. 61 and Innst.o. nr 61.
14
Possible effects on fertility will therefore not show up in the data before the beginning of 1978, at the
earliest. It is possible that mothers delivering close to July 1
st
, 1977, were able to delay their delivery. In
fact, Gans and Leigh, 2009 estimate that Australian mothers delayed child birth in response to a reform
changing incentives to fertility. Nevertheless, for the reform we study there are no significant differences
between the number of births occurring just before and just after the reform. This is shown in figure A1 in
the Appendix.
15
Verdens Gang June 30
th
,1977, Bergens Tiende June 27
th
,1977, June 30
th
,1977, Aftenposten June
30
th
,1977.
11
implemented January 1
st
, 1976. This law made it easier for women to have an abortion
within 12 weeks of conception. The first cohort to be affected by this reform is born
around July 1976. This possibly gives rise to a discontinuity in observed child outcomes
between June and July 1976 and hence we do not use 1976 as a comparison to 1977.
2.2 Institutional Background
At the time of the maternity leave reform in 1977, labour force participation for women
was relatively high in Norway. Figure 3 shows labour force participation in Norway
compared to the US from 1970 to 1990 (distinguishing Norwegian women who are
mothers from those who are not). In Norway, the labour force participation rate around
1977 was about 50 percent for married women, which are the most relevant group for our
study, and around 70 percent for non-married women. Labour force participation was
about the same in Norway and the US during the 1970s, but much higher in the former
than in the latter by 1990.
16
It is also relevant to look at the provision of public child care. In Figure 4 we
depict the development of day care coverage in Norway for children aged 0 to 2, in urban
and rural areas. In the mid 1970s, very few children aged 0 to 2 were in day care, and
there is very little difference in day care attendance between urban and rural areas (1%
vs. 0.5%). Although day care centres provided coverage for 15 percent for children aged
0 to 6 in 1977, the coverage for the first two years was very low, only 1–2 percent. This
means that the main alternative to maternal care in the early years of the child’s life was
informal care by nannies, grandparents or neighbours.
16
By 2008 the labour force participation rate in the US was around 65 percent (with small race differences).
This is comparable to the participation rate around the reform in Norway (OECD, 2008).
12
3. Empirical Strategy
Let )1(
i
y be the outcome for child i in the presence of the reform, and )0(
i
y be
the outcome for child i in the absence of the reform. Our main goal is to estimate the
average impact of the reform on the long term outcomes of children:
()
)0()1(
ii
yyE =
α
.
In order to estimate this parameter we compare children born just before and just
after the reform, which should be similar except for the fact that mothers of those in the
latter group benefit from the change in maternity leave entitlements taking place on July
1
st
, 1977. In particular, we use a regression discontinuity (RD) estimator to assess the
difference in outcomes between children born in June and July of that year.
For those women giving birth in 1977, eligibility for the new maternity leave
entitlements (
i
E ) is a deterministic function of month of birth ( )X
i
:
}cX{1E
ii
>
=
, (1)
where c is the cut-off point of July 1
st
, 1977. Therefore, all mothers giving birth to a child
after c potentially receive the treatment defined by new maternity leave entitlements,
while those giving birth before c are assigned to the control group. We use only eligible
mothers based in our main analysis as defined in Section 2.
17
The RD estimator for
α
is given by:
]|)0([]|)1([ cXycXy
iiiiRD
=
=
Ε=
α
. (2)
As in any RD estimator we are only able to identify a local effect for those born just
around the reform. However, this is one case where it is reasonable to conjecture that the
17
See Appendix, Table A2 for a comparison of results using the total versus the eligible sample.
13
effects of the reform do not vary substantially with month of birth, in which case
RD
α
would be a consistent estimator of
α
.
Assuming that ]|)1([ cXy
ii
=
Ε and ]|)0([ cXy
ii
=
are continuous in x
(continuity at x=c is all that is needed) we can estimate them as:
]|[lim]|)0([
]|[lim]|)1([
xXyEcXy
xXyEcXy
ii
cx
ii
ii
cx
ii
===Ε
=
=
=
Ε
Outcomes of interest for the child include dropping out of high school, college
attendance (both measured by age 30), and earnings at age 30 (in the appendix, we
examine also the probability of having a child before age 19 for women, and IQ and
height for men). Outcomes of interest for the mother include months of unpaid leave, and
employment and earnings 5 years after giving birth. These are mainly interesting because
we can check for changes in home environments, which can account for the effect of the
reform on child outcomes.
We estimate ]|[lim]|[lim xXyExXyE
ii
cx
ii
cx
RD
=
=
=
α
by taking the
difference between the boundary points of two regression functions of y on x: one for
eligibles (xc) and one for ineligibles (x>c). We estimate these regression functions with
local linear regression (LLR) as in Fan (1992), Hahn, Todd and Van der Klaauw (2001),
and Porter (2003). Hahn, et al. (2001) show that LLR outperforms general kernel
regression methods in terms of bias. Defining h as the bandwidth, we estimate (α, β, γ, τ):
2
1
,,,
))()((min
iiiii
i
N
i
EcXEcXy
h
cX
K
=
γτβη
γτβα
, (3)
RD
α
is estimated as
τ
α
ˆ
ˆ
=
RD
(4)
14
We use the triangle kernel which is shown to be boundary optimal (Cheng, Fan
and Marron, 1997). We obtain standard errors using the formulas in Porter (2003).
18
The
choice of bandwidth is important, as usual. In the main text we present results using a
bandwidth of 3, and in the Appendix we present further results using a bandwidth of 5.
19
It is possible that a simple comparison of outcomes for children born in different
months is contaminated by month of birth effects due, for instance, to the fact that the age
at which children start school depends on their month of birth and is potentially related to
adult education and earnings (see Black, et al., 2008, for evidence for Norway). In this
case
RD
α
converges to
Bith
λ
α
+ , where
Bith
λ
is a month of birth effect, which does not vary
across years. Therefore we combine RD with difference-in-differences (DD) by
constructing three types of control groups: one consists of children born in 1975 of
eligible mothers; another consists of children born in 1979 of eligible mothers; and
another consists of children born in 1977 of ineligible mothers.
We use the first one in our main specification, and the other two in robustness
checks (shown in the appendix).
20
We begin by estimating equation (3) for those born in
1975 and those born in 1977. Then we calculate:
BirthRDBirthRD
λ
α
τ
α
λ
τ
α
+
=
=
==
19771977,19751975,
ˆ
ˆ
;
ˆ
ˆ
18
We verify the results by using the paired-bootstrap percentile-T procedure with 2000 replications.
Cameron and Trivedi (2005), show that the bootstrap percentile-T procedure may outperform the analytical
standard errors. One reason for this might be the difficulty in estimating parts of the formulas from Porter.
From our results we do not see any significant difference between the two methods (if anything there are
slightly lower standard errors when using Porter), hence we will use the analytical formulas.
19
Using cross validation as in Imbens and Lemieux (2008) we get an optimal bandwidth of 3. However,
Ludwig and Miller (2007) point to different problems using cross validation. Therefore, we examine the
sensitivity of our results to different bandwidths.
20
As we argued earlier we cannot use 1976 because of a reform in the abortion system. For symmetry we
also try 1979 as our second control group and obtain very similar results. We also present figures in the
appendix using eligible mothers in 1974 as an additional robustness test.
15
Since there is no reform in 1975
1975,
ˆ
RD
α
should only capture month of birth effects (June
vs. July birth). On the other end,
1977,
ˆ
RD
α
confounds effects of the reform with potential
month of birth effects. Under the relatively mild assumptions that the two effects do not
interact, and that month of birth effects are the same (around July) for those born in 1975
and 1977, we can estimate the effect of the reform as
1975,1977,
ˆˆˆ
RDRDDDRD
α
α
α
=
.
21
We use the formulas in Porter (2003) for the standard errors of
1977,
ˆ
RD
α
and
1975,
ˆ
RD
α
. In order to get the standard errors for
DDRD
α
ˆ
, we assume that
1977,
ˆ
RD
α
and
1975,
ˆ
RD
α
are independent (since these are completely different cohorts of children). We
obtain similar results if instead we use the bootstrap, which relaxes independence.
Before we proceed to the next section it is important to clarify what questions we
can and cannot answer with this empirical strategy. We can answer questions about the
outcomes of children benefiting from different amounts of time with the mother early in
life, induced by changes in maternity leave entitlements. However, maternity leave
reform is about much more than that. For example, it may also affect fertility and labour
supply decisions in the medium run, but the full adjustment of these behaviours to the
new maternity leave regime is likely to happen slowly.
Therefore, we cannot fully learn about the outcomes of children living under
different maternity leave regimes, since this would require waiting for the full adjustment
of fertility and labour supply of women (and possibly their spouses). In fact, mothers of
children born in both June and July of 1977 are likely to engage in the similar
adjustments to fertility and labour supply in the medium run, especially if they are
21
It is useful to examine graphs comparing outcomes of eligible mothers in 1977 with those of eligible
mothers in 1975 and eligible mothers in 1979. The pre and post-reform trends are very similar.
16
considering having more children. What we can answer is a narrower question about the
importance of the time that mothers spend with their children in their first year of life,
which is the main difference in the early experiences of children born in June and July,
1977 (in the appendix we show that there are no differences in completed fertility and
labour supply between these two groups).
()
)0()1(
ii
yyE =
α
is an intent-to-treat estimate of the impact of being born in
the new maternity leave regime. In addition to this it would be important to estimate the
impact of the reform on the amount of time spent at home by mothers, which would give
us an idea of the intensity of the treatment. As mentioned before we do not have direct
measures of time worked each year in the data, but it is possible to infer this quantity
from information on annual income. We discuss this in the next section and in Section 6.
4. Data description
Our data source is the Norwegian Registry data maintained by Statistics Norway. It is a
linked administrative dataset that covers the population of Norwegians up to 2007 and is
a collection of different administrative registers providing information about month and
year of birth, educational attainment, labour market status, earnings, and a set of
demographic variables (age, gender) as well as information on families. To ensure that all
individuals studied went through the Norwegian educational system, we include only
individuals born in Norway. We are able to link individuals to their parents, and it is
possible to gather labour market information for both.
The main outcome variables we consider for children are dropout rates from high
school, college attendance and earnings at age 30. In terms of educational attainment, we
17
measure education at the oldest age possible for each individual,
i.e., in 2007.
22
High
School dropouts are defined as all children not obtaining a three year high school
diploma, and college attendance is defined from the annual education files identifying
whether a person ever started college. Earnings are measured as total gross pension-
qualifying earnings reported in the tax registry and are available from 1967 to 2007.
These are not top-coded and include labour earnings, taxable sick benefits,
unemployment benefits, and parental leave payments.
We also collect data on maternal income 2 and 5 years after taking birth. These
are useful to examine possible channels through which the maternity leave may affect
child outcomes, namely by promoting attachment of women to the labour market.
In the appendix we discuss the construction of additional outcome variables,
which we use in our paper but they are not part of the main analysis. These are IQ and
height (for males), teenage pregnancy (for females), place of residence, distance to
grandparents, part time work for mothers and completed fertility of mothers.
In order to construct unpaid leave we start by calculating a measure of pre-birth
monthly income by dividing 1976 earnings by 12. Then we calculate total earnings in
1977–1980, and divide them by 1976 monthly income, thereby obtaining a measure of
number of months of unpaid leave during the first 36 months after birth. For this
calculation to work, the assumption is that 1976 earnings are a good approximation for
maternal potential post-birth earnings (the earnings she would get had she not gone on
22
Our measure of child educational attainment is reported by the educational establishment directly to
Statistics Norway, thereby minimizing any measurement error due to misreporting. This educational
register started in 1970.
18
unpaid leave), adjusted for inflation.
23
We limit ourselves to a window of 36 months
because the further away we move from pre-birth earnings, the more likely earnings may
differ because of change of job, part time work, presence of new children, and other
factors unrelated to the 1977 reform.
24
We assume that paid leave has a take-up rate of
100% for those giving birth after July 1977. In Section 6 and the appendix we provide
evidence for our claim that there was full take-up of paid leave. We also argue that our
estimates of unpaid leave are reasonable. Furthermore, in section 6 and the appendix we
show that the estimated impact of the reform on unpaid leave is robust to whichever
measure of leave we consider.
5. Results
5.1 Descriptive statistics
We focus only on mothers who are eligible for the reform, and therefore it is important to
show how they compare to those who are not eligible. We saw from Figure 2 that the
proportion of mothers who are eligible for maternity leave entitlements was about 65% in
the year of the reform. This means that 35% of mothers and children giving birth in that
year are not accounted for in our estimates of the impact of the reform on child outcomes,
because the mother is not eligible for maternity leave. Interestingly, current labour force
participation rates in OECD countries are generally not much higher than 65%, except in
23
It is useful to illustrate with a specific example. If the child is born in June 1977 we subtract six months
of 1976 monthly earnings from 1977 earnings and compare the remaining earnings in 1977 and 1978 to the
1976 earnings. If the mother earns half of 1976 earnings in the twelve months after birth she has taken six
months of unpaid leave. If she earns nothing and takes all twelve months of leave we will continue and use
earnings in 1979 and 1980 to construct leave up to 36 months after birth.
24
However, remember that we will show that all these factors are the same for mothers giving birth before
and after the reform, so they will potentially only affect the estimate of the level of unpaid leave and not the
difference (effect of the reform).
19
the Scandinavian countries where they are often above 80%. Furthermore, roughly 25%
of working women in the OECD are working only part-time.
Table 2 displays the main characteristics of eligible mothers and their children
(born in 1977) as compared to those of ineligible mothers and their children. It is clear
that eligible mothers are more highly educated than ineligible mothers. They are also
more likely to be employed after birth than ineligible mothers, and as a consequence,
their income is higher during that period. Their income 2 years before giving birth is 9
times larger than that of ineligible mothers, presumably because many of the latter do not
work. Children of eligible mothers have lower high school dropout rates and higher
college attendance rates, however similar earnings at age 30. Eligible and non-eligible
mothers and their children are two very different groups. This means that we cannot
safely extrapolate our findings to the latter group of mothers and their children.
25
The average level of unpaid maternity leave taken at the time is quite high, even
for those mothers having children before the reform is implemented. For our preferred
measure, average unpaid leave is 8 months for those delivering their children before July
1977, and it barely changes for those delivering after this date. The 25
th
percentile is
about 2 months, and the 75
th
percentile is about 11 months. Any expansion in the time
mothers spend with their newborn children resulting from the reform is in addition to this
pre-existing level of leave. The fact that unpaid leave did not change in response to the
reform is robust to the measure of leave used, and depends solely on the fact that annual
income is similar for mothers giving birth before and after the reform date, which means
that both groups of mothers are taking the same amount of unpaid leave.
25
For the narrower question of whether maternity leave is important for children of those mothers affected
by the reform (eligible mothers) we have the right population.
20
Notice that, even if the reform leads to no change in family resources during the
initial period of the child’s life, it induces a slight change in the timing of these resources.
Paid leave allows mothers to receive benefits right after their child is born, whereas
unpaid leave does not. However, it is not likely that this change in the timing of benefits
dramatically impacts child outcomes, unless we are under an extreme case of credit
constraints. In order to investigate this further, in the appendix we present an analysis of
the effects of the reform for mothers with different levels of pre-reform income. Poorer
mothers are more likely to be credit constrained, so our idea is to use pre-reform income
as an indicator of the severity of such constraints (which we find to be unimportant).
Before proceeding to the results, we would like to check whether the treatment
and control groups are balanced in terms of the (pre-reform) characteristics we observe.
Imbalance may indicate a threat to the validity of our method since it would indicate the
possibility that a non-random set of mothers manipulate the date of birth of their children
(see Gans and Leigh, 2009). The various panels of Figure 5 show how observable pre-
reform characteristics of mothers vary with the month they gave birth in, and allow us to
check whether they are identical for mothers having children just before and just after the
reform. Maternal years of education, age at birth and income in 1975 are stable across
birth months and we see no discontinuity after July 1
st
, 1977. In addition, there is also no
discontinuity in the urban location of the parents in 1976 and the distance to grandparents
in 1980 (although this variable is only available in 1980). Moreover, Figure A1 in the
Appendix shows very similar numbers of births just before and after the reform was
implemented. In sum, selective manipulation of month of birth is not likely to be a
21
serious concern in our data. This is quite reasonable given that in 1977 (and even today) it
was not easy to delay childbirth much beyond the due date.
5.2 Children's outcomes
In table 3 we present estimates of the impact of the reform on a set of children’s
outcomes.
26
The first column shows the RD results while the second column presents the
DD results using the cohorts born in 1975 as a control group. In the first column we see a
negative effect of the reform of about 2 percentage points in children’s dropout rates,
however this variable is only significant at the ten percent level. When taking into
account potential month of birth effects in the DD specifications in column 2 we see an
increase in the effect to 2.7 percentage points (because the month of birth effect is
negative in 1975). We see the same pattern for college attendance: an increase of 3.6
percentage points, which is only significant in the DD specification. In addition we see a
positive effect on earnings at age 30 of 4.8 % which increases to 5.5 % in the DD
specification.
2728
In Figure 6 we present graphically the results corresponding to the
second column in Table 3 (Appendix Figure A2 shows the single difference results.)
29
We clearly see that the reform induced discontinuities (that do not occur in 1975) in
26
See Appendix, Table A3, for additional outcomes, namely IQ for males, and teenage pregnancy for
females, for which we have less robust results. See also Table A2, for a comparison of results using the
total sample. Those results compare well with the sample of eligible mothers, although they are weaker.
27
Interestingly, in the appendix, Table A3, there is also a positive effect on IQ. IQ scores are only available
for men, but due to the large sample sizes we can still get precise estimates of the effect on the reform on
IQ. The RD shows an effect of 0.11, or 5% of a standard deviation. This effect is around 0.24 in column 2
which is 12% of a standard deviation. Using estimates of the effect of IQ on wages from wage regressions
estimated on slightly older cohorts of individuals, this translates into more than a 1% in difference in
earnings as an adult. We do not see any effect of the reform on teenage pregnancy in any of the
specifications. In Table A4 in the Appendix we report results with a bandwidth of 5 corresponding to more
smoothing of the data. We see the same patterns in coefficients however the results are weaker, especially
for the RD results. This can be a feature of the possible effect of birth month on outcomes hence we will
focus on differences-in-differences for the rest of the paper.
28
See the end of Appendix A for additional robustness tests.
29
In addition we present DD figures using 1974 and 1979 as control groups in the Appendix, Figure A3
and A4. These figures show that results are very robust, the effects are even slightly higher and more
persistent than when using 1975.
22
dropout rates and earnings at age 30 as a function of month of birth.
30
We also see that
there are monthly trends for the different outcomes. The effects on dropout rates are
present for all birth months after the reform and for the most part this is also the case for
earnings. The effect on college attendance is not as robust. Therefore, most of the impact
of the reform seems to be at the low end of the education distribution, with treated
children dropping out less from high school and this show up in higher returns on
earnings at age 30.
31
6. Interpretation of empirical results and suggestive mechanisms
In the previous section we established that the maternity leave reform had a substantial
impact on schooling and earnings of children. In this section we attempt to understand the
mechanisms by which this happened, using limited information from the administrative
records we use. The results we present in this section are not individually decisive, but
together they tell a consistent story.
6.1 Time with the child
The main problem of our dataset is that it does not have a direct measure of maternal
labour supply nor of leave taking behaviour. So how can we be confident that the reform
is significantly affecting leave taking behaviour by mothers?
First, Rønsen and Sundström (1996) show that for the 1968-1988 mothers in
Norway, almost no one returned to work before 4 months after birth. Secondly, in a
survey conducted in 1977 on fertility behavior of women in Norway (Statistics Norway),
30
We also see less robust patterns for the outcomes relegated to the appendix, Figure A5: IQ, and teenage
pregnancy.
31
It is worthwhile pointing out that if we use earlier measures of earnings we cannot detect this effect. It is
important to wait until individuals have reached some maturity in the labour market.
23
60% of respondents answered that they thought mothers should stay home for the first 2
years after giving birth to a child. In addition, the coverage was 100%, which gives strong
incentives for full take up. Third, since we observe days of paid leave after 1992 we are
able to check to what extent eligible mothers take up this benefit, and how the take up
reacts to subsequent reforms in 1992 and 1993 (see Appendix, Figure A6, showing the
following description). Before the April 1992 reform, mothers are able to take 224 days
at full coverage or 280 days at 80% coverage. For mothers delivering children in March
of 1992, the average take up of paid leave was 250 days. After April 1992 there is an
increase in maternity leave entitlements to 245 days of full coverage or 310 days of 80%
coverage. We observe that average paid leave taken was 275 days for mothers of those
born in April 1992. This figure is slightly higher at 280 in March 1993, just before the
1993 reform which increased paid leave to 266 days of full coverage or 336 days of 80%
coverage. By April of 1993 average leave taken was almost 310 days. Given the high
levels of leave and strong reactions to reforms, it is reasonable to assume that the take up
of paid leave is close to 100%.
32
Therefore, we are confident that after the 1977 reform all mothers were taking 4
months of paid leave. So the follow-up question is: what was the change in unpaid leave
as a result of the reform? One way we can answer this question is by studying what
happened to maternal income before and after the reform.
33
An increase in maternal
income in the period right after birth may indicate a reduction in unpaid leave taken, and
the opposite could be inferred from a decrease in maternal income (perhaps in
32
We should also point out that the analyzes of other reforms in other countries for which there is data
available on labor supply of mothers all indicate a substantial increase in the amount of leave taken after
each reform.
33
Remember that all maternity benefits are part of our measure of income.
24
substitution of the additional paid leave mothers become entitled to). We examined
maternal income in the years surrounding the reform for those delivering children just
after and before the reform and we found no impact of the reform on these variables. This
is shown in table 4, and it indicates that there was no change in unpaid leave taken by
mothers. This is true independently of the measure of earnings we take: income in 1977,
average income between 1976 and 1978, or average income between 1975 and 1979.
34
This is true not only of the mean, but of the whole distribution of income.
In addition, as discussed above, using this data it is possible to predict how much
unpaid leave was taken by each mother, by comparing her usual earnings in a year with
no childbirth to earnings in a year (and subsequent years) with one.
We find no effects of the reform on the amount of unpaid leave taken by mothers
as shown in the first column of Table 5. This is not surprising since we emphasized
before that there is no change in average annual income for mothers giving birth just
before and just after the date of the reform, independently of the measure of earnings we
take.
In summary, this means that, whatever the measure of unpaid leave is, there is no
change in the amount of unpaid time taken off work for mothers giving birth before or
after the reform, otherwise there would be an increase in their income. Therefore, even if
our measure of unpaid leave is not exactly right, we can be confident that there is no large
change in unpaid leave as a result of the reform. We can rule out any responses that vary
more than within one month so this reform was mostly about more paid leave which since
it is fully covered means no effect on income. Even with no average response in unpaid
34
Note that the small significant effect on income at year of birth in the RD result is only a month effect of
giving birth to the child later in the year and have more months to work before giving birth. When
controlling for birth month using eligible mothers in 1975 there is no effect on income year of birth.
25
leave it is interesting to see if there are any effects across the distribution of unpaid leave.
In the Appendix, Figure A7, we see no such responses. We cannot rule out that not all
mothers took 4 months of paid leave, although the earlier evidence provided in this paper
shows that this was likely the case (Statistics Norway, fertility survey of 1977).
6.2 Maternal Labour Market Outcomes
It is possible that the reform increased labour market attachment of mothers. This is
because of the extensive job protection they became entitled to, which allowed them to
come back to their old job long after they gave birth. Therefore, it is conceivable that
children born in the post-reform period had better outcomes not only because they spent
more time with their mothers, but also because their mothers became more attached to the
labour market in the medium and long run, thereby being able to generate more income
but also spending more time at work.
Table 5 shows our main results. We do not find any long term effects of the
reform on mother’s employment two and five years after it took place, or on earnings
35
five years after. This supports the idea that our estimates of the impact of the reform on
children’s outcomes can be directly related to mother’s time investments in the child
during its first year of life.
In Figure 7 we present the differences in differences results of Table 5
graphically. The figures confirm the results of the table. There is no discontinuity in long
term labour market outcomes.
6.3 Maternal Education
35
We have also played around with mother’s earnings between one and ten years after birth and this gives
similar results of no long term effect on income.
26
We check whether the maternity leave extension had a different effect on mothers
with different educational backgrounds.
36
We split the sample in two; mothers with less
than 10 years of education versus mothers with 10 years or more of education. We see,
from the last two columns of Table 6, that the effects on mothers are very similar for the
two groups: there is no effect on unpaid leave and no significant effects on the long term
labour market outcomes. For children we see that the fall in dropout rate is 5.2 percentage
points for children of mothers with less than 10 years of education while it is around 2
percentage points for children of higher educated mothers. The pattern is similar for
earnings at age 30. However, none of these differences across maternal education groups
are significantly different. We can still take them as suggestive given their magnitude,
and the fact that the effect of the reform is larger at the bottom of the maternal education
distribution is consistent with the fact that the most robust effect of the reform is on high
school dropout rates, which is a fairly low qualification.
6.2 Results by quartiles of mother’s unpaid leave.
Table 7 presents results on mother’s and children’s outcomes by quartiles of unpaid
leave. In principle this variable should be affected by the reform and therefore we should
not condition on it. In practice, we saw that the reform has no effect on unpaid leave.
Furthermore, if the ranking of mothers in terms of unpaid leave does not depend on the
reform, we can interpret these estimates as the effects of the reform for mothers who
would take different levels of unpaid leave in the absence of the reform.
We see no effect on mother’s outcomes at any quartile.
37
This indicates a
substantial increase in mother’s time spent at home across the distribution of eligible
36
In the appendix, Table A5, we present results by distance to grandparents and centralization.
37
This happens because the distribution of annual income is roughly the same for mothers giving birth just
before and after the reform
27
mothers (since paid leave has increased for all of them). For children we see that the
effect on dropout rates is very large for the first and second quartiles, with 9 and 5
percentage points respectively, while we see no effect in the third and fourth quartiles.
This is also confirmed by the earnings results which suggest around 10 % higher earnings
in the first and second quartile and no effect in the last two quartiles.
Mothers in the first two quartiles have levels of unpaid leave much below the
average (0.4 and 5.1 months, respectively). The fact that it is for these mothers that we
see the largest effects on dropout rates and earnings (the outcomes for which our results
are the most robust) suggests that additional time with the child is mainly important
during the earliest months of the child’s life. It is possible that these differences do not
come entirely from increases in health (say, due to breastfeeding; see also the evidence
discussed in Appendix B). There may also be an impact on maternal-child attachment and
less stress in the home, leading to changes in personality traits that make these children
less likely to drop out of high school.
6.3 Any substantial differences in the impact of the reform according to other criteria?
We have checked and found no differences in the effect of the reform according to pre-
reform family income and the state of the local labour market at the time of birth.
38
In
contrast to maternal education, these are relatively short run measures of household
environments. Additional time with the child does not seem to be especially important for
dropout rates of children born in very poor households, unless they are also born in
households where mothers have low levels of maternal education.
Above we mentioned that the reform could also have an effect by shifting the
availability of income towards those months right after birth, even if there is no change in
38
See Appendix, Table A6, for results by quartiles of family income.
28
total income. If some households are severely credit constrained this may make a
difference to the child. According to our results, this is unlikely to be the case, if those
with low levels of pre-reform income are the most likely to be credit constrained.
In addition we studied completed fertility and marital stability of mothers to the
children affected by the reform. We see no effects on any of these outcomes when the
children are age 30 (see Appendix, Table A7).
We also analyzed the impact of the reform on older siblings (see Appendix, Table
A8). The fact that mothers spend additional time in the home could benefit other siblings
as well. However, this is not the case, which suggests that what drives the impact of the
reform is specific to the relationship between the mother and the newborn child (perhaps
because of a stronger attachment between the two, with benefits for mother and child). In
addition, we did not find any difference in dropout rates by gender of the child; although
the effect on wages at age 30 is driven by males (see Appendix, Table A9).
6.4 A simple model of the high school dropout decision
Finally, we studied the determinants of the dropout decision. We use it to understand the
impact of the reform relatively to that of other variables, and to understand how the
impact of other variables changes as a result of the reform.
We started by running a regression of whether an individual is a high school
dropout on years of mother’s education (measured in 1980), mother’s age at birth,
whether the mother is married (in 1980), family size, log of the present value of the sum
of mother’s and father’s income between the ages of 0 and 13, and whether the child was
born in an urban area. In addition, we included IQ and height, which means that we only
estimated this model for males (notice that throughout the paper we did not find
29
differential impacts of the reform by gender). We used a linear probability model on the
sample of all males born in 1975 or 1977 to a mother eligible to maternity leave (
i
denotes individual,
t denotes year of birth):
itititit
itititititit
UrbaneTotalIncomFamilySize
MarriedAgeatBirthsEducationMotherHeightAbilityDropout
εβββ
β
β
β
β
β
β
++++
+
+
+
++=
876
543210
'
Estimates from this model are shown in the first column of table 8. Dropout rates are
lower by: 6.6 percentage points (pp) for each additional ability point; 0.2 pp for each
centimetre in height; 1.3 pp for each year of maternal education; 0.3 pp for each year of
age at birth of the mother; 12 pp for having a married rather than an unmarried mother;
1.3 pp for a reduction of one in family size; 3.8 pp for a doubling of total maternal and
paternal income; and 1.6 pp for being in a rural rather than in an urban area. These are
substantial effects, and apart from the urban coefficient, they are largely unsurprising.
In order to understand how the reform affects the dropout decision we start by
adapting the empirical strategy laid out in section 4 to this parametric model. We add to
the regression of table 8 a parametric function of month of birth (MB, normalizing July =
0, so December = 5 and January = -6), a dummy for being born in 1977 (Y77), and a
dummy for being born in July (REFORM), to approximate the nonparametric regression
discontinuity estimator of section 4 with a parametric model:
ititititit
2
it10
ititit9it
2
it8itit7it6it
2
it5
itit4it3
2
it2it1it8it7it6
it5it4it3it2it10it
)REFORM*77Y()REFORM*77Y*MB(
)REFORM*77Y*MB()77Y*MB()77Y*MB(77Y)REFORM*MB(
)REFORM*MB(REFORMMBMBUrbaneTotalIncomFamilySize
MarriedAgeatBirthsEducation'MotherHeightAbilityDropout
ε+η+γ+
γ+γ+γ+γ+γ+
γ+γ+γ+γ+β+β+β+
β+
β
+
β
+
β+β+β=
The effect of the reform is given by
η.
30
Estimates of this model are shown in the second column of table 7. The effect of
the reform is a bit larger (5.5%) than in our original results, perhaps because we have
additional controls, or perhaps because of the parametric method.
Notice that we control for two variables that are possibly affected by the reform:
ability and height. Therefore, if anything this parametric model is understating the effect
of the reform. However, it is striking that the coefficients on these two variables are
essentially unchanged from column 1 to column 2 of this table. This says that even if
there is an effect of the reform on ability and height, it is not substantial enough to change
the coefficients on these variables. Furthermore, it says that the large effect of the reform
on dropout rates does not occur primarily through a change in IQ or height, but through a
change in another type of skill, perhaps a non-cognitive skill. It is not surprising that there
is no change in the coefficients in the other controls since they are orthogonal to month
and year of birth.
Finally, we interact
(Y77*REFORM) with all the controls (after demeaning,
denoted by D), so the reform can change the way the controls affect the dropout decision:
itititit8
ititit7ititit6
ititit5ititit4
ititit3ititit2
ititit1itit0itit
2
it10
ititit9it
2
it8itit7it6it
2
it5
itit4it3
2
it2it1it8it7it6
it5it4it3it2it10it
)DUrban*REFORM*77Y(
)meDTotalInco*REFORM*77Y()eDFamilySiz*REFORM*77Y(
)DMarried*REFORM*77Y()hDAgeatBirt*REFORM*77Y(
)sEducation'DMother*REFORM*77Y()DHeight*REFORM*77Y(
)DAbility*REFORM*77Y()REFORM*77Y()REFORM*77Y*MB(
)REFORM*77Y*MB()77Y*MB()77Y*
MB(77Y)REFORM*MB(
)REFORM*MB(REFORMMBMBUrbaneTotalIncomFamilySize
MarriedAgeatBirthsEducation'MotherHeightAbilityDropout
ε+η+
η+η+
η+η+
η+η+
η+η+γ+
γ+γ+γ+γ+γ+
γ+γ+γ+γ+β+β+β+
β+
β
+
β
+
β+β+β=
Since we demean the controls before interacting them with
(Y77*REFORM) we
can read the average effect of the reform from the coefficient on
(Y77*REFORM). The
impact of each control variable on dropping out of high school for those not affected by
31
the reform can be read from the coefficient on the controls. The impact of each variable
for those affected by the reform is obtained by adding the coefficient on the variable with
the coefficient on the interaction.
Results are displayed in the third column of table 8. Notice that, once again, there
is hardly any change in the effects of each of the control variables for those not benefiting
from the reform. When we look at those affected by the reform, here is little change on
the coefficients on ability, height, maternal education, and log total family income. There
are a few changes on the coefficients on maternal age at birth (amplifying its effect) and
maternal marital status (dampening the effect), and both remain statistically significant.
However, there is substantial dampening of the effects of family size and being born in an
urban area, which become insignificant for those benefiting from the reform.
Even though this is a reduced form model for the dropout decision, in interpreting
these results it is natural to think of returns and costs to high school graduation. Although
we can only speculate about it, we believe that it is unlikely that the reform is changing
much the returns to a high school diploma. These returns should be affected by most of
the control variables, especially ability and maternal education, and we see no general
pattern of interactions of the reform with all variables, let alone one these two in
particular. If we think about costs, we see the main impacts of the reform on urban status
and family size. Once again, we can speculate that the change in the urban coefficient is
another indication that the reform is operating through non-cognitive skills, if the reason
why urban children are more likely to drop out of high school is because they are exposed
to and tempted to engage in a wider variety of risky behaviours than those living in rural
areas. The existence of a family size – and inexistence of a family income - reform
32
interaction may indicate that the effect of family resources on dropout rates is changed by
the reform but that it is not financial resources. Instead, it could be time resources, which
decrease on a per-capita basis as the number of children increases and cannot be adjusted
as easily as financial resources. This makes sense given the nature of the reform, which is
essentially increasing time available for activities with children.
7. Concluding remarks
We investigate the long term consequences of time investments in children during their
first year of life using a maternity leave reform in Norway, offering up to 4 months of
paid leave and an additional 1 year of unpaid leave, which shows substantial positive
effects of having mother at home, compared to informal care alternatives. 2.7 percent
more children complete high school (and 5 % higher earnings at age 30), going up to 5.2
percent (8 % higher earnings) for those whose mothers have less than 10 years of
education.
The alternative for staying home with mothers around the time of the reform is
crucial to understand the results. There was almost no available high quality child care for
under-two year olds available so the alternative was grandparents or other informal care
which is not necessarily a good substitute to mother’s time at this period of a child’s life.
Note that this was different for the two papers from the Nordic countries using registry
data. In addition, the Swedish reform for instance was an extension from one year to
almost a year and a half, while the Norwegian reform was a reform for much younger
children and biting most for mothers taking short leaves. The positive effect of early
investments in children on medium to long term outcomes also resembles the relatively
33
large effects found recently from other early investments in children such as the Perry
programme and the project STAR (Chetty, Friedman, Hilger, Saez, Schanzenbach and
Yagan, 2010; Heckman, Moon, Pinto, Savelyev and Yavitz, 2010).
For policy implications we conclude that fostering policies to increase parents’
time with children the first year after birth may have an impact on children’s abilities
later in life. This effect has been an important part of the goals behind expansions in
maternity leave across countries; however this study is the first to show that this may
actually be achieved. The situation with maternity leave is remarkably similar in the US
today as it was in Norway before the reform. Parental leave is currently under debate in
the US
39
and an introduction of 4 months of paid leave and better job protection are
typically within feasible policies.
40
Using the rich set of family background variables to
address heterogeneity of effects also gives us the advantage of making the study less
dependent on institutional settings in Norway. For example by showing that the effects
are bigger for children from lower educated households this may be important for policy
discussions related to lowering inequalities in general. Many countries, like the US,
Britain, and South America have a substantial inequality in education and income. While
increasing maternity leave for women and men in these countries will not solve these
problems we have shown that it might reduce the existing gap.
39
USA today July 26
th
2005, The New York Times April 16
th
2008
40
http://www.govtrack.us/congress/bill.xpd?bill=h110-3799
34
35
References
BAKER, M., and K. MILLIGAN (2008a): "Maternal Employment, Breastfeeding, and
Health: Evidence from Maternity Leave Mandates."
Journal of Health
Economics
, 27, 871-887.
— (2008b): "Evidence from Maternity Leave Expansions of the Impact of Maternal Care
on Early Child Development." NBER Working Paper Series, 13826.
B
ERNAL, R. (2008): "The Effect of Maternal Employment and Child Care on Children’s
Cognitive Development."
International Economic Review, 49, 1173-1209.
B
ERNAL, R., and M. P. KEANE (2010): "Quasi-Structural Estimation of a Model of Child
Care Choices and Child Cognitive Ability Production."
Journal of Econometrics,
156, 164-189.
B
LACK, S., P. J. DEVEREUX, and K. SALVANES (2008): "Too Young to Leave the Nest:
The Effects of School Starting Age." NBER Working Paper No. 13969.
B
LAU, D., and J. CURRIE (2006): "Pre-School, Day Care, and after-School Care: Who's
Minding the Kids?"
Handbook of the economics of education, 2, 1163-1278.
B
ROOKS-GUNN, J., W. J. HAN, and J. WALDFOGEL (2010): "First-Year Maternal
Employment and Child Development in the First 7 Years."
Monographs of the
Society for Research in Child Development
, 75, 147.
C
AMERON, A. C., and P. K. TRIVEDI (2005): Microeconometrics: Methods and
Applications
. Cambridge Univiversity Press.
C
HENG, M. Y., J. FAN, and J. S. MARRON (1997): "On Automatic Boundary Corrections."
The Annals of Statistics, 25, 1691-1708.
C
HETTY, R., J. N. FRIEDMAN, N. HILGER, E. SAEZ, D. W. SCHANZENBACH, and D.
Y
AGAN (2010): "How Does Your Kindergarten Classroom Affect Your Earnings?
Evidence from Project Star." NBER Working Paper No. 16381.
C
RONBACH, L., and J. LEE (1964): Essentials of Psychological Testing, 2nd Edition.:
London, UK: Harper and Row, 1964.
D
USTMANN, C., and U. SCHÖNBERG (2008): "The Effect of Expansions in Maternity
Leave Coverage on Children's Long-Term Outcomes." IZA Discussion Papers
No. 3605.
F
AN, J. (1992): "Design-Adaptive Nonparametric Regression." Journal of the American
Statistical Association
, 87, 998-1004.
G
ANS, J. S., and A. LEIGH (2009): "Born on the First of July: An (Un) Natural
Experiment in Birth Timing."
Journal of Public Economics, 93, 246-263.
G
REGG, P., and J. WALDFOGEL (2005): "Symposium on Parental Leave, Early Maternal
Employment and Child Outcomes: Introduction."
The Economic Journal, 115, 1-
6.
G
REGG, P., E. WASHBROOK, C. PROPPER, and S. BURGESS (2005): "The Effects of a
Mother's Return to Work Decision on Child Development in the Uk*."
The
Economic Journal
, 115, 48-80.
H
AHN, J. Y., P. TODD, and W. VAN DER KLAAUW (2001): "Identification and Estimation
of Treatment Effects with a Regression-Discontinuity Design."
Econometrica, 69,
201-209.
36
H
ECKMAN, J., S. H. MOON, R. PINTO, P. SAVELYEV, and A. YAVITZ (2010): "Analyzing
Social Experiments as Implemented: A Reexamination of the Evidence from the
Highscope Perry Preschool Program."
Quantitative Economics, 1, 1-46.
I
MBENS, G., and T. LEMIEUX (2008): "Special Issue Editors' Introduction: The Regression
Discontinuity Design - Theory and Applications."
Journal of Econometrics, 142,
611-614.
L
IESTØL, K., M. ROSENBERG, and L. WALLØE (1988): "Breast-Feeding Practice in
Norway 1860–1984."
Journal of biosocial science, 20, 45-58.
L
IU, Q., and O. N. SKANS (2010): "The Duration of Paid Parental Leave and Children's
Scholastic Performance."
The BE Journal of Economic Analysis & Policy, 10.
L
UDWIG, J., and D. L. MILLER (2007): "Does Head Start Improve Children's Life
Chances? Evidence from a Regression Discontinuity Design."
Quarterly Journal
of Economics
, 122, 159-208.
OECD (2008): "Statistics on Labour Force Participation." OECD, Paris.
P
ORTER, J. (2003): "Estimation in the Regression Discontinuity Model." Unpublished
Manuscript, Department of Economics, University of Wisconsin at Madison.
R
ASMUSSEN, A. W. (2010): "Increasing the Length of Parents' Birth-Related Leave: The
Effect on Children's Long-Term Educational Outcomes."
Labour Economics, 17,
91-100.
R
OSSIN, M. (2011): "The Effects of Maternity Leave on Children's Birth and Infant
Health Outcomes in the United States."
Journal of Health Economics, 30, 221-
239
R
ØNSEN, M., and M. SUNDSTRÖM (1996): "Maternal Employment in Scandinavia: A
Comparison of the after-Birth Employment Activity of Norwegian and Swedish
Women."
Journal of Population Economics, 9, 267-285.
— (2002): "Family Policy and after-Birth Employment among New Mothers–a
Comparison of Finland, Norway and Sweden."
European Journal of
Population/Revue europeenne de demographie
, 18, 121-152.
S
UNDET, J. M., D. G. BARLAUG, and T. M. TORJUSSEN (2004): "The End of the Flynn
Effect?:A Study of Secular Trends in Mean Intelligence Test Scores of Norwegian
Conscripts During Half a Century."
Intelligence, 32, 349-362.
S
UNDET, J. M., K. TAMBS, J. R. HARRIS, P. MAGNUS, and T. M. TORJUSSEN (2005):
"Resolving the Genetic and Environmental Sources of the Correlation between
Height and Intelligence: A Study of Nearly 2600 Norwegian Male Twin Pairs."
Twin Research and Human Genetics, 8, 307-311.
T
ANAKA, S. (2005): "Parental Leave and Child Health across Oecd Countries." Economic
Journal
, 115, 7-28.
37
Figure 1
The 1977 reform
0 10 20 30 40 50
weeks
1956 1.7.1977
The 1977 parental leave reform
paid leave unpaid leave
Source: regjeringen.no, lovdata.no
Figure 2
Proportion of mothers eligible for maternity leave
38
Figure 3
Female employment in Norway and the US 1970-1990
Source: Statistics Norway, Bureau of Labor Statistics (projected from Population
Bulletin, Vol 63 (2008), OECD
Figure 4
Day-care coverage in Norway split by age and urban-rural areas
Data source: NSD municipality
0 10 20 30 40 50 60 70 80
percent
1
9
71
1973
1
9
75
197
7
1
9
79
1
98
1
1
9
83
1
98
5
1987
1
98
9
1970
19
72
1974
19
7
6
1978
19
8
0
1
9
82
19
8
4
1
9
86
1988
19
90
year
all Norwegian women all US women
Married Norway Non-married Norway
1970-1990
Female labour force participation
39
Figure 5
Pre-reform characteristics
Note: Each graph shows the estimated mean for mother’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of three and the dashed lines are the
corresponding 95 % confidence intervals. The y-axis includes outcomes within +/- .15 of a standard
deviation around the mean.
40
Figure 6
Children’s outcomes by birth month, eligible mothers 1977 versus 1975
Note: Each graph shows the estimated mean for children’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of three. The y-axis includes outcomes within
+/- .15 of a standard deviation around the mean.
41
Figure 7
Mother’s outcomes by birth month, eligible mothers 1977 versus 1975
Note: Each graph shows the estimated mean for mother’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of three. The y-axis includes outcomes within
+/- .15 of a standard deviation around the mean.
42
Table 1
Parametric regressions – using only children born in June and July
Birth month Single
Difference
Differences-in-
differences
using 1975 as controls
Children
High School Dropout
-.020*
(.011)
-.025*
(.016)
College Attendance
.094
(.069)
.131
(.098)
Log Earnings at Age 30
.045**
(.022)
.055*
(.031)
Mothers
Pre-Reform Characteristics
Years of Education -.023
(.063)
-.013
(.088)
Log Income Two Years Prior to
the Birth of the Child
-.014
(.031)
.027
(.040)
Outcomes
Average Log Income in the Year
of Birth and the Year After Birth
.148*
(.080)
-.030
(.116)
Employed 5 Years After the Birth
of the Child
-.002
(.012)
-.006
(.017)
Log Income 5 Years after the
Birth of the Child
-.018
(.138)
-.068
(.194)
The second column of this table shows coefficients of a regression of each of the variables in the first
column on an indicator for being born in July 1977. The sample includes only individuals born in June and
July of 1977. For the third column of the table we add to the sample those born in June and July of 1975,
and we regress each of the variables in the first column on a year indicator, a month of birth indicator, and
the interaction of the two. We report the coefficient on the latter.
43
Table 2
Characteristics of eligible and non-eligible mothers
Eligibility status Eligible 1977 Non-eligible 1977
Children
High School Dropout
.186
(.388)
.276
(.447)
College attendance
.46
(.50)
.35
(.48)
Log Earnings at Age 30
12.6
(.74)
12.5
(.76)
Mothers
Years of Education
10.63
(2.18)
9.61
(1.72)
Age at Birth
(in years)
26.1
(.028)
26.5
(.041)
Income in 1975
(in NOK)
25216
(18390)
2831
(7080)
Employed 2 years After Birth
.725
(.447)
.362
(.481)
Employed 5 years After Birth
.758
(.428)
.534
(.499)
Income in 1982
(in NOK)
71216
(73324)
29434
(48202)
44
Table 3
Children’s outcomes
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Dropout rate
.19
-.019*
(.010)
-.027**
(.014)
College attendance
.46
.018
(.013)
.036**
(.018)
Ln(earnings) at age
30
12.6
.048**
(.020)
.055*
(.029)
N
29163
59564
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
45
Table 4
Mother’s income around time of birth
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-differences
using 1975 as controls
Bandwidth 3 3
Ln(income)
Year of birth
.191**
(.083)
-.067
(.120)
Ln(income)
+/- one year around year
of birth
.036
(.025)
-.001
(.035)
Ln(income)
+/-two years around year
of birth
.020
(.024)
.005
(.034)
N
29163
59564
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
Table 5
Mother’s labor supply
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Predicted months
of unpaid leave
7.81
-.276
(.198)
.121
(.291)
Employed 2 years
after birth
.73
-.014
(.012)
-.018
(.017)
Employed 5 years
after birth
.76
-.004
(.011)
-.004
(.016)
Ln(Income) 5
years after birth
8.31
-.039
(.126)
-.068
(.178)
N
29163
59564
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
47
Table 6
Differences-in-differences using eligible mothers in 1975 as control group; Results
by mother’s education
Variables Nonparametric differences-in-differences
Bandwidth 3
Mother’s education
subgroups
Less than 10 years 10 years or more
Children
Dropout rate
-.052**
(.026)
-.019
(.016)
College attendance
.068**
(.028)
.026
(.023)
Ln(earnings) at age
30
.089**
(.045)
.033
(.037)
Mothers
Predicted months of
unpaid leave
-.259
(.524)
.157
(.337)
Employed 2 years
after birth
-.008
(.029)
-.018
(.020)
Employed 5 years
after birth
.004
(.028)
-.004
(.019)
Ln(Income) 5 years
after birth
.098
(.305)
-.093
(.216)
N
22067
37497
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
48
Table 7
Differences-in-differences using eligible mothers in 1975 as control group; Results
by quartiles of mother’s months of unpaid leave.
Variables Nonparametric differences-in-differences
Bandwidth 3
Quartiles of mothers months of unpaid leave
Quartiles
1 (lowest) 2 3
4 (highest)
Average levels
of unpaid leave
(Std.Dev)
.40
(.67)
5.14
(1.67)
9.46
(.92)
18.02
(10.2)
N
14894
14894
14889
14887
Children
Dropout rate
-.090***
(.026)
-.050*
(.027)
.008
(.029)
.015
(.032)
College
attendance
.077**
(.036)
.001
(.036)
.018
(.036)
.054
(.035)
Ln(earnings) at
age 30
.107*
(.060)
.123**
(.056)
.005
(.056)
-.006
(.057)
Mothers
Predicted months
of unpaid leave
.008
(.043)
-.059
(.118)
-.018
(.057)
.031
(.725)
Employed 2
years after birth
-.004
(.012)
-.018
(.022)
-.027
(.036)
-.010
(.035)
Employed 5
years after birth
.040*
(.021)
-.035
(.027)
-.024
(.035)
.011
(.036)
Ln(Income) 5
years after birth
.473*
(.251)
-.505*
(.304)
-.279
(.371)
.168
(.381)
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
49
Table 8
The high school dropout decision for boys
Parametric
Differences-in-differences
using 1975 as controls
High school
dropout
Model 1
Model 2
Model 3
Ability
-.066***
(.001)
-.066***
(.001)
-.065***
(.003)
Height
-.002***
(.000)
-.002***
(.001)
-.001*
(.001)
Mothers education
-.013***
(.001)
-.013***
(.001)
-.012***
(.003)
Mothers age at
birth
-.003***
(.001)
-.003***
(.001)
-.004***
(.001)
Parents married in
1980
-.121***
(.008)
-.120***
(.008)
-.102***
(.019)
Family size
.013***
(.002)
.013***
(.002)
.007
(.005)
Family income
-.038***
(.005)
-.038***
(.005)
-.038***
(.011)
Urban location
.016***
(.005)
.016***
(.005)
.010
(.010)
Reform*year77
-
-.055*
(.031)
-.055*
(.031)
Include
interactions of
reform, year and
month controls
no
yes
yes
Interact reform
effect with all
control variables
No
no
yes
N
26378
26378
26378
***significant at 1 %, **significant at 5%, *significant at 10%
50
Appendix A
Construction of Additional Variables
The IQ data is taken from the Norwegian military records for the relevant cohorts,
tested at the age of 18-19. Military service is compulsory for every able young man. IQ at
these ages is particularly interesting as it is about the time of entry into higher education
(or into the labour market for those who decide not to go to university).
The IQ measure is a composite score from three speed IQ tests, arithmetic, word
similarities, and figures (see Sundet, Barlaug and Torjussen, 2004, for details). The
figures test is similar to the Raven Progressive Matrix test (Cronbach and Lee, 1964) the
arithmetic test is quite similar to the arithmetic test in the Wechsler Adult Intelligence
Scale (WAIS) (Sundet, Tambs, Harris, Magnus and Torjussen, 2005, Cronbach and Lee,
1964) and the word test is similar to the vocabulary test in WAIS. The composite IQ test
score is an un-weighted mean of the three subtests. The IQ score is reported in stanine
(Standard Nine) units, a method of standardizing raw scores into a nine point standard
scale that has a discrete approximation to a normal distribution, a mean of 5, and a
standard deviation of 2. Height (in cm) is obtained from the same military records as IQ.
Teenage pregnancy is constructed as a dummy equal to one if the girl has given
birth to a child before she turns 20 years old, and zero otherwise.
Distance to grandparents is created by tracking the postcode information for the
parents of each child in the study with the postcode information for both sets of
respective grandparents in 1980. Living in the same postcode area means that you live
within maximum a few blocks from each other which means it is possible to have daily
contact. We have postcode information for about 80% of the sample. We create a
51
distance dummy equal to one if the couple lives in the same postcode area as at least one
set of grandparents, and 0 otherwise. The rural-urban variable is constructed using
information from Statistics Norway on the degree of centralization of municipalities in
Norway. Urban municipalities include all municipalities with a large city centre or close
to a large city centre while rural municipalities have small or almost non-existing city
centres.
The working part time variable is constructed using information from the 1980
census on whether mother work full time, part time or not work at all. We define working
part time in 1980 as working between 10 and 1300 hours per year, versus the alternative
of not working or working more than 1300 hours per year. The completed fertility of
mothers is constructed by using the population files in 2007 with information on total
number of children. As we measure total number of children 30 years after the reform,
this should capture completed fertility for all mothers, even teenage mothers in 1977.
We would like to have direct information on months of leave, but this is only
available in Norway from 1992 and onwards. Even then we only have information on
paid leave. Therefore, in order to compute total leave taken by each mother we proceed in
the following way. First, we assume that the take-up of paid leave was 100% when it was
first introduced in 1977, which is a plausible assumption.
41
This was in fact what
happened in response to the 1992 and 1993 reforms to paid leave. . Before the April 1992
reform, mothers are able to take 224 days at full coverage or 280 days at 80% coverage.
41
Firstly, Rønsen and Sundström, 1996 show that for the 1968-1988 mothers in Norway almost no one
returned to work before 4 months after the birth. Secondly, from a survey conducted in 1977 on fertility
behavior of women in Norway (Statistics Norway), 60% answered that they thought mothers should stay
home for the first 2 years after giving birth to a child. In addition, the coverage was 100% which gives
strong incentives for full take up. Third, since we observe days of paid leave after 1992 we are able to
check to what extent eligible mothers take up this benefit, and how the take up reacts to subsequent
reforms.
52
For mothers delivering children in March of 1992, the average take up of paid leave was
250 days. After April 1992 there is an increase in maternity leave entitlements to 245
days of full coverage or 310 days of 80% coverage. We observe that average paid leave
taken was 275 days for mothers of those born in April 1992. This figure is slightly higher
at 280 in March 1993, just before the 1993 reform which increased paid leave to 266 days
of full coverage or 336 days of 80% coverage. By April of 1993 average leave taken was
almost 310 days. Given the high levels of leave and strong reactions to reforms, it is
reasonable to assume that the take up of paid leave is close to 100%.
In order to construct unpaid leave we start by calculating a measure of pre-birth
monthly income by dividing 1976 earnings by 12. Then we calculate total earnings in
1977–1980, and divide them by 1976 monthly income, thereby obtaining a measure of
number of months of unpaid leave during the first 36 months after birth. For this
calculation to work, the assumption is that 1976 earnings are a good approximation for
maternal potential post-birth earnings (the earnings she would get had she not gone on
unpaid leave), adjusted for inflation.
42
We limit ourselves to a window of 36 months
because the further away we move from pre-birth earnings, the more likely earnings may
differ because of change of job, part time work, presence of new children, and other
factors unrelated to the 1977 reform.
43
42
It is useful to illustrate with a specific example. If the child is born in June 1977 we subtract six months
of 1976 monthly earnings from 1977 earnings and compare the remaining earnings in 1977 and 1978 to the
1976 earnings. If the mother earns half of 1976 earnings in the twelve months after birth she has taken six
months of unpaid leave. If she earns nothing and takes all twelve months of leave we will continue and use
earnings in 1979 and 1980 to construct leave up to 36 months after birth.
43
One problem with our approach can be that mothers may return to part time work and hence some of our
estimated leave is not absence from work but rather lower earnings due to part time work. This is not a
problem as long as the reform in itself does not effect this transition, as it will only affect levels and not the
change. As we see no effects on earnings five years later, this is not likely to be of large concern.
53
Figure A1
Number of children born to eligible mothers, by birth month, 1975-1979.
54
Figure A2
Children’s outcomes by birth month, eligible mothers 1977
Note: Each graph shows the estimated mean for children’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of three and the dashed lines are the
corresponding 95 % confidence intervals. The y-axis includes outcomes within +/- .15 of a standard
deviation around the mean.
55
Figure A3
Children’s outcomes by birth month, eligible mothers 1977 versus 1979
Note: Each graph shows the estimated mean for children’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of 3. The y-axis includes outcomes within +/-
.15 of a standard deviation around the mean.
56
Figure A4
Children’s outcomes by birth month, eligible mothers 1977 versus 1974
Note: Each graph shows the estimated mean for children’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of 3. The y-axis includes outcomes within +/-
.15 of a standard deviation around the mean.
57
Figure A5
Children’s outcomes by birth month, eligible mothers 1977 versus 1975: IQ and
Teenage pregnancy
Note: Each graph shows the estimated mean for mother’s outcomes by birth month. The solid line is non-
parametrically fitted using triangle kernel with a bandwidth of three. The y-axis includes outcomes within
+/- .15 of a standard deviation around the mean.
58
Figure A6:
Days of paid leave in 1992 and 1993
59
Figure A7
Quantiles of unpaid leave: show no action in unpaid leave across any quantiles of
unpaid leave.
60
Table A1
Parametric regressions – using only children born in June and July – additional
outcomes
Birth month Single
Difference
Differences-in-
differences
using 1975 as controls
Children
Teenage pregnancy
.002
(.009)
.009
(.013)
IQ (males)
.142*
(.074)
.295***
(.102)
Height (males)
.499*
(.281)
.503
(.384)
Mothers
Pre-characteristics
Age at birth
(in years)
-.096
(.134)
.051
(.187)
Urban location in 1976 .009
(.014)
.009
(.020)
Distance to grandparents in 1980 .004
(.014)
-.019
(.020)
The second column of this table shows coefficients of a regression of each of the variables in the first
column on an indicator for being born in July 1977. The sample includes only individuals born in June and
July of 1977. For the third column of the table we add to the sample those born in June and July of 1975,
and we regress each of the variables in the first column on a year indicator, a month of birth indicator, and
the interaction of the two. We report the coefficient on the latter.
61
Table A2
Mother’s labor supply and children’s outcomes, total sample of all mothers and
children in 1977 with control groups in 1975
Variables
Nonparametric regression discontinuity
Bandwidth 3 3
Control group
RD
1975
Children
Dropout rate
-.013
(.009)
-.012
(.012)
College attendance
.009
(.010)
.016
(.014)
Ln(earnings) at age
30
.024
(.016)
.028
(.023)
Mothers
Predicted months
of unpaid leave
-.288*
(.158)
-.004
(.227)
Employed 2 years
after birth
-.006
(.010)
-.010
(.014)
Employed 5 years
after birth
-.005
(.010)
-.009
(.014)
Ln(Income) 5
years after birth
-.057
(.108)
-.125
(.149)
N
46245
97312
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
62
Table A3
Children’s outcomes – teenage pregnancy and IQ
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Teenage
pregnancy
.052
.002
(.008)
.008
(.012)
IQ
(males)
5.39
.110*
(.067)
.240***
(.094)
N
14070 (TP-girls)
13150 (IQ-boys)
29042 (TP-girls)
27304 (IQ-boys)
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
63
Table A4
Children’s outcomes using bandwidth 5
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 5 5
Mean
Dropout rate
.19
-.012
(.008)
-.019*
(.012)
College attendance
.46
.008
(.011)
.025*
(.015)
Ln(earnings) age
30
12.6
.036*
(.020)
.037
(.028)
N
29163
59564
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
64
Table A5
Differences-in-differences using eligible mothers in 1975 as control group; Results
by urbanization and distance to grandparents
Variables Nonparametric differences-in-differences
Bandwidth 3 3
Distance to grandparents
Centralization
subgroups Close Not-close Urban Rural
Children
Dropout rate
-.050*
(.029)
-.003
(.019)
-.025
(.020)
-.028
(.021)
College
attendance
.039
(.038)
.033
(.024)
.050**
(.026)
.019
(.026)
Ln(earnings)
at age 30
.054
(.056)
.054
(.039)
.052
(.041)
.058
(.040)
Mothers
Predicted
months of
unpaid leave
1.12*
(.604)
.083
(.387)
-.036
(.399)
.344
(.425)
Employed 2
years after
birth
-.048
(.035)
-.014
(.022)
-.012
(.023)
-.025
(.024)
Employed 5
years after
birth
.002
(.034)
-.006
(.021)
-.023
(.22)
.015
(.23)
Ln(Income)
5 years after
birth
.037
(.371)
-.136
(.239)
-.246
(.248)
.100
(.254)
N
13824
33704
30314
29250
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
65
Table A6
Differences-in-differences using eligible mothers in 1975 as control group; Results
by quartiles of family income two years prior to birth.
Variables Nonparametric differences-in-differences
Bandwidth 3
Quartiles of ln(family income) two years prior to birth
Quartiles
1 (lowest) 2 3
4 (highest)
Mean ln(family
income) two
years before
(Std.Dev)
6.6
(3.2)
9.7
(.17)
10.0
(.09)
10.4
(.20)
N
14894
14898
14886
14847
Children
Dropout rate
.012
(.030)
-.014
(.033)
-.077***
(.029)
-.022
(.029)
College
attendance
.011
(.037)
.005
(.039)
.111***
(.036)
.029
(.042)
Ln(earnings) at
age 30
.072
(.063)
.033
(.062)
.074
(.063)
.051
(.067)
Mothers
Predicted months
of unpaid leave
-.112
(.589)
.852
(.621)
.558
(.567)
-.319
(.720)
Employed 2
years after birth
.006
(.033)
-.062*
(.036)
-.023
(.033)
-.050
(.039)
Employed 5
years after birth
-.005
(.032)
-.000
(.034)
.012
(.032)
-.046
(.036)
Ln(Income) 5
years after birth
-.065
(.364)
-.049
(.383)
-.043
(.355)
-.357
(.403)
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
66
Table A7
Mother’s outcomes – part time in 1980, completed fertility (number of children in
2007) and marital stability in 2007.
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Working part time
in 1980
.42
-.000
(.013)
-.007
(.018)
Completed fertility
in 2007
2.5
-.022
(.026)
-.034
(.036)
Parents are married
in 2007
.73
-.001
(.012)
-.013
(.016)
N
29163
59564
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
67
Table A8
Older sibling’s outcomes
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Dropout rates older
siblings
.20
-.011
(.016)
-.010
(.023)
College attendance
older siblings
.49
.007
(.020)
.006
(.029)
Ln(earnings) in
2007 older siblings
12.7
-.045
(.031)
-.034
(.043)
N
12046
23875
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
68
Table A9
Differences-in-differences using eligible mothers in 1975 as control group; Results
by gender
Variables Nonparametric differences-in-differences
Bandwidth 3
Gender
subgroups
Females Males
Children
Dropout rate
-.027
(.020)
-.026
(.021)
College attendance
.029
(.026)
.041*
(.025)
Ln(earnings) at age
30
-.001
(.042)
.106***
(.037)
Mothers
Predicted months of
unpaid leave
-.161
(.401)
.417
(.423)
Employed 2 years
after birth
-.021
(.023)
-.015
(.024)
Employed 5 years
after birth
-.014
(.022)
.006
(.023)
Ln(Income) 5 years
after birth
-.019
(.257)
-.119
(.246)
N
29042
30522
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
69
Further checks to the validity of the procedure
Table A10 in the Appendix shows an analysis of two populations that should not be
affected by the reform: eligible mothers in 1975 and non-eligible mothers in 1977.
Therefore, if we estimate the RD model of this section on these two populations we
should not find any effects. This is certainly the case for mother’s outcomes in these
years. The results are not statistically significant and the effect on children has the
opposite sign of the reform effect. This is as we have seen earlier due to the month effect
in child outcomes, children born earlier in the year have better outcomes than children
born later in the year.
Table A11 and A12 runs the parametric regressions of Table 1 in the paper using placebo
months (April versus May and August versus September) as discontinuities. There should
be no effect on children’s outcomes as there are no reforms between these birth months.
And this is indeed the case, as we see from the tables there is no effect on the outcomes.
The outcomes for children are of the opposite sign as the reform effect again reflecting
the effect of birth of months on outcomes.
70
Table A10
Placebo results: Mother’s labor supply and children’s outcomes
Eligible mothers 1975 and non-eligible mothers 1977
Variables
Nonparametric regression discontinuity
Bandwidth 3 3
Control group Eligible
1975
Non-eligible
1977
Children
Dropout rate
.007
(.010)
.001
(.015)
College attendance
-.018
(.013)
-.009
(.016)
Ln(earnings) age
30
-.007
(.020)
-.020
(.027)
Mothers
Predicted months
of unpaid leave
-.318
(.214)
-
Employed 2 years
after birth
.007
(.012)
-.002
(.016)
Employed 5 years
after birth
.001
(.011)
-.010
(.017)
Ln(Income) 5
years after birth
.029
(.125)
-.121
(.180)
N
30401
17082
Each cell presents the estimated discontinuity in the outcomes. The second column shows results for July 1
st
1975 and
the third column for July 1977. We estimate regressions using local linear regression as in Hahn et al. (2003) and derive
analytic standard errors based on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant
at 5%, *significant at 10%
71
Table A11
Placebo: Parametric regressions – using only children born in April and May
Birth month Single
Difference
Differences-in-
differences
using 1975 as controls
Children
Dropout rates
.007
(.011)
.009
(.016)
College attendance
-.017
(.014)
-.031
(.019)
Ln(earnings) at age 30
-.019
(.020)
-.012
(.028)
Mothers
Pre-Reform Characteristics
Years of education -.063
(.061)
-.060
(.084)
Ln(Income) two years prior to
birth
-.019
(.087)
.010
(.130)
Outcomes
Average Ln(Income) year of birth
and year after birth
.162
(.100)
-.228
(.146)
Employed 5 years after -.015
(.012)
-.026
(.016)
Ln(Income) 5 years after birth -.132
(.131)
-.308
(.184)
The second column of this table shows coefficients of a regression of each of the
variables in the first column on an indicator for being born in May 1977. The sample
includes only individuals born in April and May of 1977. For the third column of the
table we add to the sample those born in April and May of 1975, and we regress each of
the variables in the first column on a year indicator, a month of birth indicator, and the
interaction of the two. We report the coefficient on the latter.
72
Table A12
Placebo: Parametric regressions – using only children born in August and
September
Birth month Single
Difference
Differences-in-
differences
using 1975 as controls
Children
Dropout rates
.006
(.011)
.020
(.016)
College attendance
-.018
(.014)
-.033
(.020)
Ln(earnings) at age 30
-.029
(.021)
-.030
(.031)
Mothers
Pre-Reform Characteristics
Years of education .001
(.061)
.081
(.087)
Ln(Income) two years prior to
birth
-.120
(.097)
-.133
(.143)
Outcomes
Average Ln(Income) year of birth
and year after birth
.165
(.094)
.047
(.138)
Employed 5 years after .015
(.012)
.021
(.018)
Ln(Income) 5 years after birth .219
(.139)
.313
(.198)
The second column of this table shows coefficients of a regression of each of the
variables in the first column on an indicator for being born in September 1977. The
sample includes only individuals born in August and September of 1977. For the third
column of the table we add to the sample those born in August and September of 1975,
and we regress each of the variables in the first column on a year indicator, a month of
birth indicator, and the interaction of the two. We report the coefficient on the latter.
73
Appendix B
Breastfeeding
Using a survey from mainly one maternity hospital in Norway over time (Liestøl,
Rosenberg and Walløe, 1988) show the pattern of breastfeeding for about 150 years in
Norway. They show that breastfeeding in Norway started to decline around 1920 and
reached its lowest point around 1967 when only 30 percent of women breastfed for 3
months and as few as 5 percent for 9 months. In the late 1970s, the level of breastfeeding
in Norway was back to the level of around 1940 after a decline from the 1920s onwards.
Around the period of the maternity leave reform we are using, about 75 percent breastfed
for 3 months, 50 percent for 6 months and 25 percent of mothers where breastfeeding for
9 months or more. Clearly there is an increase in breastfeeding in this period if we only
study this data set.
We use survey data for mothers being asked about their breastfeeding for all of
their children, and create average months of breastfeeding. The survey is from a health
data set covering all 40 year olds in the early 1990s (“The 40 year old survey”). We are
able to match about 5% of the children in our sample. However, we have the whole
population of children so we still have more than 100 observations in each month cell.
This is too little data to establish a convincing regression design as with our other results,
but in Figure B1 we show the average months of breastfeeding across months of birth for
eligible mothers in 1977 and 1975. Firstly this shows that breastfeeding has increased
from 1975 to 1977 as is consistent with the data from Bernal and Keane, 2010. However
74
there is no increase in breastfeeding after the reform in 1977.
44
If anything there is a small
decline in average months of breastfeeding across birth months in 1977. This indicates
that breastfeeding is not the most important mechanism to explain the positive results on
children’s outcomes.
We present the results for the effect on maternity leave on the height of men at the
age of 18–19, which is an outcome linked to better health. In Table B1 we present the
results both from the RD design and the DD results using eligible mothers from 1975 as
comparison group. The results suggest that there is a positive effect of about 0.5
centimetres for men born post-reform. The increase per decade in height among men
measured at 18 was about one centimetre for cohorts born from 1950 to 1990 in Norway,
so the 0.5 centimetre is quite substantial. This clearly indicates that there is a positive
effect of the reform through better health. Given that we do not see an increase in
breastfeeding around the reform this is likely to come from the mother investing more
time at home the first year of the child’s life, providing a more stable and less stressful
environment.
44
We have also tried different measures as an indicator variable for breastfeeding at least 6, 8 and 9 months
and we obtain similar results. There is no clear pattern across birth months for eligible mothers in 1977 (or
on our control groups of eligible mothers in 1975 and non-eligible mothers in 1977).
75
Figure B1
Breast Feeding in Norway – eligible mothers 1977
76
Table B1
Height (males only)
Variables
Nonparametric
Regression
discontinuity
Nonparametric
Differences-in-
differences
using 1975 as controls
Bandwidth 3 3
Mean
Height
(male)
180 cm
.48*
(.27)
.63*
(.37)
N
13541
28371
Each cell presents the estimated discontinuity in the outcomes as a result of the maternity leave reform July 1
st
1977.
We estimate regressions using local linear regression as in Hahn et al. (2003) and derive analytic standard errors based
on formulas in Porter (2003) using a triangle kernel. ***significant at 1 %, **significant at 5%, *significant at 10%
... Indeed, most studies that explore the effects of early-life pollution exposure investigate the immediate effects on child birth outcomes (see e.g., Chay and Greenstone, 2003;Currie and Neidell, 2005;Almond et al., 2009;Currie, 2009;Jayachandran, 2009;Currie and Walker, 2011;Knittel et al., 2016;Sanders and Stoecker, 2015;Arceo et al., 2016;Hanlon, 2018;Jia and Ku, 2019;Rangel and Vogl, 2019), with only few exploring potential effects in childhood or early adulthood (see e.g., Reyes, 2007;Bharadwaj et al., 2017;Almond et al., 2009;Sanders, 2012;Black et al., 2013;Isen et al., 2017) and even fewer focusing on outcomes in older age (Bharadwaj et al., 2016;Ball, 2018a). As such, ignoring potential long-term effects 1 For example, research has explored the importance of maternal physical health (Behrman and Rosenzweig, 2004;Almond, 2006;Almond and Mazumder, 2005), maternal mental health (von Hinke et al., 2019), maternal health behaviours (Nilsson, 2017;von Hinke et al., 2014), maternal nutrition (van den Berg et al., 2021), the economic environment (Van den Berg et al., 2006;Banerjee et al., 2010), the early life health environment (Bleakley, 2007;Case and Paxson, 2009;Cattan et al., 2021), or the home environment (Carneiro et al., 2015). of pollution may lead to a substantial underestimation of the total welfare effects caused by exposure to environmental toxins. ...
Preprint
Full-text available
This paper uses a large UK cohort to investigate the impact of early-life pollution exposure on individuals' human capital and health outcomes in older age. We compare individuals who were exposed to the London smog in December 1952 whilst in utero or in infancy to those born after the smog and those born at the same time but in unaffected areas. We find that those exposed to the smog have substantially lower fluid intelligence and worse respiratory health, with some evidence of a reduction in years of schooling.
... L'hétérogénéité importante existante entre pays dans les caractéristiques de leurs systèmes de congés parentaux (montant de la prestation, durée du congé légal, protection de l'empl