Policy Research Working Paper 9105
Free Primary Education, Fertility,
and Women’s Access to the Labor Market
Evidence from Ethiopia
Luke Chicoine
Development Economics
Knowledge and Strategy Team
January 2020
Policy Research Working Paper 9105
Abstract
This article investigates the causal relationship between school fees led to an increase in schooling for Ethiopian
women’s schooling and fertility by exploiting variation women and that each additional year of schooling led to
generated by the removal of school fees in Ethiopia. The a reduction in fertility. An investigation of the underlying
increase in schooling caused by the reform is identified mechanisms linking schooling and fertility finds that the
using both geographic variation in the intensity of its decline in fertility is associated with an increase in labor
impact and temporal variation generated by the timing of market opportunity and a reduction in women’s ideal
the implementation. The model finds that the removal of number of children.
This paper is a product of the Knowledge and Strategy Team, Development Economics. It is part of a larger effort by the
World Bank to provide open access to its research and make a contribution to development policy discussions around the
world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The author may
be contacted at lchicoin@bates.edu.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
Free Primary Education, Fertility, and Women’s Access to the
Labor Market: Evidence from Ethiopia
Luke Chicoine∗
JEL classiﬁcation: O55, J13, I25, I26
Keywords: free primary education; returns to schooling; fertility; Ethiopia
∗ Department of Economics, Bates College and IZA - Institute of Labor Economics. Address: Pettengill Hall, 4 Andrews
Rd, Lewiston, ME 04240. Email: lchicoin@bates.edu.The author would like to thank Anjali Adukia, Lori Beaman, Seema
Jayachandran, Adrienne Lucas, Kazuya Masuda, and Sarah Pearlman for their extremely helpful feedback; the Department of
Social Policy at the London School of Economics and Political Science for their support, especially from Berkay Ozcan and
Jorge Garcia-Hombrados; and members of the Navarra Center for International Development and participants at the 2018
Africa Meeting of the Econometric Society, the 2018 Annual Conference of the European Society for Population Economics,
the 2018 Nordic Conference on Development Economics, the 2017 Midwest International Economic Development Conference,
the 2017 Centre for the Study of African Economies Conference, the 2017 Liberal Arts Colleges Development Conference, and
the 2017 Maine Economic Conference for helpful comments.
1 Introduction
Prominently positioned among the Millennium Development Goals, universal primary education has become
a central tenet of the international development eﬀort. As far back as the 1970s, the most readily available
policy tool for promoting enrollment has been the removal of school fees. This type of policy was implemented
in Kenya and Nigeria in the 1970s, in Zimbabwe and Tanzania in the 1980s, and in Ethiopia, Malawi, and
Uganda in the 1990s. More recently, this policy has been aggressively pursued by international development
organizations as a key tool in achieving the goal of universal primary education, as evidenced by over a dozen
additional countries removing school fees since 2000 (Kattan and Burnett, 2004; World Bank, 2009). The
accelerated proliferation of these fee-removal programs over the past few decades highlights the importance
of gaining a greater understanding of the consequences of these reforms; however, a recent systematic review
(Snilstveit et al., 2016) by the International Initiative for Impact Evaluation (3ie) concluded that little is
known about the long-term impact of reducing school fees.
This article evaluates the returns to a nationwide free primary education (FPE) program in Ethiopia.
The removal of school fees in grades one through ten is found to generate an increase in schooling. Using a
two-stage least squares (2SLS) model, each additional year of schooling generated by the reform is found to
reduce fertility by more than 0.4 births.
The article uses data from the Ethiopian census and from three rounds of the Demographic and Health
Survey (DHS). The increase in schooling is identiﬁed by combining two dimensions of variation, the timing of
the reform and geographic variation in schooling outcomes for cohorts who completed their education prior
to the introduction of the reform. Motivated by the work of Bleakley (2010), Lucas (2010, 2013), and Lucas
and Mbiti (2012a,b), the identiﬁcation relies on the concept that although the FPE policy itself is applied
uniformly across the country, the intensity of the reform in a speciﬁc location depends on the pre-existing
characteristics of that area. In this setting, as proposed in Chicoine (2019), removing school fees from an
area of high pre-reform educational attainment will have a small impact relative to removing the same fees
in an area with a low pre-reform education level.
Investigating the mechanisms through which the increase in schooling leads to a reduction in fertility for
Ethiopian women can yield an increased understanding of the household fertility decision-making process.
The increase in schooling generated by the FPE reform is found to increase literacy rates, and it leads
to women working in higher-quality jobs and wanting fewer children. However, as exposure to the reform
increases, women are no more likely to use contraception, and there is no evidence of an increase in em-
powerment. The totality of these ﬁndings suggests that the decline in the ideal number of children and the
associated increase in economic activity are the central mechanisms through which the increase in schooling
1
has reduced fertility for Ethiopian women.
Earlier literature (Ainsworth et al., 1996; Lam and Duryea, 1999; Schultz, 1994, 1997) has documented
the negative relationship between schooling and fertility that exists in the data. To identify the eﬀect
of education, Osili and Long (2008) used school construction in Nigeria, and Keats (2018) exploited a
discontinuity around the implementation of FPE in Uganda. Both articles found that education led to a
reduction in fertility of between 0.263 and 0.36 births for each additional year of schooling. Ozier (2018) also
showed that access to secondary school reduced teen pregnancy in Kenya, and Zenebe Gebre (2018) found
that an FPE reform in Malawi led to reductions in fertility through the age of 25. Although these articles
all found a similar relationship between schooling and fertility, the mechanisms vary. In addition to the
timing of marriage and evidence of increased labor market productivity, Keats (2018) also found evidence of
increased use of contraceptives. In Malawi, Zenebe Gebre (2018) documented strong evidence of an increase
in the use of contraception and a move away from agricultural employment.
Outside of Africa, the evidence of a causal relationship between schooling and fertility has been more
mixed. Fort et al. (2016) discovered evidence of a positive causal relationship between schooling and fertility
in continental Europe but a negative relationship in the United Kingdom, and Clark and Bono (2016) found
that school quality in the United Kingdom had a signiﬁcant positive impact on women’s earnings and a
negative eﬀect on fertility. Exploiting discontinuities at starting ages, McCrary and Royer (2011) found no
evidence that schooling aﬀected the probability of motherhood.
The results of this article are signiﬁcant in three ways. First, the article presents an application of a
diﬀerence-in-diﬀerences identiﬁcation strategy that can be used in a variety of settings to study national-level
reforms with a minimal amount of pre-reform information needed for identiﬁcation. Measuring the impact
of national-level removal of user fees was one of the key categories found to need further study in the 3ie
systematic review of education policy (Snilstveit et al., 2016). Second, this article ﬁnds strong evidence
of a negative relationship between schooling and fertility, and investigates the detailed pathways through
which this relationship develops. Third, the article ﬁnds signiﬁcant evidence of positive returns to schooling,
through both reduction in fertility and improvement in labor market outcomes. This result suggests that
increased schooling generated by the removal of school fees led to lasting increases in education that bettered
the day-to-day lives of Ethiopian women.
2 Background and Education Reform
After 17 years of military and communist rule, the Ethiopian People’s Revolutionary Democratic Front took
power in 1991 and quickly established a transitional government (Ofcansky and Berry, 1993). This govern-
2
ment introduced a new federal structure with nine regional governments and two independent administrative
councils in Addis Ababa and Dire Dawa. The 11 regions were established along historical ethnic lines, with
each region representing the ﬁrst administrative area level within the county, similar to a state or province.
The regions were then divided into 60 zones, as shown in ﬁg. 1.
Before the start of the 1995 school year, the new government introduced the Education and Training
Policy, which removed school fees for grades one through ten in all government-run schools. At the time
the policy was enacted, these schools educated over 90 percent of primary school students in Ethiopia, and
although there was no formal tuition fee prior to 1995, schools often imposed per-student fees to cover the
cost of operation. The reform itself had no enforcement mechanism, but most of the country had complied
with the decree by 1996 (Negash, 1996; Oumer, 2009; World Bank, 2009; UNESCO, 2007).
In addition to the removal of school fees, between 1991 and 1995 the transitional government also in-
troduced local language instruction in four of the country’s 11 regions. The introduction of local language
instruction was complicated, and the literature ﬁnds mixed evidence of the consequences of mother tongue
instruction (MTI) on schooling in Ethiopia. Although two previous articles found evidence of a positive
impact of MTI in Ethiopia (Seid [2016] and, conditional on enrollment, Ramachandran [2017]), Zenebe Ge-
bre (2014) exploited variation in the timing of the introduction of MTI in each language and found that
MTI had a negative impact on schooling. Chicoine (2019) further isolated the negative impact of MTI to
regions of Ethiopia that introduced the languages with translations using the Roman script, an alphabet
never previously used in translations of the new languages of instruction.
To focus the analysis on the consequences of an increase in schooling, the main body of this article
examines the returns to schooling generated by the removal of school fees in the seven regions of the country
that did not change the language of instruction during this period. With this restriction, the identiﬁcation
strategy exploits variation in the pre-reform levels of the remaining 32 zones of the country. Isolating the
eﬀect of the FPE program yields a more focused analysis within the main text of the article but does not
diminish the importance of considering the MTI reforms. The combined eﬀect of the two reforms, their
impact on schooling and fertility, and the potential mechanisms through which schooling impacts Ethiopian
women’s decisions are considered in detail in Appendix Section C.
3 Identiﬁcation Strategy
This section describes and expands on the method proposed by Chicoine (2019) for identifying the impact of
the FPE reform on schooling in Ethiopia. The intensity of the reform is jointly determined by both location
within Ethiopia and the timing of the reform’s implementation. Although the Education and Training Policy
3
removed school fees in grades one through ten throughout Ethiopia, the local magnitude of the reform’s
impact depends on pre-reform levels of education in each part of the country. This concept is similar to
that underlying the strategy of Bleakley (2010) and Lucas (2010, 2013), which used pre-eradication levels of
malaria to identify local variation in the impact of eradication programs; Lucas and Mbiti (2012a,b) applied
the same concept more directly to the post-2000 removal of school fees in Kenya. A similar diﬀerence-
in-diﬀerences identiﬁcation strategy can be applied to Ethiopia. Following the reform, ten years of fee-free
schooling became available to every single student; however, prior to the reform, some portion of these grades
were already being completed. In areas of the country where schooling levels were high before the reform’s
implementation, the removal of school fees would have had only a small impact relative to regions where
few students attended school in the pre-reform period. Across Ethiopia, this pre-reform level of schooling is
evaluated for each of the 60 zones in the country.
In each zone, z , the maximum potential magnitude of the reform, Mz , is calculated using information
on schooling of individuals from that zone born between 1966 and 1969.1 Only data for women are used
to calculate the measures described in this article. Women with birth dates between 1966 and 1969 were
born signiﬁcantly prior to the implementation of the reform, such that even if they entered primary school
ﬁve years late and completed ten years of education, they would not have had access to any free schooling.
In each zone, some fraction of the population, Fz,0 , never enters school; in other words, they complete zero
years of schooling. For this subset of the population, the reform has the potential to increase schooling by
ten years. An additional portion of the population, Fz,1 , dropped out after completing one year of schooling;
this subset of the population could gain as many as nine years of additional schooling, and so on. The
maximum potential magnitude of the reform in zone z is then calculated as the product of the number of
potential additional years of schooling and the fraction of the population that dropped out after each grade,
9
FPE
Mz = (10 − g )Fz,g . (1)
g =0
Equation (1) represents the number of free years of schooling made available by the reform in each zone,
beyond what was being completed prior to the reform’s implementation. For students who make the decision
to enter school following the reform’s implementation, the reform can directly increase schooling for them by
as much as Mz , relative to the pre-reform level of schooling in their zone. This maximum magnitude of the
reform applies to all individuals entering school in 1995 or later, or, if starting school on time at age seven,
cohorts born no earlier than 1987. Cohorts born in each successive year prior to 1987 beneﬁted from one
1 These data are from the 1994 Ethiopian census, collected by the Ethiopian Central Statistical Agency and made available
as part of the Integrated Public Use Microdata Series (IPUMS) International by the Minnesota Population Center and the
Ethiopian Central Statistical Agency (2017).
4
less year of free schooling; therefore, on-time entrants in the 1986 cohort beneﬁted only after completing at
least grade one. Finally, on-time entrants born in 1977 or earlier would have completed all ten grades prior
to 1995 and so gained zero years of free schooling.2 Assuming on-time entrance into school and continuous
progression, the maximum beneﬁt of the FPE reform for each cohort is as follows:
9
(10 − g ) · Fz,g if y ≥ 1987,
g =0
9
FPE
Mzy = (10 − g ) · Fz,g if 1978 ≤ y ≤ 1986, (2)
g =1987−y
0 if y ≤ 1977.
Data from the Ethiopian census provide zone-speciﬁc information on the actual starting age of children
in Ethiopia, allowing the on-time entry assumption to be relaxed. Because students often enter school at
ages other than the legal starting age of seven, an individual’s year of birth does not determine their year of
school entry but rather a possible range of years in which the individual could enter school. The calculation
of the reform’s impact can be adjusted to take into account the possibilities of starting school as early as age
six, one year early, and as late as age 12, ﬁve years late. The central assumption made regarding variation in
entry probability is that the relative age distribution within each zone is constant over time; this means that
even though all ages are more likely to enter school in the post-reform period, if a seven-year-old is twice
as likely to enter school relative to a six-year-old in the census data, then a seven-year-old remains twice as
likely to enter school in both the pre- and the post-reform states of the world.3 Following the removal of
school fees, the maximum impact of the reform would be that every student could potentially enter school.
To represent this possibility, the starting probabilities, Sz,a , are assumed to sum to 1, holding the relative
probabilities from the data constant across each age:
12
Sz,a = 1. (3)
a=6
The following set of equations (4) uses the 1985 cohort as an example to illustrate how starting ages are
included in the calculation of the impact of FPE on each cohort. First, some fraction of the cohort, Sz,6 ,
2 The timing of the impact on each cohort, assuming on-time entry, is shown in Appendix Table A.1.
3 Justiﬁcation for this assumption is shown in Appendix Figure B.1; when scaled to full entry, age speciﬁc probabilities
yield a consistent pattern (Appendix Figure B.1c). Data from the 1984, 1994, and 2007 rounds of the census are compared.
The 1984 census is the only fully pre-reform round of the census, but the administrative boundaries were changed at the
beginning of the transitional administration. The starting-age probabilities in the 1994 census were likely directly aﬀected
by the ongoing reforms. Therefore, the 2007 census, which is made up of a set of respondents whose entry decisions were
made after the post-reform equilibrium had been established, is used in the main body of the article to calculate starting
ages. Under the assumption that relative starting ages should be consistent over time, this round of the survey provides the
clearest representation of the equilibrium relative starting ages within the current administrative boundaries. Estimates using
alternative starting-age calculations from the other census rounds, with the 1984 values weighted by overlapping area into the
1994 boundaries (panel F) and the 1994 start values (panel G), can be found in Appendix Section D.2.
5
would start school one year early at age 6; this fraction of the sample will be assigned the magnitude, from
equation (2), for the previous year, 1984.
9
FPE
Age 6: Sz,6 Mz,1984 = Sz,6 (10 − g )Fz,g , (4a)
g =3
9
FPE
Age 7: Sz,7 Mz,1985 = Sz,7 (10 − g )Fz,g , (4b)
g =2
9
FPE
Age 8: Sz,8 Mz,1986 = Sz,8 (10 − g )Fz,g , (4c)
g =1
8
FPE
Summarized: Sz,a Mz,(1985+a−7) . (4d)
a=6
As the birth year assigned to the magnitude measure in (4) moves later, an extra grade of free schooling is
added to the calculation. This pattern continues through age 8, the last age at which the school entry decision
is made in the pre-FPE environment. The age-speciﬁc calculations for these three ages are summarized in a
single term in equation (4d).
For starting ages 9 to 12, the ﬁrst part of the calculation simply interacts the entry probability with the
maximum potential magnitude from equation (1). In addition to entrants at each of these ages, there exists
a stock of marginal students who would have entered school between the ages of 6 and 8 if they could have
done so for free, but faced a fee when they made their initial decision. At each age, this stock of students is
denoted by (Sz,a − Sz,a,pre ). As in equation (3), Sz,a,pre is a set of relative starting ages but scaled to equal the
12
fraction of students who entered school in the pre-reform environment, such that a=6 Sz,a,pre = (1 − Fz,0 ).
The reform’s impact for entrants of ages 9 to 12 is then written as
12 8
FPE 1
Mz Sz,a + [(10)Fz,0 ] (Sz,a − Sz,a,pre ) (5)
a=9
e9−7 a=6
where, in addition to the post-reform entrants, the stock of outstanding would-be entrants makes the decision
on whether to enter exactly one time, at the youngest possible age. These marginal students who are able
to enter in the post-reform period for the ﬁrst time at age 9 also gain 10 free years of schooling. Finally, by
6
delaying entry, it is likely that some fraction of would-be entrants are now tied to other responsibilities and
1
constrained from entering at later ages. This constraint is represented by the fraction ea−7 , where a is equal
to the age of entry being considered, in this case 9. As the post-reform age gets closer to 7, the legal age of
entry, this constraint approaches 1 and binds fewer students from delayed entry.
The full starting-age-adjusted intensity of the FPE reform for the 1985 cohort is then
8 12 8
FPE FPE FPE 1
Iz,1985 = Sz,a Mz,(1985+a−7) + Mz Sz,a + [(10)Fz,0 ] (Sz,a − Sz,a,pre ). (6)
a=6 a=9
e 9−7 a=6
Equation (6) is a combination of (4d) for pre-FPE starting ages and (5) for post-FPE starting ages. Iterating
equation (6) forward three years, when even six-year-old entrants are post-reform, demonstrates that the
1988 cohort is the ﬁrst fully post-reform cohort:
12 12
FPE FPE FPE FPE
Iz,1988 = Sz,a Mz = Mz Sz,a = Mz . (7)
a=6 a=6
Every cohort born in 1988 or later is aﬀected by the maximum potential magnitude of the reform, Mz .4
FPE
The average of the FPE intensity measure Izy in regions without any MTI introduction is shown as
the solid black line in ﬁg. 2, for each cohort. For comparison and to demonstrate the type of variation in
the intensity measure, the FPE measure is also shown for Addis Ababa, the most educated region in the
country. The height of each line can be considered the number of additional free years generated by the FPE
reform. Prior to the reform, there were more students in higher grades in Addis Ababa, leading to a greater
eﬀect size in earlier cohorts; but due to the higher level of initial schooling, the maximum magnitude of the
eﬀect in Addis Ababa is much smaller. Not only does the intensity measure predict larger increases in areas
with lower levels of initial schooling, but it also generates signiﬁcant variation in the path of the predicted
eﬀect across cohorts prior to the post-reform period.
4 Data
4.1 Data Sources
Individual-level outcome data for Ethiopian women are from the 2005, 2011, and 2016 rounds of the Ethiopian
Demographic and Health Survey (DHS) (Central Statistical Agency of Ethiopia 2005; 2011; 2016). The DHS
data used in this article are from the merged individual women and birth history datasets, and include data
4 The explicit set of equations used for all cohort-speciﬁc intensity calculations can be found in Appendix Section A.
7
from 58 of Ethiopia’s 60 zones and 30 of the 32 non-MTI zones.5 The data available for individual women
in the DHS include detailed information on birth date, district of residence, education, health, contraceptive
use, and employment. To further analyze the main outcome of the study with an alternative sample, data
from the 2007 Ethiopian census are also combined with the DHS data. The census data include information
on age, schooling, and total number of births. These data can be used to demonstrate that the conclusions
of this study are not unique to the DHS sample.
4.2 Summary Statistics
The summary statistics for the DHS data used in this article are presented in table 1. The table shows
information for women in the last three fully pre-reform cohorts, 1968 to 1970, and the ﬁrst three entirely
post-reform cohorts, 1988 to 1990. Although later cohorts are younger, they have higher levels of schooling
and literacy and far fewer births. The extremely low number of births in the post-reform sample is likely due
to women in this sample being no older than 26. For this reason, it can be informative to examine whether
the reduction in fertility is also seen at speciﬁc ages; these samples include only women older than the stated
age, and allow for a more direct comparison. The number of births to women at the ages of 20 and 25 are
also found to decline by over 60 percent between the two cohort groups. This magnitude is consistent with
the observed decline in ideal family size. Younger women are signiﬁcantly less likely to work in the unskilled
manual or agricultural sectors, and are slightly more likely to work in either of the other two employment
categories, skilled manual or professional and service or sales. Finally, probably because of the increased
literacy rate, younger cohorts are more likely to have recently read about family planning, although they are
no more likely to report knowledge of modern contraceptive methods.
5 Estimation Strategy
The central estimating model is a 2SLS model. The ﬁrst stage is deﬁned by the equation
3
p
FPE
YrsSchlizy = θ0 + θ1 Izy + θ2 Agep
izy + δz + τy + δz Trendy + νizy . (8)
p=1
FPE
The dependent variable is YrsSchlizy , the years of schooling for person i from zone z born in year y ; Izy
is the zone- and birth-year-speciﬁc estimated intensity of FPE, as described in Section 3. The ﬁrst-stage
estimate of θ1 can be interpreted as the impact of providing an additional fee-free year of school. A third-
5 DHS geocodes and administrative district data are cross-referenced with administrative boundaries using two sources:
IPUMS International (2017) and the Food and Agriculture Organization GeoNetwork’s Global Administrative Unit Layers
(GAUL) maps (2015).
8
order polynomial in age is included to take into account the fact that three waves of the DHS survey are
being used, and τy is a set of birth-year-speciﬁc ﬁxed eﬀects that capture any cohort-speciﬁc eﬀects of the
reform; δz is a vector of zone-speciﬁc ﬁxed eﬀects that capture any time-invariant characteristics of the
diﬀerent areas throughout Ethiopia, and δz Trendy is a set of zone-speciﬁc linear trends that captures secular
changes over time within each zone of Ethiopia.6
This ﬁrst-stage equation is used to estimate the exogenous increase in schooling generated by the removal
of school fees in Ethiopia. The predicted increase in schooling can then be used in the second stage to estimate
the causal relationship between schooling and births or any other outcome of interest:
3
p
Bizy = α0 + β YrsSchlizy + α2 Agep
izy + φz + µy + φz Trendy + εizy . (9)
p=1
The dependent variable Bizy is the outcome of interest, initially the number of births to person i from zone
z born in year y . The second-stage equation uses the same set of control variables as equation (8), and
the coeﬃcient on the predicted years of schooling, β , captures the causal impact of one additional year of
schooling exogenously generated by the education reform. The baseline speciﬁcation used throughout the
article includes all women born between 1970, the ﬁrst fully pre-reform cohort, and 1988, the fully post-
reform cohort. Standard errors are clustered by zone to allow for within-zone correlation (Bertrand et al.,
2004).
The ordinary least squares (OLS) relationship between schooling and fertility can be studied using a
modiﬁed version of equation (9), where the predicted level of schooling is replaced with each individual’s
actual level of schooling, YrsSchlizy . However, the OLS estimates are likely biased if schooling is correlated
with unobservable characteristics that also aﬀect the number of children women choose to have. If women
who are more likely to achieve higher levels of schooling also have higher economic ambition and lower levels of
desired fertility, the OLS estimates would be biased upward, overstating the true relationship. Alternatively,
measurement error in schooling could lead to a downward bias of the OLS estimate that may even be larger
than the ability bias that is more often discussed (Card, 2001). In fact, causal work in sub-Saharan Africa
that directly compared OLS and instrumental variables (IV) estimates found evidence that OLS estimates
signiﬁcantly underreport the relationship between schooling and fertility (Osili and Long, 2008).
The central assumption underlying this identiﬁcation strategy is that education reforms in Ethiopia,
such as the removal of school fees, impact women’s fertility decisions only through the eﬀect on their level
of schooling. This requires that contemporaneous changes in government policy and the conclusion of the
6 The set of ﬁxed eﬀects and trends is similar to what was used in the empirical strategy employed by a number of previous
studies to evaluate education reforms, including Black et al. (2005), Bleakley (2010), Lucas and Mbiti (2012a,b), Fort et al.
(2016), Holmlund et al. (2011), and Lundborg et al. (2014).
9
Ethiopian civil war not be correlated with year of birth and pre-reform levels of schooling in the same way
as the FPE reform. Potential bias generated by contemporaneous changes in educational investments, the
impact of the civil war, and the 2000 law banning marriage for those under 18 are explored in more detail
in Section 6.4. Additionally, it would be problematic if women and families relocated at the time of the
reform’s implementation in such a way that higher-ability students sorted into areas with higher predicted
intensity of the reform. However, this type of sorting is unlikely to occur in the studied setting. First, data
from the 2016 Living Standards Measurement Study show that 86 percent of respondents in the relevant
cohort range live in their region of birth. Furthermore, the intensity measure is explicitly designed to predict
a greater impact of the reform in areas with lower initial levels of schooling. A violation of this assumption
would entail the unlikely scenario that higher-ability students’ families were moving to areas that were worse
oﬀ at the time of reform implementation, even though they could have received the same reduction in fees
in their original education zone.
6 Results
6.1 Eﬀect of FPE on Years of Schooling and Fertility
To begin the analysis, the OLS relationship documents the general correlation seen in the data. This is
done by estimating equation (9) using the reported years of schooling from the data, not the predicted level
from the ﬁrst stage. A negative relationship between fertility and schooling has been well documented in
the literature (Ainsworth et al., 1996; Lam and Duryea, 1999; Schultz, 1994, 1997). The OLS estimates
are shown in column 1 of table 2. The estimates in panel A use data from both the census and the DHS,
those in panel B use census data only, and those in panel C use only DHS data. Unsurprisingly, the OLS
model estimates a strong negative relationship between schooling and fertility. However, these estimates are
unlikely to describe a causal relationship between schooling and fertility if unobserved characteristics that
impact women’s schooling also aﬀect the fertility decision. To address this concern, an exogenous increase in
schooling generated by the FPE reform in Ethiopia is identiﬁed, and an IV technique is used to investigate
the impact of this increase in education on women’s fertility.
To examine whether exposure to FPE in Ethiopia generated an identiﬁable increase in years of schooling,
the ﬁrst-stage equation, equation (8), is estimated using each combination of data. The estimates in column
2 of table 2 show that for each additional year of free schooling made available, years of schooling increased
by over one-tenth of a year. This relationship is statistically signiﬁcant at the 95 percent conﬁdence level
for all three samples, and at the 99 percent conﬁdence level when the census data are included. The ﬁrst-
10
stage F -statistic ranges from 5.93 to 37.62.7 For all three samples, the intensity measure predicts a strong
negative relationship between exposure to the FPE reform and number of children born. In column 3,
reduced-form estimates show that each additional year of free schooling made available reduces the number
of births by between 0.057 and 0.064. Estimates across all three samples yield values that are qualitatively
and quantitatively similar, providing evidence that the associations found in table 2 are not reliant on any
one source of data.
The results from the ﬁrst stage demonstrate a broad strength in the intensity measure’s ability to identify
the increase in schooling generated by FPE in Ethiopia. Estimating the second stage of the 2SLS model
focuses on the relationship between the predicted level of schooling and birth rates, as described by equation
(9). The results in column 4 of table 2 demonstrate that the exogenous increase in schooling generated
by the reform led to a reduction in fertility of 0.437 births for each additional year of schooling when
using the combined sample. The estimate is larger when using the more recent data from the DHS, but
remains similar across all three data combinations. Each estimate is statistically signiﬁcant at the 99 percent
conﬁdence level.8 Consistent with the ﬁndings in table 2, 2SLS estimates obtained by Osili and Long (2008)
for Nigeria, Keats (2018) for Uganda, and Fort et al. (2016) for the United Kingdom are signiﬁcantly larger
in magnitude than the negative OLS relationship, and Zenebe Gebre (2018) found similar evidence in Malawi
linking schooling to reductions in fertility.
6.2 Mechanisms
The results in table 2 provide evidence that additional schooling generated by the removal of school fees led
to a reduction in fertility for Ethiopian women. This subsection explores in greater detail the decisions and
changes in behavior that may be driving the relationship between schooling and fertility.
Once married, there are three broad, but not mutually exclusive, avenues through which the household
fertility decision is made. First, the increase in schooling could increase a woman’s opportunity cost of time,
impacting her desired number of children. Second, increased schooling could potentially lead to a change in
relative bargaining power over the joint fertility decision, and this is likely to lower fertility rates because
women generally desire fewer children than their husbands.9 A change in the use of contraception, especially
of forms not visible to the spouse, could be one way in which a change in the bargaining position might be
observable in the data. Finally, higher levels of schooling could lead to diﬀerent outcomes in the marriage
7 The ﬁrst stage F -statistics for the DHS-only data are less than 10; therefore, Anderson and Rubin (1949) conﬁdence sets
are given in the supplementary online appendix for all DHS-only 2SLS estimates from the main body of the article.
8 Across all three samples, the upper bound of the Anderson-Rubin weak IV robust conﬁdence sets is never more positive
than −0.265 (Appendix Table B.1).
9 More than one in three women from pre-reform cohorts report their husband wanting more children than they do, while
only nine percent report that they would like to have a larger family than their husband.
11
market, potentially aﬀecting the characteristics of a woman’s husband and his ideal family size. In addition,
it is important to consider that the schooling reform in Ethiopia may also have directly aﬀected the extensive
margin decision to marry and the timing of a woman’s ﬁrst birth.
The ﬁrst two points are directly investigated in the following subsections using data available from the
DHS. However, with only the non-MTI regions of the country and the timing of the reform relative to the data
collection, restricting the sample to examine the characteristics of only married women and their husbands
removes too many post-reform women from the latest cohorts and signiﬁcantly weakens the predictive power
of the ﬁrst-stage estimate. Therefore, the discussion regarding impact of schooling on the timing of marriage
and births and on the characteristics of husbands will be in the context of the national sample after also
taking the MTI reforms into account.10
Examining the ﬁrst two potential channels yields evidence that additional schooling leads to women being
more literate, less tolerant of domestic abuse, and increasingly likely to work in more productive sectors of
the economy. The increase in the opportunity cost of their time generated by this increase in productivity is
associated with a decline in the women’s ideal number of children. However, there is no consistent evidence
of changes in contraception use, investments in health, or control over household decisions. These ﬁndings
largely isolate economic motivations such as the increased opportunity cost of time, which is associated with
a woman’s decreased demand for children, as the central driver of the reduction in fertility.
6.2.1 Knowledge, Beliefs, and Contraception Use
To form any expectation that the increase in schooling could lead to improved labor market access or
understanding of healthcare, it is important to ﬁrst demonstrate that learning occurred for Ethiopian women
during their additional time in school. Estimates in column 1 of table 3 demonstrate that the additional
schooling generated by the reform led to a large increase in literacy, and the estimate in column 2 provides
evidence that each additional year of schooling led to an increased likelihood of 4.7 percentage points of
reading about family planning in a periodical.11 Although the increase in schooling led to an increased
likelihood of reading about family planning, general knowledge of family planning methods is widespread
and unaﬀected by the reform, as shown by column 3. Additionally, the increase in literacy and access to
information did not lead to statistically signiﬁcant changes in health, as measured by body mass index (BMI)
and, to take into consideration early-in-life investments, height.
10 Husbands are, on average, more than seven years older than their wives; therefore, unless the reform reduces the age of the
matched husband, even women born in the latest year of the sample, 1988, will have husbands who are on average not greatly
aﬀected by the removal of school fees. The median age diﬀerence ranges from six to seven years throughout the sample.
11 The literacy variable is equal to 1 if the respondent demonstrates that they are able to read a complete sentence and is equal
to 0 otherwise. These outcomes are shown for a combined sample of men and women in Chicoine (2019), and the increased
eﬀect on literacy is consistent with the ﬁndings from a combined sample that includes observations from the 2007 census and
the 2016 Living Standards and Measurement Study.
12
One of the three key channels through which schooling could impact fertility is via an increase in women’s
control over household decisions. The estimate in column 6 of table 3 suggests that the increase in schooling
may change the way women view their marriage partnership. In the DHS, women were asked about ﬁve
possible justiﬁcations for domestic violence, and each additional year of schooling decreased the number of
reasons women ﬁnd acceptable. This is largely driven by reductions in accepting the refusal of sex or burning
of food as acceptable reasons.12 With an updated view on marriage and increased access to knowledge, a
possible way for Ethiopian women to increase control over the fertility decision is through increased use of
contraception. However, the estimate in column 7 shows no evidence that additional schooling led to an
increase in the use of modern methods of contraception. The possibility that this null ﬁnding is driven by
the husband’s preferences might mean that women become more likely to conceal their contraception use.
To investigate this possibility, the indicator variable used in column 8 is only set equal to 1 if the method of
contraception used is not visible to the husband (Ashraf et al., 2014). The estimated eﬀect of schooling on
hidden contraception use is again not statistically signiﬁcant.
The results in table 3 show that while the increase in schooling improved women’s literacy and access
to healthcare information, it did not lead to increased use of available healthcare resources to exert higher
levels of control over their fertility decisions. These ﬁndings provide initial evidence that schooling did not
improve the power of Ethiopian women to make household fertility decisions. In addition, among married
women throughout Ethiopia, exposure to the reform appears not to increase their belief that they should be
able to make decisions about traveling to see family, personal healthcare, and household purchases.13 Like
the healthcare results, these ﬁndings reinforce the idea that the increase in schooling has helped women to
better understand their opportunities and their right not to fear violence within their household, but that
it has not led to improvements in their ability to control household decisions. This evidence suggests that
increased bargaining power is unlikely to play a role in post-marriage reductions in fertility.
6.2.2 Eﬀect on the Labor Market
Although the results in Section 6.2.1 provide evidence that there is no improvement in women’s relative
position within the household, the increase in schooling could still lead to women exerting increased inﬂuence
on the household fertility decision by lowering their desired number of children. If the reform is not merely
increasing schooling but also generating learning, as is suggested by the evidence in table 3, this could also
lead to improved labor market outcomes for Ethiopian women. This increase in productivity would generate
an increase in the cost of the women’s time—and an increase in their opportunity cost of raising children.
12 Coeﬃcient estimates for the ﬁve separate justiﬁcations of domestic violence can be found in Appendix Table B.4.
13 Coeﬃcient estimates for the empowerment outcomes can be found Coeﬃcient estimates for the empowerment outcomes
can be found in Appendix Table C.15.
13
An increase in opportunity cost would manifest itself in a reduced demand for children and a smaller ideal
family size. This reduction in women’s bargaining position would have the eﬀect of lowering household
fertility levels and is explored in this subsection.
Table 4 examines the impact of increased schooling on labor market outcomes in columns 1–4, and on
a woman’s ideal number of children in column 5. The estimated impact of schooling on the likelihood of
working is large but not statistically signiﬁcant at the 90 percent conﬁdence level. However, each addi-
tional year of schooling does increase the likelihood of working in a professional or skilled occupation by
5.9 percentage points; this result is statistically signiﬁcant at the 95 percent conﬁdence level. The category
of skilled/professional occupations includes the professional, clerical, and skilled manual job groups in the
DHS; common occupations in these groups include teaching, healthcare-related work, associate business ad-
ministration, and crafts, garment, and trade work. The increase in employment in the skilled and professional
sector seems to be driven by a reduction in the likelihood of employment in the unskilled and agriculture
sectors, although the estimated eﬀect in these sectors is not statistically signiﬁcant at conventional levels.
Furthermore, the employment results are not being driven by employment decisions of the husband. Only
14 percent of women in Ethiopia work at the same job as their husband; when they are removed from
the sample, the estimated eﬀect of a year of schooling on the likelihood of skilled/professional employment
remains large, 0.058, and statistically signiﬁcant at the 95 percent conﬁdence level.14
The ﬁnal column of table 4 examines whether the increase in education generates the expected negative
relationship between the opportunity cost of time and ideal family size. The estimate in column 5 indicates
that each additional year of schooling reduces a woman’s ideal number of children by 0.786. The magnitude
of this change is larger than the estimated reduction in number of births in table 2.15 This provides evidence
that the increased labor market productivity is leading to women desiring fewer children, one of the three
pathways through which the household fertility decision is made, but also that they may be constrained
away from fully adjusting the number of births to match their desired change.
6.3 National Results with Consideration of Mother Tongue Instruction
The analysis is repeated by including the four MTI regions and adding intensity measures for the predicted
exposure to the new language of instruction. Detailed discussion of the impact of the MTI program on
schooling outcomes can be found in Chicoine (2019), and the calculations of the region-speciﬁc intensity
measures can be found in Appendix Section C. The inclusion of the MTI regions and consideration of the
combined eﬀect of the FPE and MTI reforms yields estimates that are consistent with those discussed in the
14 These estimates can be found in Appendix Table C.12.
15 Ideal number of children is censored at 20; no women in the DHS report having more than 18 children. Non-numerical
responses are assigned the maximum value, and a tobit model is estimated.
14
preceding subsections.
The estimated eﬀect of each additional year of schooling on fertility is smaller when the MTI regions
are included: each additional year of schooling yields 0.273 fewer births. However, the estimated eﬀect is
statistically signiﬁcant, at no less than the 95 percent conﬁdence level across all three dataset combinations.
The national estimates also yield similar conclusions for literacy, the likelihood of reading about family
planning, and reduced acceptance of domestic violence. Similar to the FPE-only estimates, inclusion of the
MTI reform and regions in the analysis also generates evidence that women become more likely to work in
skilled/professional occupations, and produces slightly larger point estimates in the reduction of ideal family
size, although with a p-value of 0.16.16
The introduction of the MTI reform in the analysis both increases the size of the dataset and adds extra
sources of variation via the region-speciﬁc introductions of the new languages of instruction. In addition to
replicating the previous analysis, the inclusion of the MTI reform allows for analyses that focus on subsets
of the data. First, to study the impact of schooling on the marriage market, the sample is restricted to
include only married women. This analysis ﬁnds evidence that the increase in women’s schooling leads to
their marrying men with higher levels of schooling, even though husbands are an average of seven years older
than their wives and largely unaﬀected by the reforms. Furthermore, husbands are more likely to be working
in service and sales sectors, and are no more likely to want more children than their wives, even when she
desires fewer children. These ﬁndings again suggest that improvements in labor market outcomes and an
increased opportunity cost of time are likely the drivers of reductions in fertility.
The extension also allows for an examination of how schooling impacts the timing of birth and marriage
decisions at speciﬁc ages. These results provide evidence that the reforms reduce the likelihood of a woman
being married at the ages of 21 to 24, and they reduce the likelihood of a woman’s ﬁrst birth occurring
between ages 23 and 25.17 This suggests that the reforms are leading to a postponement of marriage and
ﬁrst birth for women in their early twenties. This timing, signiﬁcantly after the completion of primary
school, reduces the possibility of an incarceration eﬀect driving the results. A remaining concern is that
postponements of early fertility decisions tend to be replaced by additional births at later ages (Black et al.,
2008; Geruso and Royer, 2018). However, the reforms in Ethiopia have a greater negative eﬀect on fertility
at each subsequent age from 22 through 29; the evidence suggests that the reduction in fertility actually
increases as women age.18
16 Allestimates of joint eﬀect of both reforms can be found in Appendix Section C.3.
17 These results can be found in Appendix Figure C.2.
18 These results can be found in Appendix Figure C.1.
15
6.4 Threats to Validity
6.4.1 Contemporaneous Investment in Lagging Areas
The post-reform magnitudes from equation (1) are inversely related to pre-reform levels of schooling. If
the government matched the FPE program with increased levels of investment in lagging regions of the
country, these investments would be correlated with post-reform levels of the intensity measure. Examining
the correlation between pre-reform education levels and the change in regional spending on education would
provide insight into how funding was allocated following the implementation of the reform; ﬁnding a strong
negative correlation would suggest a disproportionate increase in funding to areas with lower pre-reform levels
of schooling. Levels of regional per-student spending in 1993, the ﬁrst year for which data are available,
exhibit a strong positive correlation with pre-reform education levels, as would be expected. Then comparing
pre-reform education levels with the growth in spending through 1996, as the reforms are implemented, and
through 2001, well after the implementation, yields correlations of 0.01 and 0.17, respectively (World Bank,
2005). This indicates that there is very little relationship between pre-reform education levels and the
post-reform investment decisions of the regional governments.
Furthermore, the inclusion of the MTI reform in the analysis does not change the article’s main results.
A beneﬁcial characteristic of including the MTI reform is that the identiﬁcation strategy of the combination
of the FPE and MTI reforms does not simply exploit a change in policy at a single point in time; the
variation exploited by the joint reforms is introduced at four points in time, in diﬀerent parts of the country.
The initial returns to MTI in Tigray were found to be positive; this was followed by the introduction of
MTI with script change in three additional regions, which initially put downward pressure on schooling
prior to the removal of school fees (Chicoine, 2019). The remaining seven regions of the country were then
positively aﬀected by the removal of school fees in 1995. The pattern of results for schooling, fertility, and
the mechanisms linking the two are largely consistent; therefore, any alternative explanation would have to
follow this pattern, signiﬁcantly reducing the possibility that the intensity measure is capturing spurious
correlations that could be assigned to a competing policy. The pattern of implementation is likely unique to
the combination of the FPE and MTI reforms.
6.4.2 Quality
One concern is that increases in class size that occurred after the implementation of the reform could lead
to a reduction of quality of education following the reform. However, this reduction in education quality
would not directly impact the ﬁrst stage, which measures years of schooling, not learning. Any reductions
in quality of education correlated with larger increases in enrollment would simply make it less likely that
16
there is any impact of the increase in schooling on later-in-life outcomes. It is doubtful that less learning
(lower-quality education) in the early years of primary school would lead to reduced fertility and improved
future labor market outcomes. If anything, even for students that would have attended school anyway, this
would likely attenuate estimates toward zero. The evidence of these long-term improvements and evidence of
signiﬁcant increases in literacy suggest that learning occurred at a level suﬃcient to generate consequential
later-in-life improvements.
6.4.3 Conclusion of Ethiopian Civil War
The long-simmering conﬂict in Ethiopia erupted in the late 1980s, with a vast majority of the ﬁghting
occurring to the north of the capital, Addis Ababa. Geocoded data from the Uppsala Conﬂict Data Program
(Sundberg and Melander, 2013; Croicu and Sundberg, 2015; Allansson et al., 2017) make it possible to match
deaths related to “organized violence” that occurred as early as 1989 to the zones used in the study.
To investigate whether characteristics of the areas most aﬀected by the civil war are driving the results,
zones are removed at two separate cutoﬀs and the models re-estimated. The ﬁrst cutoﬀ removes four zones
that had over 4,000 deaths between 1989 and 1991; these zones contained over 75 percent of all deaths
during this period. A less restrictive cutoﬀ removes all zones with at least 500 deaths related to organized
violence; these zones account for 96 percent of all deaths included in the data. Removing these zones from
the data and re-estimating equations (8) and (9) generates similar sets of ﬁndings. The results reveal a
similar pattern: 12 diﬀerent ﬁrst-stage speciﬁcations yield a consistent eﬀect of the FPE intensity measure
that ranges between 0.098 and 0.135. The 2SLS estimates show that each additional year of schooling led
to between 0.254 and 0.550 fewer births for Ethiopian women, estimates that remain both qualitatively and
quantitatively similar to the baseline ﬁndings of the article.19
6.4.4 Child Marriage Law
In 2000, Ethiopia changed the minimum legal age of marriage from 15 to 18. In the following decade,
regions throughout the country adopted the law (McGavock, 2015; Garcia Hombrados, 2018). However, the
combined FPE and MTI analysis ﬁnds statistically signiﬁcant delays in marriage only between the ages of
21 and 24.20 This timing means that the result is unlikely to be related to the law outlawing marriage prior
to age 18.
19 These estimates, and those for the paper’s other outcomes, using the mortality cutoﬀs can be found in Panel L and Panel
M of tables throughout Appendix Section D.2.
20 These results can be found in Appendix Figure C.2.
17
7 Conclusion
This article ﬁnds evidence that free primary education led to an increase in schooling in Ethiopia, and
that the increase in schooling led to a signiﬁcant reduction in the number of births for Ethiopian women.
This reduction is partially generated through a delay in ﬁrst marriage and birth, and a reduced demand
for children is also found to be associated with new labor market opportunities. There is no evidence of
increased empowerment for women or of any change in the likelihood of contraception use. The totality of
the evidence suggests that the central mechanism through which the increase in education generated by the
removal of school fees reduces fertility is via the increase in women’s labor market activity and the associated
reduction in their ideal number of children.
The identiﬁcation strategy employed in this article can be used to causally identify the returns to increased
levels of schooling generated by national-level reforms. It provides a powerful tool for examining the return
to free primary education in any number of countries, which is an area of research in need of additional
attention, as highlighted by 3ie’s report (Snilstveit et al., 2016). The results of this article suggest that large
increases in enrollment, often generated by the removal of school fees, are able to outweigh any possible
negative eﬀect of declining education quality, a ﬁnding that is consistent with recent work of Keats (2018)
and Zenebe Gebre (2018). This is an important ﬁnding because the removal of school fees is a policy lever
that has been used in many parts of the world.
18
References
Ainsworth, M., Beegle, K., and Nyamete, A. (1996). The Impact of Women’s Schooling on Fertility and
Contraceptive Use: A Study of Fourteen sub-Saharan African Countries. World Bank Economic Review,
10(1):85–122.
Allansson, M., Melander, E., and Themner, L. (2017). Organized Violence, 1989-2016. Journal of Peace
Research, 54(4):574–587.
Anderson, T. and Rubin, H. (1949). Estimation of the Parameters of a Single Equation in a Complete System
of Stochastic Equations. Annals of Mathematical Statistics, 20(1):46–63.
Ashraf, N., Field, E., and Lee, J. (2014). Household Bargaining and Excess Fertility: An Experimental
Study in Zambia. American Economic Review, 104(7):2210–2237.
Bertrand, M., Duﬂo, E., and Mullainathan, S. (2004). How Much Should We Trust Diﬀerences-In-Diﬀerences
Estimates? Quarterly Journal of Economics, 119(1):249–275.
Black, S. E., Devereux, P. J., and Salvanes, K. G. (2005). Why the Apple Doesn’t Fall Far: Understanding
Intergenerational Transmission of Human Capital. American Economic Review, 95(1):437–449.
Black, S. E., Devereux, P. J., and Salvanes, K. G. (2008). Staying in the Classroom and Out of the Maternity
Ward? The Eﬀect of Compulsory Schooling Laws on Teenage Births. Economic Journal, 118(530):1025–
1054.
Bleakley, H. (2010). Malaria Eradication in the Americas: A Retrospective Analysis of Childhood Exposure.
American Economic Journal: Applied Economics, 2(2):1–45.
Boothe, K. and Walker, R. (1997). Mother Tongue Education in Ethiopia: From Policy to Implementation.
Language Problems and Language Planning, 21(1):1–19.
Card, D. (2001). Estimating the Return to Schooling: Progress on Some Persistent Econometric Poblems.
Econometrica, 69(5):1127–1160.
Central Statistical Authority (Ethiopia) and ICF (2016). Ethiopia Demographic and Health Survey 2016.
Central Statistical Authority (Ethiopia) and ICF International (2011). Ethiopia Demographic and Health
Survey 2011.
Central Statistical Authority (Ethiopia) and ORC Macro (2005). Ethiopia Demographic and Health Survey
2005.
19
Chicoine, L. (2019). Schooling with Learning: The Eﬀect of Free Primary Education and Mother Tongue
Instruction Reforms in Ethiopia. Economics of Education Review, 69(April):94–2017.
Clark, D. and Bono, E. D. (2016). The Long-run Eﬀects of Attending an Elite School: Evidence from the
United Kingdom. American Economic Journal: Applied Economics, 8(1):150–176.
Croicu, M. and Sundberg, R. (2015). UCDP GED Codebook version 2.0. Department of Peace and Conﬂict
Research, Uppsala University.
Finlay, K., Magnusson, L., and Schaﬀer, M. E. (2013). WEAKIV: Stata Module to Perform Weak-instrument-
robust Tests and Conﬁdence Intervals for Instrumental-variable (IV) Estimation of Linear, Probit and
Tobit Models.
Food and Agriculture Organization of the United Nations. FAO GeoNetwork. (2015). Global Administrative
Unit Layers (GAUL).
Fort, M., Schneeweis, N., and Winter-Ebmer, R. (2016). Is Education Always Reducing Fertility? Evidence
from Compulsory Schooling Reforms. Economic Journal, 126(595):1823–1855.
Garcia Hombrados, J. (2018). Child Marriage and Infant Mortality: Evidence from Ethiopia. Working Paper
No. 1317. Department of Economics, University of Sussex.
Geruso, M. and Royer, H. (2018). The Impact of Education on Family Formation: Quasi-Experimental
Evidence from the UK. NBER Working Paper No. 24332.
Heugh, D. K., Benson, C., Bogale, B., and Yohannes, M. A. G. (2007). Study on Medium of Instruction in
Primary Schools in Ethiopia. Final Report. Commissioned by the Ministry of Education.
Holmlund, H., Lindahl, M., and Plug, E. (2011). The Causal Eﬀect of Parents’ Schooling on Children’s
Schooling: A Comparison of Estimation Methods. Journal of Economic Literature, 49(3):615–651.
Kattan, R. B. and Burnett, N. (2004). User Fees in Primary Education. World Bank Education Advisory
Service.
Keats, A. (2018). Women’s Schooling, Fertility, and Child Health Outcomes: Evidence from Uganda’s Free
Primary Education Program. Journal of Development Economics. Forthcoming.
Lam, D. and Duryea, S. (1999). Eﬀects of Schooling on Fertility, Labor Supply, and Investments in Children,
with Evidence from Brazil. Journal of Human Resources, 34(1):160–192.
20
Lucas, A. M. (2010). Malaria Eradication and Educational Attainment: Evidence from Paraguay and Sri
Lanka. American Economic Journal: Applied Economics, 2(2):46.
Lucas, A. M. (2013). The Impact of Malaria Eradication on Fertility. Economic Development and Cultural
Change, 61(3):607–631.
Lucas, A. M. and Mbiti, I. M. (2012a). Access, Sorting, and Achievement: The Short-Run Eﬀects of Free
Primary Education in Kenya. American Economic Journal: Applied Economics, 4(4):226–253.
Lucas, A. M. and Mbiti, I. M. (2012b). Does Free Primary Education Narrow Gender Diﬀerences in School-
ing? Evidence from Kenya. Journal of African Economies, 21(5):691–722.
Lundborg, P., Nilsson, A., and Rooth, D.-O. (2014). Parental Education and Oﬀspring Outcomes: Evidence
from the Swedish Compulsory School Reform. American Economic Journal: Applied Economics, 6(1):253–
278.
McCrary, J. and Royer, H. (2011). The Eﬀect of Female Education on Fertility and Infant Health: Evidence
from School Entry Policies Using Exact Date of Birth. American Economic Review, 101(1):158–195.
McGavock, T. J. (2015). Child Brides, Bargaining Power, and Reform of Ethiopia’s Family Law. Working
Paper.
Minnesota Population Center and the Ethiopian Central Statistical Agency (2017). Integrated Public Use
Microdata Series, International: Version 6.5.
Negash, T. (1996). Rethinking Education in Ethiopia. Uppsala. Nordiska Afrikainstituet.
Ofcansky, T. P. and Berry, L. B. (1993). Ethiopia, A Country Study. Washington DC. Federal Research
Division, Library of Congress.
Osili, U. O. and Long, B. T. (2008). Does Female Schooling Reduce Fertility? Evidence from Nigeria.
Journal of Development Economics, 87(1):57–75.
Oumer, J. (2009). The Challenges of Free Primary Education in Ethiopia. UNESCO: International Institute
for Educational Planning.
Ozier, O. (2018). The Impact of Secondary Schooling in Kenya: A Regression Discontinuity Analysis. Journal
of Human Resources, 53(1):157–188.
Ramachandran, R. (2017). Language Use in Education and Human Capital Formation: Evidence from the
Ethiopian Educational Reform. World Development, 98:195–213.
21
Schultz, T. P. (1994). Human Capital, Family Planning, and Their Eﬀects on Population Growth. American
Economic Review, 84(2):255–260.
Schultz, T. P. (1997). Demand for Children in Low Income Countries. Handbook of Population and Family
Economics, 1:349–430.
Seid, Y. (2016). Does learning in mother tongue matter? Evidence from a natural experiment in Ethiopia.
Economics of Education Review, 55:21–38.
Snilstveit, B., Stevenson, J., Menon, R., Phillips, D., Gallagher, E., Geleen, M., Jobse, H., Schmidt, T., and
Jimenez, E. (2016). The Impact of Education Programmes on Learning and School Participation in Low-
and Middle-Income Countries. 3ie Systematic Review Summary, 7.
Sundberg, R. and Melander, E. (2013). Introducing the UCDP Georeferenced Event Dataset. Journal of
Peace Research, 50(4):523–532.
UNESCO (2007). World Data on Education. Geneva. UNESCO: International Bureau of Education.
USAID (1987). Final Disaster Report: The Ethiopian Drought/Famine. Technical report, USAID, Addis
Ababa.
World Bank (2005). Education in Ethiopia: Strengthening the Foundation for Sustainable Progress. Wash-
ington D.C. World Bank.
World Bank (2009). Abolishing School Fees in Africa: Lessons from Ethiopia, Ghana, Kenya, Malawi, and
Mozambique. Washington D.C. World Bank.
Zenebe Gebre, T. (2014). Eﬀects of Mother Tongue Education on Schooling and Child Labor Outcomes.
Working Paper.
Zenebe Gebre, T. (2018). Free Primary Education, Timing of Fertility, and Total Fertility. World Bank
Economic Review, Forthcoming.
22
Figures and Tables
Figure 1: Map of Ethiopian Regions (Dark Border) and Zones.
Source: Author’s creation using spatially harmonized ﬁrst- and second-level administrative boundaries from
IPUMS International.
Note: This ﬁgure is a reproduction of ﬁg. 1 in Chicoine (2019).
23
Figure 2: FPE Intensity Measure in Non-MTI Regions, by Birth Year
Source: Author’s analysis based on years-of-schooling data from the 1994 Ethiopian census and school-
attendance data from the 2007 Ethiopian census.
Note: The ﬁgure shows the average maximum number of school years gained following the removal of school
fees throughout Ethiopia in 1995. The data are from 30 zones within seven regions—Addis Ababa, Aﬀar,
Amhara, Benishangul-Gumuz, Gambela, Harari, Somali—that did not introduce mother tongue instruction
(MTI) prior to implementation of the free primary education (FPE) program.
24
Table 1: Summary Statistics
Birth Cohorts 1968 to 1970 1988 to 1990
N Mean N Mean
Years of Schooling 1,448 1.477 2,950 4.140
Literacy 1,430 0.164 2,857 0.427
Number of Births 1,448 5.578 2,950 0.941
Births by Age 20 1,448 1.317 1,654 0.813
Births by Age 25 1,448 2.741 674 1.749
Ideal Number of Children 1,448 7.704 2,950 4.527
Currently Working 1,446 0.321 2,945 0.333
Sector of Current Work
Skilled Manual or Professional 1,429 0.082 2,912 0.098
Service or Sales 1,429 0.100 2,912 0.138
Unskilled Manual or Agriculture 1,429 0.312 2,912 0.240
Read About Family Planning 1,448 0.063 2,947 0.115
Knowledge of Modern
1,448 0.920 2,950 0.911
Family Planning Method
Source: Author’s analysis based on data for women in the 2005, 2011, and 2016
rounds of the Ethiopian Demographic and Health Survey (DHS).
Note: Ideal number of children is censored at 20; no women in the DHS report hav-
ing more than 18 children, and non-numerical responses are assigned the maximum
value. Skilled manual or professional jobs include professional, clerical, and skilled
manual occupation groups; the other categories exactly describe the included job
groups.
25
Table 2: Eﬀect of Years of Schooling on Number of Children Born
Number of Years of Number of Number of
Children Born Schooling Children Born Children Born
(OLS) (First Stage) (Reduced Form) (2SLS)
(1) (2) (3) (4)
A. Census + DHS
Years of Schoolingizy -0.120
(0.016)
[0.000]
Add’l Years of Free 0.131 -0.057
FPE
Schooling Izy (0.034) (0.016)
[0.001] [0.001]
Years of Schoolingizy -0.437
(0.090)
[0.000]
First Stage F-Statistic 14.80 14.80
Number of Clusters 32 32 32 32
N 83,005 83,005 83,005 83,005
B. Census Only
Years of Schoolingizy -0.097
(0.015)
[0.000]
Add’l Years of Free 0.154 -0.064
FPE
Schooling Izy (0.025) (0.014)
[0.001] [0.000]
Years of Schoolingizy -0.417
(0.074)
[0.000]
First Stage F-Statistic 37.62 37.62
Number of Clusters 32 32 32 32
N 69,083 69,083 69,083 69,083
C. DHS Only
Years of Schoolingizy -0.130
(0.017)
[0.000]
Add’l Years of Free 0.112 -0.059
FPE
Schooling Izy (0.046) (0.021)
[0.021] [0.007]
Years of Schoolingizy -0.529
(0.165)
[0.001]
First Stage F-Statistic 5.93 5.93
Number of Clusters 30 30 30 30
N 13,922 13,922 13,922 13,922
Source: Author’s analysis in panel A is based on data from the Ethiopian census of 2007
and from the Demographic and Health Survey (DHS) in years 2005, 2011, and 2016; each
data source is used separately in panels B and C.
Note: The dependent variable is years of schooling in column 2 and is number of births
in the other three columns. Years of Schoolingizy is the reported number of years of
schooling from the data; Years of Schoolingizy is the predicted number of years of school-
ing, instrumented with the free primary education (FPE) intensity measure, Izy FPE . All
samples include women in birth cohorts from 1970 to 1988. All regressions include birth
year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age when multiple
survey waves are included. Standard errors are clustered at the zone level and shown in
parentheses; p-values are shown in square brackets.
26
Table 3: Eﬀect of Years of Schooling on Knowledge and Health
Read about Know about BMI Height Acceptable Reasons for Use Modern Use Hidden
Literacy Fam. Planning Fam. Planning (z-score) (z-score) Domestic Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
Years of Schoolingizy 0.092 0.048 -0.013 0.316 -0.271 -0.361 -0.018 -0.035
(0.028) (0.029) (0.024) (0.355) (0.302) (0.211) (0.051) (0.042)
[0.001] [0.097] [0.594] [0.374] [0.369] [0.087] [0.721] [0.402]
Mean of Dependent 0.164 0.063 0.920 0.085 -0.160 2.318 0.193 0.142
(Pre-Reform Cohorts)
27
First Stage F-Statistic 6.10 5.92 5.93 1.91 2.22 5.67 5.93 5.93
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: BMI = body mass index. The dependent variable is described at the top of each of the eight columns. In columns 1–3, 7, and 8 it is an indicator that equals 1 if true; in
columns 4 and 5 it is a standardized value of the described outcome; and in column 6 it is the count from 0 to 5 of acceptable reasons for domestic violence (going out without
permission, neglecting children, arguing with husband, refusing sex, burning food). Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary
education (FPE) intensity measure, IzyFPE . All samples include women in birth cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc
linear trends, and a cubic for age. Each estimate is from a unique regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in
square brackets.
Table 4: Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference
Sector of Work
Skilled / Service / Agriculture / Ideal Number
Working Professional Sales Unskilled Manual of Children
(1) (2) (3) (4) (5)
Years of Schoolingizy 0.093 0.059 0.064 -0.048 -0.786
(0.058) (0.028) (0.047) (0.031) (0.468)
[0.107] [0.033] [0.169] [0.116] [0.093]
Mean of Dependent 0.321 0.082 0.100 0.312 7.704
(Pre-Reform Cohorts)
First Stage F-Statistic 6.06 6.63 6.63 6.63 6.26
N 13,909 13,755 13,755 13,755 13,789
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: The dependent variable is described at the top of each of the ﬁve columns. In columns 1–4 it is an indicator that
equals 1 if true, and in column 5 it is the ideal number of children. Skilled/Professional jobs include professional, clerical, and
skilled manual job groups; the other categories exactly describe the occupation groups included. Ideal number of children is
censored at 20; no women in the Demographic and Health Survey report having more than 18 children, and non-numerical
responses are assigned the maximum value. Years of Schoolingizy is the predicted level of schooling, instrumented with the
free primary education (FPE) intensity measure, IzyFPE . All samples include women in birth cohorts from 1970 to 1988. All
regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a
unique regression, and the second-stage estimate in column 5 is generated using a tobit model. Standard errors are clustered
at the zone level and shown in parentheses; p-values are shown in square brackets.
28
Appendix
Summary of Appendix
Sections
A Timing and Equations for FPE Intensity Measure
B Additional Results: Tables and Figures
C Mother Tongue Instruction
C.1 Background
C.2 MTI Intensity Measure, by Region
C.2.1 Oromia
C.2.2 SNNPR
C.2.3 Dire Dawa
C.2.4 Tigray
C.3 National Estimates: Accounting for FPE and MTI
C.4 Combined Instrument and Reduced Form Estimates, by Age
D Alternative Samples and Speciﬁcations
D.1 Pre-Treatment Trends and Placebo Estimates
D.2 Alternative Samples and Speciﬁcations
A.1
Figures
B.1 School Entry Probabilities from Three Census Rounds
C.1 Eﬀect of Reforms on Number of Births, by Age
C.2 Eﬀect of Reforms on Timing of First Birth, Marriage, and Intercourse, by Age
D.1 Comparison of Pre-treatment Trends
Tables
A.1 Timing of FPE Reform with On Time Entry, by Birth Year
Anderson-Rubin Weak IV Conﬁdence Sets
B.1 Eﬀect of FPE on Years of Schooling and Number of Children Born
B.2 Eﬀect of Years of Schooling on Knowledge and Health
B.3 Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference
B.4 Eﬀect of Schooling on Beliefs Regarding Domestic Violence
C.1 Oromia – MTI Implementation
C.2 SNNPR – MTI Implementation
C.3 Dire Dawa – MTI Implementation
C.4 Tigray – MTI Implementation
C.5 MTI+Script Regions: Timing of FPE and MTI Reforms with On Time Entry, by Birth Year
C.6 Tigray: Timing of FPE and MTI Reforms with On Time Entry, by Birth Year
National Estimates with Separate Instruments: Accounting for FPE and MTI
C.7 Eﬀect of Reforms on Years of Schooling and Number of Births (Census + DHS)
C.8 Eﬀect of Reforms on Years of Schooling and Number of Births (Census)
C.9 Eﬀect of Reforms on Years of Schooling and Number of Births (DHS)
C.10 Eﬀect of Years of Schooling on Knowledge and Health
C.11 Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference
C.12 Eﬀect of Years of Schooling on Sector of Employment - Job Diﬀerent than Husband
C.13 Eﬀect of Schooling on Husband’s Characteristics
C.14 Eﬀect of Schooling Regarding Domestic Violence (Married Women Only)
C.15 Eﬀect of Schooling on Beliefs Regarding Women’s Empowerment
A.2
Tables continued...
C.16 Testing for Diﬀerences Across Reform Instruments
National Estimates with Combined Instrument: Accounting for FPE and MTI
C.17 Eﬀect of Reforms on Years of Schooling and Number of Births (DHS)
C.18 Eﬀect of Years of Schooling on Knowledge and Health
C.19 Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference
C.20 Eﬀect of Schooling on Husband’s Characteristics
C.21 Eﬀect of Reforms on Likelihood of First Birth, Marriage, and Intercourse, by Age
D.1 Placebo Estimates of Misplaced Timing of Reform Prior to Actual Implementation
D.2 Additional Language(s) in Boothe and Walker (1997) Deﬁnition
D.3 Additional Language(s) in Zenebe Gebre (2014) Deﬁnition
D.4 Eﬀect of FPE and MTI Reforms on Years of School
D.5 Eﬀect of Reforms on Years of Schooling: Joint Intensity Measure
D.6 Eﬀect of Years of Schooling on Knowledge and Health
D.7 Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference
A.3
A Timing and Equations for FPE Intensity Measure
FPE
Explicit equations used to calculate the FPE intensity measure Izy for women in each zone z , and birth
year y , are listed below. The magnitude (Mz,y ) and start age (Sz,a ) variables used in the calculation are
described in Section 3. The timing of how the reform impacts each cohort, assuming school entrance at age
seven, is outlined in Appendix Table A.1. Those born in 1972 and who enter school at age 12, ﬁve years late,
would still complete all ten years of schooling prior to the implementation of the reform (reference the 1977
(= 1972 + 5) birth cohort in Appendix Table A.1). Members of the 1972 birth cohort, or any previous cohort,
could start school at any relevant age, from six to 12, and still not be aﬀected by the reform; therefore,
FPE
Iz,1972 = 0.
Those born in 1973 and entering school at age 12 would potentially receive their tenth year of education
for free, but only if they made it through the ﬁrst nine grades. Those born in 1974 and starting at 12 could
potentially have up to two free years of schooling, only if they have completed the ﬁrst eight grades, and if
starting at age 11 only one free year of school, and so on:
FPE FPE
Iz, 1973 = Sz,12 Mz,1978 = Sz,12 (10 − 9) Fz,9 ,
9
FPE FPE
Iz, 1974 = Sz,12 Mz,1979 + Sz,11 Mz,1978 = Sz,12 (10 − g ) Fz,g + Sz,11 (10 − 9) Fz,9 .
g =8
This iteration continues in the same way through the 1981 cohort. All students born through 1981, even
those who start school ﬁve years late, at age 12, will make the school entry decision in the pre-reform period.
FPE FPE FPE FPE
Iz,1975 = Sz,12 Mz,1980 + Sz,11 Mz,1979 + Sz,10 Mz,1978 ,
FPE FPE FPE FPE FPE
Iz,1976 = Sz,12 Mz, 1981 + Sz,11 Mz,1980 + Sz,10 Mz,1979 + Sz,9 Mz,1978 ,
FPE FPE FPE FPE FPE FPE
Iz, 1977 = Sz,12 Mz, 1982 + Sz,11 Mz,1981 + Sz,10 Mz,1980 + Sz,9 Mz,1979 + Sz,8 Mz,1978 ,
FPE FPE FPE FPE FPE FPE FPE
Iz, 1978 = Sz,12 Mz,1983 + Sz,11 Mz,1982 + Sz,10 Mz,1981 + Sz,9 Mz,1980 + Sz,8 Mz,1979 + Sz,7 Mz,1978 ,
FPE FPE FPE FPE FPE
Iz,1979 = Sz,12 Mz,1984 + Sz,11 Mz,1983 + Sz,10 Mz,1982 + Sz,9 Mz,1981
FPE FPE FPE
+Sz,8 Mz, 1980 + Sz,7 Mz,1979 + Sz,6 Mz,1978 ,
A.4
FPE FPE FPE FPE FPE
Iz,1980 = Sz,12 Mz,1985 + Sz,11 Mz,1984 + Sz,10 Mz,1983 + Sz,9 Mz,1982
FPE FPE FPE
+Sz,8 Mz, 1981 + Sz,7 Mz,1980 + Sz,6 Mz,1979 ,
FPE FPE FPE FPE FPE
Iz,1981 = Sz,12 Mz,1986 + Sz,11 Mz,1985 + Sz,10 Mz,1984 + Sz,9 Mz,1983
FPE FPE FPE
+Sz,8 Mz, 1982 + Sz,7 Mz,1981 + Sz,6 Mz,1980 ,
The 1982 cohort is the ﬁrst to incorporate the possibility of post-reform entry, as described in detail in
Section 3 using the 1985 cohort as an example. As described with the 1985 example, there is a stock of
students at each age that does not enter school when fees are in place, but would have entered if given the
opportunity to enter for free (Sz,a − Sz,a,pre ). For the 1982 cohort, these students have the opportunity to
enter at age 12; at this late age, there is a possibility that they may be tied to some other activity that
constrains them from entering school. The further this earliest post-reform entry age is from the legal entry
age of seven, the greater the decline in entry for would-be post-reform entrants 1/ea−7 :
FPE FPE FPE FPE FPE
Iz, 1982 = Sz,6 Mz,1981 + Sz,7 Mz,1982 + Sz,8 Mz,1983 + Sz,9 Mz,1984
11
FPE FPE FPE 1
+ Sz,10 Mz,1985 + Sz,11 Mz,1986 + Mz Sz,12 + [(10) Fz,0 ] (Sz,a − Sz,a,pre ) ,
e12−7 a=6
FPE FPE FPE FPE FPE
Iz,1983 = Sz,6 Mz,1982 + Sz,7 Mz,1983 + Sz,8 Mz,1984 + Sz,9 Mz,1985
12 10
FPE FPE 1
+ Sz,10 Mz, 1986 + Mz Sz,a + [(10) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=11
e11−7 a=6
FPE FPE FPE FPE FPE
Iz,1984 = Sz,6 Mz,1983 + Sz,7 Mz,1984 + Sz,8 Mz,1985 + Sz,9 Mz,1986
12 9
FPE 1
+ Mz Sz,a + [(10) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=10
e10−7 a=6
12 8
FPE FPE FPE FPE FPE 1
Iz, 1985 = Sz,6 Mz,1984 + Sz,7 Mz,1985 + Sz,8 Mz,1986 + Mz Sz,a + [(10) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=9
e9−7 a=6
12 7
FPE FPE FPE FPE 1
Iz,1986 = Sz,6 Mz,1985 + Sz,7 Mz,1986 + Mz Sz,a + [(10) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=8
e 8−7 a=6
12
FPE FPE FPE
Iz,1987 = Sz,6 Mz,1986 + Mz Sz,a + [(10) Fz,0 ] (Sz,6 − Sz,6,pre ) ,
a=7
FPE FPE
Iz, 1988 = Mz
A.5
Table A.1: Timing of FPE Reform with On Time Entry, by Birth Year
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
1977 Born 1978 Born 1979 Born 1980 Born 1981 Born 1982 Born
1978 0 1979 0 1980 0 1981 0 1982 0 1983 0
1979 1 1980 1 1981 1 1982 1 1983 1 1984 1
1980 2 1981 2 1982 2 1983 2 1984 2 1985 2
1981 3 1982 3 1983 3 1984 3 1985 3 1986 3
1982 4 1983 4 1984 4 1985 4 1986 4 1987 4
1983 5 1984 5 1985 5 1986 5 1987 5 1988 5
1984 6 1985 6 1986 6 1987 6 1988 6 1989 6
1985 G1 7 1986 G1 7 1987 G1 7 1988 G1 7 1989 G1 7 1990 G1 7
1986 G2 8 1987 G2 8 1988 G2 8 1989 G2 8 1990 G2 8 1991 G2 8
1987 G3 9 1988 G3 9 1989 G3 9 1990 G3 9 1991 G3 9 1992 G3 9
1988 G4 10 1989 G4 10 1990 G4 10 1991 G4 10 1992 G4 10 1993 G4 10
1989 G5 11 1990 G5 11 1991 G5 11 1992 G5 11 1993 G5 11 1994 G5 11
1990 G6 12 1991 G6 12 1992 G6 12 1993 G6 12 1994 G6 12 1995 G6 12 FPE
1991 G7 13 1992 G7 13 1993 G7 13 1994 G7 13 1995 G7 13 FPE 1996 G7 13 FPE
1992 G8 14 1993 G8 14 1994 G8 14 1995 G8 14 FPE 1996 G8 14 FPE 1997 G8 14 FPE
1993 G9 15 1994 G9 15 1995 G9 15 FPE 1996 G9 15 FPE 1997 G9 15 FPE 1998 G9 15 FPE
1994 G10 16 1995 G10 16 FPE 1996 G10 16 FPE 1997 G10 16 FPE 1998 G10 16 FPE 1999 G10 16 FPE
A.6
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
1983 Born 1984 Born 1985 Born 1986 Born 1987 Born 1988 Born
1984 0 1985 0 1986 0 1987 0 1988 0 1989 0
1985 1 1986 1 1987 1 1988 1 1989 1 1990 1
1986 2 1987 2 1988 2 1989 2 1990 2 1991 2
1987 3 1988 3 1989 3 1990 3 1991 3 1992 3
1988 4 1989 4 1990 4 1991 4 1992 4 1993 4
1989 5 1990 5 1991 5 1992 5 1993 5 1994 5
1990 6 1991 6 1992 6 1993 6 1994 6 1995 6
1991 G1 7 1992 G1 7 1993 G1 7 1994 G1 7 1995 G1 7 FPE 1996 G1 7 FPE
1992 G2 8 1993 G2 8 1994 G2 8 1995 G2 8 FPE 1996 G2 8 FPE 1997 G2 8 FPE
1993 G3 9 1994 G3 9 1995 G3 9 FPE 1996 G3 9 FPE 1997 G3 9 FPE 1998 G3 9 FPE
1994 G4 10 1995 G4 10 FPE 1996 G4 10 FPE 1997 G4 10 FPE 1998 G4 10 FPE 1999 G4 10 FPE
1995 G5 11 FPE 1996 G5 11 FPE 1997 G5 11 FPE 1998 G5 11 FPE 1999 G5 11 FPE 2000 G5 11 FPE
1996 G6 12 FPE 1997 G6 12 FPE 1998 G6 12 FPE 1999 G6 12 FPE 2000 G6 12 FPE 2001 G6 12 FPE
1997 G7 13 FPE 1998 G7 13 FPE 1999 G7 13 FPE 2000 G7 13 FPE 2001 G7 13 FPE 2002 G7 13 FPE
1998 G8 14 FPE 1999 G8 14 FPE 2000 G8 14 FPE 2001 G8 14 FPE 2002 G8 14 FPE 2003 G8 14 FPE
1999 G9 15 FPE 2000 G9 15 FPE 2001 G9 15 FPE 2002 G9 15 FPE 2003 G9 15 FPE 2004 G9 15 FPE
2000 G10 16 FPE 2001 G10 16 FPE 2002 G10 16 FPE 2003 G10 16 FPE 2004 G10 16 FPE 2005 G10 16 FPE
Source: Author’s summary based on timing of FPE reform and school entry age.
B Additional Results: Tables and Figures
Table B.1: Eﬀect of Years of Schooling on Number of Children Born – Replication of Table 2:
Anderson-Rubin Weak IV Robust Conﬁdence Sets
Census Census DHS
+ DHS Only Only
(1) (2) (3)
Coeﬃcient on Years of Schoolingizy
Weak IV Robust 95% -0.676, -0.275 -0.576, -0.285 -1.567, -0.265
Conﬁdence Set [0.000] [0.000] [0.003]
First Stage F-Statistic 14.80 37.62 5.93
Number of Clusters 32 32 30
N 83,005 69,083 13,922
Source: Author’s analysis based on data from the Ethiopian census of 2007 and the Demo-
graphic and Health Survey (DHS) in years 2005, 2011, and 2016.
Note: IV = instrumental variables. The dependent variable is number of children born.
Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary
education (FPE) intensity measure, IzyFPE . All samples include women in birth cohorts from
1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear
trends, and a cubic for age when multiple survey waves are included. Column 1 combines
data from the Ethiopian census (2007) and the DHS (2005, 2011, and 2016); each data source
is used separately in columns 2 and 3. Standard errors are clustered at the zone level, and
Anderson and Rubin (1949) conﬁdence sets of the 2SLS estimate are shown along with the
p-value from the associated chi-squared test, given in square brackets.
A.7
Table B.2: Eﬀect of Years of Schooling on Knowledge and Health – Replication of Table 3: Anderson-Rubin Weak IV Robust Conﬁdence Sets
Read about Know about BMI Height Number of Acceptable Reason Use Modern Use Hidden
Literacy Fam. Planning Fam. Planning (z-score) (z-score) for Domestic Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
Coeﬃcient on Years of Schoolingizy
Weak IV Robust 95% -0.046, 0.146 0.014, 0.278 -0.090, 0.062 — — -1.777, -0.015 -0.082, 0.366 -0.090, 0.278
Conﬁdence Set [0.088] [0.006] [0.601] — — [0.042] [0.749] [0.510]
First Stage F-Statistic 6.10 5.92 5.93 1.91 2.22 5.67 5.93 5.93
A.8
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: IV = instrumental variables; BMI = body mass index. The dependent variable is described at the top of each of the eight columns. In columns 1–3, 7, and 8 it is
an indicator that equals 1 if true; in columns 4 and 5 it is a standardized value of the described outcome; and in column 6 it is the count from 0 to 5 of acceptable reasons
for domestic violence (going out without permission, neglecting children, arguing with husband, refusing sex, burning food). Years of Schoolingizy is the predicted level of
schooling, instrumented with the free primary education (FPE) intensity measure, IzyFPE . All samples include women in birth cohorts from 1970 to 1988. All regressions include
birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Standard errors are clustered at the zone level, and Anderson and Rubin (1949) conﬁdence
sets of the 2SLS estimate are shown along with the p-value from the associated chi-squared test, given in square brackets.
Table B.3: Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference –
Replication of Table 4: Anderson-Rubin Weak IV Robust Conﬁdence Sets
Sector of Work
Skilled / Service / Agriculture / Ideal Number
Working Professional Sales Unskilled Manual of Children
(1) (2) (3) (4) (5)
Coeﬃcient on Years of Schoolingizy
Weak IV Robust 95% -0.018, 0.382 0.018, 0.218 -0.074, 0.198 -0.178, 0.014 —
Conﬁdence Set [0.093] [0.005] [0.228] [0.107] —
First Stage F-Statistic 6.06 6.63 6.63 6.63 —
N 13,909 13,755 13,755 13,755 13,789
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: IV = instrumental variables. The dependent variable is described at the top of each of the ﬁve columns. In columns
1–4 it is an indicator that equals 1 if true, and in column 5 it is the ideal number of children. Skilled/Professional jobs include
professional, clerical, and skilled manual job groups; the other categories exactly describe the occupation groups included. Ideal
number of children is censored at 20; no women in the Demographic and Health Survey report having more than 18 children, and
non-numerical responses are assigned the maximum value. Years of Schoolingizy is the predicted level of schooling, instrumented
with the free primary education (FPE) intensity measure, Izy FPE . All samples include women in birth cohorts from 1970 to
1988. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Standard errors
are clustered at the zone level, and Anderson and Rubin (1949) conﬁdence sets of the 2SLS estimate are shown along with
the p-value from the associated chi-squared test, given in square brackets. The tobit model with clustered standard errors in
column 5 is not compatible with conﬁdence set calculations from Finlay et al. (2013).
Table B.4: Eﬀect of Schooling and Reforms on Beliefs Regarding Domestic Violence
Beating justiﬁed if wife: Goes Out Neglects Argues with
w/out Permission Children Husband Refuses Sex Burns Food
(1) (2) (3) (4) (5)
Years of Schoolingizy -0.045 -0.061 -0.013 -0.099 -0.109
(0.051) (0.072) (0.042) (0.055) (0.061)
[0.380] [0.399] [0.753] [0.074] [0.073]
Mean of Dependent 0.537 0.535 0.491 0.408 0.491
(Pre-Reform Cohorts)
First Stage F-Statistic 5.66 6.20 6.05 5.05 5.32
N 13,805 13,803 13,763 13,589 13,800
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: The dependent variable in each column is an indicator that equals 1 if the statement is believed to be true and equals 0
otherwise. The sample includes all women born between 1970 and 1988. A 2SLS model is estimated where Years of Schoolingizy
FPE . All regressions
is the predicted level of schooling, instrumented with the free primary education (FPE) intensity measure, Izy
include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique
regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
A.9
(a) Raw School Entry Probabilities, By Age (b) Raw Cumulative School Entry Probabilities, By Age
A.10
(c) Scaled School Entry Probabilities, By Age (d) Scaled Cumulative School Entry Probabilities, By Age
Figure B.1: School Entry Probabilities from Three Census Rounds
Source: Author’s analysis based on data from the Ethiopian census in years 1984, 1994, and 2007.
C Mother Tongue Instruction
C.1 Background
Ethiopia has more than 80 languages (IPUMS, 2017), and in 1991, for the ﬁrst time in more than a generation,
the transitional government introduced classroom instruction in languages other than Amarigna. Amarigna
was the language preferred by the government throughout the previous decades; however, by the end of the
Transitional Government’s second year in power, the mother tongue language for 80 percent of the country’s
residents had been introduced as a language of instruction (Boothe and Walker, 1997; Zenebe Gebre, 2014).
The introduction of the mother tongue instruction program was complex.21
In 1991, the Transitional Government’s Council of Representatives selected the ﬁrst four languages to
be introduced in the mother tongue instruction program. These languages were selected from a list of
14 that were originally used in a 1979 adult literacy campaign (Boothe and Walker, 1997). The most
common language selected was Oromigna, it was spoken by 31 percent of the population at the time of
the 1994 population census, roughly the same portion of the population that spoke Amarigna. The other
three languages selected for the initial wave of the program were Tigrigna (6%), Sidamigna (3.5%), and
Wolayitigna (2.3%). These languages were respectively the fourth, ﬁfth and sixth most common mother
tongues in the country at the time. The selection of what was almost precisely the largest languages reduces
concern of favoritism from the central government, and because the new regional boundaries were drawn along
traditional ethnolinguistic borders, there was little variability of where the languages could be introduced.
The following round of languages were introduced in 1993, this group comprised of Somaligna (6%) and
three languages within the diverse Southern Nations, Nationalities, and Peoples’ Region (SNNPR) totaling
an additional 4 percent of the population (Boothe and Walker, 1997; Zenebe Gebre, 2014).22 At this point
every language spoken by at least two percent of the population had been introduced, and as time passed
political calculations are likely to have a role in the selection process for less prominent languages. Therefore,
this paper focuses on the implementation of the initial phase of the MTI project.
The ﬁnal key aspect of the introduction of the MTI program was the need to translate the primary school
material from Amarigna into the new languages. Up to this point in time, any written translations of these
languages had used the Ge’ez script; however, due to the historical connotations of the use of Amarigna, each
region was given the option to use a diﬀerent script in translation, and every region outside of Tigray selected
to translate their schooling material into the Roman script. For Tigray, this allowed them to translate all
of the material locally and introduce the MTI program on schedule in 1991 (Boothe and Walker, 1997;
21 For consistency, each language will be referenced using the name from the 2007 Ethiopian Census.
22 The languages were Hadiyigna, Gedeogna, and Kembatigna.
A.11
Zenebe Gebre, 2014). The translation for the other languages was more problematic and undertaken in a
centralized conference in Addis Ababa; to this point none of these languages had widely been translated into
the Roman script (Boothe and Walker, 1997; Heugh et al., 2007; Zenebe Gebre, 2014). This process led to
the delay of the initial implementation of the other three ﬁrst-wave languages to the 1992 school year, and
the eventual repeating of the process and 1993 introduction of the following four languages.
The analysis of MTI uses dates and languages that are independently corroborated by Boothe and
Walker (1997) and Zenebe Gebre (2014). Boothe and Walker (1997) is the most contemporaneous source
of information of which I am aware, and Zenebe Gebre (2014) directly contacted each region’s education
department more than a decade and a half later to gather information on the implementation of MTI
throughout the country. Both sources corroborate the introduction of seven of the eight languages outlined
here, only Somaligna is missing in Zenebe Gebre (2014), and is removed from the main analysis of the paper
to ensure as much accuracy as possible.23
C.2 MTI Intensity Measure, by Region
C.2.1 Oromia
Table C.1: Oromia – MTI Implementation
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Oromigna 1992 Oromia 1-8 0.93 0.84
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: MTI = mother tongue instruction; MT = mother tongue.
Assuming that students enter the classroom on time, at age 7, the following equation describes the maximum
magnitude of the MTI eﬀect for each cohort in the Oromia region.
7
M T I −O
Mz,post = wz (8 − g ) · Fz,g if y ≥ 1984
g =0
7
M T I −O
Mzy = wz (8 − g ) · Fz,g if 1977 ≤ y ≤ 1983 .
g =(1984−y )
M T I −O
Mzy =0 if y ≤ 1976
23 Alternative deﬁnitions of the MTI implementation are explored in Appendix Section D. Results using the speciﬁcation that
includes Somaligna as deﬁned by Boothe and Walker (1997) can be found in Panel P of Appendix Tables D.4 and D.5, and
another includes an additional early wave of language introductions through 1994, as deﬁned by Zenebe Gebre (2014), these
estimates can be found in Panel Q.
A.12
Table C.1 deﬁnes the characteristics used in the above Oromia magnitude equation. Grades one through
eight use the new language of instruction; therefore, only students who would have completed fewer than
eight years are aﬀected. The wz scalar is the fraction of the population in each zone that speaks the language
being introduced, and the post-MTI cohorts begin in 1984, three years prior to the ﬁrst post-FPE cohort.
The timing of MTI in Oromia for on time starters for each birth year-grade combination is shown in Appendix
Table (C.5), denoted by parentheses.
M T I −O
The cohort speciﬁc magnitudes for on time starters for Oromia Mzy are translated to the MTI
intensity measure to allow for starting age variation using a process similar to that for the FPE intensity
measure. The following set of equations describe the explicit calculations used for cohorts in Oromia.
MT I
Iz,1971 = 0,
MT I M T I −O
Iz, 1972 = Sz,12 Mz,1977 ,
MT I M T I −O M T I −O
Iz,1973 = Sz,12 Mz, 1978 + Sz,11 Mz,1977 ,
MT I M T I −O M T I −O M T I −O
Iz, 1974 = Sz,12 Mz,1979 + Sz,11 Mz,1978 + Sz,10 Mz, 1977 ,
MT I M T I −O M T I −O M T I −O M T I −O
Iz,1975 = Sz,12 Mz,1980 + Sz,11 Mz,1979 + Sz,10 Mz,1978 + Sz,9 Mz, 1977 ,
MT I M T I −O M T I −O M T I −O M T I −O M T I −O
Iz,1976 = Sz,12 Mz,1981 + Sz,11 Mz,1980 + Sz,10 Mz,1979 + Sz,9 Mz,1978 + Sz,8 Mz, 1977 ,
MT I M T I −O M T I −O M T I −O M T I −O M T I −O M T I −O
Iz,1977 = Sz,12 Mz,1982 + Sz,11 Mz,1981 + Sz,10 Mz,1980 + Sz,9 Mz,1979 + Sz,8 Mz, 1978 + Sz,7 Mz, 1977 ,
MT I M T I −O M T I −O M T I −O M T I −O
Iz,1978 = Sz,12 Mz,1983 + Sz,11 Mz, 1982 + Sz,10 Mz, 1981 + Sz,9 Mz,1980
M T I −O M T I −O M T I −O
+ Sz,8 Mz, 1979 + Sz,7 Mz, 1978 + Sz,6 Mz,1977 ,
MT I M T I −O M T I −O M T I −O M T I −O
Iz, 1979 = Sz,11 Mz,1983 + Sz,10 Mz,1982 + Sz,9 Mz,1981 + Sz,8 Mz, 1980
11
M T I −O M T I −O M T I −O 1
+ Sz,7 Mz,1979 + Sz,6 Mz,1978 + Mz,post Sz,12 + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
e12−7 a=6
MT I M T I −O M T I −O M T I −O M T I −O
Iz, 1980 = Sz,10 Mz,1983 + Sz,9 Mz,1982 + Sz,8 Mz, 1981 + Sz,7 Mz, 1980
12 10
M T I −O M T I −O 1
+ Sz,6 Mz,1979 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=11
e11−7 a=6
A.13
MT I M T I −O M T I −O M T I −O M T I −O
Iz, 1981 = Sz,9 Mz,1983 + Sz,8 Mz,1982 + Sz,7 Mz, 1981 + Sz,6 Mz, 1980
12 9
M T I −O 1
+ Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=10
e10−7 a=6
12 8
MT I M T I −O M T I −O M T I −O M T I −O 1
Iz, 1982 = Sz,8 Mz,1983 + Sz,7 Mz,1982 + Sz,6 Mz,1981 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=9
e 9−7 a=6
12 7
MT I M T I −O M T I −O M T I −O 1
Iz,1983 = Sz,7 Mz,1983 + Sz,6 Mz,1982 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=8
e 8−7 a=6
12
MT I M T I −O M T I −O
Iz, 1984 = Sz,6 Mz,1983 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,6 − Sz,6,pre ) ,
a=7
MT I M T I −O
Iz,1985 = Mz,post .
C.2.2 SNNPR
Table C.2: SNNPR – MTI Implementation
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Sidamigna 1992 SNNPR 1-4 0.99 0.18
Wolayitigna 1992 SNNPR 1-4 0.97 0.11
Hadiyigna 1993 SNNPR 1-4 0.96 0.08
Gedeogna 1993 SNNPR 1-4 0.73 0.04
Kembatigna 1993 SNNPR 1-4 0.92 0.04
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: SNNPR = Southern Nations, Nationalities, and Peoples’ Region; MTI = mother tongue instruction;
MT = mother tongue.
Unlike the other regions in the Ethiopia, SNNPR introduces new languages of instruction at two points
within the time period MTI is considered. The timing of the 1992 implementation is denoted by brackets in
Appendix Table (C.5); however, the following year two additional languages are introduced. This will have
the eﬀect of introducing two weights one for the 1992 set of languages (wz,1992 ) and another for the 1993
languages (wz,1993 ). The magnitude for each two introductions is oﬀset by one year, this yields the familiar
calculation of the maximum impact for the on time entrants:
A.14
3
M T I −S
Mzy = wz,1992 (4 − g ) · Fz,g
g =(1984−y )
3
+wz,1993 (4 − g ) · Fz,g if 1982 ≤ y ≤ 1983
g =(1985−y )
.
3
M T I −S
Mzy = wz,1992 (4 − g ) · Fz,g if y = 1981
g =(1984−y )
M T I −S
Mzy =0 if y ≤ 1980
After taking into account the diﬀerential timing of 1992 and 1993 language introductions, the consideration
of starting age variation is similar for the cohorts prior to the post-MTI entry decision.
MT I
Iz,1975 = 0,
MT I M T I −S
Iz,1976 = Sz,12 Mz,1981 ,
MT I M T I −S M T I −S
Iz,1977 = Sz,12 Mz,1982 + Sz,11 Mz,1981 ,
MT I M T I −S M T I −S M T I −S
Iz,1978 = Sz,12 Mz,1983 + Sz,11 Mz,1982 + Sz,10 Mz,1981 .
Beginning with age 12 entrants in the 1979 cohort, the entry decision for speakers of the 1992 languages
are aﬀected by the MTI reform; however, the entry decision for speakers of the 1993 languages remains
unaﬀected until the 1980 cohort. This means that the post-reform magnitude must be separated for each
of the language implementations. To incorporate this variation in access to MTI, two additional maximum
magnitude terms are deﬁned in the following way:
3
M T I −S
Mz,post = (4 − g ) · Fz,g if y ≥ 1984
g =0
3
M T I −S
Mz,1984,93 = wz,1993 (4 − g ) · Fz,g if y = 1984
g =1
The 1992 languages will be post-reform for on time school entrants beginning with the 1984 cohort, and for
simplicity, weighting adjustments to the post-reform magnitude are included within the intensity notation
below. The intensity measure for the following cohorts can then be calculated as:
MT I M T I −S M T I −S M T I −S
Iz, 1979 = Sz,11 Mz,1983 + Sz,10 Mz,1982 + Sz,9 Mz,1981
11
M T I −S M T I −S 1
+ Sz,12 Mz,1984,93 + wz,92 Mz,post Sz,12 + [(4) Fz,0 ] (Sz,a − Sz,a,pre )
e12−7 a=6
A.15
MT I M T I −S M T I −S M T I −S M T I −S
Iz, 1980 = Sz,11 Mz,1984,93 + Sz,10 Mz,1983 + Sz,9 Mz,1982 + Sz,8 Mz,1981
12
M T I −S
+ Mz,post wz,93 Sz,12 + wz,92 Sz,a
a=11
11
1
+ [(4) Fz,0 ] wz,93 (Sz,a − Sz,a,pre )
e12−7 a=6
10
1
+wz,92 (Sz,a − Sz,a,pre ) ,
e11−7 a=6
MT I M T I −S M T I −S M T I −S M T I −S
Iz,1981 = Sz,10 Mz,1984,93 + Sz,9 Mz,1983 + Sz,8 Mz, 1982 + Sz,7 Mz, 1981
12 12
M T I −S
+ Mz,post wz,93 Sz,a + wz,92 Sz,a
a=11 a=10
10
1
+ [(4) Fz,0 ] wz,93 (Sz,a − Sz,a,pre )
e11−7 a=6
9
1
+wz,92 (Sz,a − Sz,a,pre ) ,
e10−7 a=6
MT I M T I −S M T I −S M T I −S M T I −S
Iz,1982 = Sz,9 Mz,1984,93 + Sz,8 Mz,1983 + Sz,7 Mz,1982 + Sz,6 Mz,1981
12 12
M T I −S
+ Mz,post wz,93 Sz,a + wz,92 Sz,a
a=10 a=9
9
1
+ [(4) Fz,0 ] wz,93 (Sz,a − Sz,a,pre )
e10−7 a=6
8
1
+wz,92 (Sz,a − Sz,a,pre ) ,
e 9−7 a=6
MT I M T I −S M T I −S M T I −S
Iz,1983 = Sz,8 Mz,1984,93 + Sz,7 Mz,1983 + Sz,6 Mz, 1982
12 12
M T I −S
+ Mz,post wz,93 Sz,a + wz,92 Sz,a
a=9 a=8
8
1
+ [(4) Fz,0 ] wz,93 (Sz,a − Sz,a,pre )
e9−7 a=6
7
1
+wz,92 (Sz,a − Sz,a,pre ) ,
e8−7 a=6
A.16
MT I M T I −S M T I −S
Iz, 1984 = Sz,7 Mz,1984,93 + Sz,6 Mz,1983
12 12
M T I −S
+ Mz,post wz,93 Sz,a + wz,92 Sz,a
a=8 a=7
7
1
+ [(4) Fz,0 ] wz,93 (Sz,a − Sz,a,pre )
e 8−7 a=6
+wz,92 (Sz,6 − Sz,6,pre )} ,
12
MT I M T I −S M T I −S
Iz, 1985 = Sz,6 Mz,1984,93 + Mz,post wz,93 Sz,a + wz,92 + [(4) Fz,0 ] wz,93 (Sz,6 − Sz,6,pre )
a=7
MT I M T I −S
Iz,1986 = Mz,post
C.2.3 Dire Dawa
Table C.3: Dire Dawa – MTI Implementation
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Oromigna 1992 Dire Dawa 1-6 < 0.01 0.47
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: MTI = mother tongue instruction; MT = mother tongue.
The one time implementation of Oromigna in Dire Dawa is similar to that in Oromia, but with two key
diﬀerences. First, the language was only introduced in the ﬁrst six years of school in Dire Dawa. This will
again change the summation of the number of grades aﬀected, and delay the ﬁrst cohort to be introduced
to the reform by two years, from 1977 to 1979. This can be seen in the following magnitude calculations for
the on time starters in Dire Dawa:
5
M T I −DD
Mz,post = wz (6 − g ) · Fz,g if y ≥ 1984
g =0
5
M T I −DD
Mzy = wz (6 − g ) · Fz,g if 1979 ≤ y ≤ 1983 .
g =(1984−y )
M T I −DD
Mzy =0 if y ≤ 1978
The second diﬀerence is that only 47 percent of the population of Dire Dawa speaks the language being
introduced. This reduces the magnitude measure for each cohort through a smaller value of wz , but does not
impact the equations being used. The timing of MTI in Dire Dawa for on time starters for each birth year-
grade combination is shown in Appendix Table (C.5), denoted by the curled brackets. The MTI intensity
A.17
measure for cohorts in Dire Dawa is described by the following equations:
MT I
Iz,1973 = 0,
MT I M T I −DD
Iz,1974 = Sz,12 Mz, 1979 ,
MT I M T I −DD M T I −DD
Iz,1975 = Sz,12 Mz, 1980 + Sz,11 Mz,1979 ,
MT I M T I −DD M T I −DD M T I −DD
Iz, 1976 = Sz,12 Mz, 1981 + Sz,11 Mz,1980 + Sz,10 Mz,1979 ,
MT I M T I −DD M T I −DD M T I −DD M T I −DD
Iz,1977 = Sz,12 Mz,1982 + Sz,11 Mz,1981 + Sz,10 Mz,1980 + Sz,9 Mz, 1979 ,
MT I M T I −DD M T I −DD M T I −DD M T I −DD M T I −DD
Iz,1978 = Sz,12 Mz,1983 + Sz,11 Mz,1982 + Sz,10 Mz, 1981 + Sz,9 Mz,1980 + Sz,8 Mz,1979 ,
MT I M T I −DD M T I −DD M T I −DD M T I −DD
Iz, 1979 = Sz,11 Mz,1983 + Sz,10 Mz,1982 + Sz,9 Mz, 1981 + Sz,8 Mz,1980
11
M T I −DD M T I −DD 1
+ Sz,7 Mz,1979 + Mz,post Sz,12 + [(6) Fz,0 ] (Sz,a − Sz,a,pre ) ,
e12−7 a=6
MT I M T I −DD M T I −DD M T I −DD M T I −DD
Iz, 1980 = Sz,10 Mz,1983 + Sz,9 Mz,1982 + Sz,8 Mz, 1981 + Sz,7 Mz,1980
12 10
M T I −DD M T I −DD 1
+ Sz,6 Mz,1979 + Mz,post Sz,a + [(6) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=11
e11−7 a=6
MT I M T I −DD M T I −DD M T I −DD M T I −DD
Iz, 1981 = Sz,9 Mz,1983 + Sz,8 Mz,1982 + Sz,7 Mz, 1981 + Sz,6 Mz,1980
12 9
M T I −DD 1
+ Mz,post Sz,a + [(6) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=10
e10−7 a=6
12 8
MT I M T I −DD M T I −DD M T I −DD M T I −DD 1
Iz, 1982 = Sz,8 Mz,1983 + Sz,7 Mz,1982 + Sz,6 Mz, 1981 + Mz,post Sz,a + [(6) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=9
e9−7 a=6
12 7
MT I M T I −DD M T I −DD M T I −DD 1
Iz,1983 = Sz,7 Mz,1983 + Sz,6 Mz,1982 + Mz,post Sz,a + [(6) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=8
e 8−7 a=6
12
MT I M T I −DD M T I −DD
Iz, 1984 = Sz,6 Mz,1983 + Mz,post Sz,a + [(6) Fz,0 ] (Sz,6 − Sz,6,pre ) ,
a=7
MT I M T I −DD
Iz, 1985 = Mz,post .
A.18
C.2.4 Tigray
Table C.4: Tigray – MTI Implementation
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Tigrigna 1991 Tigray 1-8 0.93 0.95
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: MTI = mother tongue instruction; MT = mother tongue.
The initial calculations of the maximum impact of the MTI reform for on time starters are similar to that
of Oromia. In Tigray, the reform is implemented one year earlier due to the use of the Ge’ez script, but
both provinces introduced a single language for eight years of primary school. The timing can also be seen
in Appendix Table C.6.
7
M T I −T
Mz,post = wz (8 − g ) · Fz,g if y ≥ 1983
g =0
7
M T I −T
Mzy = wz (8 − g ) · Fz,g if 1976 ≤ y ≤ 1982 .
g =(1983−y )
M T I −T
Mzy =0 if y ≤ 1975
The birth year speciﬁc MTI intensity measure for Tigray again precedes the timing of Oromia by one year,
and is described by the following equations:
M T I −T
Iz, 1970 = 0,
M T I −T M T I −T
Iz,1971 = Sz,12 Mz, 1976 ,
M T I −T M T I −T M T I −T
Iz, 1972 = Sz,12 Mz, 1977 + Sz,11 Mz,1976 ,
M T I −T M T I −T M T I −T M T I −T
Iz, 1973 = Sz,12 Mz,1978 + Sz,11 Mz,1977 + Sz,10 Mz, 1976 ,
M T I −T M T I −T M T I −T M T I −T M T I −T
Iz, 1974 = Sz,12 Mz, 1979 + Sz,11 Mz,1978 + Sz,10 Mz, 1977 + Sz,9 Mz,1976 ,
M T I −T M T I −T M T I −T M T I −T M T I −T M T I −T
Iz,1975 = Sz,12 Mz, 1980 + Sz,11 Mz,1979 + Sz,10 Mz,1978 + Sz,9 Mz, 1977 + Sz,8 Mz,1976 ,
M T I −T M T I −T M T I −T M T I −T M T I −T M T I −T M T I −T
Iz,1976 = Sz,12 Mz, 1981 + Sz,11 Mz,1980 + Sz,10 Mz,1979 + Sz,9 Mz, 1978 + Sz,8 Mz,1977 + Sz,7 Mz, 1976 ,
A.19
MT I M T I −T M T I −T M T I −T M T I −T
Iz,1977 = Sz,12 Mz,1982 + Sz,11 Mz,1981 + Sz,10 Mz, 1980 + Sz,9 Mz,1979
M T I −T M T I −T M T I −T
+Sz,8 Mz,1978 + Sz,7 Mz, 1977 + Sz,6 Mz,1976 ,
M T I −T M T I −T M T I −T M T I −T M T I −T
Iz,1978 = Sz,11 Mz,1982 + Sz,10 Mz,1981 + Sz,9 Mz, 1980 + Sz,8 Mz,1979
11
M T I −T M T I −T M T I −T 1
+ Sz,7 Mz,1978 + Sz,6 Mz, 1977 + Mz,post Sz,12 + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
e12−7 a=6
M T I −T M T I −T M T I −T M T I −T M T I −T
Iz,1979 = Sz,10 Mz,1982 + Sz,9 Mz, 1981 + Sz,8 Mz,1980 + Sz,7 Mz, 1979
12 10
M T I −T M T I −T 1
+ Sz,6 Mz,1978 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=11
e11−7 a=6
M T I −T M T I −T M T I −T M T I −T M T I −T
Iz,1980 = Sz,9 Mz,1982 + Sz,8 Mz,1981 + Sz,7 Mz, 1980 + Sz,6 Mz,1979
12 9
M T I −T 1
+ Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=10
e10−7 a=6
12 8
M T I −T M T I −T M T I −T M T I −T M T I −T 1
Iz, 1981 = Sz,8 Mz, 1982 + Sz,7 Mz,1981 + Sz,6 Mz,1980 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=9
e 9−7 a=6
12 7
M T I −T M T I −T M T I −T M T I −T 1
Iz, 1982 = Sz,7 Mz, 1982 + Sz,6 Mz,1981 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,a − Sz,a,pre ) ,
a=8
e 8−7 a=6
12
M T I −T M T I −T M T I −T
Iz,1983 = Sz,6 Mz,1982 + Mz,post Sz,a + [(8) Fz,0 ] (Sz,6 − Sz,6,pre ) ,
a=7
M T I −T M T I −T
Iz,1984 = Mz,post .
A.20
Table C.5: MTI+Script Regions: Timing of FPE and MTI Reforms with On Time Entry, by Birth Year
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
1977 Born 1978 Born 1979 Born 1980 Born 1981 Born 1982 Born
1978 0 1979 0 1980 0 1981 0 1982 0 1983 0
1979 1 1980 1 1981 1 1982 1 1983 1 1984 1
1980 2 1981 2 1982 2 1983 2 1984 2 1985 2
1981 3 1982 3 1983 3 1984 3 1985 3 1986 3
1982 4 1983 4 1984 4 1985 4 1986 4 1987 4
1983 5 1984 5 1985 5 1986 5 1987 5 1988 5
1984 6 1985 6 1986 6 1987 6 1988 6 1989 6
1985 G1 7 1986 G1 7 1987 G1 7 1988 G1 7 1989 G1 7 1990 G1 7
1986 G2 8 1987 G2 8 1988 G2 8 1989 G2 8 1990 G2 8 1991 G2 8
1987 G3 9 1988 G3 9 1989 G3 9 1990 G3 9 1991 G3 9 1992 G3 9 [ { (MTI) }]
1988 G4 10 1989 G4 10 1990 G4 10 1991 G4 10 1992 G4 10 [ { (MTI) } ] 1993 G4 10 [ { (MTI) }]
1989 G5 11 1990 G5 11 1991 G5 11 1992 G5 11 { (MTI) } 1993 G5 11 { (MTI) } 1994 G5 11 { (MTI) }
1990 G6 12 1991 G6 12 1992 G6 12 { (MTI) } 1993 G6 12 { (MTI) } 1994 G6 12 { (MTI) } 1995 G6 12 { (FPE) }
1991 G7 13 1992 G7 13 (MTI) 1993 G7 13 (MTI) 1994 G7 13 (MTI) 1995 G7 13 (FPE) 1996 G7 13 (FPE)
1992 G8 14 (MTI) 1993 G8 14 (MTI) 1994 G8 14 (MTI) 1995 G8 14 (FPE) 1996 G8 14 (FPE) 1997 G8 14 (FPE)
1993 G9 15 1994 G9 15 1995 G9 15 FPE 1996 G9 15 FPE 1997 G9 15 FPE 1998 G9 15 FPE
1994 G10 16 1995 G10 16 FPE 1996 G10 16 FPE 1997 G10 16 FPE 1998 G10 16 FPE 1999 G10 16 FPE
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
1983 Born 1984 Born 1985 Born 1986 Born 1987 Born 1988 Born
A.21
1984 0 1985 0 1986 0 1987 0 1988 0 1989 0
1985 1 1986 1 1987 1 1988 1 1989 1 1990 1
1986 2 1987 2 1988 2 1989 2 1990 2 1991 2
1987 3 1988 3 1989 3 1990 3 1991 3 1992 3
1988 4 1989 4 1990 4 1991 4 1992 4 1993 4
1989 5 1990 5 1991 5 1992 5 1993 5 1994 5
1990 6 1991 6 1992 6 1993 6 1994 6 1995 6
1991 G1 7 1992 G1 7 [ { (MTI) } ] 1993 G1 7 [ { (MTI) } ] 1994 G1 7 [ { (MTI) } ] 1995 G1 7 [ { (FPE) } ] 1996 G1 7 [ { (FPE) } ]
1992 G2 8 [ { (MTI) } ] 1993 G2 8 [ { (MTI) } ] 1994 G2 8 [ { (MTI) } ] 1995 G2 8 [ { (FPE) } ] 1996 G2 8 [ { (FPE) } ] 1997 G2 8 [ { (FPE) } ]
1993 G3 9 [ { (MTI) } ] 1994 G3 9 [ { (MTI) } ] 1995 G3 9 [ { (FPE) } ] 1996 G3 9 [ { (FPE) } ] 1997 G3 9 [ { (FPE) } ] 1998 G3 9 [ { (FPE) } ]
1994 G4 10 [ { (MTI) } ] 1995 G4 10 [ { (FPE) } ] 1996 G4 10 [ { (FPE) } ] 1997 G4 10 [ { (FPE) } ] 1998 G4 10 [ { (FPE) } ] 1999 G4 10 [ { (FPE) } ]
1995 G5 11 { (FPE) } 1996 G5 11 { (FPE) } 1997 G5 11 { (FPE) } 1998 G5 11 { (FPE) } 1999 G5 11 { (FPE) } 2000 G5 11 { (FPE) }
1996 G6 12 { (FPE) } 1997 G6 12 { (FPE) } 1998 G6 12 { (FPE) } 1999 G6 12 { (FPE) } 2000 G6 12 { (FPE) } 2001 G6 12 { (FPE) }
1997 G7 13 (FPE) 1998 G7 13 (FPE) 1999 G7 13 (FPE) 2000 G7 13 (FPE) 2001 G7 13 (FPE) 2002 G7 13 (FPE)
1998 G8 14 (FPE) 1999 G8 14 (FPE) 2000 G8 14 (FPE) 2001 G8 14 (FPE) 2002 G8 14 (FPE) 2003 G8 14 (FPE)
1999 G9 15 FPE 2000 G9 15 FPE 2001 G9 15 FPE 2002 G9 15 FPE 2003 G9 15 FPE 2004 G9 15 FPE
2000 G10 16 FPE 2001 G10 16 FPE 2002 G10 16 FPE 2003 G10 16 FPE 2004 G10 16 FPE 2005 G10 16 FPE
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: MTI = mother tongue instruction; FPE = free primary education. Grades with ( ) mean that MTI is in place in Oromia, { } indicates MTI in Dire Dawa, and [ ]
indicates the initial introduction of MTI in the Southern Nations, Nationalities, and Peoples’ Region (SNNPR). Years with MTI mean that no FPE is in place; FPE within
brackets indicates that both FPE and MTI are in place in the speciﬁed region(s); and FPE without any type of brackets marks grades with FPE but no MTI.
Table C.6: Tigray: Timing of FPE and MTI Reforms with On Time Entry, by Birth Year
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
1977 Born 1978 Born 1979 Born 1980 Born 1981 Born 1982 Born
1978 0 1979 0 1980 0 1981 0 1982 0 1983 0
1979 1 1980 1 1981 1 1982 1 1983 1 1984 1
1980 2 1981 2 1982 2 1983 2 1984 2 1985 2
1981 3 1982 3 1983 3 1984 3 1985 3 1986 3
1982 4 1983 4 1984 4 1985 4 1986 4 1987 4
1983 5 1984 5 1985 5 1986 5 1987 5 1988 5
1984 6 1985 6 1986 6 1987 6 1988 6 1989 6
1985 G1 7 1986 G1 7 1987 G1 7 1988 G1 7 1989 G1 7 1990 G1 7
1986 G2 8 1987 G2 8 1988 G2 8 1989 G2 8 1990 G2 8 1991 G2 8 MTI
1987 G3 9 1988 G3 9 1989 G3 9 1990 G3 9 1991 G3 9 MTI 1992 G3 9 MTI
1988 G4 10 1989 G4 10 1990 G4 10 1991 G4 10 MTI 1992 G4 10 MTI 1993 G4 10 MTI
1989 G5 11 1990 G5 11 1991 G5 11 MTI 1992 G5 11 MTI 1993 G5 11 MTI 1994 G5 11 MTI
1990 G6 12 1991 G6 12 MTI 1992 G6 12 MTI 1993 G6 12 MTI 1994 G6 12 MTI 1995 G6 12 (FPE)
1991 G7 13 MTI 1992 G7 13 MTI 1993 G7 13 MTI 1994 G7 13 MTI 1995 G7 13 (FPE) 1996 G7 13 (FPE)
1992 G8 14 MTI 1993 G8 14 MTI 1994 G8 14 MTI 1995 G8 14 (FPE) 1996 G8 14 (FPE) 1997 G8 14 (FPE)
1993 G9 15 1994 G9 15 1995 G9 15 FPE 1996 G9 15 FPE 1997 G9 15 FPE 1998 G9 15 FPE
1994 G10 16 1995 G10 16 FPE 1996 G10 16 FPE 1997 G10 16 FPE 1998 G10 16 FPE 1999 G10 16 FPE
Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform Birth Reform
Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status Year Grade Age Status
A.22
1983 Born 1984 Born 1985 Born 1986 Born 1987 Born 1988 Born
1984 0 1985 0 1986 0 1987 0 1988 0 1989 0
1985 1 1986 1 1987 1 1988 1 1989 1 1990 1
1986 2 1987 2 1988 2 1989 2 1990 2 1991 2
1987 3 1988 3 1989 3 1990 3 1991 3 1992 3
1988 4 1989 4 1990 4 1991 4 1992 4 1993 4
1989 5 1990 5 1991 5 1992 5 1993 5 1994 5
1990 6 1991 6 1992 6 1993 6 1994 6 1995 6
1991 G1 7 MTI 1992 G1 7 MTI 1993 G1 7 MTI 1994 G1 7 MTI 1995 G1 7 (FPE) 1996 G1 7 (FPE)
1992 G2 8 MTI 1993 G2 8 MTI 1994 G2 8 MTI 1995 G2 8 (FPE) 1996 G2 8 (FPE) 1997 G2 8 (FPE)
1993 G3 9 MTI 1994 G3 9 MTI 1995 G3 9 (FPE) 1996 G3 9 (FPE) 1997 G3 9 (FPE) 1998 G3 9 (FPE)
1994 G4 10 MTI 1995 G4 10 (FPE) 1996 G4 10 (FPE) 1997 G4 10 (FPE) 1998 G4 10 (FPE) 1999 G4 10 (FPE)
1995 G5 11 (FPE) 1996 G5 11 (FPE) 1997 G5 11 (FPE) 1998 G5 11 (FPE) 1999 G5 11 (FPE) 2000 G5 11 (FPE)
1996 G6 12 (FPE) 1997 G6 12 (FPE) 1998 G6 12 (FPE) 1999 G6 12 (FPE) 2000 G6 12 (FPE) 2001 G6 12 (FPE)
1997 G7 13 (FPE) 1998 G7 13 (FPE) 1999 G7 13 (FPE) 2000 G7 13 (FPE) 2001 G7 13 (FPE) 2002 G7 13 (FPE)
1998 G8 14 (FPE) 1999 G8 14 (FPE) 2000 G8 14 (FPE) 2001 G8 14 (FPE) 2002 G8 14 (FPE) 2003 G8 14 (FPE)
1999 G9 15 FPE 2000 G9 15 FPE 2001 G9 15 FPE 2002 G9 15 FPE 2003 G9 15 FPE 2004 G9 15 FPE
2000 G10 16 FPE 2001 G10 16 FPE 2002 G10 16 FPE 2003 G10 16 FPE 2004 G10 16 FPE 2005 G10 16 FPE
Source: Author’s analysis based on information from Boothe and Walker (1997) and Zenebe Gebre (2014).
Note: FPE = free primary education; MTI = mother tongue instruction. Years with (FPE) indicate grade-year combinations in which both MTI and FPE are in place.
C.3 National Estimates: Accounting for FPE and MTI
Table C.7: National Estimates of Eﬀect of Years of Schooling on
Number of Children Born - Census + DHS
Number of Years of Number of Number of
Children Born Schooling Children Born Children Born
(OLS) (First Stage) (Reduced Form) (2SLS)
(1) (2) (3) (4)
Years of Schoolingizy -0.149
(0.015)
[0.000]
Add’l Years of Free 0.115 -0.051
FPE
Schooling Izy (0.044) (0.016)
[0.011] [0.002]
Add’l Year of MTI 0.183 -0.007
No Script Change (0.070) (0.018)
M T I −T
Izy [0.012] [0.687]
Add’l Year of MTI -0.119 0.064
with Script Change (0.052) (0.028)
MT I
Izy [0.025] [0.028]
FPE M T I −T
Izy × Izy -0.008 -0.001
(0.006) (0.002)
[0.169] [0.630]
FPE MT I
Izy × Izy 0.001 -0.004
(0.005) (0.002)
[0.767] [0.028]
Years of Schoolingizy -0.273
(0.109)
[0.012]
First Stage F-Statistic 14.09 14.09
Number of Clusters 60 60 60 60
N 205,141 205,141 205,141 205,141
Source: Author’s analysis based on data from the Ethiopian census of 2007 and the Demographic and Health
Survey (DHS) in years 2005, 2011, and 2016.
Note: The dependent variable is either years of schooling in column 2 or number of births in the other three columns.
Years of Schoolingizy is the reported years of schooling from the data; Years of Schoolingizy is the predicted level
of schooling, instrumented with the free primary education (FPE) intensity measure Izy FPE , two mother tongue
instruction (MTI) intensity measures Izy MTI-T and I MTI , which denote the measures for MTI regions without and
zy
with script change, respectively, and the interactions for regions in which two interventions occurred. All samples
include women in birth cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-
speciﬁc linear trends, and a cubic for age. Standard errors are clustered at the zone level and shown in parentheses;
p-values are shown in square brackets.
A.23
Table C.8: National Estimates of Eﬀect of Years of Schooling on
Number of Children Born - Census
Number of Years of Number of Number of
Children Born Schooling Children Born Children Born
(OLS) (First Stage) (Reduced Form) (2SLS)
(1) (2) (3) (4)
Years of Schoolingizy -0.120
(0.012)
[0.000]
Add’l Years of Free 0.120 -0.060
FPE
Schooling Izy (0.037) (0.017)
[0.002] [0.001]
Add’l Year of MTI 0.201 -0.011
No Script Change (0.052) (0.011)
M T I −T
Izy [0.000] [0.321]
Add’l Year of MTI -0.081 -0.001
with Script Change (0.024) (0.011)
MT I
Izy [0.002] [0.894]
FPE M T I −T
Izy × Izy -0.003 0.005
(0.003) (0.001)
[0.318] [0.001]
FPE MT I
Izy × Izy 0.005 -0.008
(0.003) (0.001)
[0.062] [0.000]
Years of Schoolingizy -0.297
(0.144)
[0.039]
First Stage F-Statistic 9.72 9.72
Number of Clusters 60 60 60 60
N 180,243 180,243 180,243 180,243
Source: Author’s analysis based on data from the Ethiopian census of 2007.
Note: The dependent variable is either years of schooling in column 2 or number of births in the other three columns.
Years of Schoolingizy is the reported years of schooling from the data; Years of Schoolingizy is the predicted level
of schooling, instrumented with the free primary education (FPE) intensity measure Izy FPE , two mother tongue
instruction (MTI) intensity measures IzyMTI-T MTI
and Izy , which denote the measures for MTI regions without and
with script change, respectively, and the interactions for regions in which two interventions occurred. All samples
include women in birth cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects and
zone-speciﬁc linear trends. Standard errors are clustered at the zone level and shown in parentheses; p-values are
shown in square brackets.
A.24
Table C.9: National Estimates of Eﬀect of Years of Schooling on
Number of Children Born - DHS
Number of Years of Number of Number of
Children Born Schooling Children Born Children Born
(OLS) (First Stage) (Reduced Form) (2SLS)
(1) (2) (3) (4)
Years of Schoolingizy -0.162
(0.016)
[0.000]
Add’l Years of Free 0.115 -0.057
FPE
Schooling Izy (0.053) (0.021)
[0.035] [0.009]
Add’l Year of MTI 0.175 -0.012
No Script Change (0.087) (0.025)
M T I −T
Izy [0.049] [0.636]
Add’l Year of MTI -0.128 0.092
with Script Change (0.073) (0.041)
MT I
Izy [0.085] [0.031]
FPE M T I −T
Izy × Izy -0.010 -0.003
(0.008) (0.002)
[0.239] [0.204]
FPE MT I
Izy × Izy -0.001 -0.002
(0.007) (0.003)
[0.930] [0.342]
Years of Schoolingizy -0.365
(0.147)
[0.001]
First Stage F-Statistic 6.91 6.91
Number of Clusters 58 58 58 58
N 24,898 24,898 24,898 24,898
Source: Author’s analysis based on data from the Demographic and Health Survey (DHS) in years 2005, 2011, and
2016.
Note: The dependent variable is either years of schooling in column 2 or number of births in the other three columns.
Years of Schoolingizy is the reported years of schooling from the data; Years of Schoolingizy is the predicted level
of schooling, instrumented with the free primary education (FPE) intensity measure Izy FPE , two mother tongue
instruction (MTI) intensity measures IzyMTI-T MTI
and Izy , which denote the measures for MTI regions without and
with script change, respectively, and the interactions for regions in which two interventions occurred. All samples
include women in birth cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-
speciﬁc linear trends, and a cubic for age. Standard errors are clustered at the zone level and shown in parentheses;
p-values are shown in square brackets.
A.25
Table C.10: National Estimates of Eﬀect of Years of Schooling on Knowledge and Health
Read about Know about BMI Height Acceptable Reasons for Use Modern Use Hidden
Literacy Fam. Planning Fam. Planning (z-score) (z-score) Domestic Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
Years of Schoolingizy 0.122 0.053 0.023 0.206 0.132 -0.254 -0.006 -0.011
(0.019) (0.018) (0.020) (0.096) (0.104) (0.124) (0.030) (0.025)
[0.000] [0.004] [0.255] [0.031] [0.204] [0.041] [0.830] [0.655]
Mean of Dependent 0.115 0.046 0.935 -0.020 -0.168 2.681 0.162 0.127
(Pre-Reform Cohorts)
First Stage F-Statistic 7.73 6.91 6.91 6.13 5.01 8.42 6.91 6.91
A.26
N 24,480 24,885 24,898 19,491 19,879 24,052 24,898 24,898
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: BMI = body mass index. The dependent variable is described at the top of each of the eight columns. In columns 1–3, 7, and 8 it is an indicator that equals 1 if true; in
columns 4 and 5 it is a standardized value of the described outcome; and in column 6 it is the count from 0 to 5 of acceptable reasons for domestic violence (going out without
permission, neglecting children, arguing with husband, refusing sex, burning food). Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary
education (FPE) intensity measure Izy FPE , two mother tongue instruction (MTI) intensity measures I MTI-T and I MTI , which denote the measures for MTI regions without
zy zy
and with script change, respectively, and the interactions for regions in which two interventions occurred. All samples include women in birth cohorts from 1970 to 1988. All
regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique regression. Standard errors are clustered at
the zone level and shown in parentheses; p-values are shown in square brackets.
Table C.11: National Estimates of Eﬀect of Years of Schooling on
Labor Market Outcomes and Fertility Preference
Sector of Work
Skilled / Service / Agriculture / Ideal Number
Working Professional Sales Unskilled Manual of Children
(1) (2) (3) (4) (5)
Years of Schoolingizy 0.010 0.033 0.020 -0.034 -0.923
(0.042) (0.017) (0.025) (0.034) (0.658)
[0.809] [0.053] [0.433] [0.319] [0.160]
Mean of Dependent 0.342 0.072 0.126 0.279 7.623
(Pre-Reform Cohorts)
First Stage F-Statistic 6.93 7.91 7.91 7.91 6.60
N 24,882 24,607 24,607 24,607 24,649
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: The dependent variable is described at the top of each of the ﬁve columns. In columns 1–4 it is an indicator that equals
1 if true, and in column 5 it is the ideal number of children. Skilled/Professional jobs include professional, clerical, and skilled
manual job groups; the other categories exactly describe the occupation groups included. Ideal number of children is censored
at 20; no women in the Demographic and Health Survey report having more than 18 children, and non-numerical responses
are assigned the maximum value. Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary
FPE , two mother tongue instruction (MTI) intensity measures I MTI-T and I MTI , which
education (FPE) intensity measure Izy zy zy
denote the measures for MTI regions without and with script change, respectively, and the interactions for regions in which
two interventions occurred. All samples include women in birth cohorts from 1970 to 1988. All regressions include birth year
and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique regression, and the
second-stage estimate in column 5 is generated using a tobit model. Standard errors are clustered at the zone level and shown
in parentheses; p-values are shown in square brackets.
A.27
Table C.12: Eﬀect of Years of Schooling on Sector of Employment:
Job Diﬀerent than Husband
Sector of Current Work
Agriculture
Skilled / Service / Unskilled
Professional / Sales Manual
(1) (2) (3)
A. Non-MTI Regions Only
Years of Schoolingizy 0.058 0.062 -0.027
(0.023) (0.034) (0.019)
[0.011] [0.071] [0.250]
First Stage F-Statistic 12.48 12.48 12.48
N 11,870 11,870 11,870
B. National Estimates (FPE + MTI)
Years of Schoolingizy 0.035 0.014 0.004
(0.015) (0.024) (0.029)
[0.019] [0.569] [0.879]
First Stage F-Statistic 8.22 8.22 8.22
N 21,242 21,242 21,242
Source: Author’s analysis based on data from the Ethiopian Demographic and
Health Survey (in years 2005, 2011, and 2016).
Note: MTI = mother tongue instruction; FPE = free primary education. The
dependent variable is an indicator that equals 1 if employed in the denoted sec-
tor and 0 otherwise. In panel A, Years of Schoolingizy is the predicted level of
schooling, instrumented with only the FPE intensity measure, Izy FPE ; in panel B
additional instruments include two MTI intensity measures, Izy MTI-T MTI ,
and Izy
which denote the measures for MTI regions without and with script change, re-
spectively, and the interactions for regions in which two interventions occurred.
All samples include women in birth cohorts from 1970 to 1988 working in diﬀer-
ent jobs than their husbands. All regressions include birth year and zone ﬁxed
eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a
unique regression. Standard errors are clustered at the zone level and shown in
parentheses; p-values are shown in square brackets.
A.28
Table C.13: National Estimates of Eﬀect of Wife’s Schooling on Husband’s Characteristics, Married Women Only
Wife’s Years Husband’s Occupation
of Schooling Husband’s Agriculture Husband
[ First Stage Husband’s Years of Skilled / Service / Unskilled Wants More
- Married Only ] Age Schooling Professional / Sales Manual Children
(1) (2) (3) (4) (5) (6) (7)
Add’l Years of Free 0.009
FPE
Schooling Izy (0.063)
[0.891]
Add’l Year of MTI 0.066
No Script Change (0.058)
M T I −T
Izy [0.261]
Add’l Year of MTI -0.212
with Script Change (0.080)
MT I
Izy [0.010]
FPE M T I −T
Izy × Izy 0.001
(0.010)
[0.947]
FPE MT I
Izy × Izy 0.009
A.29
(0.007)
[0.207]
Years of Schoolingizy 1.009 1.116 0.026 0.099 -0.130 -0.029
(0.884) (0.372) (0.028) (0.031) (0.045) (0.052)
[0.254] [0.003] [0.350] [0.001] [0.004] [0.570]
Mean of Dependant 1.15 49.48 2.28 0.083 0.081 0.792 0.370
(Pre-Reform Cohorts)
First Stage F-Statistic 4.69 6.35 7.10 6.40 6.40 6.40 5.89
N 20,959 18,174 19,784 19,814 19,814 19,814 12,761
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: MTI = mother tongue instruction. The dependent variable is described at the top of each of the six columns. The ﬁrst-stage estimate of the
eﬀect of the reforms on years of schooling for married women is shown in column 1. The dependent variables in columns 4–7 are indicator variables
that equal 1 if true. Skilled/Professional jobs include professional, clerical, and skilled manual job groups; the other categories exactly describe the
occupation groups included. Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary education (FPE) intensity
FPE , two mother tongue instruction (MTI) intensity measures I MTI-T and I MTI , which denote the measures for MTI regions without and
measure Izy zy zy
with script change, respectively, and the interactions for regions in which two interventions occurred. All samples include married women in birth
cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is
from a unique regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
Table C.14: National Estimates of Eﬀect of Schooling on
Beliefs Regarding Domestic Violence, Married Women Only
Acceptable Reasons for Beating justiﬁed if wife: Goes Out Neglects Argues with
Domestic Violence (of 5) w/out Permission Children Husband Refuses Sex Burns Food
(1) (2) (3) (4) (5) (6)
Years of Schoolingizy -0.399 -0.080 -0.020 -0.117 -0.089 -0.087
(0.171) (0.044) (0.052) (0.055) (0.048) (0.045)
[0.020] [0.071] [0.698] [0.033] [0.106] [0.051]
Mean of Dependent 2.80 0.582 0.617 0.553 0.485 0.574
(Pre-Reform Cohorts)
A.30
First Stage F-Statistic 7.94 6.64 6.82 6.31 7.92 6.85
N 17,658 18,067 18,065 18,027 17,873 18,072
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: The dependent variable in column 1 is the count from 0 to 5 of acceptable reasons for domestic violence, and the dependent variables in columns 2–6 are indicators
that equal 1 if the statement is believed to be true and 0 otherwise. The sample includes all married women born between 1970 and 1988. A 2SLS model is estimated where
FPE , two mother tongue instruction (MTI)
Years of Schoolingizy is the predicted level of schooling, instrumented with the free primary education (FPE) intensity measure Izy
MTI-T
intensity measures Izy MTI
and Izy , which denote the measures for MTI regions without and with script change, respectively, and the interactions for regions in which two
interventions occurred. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique regression.
Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
Table C.15: National Estimates of Eﬀect of Schooling on
Beliefs Regarding Women’s Empowerment, Married Women Only
Travel to Visit Personal Large Household
Family / Friends Healthcare Purchases
(1) (2) (3)
A. Wife should at least have a say in decision
Years of Schoolingizy -0.056 -0.066 -0.065
(0.043) (0.040) (0.035)
[0.194] [0.097] [0.066]
Mean of Dependent 0.783 0.721 0.650
(Pre-Reform Cohorts)
First Stage F-Statistic 6.60 6.59 6.59
N 18,138 18,139 18,139
B. Wife should be able to make decision alone
Years of Schoolingizy -0.064 -0.029 -0.041
(0.039) (0.030) (0.027)
[0.102] [0.334] [0.118]
Mean of Dependent Variable 0.163 0.161 0.128
(Pre-Reform Cohorts)
First Stage F-Statistic 6.60 6.59 6.59
N 18,138 18,139 18,139
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years
2005, 2011, and 2016).
Note: The dependent variable in each column is an indicator that equals 1 if the statement is believed to
be true and 0 otherwise. The sample includes all married women born between 1970 and 1988. A 2SLS
model is estimated where Years of Schoolingizy is the predicted level of schooling, instrumented with the
free primary education (FPE) intensity measure Izy FPE , two mother tongue instruction (MTI) intensity
MTI-T and I MTI , which denote the measures for MTI regions without and with script change,
measures Izy zy
respectively, and the interactions for regions in which two interventions occurred. All regressions include
birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a
unique regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are
shown in square brackets.
A.31
C.4 Combined Instrument and Reduced Form Estimates, by Age
The previous subsection, Appendix Section C.3, estimates the eﬀect of the FPE and MTI reforms on schooling
using ﬁve instruments. However, reduced form analysis across diﬀerent ages would become cumbersome
with the large number of variables; combining the impact of the two reforms would allow for easier graphical
representation of the impact of the reforms across diﬀerent ages. Cutting the data by age is only possible
with the inclusion of both reforms, this allows for the use of the national sample and additional variation
from the timing of the MTI implementation.
Creating a single instrument that measures exposure to both the FPE and MTI reforms, and predicts
changes in schooling, requires a simplifying assumption from the ﬁve-instrument model used in Appendix
Section C.3. The necessary baseline assumption is that the magnitude of the eﬀect of an additional year
of FPE on schooling is the same as each additional year of MTI, in either a positive (no script change) or
negative (with script change) direction. For simplicity, it is also assumes that there is no interaction between
the two reforms, only one of the six interaction terms in the ﬁrst stage of Appendix Tables C.7 to C.9 has a
p-value below 0.15. These assumptions yield a combined intensity measure that can be expressed using the
following equation:
FPE M T I −T MT I
∆Izy = Izy + Izy − Izy . (C.1)
M T I −T
Izy is zero for all non-Tigray regions, and is shown to be positive in Appendix Tables C.7 to C.9.
MT I
Similarly, Izy , is only non-zero for MTI regions other than Tigray (as described in Appendix Section
C.2), and shown to be negatively associated with schooling in Appendix Tables C.7 to C.9.24 Additionally,
the diﬀerences in the estimated eﬀect of each intensity measure can also be statistically tested. Using the
estimates from the DHS in Appendix Table C.9, the survey that contains all age speciﬁc outcomes of interest,
the null hypothesis that each combination of the three intensity measures has an equal eﬀect on schooling
is tested in Appendix Table C.16. Across the four tests, the null hypothesis of equality cannot be rejected,
and no p-value is larger than 0.5.
In addition to testing the equality of each ﬁrst stage coeﬃcient, the combined intensity measure (∆Izy )
is also used to re-estimate the eﬀect of years of schooling on number of children born, and all outcomes from
Appendix Tables C.10, C.11, and C.13. The estimates using the combined intensity measure can be found
in Appendix Tables C.17 to C.20; in each case, the combined intensity measure yields estimates similar to
the model using ﬁve instruments. Most importantly, the estimated eﬀect of an additional year of schooling
24 The additive portion of ∆I
zy does not double count years in which both FPE and MTI are available in the region. Both
reforms making one additional year of schooling available can never increase schooling by more than a single year.
A.32
on number of births is within 44-thousandths (-0.321 versus -0.365), a nearly identical result. The same
pattern is seen across nearly every estimate in the tables examining health and knowledge, the labor market,
and the marriage market. The conclusions generated by the two diﬀerent instrument strategies are both
quantitatively and qualitatively consistent. The most signiﬁcant diﬀerence is that the ﬁrst stage F-statistic
for the combined measure is more than 2.5 times larger.
The single variable that is able to capture the variation in the introduction of both the FPE and MTI
reforms can be used to estimate a set of reduced form models to quantify the impact of the reforms on central
outcomes related to fertility at speciﬁc ages. For example, the reduced form eﬀect of the reforms on number
of births at each age is shown in Appendix Figure C.1. The downward sloping black line is the coeﬃcient
estimate on the combined estimator, the 90 and 95 percent conﬁdence intervals are shown with the dashed
and solid gray lines, respectively. At the younger ages, 15 through 19, there is no eﬀect of schooling on
number of births. At the age of 20, the eﬀect is slightly larger, and becomes statistically signiﬁcant at the 90
percent conﬁdence level. The estimated eﬀect becomes increasingly negative and statistically signiﬁcant at
the 95 percent conﬁdence level at the age of 23, increasing in magnitude by 58 percent relative to the eﬀect
at 22. The eﬀect continues to become increasingly negative through the age of 29, and remains statistically
signiﬁcant.
The combination of the low levels of schooling attainment and the reduction in fertility manifesting itself
in the women’s early twenties makes it unlikely that any type of incarceration eﬀect, women physically
being in the classroom, is aﬀecting the results in the paper. Furthermore, Black et al. (2008) and Geruso
and Royer (2018) also ﬁnd that reductions in teenage fertility tend to be replaced by increases in additional
births at later ages. Although the eﬀects seen in Appendix Figure C.1 are not for completed fertility, the later
introduction of the eﬀect and the continued growth in magnitude through age 29 makes the retrenchment
seen in the teenage fertility literature less likely to occur in this setting.
The same type of reduced form model is then used to examine the eﬀect of the reform on the timing of
ﬁrst birth, intercourse, and marriage in Appendix Figure C.2. The coeﬃcients shown in the ﬁgure are from
reduced form estimates using the combined intensity measure. The eﬀect of an additional year of FPE and
MTI without script change on the timing of a woman’s ﬁrst birth (black bars), ﬁrst marriage (white), and
sexual intercourse (dotted) are shown. The ﬁrst statistically signiﬁcant changes are the reductions in the
likelihood of ﬁrst marriage and intercourse by the age of 21. The magnitude of these changes become larger at
the age of 22; evaluating the magnitudes at the post-reform average of the joint intensity measure suggests
reductions in the likelihood of ﬁrst marriage and intercourse of 7.2 percentage points and 4.8 percentage
points, respectively. The eﬀect of the reform on these two outcomes remains statistically signiﬁcant through
the age of 24, before a substantial reduction in magnitude and loss of statistical signiﬁcance beginning at
A.33
the age of 25.
As would be expected, the impact of the reform on the timing of ﬁrst birth lags the eﬀect on ﬁrst marriage
and intercourse. This timing suggests that the reform’s impact on the marriage decision is leading to an initial
delay in women’s fertility. Again, using the post-reform average of the joint intensity measure, the estimated
eﬀect at the age of 23 suggests that the reform reduced the likelihood of ﬁrst birth by 8.5 percentage points.
In fact, the reduction in the likelihood of ﬁrst birth by the age of 23 coincides with the large reduction
in number of births by this age seen in Appendix Figure C.1. The eﬀect on the likelihood of ﬁrst birth
remains statistically signiﬁcant through the age of 25, again one year later than the eﬀect on ﬁrst marriage
and intercourse. While these changes help explain some of the reduction in fertility, the magnitude of the
reductions in number of births continues to grow beyond the age of 25 suggesting post-marriage decisions
are likely changing, as well. The values of the reduced form coeﬃcient estimates in Appendix Figures C.1
and C.2 can be found in Appendix Table C.21.
A.34
Table C.16: National Estimates of Eﬀect of Years of Schooling on Number of Children Born - DHS
Years of Schooling
(First Stage)
(1)
Add’l Years of Free 0.115
FPE
Schooling Izy (0.053) F-Test F-Statistic P-Value
[0.035]
FPE M T I −T
Izy = Izy 0.46 [0.501]
Add’l Year of MTI 0.175
No Script Change (0.087)
M T I −T FPE MT I
Izy [0.049] Izy = (−1) × Izy 0.01 [0.905]
Add’l Year of MTI -0.128
M T I −T MT I
with Script Change (0.073) Izy = (−1) × Izy 0.12 [0.731]
MT I
Izy [0.085]
FPE M T I −T FPE M T I −T
Izy × Izy -0.010 Izy = Izy 0.23 [0.794]
MT I
(0.008) = (−1) × Izy
[0.239]
FPE MT I
Izy × Izy -0.001
(0.007)
[0.930]
N 24,898
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (DHS)
in years 2005, 2011, and 2016.
Note: MTI = mother tongue instruction. Column 1 is reproduced from Table C.9 in this supple-
mentary online appendix. The dependent variable is years of schooling. The sample includes women
in birth cohorts from 1970 to 1988. The regression includes birth year and zone ﬁxed eﬀects, zone-
speciﬁc linear trends, and a cubic for age when multiple survey waves are included. Standard errors
are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
A.35
Table C.17: National Estimates of Eﬀect of Years of Schooling on
Number of Children Born - DHS - Combined Instrument
Years of Number of Number of
Schooling Children Born Children Born
(First Stage) (Reduced Form) (2SLS)
(1) (2) (3)
Add’l Years of FPE or MTI 0.115 -0.037
w/out Script Change (0.026) (0.019)
(∆Izy ) [0.000] [0.019]
Years of Schoolingizy -0.321
(0.156)
[0.040]
First Stage F-Statistic 19.56 19.56
Number of Clusters 58 58 58
N 24,898 24,898 24,898
Source: Author’s analysis based on data from the Demographic and Health Survey (DHS)
in years 2005, 2011, and 2016.
Note: The dependent variable is years of schooling in column 1 and is number of births
in the other two columns. Years of Schoolingizy is the predicted number of years of
schooling, instrumented with the combined intensity measure, ∆Izy . All samples include
women in birth cohorts from 1970 to 1988. All regressions include birth year and zone
ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age when multiple survey waves
are included. Standard errors are clustered at the zone level and shown in parentheses;
p-values are shown in square brackets.
A.36
Table C.18: National Estimates of Eﬀect of Years of Schooling on Knowledge and Health - Combined Instrument
Read about Know about BMI Height Acceptable Reasons for Use Modern Use Hidden
Literate Fam. Planning Fam. Planning (z-score) (z-score) Domestic Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
Years of Schoolingizy 0.121 0.047 0.023 0.234 0.031 -0.310 -0.009 -0.017
(0.019) (0.021) (0.022) (0.132) (0.142) (0.121) (0.031) (0.027)
[0.000] [0.027] [0.287] [0.076] [0.830] [0.010] [0.773] [0.535]
First Stage F-Statistic 20.27 19.55 19.56 9.70 8.22 19.45 19.56 19.56
A.37
N 24,480 24,885 24,898 19,491 19,879 24,052 24,898 24,898
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: BMI = body mass index. The dependent variable is described at the top of each of the eight columns. In columns 1–3, 7, and 8 it is an indicator that equals 1 if true; in
columns 4 and 5 it is a standardized value of the described outcome; and in column 6 it is the count from 0 to 5 of acceptable reasons for domestic violence (going out without
permission, neglecting children, arguing with husband, refusing sex, burning food). Years of Schoolingizy is the predicted level of schooling, instrumented with the combined
intensity measure, ∆Izy . All samples include women in birth cohorts from 1970 to 1988. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends,
and a cubic for age. Each estimate is from a unique regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
Table C.19: National Estimates of Eﬀect of Years of Schooling on
Labor Market Outcomes and Fertility Preference - Combined Instrument
Sector of Work
Skilled / Service Agriculture / Ideal Number
Working Professional / Sales Unskilled Manual of Children
(1) (2) (3) (4) (5)
Years of Schoolingizy 0.017 0.044 0.016 -0.039 -0.902
(0.043) (0.015) (0.028) (0.033) (0.493)
[0.692] [0.004] [0.573] [0.241] [0.068]
First Stage F-Statistic 19.68 20.64 20.64 20.64 19.37
N 24,882 24,607 24,607 24,607 24,649
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: The dependent variable is described at the top of each of the ﬁve columns. In columns 1–4 it is an indicator that equals
1 if true, and in column 5 it is the ideal number of children. Skilled/Professional jobs include professional, clerical, and skilled
manual job groups; the other categories exactly describe the occupation groups included. Ideal number of children is censored
at 20; no women in the Demographic and Health Survey report having more than 18 children, and non-numerical responses
are assigned the maximum value. Years of Schoolingizy is the predicted level of schooling, instrumented with the combined
intensity measure, ∆Izy . All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for
age. Each estimate is from a unique regression, and the second-stage estimate in column 5 is generated using a tobit model.
Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
A.38
Table C.20: National Estimates of Eﬀect of Wife’s Exposure to Reforms on Husband’s Characteristics, Married Women Only - Combined Instrument
Wife’s Years Husband’s Occupation
of Schooling Husband’s Agriculture Husband
[ First Stage Husband’s Years of Skilled / Service / Unskilled Wants More
- Married Only ] Age Schooling Professional / Sales Manual Children
(1) (2) (3) (4) (5) (6) (7)
Add’l Years of FPE or MTI 0.097
w/out Script Change (0.041)
(∆Izy ) [0.020]
Years of Schoolingizy -0.017 0.869 -0.039 0.121 -0.129 -0.069
(0.976) (0.475) (0.044) (0.047) (0.043) (0.078)
A.39
[0.986] [0.067] [0.375] [0.010] [0.003] [0.374]
First Stage F-Statistic 5.70 5.70 6.94 6.97 6.97 6.97 5.86
N 18,174 18,174 17,998 19,814 19,814 19,814 12,761
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: FPE = free primary education; MTI = mother tongue instruction. The dependent variable is described at the top of each of the six columns. The ﬁrst-stage estimate
of the eﬀect of the reforms on years of schooling for married women is shown in column 1. The dependent variables in columns 4–7 are indicator variables that equal 1 if true.
Skilled/Professional jobs include professional, clerical, and skilled manual job groups; the other categories exactly describe the occupation groups included. Years of Schoolingizy
is the predicted level of schooling, instrumented with the combined intensity measure, ∆Izy . All samples include married women in birth cohorts from 1970 to 1988. All regressions
include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique regression. Standard errors are clustered at the zone
level and shown in parentheses; p-values are shown in square brackets.
Table C.21: Eﬀect of Reform on Number of Births and Likelihood of First Birth, Marriage, and Intercourse, by Age
Coeﬃcient Estimates from Appendix Figures C.1 and C.2
Age: 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15)
A. Number of Births
Add’l Years of FPE or MTI -0.006 -0.005 -0.000 -0.009 -0.008 -0.019 -0.023 -0.031 -0.049 -0.056 -0.059 -0.078 -0.085 -0.084 -0.106
w/out Script Change (0.005) (0.005) (0.008) (0.009) (0.010) (0.012) (0.012) (0.016) (0.019) (0.020) (0.025) (0.031) (0.037) (0.043) (0.044)
(∆Izy ) [0.228] [0.347] [0.987] [0.327] [0.445] [0.103] [0.058] [0.054] [0.011] [0.007] [0.023] [0.014] [0.024] [0.055] [0.019]
N 24,898 24,616 24,058 23,221 22,791 21,842 21,514 20,582 19,710 18,919 17,005 16,106 14,949 13,063 12,189
B. First Birth
Add’l Years of FPE or MTI -0.005 -0.004 -0.003 -0.003 -0.004 -0.006 -0.006 -0.006 -0.013 -0.011 -0.011 -0.007 -0.006 -0.006 -0.002
w/out Script Change (0.004) (0.003) (0.005) (0.006) (0.005) (0.005) (0.006) (0.005) (0.006) (0.005) (0.005) (0.004) (0.005) (0.005) (0.006)
(∆Izy ) [0.204] [0.166] [0.500] [0.550] [0.368] [0.270] [0.257] [0.269] [0.019] [0.024] [0.018] [0.127] [0.263] [0.245] [0.680]
N 24,898 24,616 24,058 23,221 22,791 21,842 21,514 20,582 19,710 18,919 17,005 16,106 14,949 13,063 12,189
C. First Marriage
A.40
Add’l Years of FPE or MTI -0.001 0.001 -0.002 -0.003 -0.004 -0.005 -0.008 -0.011 -0.010 -0.008 -0.003 -0.000 -0.002 0.000 0.004
w/out Script Change (0.005) (0.006) (0.006) (0.004) (0.004) (0.004) (0.004) (0.004) (0.005) (0.004) (0.003) (0.004) (0.004) (0.004) (0.003)
(∆Izy ) [0.904] [0.849] [0.798] [0.428] [0.320] [0.303] [0.024] [0.005] [0.027] [0.038] [0.324] [0.902] [0.680] [0.948] [0.189]
N 24,898 24,616 24,058 23,221 22,791 21,842 21,514 20,582 19,710 18,919 17,005 16,106 14,949 13,063 12,189
D. First Intercourse
Add’l Years of FPE or MTI -0.001 -0.003 -0.004 0.000 -0.001 -0.003 -0.006 -0.008 -0.008 -0.004 -0.001 -0.000 0.001 0.003 0.003
w/out Script Change (0.005) (0.005) (0.005) (0.004) (0.004) (0.004) (0.003) (0.003) (0.003) (0.003) (0.002) (0.003) (0.002) (0.002) (0.003)
(∆Izy ) [0.864] [0.508] [0.420] [0.983] [0.807] [0.411] [0.062] [0.006] [0.017] [0.095] [0.719] [0.921] [0.753] [0.244] [0.296]
N 24,898 24,616 24,058 23,221 22,791 21,842 21,514 20,582 19,710 18,919 17,005 16,106 14,949 13,063 12,189
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: FPE = free primary education; MTI = mother tongue instruction. In panel A the dependent variable is number of births by the stated age; in the remaining panels,
the dependent variable is an indicator that equals 1 if the event occurred by the denoted age and 0 otherwise. All samples include women in birth cohorts from 1970 to 1988
who are older than the denoted age. All regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Each estimate is from a unique
regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
Figure C.1: Reduced form Estimates of the Eﬀect of the Reforms on Number of Births, by Age
Source: Author’s analysis of data from the Ethiopian Demographic and Health Survey (2005, 2011, and 2016).
Note: The dependent variable is the number of births by the stated age. Coeﬃcient estimates are from a reduced-
form model using the combined intensity measure ∆Izy and a sample of women older than the age stated.
Figure C.2: Reduced form Estimates of the Eﬀect of the Reforms on
the Timing of First Birth, Marriage, and Intercourse, by Age
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (2005, 2011, and 2016).
Note: The dependent variable is an indicator that equals 1 if the ﬁrst instance of the deﬁned event occurred at or before each
age. Coeﬃcient estimates are from a reduced-form model using the combined intensity measure ∆Izy and a sample of women
older than the age stated. The 90 percent conﬁdence intervals are shown with solid gray bars and the 95 percent conﬁdence
intervals with dashed gray bars.
A.41
D Alternative Samples and Speciﬁcations
D.1 Pre-Treatment Trends and Placebo Estimates
Figure D.1: Comparison of Pre-Treatment Trends in Years of Schooling, by Birth Year
Note: Source: Author’s analysis based on data from the Ethiopian census of 2007 and the Demographic and Health
Survey in years 2005, 2011, and 2016.
Note: Data in the ﬁgure are from non-MTI regions only and are sorted by zone-level pre-1970 schooling, the data
used to create the free primary education (FPE) intensity measure. Above-median observations are represented by
circles and below-median observations by squares. The trends are calculated using the 1970–1979 cohorts, those
for which there is only a nominal change in the intensity measure seen in ﬁg. 2 in the main article. Post-1979
cohorts are indicated by larger markers.
A.42
Table D.1: Placebo Estimates Using Pre-FPE Cohorts and Misplaced Timing of Intensity Measure
First (Misplaced) Post-Reform Cohort
1979 1978 1977 1976 1975
(1) (2) (3) (4) (5)
A.i. Non-MTI Regions: Years of Schooling
Add’l Years of Free -0.042 -0.048 -0.045 0.067 0.056
FPE
Schooling Izy (0.049) (0.064) (0.072) (0.061) (0.056)
[0.396] [0.462] [0.540] [0.280] [0.327]
A.ii. Non-MTI Regions: Number of Children Born
Add’l Years of Free -0.015 -0.051 -0.004 -0.005 0.047
FPE
Schooling Izy (0.022) (0.029) (0.031) (0.045) (0.049)
[0.501] [0.093] [0.891] [0.915] [0.346]
N 8,074 8,074 8,074 8,074 8,074
B.i. National Estimates (FPE + MTI): Years of Schooling
Add’l Years of FPE or MTI -0.008 -0.016 -0.009 0.057 0.042
w/out Script Change (0.040) (0.039) (0.036) (0.045) (0.035)
(∆Izy ) [0.844] [0.674] [0.815] [0.216] [0.241]
B.ii. National Estimates (FPE + MTI): Number of Children Born
Add’l Years of FPE or MTI 0.013 -0.020 -0.016 -0.025 -0.028
w/out Script Change (0.030) (0.023) (0.022) (0.020) (0.022)
(∆Izy ) [0.657] [0.380] [0.458] [0.226] [0.208]
N 14,833 14,833 14,833 14,833 14,833
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: MTI = mother tongue instruction; FPE = free primary education. The dependent variable is years of schooling in panels
A.i and B.i, and is number of births in panels A.ii and B.ii. All samples include women in birth cohorts from 1963 to 1979,
the same number of cohorts as the baseline estimates but moved nine years back. In the baseline sample the ﬁrst post-reform
cohort is 1988, also at least a nine-cohort diﬀerence. This timing matches the period in which the FPE intensity measure
FPE ) predicts little impact of the reform, as seen in ﬁg. 2 of the main article. All regressions include birth year and zone ﬁxed
(Izy
eﬀects, zone-speciﬁc linear trends, and a cubic for age. Standard errors are clustered at the zone level and shown in parentheses;
p-values are shown in square brackets.
A.43
D.2 Alternative Samples and Speciﬁcations
The results shown throughout Appendix Section D.2 re-estimate the output shown in the main body of the
paper using a number of alternative cohort ranges, speciﬁcations, and samples. Each panel is described
below, and the results in the following tables display the panels in a consistent order.
Panel A – Baseline
For reference, Panel A reproduces the results used in the paper. These estimates use a sample of all Ethiopian
women born between 1970 and 1988. All estimates in this section include a cubic in age, and birth year and
FPE
district ﬁxed eﬀects. The baseline estimates use the FPE instrument, Izy , starting age data from the 2007
census, and a district speciﬁc linear time trend.
Panels B to E – Alternative Cohort Ranges
These panels include two expanded samples, 1968 to 1992 (Panel B) and 1969 to 1989 (Panel C), and two
more restrictive ranges. In Panel D, one cohort from each end of the baseline sample is removed, yielding
a range from 1971 to 1987. This sample no longer includes any fully post-reform cohorts. The data are
restricted to 1972 in Panel E, the ﬁnal fully pre-FPE cohort. Removing additional cohorts on the later end
of the range would remove signiﬁcant and necessary identifying variation; therefore, the 1987 cohort remains
the cutoﬀ on the upper end of the range in Panel E.
Panel F – Matched 1984 Start Ages
Intensity measures are constructed using starting age information from the 1984 census. While the pre-
reform timing of these data are ideal, the administrative boundaries are not consistent between the 1984 and
post-1991 periods. Therefore, while there is starting age information contained in the 1984 census, the level
two administrative information does not match with the zones used in the study. To adjust the 1984 data
to the 1994 geographical boundaries, shapeﬁles from each time period provided by Minnesota Population
Center and the Ethiopian Central Statistical Agency (2017) are overlaid, and new start values for the post-
1991 boundaries are calculated as the weighted averages of the start age value from the 1984 area and the
portion of the post-1991 zone that is made up of that 1984 area. Unfortunately, this requires an unrealistic
assumption of a consistent distribution of population within geographic area, and introduces a signiﬁcant
amount of measurement error into the start age calculations.
A.44
Panel G – 1994 Start Ages
Start ages from the 1994 census are used to calculate all intensity measures. The timing of this survey is
problematic in the sense that the MTI implementation had already begun at this time. This along with any
anticipation of the forthcoming FPE program could alter the decision to enter school in 1994.
Panel H – Three Part Trend
The district-speciﬁc linear trends are replaced with a set of district-speciﬁc trends that are allowed to change
slope at two points, in 1978 and in 1987. On time entrants are partially treated beginning with the 1978
cohort, and fully treated beginning with the 1987 cohort.
Panel I – Regional Trends
The district-speciﬁc linear trend is replaced with a region-speciﬁc linear trend.
Panel J – No Trends
All trend variables are removed from the estimating equations.
Panel K – Only Zones in All Rounds of the DHS
Data are restricted only to zones with observations in all three of the rounds of the DHS survey. This
includes 25 of 30 zones in the non-MTI regions, and 48 of the 60 zones throughout Ethiopia.
Panel L – Zones with Fewer than 4,000 Organized Violence Deaths (1989 to 1991)
Data from the four zones with the highest level of pre-independence violence, three of which are in the
non-MTI sample, are removed from the sample. These zones contain more than 75 percent of all deaths
included in the data over this time period.
Panel M – Zones with Fewer than 500 Organized Violence Deaths (1989 to 1991)
Data from 13 zones with more than 500 deaths related to organized violence in the pre-independence period,
eight of which are in the non-MTI sample, are removed from the sample. These zones contain more than 96
percent of all deaths included in the data over this time period.
Panel N – Zones without High Intensity of Famine (1985)
Areas of Ethiopia (Tigray, Afar, Somali regions or the zones of Gonder) from which over 90 percent of
individuals who registered with international shelters and camps at the height of the famine are removed
from the sample (USAID, 1987). This includes 15 of the 30 zones in non-MTI regions.
A.45
The following panels are included when the eﬀect of the combined FPE and MTI reforms are studied using
the national sample, in Appendix Tables D.4 and D.5.
Panel P – No Tigray
Observations from the Tigray region are dropped. Tigray is the region for which the model estimates a
positive return to the MTI reform.
Panel Q – Boothe and Walker (1997) MTI Deﬁnition
In addition to the corroborated set of languages included in the paper’s joint deﬁnition, Boothe and Walker
(1997) also ﬁnd evidence that Somaligna was introduced in the Somali region for the ﬁrst six grades in 1993,
during the second round of translation introduced by the Council of Representatives.
Table D.2: Additional Language(s) in Boothe and Walker (1997) Deﬁnition
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Somaligna 1993 Somali 1-6 0.96 0.95
Source : Author’s summary based on information from Boothe and Walker (1997).
Note : MT = mother tongue.
A non-zero MTI measure for the Somali region is introduced in the calculation of new MTI and joint intensity
measures.
Panel R – Zenebe Gebre (2014) MTI Deﬁnition
In addition to the corroborated set of languages included in the paper’s joint deﬁnition, Zenebe Gebre (2014)
also ﬁnds evidence of six additional languages being introduced prior to the 1995 fee removal. Three of these
languages were smaller languages introduced in SNNPR in 1992 and 1993. The other three are found to
be introduced in 1994, one is an expansion of Oromigna in the Amhara region, and the ﬁnal two are new
languages in new regions.
Table D.3: Additional Language(s) in Zenebe Gebre (2014) Deﬁnition
Fraction of:
MT Speakers Region Speaking
Language Year Region Grades Living in Region Language as MT
Gamogna 1992 SNNPR 1-4 0.96 0.07
Goﬃgna 1992 SNNPR 1-4 1.00 0.02
Dawurogna 1993 SNNPR 1-4 0.88 0.03
Oromigna 1994 Amhara 1-8 0.02 0.03
Anyiwakgna 1994 Gambela 1-4 0.98 0.27
Hareriegna 1994 Harari 1-6 0.47 0.07
Source : Author’s summary based on information from Zenebe Gebre (2014).
Note : MT = mother tongue; SNNPR = Southern Nations, Nationalities, and Peoples’ Region.
A.46
A non-zero MTI measure for Amhara, Gambela, and Harari are introduced in the calculation of new MTI
and joint intensity measures, along with the necessary adjustments to the SNNPR region.
A.47
Table D.4: Eﬀect of FPE and MTI Reforms on Years of Schooling:
Alternative Samples and Speciﬁcations - Analysis of First Stage Results from Table 2
A B C D E G H I J K L M N O
1968 1969 1971 1972 Start Ages Trends Consistent < 4000 < 500 No Main
Baseline 1990 1989 1987 1987 1984 1994 Three-Part Regional None Zones Deaths Deaths Famine Zones
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)
I. Non-MTI Regions Only - Eﬀect of FPE
i. Census + DHS
Add’l Years of Free 0.131 0.164 0.149 0.116 0.117 0.146 0.100 0.139 0.102 0.259 0.134 0.129 0.135 0.185
FPE
Schooling Izy (0.034) (0.045) (0.039) (0.028) (0.030) (0.043) (0.059) (0.047) (0.052) (0.046) (0.034) (0.034) (0.034) (0.035)
[0.001] [0.001] [0.001] [0.000] [0.001] [0.002] [0.101] [0.006] [0.056] [0.000] [0.001] [0.001] [0.001] [0.000]
A.48
F-Statistic 14.80 13.36 14.34 17.21 14.94 11.54 2.86 8.77 3.93 31.36 15.28 14.53 15.87 28.05
N 83,005 101,702 95,469 76,674 74,694 83,005 83,005 83,005 83,005 83,005 78,149 66,143 45,162 55,208
ii. DHS Only
Add’l Years of Free 0.112 0.146 0.122 0.106 0.097 0.141 0.082 0.181 0.082 0.244 0.116 0.109 0.126 0.170
FPE
Schooling Izy (0.046) (0.053) (0.048) (0.045) (0.048) (0.057) (0.075) (0.043) (0.066) (0.053) (0.046) (0.050) (0.048) (0.056)
[0.021] [0.010] [0.016] [0.027] [0.055] [0.019] [0.286] [0.000] [0.224] [0.000] [0.018] [0.038] [0.015] [0.009]
F-Statistic 5.93 7.54 6.51 5.44 3.99 6.16 1.18 17.98 1.54 21.47 6.44 4.77 6.95 9.26
N 13,922 16,481 15,108 12,299 11,967 13,922 13,922 13,922 13,922 13,922 13,638 12,819 9,913 10,161
Table D.4: (... continued) Eﬀect of FPE and MTI Reforms on Years of Schooling:
Alternative Samples and Speciﬁcations - Analysis of First Stage Results from Table 2
A B C D E F G H I J K L M N O P Q
1968 1969 1971 1972 Start Ages Trends Consistent < 4000 < 500 No Main No BW (1997) TZG (2014)
Baseline 1990 1989 1987 1987 1984 1994 Three-Part Regional None Zones Deaths Deaths Famine Zones Tigray Deﬁnition Deﬁnition
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17)
II. National Sample - Eﬀect of FPE and MTI (Separate Instruments)
i. Census + DHS
Add’l Years of Free 0.115 0.149 0.130 0.108 0.111 0.116 0.098 0.132 0.082 0.236 0.132 0.114 0.099 0.150 0.113 0.114 0.114
FPE
Schooling Izy (0.044) (0.046) (0.046) (0.033) (0.033) (0.057) (0.053) (0.062) (0.057) (0.073) (0.039) (0.044) (0.052) (0.048) (0.044) (0.047) (0.044)
[0.011] [0.002] [0.007] [0.002] [0.001] [0.045] [0.072] [0.038] [0.155] [0.002] [0.001] [0.011] [0.065] [0.003] [0.012] [0.019] [0.012]
Add’l Year of MTI 0.183 0.226 0.204 0.134 0.130 0.162 0.173 0.222 0.154 0.271 0.190 0.194 — — — 0.186 0.185
No Script Change (0.070) (0.060) (0.062) (0.066) (0.062) (0.087) (0.051) (0.077) (0.067) (0.056) (0.071) (0.088) — — — (0.071) (0.071)
M T I −T
A.49
Izy [0.012] [0.000] [0.002] [0.046] [0.041] [0.068] [0.001] [0.005] [0.025] [0.000] [0.010] [0.032] — — — [0.011] [0.011]
Add’l Year of MTI -0.119 -0.108 -0.097 -0.163 -0.182 -0.086 -0.054 -0.151 -0.114 -0.072 -0.114 -0.110 -0.101 -0.105 -0.120 -0.100 -0.106
with Script Change (0.052) (0.044) (0.042) (0.058) (0.062) (0.064) (0.063) (0.066) (0.047) (0.040) (0.053) (0.058) (0.076) (0.061) (0.052) (0.050) (0.052)
MT I
Izy [0.025] [0.016] [0.022] [0.007] [0.005] [0.189] [0.387] [0.025] [0.018] [0.076] [0.037] [0.063] [0.191] [0.093] [0.025] [0.052] [0.044]
FPE M T I −T
Izy × Izy -0.008 -0.014 -0.013 -0.005 -0.005 -0.009 0.000 -0.017 -0.009 -0.011 -0.009 -0.008 — — — -0.009 -0.009
(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.009) (0.009) (0.006) (0.005) (0.006) (0.008) — — — (0.006) (0.006)
[0.169] [0.023] [0.030] [0.419] [0.400] [0.170] [0.962] [0.067] [0.120] [0.039] [0.136] [0.337] — — — [0.154] [0.169]
FPE MT I
Izy × Izy 0.001 0.001 0.000 0.005 0.004 -0.001 -0.002 0.008 0.001 0.002 0.001 0.002 0.003 -0.002 0.002 0.000 0.001
(0.005) (0.004) (0.004) (0.004) (0.004) (0.005) (0.005) (0.009) (0.005) (0.005) (0.005) (0.005) (0.006) (0.005) (0.005) (0.005) (0.005)
[0.767] [0.792] [0.999] [0.247] [0.312] [0.899] [0.727] [0.390] [0.789] [0.783] [0.851] [0.699] [0.599] [0.664] [0.743] [0.995] [0.818]
F-Statistic 14.09 14.64 13.93 13.46 13.39 5.19 6.46 6.91 11.88 14.48 15.61 13.39 7.94 11.60 10.67 13.63 13.41
N 205,141 249,768 235,691 190,989 186,219 205,141 205,141 205,141 205,141 205,141 198,667 184,888 141,172 164,252 192,049 205,141 205,141
Table D.4: (... continued) Eﬀect of FPE and MTI Reforms on Years of Schooling:
Alternative Samples and Speciﬁcations - Analysis of First Stage Results from Table 2
A B C D E F G H I J K L M N O P Q
1968 1969 1971 1972 Start Ages Trends Consistent < 4000 < 500 No Main No BW (1997) TZG (2014)
Baseline 1990 1989 1987 1987 1984 1994 Three-Part Regional None Zones Deaths Deaths Famine Zones Tigray Deﬁnition Deﬁnition
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17)
II. National Sample - Eﬀect of FPE and MTI (Separate Instruments)
ii. DHS Only
Add’l Years of Free 0.115 0.155 0.124 0.110 0.107 0.116 0.111 0.178 0.093 0.239 0.137 0.117 0.098 0.152 0.113 0.112 0.114
FPE
Schooling Izy (0.053) (0.050) (0.052) (0.049) (0.050) (0.070) (0.064) (0.074) (0.065) (0.070) (0.049) (0.056) (0.065) (0.063) (0.053) (0.057) (0.054)
[0.035] [0.003] [0.020] [0.028] [0.036] [0.105] [0.085] [0.019] [0.157] [0.001] [0.008] [0.043] [0.143] [0.020] [0.039] [0.053] [0.040]
Add’l Year of MTI 0.175 0.222 0.202 0.105 0.095 0.141 0.177 0.223 0.143 0.268 0.184 0.175 — — — 0.179 0.179
No Script Change (0.087) (0.071) (0.075) (0.082) (0.081) (0.111) (0.069) (0.090) (0.082) (0.054) (0.087) (0.105) — — — (0.087) (0.087)
M T I −T
A.50
Izy [0.049] [0.003] [0.009] [0.207] [0.246] [0.207] [0.013] [0.017] [0.086] [0.000] [0.038] [0.103] — — — [0.045] [0.045]
Add’l Year of MTI -0.128 -0.112 -0.094 -0.210 -0.237 -0.082 -0.042 -0.190 -0.124 -0.079 -0.122 -0.111 -0.104 -0.115 -0.130 -0.107 -0.111
with Script Change (0.073) (0.063) (0.063) (0.090) (0.093) (0.089) (0.086) (0.090) (0.062) (0.053) (0.075) (0.082) (0.108) (0.086) (0.074) (0.071) (0.074)
MT I
Izy [0.085] [0.077] [0.140] [0.023] [0.014] [0.360] [0.626] [0.038] [0.051] [0.143] [0.109] [0.182] [0.341] [0.192] [0.086] [0.138] [0.136]
FPE M T I −T
Izy × Izy -0.010 -0.016 -0.016 -0.006 -0.006 -0.010 -0.002 -0.021 -0.011 -0.012 -0.011 -0.008 — — — -0.010 -0.010
(0.008) (0.007) (0.007) (0.009) (0.009) (0.008) (0.012) (0.013) (0.008) (0.007) (0.008) (0.011) — — — (0.008) (0.008)
[0.239] [0.025] [0.038] [0.518] [0.504] [0.238] [0.875] [0.105] [0.195] [0.114] [0.201] [0.472] — — — [0.239] [0.242]
FPE MT I
Izy × Izy -0.001 -0.001 -0.003 0.007 0.006 -0.004 -0.005 0.010 0.000 0.000 -0.001 0.000 0.002 -0.004 0.000 -0.001 -0.001
(0.007) (0.006) (0.006) (0.007) (0.007) (0.008) (0.008) (0.012) (0.007) (0.008) (0.007) (0.008) (0.008) (0.007) (0.007) (0.007) (0.007)
[0.930] [0.861] [0.659] [0.325] [0.400] [0.627] [0.524] [0.382] [1.000] [0.955] [0.868] [0.975] [0.855] [0.573] [0.962] [0.860] [0.910]
F-Statistic 6.91 20.55 11.63 7.53 6.56 2.07 3.34 5.82 5.27 12.43 8.00 5.72 4.41 5.83 6.18 6.39 6.33
N 24,898 29,547 27,106 22,023 21,390 24,898 24,898 24,898 24,898 24,898 24,478 23,188 17,942 18,739 22,500 24,898 24,898
Table D.4: (... continued) Eﬀect of FPE and MTI Reforms on Years of Schooling:
Alternative Samples and Speciﬁcations - Analysis of First Stage Results from Table 2
A B C D E F G H I J K L M N O P Q
1968 1969 1971 1972 Start Ages Trends Consistent < 4000 < 500 No Main No BW (1997) TZG (2014)
Baseline 1990 1989 1987 1987 1984 1994 Three-Part Regional None Zones Deaths Deaths Famine Zones Tigray Deﬁnition Deﬁnition
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17)
III. National Sample - Eﬀect of FPE and MTI (Combined Instrument)
i. Census + DHS
Add’l Years of FPE or MTI 0.109 0.130 0.118 0.105 0.111 0.096 0.075 0.139 0.095 0.129 0.114 0.107 0.098 0.132 0.113 0.106 0.106
w/out Script Change (0.018) (0.021) (0.018) (0.019) (0.021) (0.024) (0.022) (0.027) (0.020) (0.028) (0.017) (0.019) (0.023) (0.026) (0.025) (0.017) (0.018)
(∆Izy ) [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000]
F-Statistic 38.28 38.79 41.97 29.36 28.81 16.14 11.12 25.56 22.05 20.67 43.86 32.37 18.69 25.66 20.58 37.05 35.49
N 205,141 249,768 235,691 190,989 186,219 205,141 205,141 205,141 205,141 205,141 198,667 184,888 143,377 164,252 192,049 205,141 205,141
A.51
ii. DHS Only
Add’l Years of FPE or MTI 0.115 0.138 0.123 0.111 0.121 0.096 0.078 0.160 0.104 0.136 0.121 0.107 0.097 0.141 0.122 0.109 0.110
w/out Script Change (0.026) (0.026) (0.024) (0.021) (0.033) (0.033) (0.030) (0.037) (0.027) (0.029) (0.026) (0.028) (0.032) (0.043) (0.039) (0.025) (0.026)
(∆Izy ) [0.000] [0.000] [0.000] [0.000] [0.001] [0.005] [0.012] [0.000] [0.000] [0.000] [0.000] [0.000] [0.004] [0.002] [0.003] [0.000] [0.000]
F-Statistic 19.56 27.76 26.44 14.09 13.31 8.67 6.77 18.47 14.31 21.77 21.91 14.22 9.17 11.05 9.69 18.39 17.86
N 24,898 29,547 27,106 22,023 21,390 24,898 24,898 24,898 24,898 24,898 24,478 23,188 18,396 18,739 22,500 24,898 24,898
Source: Author’s analysis based on data from the Ethiopian census of 2007 and the Demographic and Health Survey (DHS) in years 2005, 2011, and 2016.
Note: FPE = free primary education; MTI = mother tongue instruction; BW (1997) = Boothe and Walker (1997); TZG (2014) = Zenebe Gebre (2014). The dependent variable
is years of schooling. All sample and speciﬁcation deﬁnitions can be found in Section D.2 of this supplementary online appendix. Unless otherwise noted, all regressions include
birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. Panel I includes only observations from non-MTI regions; all regions are included in panels
II and III. Estimates in each column and panel are from a unique regression. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in
square brackets.
Table D.5: Eﬀect of Years of Schooling on Number of Children Born:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 2
Non-MTI National Sample National Sample
Regions Only Eﬀect of FPE and MTI Eﬀect of FPE and MTI
Eﬀect of FPE (Separate Instruments) (Combined Instrument)
Census DHS Census DHS Census DHS
+ DHS Only + DHS Only + DHS Only
(1) (2) (3) (4) (5) (6)
A. Baseline
Years of Schoolingizy -0.437 -0.529 -0.273 -0.365 -0.257 -0.321
(0.090) (0.165) (0.109) (0.147) (0.117) (0.156)
[0.000] [0.001] [0.012] [0.001] [0.028] [0.040]
First Stage F-Statistic 14.796 5.93 14.09 6.91 38.28 19.56
N 69,083 13,922 205,141 24,898 205,141 24,898
B. Cohorts: 1968 - 1990
Years of Schoolingizy -0.422 -0.406 -0.212 -0.178 -0.198 -0.178
(0.092) (0.115) (0.101) (0.117) (0.109) (0.133)
[0.000] [0.000] [0.035] [0.131] [0.069] [0.181]
First Stage F-Statistic 13.36 7.54 14.64 20.55 38.79 27.76
N 101,702 16,481 249,768 29,547 249,768 29,547
C. Cohorts: 1969 - 1989
Years of Schoolingizy -0.429 -0.455 -0.189 -0.142 -0.188 -0.168
(0.095) (0.149) (0.106) (0.123) (0.113) (0.141)
[0.000] [0.002] [0.074] [0.250] [0.096] [0.235]
First Stage F-Statistic 14.34 6.51 13.93 11.63 41.97 26.44
N 95,469 15,108 235,691 27,106 235,691 27,106
D. Cohorts: 1971 - 1987
Years of Schoolingizy -0.427 -0.489 -0.359 -0.472 -0.280 -0.364
(0.105) (0.184) (0.113) (0.147) (0.112) (0.150)
[0.000] [0.008] [0.001] [0.001] [0.013] [0.015]
First Stage F-Statistic 17.21 5.44 13.46 7.53 29.36 14.09
N 76,674 12,299 190,989 22,023 190,989 22,023
E. Cohorts: 1972 - 1987
Years of Schoolingizy -0.380 -0.413 -0.326 -0.422 -0.230 -0.313
(0.099) (0.174) (0.124) (0.175) (0.121) (0.157)
[0.000] [0.017] [0.009] [0.016] [0.049] [0.046]
First Stage F-Statistic 14.94 3.99 13.39 6.56 28.81 13.31
N 74,694 11,967 186,219 21,390 186,219 21,390
A.52
Table D.5: (... continued) Eﬀect of Years of Schooling on Number of Children Born:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 2
Non-MTI National Sample National Sample
Regions Only Eﬀect of FPE and MTI Eﬀect of FPE and MTI
Eﬀect of FPE (Separate Instruments) (Combined Instrument)
Census DHS Census DHS Census DHS
+ DHS Only + DHS Only + DHS Only
(1) (2) (3) (4) (5) (6)
F. Using 1984 Census Matched Start Ages
Years of Schoolingizy -0.539 -0.580 -0.261 -0.314 -0.257 -0.270
(0.144) (0.187) (0.123) (0.172) (0.139) (0.188)
[0.000] [0.002] [0.035] [0.068] [0.065] [0.150]
First Stage F-Statistic 11.54 6.16 5.19 2.07 16.14 8.67
N 83,005 13,922 205,141 24,898 205,141 24,898
G. Using 1994 Census Start Ages
Years of Schoolingizy -0.550 -0.719 -0.224 -0.298 -0.231 -0.218
(0.196) (0.416) (0.111) (0.137) (0.166) (0.218)
[0.005] [0.084] [0.043] [0.030] [0.163] [0.317]
First Stage F-Statistic 2.86 1.18 6.46 3.34 11.12 6.77
N 83,005 13,922 205,141 24,898 205,141 24,898
H. Three-Part District Trends
Years of Schoolingizy -0.480 -0.372 -0.303 -0.367 -0.292 -0.368
(0.111) (0.143) (0.112) (0.129) (0.115) (0.140)
[0.000] [0.009] [0.007] [0.004] [0.011] [0.009]
First Stage F-Statistic 8.77 17.98 6.91 5.82 25.56 18.47
N 83,005 13,922 205,141 24,898 205,141 24,898
I. Regional Trends
Years of Schoolingizy -0.667 -0.822 -0.317 -0.371 -0.331 -0.358
(0.307) (0.530) (0.137) (0.188) (0.147) (0.191)
[0.030] [0.121] [0.021] [0.048] [0.024] [0.060]
First Stage F-Statistic 3.93 1.54 11.88 5.27 22.05 14.31
N 83,005 13,922 205,141 24,898 205,141 24,898
J. No Trends
Years of Schoolingizy -0.837 -0.963 -0.336 -0.351 -0.187 -0.216
(0.092) (0.152) (0.224) (0.238) (0.165) (0.176)
[0.000] [0.000] [0.134] [0.139] [0.258] [0.219]
First Stage F-Statistic 31.36 21.47 14.48 12.43 20.67 21.77
N 83,005 13,922 205,141 24,898 205,141 24,898
A.53
Table D.5: (... continued) Eﬀect of Years of Schooling on Number of Children Born:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 2
Non-MTI National Sample National Sample
Regions Only Eﬀect of FPE and MTI Eﬀect of FPE and MTI
Eﬀect of FPE (Separate Instruments) (Combined Instrument)
Census DHS Census DHS Census DHS
+ DHS Only + DHS Only + DHS Only
(1) (2) (3) (4) (5) (6)
K. Only Zones in All DHS Rounds (25 of 30; 48 of 60) National Sample National Sample
Years of Schoolingizy -0.425 -0.499 -0.260 -0.343 -0.230 -0.303 Eﬀect of FPE and MTI Eﬀect of FPE and MTI
(0.087) (0.157) (0.104) (0.135) (0.113) (0.149) (Separate Instruments) (Combined Instrument)
[0.000] [0.001] [0.013] [0.011] [0.035] [0.042]
Census DHS Census DHS
First Stage F-Statistic 15.28 6.44 15.61 8.00 43.86 21.91 + DHS Only + DHS Only
N 78,149 13,638 198,667 24,478 198,667 24,478 (3) (4) (5) (6)
L. Less than 4,000 Organized Violence Deaths: 1989 to 91 O. No Tigray
Years of Schoolingizy -0.456 -0.550 -0.264 -0.342 -0.254 -0.312 -0.443 -0.565 -0.419 -0.564
(0.103) (0.200) (0.119) (0.156) (0.124) (0.169) (0.152) (0.237) (0.154) (0.223)
[0.000] [0.006] [0.027] [0.028] [0.040] [0.064] [0.004] [0.017] [0.006] [0.012]
First Stage F-Statistic 14.53 4.77 13.39 5.72 32.37 14.22 10.67 6.18 20.58 9.69
N 66,143 12,819 184,888 23,188 184,888 23,188 192,049 22,500 192,049 22,500
M. Less than 500 Organized Violence Deaths: 1989 to 91 P. Boothe and Walker (1997) Deﬁnition
A.54
Years of Schoolingizy -0.508 -0.515 -0.436 -0.394 -0.336 -0.345 -0.211 -0.303 -0.201 -0.276
(0.186) (0.252) (0.184) (0.268) (0.178) (0.238) (0.109) (0.142) (0.119) (0.159)
[0.006] [0.041] [0.018] [0.142] [0.060] [0.148] [0.053] [0.033] [0.091] [0.081]
First Stage F-Statistic 15.87 6.95 7.94 4.41 18.69 9.17 13.63 6.39 37.05 18.39
N 45,162 9,913 141,172 17,942 143,377 18,396 205,141 24,898 205,141 24,898
N. No Regions of Highest Famine Concentration Q. Zenebe Gebre (2014) Deﬁnition
Years of Schoolingizy -0.334 -0.374 -0.341 -0.442 -0.380 -0.510 -0.264 -0.344 -0.250 -0.312
(0.079) (0.118) (0.139) (0.201) (0.150) (0.209) (0.110) (0.145) (0.119) (0.161)
[0.000] [0.002] [0.014] [0.027] [0.011] [0.015] [0.016] [0.018] [0.036] [0.053]
First Stage F-Statistic 28.05 9.26 11.60 5.83 25.66 11.05 13.41 6.33 35.49 17.86
N 55,208 10,161 164,252 18,739 164,252 18,739 205,141 24,898 205,141 24,898
Source: Author’s analysis based on data from the Ethiopian census of 2007 and the Demographic and Health Survey (DHS) in years 2005, 2011, and
2016.
Note: FPE = free primary education; MTI = mother tongue instruction. The dependent variable is the number of births. In columns 1 and
FPE ; in columns 3 and 4 additional
2, Years of Schoolingizy is the predicted level of schooling instrumented with the FPE intensity measure Izy
instruments include two MTI intensity measures Izy MTI-T and I MTI , which denote the measures for MTI regions without and with script change,
zy
respectively, and the interactions for regions in which two interventions occurred. In columns 5 and 6, the joint intensity measure ∆Izy is used. All
sample and speciﬁcation deﬁnitions can be found in Section D.2 of this supplementary online appendix. Unless otherwise noted, all regressions include
birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends, and a cubic for age. The ﬁrst two columns include only observations from non-MTI
regions; all regions are included in the ﬁnal four columns. Each estimate is from a unique regression. Standard errors are clustered at the zone level
and shown in parentheses; p-values are shown in square brackets.
Table D.6: Eﬀect of Years of Schooling on Knowledge and Health in non-MTI Regions:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 3
Acceptable Reasons
Read about Know about BMI Height for Domestic Use Modern Use Hidden
Literacy Fam. Planning Fam. Planning (z-score) (z-score) Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
A. Baseline
Years of Schoolingizy 0.092 0.048 -0.013 0.316 -0.271 -0.361 -0.018 -0.035
(0.028) (0.029) (0.024) (0.355) (0.302) (0.211) (0.051) (0.042)
[0.001] [0.097] [0.594] [0.374] [0.369] [0.087] [0.721] [0.402]
First Stage F-Statistic 6.10 5.92 5.93 1.91 2.22 5.67 5.93 5.93
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
B. Cohorts: 1968 - 1990
Years of Schoolingizy 0.107 0.031 -0.011 0.211 -0.121 -0.35 -0.001 -0.023
(0.018) (0.016) (0.016) (0.205) (0.166) (0.179) (0.038) (0.034)
[0.000] [0.047] [0.499] [0.304] [0.467] [0.051] [0.987] [0.492]
First Stage F-Statistic 7.86 7.54 7.54 3.54 3.77 7.41 7.54 7.54
N 16,157 16,471 16,481 12,962 13,286 15,848 16,481 16,481
C. Cohorts: 1969 - 1989
Years of Schoolingizy 0.105 0.034 -0.015 0.285 -0.178 -0.424 0.002 -0.026
(0.026) (0.021) (0.020) (0.289) (0.233) (0.229) (0.050) (0.042)
[0.000] [0.110] [0.450] [0.324] [0.443] [0.064] [0.966] [0.546]
First Stage F-Statistic 7.05 6.50 6.51 2.55 2.64 6.46 6.51 6.51
N 14,832 15,098 15,108 11,802 12,086 14,528 15,108 15,108
D. Cohorts: 1971 - 1987
Years of Schoolingizy 0.095 0.056 -0.034 0.152 -0.271 -0.216 -0.052 -0.072
(0.028) (0.033) (0.033) (0.285) (0.334) (0.191) (0.052) (0.042)
[0.001] [0.086] [0.294] [0.592] [0.418] [0.259] [0.322] [0.086]
First Stage F-Statistic 5.37 5.41 5.44 1.36 1.76 4.40 5.44 5.44
N 12,078 12,292 12,299 9,653 9,888 11,872 12,299 12,299
E. Cohorts: 1972 - 1987
Years of Schoolingizy 0.079 0.059 -0.038 0.290 -0.420 -0.173 -0.041 -0.078
(0.031) (0.036) (0.038) (0.436) (0.550) (0.207) (0.054) (0.040)
[0.010] [0.101] [0.320] [0.507] [0.445] [0.405] [0.456] [0.054]
First Stage F-Statistic 3.92 3.97 3.99 1.02 1.15 3.13 3.99 3.99
N 11,751 11,960 11,967 9,375 9,608 11,548 11,967 11,967
F. Using 1984 Census Matched Start Ages
Years of Schoolingizy 0.083 0.045 -0.028 0.139 -0.206 -0.251 0.017 -0.002
(0.026) (0.030) (0.022) (0.224) (0.294) (0.125) (0.062) (0.050)
[0.002] [0.123] [0.212] [0.536] [0.485] [0.044] [0.783] [0.970]
First Stage F-Statistic 6.14 6.14 6.16 1.46 1.72 5.85 6.16 6.16
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
G. Using 1994 Census Start Ages
Years of Schoolingizy 0.096 0.038 -0.019 0.267 0.055 -0.217 -0.060 -0.084
(0.033) (0.029) (0.033) (0.528) (0.285) (0.232) (0.070) (0.086)
[0.004] [0.181] [0.562] [0.614] [0.848] [0.351] [0.392] [0.328]
First Stage F-Statistic 1.20 1.18 1.18 0.49 0.49 0.84 1.18 1.18
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
H. Three-Part District Trends
Years of Schoolingizy 0.079 0.049 -0.003 0.171 -0.124 -0.322 -0.011 -0.002
(0.023) (0.020) (0.023) (0.163) (0.212) (0.181) (0.052) (0.046)
[0.001] [0.016] [0.886] [0.296] [0.558] [0.076] [0.827] [0.958]
First Stage F-Statistic 17.48 17.68 17.98 8.53 8.54 11.35 17.98 17.98
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
A.55
Table D.6: (... continued) Eﬀect of Years of Schooling on Knowledge and Health:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 3
Acceptable Reasons
Read about Know about BMI Height for Domestic Use Modern Use Hidden
Literacy Fam. Planning Fam. Planning (z-score) (z-score) Violence (of 5) Contraception Contraception
(1) (2) (3) (4) (5) (6) (7) (8)
I. Regional Trends
Years of Schoolingizy 0.102 0.072 -0.026 0.329 -0.426 -0.388 -0.038 -0.050
(0.034) (0.066) (0.035) (0.778) (1.072) (0.329) (0.053) (0.049)
[0.003] [0.275] [0.467] [0.673] [0.691] [0.238] [0.467] [0.313]
First Stage F-Statistic 1.60 1.55 1.54 0.25 0.25 1.45 1.54 1.54
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
J. No Trends
Years of Schoolingizy 0.075 0.032 -0.010 0.432 -0.066 -0.167 0.006 -0.013
(0.007) (0.009) (0.007) (0.138) (0.057) (0.036) (0.014) (0.013)
[0.000] [0.000] [0.158] [0.002] [0.248] [0.000] [0.682] [0.308]
First Stage F-Statistic 21.87 21.57 21.47 8.15 7.83 19.77 21.47 21.47
N 13,672 13,912 13,922 10,941 11,207 13,405 13,922 13,922
K. Only Zones in All DHS Rounds (25 of 30; 48 of 60)
Years of Schoolingizy 0.089 0.046 -0.015 0.276 -0.240 -0.338 -0.023 -0.040
(0.028) (0.027) (0.023) (0.292) (0.268) (0.186) (0.046) (0.038)
[0.002] [0.085] [0.515] [0.345] [0.370] [0.070] [0.618] [0.296]
First Stage F-Statistic 6.57 6.43 6.44 2.11 2.44 6.13 6.44 6.44
N 13,391 13,628 13,638 10,667 10,933 13,121 13,638 13,638
L. Less than 4,000 Organized Violence Deaths: 1989 to 91
Years of Schoolingizy 0.082 0.055 -0.021 0.249 -0.166 -0.344 -0.022 -0.049
(0.029) (0.032) (0.026) (0.303) (0.219) (0.196) (0.056) (0.048)
[0.005] [0.084] [0.403] [0.411] [0.448] [0.079] [0.703] [0.306]
First Stage F-Statistic 4.82 4.74 4.77 1.73 1.75 4.24 4.77 4.77
N 12,569 12,810 12,819 10,085 10,335 12,357 12,819 12,819
M. Less than 500 Organized Violence Deaths: 1989 to 91
Years of Schoolingizy 0.077 0.041 0.010 0.214 -0.258 -0.272 -0.008 -0.021
(0.030) (0.021) (0.023) (0.248) (0.339) (0.152) (0.046) (0.040)
[0.011] [0.055] [0.670] [0.389] [0.447] [0.073] [0.868] [0.607]
First Stage F-Statistic 6.97 7.00 6.95 2.55 2.00 8.59 6.95 6.95
N 9,677 9,908 9,913 7,826 8,020 9,539 9,913 9,913
N. No Regions of Highest Famine Concentration
Years of Schoolingizy 0.114 0.030 -0.010 0.203 -0.095 -0.203 -0.041 -0.054
(0.021) (0.020) (0.024) (0.174) (0.154) (0.111) (0.036) (0.032)
[0.000] [0.128] [0.694] [0.244] [0.534] [0.069] [0.262] [0.092]
First Stage F-Statistic 9.14 9.19 9.26 3.73 3.51 6.53 9.26 9.26
N 9,951 10,155 10,161 7,871 8,064 9,826 10,161 10,161
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: BMI = body mass index. The dependent variable is described at the top of each of the eight columns. In columns 1–3, 7,
and 8 it is an indicator that equals 1 if true; in columns 4 and 5 it is a standardized value of the described outcome; and in column
6 it is the count from 0 to 5 of acceptable reasons for domestic violence (going out without permission, neglecting children,
arguing with husband, refusing sex, burning food). Years of Schoolingizy is the predicted level of schooling, instrumented with
FPE . All sample and speciﬁcation deﬁnitions can be found in Section D.2
the free primary education (FPE) intensity measure, Izy
of this supplementary online appendix. Unless otherwise noted, all regressions include birth year and zone ﬁxed eﬀects, zone-
speciﬁc linear trends, and a cubic for age. Each estimate is from a unique regression. Standard errors are clustered at the zone
level and shown in parentheses; p-values are shown in square brackets.
A.56
Table D.7: Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 4
Sector of Work
Skilled / Service / Agriculture / Ideal Number Ideal Number Ideal Number
Working Professional Sales Unskilled Manual of Children of Children of Children
(1) (2) (3) (4) (5) (6) (7)
Non-MTI Regions – FPE Only National
(Separate) (Combined)
A. Baseline
Years of Schoolingizy 0.093 0.059 0.064 -0.048 -0.786 -0.923 -0.902
(0.058) (0.028) (0.047) (0.031) (0.468) (0.658) (0.493)
[0.107] [0.033] [0.169] [0.116] [0.093] [0.160] [0.068]
First Stage F-Statistic 6.06 6.63 6.63 6.63 6.63 6.60 19.37
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
B. Cohorts: 1968 - 1990
Years of Schoolingizy 0.076 0.052 0.040 -0.044 -0.661 -0.883 -0.834
(0.039) (0.023) (0.031) (0.030) (0.567) (0.400) (0.320)
[0.052] [0.023] [0.196] [0.146] [0.244] [0.027] [0.009]
First Stage F-Statistic 7.68 7.61 7.61 7.61 7.66 21.26 27.72
N 16,465 16,276 16,276 16,276 16,323 29,253 29,253
C. Cohorts: 1969 - 1989
Years of Schoolingizy 0.110 0.057 0.058 -0.032 -0.790 -0.839 -0.837
(0.056) (0.029) (0.042) (0.030) (0.737) (0.565) (0.440)
[0.050] [0.050] [0.166] [0.288] [0.284] [0.138] [0.057]
First Stage F-Statistic 6.67 7.20 7.20 7.20 6.76 12.01 26.80
N 15,093 14,927 14,927 14,927 14,963 26,832 26,832
D. Cohorts: 1971 - 1987
Years of Schoolingizy 0.115 0.081 0.071 -0.049 -1.073 -1.144 -1.110
(0.082) (0.043) (0.058) (0.048) (0.708) (0.583) (0.506)
[0.158] [0.060] [0.221] [0.305] [0.130] [0.050] [0.028]
First Stage F-Statistic 3.47 3.99 3.99 3.99 3.64 7.56 13.80
N 13,195 13,045 13,045 13,045 13,084 21,802 21,802
E. Cohorts: 1972 - 1987
Years of Schoolingizy 0.158 0.109 0.081 -0.061 -1.087 -0.950 -0.944
(0.107) (0.058) (0.067) (0.055) (1.848) (0.582) (0.571)
[0.140] [0.062] [0.229] [0.269] [0.556] [0.103] [0.098]
First Stage F-Statistic 2.70 3.19 3.19 3.19 2.88 6.56 13.07
N 12,863 12,715 12,715 12,715 12,754 21,177 21,177
A.57
Table D.7: (... continued) Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 4
Sector of Work
Skilled / Service / Agriculture / Ideal Number Ideal Number Ideal Number
Working Professional Sales Unskilled Manual of Children of Children of Children
(1) (2) (3) (4) (5) (6) (7)
National
(Separate) (Combined)
F. Using 1984 Census Matched Start Ages
Years of Schoolingizy 0.071 0.027 0.088 -0.061 -0.345 -1.116 -1.198
(0.051) (0.014) (0.026) (0.036) (0.518) (0.933) (0.472)
[0.160] [0.057] [0.001] [0.089] [0.505] [0.232] [0.011]
First Stage F-Statistic 6.22 5.79 5.79 5.79 6.30 1.99 6.74
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
G. Using 1994 Census Start Ages
Years of Schoolingizy 0.083 0.014 0.112 -0.028 -1.326 -1.333 -1.605
(0.056) (0.029) (0.048) (0.045) (2.144) (0.675) (0.794)
[0.139] [0.613] [0.020] [0.531] [0.536] [0.048] [0.043]
First Stage F-Statistic 1.23 1.36 1.36 1.36 1.49 3.45 18.70
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
H. Three-Part District Trends
Years of Schoolingizy 0.082 0.090 0.020 -0.008 -0.648 -0.728 -0.746
(0.060) (0.024) (0.046) (0.033) (0.446) (0.618) (0.426)
[0.171] [0.000] [0.670] [0.806] [0.146] [0.239] [0.080]
First Stage F-Statistic 17.90 16.29 16.29 16.29 17.91 5.89 18.70
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
I. Regional Trends
Years of Schoolingizy 0.133 0.078 0.083 -0.063 -1.163 -1.032 -0.946
(0.107) (0.065) (0.057) (0.066) (0.768) (0.715) (0.561)
[0.217] [0.229] [0.144] [0.344] [0.130] [0.149] [0.092]
First Stage F-Statistic 1.60 1.57 1.57 1.57 2.06 4.95 14.49
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
J. No Trends
Years of Schoolingizy 0.048 0.042 0.015 -0.026 -0.711 -0.675 -0.950
(0.016) (0.011) (0.011) (0.016) (0.220) (0.245) (0.311)
[0.002] [0.000] [0.185] [0.107] [0.001] [0.006] [0.002]
First Stage F-Statistic 21.77 21.00 21.00 21.00 23.08 13.20 21.45
N 13,909 13,755 13,755 13,755 13,789 24,649 24,649
A.58
Table D.7: (... continued) Eﬀect of Years of Schooling on Labor Market Outcomes and Fertility Preference:
Alternative Samples and Speciﬁcations - Analysis of Results from Table 4
Sector of Work
Skilled / Service / Agriculture / Ideal Number Ideal Number Ideal Number
Working Professional Sales Unskilled Manual of Children of Children of Children
(1) (2) (3) (4) (5) (6) (7)
National
(Separate) (Combined)
K. Only Zones in All DHS Rounds (25 of 30; 48 of 60)
Years of Schoolingizy 0.099 0.059 0.069 -0.047 -0.801 -0.920 -0.955
(0.058) (0.027) (0.044) (0.030) (0.633) (0.572) (0.477)
[0.090] [0.032] [0.115] [0.116] [0.206] [0.108] [0.045]
First Stage F-Statistic 6.57 7.17 7.17 7.17 6.63 7.73 21.68
N 13,625 13,473 13,473 13,473 13,507 24,232 24,232
L. Less than 4,000 Organized Violence Deaths: 1989 to 91
Years of Schoolingizy 0.113 0.045 0.067 -0.043 -0.315 -0.891 -0.728
(0.065) (0.028) (0.036) (0.041) (0.692) (1.486) (0.587)
[0.079] [0.113] [0.064] [0.292] [0.649] [0.549] [0.215]
First Stage F-Statistic 4.86 5.12 5.12 5.12 5.07 5.44 14.04
N 12,806 12,661 12,661 12,661 12,710 22,982 22,982
M. Less than 500 Organized Violence Deaths: 1989 to 91
Years of Schoolingizy 0.125 0.014 0.077 0.005 -0.069 -0.377 -0.693
(0.038) (0.026) (0.046) (0.041) (0.403) (0.675) (1.059)
[0.001] [0.595] [0.094] [0.912] [0.864] [0.576] [0.513]
First Stage F-Statistic 7.12 7.32 7.32 7.32 7.57 4.51 9.68
N 9,904 9,782 9,782 9,782 9,859 18,262 18,262
N. No Regions of Highest Famine Concentration
Years of Schoolingizy 0.127 0.042 0.082 -0.053 -0.531 -0.432 -0.604
(0.052) (0.018) (0.035) (0.032) (0.422) (1.466) (0.596)
[0.015] [0.023] [0.019] [0.101] [0.208] [0.768] [0.311]
First Stage F-Statistic 9.43 9.71 9.71 9.71 10.73 6.02 11.87
N 10,152 10,022 10,022 10,022 10,068 18,562 18,562
Source: Author’s analysis based on data from the Ethiopian Demographic and Health Survey (in years 2005, 2011, and 2016).
Note: MTI = mother tongue instruction; FPE = free primary education. The dependent variable is described at the top of
each of the ﬁve columns. In columns 1–4 it is an indicator that equals 1 if true, and in columns 5–7 it is the ideal number
of children. Skilled/Professional jobs include professional, clerical, and skilled manual job groups; the other categories exactly
describe the occupation groups included. Ideal number of children is censored at 20; no women in the Demographic and
Health Survey report having more than 18 children, and non-numerical responses are assigned the maximum value. In columns
1–5, Years of Schoolingizy is the predicted level of schooling instrumented with the FPE intensity measure Izy FPE ; in column
6 additional instruments include two MTI intensity measures Izy MTI-T and I MTI , which denote the measures for MTI regions
zy
without and with script change, respectively, and the interactions for regions in which two interventions occurred; in column 7 the
joint intensity measure ∆Izy is used. All sample and speciﬁcation deﬁnitions can be found in Section D.2 of this supplementary
online appendix. Unless otherwise noted, all regressions include birth year and zone ﬁxed eﬀects, zone-speciﬁc linear trends,
and a cubic for age. The ﬁrst ﬁve columns include only observations from non-MTI regions; all regions are included in the ﬁnal
two columns. Each estimate is from a unique regression, and the second-stage estimate in column 5 is generated using a tobit
model. Standard errors are clustered at the zone level and shown in parentheses; p-values are shown in square brackets.
A.59