The effect of mother-tongue instruction on educational attainment in Ethiopia

(1)

The effect of mother-tongue

instruction on educational attainment

in Ethiopia

Lena Van Meensel

MSc Economics - University of Amsterdam Under the supervision of Prof. Dr. Hessel Oosterbeek

Abstract

This paper empirically examines the effect of mother-tongue instruction during primary school on the enrolment rate and the years of schooling. Despite its widespread diver-sity, Ethiopia has a long history of monolingual policies of education. In 1994, when the Educational and Training Policy was signed into law, the country was introduced to mother-tongue instruction. The reform led to an exogenous source of variation in access to mother-tongue instruction across states and birth cohorts, which allows for a difference-in-differences identification strategy. The estimates from the preferred specification suggest that the change in the medium of instruction increased the years of schooling of the affected cohort by 0.245 years. This increase can at least partially be attributed to changes on the extensive margin, as the enrolment rate is estimated to increase by 2.4 percentage points. Results from alternative specifications indicate that the findings are robust to modifica-tions in the sample of interest. The present study therefore supports the argument that mother-tongue instruction enhances the educational attainment in multilingual develop-ing countries. Nevertheless, the effect is heterogeneous across gender; whereas women are positively impacted by mother-tongue instruction, men’s years of schooling and enrolment rate decrease. Additional research on the mechanisms at play for each gender needs to be conducted to pinpoint the cause of this heterogeneity.

(2)

Statement of originality

This document is written by student Lena Van Meensel, who declares to take full respon-sibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

I. Introduction

In pursuit of the second Millennium Development Goal of the United Nations, which strives for universal primary education by 2015, governments worldwide have expanded their ef-forts to ensure access to primary education for all. As a result, the enrolment rate in developing regions climbed from 83% in 2000 up to 91% in 2015. Despite this remarkable progress,UNICEF and UNESCO(2015) concludes that the Millennium Development Goal has not been met; by 2015, an estimated 58 million children of primary school age are out of school. What is even more worrisome, is the geographic concentration of these children. A relatively small number of countries, the majority of which are located in Sub-Saharan Africa, accounts for a disproportionately large number of out of school children. While some developing regions have nearly achieved universal primary education, Sub-Saharan Africa continues to lag behind in terms of primary school enrolment and attainment, partially due to a rapidly increasing school-age population.

Many different factors contribute to Sub-Saharan Africa’s poor performance in terms of educational attainment and enrolment. The present study examines the role of the language policy of the education system. Despite being characterised by multilingualism, the majority of countries in Sub-Saharan Africa use the language of the government or the dominant ethnic group as the unique medium of instruction (MOI) during primary school (UNICEF and UNESCO, 2015). This results in a large group of children who do not have access to education in their mother tongue and who are forced to learn in a second language upon enrolment. According to several pedagogical publications (see e.g.Mehrotra,

1998; Trudell, 2005; Heugh et al., 2007), this linguistic mismatch poses a threat for the accessibility and equity of the education system. First, language-minority children might face a greater barrier for enrolling in primary school if all communication is in a language they are not familiar with. Second, there is a correlation between being part of a language minority and an individual’s socioeconomic status; households of language minorities are significantly more prone to poverty than others. Any policy that impedes them from attending school can be perceived as inequitable and might lead to economic divergence within a country. The publications thus identify the language education policy, particularly the language of instruction, as a key factor in achieving universal primary education. In

(5)

this context, a number of developing countries have attempted to accommodate language minorities by implementing mother-tongue education (MTE).1

Together with the landmark publication ofUNESCO(1953), which strongly campaigns for mother-tongue education, this trend in developing countries has increased the academic interest in the education language policy over the last decades. The existing literature is predominantly focused on the relationship between MTE and the quality of schooling. Classroom observations by Taylor and Vinjevold (1999) and Trudell (2005) demonstrate the positive link between MTE and classroom participation; mother-tongue instruction facilitates the communication between the teacher and his or her students and allows for more student-centred learning. Trudell (2005) also argues that, as students no longer need to learn a second language and new subject content simultaneously, MTE results in improved academic performance. This statement is confirmed in several papers, in a variety of both developed and developing countries (see e.g. Walter and Dekker, 2013;

Ivlevs and King,2014;Eriksson,2014;Hynsj¨o and Damon,2016;Piper et al.,2016;Laitin et al.,2016). An issue of concern regarding MTE is its effect on national and international language proficiency. Often, the second and third languages primarily used as MOI are taught as a separate subject under mother-tongue education. Opponents of MTE claim that, as students are less exposed to other languages, MTE might result in lower language competence. However, a large body of evidence suggests that MTE not only has a positive effect on the native language skills, but also enhances the proficiency in second and third languages (Walter and Dekker,2013;Eriksson,2014;Hynsj¨o and Damon,2016;Laitin et al.,

2016). Another strand of literature focuses on the impact of multilingual education on long-term economic outcomes. The bulk of research finds a positive relationship between MTE and labour market outcomes; through improved academic performance and language skills, MTE results in a increased probability of being employed, increased wages and increased job satisfaction, even after controlling for educational attainment (Eriksson, 2014; Seid,

2017; Cappellari and Di Paulo, 2018). An exception to this is the article by Angrist and Lavy(1997), which studies the change in the medium of instruction from French to Arabic

1

MTE is defined as the use of a variety of languages as medium of instruction throughout the country, so that the classroom language matches the home language of the students (UNESCO,1953)

(6)

in Morocco. They find negative effects on both academic achievement and later labour market outcomes. However, Arabic is not the native language of the Berber ethnic group, which is the second largest ethnicity in Morocco. The study thus merely examines the impact of a language change rather than the impact of MTE.

Although a large body of literature is thus readily available on the effect of MTE on several outcome variables, little quantitative research has been dedicated towards exam-ining whether this is an effective policy in reaching universal primary education. This study exploits an education policy reform in Ethiopia to investigate the causal effect of the mother-tongue education on years of schooling and primary school enrolment.

Before 1994, Amharic was the unique official language of the Ethiopian federal gov-ernment. This means that all official institutions, including the formal schooling system, used Amharic as the sole medium of communication. In 1994, after a new government came to power, the Training and Education Policy was signed into law. Amongst other things, the policy allowed state governments to adopt regional languages as the MOI dur-ing primary school in their jurisdiction. This discretion led to a change of the language of instruction in most states. An exception is Amhara state, where both before and after the reform Amharic was used as the MOI, as this is the mother tongue of the majority of the state’s residents.2 The reform thus had a differential impact across birth cohorts and states. First, only children who entered school after 1994 were exposed to MTE; students who had already graduated from primary school before 1994 attended their com-plete primary schooling in Amharic and were not impacted by the reform. Second, the reform did not affect individuals in Amhara, as there was no change in the medium of instruction. This variation in exposure to mother-tongue instruction allows us to use a difference-in-differences approach as the estimation strategy.

Two recent studies,Eriksson(2014) andSeid(2016), have made important advances in this field and relate closely to my study. Eriksson(2014) examines a change in the language of instruction in South-Africa, where the duration of mother-tongue instruction increased

2

The zones Oromiya, Wag Hemra and Awi changed the language of instruction to Oromiffa, Agew and Awngi, respectively, which are the native languages of a majority of the zones’ inhabitants. All other zones continued the use of Amharic as medium of instruction (Heugh et al.,2007).

(7)

from 4 years to 8 years in 1955. The prolongation of MTE increases the years of schooling by 0.10 years for men and by 0.18 years for women. However, it is debatable whether the results from the study are generalisable to the Ethiopian context. First, Eriksson

(2014) examines the effect of longer exposure to MTE during primary school, whereas the present study focuses on an introduction of MTE. Second, South-Africa is significantly more developed than Ethiopia, even when the timing of each reform is taken into account. Not only is the GDP per capita in South-Africa higher by the time of the reform, also the initial enrolment rates and years of schooling at baseline exceed those of Ethiopia.3 In contrast, Seid (2016) focuses on the Ethiopian context, which eliminates any external validity concerns. The author finds that the introduction of MTE positively influences the probability of enrolment and the probability of being in the correct grade-for-age. The effect on years of schooling, however, is not examined. Furthermore, the reported results should be interpreted with caution as there is no analysis of the identifying assumption of the estimation strategy. As the author applies a difference-in-differences approach, thorough investigation of the common trend assumption is necessary, yet missing from the paper. In conclusion, my paper makes two main contributions to the existing literature. To my best of knowledge, it is the first quantitative research on the effect of mother-tongue education on years of schooling in a Sub-Saharan African country. As the region still struggles with low educational attainment and enrolment, the results of my research are valuable for further policy recommendations. In addition, the data used in the paper allows for close examination of the identifying assumptions of the estimation strategy, which strongly strengthens the causal interpretation. The results indicate that mother-tongue instruction leads to an increase in the mean years of schooling, which is at least partially caused by an rise in the enrolment rate.

The remainder of the paper is organised as follows. The next section provides back-ground information on the educational and linguistic history of Ethiopia. SectionIII dis-cusses the channels through which mother-tongue instruction theoretically impacts years

3

According to the World Bank data, the South-African GDP per capita was 434 USD in 1960, while the Ethiopian GDP was 125 USD in 1994. At the time of the reform, the percentage of the population without any education was 38.5% and 71.3% in South-Africa and Ethiopia, respectively.

(8)

of schooling and enrolment. SectionIV describes the data used in the paper. Section V

presents the identification strategy and the empirical model. SectionVI discusses the re-sults, including additional robustness checks, and examines the heterogeneity of the effect along gender and rurality. SectionVIIconcludes.

II. Educational and linguistic history of Ethiopia

According to the 1994 census, Ethiopia has a population of 63 million citizens, of whom 12 million are of primary school age (7 to 14 years old) and 5 million of secondary school age (15 to 18 years old). The population is highly diverse, compromising over 80 officially recognised ethnic groups. As almost every ethnicity features its own language, the country is characterised by multilingualism.

Despite its widespread diversity, Ethiopia has a long history of monolingual policies of education. Under the reign of Emperor Haile Selassie (1941-1794), there was a strong political emphasis on bringing unity to the country and establishing a centralised federa-tion. To this end, Amharic was implemented as the official language of all governmental institutions, including formal elementary schools. Amharic is the first language of the Amharas, an ethnic group that holds the reins of economic and political power, despite being a minority group in terms of population. In secondary schools, English served as the language of instruction.

Political turmoil and social unrest led to an overthrow of the Emperor by the mili-tary socialist group Derg. Under the Derg administration (1974-1991), there were several attempts of introducing additional languages to the education system. However, these attempts have only proven to be successful in non-formal education programs. Hence, Amharic remained the only national language and the unique MOI in elementary schools. Simultaneously, growing discontent regarding the socialist regime led to the establish-ment of several ethnically-based opposition parties.4 The famines of 1984-1985, which are

4

The Tigray People’s Liberation Front (TPLF), the Oromo Liberation Front (OLF) and the South Ethiopian Peoples Democratic Coalition (SEPDC) were formed and respectively protected the rights of the state Tigray, the state Oromiya and the southern states.

(9)

Table 1: Language education policy as stipulated in Educational and Training Policy

Medium of instruction Seperate subjects

Primary school Mother tongue Amharic

English

Secondary school English National or foreign languages

estimated to have impacted over one million citizens, fueled further uprisings against the ruling power. The ethnic parties joined forces and formed the coalition party Ethiopia’s People Revolutionary Democratic Front (EPRDF), which strongly promotes increased cul-tural and political autonomy for all ethnic groups. When the EPRDF toppled the Derg regime in 1991, ethnic equality was put on the forefront of political discussions. The new government formally introduced minority languages to the education system when the Education and Training Policy of 1994 was signed into law. Amongst other things, the new policy allowed state governments to adopt regional languages as the MOI in primary schools within their jurisdiction. Table1shows the language education policy as stipulated in the Education and Training Policy. During primary school, the mother tongue is used as the language of instruction, while Amharic and English are taught as separate subjects.5 English continued to be the medium of instruction in secondary schools.

Heugh et al. (2007) thoroughly investigates whether the policy was correctly imple-mented and whether the state governments followed the guidelines of the new policy. Al-though some states only partially implemented mother-tongue instruction, for example for six years instead of the full eight years of primary school, the authors conclude that nearly every state adapted their language policy in response to the reform. For more informa-tion on each state’s revised policy, we refer to AppendixA. Furthermore, the authors find that the reform had an immediate impact; rather than a gradual transition, most states

5

The most commonly spoken language of the region is used as the new medium of instruction. It is plausible that, due to large ethnic diversity, this is not the home language of every child in the region. However, the use of the lingua franca is identified as the best practice in providing mother-tongue education (see e.g.Mehrotra,1998;Trudell,2005)

(10)

introduced a direct change in the language of instruction after the signing of the Edu-cation and Training Policy in 1994. This means that the reform divides the population into three groups, which differ in exposure to mother-tongue instruction. Children who were below primary school age and had yet to enter primary school in 1994 were fully exposed to mother-tongue instruction, whereas older age cohorts, who graduated primary school prior to the reform, were not exposed at all. The children that were enrolled in primary school when the reform took place were instructed in Amharic during their first years before switching to mother-tongue instruction; hence, they were partially exposed. An exception to this is found in Amhara, where both before and after the reform Amharic was used as the MOI (excluding the zones Oromiya, Wag Hemra and Awi). The reform thus did not affect the language policy of the state; both before and after 1994, Amharic students had access to MTE. The variation in exposure to MTE across birth cohorts and across states allows us to examine the causal effects of mother-tongue instruction on diverse outcomes. It is worth noting that, while the introduction of mother-tongue instruction was only one aspect of the broader Education and Training policy, other aspects did not have a differential impact across states (Ramachandran,2012).

III. Theoretical framework

In this section, I examine how MTE is expected to affect years of schooling. The theory of change can be formulated by closely examining the determinants of educational choices. To this end, I make use of a basic model on educational choices presented byDuflo(2009). The model is build on three assumptions. First, parents make schooling decisions for their children. Second, education is defined as an investment good, with benefits and costs, but without consumption value. That it, education itself does not enter the utility function. Finally, the model assumes a static set-up. Although the benefits from education accrue over a longer period when life expectancy is higher, which increases the total benefits, the intertemporal dimension is ignored.

(11)

III.A. Model of educational choice

Parents make educational choices for their children by optimising the following utility function, which is a function of the child’s years of schooling (S) and the child’s future earnings (y):

U (S, y) = m × log y(S) − c(S), (1)

with m the value that parents put on the child’s future earnings, y(S) the future earnings when completing S years of schooling and c(S) the cost of S years of schooling.

Following a simple Mincerian framework, the model assumes that there are constant economic returns to education, meaning that each additional year of schooling increases the future wage by a fixed percentage. The earnings equation is thus the following:

log y(S) = a + b S (2)

where b expresses the returns to schooling. For the cost function, the model includes two different types of costs. The first type concerns the direct costs of schooling, such as tuition fees, materials required and travel expenses. The direct costs are assumed to be constant per additional year of schooling. The second component refers to the opportunity cost of time, namely the foregone wages that could have been earned if the child was working instead of studying. It is reasonable to assume that these opportunity costs increase with age and education level, which makes this cost component convex with respect to the years of schooling. Combining the two types of costs results in the following cost function:

c(S) = r S + 1 2φ S

2 ₍₃₎

with r the cost per additional year schooling and φ the opportunity cost of time per year of schooling.

Plugging the wage equation (2) and the cost equation (3) into the parents’ utility function (1), the optimisation problem becomes

max

S U = m × (a + b S) − r S −

1 2 φ S

(12)

The first order condition, which equalises the derivative of the utility function with respect to S to zero, gives an expression to the optimal years of schooling:

S∗= m b − r

φ (5)

The presented model therefore identifies the value that parents put on child’s earnings m, the returns to schooling b, the direct costs r and the opportunity costs φ as the determinants of educational choices. Therefore, these are the channels through which MTE can impact the years of schooling.

The relationship between each of the determinants and mother-tongue instruction can be evaluated separately. It is, however, unlikely that the language policy influences the parental valuation of the child’s future earnings or the opportunity cost of time. The former depends on social characteristics, such as how dependent elderly are on old-age support from their children. The latter depends on the institutional context concerning child labour. Accordingly, I assume that these determinants remain constant in response to a change in the MOI. If mother-tongue instruction influences schooling decisions, it is either through its impact on the directs costs or the returns on schooling. Both channels are discussed separately below.

III.B. Change in the direct costs

The direct costs of schooling can be roughly divided into two categories: monetary costs and non-monetary costs. Both types of costs can be affected by the introduction of MTE. With respect to monetary costs, the adoption of mother-tongue instruction is often considered a costly reform. It is argued that the language of wider communication, such as Amharic in the Ethiopian context, has a comparative cost advantage over minority languages when used as MOI; a standardised language policy can generate economies of scale concerning the training of teachers and the provision of materials (Clayton,1998). In the existence of these economies, a monolingual policy can be regarded appropriate. There is, however, no conclusive evidence on whether a monolingual approach is de facto cost-saving. Rubin and Jernudd(1971) attempt to quantify the cost of both approaches and find that mother-tongue education is more expensive. A similar analysis byGrin(2005) states

(13)

the opposite and concludes that a multilingual approach is in fact cost-saving, especially in the long run. Whether MTE is consistently more expensive than a monolingual system, is therefore open for discussion. However, there is a consensus that any change in the language policy will require at least some non-recurring costs. For the implementation of MTE, this refers to the production of materials in new languages as well as additional training of teachers to prepare them for the educational change (Heugh et al.,2007).

There are several options to cover these initial investments. First, state governments can increase the yearly tuition fee of public primary schools, which directly increases the public revenue. The advantage of this option is that the generated surplus reoccurs yearly, until the tuition fee is reduced again. Especially if MTE appears to be the more costly approach in the long term, this is a desired feature. However, this would mean that the direct cost of a year of schooling goes up, which, according to the found optimal (5), decreases the optimal years of schooling. A second option is that, if the tuition fees and therefore the public revenue remain constant, the government allocates a larger part of the national budget to educational expenses. Although this would result in other public spending cuts, the burden of the costs would no longer fall on the students’ parents, leaving the optimal years of schooling unchanged. A final option, rather than shifting funds between domains of public spending, is to keep the total educational budget constant and to shift funds between the different components of educational spending. The costs of adopting MTE are then compensated by economies in other educational components. However, a review of the World Bank (2014) with regard to the Ethiopian education spending concludes on “insufficiency of spending at all levels”. Further budget cuts are explicitly discouraged. With respect to the Ethiopian context, a combination of the first two options took place. While the explicit tuition fee even decreased,Heugh et al. (2007) reports that parents took part of the cost burden as some costs were no longer included in the tuition fee, such as the purchase of new textbooks. Nonetheless, the budget for education increased steadily and heavily since 1995 (Ramachandran, 2012). Therefore, I assume that the initial investments were largely covered by government funds and that the cost burden only minimally fell on the parents’ shoulders.

(14)

The second component of the direct costs, the non-monetary costs, mainly refers to the cost of effort required to succeed in a schooling programme. While there is little quantitative research available on the non-monetary component, pedagogical publications byTaylor and Vinjevold(1999) andTrudell(2005) argue that MTE lowers the physic costs, as students no longer need to learn a second language and subject content simultaneously. Moreover, the authors state that mother-tongue instruction improves the communication between the teacher and his or her students, which facilitates the knowledge transfer within the classroom; classroom hours become more efficient.

Combining the assumption that parents were only minimally impacted by the increased monetary cost with the finding that the non-monetary costs decreased, my hypothesis is that the introduction of MTE leads to a decrease in the direct costs of schooling. Following the found optimal level of schooling (5), this would ceteris paribus lead to an increase in the years of schooling.

III.C. Change in the returns on schooling

There are a variety of factors that influence the returns on schooling. One of them is the quality of schooling: high-quality schools facilitate the acquisition of human capital, which in turn improves the students’ position on the labour market. One the other hand, if schools fail to educate their students properly, the benefits of schooling will be lost and its value will decrease. This potentially results in a situation where the average years of schooling are minimal: without substantial benefits, yet considerable costs, an individual has very little incentive to enrol in an education programme. A growing body of literature suggests that mother-tongue instruction improves the quality of education. First, overall schooling methods are improved. Qualitative research byTaylor and Vinjevold(1999) and

Trudell (2005) finds that MTE leads to enhanced classroom participation and allows for more student-centred learning. Second, even if the classroom practices remain unchanged, it potentially impacts the student’s learning experience. Trudell (2005) find that students in mother-tongue classes are more capable to understand the materials presented by the teacher. The study also argues that, as students no longer need to learn a second language and new subject content simultaneously, MTE boosts the students’ motivation. Several

(15)

quantitative papers confirm that mother-tongue instruction increases the quality of school-ing and find that it results in improved academic performance (see e.g.Walter and Dekker,

2013; Eriksson, 2014; Piper et al., 2016). In conclusion, MTE enhances the quality of schooling, which in turn is positively linked with the returns on schooling.

It is, however, important to note that what matters for educational choices, is the per-ceived returns on schooling. Rather than taking the actual returns into account, parents act on what they believe the returns are. According to Heugh et al. (2007), parents in Ethiopia commonly held the belief that the use of minority languages would diminish the value of schooling. More specifically, they were not clear on the benefits of using a language that was already known to the students, rather than learning a second language through its function of MOI. Despite scientific research showing that second language acquisition ac-tually improves through MTE (see e.g.Walter and Dekker,2013;Eriksson,2014), parents expressed their concerns about the effect of MTE on both Amharic and English proficiency. In a first stadium, the resistance towards mother-tongue instruction of key stakeholders re-sults in complications for the implementation, but even if the implementation is complete, it may have lasting effects. The negative attitude towards mother-tongue instruction po-tentially results in sub-optimal investment in schooling. If the negative effect on perceived returns overrules the positive effect on actual returns, MTE may even result in lower years of schooling.

From former literature, it is unclear exactly how widespread and persistent the resis-tance towards MTE is. First, parents were significantly more negative in states with closer economic and social ties to Amhara, possibly due to the importance of Amharic in later professional communication. In more distant states, the response was rather neutral or positive. Second, while some expressed their concerns regarding linguistic skills, the beliefs regarding numeracy skills were almost exclusively positive. Third, parents that originally expressed their concerns can later change their attitude if they witness others students’ positive experience with MTE. Taking all arguments into account, my hypothesis is that mother-tongue instruction positively influences the returns on schooling through its impact on the schooling quality, especially in the long run.

(16)

IV. Data

IV.A. Data sources and sample of interest

The analysis in the present paper uses data of the Population and Housing census of Ethiopia, administered by the Central Statistical Agency (CSA) of Ethiopia and the Min-nesota Population Center. More specifically, the data is taken from the most recent survey wave of 2007. The sampling of the census was done on the household level, with one in ten households, with a random start, included in the sample. A total of 1,567,834 households were surveyed.

Of the surveyed households, 80% received the short-form questionnaire, whereas the remaining 20% received the long-form questionnaire. As the names imply, the main dif-ference between the two types of questionnaire is the number of questions included. The former collects information on basic demographic and social characteristics, whereas the latter contains additional questions on education, disability status, economic activity and housing conditions. As the present study aims to examine the effect of MTE on educational attainment, only the long-form questionnaire contains all the required information, which restricts the sample of interest to the recipients of this type of questionnaire.

A second restriction concerns the age of the respondents. As younger respondents are also included in the survey, there is a possibility that they have not finished school at the time of the data collection in 2007. In this case, the years of schooling are underestimated for the younger cohort. To avoid issues arising from censored data, I drop all observations of individuals younger than 18 years old in 2007. At this age, assuming that everyone is in the correct grade for his or her age, the formal schooling curriculum is finished and every child is potentially graduated from secondary school.6

Although the census has national coverage, I only use data from Amhara and its neigh-bouring state Tigray. This is primarily because the reform had a differential impact across

6

It is possible that, upon graduating secondary school, some individuals pursue a tertiary degree. Taking this into account would limit the sample to individuals older than 23 in 2007. However, this restriction would eliminate any individuals who were yet to enter primary school at the time of the reform, as this group was younger than 7 in 1994 and younger than 20 in 1994. Since only a very limited group obtains tertiary degrees, less than 3% according to the 2007 census, this restriction is not imposed.

(17)

the states. As explained in SectionII, Amhara was the only state that continued its original language policy after the implementation of the new policy, which makes it adequate to serve as the control state. In contrast, the state government of Tigray introduced a change in the language of instruction from Amharic to Tigrinya. Hence, Tigray is assigned the status of treatment states. Further details on the choice of Tigray as the treatment state, rather than another state that adapted its language policy, are discussed in SectionV.

IV.B. Descriptive statistics

Table2presents the descriptive statistics for Amhara and Tigray for the exposed, partially exposed and unexposed birth cohorts.7 The characteristics of the latter, presented in Column 5 to 8, can be regarded as baseline characteristics, as the individuals included are not affected by the reform.

With regard to demographic characteristics, the table illustrates large disparities be-tween Amhara and Tigray. In general, there were more female respondents in Tigray, where the proportion of women reached 61% for the unexposed cohort.8 _{Subsequently, there is a} higher proportion of female headed households in Tigray; at baseline, one in five individu-als lives in a female headed household in Tigray, compared to one in eight in Amhara. The majority of individuals in both states are resided in rural areas with, generally speaking, poor housing conditions; most households do not have access to electricity, piped water or a radio, although the households assets and amenities are significantly better in Tigray compared to Amhara. While the characteristics of the exposed and the partially exposed cohort are different from those at baseline - the proportion of women decreases, the proportion of female headed households increases and less households are resided in rural areas -the differences between -the states remain comparable across birth cohorts.

7

The exposed cohort consists of individuals below primary school age in 1997 (below 20 in 1994). The partially exposed cohort consists of individuals aged between 8 and 14 in 1994 (between 20 and 27 in 2007). The unexposed cohort consists of individuals aged between 15 and 17 in 1994 (between 28 and 30 in 2007).

8

The gender imbalance in Tigray might be due to the Ethiopian-Eritrean war of 1998-2000. As Tigray is located near the Eritrean border, it is likely that the region was more affected by the conflict. Literature on gender imbalances lists war causalities as one of the factors influencing the human-sex ratios, as men are more likely to be lethally injured in armed conflicts (Kvasnicka and Bethmann,2007).

(18)

Table 2: Descriptive statistics by exposure and treatment status

Exposed cohort Partially exposed cohort Unexposed cohort

Amhara Tigray Difference Amhara Tigray Difference Amhara Tigray Difference

Demographic characteristics

Age 19.01 19.00 -0.003 24.08 23.94 -0.133 29.22 29.22 0.003

(0.01) (0.01) (0.014) (0.01) (0.02) (0.025) (0.01) (0.01) (0.016)

Female 0.53 0.56 0.032 0.53 0.56 0.034 0.55 0.61 0.059

(0.00) (0.01) (0.007) (0.00) (0.01) (0.007) (0.00) (0.01) (0.009)

Female household head 0.21 0.29 0.082 0.17 0.25 0.078 0.13 0.19 0.068

(0.00) (0.01) (0.006) (0.00) (0.01) (0.005) (0.00) (0.01) (0.006) Household size 4.67 4.89 0.219 4.32 4.40 0.083 4.72 4.68 -0.040 (0.01) (0.03) (0.031) (0.01) (0.02) (0.026) (0.01) (0.03) (0.032) Rural 0.79 0.73 -0.062 0.83 0.73 -0.106 0.87 0.77 -0.094 (0.00) (0.01) (0.006) (0.00) (0.01) (0.005) (0.00) (0.01) (0.006) Access to electricity 0.17 0.26 0.095 0.14 0.27 0.137 0.11 0.22 0.113 (0.00) (0.01) (0.006) (0.00) (0.01) (0.005) (0.00) (0.01) (0.006)

Access to piped water 0.28 0.48 0.198 0.26 0.49 0.230 0.22 0.45 0.227

(0.00) (0.01) (0.007) (0.00) (0.01) (0.006) (0.00) (0.01) (0.008) Owns a radio 0.33 0.40 0.069 0.34 0.43 0.091 0.31 0.39 0.078 (0.00) (0.01) (0.007) (0.00) (0.01) (0.006) (0.00) (0.01) (0.008) Educational outcomes Enrolment 0.48 0.62 0.135 0.32 0.53 0.211 0.23 0.34 0.114 (0.00) (0.01) (0.007) (0.00) (0.01) (0.006) (0.00) (0.01) (0.008) Years of schooling 2.95 3.99 1.034 2.13 3.75 1.622 1.34 2.15 0.801 (0.02) (0.05) (0.057) (0.02) (0.05) (0.052) (0.02) (0.06) (0.057)

Years of schooling conditional 6.18 6.50 0.317 6.69 7.10 0.406 5.97 6.30 0.332

on enrolment (0.03) (0.05) (0.059) (0.04) (0.06) (0.071) (0.06) (0.10) (0.118)

Mother’s years of schooling 0.59 0.54 -0.050 0.59 0.42 -0.165 0.62 0.30 -0.318

(0.03) (0.05) (0.065) (0.05) (0.06) (0.081) (0.12) (0.12) (0.191)

Father’s years of schooling 1.20 1.36 0.158 1.24 1.12 -0.114 1.25 0.75 -0.497

(0.04) (0.07) (0.085) (0.06) (0.08) (0.108) (0.17) (0.14) (0.261)

Literate 0.50 0.63 0.133 0.34 0.55 0.213 0.25 0.35 0.108

(0.00) (0.01) (0.007) (0.00) (0.01) (0.006) (0.00) (0.01) (0.008)

(19)

Moreover, Table2sheds light on the poor educational outcomes in Ethiopia, as well as the differences between Amhara and Tigray regarding these outcomes. First, enrolment, defined as ever being enrolled in a formal education program, is poor. The enrolment rate at baseline is 23% and 34% in Amhara and Tigray, respectively. This means that close to three out of four children will never set foot in a classroom, which is surprisingly low, even for Sub-Saharan African standards. Second, the average years of schooling at baseline are meagre and well beneath the years of schooling required to finish primary school. There is a large disparity between Amhara, where the average years of schooling is 1.34 years, and Tigray, where children attend school for an average of 2.15 years. The observation that the years of schooling are low is not surprising, considering the poor enrolment rate and thus the high number of respondents with zero years of schooling. When dropping all zero observations and focusing on individuals that attended formal schooling, the average years of schooling increases up to 5.97 for Amhara and 6.30 for Tigray, which, however, is still insufficient for obtaining a primary school degree. The poor enrolment rate and completed years of schooling result in a literacy rate of only 25% in Amhara and 35% in Tigray. When looking at parental educational attainment, I find that parents are, in general, less educated. Fathers and mothers have completed approximately 0.77 and 0.33 years of schooling, respectively. From this observation, it is implied that, despite the baseline years of schooling being rather unsatisfactory, the education level has been increasing over the last generation. This statement is also confirmed when focusing on the educational outcomes for the exposed and partially exposed cohorts; despite the educational outcomes for the younger cohorts continue to be low on a global scale, they consistently surpass the baseline educational outcomes.

V. Empirical strategy

The exogenous variation created by the 1994 reform allows for the use of a difference-in-differences approach as identification strategy.9 First, there is a differential impact across

9

A necessary assumption is that there is no endogenous migration in response to the reform, as this would violate the random assigned to the control or treatment state and bias the estimates. There is

(20)

states. Amhara (excluding zones Oromiya, Wag Hemra and Awi) continued the use of Amharic as MOI, and therefore was not affected by the reform. Other states, however, experienced a sudden change in their language policy in response to the reform. This means that the former can be assigned the status of control state, while the latter are assigned the status of treatment states. Second, only specific birth cohorts were affected. As the reform only concerned the language policy of primary schools, only children of primary school age or younger had access to mother-tongue instruction. The majority of their older peers were unaffected by the reform, as they potentially finished the primary school curriculum before the policy was signed into law. Hence, the older cohort is assigned the status of unexposed birth cohort, while their younger counterpart is assigned the status of exposed birth cohort. To estimate the years of schooling, I assume a linear relationship between the years of schooling and the exposure to MTE:

Sijm = β1+ β2T reatedm+ β3Exposedj+ β4(T reatedm× Exposedj) + β Xi+ i (6)

where Sijm denotes the years of schooling of individual i born in year j, living in state

m; T reatedm is a binary variable that takes value 1 for the treated states and zero

oth-erwise; Exposedj is a binary variable that takes value 1 for the exposed birth cohort and

zero otherwise and Xi is a vector of individual control variables. The interaction term,

T reatedm× Exposedj, is the primary variable of interest; the corresponding coefficient β4

expresses how much the evolution in the years of schooling in the treated state deviates from the evolution in the control state.

If MTE influences the years of schooling, hence β4is significantly different from zero, it

is interesting to examine both the intensive and extensive margin. If the changes occur on the intensive margin, the number of children attending school remains equal, but those who do attend do so for a longer period of time. It is, however, not possible to formally examine the intensive margin; in order to do so, one needs to analyse the change in the schooling decisions of students that would attend primary school with and without mother-tongue

no reason to believe this assumption is invalid, as all states offer mother-tongue instruction following the reform.

(21)

instruction. Since it is unknown whether students would have enrolled in absence of MTE, this group cannot be isolated. It is, however, also possible that the years of schooling of those enrolled remains unchanged, but that more children enrol primary school in the first place. This decreases the number of zero observations, which in turn increases the average years of schooling. This is referred to as the change on the extensive margin. In contrast to the intensive margin, the extensive margin can be formally examined by analysing the effect of MTE on enrolment rates. In my analysis, enrolment is defined as a binary variable that indicates if an individual ever attended formal schooling and is modelled through the following specification:

P (Eijm = 1) = α1+α2T reatedm+α3Exposedj+α4(T reatedm×Exposedj)+αXi+i (7)

where Eijm expresses the enrolment of individual i born in year j, living in state m.10 The

other variables are the same as in the specification of the years of schooling. Again, the interaction term T reatedm× Exposedj is the variable of interest.

In the following subsections, the division of control and treatment state, as well as exposed and unexposed birth cohort, are discussed in more detail. The results of the models for years of schooling and enrolment are presented in SectionVI.

V.A. Control and treatment state

Ethiopia is a widely diverse country with large disparities between states. This diver-sity does not necessarily poses a threat for the application of the difference-in-differences method, as the method does not require the control and treatment state to be equal at baseline. Rather, the identifying assumption is that there is a parallel trend between the control and treatment state in absence of the treatment; that is, both groups would have evolved in the same manner if the reform did not take place. In this regard, is it required to carefully examine which states are appropriate for statistical analysis.

10

Since Eijmis a binary variable, the use of a linear specification is often discouraged, as it may lead to

predictions beyond the range [0, 1]. Nonetheless, there remains a discussion around the interpretation of the found coefficients of nonlinear model in a difference-in-differences framework. For this reason, and as I do not expect the probability of enrolment to be close to zero or one, the linear regression (7) is preferred.

(22)

Amhara is assigned the status of control state. As previously discussed, Amhara continued the use of Amharic as language of instruction after the reform, as it is the native language of the majority of its inhabitants. This leaves the state unaffected by the reform; children had access to MTE both before and after the reform. An exception to this is found in the zones Oromiya, Wag Hemra and Awi. The 2007 census shows that within these states, another ethnic group forms the dominant group in terms of population. To accommodate the dominant group, the language of instruction was changed to Oromiffa, Agew and Awgni, respectively. All observations of individuals residing in these zones are dropped from the sample.

For the treatment group, I limit the sample to individuals living in Tigray at the time of the census. There are several arguments on why to focus on Tigray as opposed to other states that have adapted their language policy. First, although all states excluding Amhara changed their language policy in 1994,Heugh et al.(2007) finds that not all states followed the national guidelines.11 A number of states still partially used Amharic as MOI after the reform, even if it was not the dominant language. If these states are included in the analysis, they would be wrongfully assigned as treatment states, while they actually did not undergo the full treatment. Tigray is one of the states whose practice is entirely consistent with the national policy. Another concern is that the education reform happened simultaneously with other policy changes that potentially influence the years of schooling. As discussed in SectionII, the reform was signed after the ethnically-based opposition party EPRDF came to power, which strongly campaigns for cultural assimilation. It is reasonable to assume that the education reform was only a small part of the new political agenda. If additional measures taken in the same period also impact the years of schooling, it is not possible to isolate the effect attributable to MTE through the difference-in-differences approach. However, Tigray offers a unique political setting which eliminates this problem. While the EPRDF only came to power in 1991, the regional party TPLF performed a coup against the suppressing regime in 1986; Tigray received partial autonomy and underwent substantial cultural and social changes (Ramachandran,2012). Nevertheless, the education policy remained unchanged until 1994, as it continued to fall under national competences.

(23)

This means that Tigray faced strong cultural assimilation before the education reform took place, while other states faced both processes simultaneously. Finally, the linguistic situation of Tigray is appealing. In contrast to some other states, Tigray has a population that is relatively homogeneous in their ethno-linguistic background; 95% of the population lists Tigrigna as their native language. The language of the state can thus be considered a relatively good proxy for the individual’s native language. This minimises the possibility of children not having access to MTE despite being assigned the treatment status.

When assigning the control and treatment status, I thus aim to separate individuals who went to primary school in Amhara from those who went to primary school in Tigray. A potential problem is that the data does not contain information on the state where the respondents’ lived at primary school age. Rather, it reports the current state of residence. Under strict assumptions, this can serve as a proxy for the state where primary school was attended; the necessary assumption is that the individual did not migrate over state borders between finishing primary school and being surveyed for the census. If this assumption is not fulfilled, people may live in a state different from where they attended primary school, which makes it challenging to observe whether or not they had access to MTE.

Due to Ethiopia’s political instability over the last decades, it is estimated that a large number of households have relocated over state borders, which poses a significant threat for the no-migration assumption. With regard to Amhara and Tigray, the states of interest in the present study, especially the Eritrean-Ethiopian war between 1998 and 2000 left its mark. As Tigray is located near the Eritrean border, the state was one of the most impacted by the conflict. The war led to large displacements of civilians who attempted to flee the war zones (Akresh et al., 2012). While this migration wave is a clear violation of the no-migration assumption, it does not necessarily pose a threat for the identification strategy. As Tigray was considered an area of conflict, it is unlikely that a large number of households migrated into Tigray. Rather, the migration movement was almost exclusively out of Tigray. This means that households who were residing in Tigray at the time of the census in 2007 are likely to have never migrated, or before the war broke out. On the other hand, the displacements could lead to significant attrition, as a large number of individuals who attended primary school in Tigray are no longer included in the sample. If

(24)

this decision to flee the war zone and migrate across state borders is systematically related to other individual characteristics, the internal validity of my analysis may be threatened. One way to address the no-migration assumption is to restrict the sample to individuals that have been residing in their current locality for over 10 years, which is observed in the census. While this restriction strengthens the no-migration assumption, it may also disregard individuals that are in fact of interest for the analysis. That is, the dropped individuals have not necessarily migrated to another state recently; it is possible that they merely relocated within state borders. While this specific group does not in fact violate the no-migration assumption, they are no longer taken into account. Two other problems arise from this. First, unnecessarily restricting the sample of interest will lower the number of observations, which in turn decreases the precision of the estimates. Second, if the group of individuals residing in their current locality for over 10 years is systematically different from the group that has recently relocated within state borders, the restriction again poses a threat for the internal validity of the present study. AppendixBpresents a balancing test for the two groups and illustrates clear differences between both groups. Accordingly, the restriction is not imposed for the sample of interest in the main specification. SectionVI

includes a robustness test to investigate whether the conclusions drawn from the analysis change significantly when excluding recently relocated households.

V.B. Exposed and unexposed birth cohort

The education reform of 1994 resulted in an almost immediate change of MOI in primary schools in the treated states (Heugh et al.,2007). This leads to a variation in exposure to mother-tongue instruction across birth cohorts, where an individual is either fully exposed, partially exposed or unexposed.

The Ethiopian schooling system includes 8 years of primary school, normally between 7 and 14 years of age. Individuals below school age in 1994, hence 7 years old or younger, entered primary school after the new policy was in place; therefore, they are assigned the status of fully exposed birth cohort. In contrast, children between the age of 8 and 14 years old in 1994 were already in primary school when the reform was introduced. As they were subject to Amharic instruction during their first schooling years and later switched to

(25)

mother-tongue instruction, this birth cohort is identified as the partially exposed cohort. Finally, children older than 14 years in 1994 already finished primary school before the reform took place and were thus not affected; they form the unexposed birth cohort.

Since the census took place in 2007, this implies that respondents of 20 years or younger form the fully exposed cohort. There is, however, a possibility that younger respondents have not finished their desired schooling level in 2007. Including students who plan on con-tinuing school beyond the moment of the data collection would result in an underestimation of the years of schooling for the younger cohort. Therefore, I only consider individuals over 18 years old. At this age, the formal schooling curriculum is finished and every person is potentially graduated from secondary school. The fully exposed cohort is thus defined as individuals from 18 to 20 years old in 2007. Similarly, respondents from 21 to 27 years old are assigned to the partially exposed cohort, while those from 28 to 30 years old form the unexposed birth cohort. As this method exploits the age of the respondents as an indicator for when they attended primary school, I must make the assumption that everyone is in the correct grade for his or her age. This is a strong assumption, especially for developing countries, where the repetition rate is rather high and late enrolment is common. How-ever, I do not observe when an individual started or graduated primary school, therefore I cannot not check the validity of this assumption.

The next step is to define the value of the binary variable Exposed for each birth co-hort. For the majority of respondents, this assignment is rather straightforward; Exposed takes the value of 1 for the fully exposed cohort, and zero for the unexposed cohort. For the partially exposed cohort, there are several options. First, it is possible to maximally exploit the variation in exposure by assigning non-binary values, depending on how many years an individual was exposed to mother-tongue instruction. The variable then expresses the percentage of primary school which was instructed in the native language. For example, a respondent born in 1985 turned 7 years old and started primary school in 1992, which means that he or she was instructed in Amharic during the first two years of primary school before switching to mother-tongue instruction. The value of Exposedj with j referring to

the year 1985, is therefore 0.75. A second option is to treat the partially exposed group in the same way as the unexposed group. While official policy documents reported that the

(26)

reform led to an abrupt change in MOI for all years of primary schooling,Ramachandran

(2012) argues that the actual implementation was likely to be more gradual; children who entered primary school in 1994 were directly instructed in their native language, while older students continued to be instructed in Amharic. As the first-grade students advanced to higher schooling levels, the provision of mother-tongue instruction gradually expanded. In that case, the policy was implemented only for children who entered primary schooling from 1994 and thereafter. The value of Exposed is then 1 for the fully exposed cohort, and zero for both the partially exposed and unexposed cohort. It is worth stressing that this approach is only valid if the reform was actually implemented gradually. Publications by bothHeugh et al.(2007) andAlemu and Tekleselassie(2006) cast doubt on this assumption and argue that the reform in fact did lead to an immediate change in MOI. A third option would be to disregard the partially exposed cohort altogether and to drop any observations of individuals in this cohort. This would minimise the risk of wrongfully assigning the partially exposed cohort either non-binary values or zero values. Since the initial impact of the reform is open for discussion, as illustrated above, this is an interesting feature. How-ever, dropping all individuals of the partially exposed cohort would significantly decrease the number of observations. In particular, there would be considerably less observations of individuals who were involved in primary school right after the reform happened. This poses a threat for the common trend assumption; the longer the time between the reform and the start of the individuals’ education, the more confounding factors possibly influence the trend in educational attainment.

Considering the disadvantages of eliminating the partially exposed cohort, and the evidence in favour of an immediate change in MOI, the first option is preferred. Hence, Exposed takes 1 for the fully exposed cohort, zero for the unexposed cohort and non-binary values for the partially exposed cohort. The results following this specification are discussed in SectionVI. Additional robustness checks, where other values are used for the partially exposed cohort, are also reported and discussed.

(27)

VI. Results

This section presents the estimates of the effect of MTE on several educational outcomes. First, the difference-in-differences estimates of the general effect are discussed. Second, the validity of the common trend assumption is investigated, by providing both graphical and formal evidence for the existence of a common pre-trend. Third, I examine heterogeneity of the treatment effect across gender and rurality. Finally, robustness checks regarding migration and the definition of the Exposure variable are reported.

VI.A. Difference-in-differences regressions

Table3 displays the the results from the difference-in-differences regressions as presented in the previous section. That is, Amhara (excluding zones Oromiya, Wag Hemra and Awi) serves as the control state, while Tigray is considered the treated state. Individuals aged 18 to 20 form the fully exposed cohort, those aged 21 to 27 the partially exposed cohort and those aged 28 to 30 the unexposed cohort.

The first two columns of Table3present the model with years of schooling as outcome variable. In the first column, the unadjusted estimates are reported; no additional control variables were included in the regression. The estimated coefficient corresponding to the T reated variable is positive and significant, indicating pre-existing differences in the years of schooling across states; before the 1994 reform, children living in Tigray on average attended school longer then children living in Amhara. Furthermore, also the variable Exposed is estimated to have a significantly positive effect. This demonstrates a general upward trend in the years of schooling, where schooling in both the control and treated state increases over time. This evolution is likely the result of the recent focus on universal primary schooling, accompanied by an expanded educational budget. In contrast, the primary variable of interest T reated×Exposed is insignificant. As this variable captures the causal relationship, the estimate suggests that mother-tongue instruction has no significant impact on years of schooling.

Performing a regression without additional control variables is a rather naive approach when estimating the causal relationship. One of the appealing features of the

(28)

difference-Table 3: difference-in-differences estimates of mother-tongue instruction on educational outcomes

(1) (2) (3) (4)

Years of schooling Years of schooling Enrolment Enrolment

Treated × Exposed 0.098 0.245∗∗∗ 0.017 0.024∗∗∗ (0.081) (0.068) (0.010) (0.009) Treated 1.162∗∗∗ 0.389∗∗∗ 0.137∗∗∗ 0.065∗∗∗ (0.055) (0.046) (0.007) (0.006) Exposed 1.629∗∗∗ 1.235∗∗∗ 0.231∗∗∗ 0.191∗∗∗ (0.033) (0.027) (0.004) (0.004) Constant 1.340∗∗∗ 3.252∗∗∗ 0.249∗∗∗ 0.426∗∗∗ (0.021) (0.070) (0.003) (0.008)

Controls No Yes No Yes

Observations 82632 82632 85000 85000

Adjusted R2 0.048 0.360 0.052 0.258

Notes: ∗p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01

Robust standard errors are reported in parentheses. The regression controls included are gender, the gender of households’ head, the household size, a rural dummy variable, access to electricity, access to piped water and the ownership of a radio. The individual’s age in years is not included as control variable due to multicollinearity with the variable Exposed. The parental years of schooling are not included as control variables due to a large number of missing observations.

in-differences method is that it explicitly removes any biases occurring from time constant confounding factors, as the dummy variable T reated controls for any pre-existing differ-ences between states. It is, however, possible that there are also time-varying variables that lead to state-specific trends. The trend in the outcome variable is then either caused by the treatment, the state-specific trend or a combination of both. To distangle the effect of the treatment, the variables that potentially lead to differential trends need to be included.12

12

Figure 1 and Figure 2 graphically illustrate and validate the common trend assumption for older, unaffected birth cohorts. Despite the historical trend being parallel, it is possible that abrupt changes in the states’ characteristics lead to deviation from the common trend for the treated birth cohort.

(29)

In addition, as in other econometric models, adding control variables reduces the residual variance, which in turn increases the precision of the estimates. In the second column of Table 3, control variables are included. The inclusion of additional variables does not affect the sign or significance of the estimated coefficient of T reated and Exposed; both variables have a positive, significant effect. On the other hand, the coefficient of the in-teraction term T reated × Exposed increases sharply in magnitude and becomes significant at the 1% significance level. The estimate implies that the introduction of MTE led to an increase in the average years of schooling by 0.245. This positive effect is consistent with the hypotheses of the present study.

The next two columns of Table 3investigate whether this positive effect occurs on the extensive margin, by evaluating the probability of enrolment as outcome variable. The estimated coefficient of T reated is again positive and significant. In addition to higher educational attainment, children in Tigray were also more likely to be enrolled in a formal schooling programme before the reform, relative to children in Amhara. Also the Exposed variable is estimated to have a positive effect, which indicates a general positive evolution in the enrolment rate across states. The variable of interest remains the interaction term T reated × Exposed, as it expresses the relationship between MTE and enrolment. The estimated is coefficient is positive, yet insignificant, when no additional control are included in the regression. When covariates are added, the coefficient increases and becomes signif-icant; MTE is estimated to raise the probability of enrolment by 2.4 percentage points. As the found effects of MTE on enrolment are similar in sign and significance to those on years of schooling, I can conclude that the positive effect mother-tongue instruction on years of schooling is at least partially caused by changes on the extensive margin.

The results are in line with former literature. In particular,Seid(2016) finds that the Ethiopian educational reform of 1994 led to an increase in the probability of enrolment of up to 7.4 percentage points, depending on the specification of the treatment region. The difference in magnitude of the effect - the estimates ofSeid(2016) are more than three times as large as the estimates of the present study - is driven by differences in the specification of the treatment region and the birth cohorts. First, while the present study focuses on Tigray as the treatment state, Seid (2016) performs a within-state analysis. The zones

(30)

Oromiya, Wag Hemra and Awi are than assigned the treatment status, while the other zones of Amhara serve as the control group. As discussed in Section V.A, the formerly suppressed zones might undergo other policy changes, simultaneous to the educational reform, as a result of the new political agenda. The increase in the enrolment rate than reflects the combined effect of mother-tongue instruction as well as other policy changes that facilitate primary school enrolment. Second, instead of exploiting the full variation in exposure to mother-tongue instruction by including the partially exposed cohort, Seid

(2016) does not take the partially exposed cohort into account. This leads to considerably less observations of individuals who were involved in primary school right after the reform happened. As argued in Section V.B, this in turn poses a threat for the common trend assumption, as it increases the probability of confounding factors influencing the trend in educational attainment. Considering the two caveats in the study of Seid (2016), I argue that the estimates of the present study more precisely capture the isolated impact of mother-tongue instruction on the probability of enrolment.

VI.B. Validity of the identifying assumptions

The main analysis, presented above, documents a positive correlation between MTE and both years of schooling as well as the probability of enrolment. When arguing that the correlation is in fact a causal relationship, it is implicitly assumed that the presented estimates pick up the isolated effect of the reform rather than the effect of other factors that also influence the educational attainment. By investigating the identifying assumptions of the difference-in-differences method, I assess the validity of this claim.

As previously discussed, the causal interpretation of the difference-in-differences method hinges on the assumption that in absence of the treatment, there would have been no sys-tematic differences in the educational attainment trend in both states. Since the counter-factual cannot be directly observed, it is unknown whether the common trend assumption is valid. However, it is possible to closely examine whether there is a common pre-trend. That is, I investigate whether the years of schooling and enrolment rate in both states evolved in a similar manner prior to the reform. The existence of a common pre-trend strongly substantiates the assumption of a common trend in absence of the treatment.

(31)

Figure 1: Common pre-trend in mean years of schooling .5 1 1.5 2 2.5 Mean y ears of sc ho oling 52-54 49-51 46-48 43-45 40-42 37-39 34-36 31-33 28-30 Age in 2007 Amhara Tigray

Figure 2: Common pre-trend in enrolment rate

.15 .2 .25 .3 .35 .4 Enrollmen t rate 52-54 49-51 46-48 43-45 40-42 37-39 34-36 31-33 28-30 Age in 2007 Amhara Tigray

(32)

Table 4: Falsification test: estimates of placebo reform in 1984

(1) (2)

Years of schooling Enrolment

Treated × Placebo exposed 0.113 0.025

(0.089) (0.017) Treated -0.172∗∗ -0.0289∗∗ (0.082) (0.014) Placebo exposed 0.105∗∗∗ -0.00546 (0.040) (0.007) Constant 2.731∗∗∗ 0.435∗∗∗ (0.102) (0.014)

Controls Yes Yes

Observations 39209 41964

Adjusted R2 0.348 0.214

Notes: ∗p < 0.10,∗∗ p < 0.05,∗∗∗p < 0.01

Robust standard errors are reported in parentheses. The regression controls included are gender, the gender of households’ head, the household size, a rural dummy variable, access to electricity, access to piped water and the ownership of a radio.

One approach is to examine the pre-trend in both states graphically. Figure 1 and Figure2plot the mean years of schooling and the enrolment rate for older cohorts, respec-tively, in both Tigray and Amhara (excluding zones Oromiya, Wag Hemra and Awi). As all the individuals in these cohorts were over primary school age in 1994, hence 28 years old or older in 2007, none of them were affected by the reform. The graph shows that, while the average schooling level and the enrolment rate in Tigray consistently exceeds those in Amhara, both states follow similar trajectories. However, it is unclear whether or not the graphs serve as proof of a common pre-trend; despite the simultaneous upward and downward movement, slight differences in magnitude are clearly visible. Therefore, I deem it necessary to formally investigate whether a common pre-trend exists.

(33)

In this regard, a falsification test is performed. This test uses the difference-in-differences framework applied in the main analysis to examine the effect of a placebo policy, where the treatment period is artificially placed ten years prior to the actual re-form.13 I thus investigate the impact of a hypothetical reform, which took place in 1984, on the years of schooling and the probability of enrolment. The formation of each cohort is in analogy with the method applied in the main analysis; children between 5 and 7 years old in 1984 (between 28 and 30 in 2007) form the exposed cohort, those between 8 and 14 years old in 1984 (between 31 and 37 in 2007) form the partially exposed cohort and those between 15 and 17 years old in 1984 (between 38 and 40 in 2007) form the unexposed cohort. Since none of the respondents are in fact affected by the reform, I do not expect to find a significant effect of the placebo reform if a common pre-trend exists. The results of the falsification test are presented in Table 4. The coefficients on the primary term of interest, T reated × P lacebo exposed, are uniformly insignificant; there is no statistically significant difference in the trend in years of schooling or probability to enrolment across states prior to the reform. The existence of a common pre-trend is a strong argument in favour of the validity of the common trend assumption. This, in turn, confirms that the positive correlation found in the main analysis can be interpreted as a causal relationship; mother-tongue instruction increases the years of schooling and the probability of enrolment.

VI.C. Heterogeneous treatment effects

In addition to exploring the general effect, it is interesting to further examine whether the treatment has a differential impact across subgroups. In order to allow for heterogeneity in the difference-in-differences model, the specifications for years of schooling and probability of enrolment are the following:

yijm= γ1+ γ2T reatedm+ γ3Exposedj+ γ4Groupi+ γ5(T reatedm× Exposedj)

+γ6(T reatedm× Groupi) + γ7(Exposedj× Groupi)

+γ8(T reatedm× Exposedj× Groupi) + γ Xi+ i

(8)

13

The choice of timing for the placebo reform, ten years prior to the actual reform, prevents that individuals who were in fact affected by the reform are included in the falsification test.

(34)

where yijmdenotes either the years of schooling or the probability of enrolment of individual

i born in year j, living in state m. Groupiis a binary indicator that divides the full sample

into two subgroups, amongst which possible heterogeneous effects are examined. The triple-interaction term T reatedm× Exposedj× Groupj is the primary variable of interest;

its corresponding coefficient γ8 expresses the differential impact of the treatment across

the defined groups. The heterogeneity is examined over rurality and gender.

Table5presents the heterogeneity analysis with respect to urban and rural households; the variable Groupitakes the value of 1 if the household of individual i is located in an urban

area, and zero otherwise. the years of schooling and the probability of enrolment serve as outcome variables in the first and second column, respectively. The results illustrate the large educational differences between urban and rural households. The initial years of schooling in untreated regions is 2.066 years higher among urban households, which indicates that urban households obtain close to five times more schooling relative to rural households. This is not surprising, considering that the urban households are also 22.3 percentage points more likely to ever attend school. While there has been a general positive trend in both years of schooling and enrolment, this upward evolution was significantly stronger in urban areas. This leads to a situation of divergence, where the educational gap between urban and rural households is increasing.

The coefficient corresponding to T reated×Exposed×U rban is statistically insignificant in the specification for the years of schooling, suggesting that there are no heterogeneous treatment effects by rurality; MTE impacts the educational attainment in both urban and rural regions in a similar way. Nevertheless, when examining the probability of enrolment as outcome variable, the coefficient on the interaction term T reated × Exposed × U rban is negative and significant. This indicates that, whereas the general effect of MTE on the enrolment rate is positive, the impact is significantly more limited in urban areas; the estimates even imply that mother-tongue instruction does not influence the enrolment rate

The effect of mother-tongue instruction on educational attainment in Ethiopia