Ethiopia

Journal Pre-proof
Here waits the bride? The effect of Ethiopia's child marriage law
Tamara McGavock
PII: S0304-3878(20)30155-3
DOI: https://doi.org/10.1016/j.jdeveco.2020.102580
Reference: DEVEC 102580
To appear in: Journal of Development Economics
Received Date: 5 December 2018

Revised Date: 21 October 2020
Accepted Date: 23 October 2020
Please cite this article as: McGavock, T., Here waits the bride? The effect of Ethiopia's child marriage
law, Journal of Development Economics (2020), doi: https://doi.org/10.1016/j.jdeveco.2020.102580.
This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.
© 2020 Published by Elsevier B.V.

Here Waits the Bride? The Effect of Ethiopia’s Child Marriage Law
Tamara McGavocka
a Grinnell College, USA
ARTICLE INFO ABSTRACT

Keywords: Child marriage is still common in the developing world: in 2012, one in three women was
Early marriage married by age 18, with more than 10% married before age 15. Beginning in 2000, Ethiopia’s
Fertility semi-autonomous regions raised the legal minimum age of marriage from 15 to 18. This study
Harmful traditional practices leverages the natural experiment arising from the staggered roll out of the policy in a difference-
Women’s empowerment in-differences and event study framework. The results suggest that the reform delayed women’s
marriage, and in particular delayed marriages of girls under 16 by about 17 percent (6.8 per-
centage points) in areas where early marriage was more common prior to the reform. However,
the effect of the reform, though larger, is insignificant among women belonging to ethnic groups
with the strongest norms toward early marriage. Women’s fertility was delayed and may be lower
over their lifetimes.
f
oo
“With marriage, a girl’s childhood abruptly ends. Her health and future prospects immediately fall in jeopardy.
Tens of millions of girls in the developing world are at risk of being forced to wed as part of this deeply entrenched
pr
practice that significantly impedes progress on human rights, education, global health, and economic development."
-International Center for Research on Women1 -
re
1. Introduction
lP
Marriage under age 18 was named a human rights violation under international law in 1981.2 Still, during the
2000s more than 30% of women ages 20-24 in the developing world had been married before their 18th birthday, many
well before their 15th (2013, 2013). In Africa and South Asia, the incidence of marriage under 18 was about 50%.
na
Many assert that early marriage generates health risks and social and economic costs for women, their children, and
their communities (Chari et al., 2017; Field and Ambrus, 2008; Malhotra et al., 2011; Jensen and Thornton, 2003;
ur
Warner et al., 2014; Corno et al., 2020; Raj et al., 2009; Mathur et al., 2003; Edmeades, 2013). Ending child marriage
is thus one of the targets under the fifth Sustainable Development Goal (SDG) on gender equality. To this end, more
than half of all low- and middle-income countries have enacted legislation since 1980 prohibiting or limiting marriage
Jo
under age 18 (Kim et al., 2013).3 To date, however, none of these policies has been evaluated at scale in a causal
inference framework (Field and Ambrus, 2008; Wodon et al., 2017).
In this paper, I estimate the effect of Ethiopia’s child marriage law. Beginning in 2000, nine of Ethiopia’s eleven
regions adopted reforms to their family laws that raised the minimum age of marriage from 15 to 18 for girls only,
leaving unchanged the minimum age for boys at eighteen, and allowing marriages of girls at 16 and 17 with parental
consent. Together, my results show a sudden and persistent decrease in the prevalence of under-age marriages upon
implementing the legal reform in the regions. Using flexible difference-in-differences and event study strategies with
Demographic and Health Surveys (DHS) data, I find that the reform decreased the probability that marriages involved
a girl under age 16 by 6.8 percentage points, or 17 percent relative to the mean at baseline. Moreover, this shift did not
simply result from 17 year old girls in more progressive areas waiting until their 18th birthday to marry. Rather, I find
that the reform was most effective at delaying early teen marriages in districts where the earliest marriages were more
common pre-reform.4 There, the probability of marriage between 16 and 17 actually increased slightly, indicating
that some pre-16 marriages were delayed until at least the parental exception age of 16. In districts where the average
ORCID (s):
1 See https://www.girlsnotbrides.org/ending-child-marriage-works-look-evidence/.
2 1981 United Nations’ Convention on the Elimination of All Forms of Discrimination Against Women (UNICEF, 2013), see also Jacobson
(1992).
3 Based on a study of 120 low and middle income countries. Some nations prohibit marriage under age 18 entirely, while others, including
Ethiopia, allow exceptions, e.g. for girls at least 16 years old with parental consent.
4 To do this, I match DHS clusters to districts using publicly available geo-spatial data.
T. McGavock: Preprint submitted to Elsevier Page 1 of 35

age of marriage was already above 18, the proportion of marriages at ages 16 and 17 decreased and marriage over 18
became more common.
I also consider heterogeneity in the impact of the reform by ethnicity, a key factor associated with norms toward
early marriage in Ethiopia. I find that the post-reform declines in early marriage are larger in magnitude, but not quite
significant, among women belonging to ethnic groups with higher rates of early marriage pre-reform. Interestingly,
these groups’ pre-reform educational attainment and empowerment levels are higher as well, perhaps owing to historic
agricultural practices that imply a low importance for women’s and children’s contributions to agriculture based on
evidence from Murdock’s Ethnographic Atlas. These differences by ethnicity in the effects of the reform suggest that
efforts to address early marriage should consider local norms in the design of effective local policy.
The validity of these causal estimates rests on two assumptions: i) the reform’s timing was exogenous to regional
trends and ii) other policies and interventions cannot explain the results I observe. Although these assumptions are
inherently not testable, I provide empirical evidence in a number of ways to demonstrate that these critical conditions
likely hold in this setting. First, the timing of the reform across regions does not appear to be systematically related
to pre-existing norms or trends in rates of early marriage. My event study analysis supports this assumption, and
also tests for dynamic treatment effects. Second, while there were a number of government and NGO interventions
in Ethiopia around the same time frame as the family law reform, my results are robust to considering their potential
f
impact on child marriage. That is, my results are little changed by dropping data from the locations where NGO efforts
oo
were focused on early marriage interventions or by controlling for the effect of Ethiopia’s free primary education and
mother tongue instruction programs.
pr
This paper makes two contributions to the literature on child marriage. First, it provides the first causal analysis
of the effectiveness of an age of consent law in developing countries, to my knowledge. Though many countries have
-
amended laws on the minimum age of consent to marry, sources of plausibly exogenous variation are difficult to find
re
outside of the US. Thus most empirical evidence relies on associations between legal minimum ages of marriage and
declining early marriage rates without confirmation of causality (Arthur et al., 2018; Field and Ambrus, 2008; Wodon
et al., 2017; Maswikwa et al., 2015; Blank et al., 2009).5 The staggered roll-out of the law across regions in Ethiopia
lP
during 2000-2010 that I use in this paper provides exogenous variation required to address this question.
Programs striving to prevent early marriage via other means have met with success. A number of studies find that
na
early marriage rates drop in response to educational subsidies (Duflo et al., 2015, 2019; Chicoine, 2020), cash transfers
(Baird et al., 2010, 2011; Buchmann et al., 2018), training in both soft and hard skills for jobs (Bandiera et al., 2012,
2020) and even political role models (Castilla, 2018). “Community conversations” – in which local leaders and NGOs
ur
meet with community members to discuss the harms of traditional practices and the benefits of waiting for marriage –
have also met with some success, though the vast majority of these programs have not been evaluated using randomized
Jo
research designs (Erulkar and Muthengi, 2009; Malhotra et al., 2011; Lee-Rife et al., 2012; Chow and Vivalt, 2019).6
In some cases, these community conversations have increased participants’ polarization of views on early marriage
(Chow and Vivalt, 2019). Even considering encouraging results from interventions in partial equilibrium, it is not
always clear that interventions that work can be implemented at scale due to budgetary concerns and, even if they are,
whether the effects scale up as well (Deaton, 2010; Muralidharan and Prakash, 2017; Banerjee et al., 2017).
The second contribution of this paper is that it provides another source of plausibly exogenous identification in
estimating the causal effect of early marriage for women on fertility, education, and empowerment. Identifying causal
effects is difficult because of potential reverse causality: human capital, labor supply, and the availability of contra-
ception have all been shown to lead to delays in marriage (Wodon et al., 2017; Baird et al., 2010; Duflo et al., 2006).
A small body of work directly takes up the question of the reverse direction, the causal impact of delaying marriage
on women’s outcomes. These studies, nearly all of which cover South Asia and not sub-Saharan Africa, utilize either
i) an instrumental variable strategy to identify the timing of marriage for girls (Chari et al., 2017; Field and Ambrus,
5 Legal reforms formalizing equality between the sexes in land titling, divorce and inheritance law, and political participation have been shown
to lead to favorable changes in women’s outcomes (Chiappori et al., 2002; Adam et al., 2003; Bailey, 2006; Gray, 1998; Harari, 2019; Duflo, 2012;
Castilla, 2018). Even when laws are not easily enforced by governments with limited resources, legal reforms and the information campaigns that
advertise them can signpost changes in social norms that complement enforcement of the legal framework (North, 1990; Kumar and Quisumbing,
2015; Hallward-Driemeier and Pritchett, 2015; Lee-Rife et al., 2012).
6 Out of 66 NGO or government programs related to child marriage in 30 countries, and an additional 58 in India alone, less than 10% had
designed their studies with concern for causal identification at all (Malhotra et al., 2011). Only four of these 124 programs had a randomized
treatment and control group design. Since then, Buchmann et al. (2018) evaluated the effects of a cash transfer program conditioned on remaining
unmarried combined with an empowerment program. They find that the cash transfer successfully delayed marriage, but the empowerment program
had no effect, whether combined with cash or not.

2008; Sekhri and Debnath, 2014) or ii) evaluations of interventions specifically designed to delay marriage through
community conversations, cash transfers, or both (Erulkar and Muthengi, 2009; Baird et al., 2010). Previous estimates
of the causal effect of age at marriage on outcomes suggest that delaying marriage by one year improves human cap-
ital and health outcomes for women, measured using educational attainment, literacy, and weight-for-height. As a
result, schooling and health outcomes for their children improve as well, measured by vaccination rates, weight-for-
height, school enrollment, and test scores. However, evidence for impacts of delaying marriage on other dimensions
of well-being such as women’s empowerment measures – perceived say in decision-making, freedom to go out alone,
perceptions about the acceptability of domestic violence – is very mixed (Chari et al., 2017; Wodon et al., 2017).
By introducing an age of consent law as another source of exogeneity, this paper offers identification of these causal
effects with several novel elements. First, the group of women treated by the reform is older than the the treatment
group identified by existing IV estimates. That is, the age of consent law I study generated delays among 15 and 16
year old girls. On the other hand, previous work has identified exogenous variation in marriage age using menarche,
or the age of the onset of menstruation, as the instrumental variable, which typically generates delays in marriage
among younger girls around 12-14 years old (Chari et al., 2017; Field and Ambrus, 2008; Sekhri and Debnath, 2014).7
Delaying marriage at ages 15-16 may have different implications for the causal impacts on other outcomes relative to
delays in earlier marriages. Second, if age of consent laws are well-enforced, they may generate even larger delays in
f
marriage than do menarche or NGO interventions, which may have different implications for secondary outcomes if
oo
the effects are discontinuous or nonlinear. Third, both menarche and smaller scale NGO interventions have impacts in
partial equilibrium, but legal reforms can move marriage markets to a new equilibrium. Finally, previous studies find
pr
that both menarche and NGO-led programming around child marriage lead to gains in schooling, which can impact
outcomes independently of marriage age. On the other hand, it is possible that changing the legal age of marriage from
-
15 to 18 in a setting with very low secondary school completion rates would not lead to effects on education (Field and
re
Ambrus, 2008). In that case, using a legal change as variation could help to isolate the effect of maturity at the time
of marriage from any attendant education effects.8
lP
My results from these secondary outcomes suggest that women’s total lifetime fertility will be lower as a result of
later marriage, and that women married later are less likely to give birth at home. This may result in lower mortality
rates among infants, though this effect is not statistically significant (possibly due to low power). Women may have
na
more perceived decision-making power and may be less likely to accept spousal violence (but these results are also not
significant).
In what follows, I discuss the particular legal, political, and ethnographic environments in which my empirical
ur
approach is set. Section 3 presents the empirical strategy I use to investigate the effect of the 2000 Revised Family
Code in Ethiopia and the data I use to estimate the results. In Section 4, I present results from my estimation and
Jo
robustness considerations. Section 5 discusses the effect on secondary outcomes and mechanisms, and 6 concludes
with a discussion for future work.
2. Marriage and Ethiopia’s Family Law Reform

Across Ethiopia, marriage is most often arranged, with most brides and many grooms having no part in the choice
of spouse; it is common for young girls not to know they are being married until the day of the wedding, on which
the bride and groom meet for the first time (Pankhurst, 1992; Fafchamps and Quisumbing, 2002, 2005; Kumar and
Quisumbing, 2015). The majority of brides bring nothing to the marriage. Bride price and, to a far lesser extent,
polygamy are both practiced with some frequency in Ethiopia across ethnic and religious groups. Marriage markets
operate within tightly clustered sub-regional districts (woredas). Given Ethiopia’s federal ethnic state structure, this
means that nearly all marriages occur between members of the same ethnic group.9
In July 2000, the Federal Democratic Republic of Ethiopia amended existing law on family relation in its “Revised
Family Code” to “provide the legal basis which guarantees the equality of the spouses during the conclusion, duration,
and dissolution of marriage” (Federal Negarit Gazetta of the Federal and of Ethiopia, 2000). Within a package of
reforms to existing law, Articles 6 and 7 require “free and full consent” and raised the minimum age of marriage from
7 Recent evidence suggests IV estimates using menarche to identify marriage exogenously may underestimate the effects of delaying marriage
on outcomes: if girls with better nutrition begin menstruating earlier, later menarche is associated with lower nutritional status (Khanna, 2020). In
this case, the treatment effects of delaying marriage may be biased downward.
8 Chari et al. (2017) and Field and Ambrus (2008) examined this maturity effect by estimating the effect of delays in marriage among the never
schooled population of girls.

9 In the DHS, 87% of couples sampled in the DHS report the same ethnicity in their individual interviews.

15 to 18 years for girls, granting an exception of not more than two years for “serious cause” with parental consent.
In other words, the new legal absolute minimum age of marriage for girls was 16, but the guidance of the law aimed
toward a minimum age of 18. The law left untouched the legal minimum age of marriage for men at 18, and marriages
that occurred before the reform involving girls under age 18 remained valid. Later, in 2004, arranging the marriage
of a child under 13 was named a criminal act, punishable by up to seven years in prison.10 An important feature of
marriage law in Ethiopia is that marriages that are performed as part of a traditional customary or religious ceremony
are considered legal by the government, in addition to marriages carried out by an officer of civil status. In practice,
this means that the majority of marriages in rural Ethiopia, where nearly 80% of the population still resides as of 2020,
are executed by traditional and religious leaders, not by officers of the court. Combined with the fact that enforcement
of the legal minimum age – both before and after the reform – is weak, this means that persisting norms and customs
have historically caused the median age at marriage in Ethiopia to be substantially lower than 18. In my sample from
the nationally representative Demographic and Health Surveys (DHS), the median age at marriage among women ages
in 2000 was 16, with 30% of women married at age 14 or younger.
The Revised Family Code was approved by the federal government in July 2000 but its implementation occurred
at the regional government level and was staggered: the reform took effect immediately in the two chartered cities
(Addis Ababa and Dire Dawa). Three regions (Oromia, Amhara, and SNNPR) followed suit with adoption of the
f
reform during 2000-2005, and then Benishangul-Gumaz, Tigray, Harari, and Gambela adopted their own reformed
oo
family codes between 2005 and 2009. The two remaining regions, Afar and Somali, have not yet adopted revised
family codes of their own as of this writing. 11 My empirical strategy, described more fully in Section 3, capitalizes
pr
on this staggered roll-out of the reform to estimate the effect of raising the age of consent on the distribution of age at
marriage of women who were married just before and just after the reform date in their regions.
-
Evidence from local NGOs, ethnographers, and government documents shows that news of the reform spread via
re
the de-centralized network of local government bodies at the woreda (sub-regional district) and kebele (sub-district)
levels (Mekonnen and Aspen, 2009). In addition, international NGOs and local community organizations were active
lP
in raising awareness of the new minimum age. Newspaper, radio, and community plays advertised the reform’s im-
plications for marriage ages and the dangers of harmful traditional practices like female genital mutilation and early
marriage for girls. Local government, health, education, and religious leaders were trained in order to spread knowl-
na
edge of it (Pankhurst, 2015).
3. Empirical Strategy
ur
3.1. Data
Jo
I use household surveys from the Demographic and Health Surveys (DHS), conducted in partnership with the
Central Statistical Agency in Ethiopia, which provide nationally and regionally representative cross-sections of women
ages 15-49 and men ages 15-59 in 2000, 2005, 2011, and 2016. 12,13 Appendix Figure B1 shows the timing of these
surveys relative to the timing of reforms in the regions.
Women were asked about their marital status and their “age at first cohabitation,” which the DHS reports in its
own documentation as age at marriage for ever-married women. Details about this variable and evidence that helps
to allay concerns about measurement error are described in Appendix A. In addition to exploring the distribution
of marriage ages as outcomes, I also explore additional variables as secondary outcomes or robustness checks. These
include: women’s educational attainment (whether any school, years of schooling), age at first birth, child mortality, and
indices women’s involvement in decision-making and attitudes about domestic violence. In addressing these outcomes,
I follow Field and Ambrus (2008), Chari et al. (2017), and others. Details about these indices are in Appendix A.
Figure 1 shows that the average age of marriage between the 90s (prior to the reform) and the 00s (following the
federal reform) increased from just below age 17 to just over 18 years old. To identify where in the age distribution
delays in marriage occurred, Figure 2 Panel A shows cumulative distribution functions of the age of marriage for
10 See Giorgis (2005).
11 Appendix Table B1 shows these dates by region. Following Hallward-Driemeier and Gajigo (2011), I refer to these eleven units - the nine
regions plus the two municipalities - as regions. This is how they are defined in the DHS data as well.
12 The surveys were designed to be nationally and regionally representative for the purpose of policy planning. All of my analysis which follows
uses the appropriate sampling weights provided in the data.

13 Due to some security concerns, some enumeration areas in the Somali region were not surveyed for the 2005 and 2011 rounds, and similarly
in Afar for the 2005 round. My empirical strategy counts all women in these regions as untreated, because they have not yet revised their family
code.

women married during 1990-9 (solid black), and 2000-2009 (dashed blue). The figure shows that the probability of
marriage at all ages less than 24 decreased over time. In particular, the probability of marriage by age 15, the previous
legal minimum made illegal by the reform, dropped from about nearly 50% to less than 30%. One might argue that this
change can be attributed to Ethiopia’s development in general that brought other gains in human capital and women’s
status. Breaking the CDFs into 5-year periods in Figure 2 Panel B shows, however, that the CDFs for the two five-year
periods prior to the reform – 1990-1994 (gray, solid) and the 1995-99 period (black, solid) – are more similar to each
other than they are to the 5 years immediately following the reform (blue, dashed). The probability of marriage by age
16 fell by about 5 percentage points between 1990-94 and 1995-99 but by more than 11 percentage points (more than
double) during the next 5 year period. This visual evidence suggests that a significant portion of the change across the
90s and the 00s was not part of a grander trend in falling rates of early marriages.
Figure 1: Average Age of Marriage Pre- and Post-reform

20
f
oo
Average age of marriage
pr
18
-
re
lP
16
na
ur
14
1990-1999 2000-2009
Jo
Notes: The figure shows the average age of marriage across the decade immediately before and
immediately after the reform. Source: Author’s calculations from DHS data.
Table 1 reports summary statistics of the variables of interest in this paper for women married during the 10
years pre- and post-reform, including p-values for t-tests of the difference. These indicate that early marriage was
less common during the post-reform period, and that educational attainment also increased. My empirical strategy
identifies the portion of these gains that can be attributed to the reform.
3.2. Empirical Strategy

I estimate the effect of Ethiopia’s reform by comparing the average age of women married in years pre- and post-
reform, capitalizing on the staggered timing of the reform in each region. While this approach allows only for investi-
gation of the intensive margin of age of marriage and not marital status, I argue that little is lost since Figure 2 shows
that even after the reform 90% of women were married by age 24. The benefit of this approach is that it clearly defines

Table 1
Summary Statistics
Pre-Reform Post-Reform P-value for

Difference
Age Marr. 17.55 19.1 0

Under 16 .39 .22 0
16 to 17 .22 .23 .04
Over 18 .4 .55 0
Age at Survey 27.02 24.02 0
Divorced .08 .08 .84
Fertility
Has Any Children .87 .75 0
Age at First Birth 19.24 20.23 0
Has Lost a Child .22 .09 0
Husband wants more children than respondent .22 .2 0
f
Human Capital and Labor Supply
oo
Any School .38 .6 0
Yrs Sch. if any 6.44 7.04 0
Employed .48 .5 0
pr
Perceived Empowerment
Decision-making Index - -.21 -.16 .02
Can get medical help .71 .73 .01
re
Can ask husband to use condom .72 .75 0
Domestic Violence Index .07 1.05 0
lP
Location and Ethnicity

Lives in Addis .09 .13 0
Affar 0 0 .91
na
Amhara .28 .28 .81

Somalie .01 .02 0
Oromo .29 .31 .01
ur
Tigrie .1 .11 .2
Notes: This table shows average values of each variable for women married 10 years prior to the
Jo
reform and those married 10 years post-reform. P-values for a t-test are also shown.
the “treatment” group as women whose marriages occurred post-reform.14
Age at first marriage𝑖𝑤𝑟𝑡𝜏 =

𝛽 Post-Reform𝑖𝑤𝑟𝑡 + 𝜂𝑡 (1)
+ 𝜇𝑟 + 𝜇𝑟 ∗ 𝑡𝑟𝑒𝑛𝑑 + 𝜈𝜏 + 𝜖𝑖𝑤𝑟𝑡𝜏
where the dependent variable is the age at first marriage15 of woman 𝑖 residing in district (woreda) 𝑤 in region 𝑟
married in calendar year 𝑡 and surveyed in year 𝜏, “Post-Reform” is an indicator for whether the marriage took place
after the reform in region 𝑟, and year of marriage (𝜂𝑡 ), region (𝜇𝑟 ), and year of survey (𝜈𝜏 ) fixed effects are included.
I also include region-specific linear time trends (𝜇𝑟 ∗ 𝑡𝑟𝑒𝑛𝑑) and a vector of other controls: a rural dummy, ethnicity
fixed effects, and age at survey. Using this specification, the coefficient 𝛽 on the post-reform dummy captures the
treatment effect of the reform: the average change in marriage age generated by the reform.
14 This approach was used by Castilla (2018) and is similar to the identification strategy of Gershoni and Low (2017), without the comparison to
men.
15 I de-trend all outcomes first by estimating residuals from a regression in which the outcome variable (here, age at first marriage) is regressed
on year of birth dummies, survey year dummies, region dummies, region-specific time trends, and age control, and a rural dummy. De-trending
generates better-performing estimates in the presence of dynamic treatment effects (Borusyak and Jaravel, 2016; Goodman-bacon, 2019). Appendix
C3 shows that results are robust to using the outcome variable without de-trending, but the estimates are noisier.

The year of marriage and region fixed effects in these specifications control for any factors affecting marriage age
that are common among women married in the same year (e.g. economic crises or security issues generated by shifts
of power) or living in the same region. Economic conditions, changes in norms over time, and changes in average
educational attainment over time that are specific to regions are flexibly controlled for by the region-specific linear
time trends and the survey year fixed effects. Because women are interviewed either after they are married or before
they are married, but not both, it is not possible to include controls related to women’s families of origin which might
contribute to their timing of marriage. To the extent that the distribution of or trends in these factors vary systematically
by region or over time, these factors are implicitly controlled for by inclusion of marriage year, survey year, and region
fixed effects and region-specific time trends.16
The treatment effect of the reform 𝛽 is thus identified off of variation from differential exposure to the change in
the minimum age of marriage across regions and within marriage years as well as within regions and across marriage
years. One “experiment” set up by this strategy can be described as follows: consider all women belonging to the
Oromo ethnic group and who were married in 2001. In 2001, only Oromia and the two federal cities had reformed
their family codes. The empirical strategy employed in this study thus compares the average age at marriage (or some
other outcome) among women in the regions where the reform had already occurred to those in regions where it had
not yet occurred. If the average age of marriage after de-meaning for region differences in levels (using region fixed
f
effects) was higher in the already-reformed regions than in the not-yet-reformed regions, we would observe a positive
oo
coefficient 𝛽 from estimating equation 1.
I correct errors for heteroskedasticity and serial correlation by clustering at the district (woreda) level, but show that
pr
my results are robust in general to clustering at both smaller and higher levels. While the timing of the reform varied
at the region level, in Section 2 I described reasons that the level of treatment should be considered as the district level:
-
news of the reform spread through district (or lower) level efforts and marriage markets operate at at the district level
re
for the vast majority of women.17 Thus, my preferred results cluster standard errors at the district level, which results
in 488 clusters. For reference, I also show p-values at the region level, corrected for the small number of clusters using
lP
the Wild bootstrap procedure (Cameron et al., 2008). Appendix Table C6 shows that results are robust to clustering
at the lower DHS survey cluster level. Finally, all tables include errors adjusted for multiple hypothesis testing when
considering related outcomes.
na
Identification of causal effects relies on two assumptions. First, that the timing of legal reform was uncorrelated
with time-varying observed or unobservable factors that affect age of marriage or other outcomes, after controlling for
time-invariant region characteristics and linear region-specific time trends. The second assumption is that there are no
ur
other policy changes affecting the age at which women were marrying before and after the reform that could account
for the results obtained in this study.
Jo
Identifying Assumption 1: Exogenous Timing of Reform by Region.

A violation of this assumption would occur if regions that adopted the reform earlier were systematically more
(less) progressive or if changes over time in early marriage rates were systematically faster or slower. Normalizing
the reform date (x-axis) and pre-reform averages (y-axis) to zero, Figure 3 shows that age at marriage was relatively
stable across three groups (the Federal cities Addis Ababa and Dire Dawa, where the reform took immediate effect in
2000; the early adopting regions; and the late adopting regions). Post-reform, there is a jump apparent at year zero,
and upward trends from there. In Appendix B, I provide additional evidence from the raw data for the exogeneity of
the timing of the reform to test and reject that the timing of the reform was related to region-level early marriage and
education rates.
In addition to this direct empirical evidence from the data, Ethiopia’s political structure also warrants this assump-
tion of plausibly exogenous timing. Ethiopia’s democratic socialist republic was designed to give importance and pay
respect to Ethiopia’s ethnic groups and “nationalities,” establishing regional boundaries based on these ethnic group-
ings and granting regions significant autonomy (Kefale, 2013; Abebe, 2014). The federal government’s immediate
jurisdiction is within the two large federal cities, Addis Ababa and Dire Dawa, and its legislative proclamations take
immediate effect there, but each region enacts and approves its own laws. In practice, this pays recognition to Ethiopia’s
customary, religious, and traditional laws in Ethiopia’s complex ethnic and religious landscape, but ultimately estab-
lishes extremely similar laws throughout the country after the regional bureaucracies complete their ratification pro-
16 In Appendix Table C1 I demonstrate that in practice the inclusion of survey year and region-specific time trends affects the estimates little, but
also makes estimates less precise: the standard errors of the estimates increase.
17 Here, I follow Cameron et al. (2008) in assuming that the treatment level is the unit (woreda) and not the unit-cohort, or unit-pre/post level.

cesses (Burgess, 2012). Several regions18 are typical early adopters with some procedural lag. In the case of the family
law, security concerns delayed legislative action in SNNPR and disputes about polygamy delayed adoption in Oromia
(Pausewang et al., 2002; Smith, 2013). Together, these political economy procedures and region-specific delays unre-
lated to trends in or preferences for early marriage helped spread out the timing of the changes in the age of consent
across Ethiopia.
To more formally test for differential pre-trends as well as dynamic treatment effects, I also employ an event-study
approach by replacing the Post-Reform𝑖𝑤𝑟𝑡 indicator with a set of dummies for periods pre- and post-reform (Wolfers,
2006; Freyaldenhoven et al., 2019; Goodman-bacon, 2019). This specification is:
Age at first marriage𝑖𝑤𝑟𝑡 =

∑
𝛽𝑘 ∗ Period k relative to reform𝑖𝑐𝑟𝑡 (2)
𝑘
+ 𝜂𝑡 + 𝜇𝑟 + 𝜇𝑟 ∗ 𝑡𝑟𝑒𝑛𝑑 + 𝜈𝜏 + 𝜖𝑖𝑤𝑟𝑡
where Period k relative to reform𝑖𝑐𝑟𝑡 is a set of dummy variables for each two-year period pre- or post-reform. 19 I
f
group all observations more than 8 years prior to and after the reform into a single period because the region with the
oo
latest reform occurred exactly eight years after the federal reform and eight years before the most recent DHS sample.
Identifying Assumption 2: Accounting for Other Policies and Shocks
pr
In order to estimate the treatment effect of the change in the legal age of consent to marry, the reforms in each region
must be uncorrelated with other policy changes or shocks. Given the various fixed effects included in my specifications,
-
it is unlikely that there remains variation in outcomes which is explained by systematic correlation with other policies.
re
To test this empirically, I implement a number of robustness checks that incorporate these other potential treatment
indicators as controls. I take this question up in Section 4.4.
Selective migration to avoid the reform’s stipulations would also bias estimates of the treatment effect.20 However,
lP
a few pieces of evidence mitigate this potential problem. First, anthropological analysis suggests, and data on spousal
pairs’ ethnicity confirm, that intra-ethnic marriage is a very strong norm in Ethiopia. The ethno-linguistic diversity
na
of Ethiopia and the fact that Ethiopia’s regions were distinctly drawn along the major geographic boundaries of these
ethno-linguistic groups means that in practice cross-regional migration for marriage is very uncommon (Pankhurst,
1992). Second, I examine the data and conclude that cross-region migration is not a concern. Details are reported in
ur
Appendix A.
Jo
4. Effect of the reform on early marriages

4.1. Difference-in-differences analysis
First, I present results about the effect of the reform on the average age at marriage as described in Equation 1.
Table 2 Column 1 shows the treatment effect of the reform on age at marriage using my preferred combination of
controls: a rural dummy, age at survey, and ethnicity fixed effects, plus region-specific linear time trends. The result
in Column 1 suggests that the average age of women married after the reform in their region increased by 0.032 years,
which is a small effect. Appendix Table A1 shows that excluding these controls generates estimates that are stable in
sign and magnitude.
To explore where in the age distribution the reform was effective, I replace the outcome variable in Equation 1
with indicators for marriage in the age groups most likely to be affected by the reform. Recall that while the reform
raised the legal minimum age for marriages without parental consent from 15 to 18, it still allowed for marriages at
ages 16 and 17 with parental consent. Since marriages – and especially those of concern in the abolition of “child
marriage” – are almost always arranged by parents, it is possible that the reform reduced the probability of marriages
which were noncompliant with the exception age of 16 but had no effect on the probability of marriages over 18. That
is, parents planning to arrange a wedding for their 14 year old daughter might have been induced by the reform to delay
18 These are Amhara, Oromia, Tigray, and SNNPR.
19 Iuse two-year periods rather than smaller ranges of time (such as 6 months to one year) due to smaller sample sizes in the post period that
particularly affect later reform regions.
20 That is, early marriage rates in early adopting regions might appear to have fallen significantly when in fact those engaging in early marriage
simply moved to regions that had not yet reformed to avoid the new law.

Table 2
Effect of the Reform on the Distribution of Marriage Ages
Panel A: Combined Effect
Average % Under 16 % at 16/17 % Over 18
Post Reform 0.0315∗ -0.0348∗∗∗ 0.0125 0.0224∗∗
(0.0164) (0.0108) (0.0106) (0.00959)
Observations 41852 41852 41852 41852
p-value CGM 0.198 0.00400 0.344 0.0170
Baseline mean 16.92 0.417 0.211 0.372
Panel B: Separate Effects by District Pre-reform Compliance

Post Reform 0.0125 -0.00209 -0.0230∗ 0.0251∗∗
(0.0198) (0.0116) (0.0136) (0.0115)
Post*Noncompliant 0.0396∗ -0.0681∗∗ 0.0737∗∗ -0.00565
(0.0210) (0.0125) (0.0146) (0.0116)
f
Observations 41852 41852 41852 41852
oo
p-value CGM 0.623 0.812 0.0931 0.0611
p-value CGM (interaction) 0.0621 0 0.00701 0.0551
Baseline mean 16.92 0.417 0.211 0.372
pr
Notes: Panel A shows estimates of the treatment coefficient on “Post Reform” as described in
Equation 1 for four outcomes related to the distribution of marriage ages: the average age of
marriage plus three dummies for marriage under age 16 (made illegal by the reform), marriage at
-
ages 16 or 17 (exceptions allowed with parental consent), and marriage at or above age 18 (the new
legal minimum age). Panel B shows the same outcomes, but includes an interaction effect of
re
“Post-reform” with a dummy indicating that the woman resides in a “Noncompliant” district (woreda),
where noncompliance is determined by whether the average age at marriage in the district is below
18 during the 5 years prior to the reform in each region. A dummy for “Noncompliant” is also
lP
included, but not shown in the table for simplicity. For more details on construction of this variable,
see Appendix A. Included in each specification are: region fixed effects, year of marriage and year of
survey fixed effects, region-specific time trends, respondent’s age at interview, a rural dummy, and
ethnicity fixed effects. Robust standard errors in parentheses are clustered at the district (woreda)
level, since district-level variation is important in identifying the treatment effects, as described in
na
Section 3.2. Significance levels show * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01, where these have been
adjusted for multiple hypothesis testing using the False Discovery Rate procedure described in
Benjamini et al. (2006). p-values from clustering standard errors at the region level (with 11 clusters)
appear in the bottom rows; these p-values have been calculated using the Wild cluster bootstrap
ur
procedure of Cameron et al. (2008) to correct for the small number of clusters. The estimation
sample includes all married women ages 15-49 at the time of survey in the 9 regions that
implemented reforms plus the two that have not (Afar and Somali). Table B1 shows these dates.
Jo
Appendix C shows that estimates are robust to various alternative specifications. Source: Author’s
calculations using DHS sample, geo-spatial data on district (woreda) boundaries, and dates from
regional legal reforms.
the wedding until she was 16 in order to be compliant, but might not have been willing to delay it until after her 18th
birthday. Therefore, I estimate the treatment effect of the reform on the probability of marriages at ages under 16, at 16
or 17, and at ages above 18. These results appear in Columns 2 through 4 of Table 2 Panel A. The magnitudes of the
coefficients suggest that the probability of marriage at ages under 16 (Column 2) decreased by about 3.5 percentage
points, or just under 10 percent relative to the pre-reform baseline mean. The probability of a marriage involving a
girl aged 16 or 17 (Column 3) does not appear to have changed substantially, and is not statistically significant. On the
other hand, the probability of marriage at age 18 and above increased by about 2.24 percentage points.
These modest effects could mask a more nuanced story. Appendix Figure A4 Panels A and B show that the average
age at marriage differed substantially prior to the reform across districts in Ethiopia.21 These differences across space
imply that the exception age of 16 might have more influence in some areas where marriage under 16 was common
but in other areas where the average marriage occurred at older ages prior to the reform, the new age of 18 might have
more bite. To test this hypothesis, I classify districts as “noncompliant” by indicating whether the pre-reform average
age of marriage in each district was under age 18 immediately prior to the reform in each region.22 Then, I interact the
21 These figures show all districts combined. There is also substantial variation in pre-reform averages across districts within the same region:
the intra-region correlation of district level pre-reform mean age at marriage is 0.35, suggesting that less than half of the variance across districts in
pre-reform averages is explained by within-region characteristics.
22 Utilizing pre-policy average levels to identify treatment intensity has precedent in Lucas and Mbiti (2012), Chicoine (2020), and Bleakley

“Post Reform” treatment with this pre-reform noncompliance indicator and find that, indeed, a more nuanced pattern
of treatment effects emerges in Panel B of Table 2.
Column 1 of Table 2 Panel B shows that the increase in age at marriage was concentrated in districts where the
average marriage was “noncompliant” with the new minimum of 18 just before the reform took place. But examining
Columns 2 thru 4 reveals that the effect of the reform differed across space with respect to where exactly in the age
distribution the marriages of girls were delayed. In “compliant” districts where the average age of marriage was already
above 18 pre-reform, the reform had no effect on the probability of marriage under age 16 (which was already low
at baseline), but the probability of marriage at ages 16 and 17 decreased by about 2.3 percentage points (Column 3).
Correspondingly, the probability of marriage over 18 increased by about the same magnitude. On the other hand,
in noncompliant districts (where the average pre-reform was not yet 18), the probability of marriage under age 16
decreased by about 6.8 percentage points, or about 15% relative to the baseline mean (the interaction effect in Column
2). In those districts, the reform actually increased the probability of marriage at ages 16 and 17 (Column 3, the sum of
the main effect and the interaction effect coefficients). The reform still increased the probability of marriages over 18 in
these districts, given the sign of the main coefficient, but the effect did not differ significantly across districts. Together,
these results suggest that the reform encouraged delays in marriages across the country and shifted the distribution of
marriage ages upward, just as the raw data presented in Figure 2 suggested.
f
oo
4.2. Event study analysis
Figures 4 and 5 graphically show coefficients from estimating the treatment effect of the reform using an event study
pr
design as described in Equation 2. These figures lend credence to the assumption of no pre-trends in the probability
of marriage in these ranges since the pre-reform coefficients are all tightly concentrated around zero. Post-reform,
the effect sizes appear to grow somewhat toward the later end of the treatment period (6 to 8 years pre-reform), but
-
re
the figures also show that the reform had immediate and sizeable impacts in reducing early marriages (under 16) and
increasing the probability of marriages over 18.23 The figures for “compliant” and “noncompliant” districts reinforce
the findings in Table 2 Panel B: the reform generated a shift upward in the distribution of marriage ages that differed
lP
across space in precisely where in the distribution these changes occurred.24
4.3. Heterogeneity in Effects of the Reform by Ethnic Group

na
There are more than 80 ethnic groups in Ethiopia.25 Nearly all Ethiopian ethnic groups practice some form of a
bride price tradition, which Corno et al. (2020) show can keep marriage ages low for girls, especially in response to
ur
adverse income shocks. On the other hand, there is heterogeneity across the major ethnic groups in early marriage rates.
Appendix Figure A5 shows that women from the Amhara group have the highest rates of early marriage pre-reform,
Jo
more than double that of others, with the next highest rates among Tigrie women. Perhaps surprisingly, Amhara and
Tigrie women also have higher average educational attainment. Attendant to these education differences, women from
these groups also have higher values of the decision-making and domestic violence indices. To explore a potential
reason for these differences at baseline, I incorporate pre-industrial ethnographic data from Murdock’s Ethnographic
Atlas.26 Interestingly, the Amhara and Tigrie groups are the only major ethnic groups in Ethiopia that utilized plow
agriculture historically according to Murdock’s Atlas, which Alesina et al. (2013) and Carranza (2014) show implies
limited demand for women’s and children’s labor in agriculture.27 This historic economic difference suggests a low
opportunity cost of girls’ time, which could lead to both early marriage rates and high education rates.28
(2010), among others. See Appendix A for more details on construction of this variable. Appendix Table C2 shows that results are robust to using
pre-2000 district averages to classify them instead.
23 Sample sizes are smaller for the post-reform periods, so the confidence intervals are larger, but they almost all suggest that the effects are
statistically significantly different from zero.

24 Since Table 2 Panels A and B showed the importance of considering these age ranges in the distribution, these event study graphs are my
preferred set of results.

25 Some of them are very small. The five major groups, which together comprise more than 75% of the population, are: Affar, Amhara, Oromo,
Somalie, Tigrie.
26 Specifically, I follow Alesina et al. (2016) and match the ethnicity names in the DHS to those in the Ethnographic Atlas
27 Specifically, Alesina et al. (2013) and Carranza (2014) document both cross-country and individual level empirical evidence suggesting that
exogenous variation in the workability of soil led to adoption of plow technology by some ethnic groups and not others, which in turn generated
differences across ethnic groups in the relative returns to and demand for women’s labor. Among those societies with historic plow use, women
were less involved in agriculture and women’s outcomes tend to reflect more unequal positions in society for women.
28 If girls are not helpful to their parents in agriculture, then families do not benefit from keeping them at home as much as sending them to be
married (especially in the context of a bride price tradition) or to go to school. Alesina et al. (2011) also suggest that fertility rates would also be
lower, as the demand for children as helpers in economic life is lower when using plow technology. This may also lead to more quality investments

Table 3
Effect of the Reform on the Distribution of Marriage Ages
Panel A: Women from ethnic groups using the plow
Post Reform 0.0265 -0.00159 -0.0334 0.0350∗
(0.0411) (0.0277) (0.0309) (0.0178)
Post*Noncompliant 0.0528 -0.0803∗ 0.114∗∗ -0.0334
(0.0469) (0.0377) (0.0287) (0.0248)
Observations 9069 9069 9069 9069
p-value CGM 0.553 0.934 0.291 0.0731
p-value CGM (interaction) 0.132 0.187 0.114 0.925
Baseline mean 16.30 0.528 0.174 0.298
Panel B: Women from ethnic groups not using the plow

Post Reform -0.0168 0.00843 -0.0456∗ 0.0372
f
(0.0263) (0.0184) (0.0227) (0.0207)
oo
Post*Noncompliant 0.0281 -0.0571∗∗ 0.0784∗∗ -0.0212
(0.0376) (0.0218) (0.0262) (0.0227)
Observations 17020 17020 17020 17020
pr
p-value CGM 0.678 0.472 0.168 0.301
Baseline mean 17.06 - 0.395 0.231 0.374
re
Notes: The table reports estimates from the same specifications as in Table 2 Panel B but separately
by ethnic group depending on pre-colonial use of the plow in agriculture as indicated by Murdock’s
Ethnographic Atlas. These include an interaction effect of “Post-reform” with a dummy indicating that
lP
the woman resides in a “Noncompliant” district (woreda), where noncompliance is determined by

whether the average age at marriage in the district is below 18 during the 5 years prior to the reform
in each region. A dummy for “Noncompliant” is also included, but not shown in the table for
simplicity. For more details on construction of this variable, see Appendix A. Included in each
specification are: region fixed effects, year of marriage and year of survey fixed effects, region-specific
na
time trends, respondent’s age at interview, a rural dummy, and ethnicity fixed effects. Robust
standard errors in parentheses are clustered at the district (woreda) level, since district-level variation
is important in identifying the treatment effects. Significance levels show * 𝑝 < 0.10, ** 𝑝 < 0.05, ***
𝑝 < 0.01, where these have been adjusted for multiple hypothesis testing using the False Discovery
ur
Rate procedure described in Benjamini et al. (2006). p-values from clustering standard errors at the
region level (with 11 clusters) appear in the bottom rows; these p-values have been calculated using
the Wild cluster bootstrap procedure of Cameron et al. (2008) to correct for the small number of
clusters. The estimation sample includes all married women ages 15-49 at the time of survey in the 9
Jo
regions that implemented reforms plus the two that have not (Afar and Somali). Table B1 shows
these dates. Appendix C shows that estimates are robust to various alternative specifications.
Source: Author’s calculations using DHS sample, geo-spatial data on district (woreda) boundaries,
and dates from regional legal reforms.
These large differences in women’s marriage, education, and empowerment outcomes across ethnic groups at
baseline imply that the effects of the age of marriage reform may also be heterogeneous by ethnic group. To identify
whether the effect of the age of consent law reform was different across women in these groups, I estimate the results
from Table 2 separately for women from groups that used the plow (Table 3 Panel A, which includes Amhara and
Tigrie women as well as those from several smaller ethnic groups) and for women from groups that did not use the
plow (Table 3 Panel B). The results suggest that for women from ethnic groups using the plow (whose early marriage
rates were higher at baseline) the effects of the reform are larger relative to the effects among women from ethnic
groups with no historic plow use. However, the effects are not statistically significant after using the most stringent
corrections to p-values.
4.4. Robustness to impacts of other interventions

Ethiopia’s poverty rates are high and the agendas of the government and of development organizations are full. It
is important to ensure that the results discussed in this paper are not contaminated by other policy initiatives. While
the set of fixed effects and controls I include in my empirical analysis allays many of these concerns, I also report on a
in children’s education.

Table 4
Effect of the Reform on Human Capital Outcomes
Any School Years (incl. 0) Employed

Post Reform -0.00540 0.243 0.0222
(0.0137) (0.150) (0.0171)
Post*Noncompliant 0.0326 -0.191 -0.0310∗
(0.0198) (0.181) (0.0186)
Observations 41852 41852 41712
p-value CGM 0.723 0.107 0.180
p-value CGM (interaction) 0.289 0.745 0.789
Baseline mean 0.294 1.898 0.513
Notes: The table reports estimates from the same specifications as in Table 2 Panel B for outcomes
related to human capital and labor. These specifications include an interaction effect of
“Post-reform” with a dummy indicating that the woman resides in a “Noncompliant” district (woreda),
where noncompliance is determined by whether the average age at marriage in the district is below
18 during the 5 years prior to the reform in each region. A dummy for “Noncompliant” is also
f
included, but not shown in the table for simplicity. For more details on construction of this variable,
oo
see Appendix A. Included in each specification are: region fixed effects, year of marriage and year of
survey fixed effects, region-specific time trends, respondent’s age at interview, a rural dummy, and
ethnicity fixed effects. Robust standard errors in parentheses are clustered at the district (woreda)
level, since district-level variation is important in identifying the treatment effects. Significance levels
show * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01, where these have been adjusted for multiple hypothesis
pr
testing using the False Discovery Rate procedure described in Benjamini et al. (2006). p-values from
clustering standard errors at the region level (with 11 clusters) appear in the bottom rows; these
p-values have been calculated using the Wild cluster bootstrap procedure of Cameron et al. (2008) to
-
correct for the small number of clusters. The estimation sample includes all married women ages
re
15-49 at the time of survey in the 9 regions that implemented reforms plus the two that have not
(Afar and Somali). Table B1 shows these dates. Appendix B shows that estimates are robust to
various alternative specifications. Source: Author’s calculations using DHS sample, geo-spatial data
on district (woreda) boundaries, and dates from regional legal reforms.
lP
battery of tests that address specific policies occurring around the same time. My estimates remain stable throughout,
na
confirming that the effects on marriage ages observed in this paper cannot be explained by these other programs.
The first set of robustness tests relates to education reforms. In the 1990s, Ethiopia’s free primary education (FPE)
program removed school fees for grades primary school (grades one through ten) and implemented a mother tongue
ur
instruction (MTI) program in some regions. However, empirical evidence demonstrates that the results I estimate
in this study cannot be explained by these programs. First, Chicoine (2019) and Chicoine (2020) show that while
Jo
FPE was successful in increasing educational attainment (by 0.7 years), MTI actually interrupted schooling and had
a negative impact on educational attainment. Second, marriages for women exposed to FPE were delayed, but only
for women over age 20, thus an older group than that treated by the minimum age. Third, if education reforms were
influencing the age at marriage results I observe in this study instead, we would expect to see an increase in these two
indicators of educational attainment. However, the results in Table 4 show that women married after the family law
reform, and therefore treated by the new minimum age of marriage, are no more or less likely to have ever been in
school and have no more years of schooling on average.29 As a final check for whether trends in schooling are driving
the results on marriage I observe in this study, I employ an adjustment described in Freyaldenhoven et al. (2019) for
controlling for confounding factors: I estimate the results in Table 2 Panel B including a control for whether the woman
has any schooling (which itself would have been determined about 9 years pre-reform for girls aged 16 at the time of
the reform) but I instrument for this variable with the average ever schooling rate of women married in the two years
post-reform (leads). These results are reported in Table C5 and they show that including this instrumented control
does not substantively change the results from Table 2.
Similarly, the results in this study cannot be explained by the NGO program Berhane Hewan, which was designed
to address early marriage in a multi-faceted intervention in East and West Gojjam, two zones in the Amhara region.
The program provided girls in these rural areas with training programs, school supplies, family planning information,
and in-kind transfers to families on the condition that girls remained unmarried for two years (Erulkar and Muthengi,
29 The coefficient on the interaction of Post with “noncompliant” districts is about 3.2 percentage points, but the combined effect is smaller and
not statistically significant when the Wild bootstrap correction is performed. On years of education, I do find a small and positive impact on girls
that is mostly concentrated in “compliant” districts but is again noisily estimated. I conclude, then, that the reform’s effect on age at marriage cannot
be explained by educational advancements generated by the FPE reform.

2009). Alongside these direct engagements, the program also held community conversations on early marriage as
a harmful traditional practice, similar to the kinds of conversations that were occurring at the local level across the
country (Convention on the Elimination of All Forms of Discrimination against Women (CEDAW), 2010). While this
program has received a lot of attention in reports from NGOs on what works in fighting child marriage, it was relatively
small: only two districts in the DHS data were covered by this program (Mencha and Yilmana Densa, in West Gojjam
Zone). This represents less than 1% of the sample. My results are robust to excluding women from these districts.
Table C7 demonstrates robustness of the main results to other programs beyond education and early marriage.
Panel A shows that my results are robust to including a control for whether women live in districts where Ethiopia’s
Productive Safety Net Program (PSNP), a large food security program,30 was implemented first. While women in
early PSNP districts have younger marriage ages on average, including this dummy leaves almost unchanged the point
estimates and statistical significance of my main results of the effect of the reform on women’s early marriage. Panel
B of Table C7 shows that including controls for whether women had been exposed to family planning programming
(either on the radio or in the form of visits from extension workers) also does not change the main results from Table
2.
As an additional check to ensure that the effects I observe in this study arise from the specific timing associated
with the family law reform changes in each region, I implement a simulation-based placebo reform test. Specifically, I
f
draw 1000 sets of placebo reform dates, where each region is assigned to a random month during the period covering
oo
8 years prior to and 8 years following the federal reform in 2000.31 I estimate the treatment effect coefficients from
Equation 1 for each set of these placebo dates. As Figure 6 shows, the estimate I report for the interaction effect
pr
(Post-reform∗Non-compliant) in Column 2 of Table 2 is well outside of the 99% confidence interval of these placebo
effects.32 This strongly suggests that the effects observed in this paper on the effect of the reform on early marriage
-
are only observable when the precise dates of the family reform are used. When random dates, some of which would
re
by chance coincide with the dates of other ongoing interventions, are used, I do not estimate “effects” that are nearly
as strong.
lP
5. Secondary outcomes and mechanisms

na
The results discussed thus far suggest that the reform delayed women’s marriage, and in particular delayed mar-
riages of girls under 16 in areas where early marriage was more common prior to the reform. Section 4.4 provided
empirical evidence to suggest that these effects cannot be attributed to other ongoing policy interventions. I showed that
ur
women’s educational and employment outcomes do not appear to have systematically improved as a result of waiting
for marriage in areas where early marriage was most common. These outcomes on human capital are consistent with
Jo
the predictions from Field and Ambrus (2008) who, using simulations based on their IV estimates (using menarche as
the instrument), predicted that the gains in education resulting from an age of consent law could be achieved just as
well with a minimum age of 16 as with a minimum of 18.
Still, women’s outcomes beyond education and employment may change as a result of later marriage if i) women’s
maturity at marriage affects knowledge or behavior within marriage or ii) women who marry later attain a different
match in the marriage market (Chari et al., 2017; Field and Ambrus, 2008). Importantly, these changes need not
be improvements per se. If women who marry later feel the need to compensate in some way for the “undesirable”
quality of being older, or if their husbands’ qualities are less desirable due to the match they were able to secure when
marrying older, women might experience losses in real or perceived outcomes due to later marriage. Chari et al.
(2017) find evidence from India supporting this possibility: while later marriage due to later menarche improves some
outcomes (child health investments, enrollment, and cognitive scores, and the likelihood that women are free to go
places without permission or to eat at the same time as their husband) even among women with no education, they find
that later marriage reduces women’s self-reported say in household decisions and freedom to go places alone. In other
words, the effect of later marriage on outcomes within marriage is theoretically ambiguous, and the previous literature
suggests mixed results on these outcomes. To examine whether these patterns might hold for the Ethiopian context
30 Depending on availability of workers in the household, transfers are either unconditional (e.g. female-headed households) or conditional on
participation in public works programs during the non-growing season. This cash transfer system can be seen theoretically as generating a pure
income effect on families’ consumption of goods including schooling, and might cause an increase in schooling, a decrease in the prevalence of
early marriage, or both.
31 I choose 8 years because the last reform in the regions that have reformed occurred eight years after the federal reform.
32 Estimates for other outcomes are available on request.

Table 5
Effect of the Reform on Fertility and Health Outcomes
Panel A: Fertility Outcomes
Has Any Kids Ideal num kids Num Kids if > 0 Age First Birth
Post Reform -0.0287∗ 0.0672 0.137∗ 0.179
(0.0130) (0.0926) (0.0458) (0.0953)
Post*Noncompliant 0.0329∗∗ -0.241∗ -0.0883 -0.275
(0.0140) (0.0998) (0.0795) (0.0903)
Observations 41852 31696 41852 30979
p-value CGM 0.0340 0.519 0.125 0.180
Baseline mean 0.887 5.508 2.720 19.27
Panel B: Health Outcomes

Home delivery Antenatal visits Under 2 mortality
Post Reform -0.0815∗∗ 0.00497 -0.0121
(0.0140) (0.0159) (0.00562)
f
oo
Post*Noncompliant 0.0545* -0.0182 0.00168
(0.0194) (0.0187) (0.00594)
pr
p-value 0.00300 0.711 0.0591
Baseline mean 0.420
- 0.352 0.0835
re
related to fertility and health. Home delivery and antenatal visits are normalized by the number of
children. These specifications include an interaction effect of “Post-reform” with a dummy indicating
that the woman resides in a “Noncompliant” district (woreda), where noncompliance is determined by
lP
whether the average age at marriage in the district is below 18 during the 5 years prior to the reform
in each region. A dummy for “Noncompliant” is also included, but not shown in the table for
simplicity. For more details on construction of this variable, see Appendix A. Included in each
specification are: region fixed effects, year of marriage and year of survey fixed effects, region-specific
na
time trends, respondent’s age at interview, a rural dummy, and ethnicity fixed effects. Robust
standard errors in parentheses are clustered at the district (woreda) level, since district-level variation
is important in identifying the treatment effects. Significance levels show * 𝑝 < 0.10, ** 𝑝 < 0.05, ***
𝑝 < 0.01, where these have been adjusted for multiple hypothesis testing using the False Discovery
Rate procedure described in Benjamini et al. (2006). p-values from clustering standard errors at the
ur
region level (with 11 clusters) appear in the bottom rows; these p-values have been calculated using
the Wild cluster bootstrap procedure of Cameron et al. (2008) to correct for the small number of
clusters. The estimation sample includes all married women ages 15-49 at the time of survey in the 9
Jo
regions that implemented reforms plus the two that have not (Afar and Somali). Table B1 shows
these dates. Appendix C shows that estimates are robust to various alternative specifications.
Source: Author’s calculations using DHS sample, geo-spatial data on district (woreda) boundaries,
and dates from regional legal reforms.
arising from changes in marriage age due to the age of consent law, I explore similar outcomes to Chari et al. (2017)
in Tables 5 and 6.
5.1. Results for secondary outcomes

Table 5 shows that women married post-reform in compliant districts are about 2.9 percentage points less likely to
have any children by the time of interview (Column 1), conditional on their age. This main effect coefficient has a p-
value of 0.03. However, the interaction term combines with the main effect to a zero effect for women in noncompliant
districts. Column 2 shows that in noncompliant districts, women’s ideal number of children decreased. Conditional
on having children, the number of children increases in compliant districts (Column 3). Column 4 shows that, among
women who have children, their age at first birth was later on average in compliant districts (though noisily estimated);
in noncompliant districts, there was no effect on age at first birth.
While the positive effect on number of children and weak effect on age at first birth might seem counter-intuitive,
it is important to keep in mind that any effects observed for women with children (e.g. number of children, indicators
of child health) may be biased due to the selection effect that Column 1 implies. That is, since women married post-
reform in compliant districts are less likely to have any children at survey date (the extensive margin), those who
do have children may differ systematically from those who do not in ways that affects outcomes on the intensive as

well. Together, the extensive margin effect and the effect on the ideal number of children suggest that women married
post-reform, who married later on average, will have fewer children over their lifetime. This result is consistent with
simulations in Wodon et al. (2017) suggesting that age of consent laws can lower aggregate fertility rates.
Beyond fertility timing and quantity effects, delaying marriage may lead to improvements in women’s knowledge
about or access to better quality prenatal, delivery, or postnatal care in ways that lead to other gains in child health
(Chari et al., 2017). Panel B of Table 5 examines this. Column 1 shows that women married post-reform were 7
percentage points less likely to deliver babies at home. On the other hand, the number of prenatal visits (Column 2)
does not appear to have changed. Still, a supervised delivery outside of the home may translate into better outcomes for
children, since the coefficient on under 2 mortality rates of children suggests that it decreases (Column 3). However,
this estimate is noisily estimated, as was the estimate in Chari et al. (2017), likely due to low statistical power for this
relatively rare event. A lower mortality rate would also help to explain the positive coefficient in Panel A Column 3
on the number of children.33
In summary, the results on secondary outcomes imply that women married later delay birth and desire fewer total
children. Those who do have children have slightly more on average for their age, but this is likely due to either
selection effects (delays in birth mean that those who have given birth by a certain age are differentially selected) or
to mortality effects arising from increased probability of having a birth attended by a professional. These results are
f
consistent with the previous literature (Chari et al., 2017).
oo
5.2. Mechanisms
pr
In the absence of gains to education (which Table 4 verified), there are two potential mechanisms through which
later marriage could generate the effects on fertility and health investments described above. The first is a pure age
effect. In some ways, an age effect is simply mechanical34 , but also more mature women might be better able to delay
-
pregnancy relative to marriage, and they might desire fewer children.35
re
Another potential mechanism that could lead to fewer births and better care at the time of birth is the “quality”
of spouse. Theoretically, delayed marriage could either improve or disimprove quality of matches for some women
lP
in the marriage market in partial equilibrium. However, in a general equilibrium response to the legal change in the
age at marriage, we should not expect to observe substantial changes in spousal quality when measured using men’s
characteristics like education and employment.36 On the other hand, delaying marriage may yield a better quality
na
match for most women if match quality is increasing in time spent searching (whether on the part of parents or the
bride herself). We might measure the quality of these matches on the marriage market using the rate of divorce (Rotz,
ur
2016). Ethnographic data collected before 1992 suggests that divorce was rather common in Ethiopia (Pankhurst,
1992). However, it had been trending down by 2000, when the reform was enacted. Among women married during
Jo
the 5 years prior to the reform, 8% of them were divorced by the time of survey, and an additional 4% were separated.37
If later marriage leads to better matches for both men and women, then marital stability might also improve, resulting in
less divorce (Rotz, 2016). A lower divorce rate could also help to explain the positive effect I observe on the number of
children: if marriages of women married post-reform, at older ages on average, are more stable, then conditional on age
we might expect that women who have been married for longer would have more children. Column 1 of Table 6 shows
that women married post-reform are 2.17 percentage points less likely to be divorced at the time of survey compared
to women who were married pre-reform, when marriage under age 18 was more common (though this estimate is a
bit noisy after accounting for multiple hypothesis testing).38 However, the interaction effect combines with the main
effect to imply a zero change in divorce rates for women in noncompliant districts. It is important to note, however,
that without information on the timing of the divorce (which is not available in the DHS), it is impossible to know
whether this effect on divorce is purely mechanical (later marriage and a fixed average duration until divorce implies
33 Maternal mortality could be an omitted factor that would generate downward selection bias in these estimates, if women with high-risk births
are no longer alive to be respondents in the DHS. Such selection bias would lead the estimates observed here to overestimate the effect of the reform
on under 2 mortality rates, and might also generate concerns for selection in other outcomes.
34 If pregnancy follows with some fixed probability after consummation of a marriage, then conditional on age a woman will be less likely to
have a child on the day of survey.

35 More mature women might also have more information simply from life experience about the importance of medical assistance during delivery.
36 Chari et al. (2017) use an index composed of men’s education and employment characteristics to proxy for spousal quality.
37 These numbers are consistent with those from the Ethiopian census data available on IPUMS.
38 Since I do not have access to information about whether women were previously divorced if they are currently married, this negative effect on
divorce rates also corresponds to a small, but noisily estimated, positive effect of post-reform marriage on the likelihood that a woman is married
at the time of survey. It is important to keep in mind, though, that the results on the impact of the reform on the likelihood of marriage at different
ages focuses on first marriage.

that, conditional on age, women have not yet reached the point of divorce) or if this reflects a difference in match quality
generated by the effects of the reform.
Finally, I also investigate the effect of the reform on several outcomes related to women’s perceived empower-
ment39 , similarly to that of Field and Ambrus (2008) and Chari et al. (2017). Specifically, in Panel A Column 2 I
consider the effect of the reform on an index collecting responses about women’s self-reported participation in decision-
making about their healthcare, large purchases for the household, and visiting friends and relatives. The results suggest
that there are no changes in women’s reported say in decision-making. Column 3 shows that women in noncompliant
districts are more likely to report that their husband wants to have more children than they do, which is consistent with
responses from Table 5 on the ideal number of children. In Panel B, I report additional results from women’s responses
to questions about perceived empowerment. Column 1 of Panel B shows that women married post-reform are more
likely to report that they have the freedom to seek medical help for themselves when needed. Column 2 shows that
women in noncompliant areas are more likely to report that they feel comfortable asking their husband to use a condom
in the event that he is known to have an STI. Finally, in Column 3 I use an index to collect women’s responses to the
questions “is it your opinion that a man is justified in beating his wife if she... [burns the food, goes out without telling
him, etc.]” While positive, the results in Panel B are not statistically significant when p-values are adjusted for multiple
hypothesis testing. In assessing these results, one must bear in mind that these self-reports may be mismeasured in
f
ways that might systematically relate to other outcomes but are inherently unobservable (Glennerster et al.). Impor-
oo
tantly, measurement error in the dependent variable may lead to attenuation bias of estimates of the reform. These
results together imply that women’s perceptions about their ability to make decisions are not statistically significantly
pr
different among those married post-reform (at later ages on average), but may improve over time.
Altogether, I find that delaying marriage likely leads to lower total fertility for women and improved take-up of
-
formalized healthcare at the time of delivery. These gains are not generated by gains in education nor by marrying
re
men of better observable quality. Rather, delaying marriage and the maturity – and, potentially, empowerment – that
comes with it lead to improved outcomes for women’s fertility health. As Chari et al. (2017) find, these gains may in
lP
turn lead to improved outcomes for women’s children.
5.3. Ruling out other components of the Family Law reform

na
It is also important to discuss other components of the family law reform, which occurred in a package with the
same roll-out schedule. Specifically, the two main novel stipulations of the law besides the change to marriage age 1)
allowed women to work outside the home without spousal approval, and 2) provided for equitable division of property
ur
upon the dissolution of marriage. While I must caution the reader that I cannot entirely disentangle the effect of
the reform’s stipulation about marriage age from potential separate effects of these other components of the legal
Jo
reform, the combination of several pieces of empirical evidence strongly suggests that the marriage age stipulation
had a separate effect on its own. First, I find no statistically significant impact of being married post-reform on the
probability of employment among married women (Table 4 Column 3). This suggests that my results are not being
driven by changes in employment outcomes, which we would expect to see based on the release of married women to
work without spousal approval.40 Second, I find no evidence that women’s self-reported land ownership rates are higher
for women married post-reform (Table C7 Panel C). If the new stipulations about division of assets on divorce were
driving the results I present in this paper on marriage age and secondary outcomes, we would expect to see women’s
self-reports of land ownership also increasing.41 The effect on home ownership is negative, not positive, indicating a
negative treatment effect. These results are consistent with later marriage but are not consistent with stipulations about
property division at divorce affecting marriage ages.
Finally, while marriage ages may indeed change in response to improved property rights for women or higher rates
of employment among young unmarried women, we would expect these effects to be gradual, and to be complementary
with changes in educational attainment. However, the event study figures (Figures 4 and 5) show that the effect of the
reform on the distribution of marriage ages was sudden, not gradual. Together, these results suggest that it is unlikely
that the effects observed in this paper on age of marriage can be attributed to these other aspects of the reform.
39 I only use these DHS questions about women’s empowerment outcomes if they were included in all years of the DHS in Ethiopia.
40 I note here that Hallward-Driemeier and Gajigo (2011) observed that over the period from 2000 to 2005, women’s labor supply increased
faster in the early reform regions than in the later reform regions. However, this analysis does not employ the precision in time variation allowed by
comparison of marriage timing to the exact dates of reform.
41 Land is the most commonly held asset in Ethiopia, and the most frequently contributed by partners upon marriage
Fafchamps2005,Quisumbing2003. This finding is consistent with Kumar and Quisumbing (2015), which finds that while rural women’s
perceptions of equitable access of land improved between 2000 and 2009, there was no substantial improvement in their well-being.

Table 6
Effect of the Reform on Marital Outcomes and Women’s Empowerment
Panel A: Marital Outcomes
Divorced Decision Index Husband wants more children than respondent
Post Reform -0.0217∗∗ 0.0000251 -0.0117
(0.00983) (0.0562) (0.0127)
Post*Noncompliant 0.0266∗∗ 0.00474 0.0223
(0.0114) (0.0485) (0.0134)
p-value CGM 0.0531 0.998 0.460
Baseline mean 0.0842 -0.211 0.249
Panel B: Perceived Empowerment Outcomes

Can get medical help Can ask for condom DV Index
Post Reform 0.0341∗ -0.0127 0.0640
(0.0189) (0.0138) (0.116)
f
oo
Post*Noncompliant -0.0269 0.0275∗ 0.0252
(0.0167) (0.0150) (0.131)
pr
p-value CGM 0.194 0.373 0.686
Baseline mean 0.695
- 0.684 -0.111
re
related to marriage and women’s perceived empowerment. The Decision and DV indices are
constructed as described by Chari et al. (2017), with larger values coded as more favorable to
well-being. “Can get medical help” and “can ask for condom” are the positive binary opposites:
lP
women are asked whether getting medical help would be a problem and whether they believe that a
wife is justified in asking her husband to use a condom if it is known that he has an STI. These
specifications include an interaction effect of “Post-reform” with a dummy indicating that the woman
resides in a “Noncompliant” district (woreda), where noncompliance is determined by whether the
na
average age at marriage in the district is below 18 during the 5 years prior to the reform in each
region. A dummy for “Noncompliant” is also included, but not shown in the table for simplicity. For
more details on construction of this variable, see Appendix A. Included in each specification are:
region fixed effects, year of marriage and year of survey fixed effects, region-specific time trends,
respondent’s age at interview, a rural dummy, and ethnicity fixed effects. Robust standard errors in
ur
parentheses are clustered at the district (woreda) level, since district-level variation is important in
identifying the treatment effects. Significance levels show * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01, where
these have been adjusted for multiple hypothesis testing using the False Discovery Rate procedure
Jo
described in Benjamini et al. (2006). p-values from clustering standard errors at the region level
(with 11 clusters) appear in the bottom rows; these p-values have been calculated using the Wild
cluster bootstrap procedure of Cameron et al. (2008) to correct for the small number of clusters.
The estimation sample includes all married women ages 15-49 at the time of survey in the 9 regions
that implemented reforms plus the two that have not (Afar and Somali). Table B1 shows these
dates. Appendix C shows that estimates are robust to various alternative specifications. Source:
Author’s calculations using DHS sample, geo-spatial data on district (woreda) boundaries, and dates
from regional legal reforms.
6. Discussion
Early marriage has been a strong norm in Ethiopia for decades, as in other parts of the developing world. In
this paper, I show evidence suggesting that Ethiopia’s Revised Family Code resulted in delays in marriage for girls.
Between 1990 and 2009, the proportion of marriages in a given year that involved girls younger than 16 dropped from
50% to 25%, and my results suggest that at least a fifth of that decrease can be attributed to revision of the law. These
results are robust to consideration of several other ongoing programs related to poverty reduction, education, and land
titling. Further, results suggest that even in areas where compliance with the previous minimum age was relatively
low, the effect of the reform is evident.
Still, the magnitude of the effects in this paper and the raw evidence provided in Figure 2 also demonstrate imperfect
compliance with the law. As recently as 2016, 18% of 17 year old girls were already married. For women aged 18, the
proportion was 35%. There is little evidence available to suggest that violations of the law have been widely prosecuted
(Mekonnen and Aspen, 2009). Doing so, especially in rural, high poverty areas where child marriage is most common,
is difficult to coordinate.

One is left to wonder, then, why the law was effective at all. The key may lie in the information campaign that
accompanied the change in the legal age of consent, as documented by the Ethiopian government, ethnographic work,
NGO reports, and news articles (Boyden et al., 2013; Pankhurst, 2015; Convention on the Elimination of All Forms of
Discrimination against Women (CEDAW), 2010).42,43 Alongside the Revised Family Code, the Ethiopian government
undertook a large information campaign at all levels of government to address early marriage for girls (Mekonnen and
Aspen, 2009). At the sub-regional and sub-district levels, committees galvanized to raise awareness of the law, studying
local practices and using “awareness-creation and promotion campaigns" along with watch-dog committees to mobilize
community members to speak out against harmful traditional practices of child marriage and female circumcision.
Non-governmental organizations - both local and international - also helped to advertise the law and the harms to
girls of early marriage. While knowledge of the revised minimum age is still not universal in some rural areas (Chow
and Vivalt, 2019; Convention on the Elimination of All Forms of Discrimination against Women (CEDAW), 2010),
ethnographic data confirm that many are aware, though some are still opposed.44 This wide-sweeping information
campaign means that the estimates obtained in the paper cannot disentangle the impact – or whatever instances there
are of enforcement – of the law per se from the community based efforts to advertise it. However, without such
legal literacy campaigns about the risks of child marriage and provisions of the new law, it would likely go unknown in
communities with the highest rates of child marriage considering Ethiopia’s low literacy and radio/television ownership
f
rates. With other variation in legal changes, information, and enforcement, future work might be able to disentangle
oo
these effects.
The results from this paper, like those of others (see, for example, Harari (2019)), suggest that legal reforms and
pr
the legal literacy campaigns that accompany them can change behavior and outcomes even in the face of strong cul-
tural traditions and poor enforcement. Institutions and cultural traditions around gender norms can be slow moving,
-
but institutional reforms can also bring changes that are cumulative and self-reinforcing (North, 1990; Kumar and
re
Quisumbing, 2015; Voigt and Kiwit, 1998). This paper suggests, however, that legal reform and mass information
campaigns can also generate abrupt changes in behavior, as do other interventions targeting norms by addressing
lP
asymmetry in information (Bursztyn et al., 2020; Dhar et al., 2018; Stopnitzky, 2017; Madajewicz et al., 2007). How-
ever, the heterogeneity results in this paper suggest that for women in ethnic groups with the strongest norms toward
early marriage and low education levels for girls, the legal change and its information campaign may not have been
na
as effective. Legal signposts and information, in some cases, may not be sufficient to reach the goal of ending child
marriage by 2030.
ur
A. Data Details
Jo
I address here several aspects of the data in more detail.
A.1. Age at marriage and possible measurement error

The DHS uses reported age at first cohabitation in its documentation on ages at marriage. The survey does not
separately collect age at first marriage if it differed from age at cohabitation45 . In some parts of the world where
child marriage is common, these dates may differ by a number of years if young girls are “betrothed” and living with
her future husband’s family, but not yet engaging in the full duties of marriage. Of course, cohabitation may occur
voluntarily before marriage as well.
However, I argue that these distinctions pose little cause for concern regarding the estimates obtained in this study
for several reasons. The correlation of age at first cohabitation and age at first intercourse is very strong at 0.85.
Seventy-one percent of married women report their age at first intercourse to be the same as their age at first marriage
(cohabitation). Among the 15 percent of married women who report their first intercourse as occurring in a year later
than that of their first cohabitation, the average gap is 1.9 years; among the other 15 percent of married women who
42 Some local community organizers say that some villagers may not be aware that punishment can be incurred even after the marriage ceremony,
while still others cite the stigma associated with divorce as a reason for remaining married despite the law’s implications.
43 See Walters, A. 2014. “Ethiopian girls as young as five married off.” Aljazeera, August 2014.
http://www.aljazeera.com/news/africa/2014/08/ethiopian-girls-as-young-as-five-married-off-20148269524221721.html . Pankhurst, A. 2015.
“Keeping child marriages trending down.” New Business Ethiopia, June 2015. http://newbusinessethiopia.com/index.php/resources/18-nbe-
blog/470/keeping-child-marriages-trending-down [Accessed October 1 2015].
44 In the extreme, Mekonnen and Aspen (2009); Convention on the Elimination of All Forms of Discrimination against Women (CEDAW) (2010)
note that during fieldwork in the Maqet district of Amhara, the region where the prevalence of early marriage is highest, most individuals consulted
during interviews were aware of, but opposed to, the new law. Cosier, C. 2015. “How an Ethiopian priest changed his views on child marriage.”
Public Radio International, March 2015. http://www.pri.org/stories/2015-03-11/how-ethiopian-priest-changed-his-views-child-marriage.
45 Data for ages at subsequent marriages or cohabitation arrangements are not collected

report their first intercourse as occurring before their first marriage, the average gap is 3.4 years. In the event that
women report being unmarried but have cohabited with a partner, my analysis treats them as never married. Of course,
women might misreport the timing of marriage due to recall bias, whether from intentional, systematic misreporting
or from a difficulty of remembering or lack of accurate knowledge about dates. Such measurement error might cause
concern in use of the information for estimating the effects of the reform. I provide a few pieces of evidence to allay
concerns about such measurement error. First, Figure A2 shows the distribution of ages of marriage among women
who were 18 - 33 at the time of survey in 2000 by region, in order of the timing of the reform (left to right in each row).
The distribution of marriage ages shows that women reported ages younger than 15 (the previous minimum age) prior
to the reform and continue to do so afterward. These graphs also suggest that across regions there is some variation
in the distributions of timing of marriage, but not in a manner that is systematically related to the timing of adoption
of the reform.46 Similarly, Figure A3 shows the distribution of marriage ages by survey year for ever-married women
ages 18-33 at the time of survey. The graphs show that the distribution of marriage ages has shifted up over time, but
that women in 2016 are still reporting having married earlier than age 18 and age 15.
Several features of the DHS survey design also allay concerns about measurement error in marriage ages. First,
the DHS begins by asking respondents for the month and year of first marriage (cohabitation), and imputes values for
those who cannot readily answer with a month and year. Among women in the samples used in this study married
pre-reform, 55% of them are able to recall the month and year of their marriage and the remaining 44% are assigned
imputed values. Among those married post-reform at the time of reform in their region – the “treated” group in this
study – 79.5% of women are able to recall the month and year of marriage at the time of survey. Appendix Figure
f
A1 shows the distribution of the resulting information on age at marriage (including any imputed values) for women
oo
married pre- and post-reform, separately by ability to recall details of their marriage timing. These graphs show that
recorded marriage ages are not systematically bunched at 18 across any method of measuring timing of marriage. This
helps to minimize the risk that individuals report the legal minimum age out of fear of violation of the law despite
pr
having been married before it. These patterns – suggesting very limited misreporting in the surveys for the sake of
demonstrating compliance with the legal age – are consistent with the results of Blank et al. (2009). They document
that self-reports in recall surveys are more accurate as to marriage ages than administrative data originating from legal
-
records where compliance is important. Taking these pieces of evidence together, I conclude that there is little concern
re
of systematic measurement error in ages of marriage reported in the survey that would call into question the results in
this paper.
lP
A.2. Construction of outcome indices

I use two indices to collect women’s responses to questions about their involvement in decision-making and their
perceptions about the justifiability of domestic violence. I follow Chari et al. (2017) in construction of these indices.
na
A.3. Migration
Unfortunately, previous region of residence was not recorded in the DHS data until the latest round in 2016, so I
ur
cannot use this as a control in my analyses without dropping previous rounds. However, the most recent DHS round
(2016) asked women if they had ever lived in a different region and if so how recently they had moved. For the 10,096
Jo
women aged 18 to 33 years who were interviewed in the 2016 DHS, 80.1% report that they have never moved across
regions during their lives. The remaining 1,940 women report having moved across regions previously but more than
50% of these women moved from a previous region of residence to Addis Ababa or Dire Dawa, where the reform
happened earlier, not later. Women who have moved – whether from an early reform region to a region that reformed
later/not yet, or vice versa – are not statistically significantly more likely to have ever married, married early, or become
a mother early. Thus I conclude that selective migration for marriage or for related reasons does not pose a problem
for the identification strategy and empirical results in this paper.
A.4. Pre-reform compliance levels

I match DHS clusters to districts (woredas) using the GPS coordinates and the coordinates of perimeters of districts
available at OpenAfrica (2013). For each district, I calculate the average age at marriage among women married during
the 5 years immediately prior to the reform in their region of residence. As a robustness measure, I also calculate the
average from the years prior to the federal reform. district of current residence is not necessarily that of household of
origin (where the decision of whether to not to marry at a certain age is made), but anthropological evidence suggests
that kinship networks and marriage matches are tightly distributed within rural Ethiopia (Pankhurst, 1992). Panel A
in the following figure shows that these averages vary across districts substantially. Panel B shows that this variation
is not equivalent to a rural-urban divide.
46 Appendix Table B2 shows more formally that the timing of the reform in regions was not systematically related to rates of early marriage,
motherhood, or years of educational attainment.

Figure 2: CDFs of Age at First Marriage

Panel A: 10 Years Pre- and Post-Reform
1
.8
% of women
.4 .6
f
oo
.2
- pr
0
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
re
Age at Marriage
lP
1990-1999 2000-2009
Panel B: Pre- and Post-Reform by 5-year Periods

na
1
ur
.8
Jo
% of women
.4 .2
0 .6
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Age at Marriage
1900-1994 1995-1999
2000-2004 2005-2009
Notes: The figures show cumulative density functions of age at marriage. Panel A shows two
periods: the decade immediately before the reform and the decade immediately after. Panel B shows
the data by five-year intervals. Source: Author’s calculations from DHS data.

Figure 3: Parallel Trends Check: Average Age at Marriage

Age at Marriage rel. to 5 year pre-reform mean
6
f
oo
4
pr
2
-
re
lP
0
na
-2
ur
-4
Jo
10 yrs Pre 5 yrs Pre Reform 5 yrs Post 10 yrs Post 15 yrs Post
Federal Cities Early Adopters

Late Adopters
Notes: The figure shows age at marriage for three groups of regions: the Federal cities (Addis Ababa
and Dire Dawa) and the remaining regions grouped by timing of the reform (early and late). The
exact dates of reform are reported in Appendix Table B1. Each dot represents the average age at
marriage relative to the mean age during the 5 years pre-reform in that region, with pre-reform
means across region normalized to zero. Lines are local polynomials fitted to the data. Source:
Author’s calculations from DHS data.

Figure 4: Event Study Coefficients - All Woredas
Panel A: Under 16
.4
.3
.2
% of Marriages
-.1 0
-.2
-.3 .1
f
-.4
oo
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
Years Pre/Post
pr
Panel B: 16 - 17
.4
-
re
.3
.2
lP
% of Marriages
0 .1
na
-.1
ur
-.2
-.3
Jo -.4
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
Years Pre/Post
Panel C: Over 18
.4
.3
.2
% of Marriages
-.1 0
-.2
-.3
-.4 .1
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
Years Pre/Post
Notes: The figures show event study estimates from Equation 2, where the omitted year is year zero.
Included in each specification are: region fixed effects, year of marriage and year of survey fixed
T. McGavock: Preprint
effects,submitted to Elsevier
region-specific time trends, respondent’s age at interview, a rural dummy, and ethnicity fixed Page 22 of 35
effects. Robust standard errors in parentheses are clustered at the district (woreda) level, since
district-level variation is important in identifying the treatment effects, as described in Section 3.2.
Error bars show 95% confidence intervals. Source: Author’s calculations from DHS data and legal
documentation for reform dates.
Figure 5: Event Study Coefficients - Compliant (L) and Non-compliant (R) Districts
Panel A: Under 16 Panel A: Under 16

.4
.4
.3
.3
.2
.2
% of Marriages
% of Marriages
.1
.1
0
0
-.1
-.1
-.2
-.2
f
-.3
-.3
oo
-.4
-.4
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8 >8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
Years Pre/Post Years Pre/Post
pr
Panel B: 16 - 17 Panel B: 16 - 17
.4
.4
-
.3
.3
re
.2
.2
% of Marriages
% of Marriages
.1
.1
lP
0
0
-.1
-.1
-.2
-.2
na
-.3
-.3
-.4
-.4
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8 >8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
ur
Panel C: Over 18 Panel C: Over 18

.4
.4
Jo
.3
.3
.2
.2
% of Marriages
% of Marriages
.1
.1
0
0
-.1
-.1
-.2
-.2
-.3
-.3
-.4
-.4
>8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8 >8 7-8 5-6 3-4 1-2 0 1-2 3-4 5-6 7-8 >8
Notes: The figures show event study estimates from Equation 2 when each period’s indicator is
interacted with the non-compliant indicator used in Table 2 Panel B. The omitted year is year zero.
Included in each specification are: region fixed effects, year of marriage and year of survey fixed
effects, region-specific time trends, respondent’s age at interview, a rural dummy, and ethnicity fixed
effects. Robust standard errors in parentheses are clustered at the district (woreda) level, since
district-level variation is important in identifying the treatment effects, as described in Section 3.2.
Error bars show 95% confidence intervals. Source: Author’s calculations from DHS data and legal
documentation for reform dates.

Figure 6: Placebo Treatment Coefficients, with Reform Dates Randomly Assigned
f
oo
80
pr
60
-
re
Density
lP
40
na
20
ur
Jo
0
-.06 -.04 -.02 0

Placebo Effect of Reform
Notes: The figure plots the coefficient on “Post-reform” generated from estimating Equation 1 on
data using placebo dates for the reform. For each region, the date of the reform is randomly selected
with replacement from the interval covering 8 years pre-reform and 10 years post-reform. The blue
vertical line indicates the estimate obtained from estimating the effect using the correct dates; that is,
it takes the value that appears in Table 2 Column 1. Source: Author’s calculations from DHS data.

Figure A1: Age of Marriage Reported by Ability to Recall
Panel A: Pre-reform Panel B: Post-reform

Month and Year (N=17792) Month and Year (N=6132)
.1 .2 .3
0 .05 .1 .15 .2
0
Month and Age (N=2990) Month and Age (N=111)

.1 .2 .3
0 .05 .1 .15 .2
0
Fraction
Fraction
Year Only (N=6641) Year Only (N=678)
.1 .2 .3
0 .05 .1 .15 .2
0
Age Only (N=10290) Age Only (N=267)
f
oo
.1 .2 .3
0 .05 .1 .15 .2
0
pr
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Age at Marriage Age at Marriage
-
Notes: The figures show the distribution of age of marriage recorded in the DHS sample of women in
re
this study based on the manner in which the information about age of marriage was obtained. Panel
A shows women who were over 18 at the time of the reform in their regions of residence, while Panel
B shows those who were exposed to the reform because they were under 18 at the time (the
“intent-to-treat” group in the specifications in this paper). Age of marriage was determined based on
lP
the month and year of marriage for the majority of women; in the event that both could not be
recalled by respondents, age of marriage was imputed using the month of marriage and approximate
age, the year of marriage, or using age only (the bottom graphs). Source: DHS documentation and
author’s calculations from the DHS data combined with reform dates from legal documents.
na
ur
B. Timing of the Reform by Region

Table B1 shows the dates of the reform by region according to the legal documentation. Figure B1 shows a timeline
Jo
of the reform dates and the DHS survey round dates. Table B2 reports coefficients from a series of regressions in which
the number of months that elapsed between the federal reform (July 2000) and a region’s adoption of the reform is
regressed on woreda level rates of early marriage and motherhood and educational attainment, weighted by population.
In all of these regressions, the coefficients are statistically insignificant and quite small in magnitude, suggesting that
pre-trends are not a concern. Appendix Figure A2 also shows the distribution of marriage ages in the various regions
left to right in order of timing of the reform, showing that reform timing and pre-existing distributions in the age of
marriage are not systematically related.

Figure A2: Distribution of Marriage Ages by Region
Addis Ababa Dire Dawa Oromia Amhara
30
30
30
30
20
20
20
20
Percent
Percent
Percent
Percent
10
10
10
10
0
0
5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35
Age at Marriage Age at Marriage Age at Marriage Age at Marriage
SNNPR BG Tigray Harari

30
30
30
30
20
20
20
20
Percent
Percent
Percent
Percent
10
10
10
10
0
0
5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35
Age at Marriage Age at Marriage Age at Marriage Age at Marriage
Gambela Afar Somali

30
30
30
f
oo
20
20
20
Percent
Percent
Percent
10
10
10
0
pr
5 10 15 20 25 30 35 5 10 15 20 25 30 35 5 10 15 20 25 30 35
Age at Marriage Age at Marriage Age at Marriage
- Percent
re
lP
Figure A3: Distribution of Marriage Ages by Survey Year
2000 2005
na
20
20
10 15
10 15
ur
Percent
Percent
5
5
Jo
0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
2011 2016
20
20
10 15
10 15
Percent
Percent
5
5
0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Percent
C. Robustness of Main Results

Supplementary data associated with this article can be found in the online version at doi:XXX
CREDITS: XXX I am grateful to Francine Blau, Kaushik Basu, James Berry, and Victoria Prowse for their en-

Figure A4: Variation in pre-reform average age at marriage by district (district)
25
20 15
Percent
10
f
oo
5
pr
0
10 15 18 20 25 30
-
Pre-reform Mean Age at Marriage by District
re
All
.3
lP
na
.2
ur
Jo .1
0
10 15 20 25 30
Pre-reform Mean Age at Marriage by District
Rural Urban
By rural/urban
couragement and guidance. Thanks to Kristin Butcher, Esther Duflo, Pascaline Dupas, Andrew Green, Philip Levine,
Ellen McCullough, Eric Ohrn, Andrew Simons, Alemayehu Seyoum Taffesse, Kibrom Tafere, Mallika Thomas, two
anonymous referees, and numerous seminar participants including those at the Society of Labor Economics (SOLE)
2016 meetings, Cornell University, Wellesley College, Swarthmore College, and Grinnell College. Fitsum Getahun,
Natnael Shibru, Yi Qin, Devansh Chandgothia, and Ishaan Tibrewal provided excellent research assistance. I grate-
fully acknowledge support from Grinnell College. A version of this paper previously circulated under a different title

Figure A5: Heterogeneous Baseline Outcomes by Ethnic Group

Panel A: % of Women Married Under 16
.6 .4
% of women
.2
f
oo
pr
0
Affar Amhara Oromo Somalie Tigrie

-
re
2.5
lP
na
2
Years of Education
ur 1.5
Jo
1 .5
0
Affar Amhara Oromo Somalie Tigrie

Panel B: Years of Education
“Child Brides, Bargaining Power, and Reform of Ethiopia’s Family Law” (McGavock, 2016). All errors are my own.
References
Abebe, S.G., 2014. The Last Post-Cold War Socialist Federation: Ethnicity, Ideology, and Democracy in Ethiopia. Ashgate Publishing Limited,
Surrey.
Adam, C., Hoddinott, J., Ligon, E., 2003. Dynamic Intrahousehold Bargaining, Matrimonial Property Law, and Suicide in Canada.
Alesina, A., Brioschi, B., La Ferrara, E., 2016. Violence Against Women: A Cross-cultural Analysis for Africa.
Alesina, A., Giuliano, P., Nunn, N., 2011. Fertility and the plough. American Economic Review 101, 499–503. doi:10.1257/aer.101.3.499.
Alesina, A., Giuliano, P., Nunn, N., 2013. On the Origins of Gender Roles: Women and the Plough. The Quartely Journal of Economics 128,

Figure B1: Timing of Reform and DHS Data Collection
The figure shows a timeline of the reform dates and DHS survey round dates. Here, as elsewhere, I
use years from the international (Gregorian) calendar. Dates appear in the DHS data and official
legal documentation according to the Ethiopian Orthodox calendar, which has 13 months. For the
purpose of this paper, I follow standard convention and transform Ethiopian calendar dates to
Gregorian dates by adding 92 months unless a transformation is already explicitly documented, as on
the legal documents.
f
oo
Table B1
Dates of Reform by Region
pr
Region Date of Reform
Addis Ababa - 7/4/2000
Afar Not yet revised
re
Amhara 6/25/2003
Benishangul-Gumaz (BG) 5/1/2006
lP
Dire Dawa 7/4/2000

Gambela 1/1/2008
Harari 8/1/2008
na
Oromia 7/10/2002
Southern Nations, Nationalities, and Peoples (SNNPR) 10/23/2004
Somali Not yet revised
ur
Tigray 2/20/2007
Jo
Notes: The table shows the reform dates in the regions according to legal documentation.
469–530. doi:10.1093/qje/qjt005.Advance.
Arthur, M., Earle, A., Raub, A., Vincent, I., Atabay, E., Latz, I., Kranz, G., Nandi, A., Heymann, J., 2018. Child Marriage Laws around the
World: Minimum Marriage Age, Legal Exceptions, and Gender Disparities. Journal of Women, Politics and Policy 39, 51–74. URL: https:
//doi.org/10.1080/1554477X.2017.1375786, doi:10.1080/1554477X.2017.1375786.
Bailey, M.J., 2006. More Power to the Pill: The Impact of Contraceptive Freedom on Women’s Life Cycle Labor Supply. The Quarterly Journal of
Economics 121, 289–320.
Baird, S., Chirwa, E., McIntosh, C., Özler, B., 2010. The short-term impacts of a schooling conditional cash transfer program on the sexual behavior
of young women. Health Economics 19, 55–68. doi:10.1002/hec.1569.
Baird, S., Friedman, J., Schady, N., 2011. Aggregate Income Shocks and Infant Mortality in the Developing World. Review of Economics and
Statistics 93, 847–856. doi:10.1162/REST_a_00084.
Bandiera, O., Buehren, N., Burgess, R., Goldstein, M., Gulesci, S., Rasul, I., Sulaiman, M., 2012. Empowering Adolescent Girls: Evidence from a
Randomized Control Trial in Uganda. Technical Report.
Bandiera, O., Burgess, R., Gulesci, S., Rasul, I., Sulaiman, M., 2020. Women’s Empowerment in Action: Evidence from a Random-
ized Control Trial in Africa. American Economic Journal: Applied Economics2 12, 210–259. URL: http://www.economics.
ox.ac.uk/materials/papers/13805/csae-wps-2014-30.pdf{%}0Ahttps://stanford.idm.oclc.org/login?url=https:
//search.ebscohost.com/login.aspx?direct=true{&}db=eoh{&}AN=1496027{&}site=ehost-live{&}scope=site.
Banerjee, A., Banerji, R., Berry, J., Duflo, E., Kannan, H., Mukerji, S., Shotland, M., Walton, M., 2017. From proof of concept to scalable policies:
Challenges and solutions, with an application. Journal of Economic Perspectives 31, 73–102. doi:10.1257/jep.31.4.73.
Benjamini, Y., Krieger, A.M., Yekutieli, D., 2006. Adaptive Linear Step-up Procedures That Control the False Discovery Rate. Biometrika 93,
491–50.
Blank, R.M., Charles, K.K., Sallee, J.M., 2009. A cautionary tale about the use of administrative data: Evidence from age of marriage laws.
American Economic Journal: Applied Economics 1. doi:10.1257/app.1.2.128.

Table B2
Exogeneity of Timing of Reform
𝛽 Standard Error t-Stat
Married by 15 -.09 .45 -.19

Married by 18 .05 .5 .09
Mother by 15 -.55 .65 -.84
Mother by 18 -.27 .45 -.59
Mother by 21 -.51 .48 -1.07
Years of Education .1 .09 1.11
Notes: Coefficients 𝛽 in this table are estimated from a series of regressions where the dependent
variable takes values between 27 and 97 for the number of months that elapse between July 2000 –
the date of the federal reform – and the date of reform in each region. Regressors are woreda
(district) aggregates of the probability of early marriage and motherhood and the years of education
attained by women 18 to 33 in 2000. There are 242 woreda observations in the regressions. The
coefficients test for a systematic relationship between rates of early marriage or motherhood
(outcomes used in the analysis that follows) and the timing of the reform. The coefficients are small
and statistically insignificant and therefore lend credibility to the exogeneity of the reform timing
f
relative to trends in marriage ages and motherhood. Sources: DHS data and author’s calculations.
oo
Table C1
Effect of Excluding Controls from Main Result in Table 2 Column 2
pr
Panel A: Combined Effect
None Ethnicity FE
- Rural FE Age at Survey Survey FE
re
Post Reform -0.0348∗∗∗ -0.0345∗∗∗ -0.0350∗∗∗ -0.0644∗∗∗ -0.0299∗∗
(0.0108) (0.0110) (0.0106) (0.0160) (0.0152)
lP
Observations 41852 43804 41852 41852 41852

CGM_region 0.00400 0.00601 0.00701 0.0280 0.219
baseline 0.417 0.406 0.417 0.417 0.417
na
Panel B: Separate Effects by District Pre-reform Compliance

None Ethnicity FE Rural FE Age at Survey
Post Reform -0.00209 -0.00276 -0.00243 -0.0355∗
ur
(0.0116) (0.0116) (0.0114) (0.0182)

Post*Noncompliant -0.0681∗∗∗ -0.0658∗∗∗ -0.0677∗∗∗ -0.0601∗∗∗
Jo
(0.0125) (0.0124) (0.0129) (0.0192)

Observations 41852 43804 41852 41852
CGM_region 0.812 0.776 0.773 0.184
CGM_region_int 0 0 0 0.00501
baseline 0.417 0.406 0.417 0.417
Notes: This table shows estimates of the treatment coefficients on “Post Reform” (and, in Panel B,
“Post*Noncompliant”) as described in Equation 1 and displayed in Table 2 Panels A and B, Column
2. The first column “None” displays the main result (exactly the result in Column 2 of Table 2),
which includes ethnicity fixed effects, a rural dummy, age at survey, and survey year fixed effects.
Each subsequent column in this table shows the main result after dropping a single control. Robust
standard errors in parentheses are clustered at the woreda (district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05,
*** 𝑝 < 0.01.
Bleakley, H., 2010. Malaria eradication in the americas: A retrospective analysis of childhood exposure. American Economic Journal: Applied
Economics 2, 1–45. doi:10.1257/app.2.2.1.
Borusyak, K., Jaravel, X., 2016. Revisiting Event Study Designs. Ssrn , 1–25doi:10.2139/ssrn.2826228.
Boyden, J., Pankhurst, A., Tafere, Y., 2013. Harmful Traditional Practices and Child Protection.
Buchmann, N., Field, E., Glennerster, R., Nazneen, S., Pimkina, S., Sen, I., 2018. Power vs Money: Alternative Approaches to Reducing Child Mar-
riage in Bangladesh, a Randomized Control Trial. URL: https://www.povertyactionlab.org/sites/default/files/publications/
100{_}child-marriage-bangladesh-May2017.pdf.
Burgess, G.L., 2012. When the personal becomes political: using legal reform to combat violence against women in Ethiopia. Gender, Place &
Culture 19, 153–174. doi:10.1080/0966369X.2011.573142.
Bursztyn, L., González, A.L., Yanagizawa-Drott, D., 2020. Misperceived Social Norms: Women Working Outside the Home in Saudi Arabia.

Table C2
Testing Robustness to a Different Definition of “Noncompliant”
Panel A: Using pre-2000 averages to determine “noncompliant” districts (for the
specification in Table 2 Panel B
Post Reform 0.00163 -0.00232 -0.0146 0.0170
(0.0212) (0.0108) (0.0129) (0.0118)
marr_post_type_p2000_12 0.0548∗∗ -0.0596∗∗∗ 0.0497∗∗∗ 0.00995
(0.0213) (0.0136) (0.0151) (0.0116)
Observations 41852 41852 41852 41852
CGM_region 0.922 0.818 0.293 0.169
CGM_region_int 0.0460 0 0.0571 0.00200
baseline 16.92 0.417 0.211 0.372
Panel B: For reference (copy results from Table 2 Panel B)

f
Post Reform 0.0125 -0.00209 -0.0230∗ 0.0251∗∗
oo
(0.0198) (0.0116) (0.0136) (0.0115)
Post*Noncompliant 0.0396∗ -0.0681∗∗ 0.0737∗∗ -0.00565
(0.0210) (0.0125) (0.0146) (0.0116)
pr
Observations 41852 41852 41852 41852
p-value CGM 0.623 0.812 0.0931 0.0611
p-value CGM (interaction) 0.0621
- 0 0.00701 0.0551
re
Baseline mean 16.92 0.417 0.211 0.372
Notes: This table shows estimates of the treatment coefficients as described in Equation 1 and
displayed in Table 2 B. Panel A of this table uses pre-2000 average marriage ages to identify
lP
“noncompliant” districts. In the main specification reported in Table 2 Panel B, which is repeated
below in Panel B of this table, “noncompliant” districts are identified using pre-reform average
marriage ages, where the dates of the reform are region-specific. Robust standard errors in
parentheses are clustered at the woreda (district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01.
na
American Economic Review 110, 2997–3029. doi:10.1257/aer.20180975.

ur
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2008. Bootstrap-Based Improvements for Inference with Clustered Errors. Review of Economics and
Statistics 90, 414–427. doi:10.1162/rest.90.3.414.
Carranza, E., 2014. Soil endowments, female labor force participation, and the demographic deficit of women in India. American Economic Journal:
Jo
Applied Economics 6, 197–225. doi:10.1257/app.6.4.197.

Castilla, C., 2018. Political role models and child marriage in India. Review of Development Economics 22, 1409–1431. doi:10.1111/rode.
12513.
Chari, A.V., Heath, R., Maertens, A., Fatima, F., 2017. The causal effect of maternal age at marriage on child wellbeing: Evidence from India.
Journal of Development Economics 127, 42–55. doi:10.1016/j.jdeveco.2017.02.002.
Chiappori, P.A., Fortin, B., Lacroix, G., 2002. Marriage Market, Divorce Legislation, and Household Labor Supply. Journal of Political Economy
110, 37–72. doi:10.1086/324385.
Chicoine, L., 2019. Schooling with learning: The effect of free primary education and mother tongue instruction reforms in Ethiopia. Economics
of Education Review 69, 94–107. URL: https://doi.org/10.1016/j.econedurev.2019.01.002, doi:10.1016/j.econedurev.2019.
01.002.
Chicoine, L., 2020. Free Primary Education, Fertility, and Women’s Access to the Labor Market : Evidence from Ethiopia. World Bank Economic
Review .
Chow, V., Vivalt, E., 2019. Challenges in Changing Social Norms: Evidence from Interventions Targeting Child Marriage in Ethiopia .
Convention on the Elimination of All Forms of Discrimination against Women (CEDAW), 2010. Combined sixth and seventh periodic reports of
States parties: Ethiopia .
Corno, L., Hildebrandt, N., Voena, A., 2020. Age of Marriage, Weather Shocks, and the Direction of Marriage Payments. Econometrica 88,
879–915. doi:10.3982/ECTA15505.
Deaton, A., 2010. Learning about Development. Journal of Economic Literature 48, 424–455. doi:10.1257/jel.48.2.424.
Dhar, D., Jain, T., Jayachandran, S., 2018. Reshaping Adolescents’ Gender Attitudes: Evidence from a School-Based Experiment in India. SSRN
Electronic Journal doi:10.2139/ssrn.3301068.
Duflo, E., 2012. Women Empowerment and Economic Development. Journal of Economic Literature 50, 1051–1079.
Duflo, E., Dupas, P., Kremer, M., 2015. Education, HIV, and early fertility: Experimental evidence from Kenya. American Economic Review 105,
2757–2797. doi:10.1257/aer.20121607.
Duflo, E., Dupas, P., Kremer, M., 2019. The Impact of Free Secondary Education.

Table C3
Results of Table 2 without Detrending Outcomes
Panel A: Combined effect, without detrending outcomes
∗∗∗ ∗∗∗
Post Reform 0.0532 -0.0338 0.0125 0.0213∗∗
(0.0121) (0.0108) (0.0106) (0.00975)
Observations 41852 41852 41852 41852
CGM_region 0.00400 0.0120 0.312 0.0410
baseline 16.92 0.417 0.211 0.372
Panel B: Interaction Effects, without detrending outcomes
∗∗∗
Post Reform 0.0717 -0.000206 -0.0206 0.0208∗
(0.0158) (0.0115) (0.0134) (0.0112)
Post*Noncompliant -0.0383∗∗ -0.0698∗∗∗ 0.0687∗∗∗ 0.00105
f
(0.0169) (0.0130) (0.0144) (0.0116)
oo
Observations 41852 41852 41852 41852
CGM_region 0.00400 0.976 0.122 0.106
pr
CGM_region_int 0.135 0 0.00601 0.0480
baseline 16.92 0.417 0.211 0.372
-
re
Notes: This table shows estimates of the treatment coefficient on “Post Reform” as described in
Equation 1 and displayed in Table 2 Column 2. The first column “None” displays the main result
(exactly the result in Column 2 of Table 2), which includes ethnicity fixed effects, a rural dummy, age
lP
at survey, and survey year fixed effects. Each subsequent column in this table shows the main result
after dropping a single control. Robust standard errors in parentheses are clustered at the woreda
(district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01. p-values from the Wild cluster bootstrap
na
procedure of Cameron et al. (2008), which may further correct for serial correlation, appear in the
second to last row. The final row shows p-values from clustering at the region level, correcting for
the small number of clusters. The estimation sample includes all married women ages 15-49 in the
ur
DHS samples.
Jo
Duflo, E., Dupas, P., Kremer, M., Sinei, S., 2006. Education and HIV/AIDS Prevention: Evidence from a randomized evaluation in Western Kenya.
Technical Report. URL: http://econ.worldbank.org.
Edmeades, J., 2013. Child Brides - We Cannot Afford to Ignore Their Plight.
Erulkar, A.S., Muthengi, E., 2009. Evaluation of Berhane Hewan: A Program to Delay Child Marriage in Rural Ethiopia | Guttmacher Insti-
tute. International Perspectives on Sexual Reproductive Health URL: https://www.guttmacher.org/about/journals/ipsrh/2009/
03/evaluation-berhane-hewan-program-delay-child-marriage-rural-ethiopia.
Fafchamps, M., Quisumbing, A., 2002. Control and Ownership of Assets Within Rural Ethiopian Households. Journal of Development Studies 38,
47–82. doi:10.1080/00220380412331322581.
Fafchamps, M., Quisumbing, A., 2005. Assets at marriage in rural Ethiopia. Journal of Development Economics 77, 1–25. doi:10.1016/j.
jdeveco.2004.02.003.
Federal Negarit Gazetta of the Federal, of Ethiopia, D.R., 2000. Federal Negarit Gazetta Extra Ordinary Issue No. 1/2000: The Revised Family
Code Proclamation No. 213/2000.
Field, E., Ambrus, A., 2008. Early Marriage, Age of Menarche, and Female Schooling Attainment in Bangladesh. Journal of Political Economy
116, 881–930. doi:10.1086/593333.
Freyaldenhoven, S., Hansen, C., Shapiro, J.M., 2019. Pre-event trends in the panel event-study design. American Economic Review 109, 3307–3338.
doi:10.1257/aer.20180609.
Gershoni, N., Low, C., 2017. The Impact of Extended Reproductive Time Horizons: Evidence from Israel’s Expansion of Access to IVF.
Giorgis, G.W., 2005. Proclamation No.414/2004: The Criminal Code of the Federal Democratic Republic of Ethiopia.
Glennerster, R., Walsh, C., Diaz-Martin, L., . A practical guide to measuring women’s and girls’ empowerment in impact evaluations. Technical
Report. J-PAL.
Goodman-bacon, A., 2019. Difference-in-Differences with Variation in Treatment Timing.
Gray, J.S., 1998. Divorce-Law Changes, Household Bargaining, and Married Women’s Labor Supply. American Economic Review 88, 628–642.
doi:http://www.jstor.org/stable/116853.
Hallward-Driemeier, M., Gajigo, O., 2011. Strengthening Economic Rights and Women’s Occupational Choice The Impact of Reforming Ethiopia’s

Table C4
Results of Table 2 using Freyaldenhoven et al. (2019) procedure to control for ever-schooling rates
Panel A: Combined effect, using Freyaldenhoven et al. (2019) procedure to control
for ever-schooling rates
∗∗∗ ∗∗∗
Post Reform 0.0535 -0.0369 0.0139 0.0230∗∗
(0.0155) (0.0133) (0.0119) (0.0109)
Observations 31366 31366 31366 31366
baseline 16.96 0.416 0.204 0.380
Panel B: Interaction Effects, using Freyaldenhoven et al. (2019) procedure to control
for ever-schooling rates
∗
Post Reform 0.0358 -0.0122 -0.0170 0.0291∗∗
(0.0187) (0.0152) (0.0157) (0.0129)
f
Post*Noncompliant 0.0380∗ -0.0532∗∗∗ 0.0663∗∗∗ -0.0131
oo
(0.0220) (0.0116) (0.0154) (0.0114)
Observations 31366 31366 31366 31366
pr
baseline 16.96 0.416 0.204 0.380
Notes:Notes: This table shows estimates of the treatment coefficients on “Post Reform” (and, in
-
Panel B, “Post*Noncompliant”) as described in Equation 1 and displayed in Table 2 Panels A and B,
re
Column 2. In Table 2, outcomes are de-trended first using region and year of birth fixed effects and
region-specific time trends. Robust standard errors in parentheses are clustered at the woreda
lP
(district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01.
Table C5
na
Results of Table 2 with District level fixed effects

ur
∗ ∗∗∗
Post Reform 0.0291 -0.0338 0.0122 0.0216∗∗
(0.0164) (0.0107) (0.0105) (0.00963)
Jo
Observations 41852 41852 41852 41852

CGM_region 0.266 0.00300 0.355 0.0200
baseline 16.92 0.417 0.211 0.372
2. This table includes district-level fixed effects as an additional robustness consideration. Robust
standard errors in parentheses are clustered at the woreda (district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05,
*** 𝑝 < 0.01.
Family Law.
Hallward-Driemeier, M., Pritchett, L., 2015. How Business is Done in the Developing World: Deals versus Rules. Journal of Economic Perspectives
29, 121–140. URL: http://pubs.aeaweb.org/doi/10.1257/jep.29.3.121, doi:10.1257/jep.29.3.121.
Harari, M., 2019. Women’s inheritance rights and bargaining power: Evidence from Kenya. Economic Development and Cultural Change 68,
189–238. doi:10.1086/700630.
Jacobson, R., 1992. The Committee on the Elimination of Discrimination Against Women. The United Nations and human rights: a critical appraisal
38, 444–472.
Jensen, R., Thornton, R., 2003. Early female marriage in the developing world. Gender & Development 11, 9–19. doi:10.1080/741954311.
Kefale, A., 2013. Federalism and Ethnic Conflict in Ethiopia. Routledge, Abingdon.
Khanna, M., 2020. The Precocious Period: The Impact of Early Menarche on Schooling in India. SSRN Electronic Journal , 1–66doi:10.2139/
ssrn.3419041.
Kim, M., Longhofer, W., Boyle, E.H., Nyseth Brehm, H., 2013. When Do Laws Matter? National Minimum-Age-of-Marriage Laws, Child Rights,
and Adolescent Fertility, 1989-2007. Law and Society Review 47, 589–619. doi:10.1111/lasr.12033.

Table C6
Results of Table 2 clustering errors at the DHS cluster level
Panel A: Combined effect
∗∗ ∗∗∗
Post Reform 0.0315 -0.0348 0.0125 0.0224∗∗
(0.0145) (0.00962) (0.0112) (0.00911)
Observations 41852 41852 41852 41852
baseline 16.92 0.417 0.211 0.372
Panel B: Interaction Effects
∗
Post Reform 0.0125 -0.00209 -0.0230 0.0251∗∗
(0.0166) (0.0108) (0.0127) (0.0104)
Post*Noncompliant 0.0396∗∗ -0.0681∗∗∗ 0.0737∗∗∗ -0.00565
(0.0174) (0.0109) (0.0131) (0.0109)
f
oo
Observations 41852 41852 41852 41852
baseline 16.92 0.417 0.211 0.372
pr
Notes:Notes: This table shows estimates of the treatment coefficients on “Post Reform” (and, in
Panel B, “Post*Noncompliant”) as described in Equation 1 and displayed in Table 2 Panels A and B,
Column 2. In Table 2, outcomes are de-trended first using region and year of birth fixed effects and
-
region-specific time trends. Robust standard errors in parentheses are clustered at the woreda
re
(district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01.
lP
Kumar, N., Quisumbing, A.R., 2015. Policy Reform toward Gender Equality in Ethiopia: Little by Little the Egg Begins to Walk. World Development
67, 406–423. doi:10.1016/j.worlddev.2014.10.029.
na
Lee-Rife, S., Malhotra, A., Warner, A., Glinski, A.M., 2012. What Works to Prevent Child Marriage: A Review of the Evidence. Studies in Family
Planning 43, 287–303. doi:10.1111/j.1728-4465.2012.00327.x.
Lucas, A.M., Mbiti, I.M., 2012. Access, sorting, and achievement: The short-run effects of free primary education in Kenya. American Economic
Journal: Applied Economics 4, 226–253. doi:10.1257/app.4.4.226.
ur
Madajewicz, M., Pfaff, A., van Geen, A., Graziano, J., Hussein, I., Momotaj, H., Sylvi, R., Ahsan, H., 2007. Can information alone change behavior?
Response to arsenic contamination of groundwater in Bangladesh. Journal of Development Economics 84, 731–754. doi:10.1016/j.jdeveco.
Jo
2006.12.002.
Malhotra, A., Warner, A., McGonagle, A., Lee-Rife, S., 2011. Solutions to End Child Marriage. Technical Report. International Center for Research
on Women.
Maswikwa, B., Richter, L., Kaufman, J., Nandi, A., 2015. Minimum marriage age laws and the prevalence of child marriage and adolescent birth:
Evidence from sub-Saharan Africa. International Perspectives on Sexual and Reproductive Health 41, 58–68. doi:10.1363/4105815.
Mathur, S., Greene, M., Malhotra, A., 2003. Too Young to Wed: The Lives, Rights, and Health of Young Mar-
ried Girls. International Center for Research on Women (ICRW) URL: http://www.icrw.org/files/publications/
Too-Young-to-Wed-the-Lives-Rights-and-Health-of-Young-Married-Girls.pdf.
Mekonnen, B.M., Aspen, H., 2009. Early Marriage and the Campaign against it in Ethiopia, in: Ege, S., Aspen, H., Teferra, B., Bekele, S. (Eds.),
Proceedings of the 16th International Conference of Ethiopian Studies.
Muralidharan, K., Prakash, N., 2017. Cycling to school: Increasing secondary school enrollment for girls in India. American Economic Journal:
Applied Economics 9, 321–350. doi:10.1257/app.20160004, arXiv:arXiv:1011.1669v3.
North, D.C., 1990. Institutions, Institutional Change and Economic Performance. Cambridge University Press, Cambridge. doi:10.2307/2234910.
OpenAfrica, 2013. Ethiopia Shapefiles. URL: https://africaopendata.org/dataset/ethiopia-shapefiles.
Pankhurst, A., 2015. Keeping child marriages trending down. URL: http://newbusinessethiopia.com/index.php/resources/
18-nbe-blog/470/keeping-child-marriages-trending-down.
Pankhurst, H., 1992. Gender, Development, and Identity: An Ethiopian Study. Zed Books, London.
Pausewang, S., Tronvoll, K., Aalen, L. (Eds.), 2002. Ethiopia Since the Derg: A Decade of Democratic Pretension and Performance. Zed Books,
London.
Raj, A., Saggurti, N., Balaiah, D., Silverman, J.G., 2009. Prevalence of child marriage and its effect on fertility and fertility-control outcomes
of young women in India: a cross-sectional, observational study. The Lancet 373, 1883–1889. URL: http://dx.doi.org/10.1016/
S0140-6736(09)60246-4, doi:10.1016/S0140-6736(09)60246-4.
Rotz, D., 2016. Why have divorce rates fallen? The role of women’s age at marriage. Journal of Human Resources 51, 961–1002. doi:10.3368/
jhr.51.4.0214-6224R.
Sekhri, S., Debnath, S., 2014. Intergenerational Consequences of Early Age Marriages of Girls: Effect on Children’s Human Capital. The Journal of

Table C7
Robustness to Other Programs
Panel A: PSNP
Early PSNP District -0.00839 0.00343 0.0119 -0.0153∗∗∗
(0.0109) (0.00651) (0.00730) (0.00551)
Post Reform 0.0125 -0.00212 -0.0231∗ 0.0252∗∗
(0.0198) (0.0115) (0.0136) (0.0114)
Post*Noncompliant 0.0395∗ -0.0680∗∗∗ 0.0739∗∗∗ -0.00594
(0.0210) (0.0124) (0.0147) (0.0115)
Observations 41852 41852 41852 41852
baseline 16.92 0.417 0.211 0.372
Panel B: Family Planning

Family Planning Worker Visit -0.0114 -0.0103∗∗ 0.0116∗ -0.00135
f
(0.00750) (0.00482) (0.00606) (0.00422)
oo
Heard Family Planning on Radio -0.00302 -0.00158 -0.00508 0.00666
(0.00731) (0.00390) (0.00501) (0.00406)
pr
Post Reform 0.0120 -0.00227 -0.0229∗ 0.0252∗∗
(0.0199) (0.0116) (0.0135) (0.0114)
Post*Noncompliant 0.0399∗
- -0.0680∗∗∗ 0.0738∗∗∗ -0.00577
re
(0.0210) (0.0125) (0.0146) (0.0116)
Observations 41834 41834 41834 41834
baseline 16.92 0.417 0.211 0.372
lP
Panel C: Land or house non-owners

Owns Land Owns House
na
Post Reform -0.0144 -0.0380∗∗

(0.0157) (0.0172)
Post*Noncompliant -0.00357 0.00742
ur
(0.0192) (0.0196)
Observations 22488 22485
Jo
baseline 0.561 0.734

2, but including controls or using outcomes as indicated in Section 4.4. Robust standard errors in
parentheses are clustered at the woreda (district) level, with * 𝑝 < 0.10, ** 𝑝 < 0.05, *** 𝑝 < 0.01.
Development Studies 50, 1670–1686. URL: http://www.tandfonline.com/doi/abs/10.1080/00220388.2014.936397, doi:10.1080/

00220388.2014.936397.
Smith, L., 2013. Making Citizens in Africa: Ethnicity, Gender, and National Identity in Ethiopia. Cambridge University Press.
Stopnitzky, Y., 2017. No toilet no bride? Intrahousehold bargaining in male-skewed marriage markets in India. Journal of Development Economics
127, 269–282. URL: http://dx.doi.org/10.1016/j.jdeveco.2017.04.003, doi:10.1016/j.jdeveco.2017.04.003.
UNICEF, 2013. The State of the World’s Children. Technical Report. doi:10.1016/S0882-5963(05)80079-X.
Voigt, S., Kiwit, D., 1998. The Role and Evolution of Beliefs, Habits, Moral Norms, and Institutions, in: Merits and Limits of Markets. Springer
Berline Heidelberg, Berlin. doi:10.1007/s13398-014-0173-7.2.
Warner, A., Stoebenau, K., Glinski, A.M., 2014. More Power to Her. Technical Report. International Center for Research on Women (ICRW).
Wodon, Q., Male, C., Nayihouba, A., Savadogo, A., Yedan, A., Edmeades, J., Kes, A., John, N., Murithi, L., Steinhaus, M., Petroni, S., 2017.
Economic Impacts of Child Marriage: Global Synthesis Report. Technical Report June. The World Bank and International Center for Research
on Women. Washington, D.C.
Wolfers, J., 2006. Did unilateral divorce laws raise divorce rates? A reconciliation and new results. American Economic Review 96, 1802–1820.
doi:10.1257/aer.96.5.1802.

“Here Waits the Bride”
Highlights:
• This study leverages the natural experiment arising from the staggered
roll out of the policy in a difference-in-differences and event study framework.
• The results suggest that the reform delayed women’s marriage, and in particular delayed
marriages of girls under 16 by about 17 percent (6.8 percentage points) in areas where
early marriage was more common prior to the reform.
• The effect of the reform is larger but insignificant among women belonging to ethnic
groups with the strongest norms toward early marriage.
• Women’s fertility was delayed and may be lower over their lifetimes.
of
ro
-p
re
lP
na
ur
Jo

Ethiopia

Uploaded by

Copyright:

Available Formats

Ethiopia

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ethiopia

Uploaded by

Copyright:

Available Formats

Journal Pre-proof

To appear in: Journal of Development Economics

Received Date: 5 December 2018

© 2020 Published by Elsevier B.V.

ARTICLE INFO ABSTRACT

T. McGavock: Preprint submitted to Elsevier Page 1 of 35

T. McGavock: Preprint submitted to Elsevier Page 2 of 35

2. Marriage and Ethiopia’s Family Law Reform

schooled population of girls.

T. McGavock: Preprint submitted to Elsevier Page 3 of 35

edge of it (Pankhurst, 2015).

uses the appropriate sampling weights provided in the data.

T. McGavock: Preprint submitted to Elsevier Page 4 of 35

Figure 1: Average Age of Marriage Pre- and Post-reform

3.2. Empirical Strategy

T. McGavock: Preprint submitted to Elsevier Page 5 of 35

Pre-Reform Post-Reform P-value for

Age Marr. 17.55 19.1 0

Location and Ethnicity

Amhara .28 .28 .81

the “treatment” group as women whose marriages occurred post-reform.14

Age at first marriage𝑖𝑤𝑟𝑡𝜏 =

T. McGavock: Preprint submitted to Elsevier Page 6 of 35

Identifying Assumption 1: Exogenous Timing of Reform by Region.

T. McGavock: Preprint submitted to Elsevier Page 7 of 35

Age at first marriage𝑖𝑤𝑟𝑡 =

4. Effect of the reform on early marriages

T. McGavock: Preprint submitted to Elsevier Page 8 of 35

Panel B: Separate Effects by District Pre-reform Compliance

T. McGavock: Preprint submitted to Elsevier Page 9 of 35

across space in precisely where in the distribution these changes occurred.24

4.3. Heterogeneity in Effects of the Reform by Ethnic Group

statistically significantly different from zero.

preferred set of results.

T. McGavock: Preprint submitted to Elsevier Page 10 of 35

Panel B: Women from ethnic groups not using the plow

the woman resides in a “Noncompliant” district (woreda), where noncompliance is determined by

4.4. Robustness to impacts of other interventions

T. McGavock: Preprint submitted to Elsevier Page 11 of 35

Any School Years (incl. 0) Employed

T. McGavock: Preprint submitted to Elsevier Page 12 of 35

5. Secondary outcomes and mechanisms

T. McGavock: Preprint submitted to Elsevier Page 13 of 35

Panel B: Health Outcomes

5.1. Results for secondary outcomes

T. McGavock: Preprint submitted to Elsevier Page 14 of 35

have a child on the day of survey.

T. McGavock: Preprint submitted to Elsevier Page 15 of 35

turn lead to improved outcomes for women’s children.

5.3. Ruling out other components of the Family Law reform

T. McGavock: Preprint submitted to Elsevier Page 16 of 35

Panel B: Perceived Empowerment Outcomes

T. McGavock: Preprint submitted to Elsevier Page 17 of 35

I address here several aspects of the data in more detail.

A.1. Age at marriage and possible measurement error

T. McGavock: Preprint submitted to Elsevier Page 18 of 35

A.2. Construction of outcome indices

A.4. Pre-reform compliance levels

motherhood, or years of educational attainment.

T. McGavock: Preprint submitted to Elsevier Page 19 of 35