The World Bank Economic Review

Zenebe Gebre,, Tihtina

doi: 10.1093/wber/lhz003pmid: N/A

Abstract This study employs a difference-in-differences estimation strategy to study the causal impacts of free primary education (FPE) on women’s schooling and fertility-related outcomes using the 1990s FPE reform of Malawi. The results show that full exposure to FPE at the mean value of the reform’s potential impact led to an 18.5 percentage points increase in the probability of ever enrolling in school and 1.2 years of additional schooling. Furthermore, the reduced form estimates suggest that full exposure to FPE increased age at first marriage and at first childbirth by almost 0.79 and 0.98 years respectively. The study also finds statistically significant reductions at all ages between age 14 and 24 in the total number of children to which women give birth. By age 24, for example, full exposure to FPE at the mean value of the reform’s potential impact led to 0.28 fewer children. In addition, there is evidence showing that exposure to FPE led to increased spacing between the first two births (about six additional months) if the first child is female. The study explores several channels through which FPE may affect fertility-related outcomes. Postponement of marriage and first childbirth, increased contraceptive use, and changes in mate characteristics all play a role. free primary education, fertility, Malawi schooling 1. Introduction In many Sub-Saharan African countries cost-sharing policies that left parents bearing some or all of the direct and indirect costs of their children’s primary education were prevalent.1 Under these policies, tuition and nontuition school fees often amounted to a significant share of households’ income and posed serious obstacles to school attendance (Buston 2002 and World Bank 2004). The impact fell on girls disproportionately, typically being the first to be pulled out of school or not being allowed to enroll in the first place (Unterhalter et al. 2015). In response, since the early 1990s many governments in Sub-Saharan Africa have introduced free primary education (FPE) policies, which abolished tuition and sometimes nontuition payments. Research shows that these policies are linked with increased educational attainment of girls and other marginalized groups (Bentaouet Kattan and Burnett 2004 and Bentaouet Kattan and Nicholas 2004). Given these linkages, in developing countries where fertility rates are often much higher than optimal, increased girls’ schooling due to policies such as FPE is considered to be a key mechanism to delaying fertility, reducing total number of children, and increasing the gap between births. This paper empirically examines the impact of FPE on female educational attainment and marriage and fertility-related outcomes using the FPE reform of Malawi. When it was first implemented, the reform eliminated school fees for all cohorts entering standard 1 (i.e., grade 1) in 1991 or later. In 1994, the fee elimination was further expanded to cover all students who were attending primary school at the time, including cohorts that started primary school prior to 1991. Combining cross-cohort variations in exposure to the FPE policy with cross-district variations in the potential impact of the reform on enrollment, the study employs a difference-in-differences estimation strategy to examine the effects of FPE. The difference-in-differences approach, which is estimated using a two-way fixed effects model, allows for consistent estimation of the causal effects of FPE separated from other secular trends and systematic regional differences in schooling and fertility. The estimates for the effects of FPE on schooling outcomes suggest that full exposure to the policy (i.e., 8 years of primary education under FPE) at the mean value of the potential impact of the reform (i.e., 0.51 in the data used in this study) led to an 18.5 percentage points increase in the probability of ever enrolling in school and to 1.2 years of additional schooling.2 There are a number of studies that use simple comparison between prereform and postreform periods to examine effects of FPE on schooling in the context of developing countries, but fail to establish causal links.3 The results in this paper improve upon the findings of these studies as the difference-in-differences approach captures differences across districts that are constant over time and differences over time that are constant across districts, allowing for consistent estimation of the causal impact of FPE on schooling. Studies by Osili and Long (2008), Oyelere (2010), and Lucas and Mbiti (2012a and 2012b) also employ a difference-in-differences strategy and find positive and statistically significant effects of FPE on schooling in Nigeria and Kenya.4 The results from this study add to this group of studies, by providing important additional evidence on the impact of the new wave of FPE reforms that have become increasingly common in Sub-Saharan African countries. The paper also examines the effects of FPE on women’s marriage and fertility-related outcomes showing that full exposure to FPE at the mean value of the potential impact of the reform increased age at first marriage and at first childbirth by about 0.79 and 0.98 years respectively. In addition, the study finds that FPE exposure had negative and statistically significant effects on the probability of having the first childbirth between the ages of 14 and 17. The biggest reductions were recorded at ages 16 and 17 where the probability of having the first child was reduced by 7.3 and 8.6 percentage points respectively. Another key result is the substantial negative effect of FPE on the total number of children at all ages between age 14 and 24. By age 24, for example, full exposure to FPE at the mean value of the potential impact of the reform led to 0.28 fewer children.5 Based on the estimates for the effects of FPE on schooling and on total fertility, a simple Wald calculation suggests that an additional year of schooling led to a reduction of 0.24 children by age 24. This result is in contrast to the findings in a quasi-experimental study by Breirova and Duflo (2004) in Indonesia. Using difference-in-differences approach, the authors find no conclusive evidence for the impact of increased average schooling within the household on total fertility.6 In contrast, Osili and Long (2008) examine the question in Nigeria and find that an additional year of schooling led to the birthing of 0.26 to 0.48 fewer children by age 25.7 While the findings from Osili and Long (2008) are statistically comparable to the results of this study, the authors do not explore the channels through which such effects happen. Looking at the different mechanisms through which FPE may impact fertility related outcomes, this study finds no conclusive evidence that increased labor market participation explains the reduction in fertility in the Malawi context. However, the study finds that postponement of marriage and first childbirth, increased contraceptive use, and changes in mate characteristics all play a role. In this regard, this paper relates more closely to Chicoine (2018) and Keats (2018). Chicoine (2018) in Ethiopia using Instrumental Variable (IV) method and Keats (2018) in Uganda using a regression discontinuity method report that an additional year of schooling led to 0.27 and 0.20 fewer children by age 24 and 25 respectively. Similar to the findings of this paper, Chicoine (2018) identifies postponement of marriage and first childbirth as an important mechanism for the reduction of fertility. In contrast to the results in this paper, he finds evidence that reduction in preferred number of children and increased participation in the formal labor market might help explain the results. He also fails to find support for increased contraceptive use or changes in mate characteristics as potential channels. Keats (2018), in addition to the channels identified by Chicoine (2018), finds that increased contraceptive use and changes in mate characteristics are important mechanisms. The variations in the findings of these studies show that more empirical work needs to be done to better understand how FPE and the resulting increase in educational attainment affect fertility. The in-depth and systematic analysis in this paper provides much needed additional evidence in this discussion. The study also makes a very important contribution by providing novel evidence on the impact of FPE on birth spacing and by teasing-out the differential effect of the policy depending on the gender of the first child. Short gaps between consecutive births are linked with negative health outcomes for children and mothers (WHO (2007) and USAID (2012)). Therefore, examining the implications of FPE on this key fertility-related outcome is instrumental in understanding the long-term benefits of such education policies. The results indicate that FPE exposure led to longer gaps between consecutive births, when the first child is female but not when the first child is male. Specifically, when the first child is female, the spacing between the first two births increased by about six months due to full exposure to FPE at the mean value of the potential impact of the reform. 2. Malawi’s Education System and the FPE Policy Primary education in Malawi takes 8 years and students who reach standard 8 (i.e., grade 8) sit for the Primary School Leaving Certificate Examination (PSLCE) in order to determine whether or not they can move to secondary school. The abolition of primary school fees in Malawi was phased in beginning with standard 1 during the 1991/1992 school year.8 Under this reform, any student who started school in 1991 or later will get eight years of FPE. Following the 1994 elections, the education policy of the country entered a new phase. The new Government of Malawi (GoM) continued to make primary education a top priority, and the FPE policy was expanded to all primary standards immediately. Hence, under the 1994 FPE policy, all primary students, including cohorts who enrolled in primary school before 1991 benefited from the policy. Over the same time period, the GoM increased both the recurrent budget of the education sector and the share of the recurrent budget allocated to primary schools.9 However, this increase in expenditure was not sufficient to have significant impact on per-student expenditure, which increased only from K39 to K41 during this time period.10 To meet the growing demand for teachers, the GoM recruited a large number of untrained teachers.11 This approach helped maintain the average students-to-teacher ratio relatively stable.12 However, the students-to-qualified teacher ratio rose significantly.13 Despite these challenges, following the FPE policy, enrollment increased drastically. The trends presented in fig. 1 show that, between 1990 and 1995, both primary school total enrollment and gross enrollment ratio (GER) increased drastically.14 Figure 1. Open in new tabDownload slide Primary School Total Enrollment and Gross Enrollment Ratio (GER) Source: Author’s analysis using enrollment data from the Ministry of Education, Science, and Technology annual reports and cohort population data from the 2008 census of Malawi. Note: Figure 1 plots primary school total enrollment and GER between 1975 and 2008. Primary GER is defined as the number of children enrolled in primary school regardless of age divided by the official primary school age population (6 to 13 year olds for Malawi). Figure 1. Open in new tabDownload slide Primary School Total Enrollment and Gross Enrollment Ratio (GER) Source: Author’s analysis using enrollment data from the Ministry of Education, Science, and Technology annual reports and cohort population data from the 2008 census of Malawi. Note: Figure 1 plots primary school total enrollment and GER between 1975 and 2008. Primary GER is defined as the number of children enrolled in primary school regardless of age divided by the official primary school age population (6 to 13 year olds for Malawi). The impact of the reform on enrollment, however was not uniform across districts in the three regions of Malawi. According to the 2008 census, Malawi is divided into 31 districts that are organized into 3 broad regions, the Northern, Central, and Southern regions.15 Compared to districts in the Central and Southern regions of Malawi, districts in the Northern region historically had higher enrollment and educational attainment levels. These regional variations can be traced back to the period of colonization. Christian missionaries first introduced formal education to the country at the end of the nineteenth century. During this period, they established most of their schools in the Northern region of the country (McCracken 1977; Divala 2010; World Bank 2004). Given these prereform historical patterns in schooling, the introduction of FPE had a much higher impact in the Southern and Central districts of the country (Al-Samarrai and Zaman 2002). A comparison of prereform and postreform enrollment rates in the three regions of Malawi clearly shows the differential impact of the reform. Regional trends of primary GER presented in fig. 2 show that the Northern region of Malawi had a much higher GER compared to the Central and Southern regions prereform. Furthermore, the gains in GER postreform were higher in the Central and Southern regions compared to the Northern region, showing the differential impact of the FPE policy.16 Figure 2. Open in new tabDownload slide Primary School Gross Enrollment Ratio (GER) by Region Source: Author’s analysis using enrollment data from the Ministry of Education, Science, and Technology annual reports and cohort population data from the 2008 Malawi census. Note: Figure 2 plots primary school GER separately for the three regions of Malawi. Primary GER is defined as the number of children enrolled in primary school regardless of age divided by the official primary school age population (6 to 13 year olds for Malawi). Figure 2. Open in new tabDownload slide Primary School Gross Enrollment Ratio (GER) by Region Source: Author’s analysis using enrollment data from the Ministry of Education, Science, and Technology annual reports and cohort population data from the 2008 Malawi census. Note: Figure 2 plots primary school GER separately for the three regions of Malawi. Primary GER is defined as the number of children enrolled in primary school regardless of age divided by the official primary school age population (6 to 13 year olds for Malawi). 3. Data The data used for the main analysis of this paper come from the 2000, 2005, and 2010 rounds of the Demographic and Health Surveys (DHS) of Malawi. The data sets contain detailed information on demographic characteristics of respondents, including their age, sex, marital status, and level of education as well as information on region of residence and urban/rural status. In addition, for women between the ages of 15 and 49, detailed data on their birth history, sexual activity history, and spousal characteristics is collected through the Individual Women Survey part of the DHS.17 This aspect of the DHS makes it superior to other sources, including the Malawi census data, to explore impacts of FPE on marriage and fertility-related outcomes.18 For the main analysis, one of the variations that is exploited to establish causal links comes from differences in the potential impact of the FPE reform across the districts of Malawi. However, the Malawi DHS do not include information on district of birth or residence. Instead, they contain information on the latitude and longitude of surveying clusters, which is used to identify the corresponding district.19 Using this approach, the district of residence for each respondent in the study sample is extrapolated. 4. Empirical Strategy In order to examine the causal impact of FPE on schooling and later-in-life outcomes, simply comparing younger cohorts that were exposed to FPE against older cohorts that were not exposed is not sufficient, as any impact will be confounded with other secular trends. To isolate the impact of the policy from such trends and find consistent estimates, this study uses a difference-in-differences strategy, which is estimated using a two-way fixed effects approach. This strategy combines cross-cohort variations in exposure to FPE with cross-district variations in the potential impact of the FPE reform on enrollment, effectively controlling for systematic differences across districts and cohorts.20 In this approach, a key identifying assumption is that, in the absence of the FPE reform, the average change in schooling and other outcome variables of interest would have been the same for both the treatment and control groups (i.e., districts with high potential impact and districts with low potential impact).21 Cross-Cohort Variation in Exposure to FPE Malawi’s official enrollment age for standard 1 is age 6. Individuals born in or after 1985 would turn 6 years of age in 1991 or later. Hence, these younger cohorts would have full exposure to FPE as a result of the introduction of the reform in 1991 for standard 1 entrants or younger cohorts. Furthermore, cohorts who were born between 1986 and 1989 would have partial exposure to FPE as a result of the expansion of the reform to all primary grades in 1994. This variation in exposure to FPE is presented in table 1, where row (1) presents year of birth, row (2) presents individuals’ age at the time of the 2010 DHS survey (i.e., the latest round of the DHS included in this study), and row (3) shows respondents’ age in 1991 when the reform was first implemented. Rows (4) and (5) present individuals’ standard in 1991 and 1994 respectively. These standard calculations assume on-time enrollment and hence are referred to as potential standards. Potential exposure to FPE, which is presented in row (7) of table 1, is then defined as the fraction of primary standards under FPE. Table 1. Exposure to Free Primary Education (FPE) by Eohort (1) Year of birth (j) >1985 1985 1984 1983 1982 1981 1980 1979 1978 1977 1976 (2) Age in 2010 <25 25 26 27 28 29 30 31 32 33 >34 (3) Age in 1991 <6 6 7 8 9 10 11 12 13 14 >14 (4) Standard in 1991 . 1 2 3 4 5 6 7 8 9 >9 (5) Standard in 1994 <4 4 5 6 7 8 9 10 11 12 >12 (6) Years of schooling under FPE 8 8 4 3 2 1 0 0 0 0 0 (7) Potential FPE exposure 1 1 |$\frac{1}{2}$| |$\frac{3}{8}$| |$\frac{1}{4}$| |$\frac{1}{8}$| 0 0 0 0 0 (8) Expected FPE exposure 1 1 0.93 0.84 0.68 0.55 0.42 0.34 0.24 0.15 <0.07 (1) Year of birth (j) >1985 1985 1984 1983 1982 1981 1980 1979 1978 1977 1976 (2) Age in 2010 <25 25 26 27 28 29 30 31 32 33 >34 (3) Age in 1991 <6 6 7 8 9 10 11 12 13 14 >14 (4) Standard in 1991 . 1 2 3 4 5 6 7 8 9 >9 (5) Standard in 1994 <4 4 5 6 7 8 9 10 11 12 >12 (6) Years of schooling under FPE 8 8 4 3 2 1 0 0 0 0 0 (7) Potential FPE exposure 1 1 |$\frac{1}{2}$| |$\frac{3}{8}$| |$\frac{1}{4}$| |$\frac{1}{8}$| 0 0 0 0 0 (8) Expected FPE exposure 1 1 0.93 0.84 0.68 0.55 0.42 0.34 0.24 0.15 <0.07 Source: Author’s analysis using prereform enrollment age data from the 2004/2005 IHS. Note: Table 1 presents cross-cohort variation in exposure to FPE. Row (1) presents year of birth, while rows (2) and (3) present individuals’ age in 2010 (i.e., the year of the latest DHS survey included in the study) and in 1991 respectively. Rows (4) and (5) present individuals’ standards in 1991 and 1994 respectively. Row (6) reports years of schooling under the FPE policy assuming on time enrollment, and row (7) presents potential exposure to FPE based on the same assumption. Row (8) presents expected exposure to FPE, which is estimated taking into account the probability of overage enrollment. Open in new tab Table 1. Exposure to Free Primary Education (FPE) by Eohort (1) Year of birth (j) >1985 1985 1984 1983 1982 1981 1980 1979 1978 1977 1976 (2) Age in 2010 <25 25 26 27 28 29 30 31 32 33 >34 (3) Age in 1991 <6 6 7 8 9 10 11 12 13 14 >14 (4) Standard in 1991 . 1 2 3 4 5 6 7 8 9 >9 (5) Standard in 1994 <4 4 5 6 7 8 9 10 11 12 >12 (6) Years of schooling under FPE 8 8 4 3 2 1 0 0 0 0 0 (7) Potential FPE exposure 1 1 |$\frac{1}{2}$| |$\frac{3}{8}$| |$\frac{1}{4}$| |$\frac{1}{8}$| 0 0 0 0 0 (8) Expected FPE exposure 1 1 0.93 0.84 0.68 0.55 0.42 0.34 0.24 0.15 <0.07 (1) Year of birth (j) >1985 1985 1984 1983 1982 1981 1980 1979 1978 1977 1976 (2) Age in 2010 <25 25 26 27 28 29 30 31 32 33 >34 (3) Age in 1991 <6 6 7 8 9 10 11 12 13 14 >14 (4) Standard in 1991 . 1 2 3 4 5 6 7 8 9 >9 (5) Standard in 1994 <4 4 5 6 7 8 9 10 11 12 >12 (6) Years of schooling under FPE 8 8 4 3 2 1 0 0 0 0 0 (7) Potential FPE exposure 1 1 |$\frac{1}{2}$| |$\frac{3}{8}$| |$\frac{1}{4}$| |$\frac{1}{8}$| 0 0 0 0 0 (8) Expected FPE exposure 1 1 0.93 0.84 0.68 0.55 0.42 0.34 0.24 0.15 <0.07 Source: Author’s analysis using prereform enrollment age data from the 2004/2005 IHS. Note: Table 1 presents cross-cohort variation in exposure to FPE. Row (1) presents year of birth, while rows (2) and (3) present individuals’ age in 2010 (i.e., the year of the latest DHS survey included in the study) and in 1991 respectively. Rows (4) and (5) present individuals’ standards in 1991 and 1994 respectively. Row (6) reports years of schooling under the FPE policy assuming on time enrollment, and row (7) presents potential exposure to FPE based on the same assumption. Row (8) presents expected exposure to FPE, which is estimated taking into account the probability of overage enrollment. Open in new tab However, overage enrollment is a very common phenomenon in Malawi.22 Therefore, calculation of exposure to FPE using the official enrollment age will fail to account for a number of fully and partially affected cohorts. In order to address this concern, each cohort’s expected exposure to FPE is calculated using prereform data on enrollment age from the 2004/2005 Integrated Household Survey (IHS) of Malawi. First, for each cohort j, the probability that they started standard 1 by year t (i.e Pjt) is calculated using data from the 2004/2005 IHS. Next, expected exposure to FPE for each cohort j (i.e., FPEj) is calculated using the following equation $$\begin{eqnarray} \textit{FPE}_{j} =\sum _{t}^{1990}max\left\lbrace P_{jt}\left(\frac{8+t-1994}{8}\right), 0\right\rbrace +P_{jt}(\ge 1991)\left(\frac{8}{8}\right) \end{eqnarray}$$ (1) The results from these estimations, which are referred to as expected exposure to FPE for cohort j, are reported in row (8) of table 1. Cross-District Variation in Intensity of Impact of FPE Using the cross-cohort variation presented in table 1 as the only source of variation in exposure to FPE, it is not possible to isolate the effect of the FPE reform from other secular trends. In order to overcome this problem, this study exploits historical differences in educational attainment across districts of Malawi. As discussed in detail in section 2, districts in the Central and Southern regions of Malawi historically had a much lower educational attainment compared to districts in the Northern region. Furthermore, districts that had low levels of schooling prereform experienced the largest increases in enrollment due to the reform. This variation is clearly illustrated in fig. 3, where prereform and postreform enrollment rates are compared across districts, using data from the 1987 and 2008 censuses of Malawi. Given these observations, a variable that measures prereform educational attainment in each district is formulated to serve as a measure for the potential impact of the FPE reform. This variable is constructed using the fraction of the adult population in each district that never enrolled in school prior to the introduction of the FPE reform in 1991, using data from the 1987 census of Malawi (i.e., the latest census prior to the 1991 FPE reform). Figure 3. Open in new tabDownload slide Correlation between Prereform Enrollment Rates and Change in Enrollment Rates Following the Free Primary Education Policy (FPE) Reform Source: Author’s analysis using data from the 1987 and 2008 Malawi censuses. Note: Figure 3 plots the fraction of the population ever enrolled in school by district in the 1987 census (i.e., x-axis) against change in enrollment rate by district between the 1987 and the 2008 censuses (i.e y-axis). Figure 3. Open in new tabDownload slide Correlation between Prereform Enrollment Rates and Change in Enrollment Rates Following the Free Primary Education Policy (FPE) Reform Source: Author’s analysis using data from the 1987 and 2008 Malawi censuses. Note: Figure 3 plots the fraction of the population ever enrolled in school by district in the 1987 census (i.e., x-axis) against change in enrollment rate by district between the 1987 and the 2008 censuses (i.e y-axis). Estimation The first question this study aims to answer is what effect FPE has had on women’s educational attainment in Malawi. For this analysis, various measures of educational attainment including the probability of ever enrolling in school, overall years of schooling, probability of completing primary school, and probability of being literate are considered. The estimations for the impacts of FPE on educational outcomes are conducted using the following two-way fixed effects equation: $$\begin{eqnarray} S_{ijd}=\gamma _{o}+ \eta _{j}+\nu _{d}+\gamma _{1}\left(FPE_{j}.I_{d}\right)+ \gamma _{x}X_{ijd}+\xi _{ijd}, \end{eqnarray}$$ (2) where for individual i in cohort j and district d, Sijd is a schooling-related outcome of interest. FPEj is expected exposure to FPE, which is computed as the fraction of primary school under the FPE policy for cohort j.23Id, which measures the potential impact of the FPE reform in district d, is computed as the fraction of the adult population that never enrolled in school in the district prior to the FPE reform.24 ηj and νd are cohort and district fixed effects, respectively, and Xijd is a vector of other covariates including controls for age, mother tongue language (used as proxy for ethnicity), religion, survey year, and district-specific trends.25 The coefficient of interest, γ1, quantifies the effect on schooling of eight years of exposure to FPE in a district with full potential impact of the FPE reform (i.e., district with zero enrollment rate prior to FPE). To look at the effects of full FPE exposure at the mean value of the potential impact of the reform, γ1 is multiplied by the mean value of Id, which is 0.51 in the data. After establishing effects on women’s educational attainment, the study looks at the impacts of FPE on fertility-related later-in-life-outcomes, including age at first intercourse, marriage, childbirth, and fertility outcomes by age for ages 14 to 25. These reduced-form effects are estimated using the following two-way fixed effects equation $$\begin{eqnarray} Y_{ijd}=\gamma _{o}+ \eta _{j}+\nu _{d}+\gamma _{1}\left(FPE_{j}.I_{d}\right)+ \gamma _{x}X_{ijd}+\xi _{ijd}, \end{eqnarray}$$ (3) where for individual i in cohort j and district d, Yijd is an outcome of interest and all other variables are defined in the same way as in equation (2). All of the estimations are weighted using sample weights provided in the DHS. Furthermore, the standard errors are clustered at the district level in all of the estimations. However, significance tests based on clustered standard errors may over-reject the null when the number of clusters is too few (Bertrand, Duflo, and Mullainathan 2004; Cameron, Gelbach, and Miller 2008). Considering that there are only 26 districts in the main analysis of this study, these concerns are prevalent in any statistical inference made with basic clustering.26 To address these concerns, p-values are obtained using the wild cluster bootstrapping procedure discussed in Cameron, Gelbach, and Miller (2008), which is shown to have good size properties even with very few number of clusters.27 In all result tables, standard errors and p-values corresponding to both robust clustering at district level and the wild bootstrap clustering procedures are presented. 5. Results This section presents the main findings of the paper on the effects of FPE on schooling and fertility outcomes as well as results regarding the channels through which FPE may affect fertility outcomes. Effects of FPE on Women’s Educational Attainment The results for the effect of FPE on the probability of ever enrolling in school, which is measured using a dummy variable that takes value one if the individual ever enrolled in school and zero otherwise, are presented in column (1) of table 2. The result shows that the probability of enrolling in school increased by 18.5 percentage points due to 8 years of exposure to FPF at the mean value of the potential impact of the reform.28 Looking at the effect of the policy on total years of schooling, the results show that 8 years of exposure to FPE led to 1.2 more years of schooling at the mean value of the potential impact of the reform. The study also finds a 5.6 percentage points increase in the probability of completing primary school and a 15.5 percentage points increase in the probability of being literate due to 8 years of FPE exposure. Table 2. Effects of Free Primary Education (FPE) on Women’s Educational Outcomes Dependent variable Ever enrolled in school Years of schooling Completed Primary Literate (1=Yes, 0=No) (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) FPE exposure×Program intensity 0.362*** 2.311*** 0.110* 0.304*** Standard error (robust cluster) (0.083) (0.685) (0.076) (0.101) P-value (robust cluster) (0.000) (0.002) (0.160) (0.006) P-value (wild cluster) (0.000) (0.000) (0.100) (0.004) Effect at mean potential impact 0.185 1.179 0.056 0.155 Mean 0.793 4.858 0.253 0.636 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,529 Dependent variable Ever enrolled in school Years of schooling Completed Primary Literate (1=Yes, 0=No) (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) FPE exposure×Program intensity 0.362*** 2.311*** 0.110* 0.304*** Standard error (robust cluster) (0.083) (0.685) (0.076) (0.101) P-value (robust cluster) (0.000) (0.002) (0.160) (0.006) P-value (wild cluster) (0.000) (0.000) (0.100) (0.004) Effect at mean potential impact 0.185 1.179 0.056 0.155 Mean 0.793 4.858 0.253 0.636 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,529 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 2 presents estimates for the effects of FPE on educational outcomes. Columns (1) and (2) report effects of FPE on the probability of ever enrolling in school and on total years of schooling respectively. Column (3) reports effect of FPE on the probability of completing primary school, while column (4) presents the effect on the probability of being literate. For all estimates, standard errors are clustered at district level. P-values based on both robust cluster and wild cluster bootstrap approaches are reported. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Table 2. Effects of Free Primary Education (FPE) on Women’s Educational Outcomes Dependent variable Ever enrolled in school Years of schooling Completed Primary Literate (1=Yes, 0=No) (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) FPE exposure×Program intensity 0.362*** 2.311*** 0.110* 0.304*** Standard error (robust cluster) (0.083) (0.685) (0.076) (0.101) P-value (robust cluster) (0.000) (0.002) (0.160) (0.006) P-value (wild cluster) (0.000) (0.000) (0.100) (0.004) Effect at mean potential impact 0.185 1.179 0.056 0.155 Mean 0.793 4.858 0.253 0.636 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,529 Dependent variable Ever enrolled in school Years of schooling Completed Primary Literate (1=Yes, 0=No) (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) FPE exposure×Program intensity 0.362*** 2.311*** 0.110* 0.304*** Standard error (robust cluster) (0.083) (0.685) (0.076) (0.101) P-value (robust cluster) (0.000) (0.002) (0.160) (0.006) P-value (wild cluster) (0.000) (0.000) (0.100) (0.004) Effect at mean potential impact 0.185 1.179 0.056 0.155 Mean 0.793 4.858 0.253 0.636 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,529 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 2 presents estimates for the effects of FPE on educational outcomes. Columns (1) and (2) report effects of FPE on the probability of ever enrolling in school and on total years of schooling respectively. Column (3) reports effect of FPE on the probability of completing primary school, while column (4) presents the effect on the probability of being literate. For all estimates, standard errors are clustered at district level. P-values based on both robust cluster and wild cluster bootstrap approaches are reported. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Effect of FPE on Marriage and Fertility-Related Outcomes After establishing the effects of FPE on women’s educational outcomes, the study examines if FPE has had any impact on age at first intercourse, marriage, and childbirth. The findings presented in table 3 show that full exposure to FPE increased age at first marriage and at first childbirth by almost 0.79 and 0.98 years respectively at the mean value of the potential impact of the reform. The estimated impact on age at first intercourse, though positive, is not statistically significant. Table 3. Effects of Free Primary Education (FPE) on Age at First Intercourse, Marriage and Childbirth Dependent variables (In years) Age at first intercourse Age at first marriage Age at first birth (1) (2) (3) FPE exposure×Program intensity 0.410 1.540* 1.930*** Standard error (robust cluster) (0.675) (0.664) (0.415) P-value (robust cluster) (0.550) (0.029) (0.000) P-value (wild cluster) (0.533) (0.096) (0.002) Effect at mean potential impact 0.209 0.785 0.984 Mean 16.5 17.6 18.4 Cohort fixed effects Yes Yes Yes District fixed effects Yes Yes Yes District trends Yes Yes Yes N 41,872 38,931 37,442 Dependent variables (In years) Age at first intercourse Age at first marriage Age at first birth (1) (2) (3) FPE exposure×Program intensity 0.410 1.540* 1.930*** Standard error (robust cluster) (0.675) (0.664) (0.415) P-value (robust cluster) (0.550) (0.029) (0.000) P-value (wild cluster) (0.533) (0.096) (0.002) Effect at mean potential impact 0.209 0.785 0.984 Mean 16.5 17.6 18.4 Cohort fixed effects Yes Yes Yes District fixed effects Yes Yes Yes District trends Yes Yes Yes N 41,872 38,931 37,442 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 3 presents results for the effects of FPE on ages at first intercourse (column (1)), first marriage (column (2)), and first childbirth (column (3)). For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Table 3. Effects of Free Primary Education (FPE) on Age at First Intercourse, Marriage and Childbirth Dependent variables (In years) Age at first intercourse Age at first marriage Age at first birth (1) (2) (3) FPE exposure×Program intensity 0.410 1.540* 1.930*** Standard error (robust cluster) (0.675) (0.664) (0.415) P-value (robust cluster) (0.550) (0.029) (0.000) P-value (wild cluster) (0.533) (0.096) (0.002) Effect at mean potential impact 0.209 0.785 0.984 Mean 16.5 17.6 18.4 Cohort fixed effects Yes Yes Yes District fixed effects Yes Yes Yes District trends Yes Yes Yes N 41,872 38,931 37,442 Dependent variables (In years) Age at first intercourse Age at first marriage Age at first birth (1) (2) (3) FPE exposure×Program intensity 0.410 1.540* 1.930*** Standard error (robust cluster) (0.675) (0.664) (0.415) P-value (robust cluster) (0.550) (0.029) (0.000) P-value (wild cluster) (0.533) (0.096) (0.002) Effect at mean potential impact 0.209 0.785 0.984 Mean 16.5 17.6 18.4 Cohort fixed effects Yes Yes Yes District fixed effects Yes Yes Yes District trends Yes Yes Yes N 41,872 38,931 37,442 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 3 presents results for the effects of FPE on ages at first intercourse (column (1)), first marriage (column (2)), and first childbirth (column (3)). For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab The study also estimates the impact of FPE on the probability of first childbirth by age for ages 14 to 25. The results are presented in fig. 4, where each point estimate comes from a separate reduced form regression given by equation (3). In the figure, the 95 percent confidence interval is constructed using wild cluster bootstrapping approach, and the stars indicate the level of statistical significance based on wild cluster bootstrap pvalues.29 The results in fig. 4 suggest that FPE had negative and statistically significant effects on the probability of first childbirth between the ages of 14 and 17. The highest impacts were recorded at ages 16 and 17, where the probability of having the first child was reduced by 7.3 and 8.6 percentage points respectively due to full exposure to FPE at the mean value of the potential impact of the reform. The effect of FPE on the probability of first childbirth becomes less negative and statistically insignificant past age 17. This suggests that the FPE policy may have led to reduced probability of teenage motherhood. Figure 4. Open in new tabDownload slide Reduced Form Estimates for the Effects of Free Primary Education (FPE) on the Probability of First Child Birth by Age Source: Author’s analysis using data from the 2000, 2005, and 2010 DHS. Note: Figure 4 presents estimates for the effects of FPE on the probabilities of first childbirth by age. Each point estimate in the figure comes from a separate reduced form regression, and the 95 percent confidence interval is constructed using a wild cluster bootstrapping approach. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Figure 4. Open in new tabDownload slide Reduced Form Estimates for the Effects of Free Primary Education (FPE) on the Probability of First Child Birth by Age Source: Author’s analysis using data from the 2000, 2005, and 2010 DHS. Note: Figure 4 presents estimates for the effects of FPE on the probabilities of first childbirth by age. Each point estimate in the figure comes from a separate reduced form regression, and the 95 percent confidence interval is constructed using a wild cluster bootstrapping approach. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. In addition, the study finds negative and statistically significant effects of FPE on total fertility at all ages between ages 14 and 24. The results in fig. 5 show that the effects of FPE on total fertility becomes more negative at every age. For example, by age 24, 8 years of exposure to FPE led to 0.28 fewer children at the mean value of the potential impact of the reform. To compare the results in this study with other findings in the literature, the Wald estimate for the causal impact of one year of additional schooling on total fertility is calculated. This is done by taking the ratio between the two-way fixed effects estimates for the impact of FPE on total fertility and on years of schooling. The Wald estimate suggests that an additional year of schooling led to the birthing of 0.24 fewer children by age 24, a result that is comparable to other estimates in the literature.30 Figure 5. Open in new tabDownload slide Reduced Form Estimates for the Effects of Free Primary Education (FPE) on the Number of Children by Age Source: Author’s analysis using data from the 2000, 2005, and 2010 DHS. Note: Figure 5 presents estimates for the effect of FPE on the total number of children by age. Each point estimate in the figure comes from a separate reduced form regression, and the 95 percent confidence interval is constructed using a wild cluster bootstrapping approach. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Figure 5. Open in new tabDownload slide Reduced Form Estimates for the Effects of Free Primary Education (FPE) on the Number of Children by Age Source: Author’s analysis using data from the 2000, 2005, and 2010 DHS. Note: Figure 5 presents estimates for the effect of FPE on the total number of children by age. Each point estimate in the figure comes from a separate reduced form regression, and the 95 percent confidence interval is constructed using a wild cluster bootstrapping approach. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Effect of FPE on Birth Spacing Another key outcome variable to consider is the gap between consecutive births. There is extensive literature showing a strong link between increased birth interval and improvements in infants and children health outcomes (Rutstein et al. 2005). In addition, based on a systematic review of studies on birth timing, policy recommendation by WHO (2007) and USAID (2012) suggest that women should wait at least 24 months between consecutive births. These recommendations are expected to reduce infant and child mortality as well as improve maternal heath. In light of these important benefits of increased birth spacing, the study examines the effects of FPE on the length of birth intervals between the first and second births, the second and third births, and the third and fourth births, conditional on having at least two, three, and four children, respectively. The results, presented in table 4, show that full exposure to FPE increased the birth interval between the first two births by almost 3.4 months at the mean value of the potential impact of the reform. However, this effect masks the differential impact of FPE on birth spacing depending on the gender of the first child. When the first child is male, the study finds no conclusive evidence for the impact of FPE on birth spacing. When the first child is female, on the other hand, the study finds six months increase in birth spacing due to full exposure to FPE. Effects of FPE on the spacing between the second and third births and the third and fourth births are positive but statistically not significant. Table 4. Effects of Free Primary Education (FPE) on Birth Spacing Birth spacing Dependent variables (In months) 1st and 2nd 1st and 2nd, 1st birth male 1st and 2nd, 1st birth female 2nd and 3rd 3rd and 4th (1) (2) (3) (4) (5) FPE exposure×Program intensity 6.658** 1.646 12.000*** 5.621 6.591* Standard error (robust cluster) (2.219) (3.757) (3.332) (4.558) (3.420) P-value (robust cluster) (0.007) (0.665) (0.001) (0.229) (0.065) P-value (wild cluster) (0.012) (0.655) (0.000) (0.322) (0.074) Effect at mean potential impact 3.395 0.839 6.120 2.870 3.361 Mean 34.1 34.2 34.0 34.8 35.6 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 30,332 15,269 15,063 24,171 18,272 Birth spacing Dependent variables (In months) 1st and 2nd 1st and 2nd, 1st birth male 1st and 2nd, 1st birth female 2nd and 3rd 3rd and 4th (1) (2) (3) (4) (5) FPE exposure×Program intensity 6.658** 1.646 12.000*** 5.621 6.591* Standard error (robust cluster) (2.219) (3.757) (3.332) (4.558) (3.420) P-value (robust cluster) (0.007) (0.665) (0.001) (0.229) (0.065) P-value (wild cluster) (0.012) (0.655) (0.000) (0.322) (0.074) Effect at mean potential impact 3.395 0.839 6.120 2.870 3.361 Mean 34.1 34.2 34.0 34.8 35.6 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 30,332 15,269 15,063 24,171 18,272 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 4 presents results for the effects of FPE on birth spacing between consecutive births. Column (1) presents effect on birth spacing between the first two births, while columns (2) and (3) present effects on birth spacing between the first two births conditional on having a male and a female first child respectively. Columns (4) and (5) report effects of FPE on the spacing between the second and third births and the third and fourth births respectively. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Table 4. Effects of Free Primary Education (FPE) on Birth Spacing Birth spacing Dependent variables (In months) 1st and 2nd 1st and 2nd, 1st birth male 1st and 2nd, 1st birth female 2nd and 3rd 3rd and 4th (1) (2) (3) (4) (5) FPE exposure×Program intensity 6.658** 1.646 12.000*** 5.621 6.591* Standard error (robust cluster) (2.219) (3.757) (3.332) (4.558) (3.420) P-value (robust cluster) (0.007) (0.665) (0.001) (0.229) (0.065) P-value (wild cluster) (0.012) (0.655) (0.000) (0.322) (0.074) Effect at mean potential impact 3.395 0.839 6.120 2.870 3.361 Mean 34.1 34.2 34.0 34.8 35.6 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 30,332 15,269 15,063 24,171 18,272 Birth spacing Dependent variables (In months) 1st and 2nd 1st and 2nd, 1st birth male 1st and 2nd, 1st birth female 2nd and 3rd 3rd and 4th (1) (2) (3) (4) (5) FPE exposure×Program intensity 6.658** 1.646 12.000*** 5.621 6.591* Standard error (robust cluster) (2.219) (3.757) (3.332) (4.558) (3.420) P-value (robust cluster) (0.007) (0.665) (0.001) (0.229) (0.065) P-value (wild cluster) (0.012) (0.655) (0.000) (0.322) (0.074) Effect at mean potential impact 3.395 0.839 6.120 2.870 3.361 Mean 34.1 34.2 34.0 34.8 35.6 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 30,332 15,269 15,063 24,171 18,272 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 4 presents results for the effects of FPE on birth spacing between consecutive births. Column (1) presents effect on birth spacing between the first two births, while columns (2) and (3) present effects on birth spacing between the first two births conditional on having a male and a female first child respectively. Columns (4) and (5) report effects of FPE on the spacing between the second and third births and the third and fourth births respectively. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab In all the estimations in this paper, the main identifying assumption is that in the absence of the FPE reform, the average change in schooling and other outcome variables would have been the same for both the treatment and control groups. To provide some supportive evidence for this assumption, robustness checks are conducted by estimating effects of FPE on schooling and fertility outcomes of older unaffected cohorts. The results show that FPE had no impact on these older cohorts, suggesting that secular trends are not driving the findings.31 Additional robustness checks are conducted using alternative specifications by (1) including controls for wealth index, quarter of birth dummies, and the quadratic and cubic of age, (2) using alternative measures for FPE exposure, and (3) excluding partially exposed cohorts and internal migrants from the study sample. The findings from these robustness checks, which are discussed in detail in the supplementary online appendix, show that the main findings of the study continue to hold under these alternative specifications. Estimating the effects of FPE on cross-region and cross-district migration, the study finds no conclusive evidence that FPE had an impact on internal migration or that internal migration might be driving any of the reduced-form estimates for the effects of FPE on fertility-related outcomes. Channel Through Which Schooling Affects Fertility Outcomes From a theoretical perspective, several channels linking increased schooling (e.g., due to an FPE policy) and fertility-related outcomes have been proposed in the literature. First, education may shift a woman’s optimal fertility choices towards fewer children of higher quality by increasing her earnings and the opportunity cost of child rearing (Becker 1960; Mincer 1963; Becker and Lewis 1973). Second, positive assortative matching may lead to more educated women marrying more educated men, which will further increase the household’s permanent income (Behrman and Rosenzweig 2004), augmenting the effect of education on fertility through a multiplier effect. Third, education may improve individuals’ and households’ views and knowledge of contraceptive methods and their ability to effectively utilize them (Grossman 1972). This paper empirically examines the importance of some of these channels by looking at the effects of FPE on women’s labor market participation, use of contraceptives, and changes in the characteristics of their mates.32 The study starts by examining if FPE exposure led to higher rates of employment for affected women and if this can explain the decline in fertility. To this end, effects of FPE on several labor market outcomes including the probability of being currently employed, the probability of having worked in the last 12 months, and the probability of currently working in nonagriculture sectors are considered. The results presented in table 5 provide no conclusive evidence that FPE exposure led to changes in women’s employment status. This result is different from findings in Chicoine (2018) and Keats (2018). Table 5. Effects of Free Primary Education (FPE) on Women’s Employment Dependent variables (1=Yes, 0=No) Current employment Employment in the Last 12 months Self employment Non-agricultural employment (1) (2) (3) (4) FPE exposure×Program intensity 0.091 −0.015 0.030 0.063 Standard error (robust cluster) (0.083) (0.083) (0.072) (0.056) P-value (robust cluster) (0.284) (0.856) (0.681) (0.275) P-value (wild cluster) (0.206) (0.873) (0.677) (0.358) Effect at mean potential impact 0.046 −0.008 0.015 0.061 Mean 0.566 0.668 0.462 0.205 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 Dependent variables (1=Yes, 0=No) Current employment Employment in the Last 12 months Self employment Non-agricultural employment (1) (2) (3) (4) FPE exposure×Program intensity 0.091 −0.015 0.030 0.063 Standard error (robust cluster) (0.083) (0.083) (0.072) (0.056) P-value (robust cluster) (0.284) (0.856) (0.681) (0.275) P-value (wild cluster) (0.206) (0.873) (0.677) (0.358) Effect at mean potential impact 0.046 −0.008 0.015 0.061 Mean 0.566 0.668 0.462 0.205 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 5 presents results for the effects of FPE on the probabilities of being currently employed (column (1)), being employed in the last 12 months (column (2)), being self-employed (column (3)), and being employed in nonagricultural sectors (column (4)). For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Table 5. Effects of Free Primary Education (FPE) on Women’s Employment Dependent variables (1=Yes, 0=No) Current employment Employment in the Last 12 months Self employment Non-agricultural employment (1) (2) (3) (4) FPE exposure×Program intensity 0.091 −0.015 0.030 0.063 Standard error (robust cluster) (0.083) (0.083) (0.072) (0.056) P-value (robust cluster) (0.284) (0.856) (0.681) (0.275) P-value (wild cluster) (0.206) (0.873) (0.677) (0.358) Effect at mean potential impact 0.046 −0.008 0.015 0.061 Mean 0.566 0.668 0.462 0.205 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 Dependent variables (1=Yes, 0=No) Current employment Employment in the Last 12 months Self employment Non-agricultural employment (1) (2) (3) (4) FPE exposure×Program intensity 0.091 −0.015 0.030 0.063 Standard error (robust cluster) (0.083) (0.083) (0.072) (0.056) P-value (robust cluster) (0.284) (0.856) (0.681) (0.275) P-value (wild cluster) (0.206) (0.873) (0.677) (0.358) Effect at mean potential impact 0.046 −0.008 0.015 0.061 Mean 0.566 0.668 0.462 0.205 Cohort fixed effects Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes District trends Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 5 presents results for the effects of FPE on the probabilities of being currently employed (column (1)), being employed in the last 12 months (column (2)), being self-employed (column (3)), and being employed in nonagricultural sectors (column (4)). For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Next, the study looks at impacts of FPE on the use of contraceptives, measured through the probabilities of reporting use of or intention to use contraceptives, ever having used contraceptives, current use of contraceptives, and current use of modern contraceptives. The results, which are presented in table 6, show that full exposure to FPE at the mean value of the potential impact of the reform led to a 5.2 percentage points increase in the probability of reporting use of or intention to use contraceptives and an 11.5 percentage points increase in the probability of ever having used contraceptives. The study also finds a 9.7 percentage points and 8.3 percentage points increase in the probabilities of current use of any contraceptives and current use of modern contraceptives, respectively. These results suggest that increased practice of contraceptive methods is an important channel through which the FPE policy affects fertility. The study also finds positive effects of FPE on the probabilities of women discussing family planning (FP) options with their partners and women reporting participation in the decisions to use FP options. However, the results are not statistically significant.33 Table 6. Effects of Free Primary Education (FPE) on Contraceptives and Family Planning (FP) Options Dependent variables (1=Yes, 0=No) Contraceptives use/intention to use Ever used contraceptives Current use of contraceptives Current use of modern contraceptives Discussed FP options with partner Participated in decision to use FP options (1) (2) (3) (4) (5) (6) FPE exposure×Program intensity 0.104* 0.225** 0.190* 0.163* 0.142 0.224 Standard error (robust cluster) (0.061) (0.089) (0.118) (0.089) (0.094) (0.136) P-value (robust cluster) 0.099 (0.019) (0.118) (0.079) (0.145) (0.100) P-value (wild cluster) 0.101 (0.028) (0.100) (0.078) (0.162) (0.158) Effect at mean potential impact 0.053 0.115 0.097 0.083 0.072 0.114 Mean 0.767 0.561 0.300 0.267 0.093 0.785 Cohort fixed effects Yes Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 24,616 14,289 Dependent variables (1=Yes, 0=No) Contraceptives use/intention to use Ever used contraceptives Current use of contraceptives Current use of modern contraceptives Discussed FP options with partner Participated in decision to use FP options (1) (2) (3) (4) (5) (6) FPE exposure×Program intensity 0.104* 0.225** 0.190* 0.163* 0.142 0.224 Standard error (robust cluster) (0.061) (0.089) (0.118) (0.089) (0.094) (0.136) P-value (robust cluster) 0.099 (0.019) (0.118) (0.079) (0.145) (0.100) P-value (wild cluster) 0.101 (0.028) (0.100) (0.078) (0.162) (0.158) Effect at mean potential impact 0.053 0.115 0.097 0.083 0.072 0.114 Mean 0.767 0.561 0.300 0.267 0.093 0.785 Cohort fixed effects Yes Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 24,616 14,289 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 6 presents results for the effects of FPE on the use of contraceptives. Column (1) reports effects on the probability of reporting intention to use contraceptives, and column (2) reports effect on the probability of ever having used contraceptives. Columns (3) and (4) report effects on the probabilities of current use of any contraceptives and current use of modern contraceptives, respectively. Columns (5) and (6) preset effects on the probabilities of women discussing family planning (FP) options with partner and women participating in decisions to use contraceptives. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab Table 6. Effects of Free Primary Education (FPE) on Contraceptives and Family Planning (FP) Options Dependent variables (1=Yes, 0=No) Contraceptives use/intention to use Ever used contraceptives Current use of contraceptives Current use of modern contraceptives Discussed FP options with partner Participated in decision to use FP options (1) (2) (3) (4) (5) (6) FPE exposure×Program intensity 0.104* 0.225** 0.190* 0.163* 0.142 0.224 Standard error (robust cluster) (0.061) (0.089) (0.118) (0.089) (0.094) (0.136) P-value (robust cluster) 0.099 (0.019) (0.118) (0.079) (0.145) (0.100) P-value (wild cluster) 0.101 (0.028) (0.100) (0.078) (0.162) (0.158) Effect at mean potential impact 0.053 0.115 0.097 0.083 0.072 0.114 Mean 0.767 0.561 0.300 0.267 0.093 0.785 Cohort fixed effects Yes Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 24,616 14,289 Dependent variables (1=Yes, 0=No) Contraceptives use/intention to use Ever used contraceptives Current use of contraceptives Current use of modern contraceptives Discussed FP options with partner Participated in decision to use FP options (1) (2) (3) (4) (5) (6) FPE exposure×Program intensity 0.104* 0.225** 0.190* 0.163* 0.142 0.224 Standard error (robust cluster) (0.061) (0.089) (0.118) (0.089) (0.094) (0.136) P-value (robust cluster) 0.099 (0.019) (0.118) (0.079) (0.145) (0.100) P-value (wild cluster) 0.101 (0.028) (0.100) (0.078) (0.162) (0.158) Effect at mean potential impact 0.053 0.115 0.097 0.083 0.072 0.114 Mean 0.767 0.561 0.300 0.267 0.093 0.785 Cohort fixed effects Yes Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes Yes N 47,642 47,642 47,642 47,642 24,616 14,289 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 6 presents results for the effects of FPE on the use of contraceptives. Column (1) reports effects on the probability of reporting intention to use contraceptives, and column (2) reports effect on the probability of ever having used contraceptives. Columns (3) and (4) report effects on the probabilities of current use of any contraceptives and current use of modern contraceptives, respectively. Columns (5) and (6) preset effects on the probabilities of women discussing family planning (FP) options with partner and women participating in decisions to use contraceptives. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap p-values. Open in new tab When examining the link between women’s FPE exposure and the characteristics of their life partners, the study finds that FPE exposure increased women’s probability of marrying younger men. The results in table 7 show that full exposure to FPE at the mean level of the potential impact of the reform led to a 1.2 years reduction in the age of a woman’s life partner. The results for the effect of FPE on the partner’s educational attainment and occupation are inconclusive. Table 7. Effects of Free Primary Education (FPE) on Life Partner’s Characteristics Dependent variables Partner’s age Partner’s years of schooling (In years) Education difference (In years) Nonagricultural employment (1=Yes, 0=No) Professional, technical, & managerial employment (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) (5) FPE exposure×Program intensity −2.413* 1.265 −0.645 0.017 0.010 Standard error (robust cluster) (1.026) (0.824) (0.452) (0.107) (0.031) P-value (robust cluster) (0.027) (0.138) (0.167) (0.876) (0.747) P-value (wild cluster) (0.074) (0.182) (0.178) (0.877) (0.745) Effect at mean potential impact −1.231 0.645 −0.329 0.009 0.005 Mean 36.162 6.342 1.947 0.476 0.053 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 32,950 38,362 38,362 38,949 38,949 Dependent variables Partner’s age Partner’s years of schooling (In years) Education difference (In years) Nonagricultural employment (1=Yes, 0=No) Professional, technical, & managerial employment (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) (5) FPE exposure×Program intensity −2.413* 1.265 −0.645 0.017 0.010 Standard error (robust cluster) (1.026) (0.824) (0.452) (0.107) (0.031) P-value (robust cluster) (0.027) (0.138) (0.167) (0.876) (0.747) P-value (wild cluster) (0.074) (0.182) (0.178) (0.877) (0.745) Effect at mean potential impact −1.231 0.645 −0.329 0.009 0.005 Mean 36.162 6.342 1.947 0.476 0.053 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 32,950 38,362 38,362 38,949 38,949 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 7 presents results for the effects of FPE on partner’s characteristics. Columns (1) and (2) report effects on partner’s age and education, respectively, while column (3) presents effect on the educational difference between the couple. Column (4) reports effect of FPE on the probability of the partner’s employment in nonagricultural sectors, while column (5) reports his probability of holding professional, technical, or managerial jobs. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap P-values. Open in new tab Table 7. Effects of Free Primary Education (FPE) on Life Partner’s Characteristics Dependent variables Partner’s age Partner’s years of schooling (In years) Education difference (In years) Nonagricultural employment (1=Yes, 0=No) Professional, technical, & managerial employment (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) (5) FPE exposure×Program intensity −2.413* 1.265 −0.645 0.017 0.010 Standard error (robust cluster) (1.026) (0.824) (0.452) (0.107) (0.031) P-value (robust cluster) (0.027) (0.138) (0.167) (0.876) (0.747) P-value (wild cluster) (0.074) (0.182) (0.178) (0.877) (0.745) Effect at mean potential impact −1.231 0.645 −0.329 0.009 0.005 Mean 36.162 6.342 1.947 0.476 0.053 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 32,950 38,362 38,362 38,949 38,949 Dependent variables Partner’s age Partner’s years of schooling (In years) Education difference (In years) Nonagricultural employment (1=Yes, 0=No) Professional, technical, & managerial employment (1=Yes, 0=No) (1=Yes, 0=No) (1) (2) (3) (4) (5) FPE exposure×Program intensity −2.413* 1.265 −0.645 0.017 0.010 Standard error (robust cluster) (1.026) (0.824) (0.452) (0.107) (0.031) P-value (robust cluster) (0.027) (0.138) (0.167) (0.876) (0.747) P-value (wild cluster) (0.074) (0.182) (0.178) (0.877) (0.745) Effect at mean potential impact −1.231 0.645 −0.329 0.009 0.005 Mean 36.162 6.342 1.947 0.476 0.053 Cohort fixed effects Yes Yes Yes Yes Yes District fixed effects Yes Yes Yes Yes Yes District trends Yes Yes Yes Yes Yes N 32,950 38,362 38,362 38,949 38,949 Source: Author’s analysis using data from the 2000, 2005, and 2010 Demographic and Health Surveys of Malawi. The study sample includes women born between 1950 and 1995. Note: Table 7 presents results for the effects of FPE on partner’s characteristics. Columns (1) and (2) report effects on partner’s age and education, respectively, while column (3) presents effect on the educational difference between the couple. Column (4) reports effect of FPE on the probability of the partner’s employment in nonagricultural sectors, while column (5) reports his probability of holding professional, technical, or managerial jobs. For all estimates, standard errors are clustered at district level. P-values are calculated using both robust cluster and wild cluster bootstrap approaches. *** p < 0.01, ** p < 0.05, and * p < 0.1 based on wild cluster bootstrap P-values. Open in new tab 6. Conclusion The findings in this paper contribute to the literature, first, by providing consistent estimates for the causal effect of FPE on female schooling. Specifically, the results show that the abolition of primary school fees in Malawi had significant positive effects on women’s probability of enrollment, years of schooling, and probability of being literate. These results are important, as they provide a strong argument for making schooling more affordable in order to increase the participation of historically marginalized groups, including women. Second, this paper examines the causal link between FPE and fertility-related outcomes and the channels through which such impacts might occur. The results show that schooling has negative and statistically significant effects on age at marriage and first childbirth and on total fertility at all ages for 14-to-24-year-olds. The study also finds evidence that exposure to FPE led to increased spacing between the first two births when the first child is female. These results show that FPE policies that extend educational opportunities to girls also reap vast rewards by lowering fertility rates and by increasing birth spacing in developing countries, where population growth tends to be the fastest. Although the impact of FPE on schooling and fertility-related outcomes has been explored in other developing countries, the findings of this paper in the context of Malawi are important for a number of reasons. The World Bank’s Country Status Report for Malawi, which was published in 2004, reported that the rapid increase in primary school enrollment due to the FPE policy put the education sector under tremendous stress. Most particularly, the enrollment increase led to an acute shortage of trained teachers and teaching-learning materials. In addition to the decline in learning conditions, over the past two decades, Malawi has been one of the worst performers in Sub-Saharan Africa in terms of student outcomes. A 2010 World Bank report, for example, shows that primary completion rate for Malawi was 35 percent in 2007, while the average for Sub-Saharan African countries was 61 percent. Regarding learning outcomes, the results from Southern Africa Consortium for Monitoring Educational Quality (SACMEQ) show that, between 1998 and 2004, the fraction of Malawian students who achieved the minimum level of mastery in reading was reduced by half to just about 9 percent, the lowest percentage among SACMEQ countries. Furthermore, only about 2 percent of students possessed skills beyond basic numeracy in mathematics (World Bank 2010). Given the country’s dismal performance in key educational indicators, it is not obvious whether the causal relationships between FPE, educational attainment, and fertility-related outcomes will hold in the same way as in other developing countries. The results from this paper provide strong empirical evidence that increased schooling of women, even under poor learning conditions, leads to reduced total fertility and increased control over the timing of fertility. The results also provide some evidence as to which channels are important in explaining these causal relationships in this context. Author Biographical Tihtina Zenebe Gebre (corresponding author) is an economist in the Education Global Practice of the World Bank Group. Her email address is [email protected]. The author thanks Kasey Buckles, Joseph Kaboski, William Evans for insightful comments. A supplementary online appendix for this article can be found at The World Bank Economic Review website. Footnotes 1 In a 2004 World Bank survey, for example, 11 out of the 27 countries surveyed in the region charged tuition fees and 22 collected at least one type of nontuition payment in a form of textbook rents, exam registration fees, or parent-teacher associations dues (World Bank 2004). 2 The potential impact of the FPE reform in a given district, which is a proxy for the intensity of the impact of the reform, is measured using prereform educational attainment in each district. Districts with low educational attainment prereform experienced the highest gains in enrollment and attainment following the FPE reform. Given this inverse relationship, the fraction of each district’s population that never enrolled in school pre-FPE is used as a measure for the potential impact of the reform. The mean value of this measure in the study sample data is 0.51. Detailed discussion on this potential impact measure is provided in section 4. 3 For example, see Kadzamira and Rose (2003), Deininger (2003), Nishimura, Yamano, and Sasaoka (2008), and Brossard and Foko (2007). 4 The first two papers focus on universal education policies that abolished school fees in some Nigerian states between 1955 and 1981. In this context, Osili and Long (2008) find that full exposure to FPE led to 1.54 years of additional schooling while Oyelere (2010) finds that for every year of exposure to free education, educational attainment increased by 0.15 year, implying that eight years of exposure to FPE would lead to 1.2 years of additional schooling. Lucas and Mbiti (2012a) look at a more recent FPE reform that took place in Kenya in 2003 and find that the reform increased the number of students who completed primary school. Looking at the same policy reform in Kenya, Lucas and Mbiti (2012b) show that boys responded to the FPE reform in greater numbers than girls, widening the primary completion gap. 5 By age 25, the study finds a reduction of 0.32 children due to full exposure to FPE at the mean value of the potential impact of the reform. However, this result is not statistically significant. 6 Early work in developing countries, including Ainsworth, Beegle, and Nyamete (1996) and Schultz (1997), find negative associations between schooling and fertility, but fail to address possible endogeneity concerns in order to establish causal relationships. 7 The range of values are reported in the paper based on different estimation specifications. 8 This change in policy was in part due to the serious shift in the priorities of donors and developing countries, including Malawi, in the 1990s towards making primary education accessible and affordable to all. 9 For example, the recurrent budget for the education sector increased from 11 percent of the national budget in 1990/1991 to 24 percent in 1997. Out of the education sector’s recurrent budget, the share of primary school increased from 45 to 65 percent. 10 The kwacha, denoted by “K”, is the currency of Malawi. In 1994, the exchange rate of the kwacha against the U.S. dollar was K17 (Simwaka 2007). 11 For example, more than 18,000 secondary school leavers were recruited and assigned to schools in 1993/1994 (World Bank 2004). 12 In 1993/1994, the average students-to-teacher ratio was 62, and in 1998/99 it was 60, while the standard set by the Government of Malawi was 60. 13 The students-to-qualified teacher ratio rose from 88 in 1992/1993 to 119 in 1997/1998 (Kadzamira and Rose 2003). 14 UNICEF defines the primary gross enrollment ratio (GER) as the number of children enrolled in a primary school, regardless of age, divided by the population of the age group that officially corresponds to primary school. Due to late enrollment and grade repetition GER measures can be more than 100 percent. In fig. 1, GER for the years between 1975 and 2008 is plotted using yearly enrollment data from MOEST reports and cohort population data from the 2008 census of Malawi. 15 The Demographic and Health Surveys of Malawi, which are the main sources of data used in this paper, use an older classification of districts, which divides the country into 26 districts organized into 3 regions. For the main analysis, the DHS classification is used. 16 In the supplementary online appendix, available with this article at The World Bank Economic Review website, fig. S.1 presents trends in total primary enrollment by region. The figure shows that the change in total enrollment postreform was much higher in the Central and Southern regions of the country compared to the Northern region. 17 Summary statistics are presented in the supplementary online appendix. 18 Although the DHS is used for the main analysis, the 1987 and 2008 censuses, as well as annual reports from the MOEST, have been analyzed to more effectively study the impact of FPE and undertake robustness checks. 19 Clusters are geographic areas that are defined only for the purpose of the DHS survey. QGIS, a geographic information systems analysis program, is used to match the latitude and longitude of clusters to their corresponding districts. 20 Similar approaches have been applied by Bleakley (2010) and Lucas and Mbiti (2012a and 2012b). Bleakley (2010) utilizes pre-eradication levels of malaria to identify local variation in the impact of eradication programs while Lucas and Mbiti (2012a and 2012b) more directly applied this concept to examine effects of the 2003 FPE policy in Kenya. 21 Some evidence is provided in support of this identification assumption through a robustness check, which is presented in the supplementary online appendix. In the robustness check, all treated younger cohorts are dropped out of the sample and some of the remaining older cohorts are treated with a placebo FPE policy reform. The estimates for the impact of the placebo FPE reform on these older cohorts show no effects on schooling or fertility-related outcomes. This suggests that prereform trends are not driving the results. 22 Ravishankar et al. (2016) state that overage enrollment is common in Malawi partly due to the remoteness of schools in rural areas of the country and parents’ reservations about allowing very young children to walk long distances to school. 23 The range of values for this variable are presented in row (8) of table 1. 24 This calculation is done using enrollment data from the 1987 Malawi census. 25 Since multiple rounds of the DHS are used in the study, controlling for age is important. In the supplementary online appendix, results with additional controls for quadratic and cubic of age (i.e., age2, age3) are presented. 26 The number of districts used in the analysis is based on the district classification used in 2000 DHS. The number of districts in Malawi has increased to 31 more recently. However, in this study, the older definition of districts is used to allow the use of both older and more recent data. 27 The wild bootstrap procedure creates pseudo-samples using Rademacher weights (i.e +1 with probability 0.5 and –1 with probability 0.5) to obtain a new sampling of residuals from a restricted regression that imposes the null hypothesis. In each of the 999 replications, the Wald test statistic is calculated for the statistical significance of the coefficient of interest. The Wald test statistics are sorted and the location of the original Wald test statistic in the sorted distribution of bootstrapped Wald test statistics provides the bootstrapped p-value. The bootstrapped confidence interval is given by |$[\hat{\gamma _{1}}-se(\hat{\gamma _{1}})c_{1-\frac{\alpha }{2}}, \hat{\gamma _{1}}+se(\hat{\gamma _{1}})c_{\frac{\alpha }{2}}]$|⁠, where |$c_{1-\frac{\alpha }{2}}$| and |$c_{\frac{\alpha }{2}}$| denote the |$1-\frac{\alpha }{2}$| and |$\frac{\alpha }{2}$| quantiles on the sorted list. Since these quantiles are obtained from a sorted list that is not symmetric, the confidence interval is in general asymmetric. In all estimations, the Stata command ’boottest’, which is developed by Roodman (2015), is used to obtain the wild bootstrap p-values and confidence intervals.The study also uses its own Stata code to corroborate the boottest estimates and finds very similar values. However, the boottest command by Roodman (2015) is faster, hence the study uses it for most of its estimations. 28 The probability of enrolling in school increased by 36.2 percentage points in a district with full potential impact of the reform (i.e., district with zero enrollment pre-FPE). The mean value of the potential impact of the reform in the data is 0.51, which implies that 8 years of FPE exposure led to 18.5 percentage points increase in the probability of enrolling in school at this mean value. 29 Note that wild cluster bootstrap confidence intervals need not be symmetric. See MacKinnon (2015) for discussion. 30 See summary of findings in the literature in section 1. 31 The results, which are presented in the supplementary online appendix, show no statistically significant impact of the FPE reform on these older unaffected cohorts. 32 Additional results on effects of FPE on fertility preferences and women’s empowerment in the household are presented in the supplementary online appendix. 33 In the supplementary online appendix, the impact of FPE on women’s knowledge about contraceptives is examined. The DHS data do not provide any information specifically quantifying the level of understanding about contraceptives and ability to effectively use them. However, whether or not FPE exposure increased the likelihood of hearing about contraceptives from the radio, television, newspaper, home visits by health workers, or visits to health facilities is examined. The findings provide no conclusive evidence that these are important channels through which women learn about contraceptives. References Ainsworth M. , K. Beegle , A. Nyamete . 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 . Google Scholar Crossref Search ADS WorldCat Al-Samarrai S. , H. Zaman . 2002 . “ The Changing Distribution of Public Education Expenditure in Malawi .” Institute of Development Studies, University of Sussex, Falmer, UK, and the World Bank, Washington, DC . COPAC Becker G. S. 1960 . “An Economic Analysis of Fertility.” NBER Chapters, In Demographic and Economic Change in Developed Countries, 209 – 40 , National Bureau of Economic Research, Inc , Cambridge, MA . COPAC Becker G. S. , H. G. Lewis . 1973 . “ On the Interaction between the Quantity and Quality of Children .” Journal of Political Economy 81 ( 2, Part 2 ): S279 – S288 . Google Scholar Crossref Search ADS WorldCat Behrman J.R. , M.R. Rosenzweig . 2004 . “ Parental Allocations to Children: New Evidence on Bequest Differences Among Siblings .” Review of Economics and Statistics 86 ( 2 ): 637 – 40 . Google Scholar Crossref Search ADS WorldCat Bentaouet Kattan R. , N. Burnett . 2004 . “ School Fees: A Roadblock to Education For All ,” World Bank Report 30193, World Bank , Washington, DC . COPAC Bentaouet Kattan R. , B. Nicholas . 2004 . “ User Fees in Primary Education .” Education for all working papers, World Bank , Washington, DC . COPAC Bertrand M. , E. Duflo , S. Mullainathan . 2004 . “ How Much Should We Trust Differences-in-Differences Estimates? ” Quarterly Journal of Economics 119 ( 1 ): 249 – 75 . Google Scholar Crossref Search ADS WorldCat Bleakley H. 2010 . “ Malaria Eradication in the Americas: A Retrospective Analysis of Childhood Exposure .” American Economic Journal: Applied Economics 2 ( 2 ): 1 – 45 . Google Scholar Crossref Search ADS WorldCat Breierova L. , E. Duflo . 2004 . “ The Impact of Education on Fertility and Child Mortality: Do Fathers Really Matter Less Than Mothers ?” NBER Working Paper No. W10513 , National Bureau of Economic Research , Cambridge, MA . COPAC Brossard M. , B. Foko . 2007 . “ Coûts et Financement de l’Enseignement Supérieur en Afrique Francophone .” World Bank Seminar , Cotonou, Benin , July 2–4 , 2007 . WorldCat Buston O. 2002 . “ Every Child in School: A Challenge to Finance and Development Ministers .” Oxfam Briefing Paper 20 , Oxfam International . Washington, DC . COPAC Cameron A. C. , J. B. Gelbach , D. L. Miller . 2008 . “ Bootstrap-Based Improvements for Inference with Clustered Errors .” Review of Economics and Statistics 90 ( 3 ): 414 – 27 . Google Scholar Crossref Search ADS WorldCat Chicoine L. 2018 . “ Free Primary Education, Fertility, and Women’s Access to the Labor Market: Evidence from Ethiopia .” IZA Discussion Paper No. 10387 , Institute of Labour Economics , Bonn, Germany . COPAC Deininger K. 2003 . “ Does Cost of Schooling Affect Enrollment by the Poor? Universal Primary Education in Uganda .” Economics of Education Review 22 ( 3 ): 291 – 305 . Google Scholar Crossref Search ADS WorldCat Divala J. J. K. 2010 . “ Rethinking Higher Education Access in Malawi: Dilemmas in Achieving a Just System .” Paper presented at the Philosophy of Education Society of Great Britain Annual Conference , New College, Oxford , March 27, 2010 . COPAC Grossman M. 1972 . “ On the Concept of Health Capital and the Demand for Health .” Journal of Political Economy 80 ( 2 ): 223 – 55 . Google Scholar Crossref Search ADS WorldCat Kadzamira E. , P. Rose . 2003 . “ Can Free Primary Education Meet the Needs of the Poor? Evidence from Malawi .” International Journal of Educational Development 23 ( 5 ): 501 – 16 . Google Scholar Crossref Search ADS WorldCat Keats A. 2018 . “ Women’s Schooling, Fertility, and Child Health Outcomes: Evidence from Uganda’s Free Primary Education Program .” Journal of Development Economics 135 : 142 – 59 . Google Scholar Crossref Search ADS WorldCat Lucas A. M. , I. Mbiti . 2012a . “ Access, Sorting, and Achievement: The Short-Run Effects of Free Primary Education in Kenya .” American Economic Journal: Applied Economics 4 ( 4 ): 226 – 53 . Google Scholar Crossref Search ADS WorldCat Lucas A. M. , I. Mbiti . 2012b . “ Does Free Primary Education Narrow Gender Differences in Schooling Outcomes? Evidence From Kenya .” Journal of African Economies 21 ( 5 ): 691 – 22 . Google Scholar Crossref Search ADS WorldCat MacKinnon J. 2015 . “ Wild Cluster Bootstrap Confidence Intervals .” L’Actualité économique 91 ( 1–2 ): 11 – 33 . WorldCat McCracken J. 1977 . Politics and Christianity in Malawi, 1875 to 1940: The Impact of the Livingstonia Mission in the Northern Province . Cambridge : Cambridge University Press . Google Preview WorldCat COPAC Mincer J. 1963 . “ Market Prices, Opportunity Costs, and Income Effects .” In Measurement in Economics , edited by Christ C. , 67 – 82 . Stanford : Stanford University Press . Google Preview WorldCat COPAC Technology (MOEST) , Ministry of Education Science 1986-008 . Education Statistics . Lilongwe, Malawi : MOEST . Google Preview WorldCat COPAC National Statistical Office–NSO . 2005 . “ Malawi Second Integrated Household Survey (IHS-II) 2004–2005 ” [Datasets]. Lilongwe , Malawi . COPAC National Statistical Office–NSO/Malawi and ICF. 2005 . “ Malawi Demographic and Health Survey 2004–2005 ” [Datasets]. Zomba, Malawi : NSO and ICF . COPAC National Statistical Office–NSO/Malawi and ICF 2000 . “ Malawi Demographic and Health Survey 2000 ” [Datasets]. Zomba, Malawi : NSO and ICF . COPAC National Statistical Office–NSO/Malawi and ICF 2010 . “ Malawi Demographic and Health Survey 2010 ” [Datasets]. Zomba, Malawi : NSO and ICF . COPAC National Statistical Office, Minnesota Population Center . 1987 . “ Population and Housing Census of Malawi 1987 ” [Datasets]. Integrated Public Use Microdata Series, International . Minneapolis, MN : IPUMS . https://doi.org/10.18128/D020.V7.1. COPAC National Statistical Office, Minnesota Population Center . 2008 . “ Population and Housing Census of Malawi 2008 ” [Datasets]. Integrated Public Use Microdata Series, International . Minneapolis, MN : IPUMS . https://doi.org/10.18128/D020.V7.1. COPAC Nishimura M. , T. Yamano , Y. Sasaoka . 2008 . “ Impacts of the Universal Primary Education Policy on Educational Attainment and Private Costs in Rural Uganda .” International Journal of Educational Development 28 ( 2 ): 161 – 75 . Google Scholar Crossref Search ADS WorldCat Osili U. O. , B. T. Long . 2008 . “ Does Female Schooling Reduce Fertility? Evidence from Nigeria .” Journal of Development Economics 87 ( 1 ): 57 – 75 . Google Scholar Crossref Search ADS WorldCat Oyelere R. U. 2010 . “ Africa’s Education Enigma? The Nigerian story .” Journal of Development Economics 91 ( 1 ): 128 – 39 . Google Scholar Crossref Search ADS WorldCat Ravishankar V. , S. E. T. El-Kogali , D. Sankar , N. Tanaka , N. Rakoto-Tiana . 2016 . Primary Education in Malawi: Expenditures, Service Delivery, and Outcomes. Washington, DC : World Bank . COPAC Roodman D. 2015 . “ BOOTTEST: Stata Module to Provide Fast Execution of the Wild Bootstrap with Null Imposed .” Statistical Software Components S458121, Department of Economics , Boston College , Boston, MA . COPAC Rutstein S. O. , K. Johnson , A. Conde-Agudelo , A. Rosas-Bermudez . 2005 . “ Effect of Birth Spacing on Infant and Child Mortality: A Systematic Review and Meta-analysis .” In World Health Organization Consultation and Scientific Review of Birth Spacing . WHO , Geneva , June 13–15, 2005 . Google Preview WorldCat COPAC Schultz T. P. 1997 . “ Demand for Children in Low Income Countries .” In Handbook of Population and Family Economics , vol. 1 . part A, edited by Rosenzweig M. R. , Stark O. , 349 – 430 . Amsterdam : Elsevier . Google Preview WorldCat COPAC Simwaka K. 2007 . “ Modeling and Forecasting the Malawi Kwacha–US Dollar Nominal Exchange Rate .” Research Gate . https://www.researchgate.net/publication/24113463_Modeling_and_Forecasting_the_Malawi_Kwacha-US_Dollar_Nominal_Exchange_Rate. WorldCat United States Aid for International Development (USAID) . 2012 . “ Healthy Timing and Spacing of Pregnancies: A Family Planning Investment Strategy for Accelerating the Pace of Improvements in Child Survival .” Prepared by U.S. Agency for International Development (USAID), Bureau for Global Health, Office of Population and Reproductive Health, Washington, DC 20523-3600. May 2012 . COPAC Unterhalter E. S. , K. McCracken , S. Marquez , A. Chelstowska . 2015 . “ Empowering Women and Girls Through Education .” Brussels : Directorate General For Internal Policies Policy Department C: Citizen’s Right and Constitutional Affairs . COPAC Willis R. J. 1973 . A New Approach to the Economic Theory of Fertility Behavior . Journal of Political Economy 81 ( 2, Part 2 ): S14 – S64 . Google Scholar Crossref Search ADS WorldCat World Bank . 2004 . “ Cost, Financing, and School Effectiveness of Education in Malawi .” Africa Region Human Development Working Paper Series , World Bank , Washington, DC . COPAC World Bank . 2010 . The Education System in Malawi: Country Status Report . Washington , DC: World Bank . Google Preview WorldCat COPAC World Health Organization (WHO) . 2007 . Report of a WHO Technical Consultation on Birth Spacing : Geneva , Switzerland , June 13–15, 2005 . COPAC © The Author(s) 2019. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)

Alzúa, Maria, Laura;Batbekh,, Soyolmaa;Batchuluun,, Altantsetseg;Dalkhjav,, Bayarmaa;Galdo,, José

doi: 10.1093/wber/lhaa013pmid: N/A

Abstract The effectiveness of a demand-driven vocational-training program for disadvantaged youth in Ulaanbaatar, the capital city of Mongolia is assessed through a randomized controlled trial. Mongolia, a transitional country whose economic structure shifted from a Communist, centrally planned economy to a free-market economy over a relatively short period, offers a new setting in which to test the effectiveness of market-based active-labor-market policies. Results show short-term positive impacts on self-employment and skills match, while positive but uncertain effects emerge for employment and earnings. Substantial heterogeneity emerges as relatively older, richer, and better-educated individuals drive these positive effects. A second intervention, in which participants were randomly assigned to receive newsletters with information on market returns to vocational training, shows statistically meaningful effects on the length of exposure to the program (i.e., number of training days attended). These positive impacts, however, do not lead to higher employment or greater earnings. vocational-training programs, labor market, randomized controlled trial, employment, earnings, job quality 1. Introduction Youth unemployment is a ubiquitous problem in most developed and developing countries. Over seventy-three million youth aged 15–24 are unemployed worldwide, and around 20 percent of the world's youth are neither employed nor enrolled in an education or training program (International Labour Organization 2017). For disadvantaged youth, who disproportionately lack skills and access to decent jobs that could lift them out of poverty, labor-market conditions are even more drastic (International Labour Organization 2015). In Mongolia, the share of young individuals who were neither employed nor enrolled in an education or training program topped 25 percent (Shatz et al. 2015). This study assesses the effectiveness of a vocational-training program on labor-market outcomes of disadvantaged youth in Ulaanbaatar, the capital city of Mongolia, following a standard randomized controlled trial design (RCT). Vocational-training programs are the most widely used active-labor-market policy (ALMP) intended to mitigate unemployment among poor youth worldwide, but no evidence exists regarding their effectiveness in central Asia. The most recent meta-analyses (e.g., McKenzie 2017; Card, Kluve, and Weber 2018; Kluve et al. 2019) have revealed this important gap in the literature. Indeed, this large body of evidence shows modest (e.g., McKenzie 2017) to positive (e.g., Kluve et al. 2019) effects of vocational-training programs on employment and earnings of youth in developing settings.1 A solid pattern that has emerged from this literature is the substantial heterogeneity of estimated treatment effects as a result of variations in country settings, program design, and profile of targeted groups. Mongolia offers a new setting for an assessment of ALMPs because it is a “transitional country” – that is, one whose labor market rapidly transformed from a centrally planned to a free-market economy. In fact, Mongolia's economy shifted from a centralized, Communist system, in which youth unemployment was officially reported as being very low or nonexistent, to a free-market economy with a high share of youth not in employment, education, or training (NEET). Mongolia's is a traditional herding and agricultural economy that transitioned into a resource-based economy in a short period of time. Two demographic shifts have added pressure to the labor markets over the past 20 years. First, the share of working-age youth is relatively high because Mongolia has a population with an average age of 26. Second, sustained internal migration from rural areas to the city of Ulaanbaatar has doubled the population of the capital city over that period. As a result, the rate of youth labor-force participation in Ulaanbaatar is remarkably low at around 40 percent. Thus, assessing whether vocational-skills training would foster employment and wage growth for young, vulnerable, and unemployed individuals in such an environment is relevant to policy. The Mongolian Vocational Skills-Training Program (MVTP) follows a market-based approach in program design that is intended to match the skills that training institutions offer with those that are in demand in the market. This is sometimes known as a “demand-driven” ALMP. Such programs have become widely used in several Latin American countries, and assessment of their effectiveness shows positive effects on employment and earnings, particularly for disadvantaged women (Betcherman et al. 2007; Attanasio, Kugler, and Meghir 2011). “Demand-driven” training programs are designed to combine in-class training in selected vocational skills – where proven demand exists for those skills – with on-the-job training in the form of internships (see, e.g., Ibarrarán and Rosas 2009; Galdo and Chong 2012). In contrast to RCT training interventions implemented over the last decade in developing settings, a distinctive difference of the MVTP initiative is its short duration and low cost. The per-trainee cost ranges from USD 90 to USD 140 for trainings that vary from 20 to 45 days. Because lack of knowledge about market returns to training in a setting in transition from Communism to a market-based economy could lead participants to make uninformed choices, this study implements a complementary RCT design within a group of MVTP participants in which personalized weekly newsletters with information about the market returns to vocational training in Mongolia are distributed. The aim of the intervention was to shape choices made by MVTP participants regarding dropout or completion of vocational training. As shown Dammert and Galdo (2013) and Choe, Flores-Lagunes, and Lee (2015), lack of compliance with the treatment design is an important source of variation in the magnitude of treatment effects for training programs. Such a finding suggests the possible benefits of increased training exposure for potential dropouts. While providing information about market returns has been shown to be effective in education interventions in other developing settings (e.g., Jensen 2010; Dinkelman and Martínez 2013), to the best of the authors’ knowledge the approach has not been tested as a policy design in the context of an ALMP in a developing country. The main results from this study point out positive and meaningful effects on skills matching and self-employment. The estimated impacts for employment and earnings, in contrast, are more ambiguous. The intent-to-treat estimated employment impacts are not statistically significant and range from −0.02 to 0.054, while the absolute change in monthly earnings ranges from USD 17 to USD 32, with statistical significance varying across model specifications and time span. In line with what has been reported in the literature, these mean effects hide important heterogeneities across the sample of interest, to the extent that positive and statistically significant effects on employment, earnings, and skill match are observed for relatively older, richer, and better-educated young people. Finally, provision of information on market returns to training positively affects the time of exposure to training (i.e., number of program days attended), but does not lead to higher employment or greater earnings. This article adds to the present understanding of youth (un)employment, particularly in transitional countries. Its results are consistent with a broader body of micro-level evidence that suggests modest labor-market gains of vocational-training programs for youth. The study contributes to this evidence by assessing a low-cost, short-length training initiative, by providing some of the first evidence from central Asia, and by focusing attention on information constraints in returns to vocational training upon completion of training and employment. Indeed, these findings complement research regarding how providing information affects returns to formal education and school completion (e.g., Jensen 2010). Although there are limits to generalizability and long-term effects, these results provide insight into market-based approaches and information-based programs for ALMP in transitional settings. This paper is organized as follows. Section 2 briefly describes the Mongolian labor market; section 3 provides details about the training design; section 4 develops the evaluation experimental design; section 5 discusses the data and baseline covariate distribution for treatment groups; section 6 presents the results; and section 7 provides the concluding remarks. 2. Institutional Setting Mongolia is a small, transitional country with a population of slightly above 3 million living in a landlocked area located between Russia and China. Ethnic Mongols account for about 95 percent of the population. Mongolia is one of the least densely populated countries in the world, with almost half its population living in the city of Ulaanbaatar, the country's capital. Following the dismantling of the USSR in the early 1990s, Soviet assistance, which had accounted for almost one-third of Mongolia's GDP, disappeared almost overnight. Mongolia's economy changed dramatically from a centrally planned system to a market-based one. Mongolia, however, is resource-rich, with large deposits of copper, gold, coal, and uranium. Its mining sector accounts for almost one-fifth of GDP and 40 percent of exports, although it accounts for only 4.2 percent of employment. Mining has transformed Mongolia in a few decades from a traditional agricultural- and herding-based economy to a resource-based economy (World Bank Enterprise Survey 2013). Today, Mongolia has a per capita GDP of USD 13,000 (PPP), which makes its economy comparable to that of South Africa or Sri Lanka. Agriculture nonetheless remains the largest employer in the country and still absorbs 30 percent of the labor force (down from 46 percent in 2003), followed by services (17 percent), public administration (15 percent), and manufacturing (10 percent) (World Bank Enterprise Survey 2013). Structural economic changes have been accompanied by two important demographic shifts that have imposed additional pressure on labor markets. Mongolia's population is relatively young: the median age is 26 years, while the youth share of the working-age population is close to 55 percent. Likewise, the clearly declining trend of employment in herding and agricultural is mirrored by strong internal migration from provinces to the city of Ulaanbaatar. Indeed, since the late 1990s, population growth in Ulaanbaatar has averaged 5.8 percent, twice the country's population growth rate. In 2013, 20 percent of the youth in Ulaanbaatar were new migrants who had moved to the city in the previous five years (Shatz et al. 2015). Before the transition, the Mongolian labor market was characterized by high labor-force participation, which reached more than 75 percent in the early 1990s. According to official statistics, unemployment was nonexistent because a policy was in place that “everyone should work.” Only after 1992 did Mongolia begin measuring and reporting unemployment rates. Economic reforms implemented in the 1990s and afterward, including privatization and price liberalization, led to structural changes in Mongolia's economy, vastly transforming Mongolian labor markets. Indeed, labor-force participation began to fall consistently while the youth unemployment rates increased dramatically. In 2013, the country-level labor-force participation rate was 56 percent and 42 percent, respectively, for men and women aged 15–29. In Ulaanbaatar, the labor-participation rates of youth had fallen into the range of 30 percent for young women and 40 percent for young men (Shatz et al. 2015). Importantly, the share of inactive youth aged 15–29 has remained consistently above 20 percent. In 2013, the NEET rate for men and women aged 15–29 in Ulaanbaatar reached 13 and 27 percent. As in other developing settings, the Mongolian labor market shows stark gender disparities as a result of social norms and different employment frameworks for men and women (Pastore 2010; Schmillen and Sandig 2018). Social norms or deeply rooted expectations regarding the state's responsibility for providing jobs may have played some role in this emerging picture. In the study sample, for instance, measuring the subjective belief that it is a “government responsibility to provide jobs to us” resulted in a mean response of 85 points (on a scale of 0 to 100 points). On the demand side, the role of labor-market regulations that directly affect the costs of hiring and firing may constitute a relevant barrier to entry. Lehmann and Muravyev (2012) described the drastic transformation of labor-market legislation frameworks in transitional countries and their subsequent role in youth unemployment. According to the World Bank Enterprise Survey (World Bank 2013), less than 3 percent of Mongolian firms reported labor-market legislation as a constraint on labor demand. Yet, a quarter of sampled firms stated that the main obstacle to hiring was the inadequacy of skills in the workforce and mismatching between prospective employees’ skills and market needs. If such mismatching exists, it is important to assess the effectiveness of training programs tailored toward providing the vocational skills that are in demand in the market. One potential barrier for such investments could be the existence of compressed salary scales – which are typical under centrally planned economies – because firms might be unable to reward applicants who have additional skills. The evidence, however, is not clear in this regard. In 2012, large and increasing labor-market returns were observed for each additional year of schooling.2 Nonetheless, firm-level data showed that only 22.6 percent of Mongolian enterprises had a performance-based wage system, and only 7.9 percent had skill-based wage systems (Research Institute of Labor and Social Protection 2017). Furthermore, although there is no legislation that prevents firms from adopting new screening technologies in the labor markets, the awareness and utilization rate of government employment services is very low (less than 20 percent of idle youth in Ulaanbaatar indicated they were aware of employment websites, electronic job boards, and other types of public support; see Shatz et al. 2015). 3. The Mongolian Vocational-Training Program (MVTP) The MVTP was introduced in 2003 to counteract high levels of unemployment and idleness. Its primary goal was to help those who were poor, unemployed, or vulnerable to unemployment to find jobs through a combination of vocational-skills training and internships in business firms. This program remains the oldest and most extensive labor-market policy in effect in Mongolia. The MVTP follows key aspects of the so-called demand-driven training approach, which is designed to align prospective employees’ skills with the real needs of local employers (Betcherman et al. 2007). The rationale for demand-driven training programs is twofold. First, private institutions can offer relevant, up-to-date training that decentralizes the traditional supply of vocational training by public institutions. Established training providers therefore usually bid to offer training in trades with high labor demand. Second, by combining traditional classroom education with on-the-job training in the form of internships, MVTP can provide training in trades that the market demands.3 The Employment Promotion Service Center (EPSC) and the Metropolitan Employment Department (MED), which are both within the Mongolian Ministry of Labor, oversaw the implementation of this training program. The MVTP is financed by the State Employment Promotion Fund, and according to administrative data from the Ministry of Labor the total number of participants was 8,000 by 2011, and total program expenditure was approximately 3.5 billion MNT (USD 2.1 million). The EPSC selects training institutions through a competitive bidding process. The selection criteria for training institutions include evidence of ability to provide adequate training, legal registration, curriculum quality, teaching quality, adequacy of training sites, and, importantly, the ability to place trainees in internship positions with registered private employers. Although the MVTP began in 2003, its effectiveness has never been assessed. This paper focuses on the 2013 call that deliberately used an RCT to identify and measure MVTP impacts. In that year, the EPSC selected 47 training institutions in Ulaanbaatar that offered 141 courses in 6 family vocational skills, including construction, hairdressing, cooking, and heavy-machinery operation. The length of training varied from 20 to 45 days, depending upon vocation, with a minimum duration of 144 hours per course. According to the program's regulations, traditional classroom teaching did not exceed 30 percent of total hours, and practical, on-the-job training and internships accounted for the rest. The MVTP tuition was set at between 140,000 MNT and 220,000 MNT per participant in 2013 (approximately USD 90–140). These fees were paid directly by MED to training providers. The program offered no additional benefits such as transportation, meals, or insurance to trainees. As a result of budgetary constraints, the official number of training slots was set to 1,400 in 2013. Comparative information on course length, costs, take-up rates, attrition rates, and treatment effects across 11 youth training RCT interventions in developing countries is reported in table 1. The length of training ranged from two-and-a-half months to two years for classroom training plus two to six months for internships. The cost of training ranged from USD $330 to USD $1,700 dollars per trainee, with the exception of Uganda (Bandiera et al. 2020). Compared to other youth vocational-training programs, then, the MVTP was shorter and less expensive. Table 1. An Overview of Youth Vocational Training Programs Worldwide . . . . . Course length . . . . . . . . . . Country . NEET (%) . LFP youth (%) . Study . Target population . Classroom training . Internship . Take-up (%) . Attrition (%) . Time frame . Employment . Formal employment . Impacts on Earnings . Formal earnings . Monthly income . Cost . Argentina 19.70 38.50 Alzúa et al. (2016) Low-income youth 3.5 months up to 4 months 0.66 0 18 months n.r 8 n.r 64.9 83 1722 [0.7, 15.3] [17.1, 112.7] 0 33 months n.r 4.3 n.r 23.1 45 [−3.6, 12.1] [−15.3, 61.5] 0 48 months n.r 1.4 n.r n.r [−7.2, 10] Colombia 20.94 55.46 Attanasio, Kugler, and Meghir (2011) Low-income youth 3 months 3 months 0.99 0.18 14 months 4.5 6.4 11.6 27.1 12.8 750 [1.0, 8.0] [3.2, 9.6] [4.5, 18.7] [12.8, 41.3] Dominican 20.47 44.77 Card et al. (2011) Low-income youth max 350 hours 2 months 0.83 0.38 12 months 0.7 2.2 10.8 n.r 10 330 Republic [−4.6, 6.0] [−2.3, 6.7] [−4.2, 25.7] Ibarrarán et al. (2014) Low-income youth 225 hours 3 months 0.96 0.20 6 years −1.4 2.6 −1.9 n.r −2.3 700 [−4.4, 1.6] [−0.5,5.5] [−10.0, 6.3] Acevedo et al. (2017) Low-income youth 225 hours 240 hours n.r 0.17 3 years 0.7 n.r n.r. n.r n.r n.r. [−4.0, 5.3] Kenya 13.73 36.69 Honorati (2015) Low-income youth 3 months 3 month 0.90 0.23 14 months 5.6 n.r 29.7 n.r 47.5 1150 [0.9, 10.3] [−2.9, 62.3] Malawi 32.90 64.53 Cho et al. (2013) Low-income youth 3 months 0.91 0.46 4 months n.r n.r −19.6 n.r −5 n.r. [−63.9, 24.7] Peru 18.45 60.43 Diíz and Rosas (2016) Low-income youth 3 months 3 months 0.99 0.35 36 months 1.6 3.8 13.4 n.r n.r. 420 [−3.3, 6.5] [0.3, 7.3] [−17.6, 44.4] Brazil 21.50 54.80 Calero et al. (2014) Low-income youth 6 months 0.90 0.23 1–5 months 5.1 n.r n.r n.r 14 2225 [−7.4, 17.6] Liberia 15.10 30.50 Adoho et al. (2014) Young women 6 months 6 months 0.91 0.20 7 months 10.1 n.r 27.4 n.r 2.9 1200 [−2.1, 18.1] [−22.4, 77.3] Uganda 33.50 54.40 Bandiera et al. (2020) Adolescent girls 2 years n.r 0.18 2 years 6.8 n.r 18.9 n.r 3.1 17.9 [3.7, 9.9] [−106.1, 143.9] 0.41 4 years 4.9 308.8 50 [1.00, 8.80] [10.3, 607.2] . . . . . Course length . . . . . . . . . . Country . NEET (%) . LFP youth (%) . Study . Target population . Classroom training . Internship . Take-up (%) . Attrition (%) . Time frame . Employment . Formal employment . Impacts on Earnings . Formal earnings . Monthly income . Cost . Argentina 19.70 38.50 Alzúa et al. (2016) Low-income youth 3.5 months up to 4 months 0.66 0 18 months n.r 8 n.r 64.9 83 1722 [0.7, 15.3] [17.1, 112.7] 0 33 months n.r 4.3 n.r 23.1 45 [−3.6, 12.1] [−15.3, 61.5] 0 48 months n.r 1.4 n.r n.r [−7.2, 10] Colombia 20.94 55.46 Attanasio, Kugler, and Meghir (2011) Low-income youth 3 months 3 months 0.99 0.18 14 months 4.5 6.4 11.6 27.1 12.8 750 [1.0, 8.0] [3.2, 9.6] [4.5, 18.7] [12.8, 41.3] Dominican 20.47 44.77 Card et al. (2011) Low-income youth max 350 hours 2 months 0.83 0.38 12 months 0.7 2.2 10.8 n.r 10 330 Republic [−4.6, 6.0] [−2.3, 6.7] [−4.2, 25.7] Ibarrarán et al. (2014) Low-income youth 225 hours 3 months 0.96 0.20 6 years −1.4 2.6 −1.9 n.r −2.3 700 [−4.4, 1.6] [−0.5,5.5] [−10.0, 6.3] Acevedo et al. (2017) Low-income youth 225 hours 240 hours n.r 0.17 3 years 0.7 n.r n.r. n.r n.r n.r. [−4.0, 5.3] Kenya 13.73 36.69 Honorati (2015) Low-income youth 3 months 3 month 0.90 0.23 14 months 5.6 n.r 29.7 n.r 47.5 1150 [0.9, 10.3] [−2.9, 62.3] Malawi 32.90 64.53 Cho et al. (2013) Low-income youth 3 months 0.91 0.46 4 months n.r n.r −19.6 n.r −5 n.r. [−63.9, 24.7] Peru 18.45 60.43 Diíz and Rosas (2016) Low-income youth 3 months 3 months 0.99 0.35 36 months 1.6 3.8 13.4 n.r n.r. 420 [−3.3, 6.5] [0.3, 7.3] [−17.6, 44.4] Brazil 21.50 54.80 Calero et al. (2014) Low-income youth 6 months 0.90 0.23 1–5 months 5.1 n.r n.r n.r 14 2225 [−7.4, 17.6] Liberia 15.10 30.50 Adoho et al. (2014) Young women 6 months 6 months 0.91 0.20 7 months 10.1 n.r 27.4 n.r 2.9 1200 [−2.1, 18.1] [−22.4, 77.3] Uganda 33.50 54.40 Bandiera et al. (2020) Adolescent girls 2 years n.r 0.18 2 years 6.8 n.r 18.9 n.r 3.1 17.9 [3.7, 9.9] [−106.1, 143.9] 0.41 4 years 4.9 308.8 50 [1.00, 8.80] [10.3, 607.2] Source: Authors' elaboration based on McKenzie (2017); Card, Kluve, and Weber (2018); Kluve et al. (2019). Note: NEET is the acronym for “not in education, employment, or training.” Timeframe refers to time since the end of the intervention before measuring follow‐up outcomes. n.r. denotes not recorded.; 95 percent confidence intervals shown in parentheses. Estimates are the Intention‐to‐Treat estimates reported in different studies. Impacts on employment are in terms of percentage points, impacts on earnings in terms of percentage growth relative to control mean. Monthly income and Cost are expressed in USD. Open in new tab Table 1. An Overview of Youth Vocational Training Programs Worldwide . . . . . Course length . . . . . . . . . . Country . NEET (%) . LFP youth (%) . Study . Target population . Classroom training . Internship . Take-up (%) . Attrition (%) . Time frame . Employment . Formal employment . Impacts on Earnings . Formal earnings . Monthly income . Cost . Argentina 19.70 38.50 Alzúa et al. (2016) Low-income youth 3.5 months up to 4 months 0.66 0 18 months n.r 8 n.r 64.9 83 1722 [0.7, 15.3] [17.1, 112.7] 0 33 months n.r 4.3 n.r 23.1 45 [−3.6, 12.1] [−15.3, 61.5] 0 48 months n.r 1.4 n.r n.r [−7.2, 10] Colombia 20.94 55.46 Attanasio, Kugler, and Meghir (2011) Low-income youth 3 months 3 months 0.99 0.18 14 months 4.5 6.4 11.6 27.1 12.8 750 [1.0, 8.0] [3.2, 9.6] [4.5, 18.7] [12.8, 41.3] Dominican 20.47 44.77 Card et al. (2011) Low-income youth max 350 hours 2 months 0.83 0.38 12 months 0.7 2.2 10.8 n.r 10 330 Republic [−4.6, 6.0] [−2.3, 6.7] [−4.2, 25.7] Ibarrarán et al. (2014) Low-income youth 225 hours 3 months 0.96 0.20 6 years −1.4 2.6 −1.9 n.r −2.3 700 [−4.4, 1.6] [−0.5,5.5] [−10.0, 6.3] Acevedo et al. (2017) Low-income youth 225 hours 240 hours n.r 0.17 3 years 0.7 n.r n.r. n.r n.r n.r. [−4.0, 5.3] Kenya 13.73 36.69 Honorati (2015) Low-income youth 3 months 3 month 0.90 0.23 14 months 5.6 n.r 29.7 n.r 47.5 1150 [0.9, 10.3] [−2.9, 62.3] Malawi 32.90 64.53 Cho et al. (2013) Low-income youth 3 months 0.91 0.46 4 months n.r n.r −19.6 n.r −5 n.r. [−63.9, 24.7] Peru 18.45 60.43 Diíz and Rosas (2016) Low-income youth 3 months 3 months 0.99 0.35 36 months 1.6 3.8 13.4 n.r n.r. 420 [−3.3, 6.5] [0.3, 7.3] [−17.6, 44.4] Brazil 21.50 54.80 Calero et al. (2014) Low-income youth 6 months 0.90 0.23 1–5 months 5.1 n.r n.r n.r 14 2225 [−7.4, 17.6] Liberia 15.10 30.50 Adoho et al. (2014) Young women 6 months 6 months 0.91 0.20 7 months 10.1 n.r 27.4 n.r 2.9 1200 [−2.1, 18.1] [−22.4, 77.3] Uganda 33.50 54.40 Bandiera et al. (2020) Adolescent girls 2 years n.r 0.18 2 years 6.8 n.r 18.9 n.r 3.1 17.9 [3.7, 9.9] [−106.1, 143.9] 0.41 4 years 4.9 308.8 50 [1.00, 8.80] [10.3, 607.2] . . . . . Course length . . . . . . . . . . Country . NEET (%) . LFP youth (%) . Study . Target population . Classroom training . Internship . Take-up (%) . Attrition (%) . Time frame . Employment . Formal employment . Impacts on Earnings . Formal earnings . Monthly income . Cost . Argentina 19.70 38.50 Alzúa et al. (2016) Low-income youth 3.5 months up to 4 months 0.66 0 18 months n.r 8 n.r 64.9 83 1722 [0.7, 15.3] [17.1, 112.7] 0 33 months n.r 4.3 n.r 23.1 45 [−3.6, 12.1] [−15.3, 61.5] 0 48 months n.r 1.4 n.r n.r [−7.2, 10] Colombia 20.94 55.46 Attanasio, Kugler, and Meghir (2011) Low-income youth 3 months 3 months 0.99 0.18 14 months 4.5 6.4 11.6 27.1 12.8 750 [1.0, 8.0] [3.2, 9.6] [4.5, 18.7] [12.8, 41.3] Dominican 20.47 44.77 Card et al. (2011) Low-income youth max 350 hours 2 months 0.83 0.38 12 months 0.7 2.2 10.8 n.r 10 330 Republic [−4.6, 6.0] [−2.3, 6.7] [−4.2, 25.7] Ibarrarán et al. (2014) Low-income youth 225 hours 3 months 0.96 0.20 6 years −1.4 2.6 −1.9 n.r −2.3 700 [−4.4, 1.6] [−0.5,5.5] [−10.0, 6.3] Acevedo et al. (2017) Low-income youth 225 hours 240 hours n.r 0.17 3 years 0.7 n.r n.r. n.r n.r n.r. [−4.0, 5.3] Kenya 13.73 36.69 Honorati (2015) Low-income youth 3 months 3 month 0.90 0.23 14 months 5.6 n.r 29.7 n.r 47.5 1150 [0.9, 10.3] [−2.9, 62.3] Malawi 32.90 64.53 Cho et al. (2013) Low-income youth 3 months 0.91 0.46 4 months n.r n.r −19.6 n.r −5 n.r. [−63.9, 24.7] Peru 18.45 60.43 Diíz and Rosas (2016) Low-income youth 3 months 3 months 0.99 0.35 36 months 1.6 3.8 13.4 n.r n.r. 420 [−3.3, 6.5] [0.3, 7.3] [−17.6, 44.4] Brazil 21.50 54.80 Calero et al. (2014) Low-income youth 6 months 0.90 0.23 1–5 months 5.1 n.r n.r n.r 14 2225 [−7.4, 17.6] Liberia 15.10 30.50 Adoho et al. (2014) Young women 6 months 6 months 0.91 0.20 7 months 10.1 n.r 27.4 n.r 2.9 1200 [−2.1, 18.1] [−22.4, 77.3] Uganda 33.50 54.40 Bandiera et al. (2020) Adolescent girls 2 years n.r 0.18 2 years 6.8 n.r 18.9 n.r 3.1 17.9 [3.7, 9.9] [−106.1, 143.9] 0.41 4 years 4.9 308.8 50 [1.00, 8.80] [10.3, 607.2] Source: Authors' elaboration based on McKenzie (2017); Card, Kluve, and Weber (2018); Kluve et al. (2019). Note: NEET is the acronym for “not in education, employment, or training.” Timeframe refers to time since the end of the intervention before measuring follow‐up outcomes. n.r. denotes not recorded.; 95 percent confidence intervals shown in parentheses. Estimates are the Intention‐to‐Treat estimates reported in different studies. Impacts on employment are in terms of percentage points, impacts on earnings in terms of percentage growth relative to control mean. Monthly income and Cost are expressed in USD. Open in new tab The contractual framework for the market-based MVTP approach follows two levels. First, the EPSC signs contractual agreements with training providers regarding course length and content, timelines, and payments and internships for trainees. Second, individuals assigned to treatment are invited to sign an agreement with the corresponding labor office to secure participation. Up until Spring 2013, participants were required to sign a “trilateral” contract that involved a labor division district officer, an MED officer, and the beneficiary. This contract did not include any obligation on the trainees’ side to secure jobs after the completion of the classroom and internship phases. Subsequently, the MED changed to a “quadrilateral” agreement that added prospective employers as a fourth signatory. In practice, this meant that trainees needed to sign a job agreement with business firms to secure job placements after the completion of the program. Although these job agreements were not legally binding, the institutional change created an important slowdown in registration and enrollment, which in turn affected the program's take-up rates, as shown in section 5. A Complementary Treatment Design Because no information exists regarding the effectiveness of this ALMP, or any similar intervention in Mongolia, the average participant may not be aware of the labor-market benefits of completing training. Indeed, administrative data showed that dropout rates were particularly high for the MVTP. This lack of information about the benefits of vocational training is even more important in the context of Mongolia's very rapid transition from a state-controlled to a market-based economy. Less than 20 percent of unemployed Mongolian youth in Ulaanbaatar indicated they were aware of public-information sources regarding employment, while less than 5 percent made use of them in 2014 (Shatz et al. 2015). Thus, inspired by the work of Jensen (2010) in the context of educational interventions, an information treatment has been added to the study's evaluation framework. Participants are randomly selected to receive weekly newsletters with information about labor-market outcomes among skilled and unskilled workers in Mongolia.4 The newsletters stated market wages for occupations in sectors that were similar to those of the trainees and compared wages for jobs filled by unskilled workers. This information was estimated based on the 2012 Mongolian Household Survey. Figure S1.1 in the supplementary online appendix shows a typical newsletter submitted to trainees. This intervention combines two salient features of a relatively new stream of literature in development economics that has highlighted the role of framing and message repetition (e.g., Chong 2011). On the one hand, it provides (new) information on market returns to vocational training to participants who are typically information constrained. On the other hand, it provides repetitive (weekly) information to target individuals with potential limited attention or present-bias individuals who, given a choice between a payoff today (dropout from training for work) and a payoff in the future (complete training for subsequent work), will choose to have the payoff now (Gabaix and Laibson 2006).” Information theory states that the value of information is determined by three important factors – credibility, novelty, and ability and willingness to act based on updated beliefs – all of which involve different forces and trade-offs (Hirshleifer and Riley 1992). Individuals process new information largely based on their prior beliefs. If individuals change behavior in response to new information, then that information has value to them. Thus, the value of these newsletters among participants could not be considered a given. On the one hand, disadvantaged participants do not have easy access to labor-market information, so it is arguably the case that information on training returns was new to the average participant. On the other hand, MVTP is a government program, and the erosion of trust in government on the part of disadvantaged populations may affect their perceptions of the credibility of information. Willingness to act based on new information could also be affected by commonly reported biases among unemployed individuals regarding their employment prospects (e.g., overoptimism). Such biases distort their behavior and may cause them to do little to leave unemployment (Spinnewijn 2015). The ways in which such factors interacted each other, then, would determine the value of the information to participants. 4. Evaluation Design The randomization scheme for the two RCTs is covered in section 3 is depicted in the fig. S1.2 in the supplementary online appendix. Prospective trainees aged 15–30 were assessed at their local district office. After completing a short baseline questionnaire, each applicant was screened for eligibility by an administrative officer. All suitable applicants were then sent to a district labor office in which each applicant, alongside a labor officer, chose her preferred trade courses on a first-come, first-served basis. Individuals who chose unavailable trade courses were not included in the experiment unless they opted for an alternative available trade.5 At this stage, after prospective beneficiaries chose their courses, randomization at the individual level proceeded, and information from all district labor offices was pooled at the end of each day. The random allocation of MVTP beneficiaries took place between August 26 and November 22, 2013. Each day during that period, the research team in coordination with MED officers, randomly assigned eligible applicants to either the treatment or control groups based on a 2:1 rule. As a result, the treatment group was made up of 774 trainees (65.2 percent), and 414 (34.8 percent) were assigned to the control group for a total of 1,188 participants.6 Individuals randomly assigned to the treatment group were notified by telephone or through in-house visits about the program benefits, timelines, and regulations in the days that followed. Individuals who chose unavailable trade courses were not included in the sampling design. Individuals in the treatment group were distributed across 141 courses at 47 training centers. Class size varied highly across training courses and ranged from 3 to 30 participants. The average number of MED-funded students was 9.6 per class. Actual class sizes may have been slightly larger, however, because training institutions were entitled to recruit privately funded students if maximum class size was not reached. Though this happened occasionally, the significance was marginal, and most courses maintained small class sizes. The second random allocation, that is, the newsletter treatment, was implemented after beneficiaries were assigned to a vocational-training treatment group. The allocation of weekly newsletters was implemented at the class level, rather than at an individual level, to prevent spillover effects. Overall, 101 classes were randomly assigned to the treatment-newsletter group and 40 classes to the control group. Thus, the final sample size was not known a priori and was determined after assigning classes to newsletter treatment and control groups. The number of letters trainees received in the information-treatment group varied according to the length of training. Individuals assigned to the newsletter-treatment group received from 0 to 4 newsletters, with an average of 2.25. Across vocational trades, the number of letters varied from an average of 3.24 (cooking) to 1.66 (services). Because the total number of letters provided to participants varied depending on differences in the training, and because people who chose longer trainings may had unobservable characteristics that were correlated with training completion, the evaluation framework focused only on the comparison of the treatment and control information groups without considering the intensive margin of the treatment itself. 5. Data, Balancing Tests, and Take-Up Assessment The empirical framework used in this study is based on individual-level survey data, including a baseline, collected in Fall 2013 and two follow-up surveys administered 6 and 18 months after training. The timeline of the intervention and data collection is illustrated in fig. 1. The evaluation data includes socio-demographic variables, participation in formal schooling and training, labor-market outcomes, and detailed information on participation in the MVTP. Attrition reached 5.4 percent at baseline, 9.6 percent at the first follow-up survey, and 15 percent at the second follow-up survey. As shown in table 1, and relative to similar interventions with similar time span for the measurement of treatment effects, the attrition rate in this study is moderate.7 Figure 1. Open in new tabDownload slide Timeline of the VT Intervention Mongolian VT Program, 2014–2016 Source: Authors' elaboration. Figure 1. Open in new tabDownload slide Timeline of the VT Intervention Mongolian VT Program, 2014–2016 Source: Authors' elaboration. Tables S1.1 and S1.2 in the supplementary online appendix provide results of tests for differences in baseline variables across treatment and control units for the 6- and 18-month follow-up samples conditional on response status. For most baseline covariates, no systematic differences are observed between individuals in the treatment and control groups after conditioning on response status. Only for a small number of baseline covariates is the null of equality of means by treatment status rejected, after conditioning on response status. Finally, table S1.3 in the supplementary online appendix reports the overall rate of attrition by treatment status. The null hypothesis that the overall attrition rate is of equal magnitude between treatment and control groups cannot be rejected. Mean covariate balancing tests for the two experimental designs are shown in table 2, one for the allocation of training slots (left panel) and one for the allocation of weekly newsletters to individuals who received treatment (right panel). Baseline data show that the typical applicant is 23 years old, a woman (65 percent), and poor (82 percent lived in gers).8 Almost half of the participants are married (45 percent), have children (42 percent), and live with their parents (47 percent). On average, they show high levels of formal schooling (51 percent had completed high school) and prior labor experience (60 percent). The p-values for the coefficients of OLS models that regressed treatment status on baseline covariates (left panel) are above 0.10 in all cases, indicating that individuals in the treatment and control groups came from the same population. These baseline characteristics are broadly similar to those of poor individuals aged 15–30 in Ulaanbaatar as reported in the 2014 Mongolian socio-economic survey, which shows similar shares of idle youth, age mean composition, and average years schooling. The MVTP sample, however, shows a higher proportion of women (65 vs. 55 percent), a higher share of individuals living in gers (82 vs. 56 percent), smaller household size (4 vs. 6), and higher marriage rates (45 percent vs. 35 percent). Put differently, these differences suggest that the MVTP served, on average, a poorer set of Mongolian youth. Table 2. Balancing Test Across Experimental Groups Mongolian VT Program, 2014–2016 . Treatment I: MVTP . Treatment II: Information letters . . Treated . Control . p-value . Treated w/ letter . Treated w/o letter . p-value . Socio-demographics Gender (1 = males) 0.35 0.35 0.99 0.38 0.21 0.00 Age in years 22.97 22.94 0.76 23.43 23.46 0.97 Marital status (1 = married) 0.45 0.47 0.44 0.49 0.51 0.68 Residence (1 = Ger) 0.82 0.82 0.85 0.78 0.78 0.93 Less than high school 0.19 0.2 0.78 0.13 0.15 0.60 High school 0.51 0.5 0.51 0.58 0.51 0.22 Technical education 0.08 0.11 0.17 0.07 0.11 0.27 College + 0.21 0.19 0.67 0.22 0.24 0.75 Household size average persons 3.99 4.09 0.25 3.96 3.98 0.89 Has children 0.41 0.42 0.34 0.44 0.45 0.85 Live with parents 0.48 0.45 0.81 0.47 0.43 0.48 Parents have work 0.29 0.27 0.63 0.30 0.26 0.48 Has disability 0.05 0.05 0.98 0.06 0.05 0.87 Poverty index 0 0.01 0.87 0.11 0.14 0.84 Labor market and income Has work experience 0.61 0.62 0.68 0.68 0.68 0.86 # weeks of work experience 4.88 4.58 0.69 4.03 0.00 0.16 Previous vocational training 0.2 0.22 0.5 0.23 0.27 0.34 Out of LF (child care duties) 0.25 0.29 0.28 0.24 0.32 0.10 Out of LF (student) 0.1 0.09 0.54 0.06 0.10 0.16 Out of LF (homemaker) 0.06 0.07 0.44 0.04 0.05 0.78 No monthly income 0.54 0.55 0.9 0.49 0.48 0.85 Has income from remittances 0.06 0.06 0.79 0.06 0.09 0.26 Has labor market income 0.15 0.12 0.13 0.16 0.12 0.40 Receive welfare income 0.14 0.17 0.25 0.18 0.21 0.43 Expectations Subjective prob. of getting a job 0.78 0.8 0.28 0.77 0.80 0.31 Optimistic to get a job 0.67 0.7 0.46 0.71 0.75 0.85 Ambition to succeed in labor market 0.88 0.91 0.22 0.91 0.95 0.31 Personal responsibility to get a job 0.66 0.66 0.94 0.62 0.69 0.02 Government responsibility to provide a job 0.85 0.86 0.92 0.88 0.90 0.94 Plan to complete VT 0.94 0.96 0.45 0.94 0.95 0.39 Number of days plan to attend VT 34 35 0.19 34 34 0.93 Eligibility Eligible to VT due to unemployment status 0.85 0.84 0.58 0.83 0.88 0.13 Employment as main reason to join VT 0.78 0.81 0.21 0.77 0.79 0.67 Applied to cooking/baking VT courses 0.12 0.12 0.96 0.13 0.08 0.13 Applied to beauty/hairdressing VT courses 0.24 0.26 0.57 0.25 0.26 0.72 Applied to mechanical/machinery VT courses 0.26 0.23 0.31 0.29 0.11 0.00 Applied to craftsmanship VT courses 0.17 0.13 0.31 0.20 0.18 0.63 Applied to agriculture/gardening VT courses 0.13 0.13 0.78 0.08 0.24 0.00 Applied to services and others 0.07 0.13 0.33 0.04 0.13 0.00 . Treatment I: MVTP . Treatment II: Information letters . . Treated . Control . p-value . Treated w/ letter . Treated w/o letter . p-value . Socio-demographics Gender (1 = males) 0.35 0.35 0.99 0.38 0.21 0.00 Age in years 22.97 22.94 0.76 23.43 23.46 0.97 Marital status (1 = married) 0.45 0.47 0.44 0.49 0.51 0.68 Residence (1 = Ger) 0.82 0.82 0.85 0.78 0.78 0.93 Less than high school 0.19 0.2 0.78 0.13 0.15 0.60 High school 0.51 0.5 0.51 0.58 0.51 0.22 Technical education 0.08 0.11 0.17 0.07 0.11 0.27 College + 0.21 0.19 0.67 0.22 0.24 0.75 Household size average persons 3.99 4.09 0.25 3.96 3.98 0.89 Has children 0.41 0.42 0.34 0.44 0.45 0.85 Live with parents 0.48 0.45 0.81 0.47 0.43 0.48 Parents have work 0.29 0.27 0.63 0.30 0.26 0.48 Has disability 0.05 0.05 0.98 0.06 0.05 0.87 Poverty index 0 0.01 0.87 0.11 0.14 0.84 Labor market and income Has work experience 0.61 0.62 0.68 0.68 0.68 0.86 # weeks of work experience 4.88 4.58 0.69 4.03 0.00 0.16 Previous vocational training 0.2 0.22 0.5 0.23 0.27 0.34 Out of LF (child care duties) 0.25 0.29 0.28 0.24 0.32 0.10 Out of LF (student) 0.1 0.09 0.54 0.06 0.10 0.16 Out of LF (homemaker) 0.06 0.07 0.44 0.04 0.05 0.78 No monthly income 0.54 0.55 0.9 0.49 0.48 0.85 Has income from remittances 0.06 0.06 0.79 0.06 0.09 0.26 Has labor market income 0.15 0.12 0.13 0.16 0.12 0.40 Receive welfare income 0.14 0.17 0.25 0.18 0.21 0.43 Expectations Subjective prob. of getting a job 0.78 0.8 0.28 0.77 0.80 0.31 Optimistic to get a job 0.67 0.7 0.46 0.71 0.75 0.85 Ambition to succeed in labor market 0.88 0.91 0.22 0.91 0.95 0.31 Personal responsibility to get a job 0.66 0.66 0.94 0.62 0.69 0.02 Government responsibility to provide a job 0.85 0.86 0.92 0.88 0.90 0.94 Plan to complete VT 0.94 0.96 0.45 0.94 0.95 0.39 Number of days plan to attend VT 34 35 0.19 34 34 0.93 Eligibility Eligible to VT due to unemployment status 0.85 0.84 0.58 0.83 0.88 0.13 Employment as main reason to join VT 0.78 0.81 0.21 0.77 0.79 0.67 Applied to cooking/baking VT courses 0.12 0.12 0.96 0.13 0.08 0.13 Applied to beauty/hairdressing VT courses 0.24 0.26 0.57 0.25 0.26 0.72 Applied to mechanical/machinery VT courses 0.26 0.23 0.31 0.29 0.11 0.00 Applied to craftsmanship VT courses 0.17 0.13 0.31 0.20 0.18 0.63 Applied to agriculture/gardening VT courses 0.13 0.13 0.78 0.08 0.24 0.00 Applied to services and others 0.07 0.13 0.33 0.04 0.13 0.00 Source:Authors' calculations from Baseline Survey data. Note: p-values from OLS models of treatment status on each baseline covariate of interest. All values expressed as proportions unless otherwise specified. For treatment I, fixed-effects were included for day of random assignment. Sample size varies across covariates and ranges from 1185 to 1118 for treatment I and from 410 to 389 for treatment II. Open in new tab Table 2. Balancing Test Across Experimental Groups Mongolian VT Program, 2014–2016 . Treatment I: MVTP . Treatment II: Information letters . . Treated . Control . p-value . Treated w/ letter . Treated w/o letter . p-value . Socio-demographics Gender (1 = males) 0.35 0.35 0.99 0.38 0.21 0.00 Age in years 22.97 22.94 0.76 23.43 23.46 0.97 Marital status (1 = married) 0.45 0.47 0.44 0.49 0.51 0.68 Residence (1 = Ger) 0.82 0.82 0.85 0.78 0.78 0.93 Less than high school 0.19 0.2 0.78 0.13 0.15 0.60 High school 0.51 0.5 0.51 0.58 0.51 0.22 Technical education 0.08 0.11 0.17 0.07 0.11 0.27 College + 0.21 0.19 0.67 0.22 0.24 0.75 Household size average persons 3.99 4.09 0.25 3.96 3.98 0.89 Has children 0.41 0.42 0.34 0.44 0.45 0.85 Live with parents 0.48 0.45 0.81 0.47 0.43 0.48 Parents have work 0.29 0.27 0.63 0.30 0.26 0.48 Has disability 0.05 0.05 0.98 0.06 0.05 0.87 Poverty index 0 0.01 0.87 0.11 0.14 0.84 Labor market and income Has work experience 0.61 0.62 0.68 0.68 0.68 0.86 # weeks of work experience 4.88 4.58 0.69 4.03 0.00 0.16 Previous vocational training 0.2 0.22 0.5 0.23 0.27 0.34 Out of LF (child care duties) 0.25 0.29 0.28 0.24 0.32 0.10 Out of LF (student) 0.1 0.09 0.54 0.06 0.10 0.16 Out of LF (homemaker) 0.06 0.07 0.44 0.04 0.05 0.78 No monthly income 0.54 0.55 0.9 0.49 0.48 0.85 Has income from remittances 0.06 0.06 0.79 0.06 0.09 0.26 Has labor market income 0.15 0.12 0.13 0.16 0.12 0.40 Receive welfare income 0.14 0.17 0.25 0.18 0.21 0.43 Expectations Subjective prob. of getting a job 0.78 0.8 0.28 0.77 0.80 0.31 Optimistic to get a job 0.67 0.7 0.46 0.71 0.75 0.85 Ambition to succeed in labor market 0.88 0.91 0.22 0.91 0.95 0.31 Personal responsibility to get a job 0.66 0.66 0.94 0.62 0.69 0.02 Government responsibility to provide a job 0.85 0.86 0.92 0.88 0.90 0.94 Plan to complete VT 0.94 0.96 0.45 0.94 0.95 0.39 Number of days plan to attend VT 34 35 0.19 34 34 0.93 Eligibility Eligible to VT due to unemployment status 0.85 0.84 0.58 0.83 0.88 0.13 Employment as main reason to join VT 0.78 0.81 0.21 0.77 0.79 0.67 Applied to cooking/baking VT courses 0.12 0.12 0.96 0.13 0.08 0.13 Applied to beauty/hairdressing VT courses 0.24 0.26 0.57 0.25 0.26 0.72 Applied to mechanical/machinery VT courses 0.26 0.23 0.31 0.29 0.11 0.00 Applied to craftsmanship VT courses 0.17 0.13 0.31 0.20 0.18 0.63 Applied to agriculture/gardening VT courses 0.13 0.13 0.78 0.08 0.24 0.00 Applied to services and others 0.07 0.13 0.33 0.04 0.13 0.00 . Treatment I: MVTP . Treatment II: Information letters . . Treated . Control . p-value . Treated w/ letter . Treated w/o letter . p-value . Socio-demographics Gender (1 = males) 0.35 0.35 0.99 0.38 0.21 0.00 Age in years 22.97 22.94 0.76 23.43 23.46 0.97 Marital status (1 = married) 0.45 0.47 0.44 0.49 0.51 0.68 Residence (1 = Ger) 0.82 0.82 0.85 0.78 0.78 0.93 Less than high school 0.19 0.2 0.78 0.13 0.15 0.60 High school 0.51 0.5 0.51 0.58 0.51 0.22 Technical education 0.08 0.11 0.17 0.07 0.11 0.27 College + 0.21 0.19 0.67 0.22 0.24 0.75 Household size average persons 3.99 4.09 0.25 3.96 3.98 0.89 Has children 0.41 0.42 0.34 0.44 0.45 0.85 Live with parents 0.48 0.45 0.81 0.47 0.43 0.48 Parents have work 0.29 0.27 0.63 0.30 0.26 0.48 Has disability 0.05 0.05 0.98 0.06 0.05 0.87 Poverty index 0 0.01 0.87 0.11 0.14 0.84 Labor market and income Has work experience 0.61 0.62 0.68 0.68 0.68 0.86 # weeks of work experience 4.88 4.58 0.69 4.03 0.00 0.16 Previous vocational training 0.2 0.22 0.5 0.23 0.27 0.34 Out of LF (child care duties) 0.25 0.29 0.28 0.24 0.32 0.10 Out of LF (student) 0.1 0.09 0.54 0.06 0.10 0.16 Out of LF (homemaker) 0.06 0.07 0.44 0.04 0.05 0.78 No monthly income 0.54 0.55 0.9 0.49 0.48 0.85 Has income from remittances 0.06 0.06 0.79 0.06 0.09 0.26 Has labor market income 0.15 0.12 0.13 0.16 0.12 0.40 Receive welfare income 0.14 0.17 0.25 0.18 0.21 0.43 Expectations Subjective prob. of getting a job 0.78 0.8 0.28 0.77 0.80 0.31 Optimistic to get a job 0.67 0.7 0.46 0.71 0.75 0.85 Ambition to succeed in labor market 0.88 0.91 0.22 0.91 0.95 0.31 Personal responsibility to get a job 0.66 0.66 0.94 0.62 0.69 0.02 Government responsibility to provide a job 0.85 0.86 0.92 0.88 0.90 0.94 Plan to complete VT 0.94 0.96 0.45 0.94 0.95 0.39 Number of days plan to attend VT 34 35 0.19 34 34 0.93 Eligibility Eligible to VT due to unemployment status 0.85 0.84 0.58 0.83 0.88 0.13 Employment as main reason to join VT 0.78 0.81 0.21 0.77 0.79 0.67 Applied to cooking/baking VT courses 0.12 0.12 0.96 0.13 0.08 0.13 Applied to beauty/hairdressing VT courses 0.24 0.26 0.57 0.25 0.26 0.72 Applied to mechanical/machinery VT courses 0.26 0.23 0.31 0.29 0.11 0.00 Applied to craftsmanship VT courses 0.17 0.13 0.31 0.20 0.18 0.63 Applied to agriculture/gardening VT courses 0.13 0.13 0.78 0.08 0.24 0.00 Applied to services and others 0.07 0.13 0.33 0.04 0.13 0.00 Source:Authors' calculations from Baseline Survey data. Note: p-values from OLS models of treatment status on each baseline covariate of interest. All values expressed as proportions unless otherwise specified. For treatment I, fixed-effects were included for day of random assignment. Sample size varies across covariates and ranges from 1185 to 1118 for treatment I and from 410 to 389 for treatment II. Open in new tab For the newsletter-treatment group, on the other hand, the right panel in table 2 shows that the p-values for most variables do not reject the equality of means between experimentally determined treatment groups. Equality is rejected regarding a few variables, with the exception of those related to chosen vocations. This is expected because the random allocation of MVTP participants to the newsletter-treatment group occurred at the course level rather than at the individual level. Compared to other vocational-training programs for youth in developing countries (see table 1), administrative data reveal that lack of full compliance with the treatment is relatively high for the MVTP. Out of the 766 applicants randomly assigned to receive vocational-skills training, 439 (57 percent) did show up for training, while other studies have reported take-up rates from 77 to 99 percent. Table S1.4 in the supplementary online appendix gives detailed information on enrollment numbers. Self-reported survey information indicated that, among those who did not take up the treatment, 35 percent cited family and personal commitments (e.g., household chores or pregnancy), 30 percent had started a new job right after enrollment, and 31 percent said they were not able to comply with contractual MVTP requirements (i.e., quadrilateral contracts). From a policy standpoint, then, empirical assessment of the determinants of take-up becomes essential to gain insight both into the operation of the program (e.g., targeting, eligibility rules, and institutional requirements) and the identification and estimation of the parameters of interest. The results from linear-probability models are shown in table 3, where the dependent variable takes the value of 1 for those treatment-group individuals who received the treatment and 0 for those who did not. After including a rich set of socio-demographic and labor-market variables, MVTP institutional variables, and self-reported expectations regarding training and performance in the labor markets, results indicate that a handful of socio-demographic variables are statistically correlated with take-up decisions. On average, gender, household wealth, age, and formal schooling are important to take-up rates because women and individuals who were wealthier, older, and better-educated were more likely to participate in the program relative to male, poorer, younger, and less well-educated individuals. Importantly, findings show no meaningful statistical relationship between take-up and labor-market variables at baseline. This pattern ran counter to observations in other labor-market programs in which variables related to labor markets emerged as the main determinants of take-up. Table 3. Determinants of Take-Up for Vocational Training Program Mongolian VT Program, 2014–2016 . coeff. . std. error . Socio-demographics Age 20–24 −0.090* 0.050 Age 25–30 −0.016 0.058 Gender (1 = males) −0.094* 0.051 Marital status (1 = married) 0.023 0.050 Residence (1 = Ger) −0.047 0.046 Less than high school −0.109* 0.058 High school 0.036 0.045 Technical education 0.014 0.070 Household size −0.016 0.012 Has children 0.032 0.051 Live with parents 0.060 0.054 Parents have work 0.012 0.049 Has disability 0.087 0.085 Poverty index 0.029* 0.016 Labor market and income Has work experience 0.052 0.039 # weeks of work experience −0.000 0.000 Previous vocational training 0.004 0.041 Out of LF (childcare duties) −0.048 0.050 Out of LF (student) 0.019 0.061 Out of LF (homemaker) −0.036 0.074 No monthly income −0.004 0.057 Has income from remittances 0.071 0.086 Has labor market income 0.067 0.067 Has welfare income 0.121* 0.072 VT institutions Quadrilateral VT contracts −0.547*** 0.049 Ratio training slots/applicants −0.000 0.001 Eligible to VT due to unemployment status 0.018 0.048 Know about VT through media 0.132** 0.053 Know about VT through letter −0.073 0.052 Know about VT through Internet 0.080 0.071 Know about VT through local employment office 0.064 0.042 Applied to cooking VT courses −0.048 0.076 Applied to beauty/hairdressing VT courses −0.215*** 0.073 Applied to mechanical/machinery VT courses 0.171** 0.074 Applied to craftsmanship VT courses −0.206*** 0.075 Applied to agriculture/gardening VT courses −0.049 0.076 Expectations Optimistic to get a job 0.077** 0.038 Ambition to succeed in labor market 0.035 0.056 Personal responsibility to get a job −0.045** 0.017 Government responsibility to provide a job 0.068*** 0.017 Plan to complete VT −0.030 0.077 Number of days plan to attend VT −0.001 0.001 N 702 R2 0.32 p-value of F-test for joint demographic variables = 0 0.000 p-value of F-test for joint labor market variables = 0 0.209 p-value of F-test for joint subjective expectations variables = 0 0.000 p-value of F-test for joint VT institutions variables = 0 0.000 . coeff. . std. error . Socio-demographics Age 20–24 −0.090* 0.050 Age 25–30 −0.016 0.058 Gender (1 = males) −0.094* 0.051 Marital status (1 = married) 0.023 0.050 Residence (1 = Ger) −0.047 0.046 Less than high school −0.109* 0.058 High school 0.036 0.045 Technical education 0.014 0.070 Household size −0.016 0.012 Has children 0.032 0.051 Live with parents 0.060 0.054 Parents have work 0.012 0.049 Has disability 0.087 0.085 Poverty index 0.029* 0.016 Labor market and income Has work experience 0.052 0.039 # weeks of work experience −0.000 0.000 Previous vocational training 0.004 0.041 Out of LF (childcare duties) −0.048 0.050 Out of LF (student) 0.019 0.061 Out of LF (homemaker) −0.036 0.074 No monthly income −0.004 0.057 Has income from remittances 0.071 0.086 Has labor market income 0.067 0.067 Has welfare income 0.121* 0.072 VT institutions Quadrilateral VT contracts −0.547*** 0.049 Ratio training slots/applicants −0.000 0.001 Eligible to VT due to unemployment status 0.018 0.048 Know about VT through media 0.132** 0.053 Know about VT through letter −0.073 0.052 Know about VT through Internet 0.080 0.071 Know about VT through local employment office 0.064 0.042 Applied to cooking VT courses −0.048 0.076 Applied to beauty/hairdressing VT courses −0.215*** 0.073 Applied to mechanical/machinery VT courses 0.171** 0.074 Applied to craftsmanship VT courses −0.206*** 0.075 Applied to agriculture/gardening VT courses −0.049 0.076 Expectations Optimistic to get a job 0.077** 0.038 Ambition to succeed in labor market 0.035 0.056 Personal responsibility to get a job −0.045** 0.017 Government responsibility to provide a job 0.068*** 0.017 Plan to complete VT −0.030 0.077 Number of days plan to attend VT −0.001 0.001 N 702 R2 0.32 p-value of F-test for joint demographic variables = 0 0.000 p-value of F-test for joint labor market variables = 0 0.209 p-value of F-test for joint subjective expectations variables = 0 0.000 p-value of F-test for joint VT institutions variables = 0 0.000 Source: Authors' calculations from Baseline and First follow up data. Note: Linear probabilistic model on take-up for VT program. Dependent variable takes the value 1 for those treated units who attended program, 0 for the treated no-show units. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 3. Determinants of Take-Up for Vocational Training Program Mongolian VT Program, 2014–2016 . coeff. . std. error . Socio-demographics Age 20–24 −0.090* 0.050 Age 25–30 −0.016 0.058 Gender (1 = males) −0.094* 0.051 Marital status (1 = married) 0.023 0.050 Residence (1 = Ger) −0.047 0.046 Less than high school −0.109* 0.058 High school 0.036 0.045 Technical education 0.014 0.070 Household size −0.016 0.012 Has children 0.032 0.051 Live with parents 0.060 0.054 Parents have work 0.012 0.049 Has disability 0.087 0.085 Poverty index 0.029* 0.016 Labor market and income Has work experience 0.052 0.039 # weeks of work experience −0.000 0.000 Previous vocational training 0.004 0.041 Out of LF (childcare duties) −0.048 0.050 Out of LF (student) 0.019 0.061 Out of LF (homemaker) −0.036 0.074 No monthly income −0.004 0.057 Has income from remittances 0.071 0.086 Has labor market income 0.067 0.067 Has welfare income 0.121* 0.072 VT institutions Quadrilateral VT contracts −0.547*** 0.049 Ratio training slots/applicants −0.000 0.001 Eligible to VT due to unemployment status 0.018 0.048 Know about VT through media 0.132** 0.053 Know about VT through letter −0.073 0.052 Know about VT through Internet 0.080 0.071 Know about VT through local employment office 0.064 0.042 Applied to cooking VT courses −0.048 0.076 Applied to beauty/hairdressing VT courses −0.215*** 0.073 Applied to mechanical/machinery VT courses 0.171** 0.074 Applied to craftsmanship VT courses −0.206*** 0.075 Applied to agriculture/gardening VT courses −0.049 0.076 Expectations Optimistic to get a job 0.077** 0.038 Ambition to succeed in labor market 0.035 0.056 Personal responsibility to get a job −0.045** 0.017 Government responsibility to provide a job 0.068*** 0.017 Plan to complete VT −0.030 0.077 Number of days plan to attend VT −0.001 0.001 N 702 R2 0.32 p-value of F-test for joint demographic variables = 0 0.000 p-value of F-test for joint labor market variables = 0 0.209 p-value of F-test for joint subjective expectations variables = 0 0.000 p-value of F-test for joint VT institutions variables = 0 0.000 . coeff. . std. error . Socio-demographics Age 20–24 −0.090* 0.050 Age 25–30 −0.016 0.058 Gender (1 = males) −0.094* 0.051 Marital status (1 = married) 0.023 0.050 Residence (1 = Ger) −0.047 0.046 Less than high school −0.109* 0.058 High school 0.036 0.045 Technical education 0.014 0.070 Household size −0.016 0.012 Has children 0.032 0.051 Live with parents 0.060 0.054 Parents have work 0.012 0.049 Has disability 0.087 0.085 Poverty index 0.029* 0.016 Labor market and income Has work experience 0.052 0.039 # weeks of work experience −0.000 0.000 Previous vocational training 0.004 0.041 Out of LF (childcare duties) −0.048 0.050 Out of LF (student) 0.019 0.061 Out of LF (homemaker) −0.036 0.074 No monthly income −0.004 0.057 Has income from remittances 0.071 0.086 Has labor market income 0.067 0.067 Has welfare income 0.121* 0.072 VT institutions Quadrilateral VT contracts −0.547*** 0.049 Ratio training slots/applicants −0.000 0.001 Eligible to VT due to unemployment status 0.018 0.048 Know about VT through media 0.132** 0.053 Know about VT through letter −0.073 0.052 Know about VT through Internet 0.080 0.071 Know about VT through local employment office 0.064 0.042 Applied to cooking VT courses −0.048 0.076 Applied to beauty/hairdressing VT courses −0.215*** 0.073 Applied to mechanical/machinery VT courses 0.171** 0.074 Applied to craftsmanship VT courses −0.206*** 0.075 Applied to agriculture/gardening VT courses −0.049 0.076 Expectations Optimistic to get a job 0.077** 0.038 Ambition to succeed in labor market 0.035 0.056 Personal responsibility to get a job −0.045** 0.017 Government responsibility to provide a job 0.068*** 0.017 Plan to complete VT −0.030 0.077 Number of days plan to attend VT −0.001 0.001 N 702 R2 0.32 p-value of F-test for joint demographic variables = 0 0.000 p-value of F-test for joint labor market variables = 0 0.209 p-value of F-test for joint subjective expectations variables = 0 0.000 p-value of F-test for joint VT institutions variables = 0 0.000 Source: Authors' calculations from Baseline and First follow up data. Note: Linear probabilistic model on take-up for VT program. Dependent variable takes the value 1 for those treated units who attended program, 0 for the treated no-show units. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Moreover, institutional variables related to the operation of the MVTP emerge as important take-up predictors. Individuals who were required to present quadrilateral contracts, for example, are 54 percentage points less likely to take up the treatment relative to individuals asked to present trilateral contracts. This institutional requirement constituted a critical barrier for take-up. Likewise, the likelihood of attending the training is statistically related to the vocational-skills courses that participants chose. Individuals who initially selected courses related to hairdressing and craftsmanship are less likely to attend training (−20 percentage points), while individuals who selected mechanical- and machinery-related courses are more likely to participate (+18 percentage points). Results also show that take-up decisions are associated with self-reported expectations regarding training, labor-markets, and the role of government in facilitating jobs for youth (see table 3). Individuals who felt optimistic about getting a job or those who believed that the government should play a major role in helping youth find jobs took up the treatment disproportionately. In contrast, individuals who believe that getting a job is a personal responsibility are less likely to take up the treatment. Overall, as shown by the p-values of joint significance at the bottom of table 3, socio-demographic and institutional variables (i.e., contractual agreements) are the most important predictors of take-up rates, while prior labor-market outcomes are not statistically associated with take-up decisions. 6. Empirical Framework and Findings The study estimates two parameters of interest for the sample with complete survey information: intent-to-treat (ITT) and effective treatment-on-the-treated (TOT). The estimate of the intent-to-treat parameter follows a standard, multivariate linear regression function of the form: $$\begin{equation}{Y_i} = \alpha + \beta {Z_i} + X_i^\prime \gamma + {\tau _i} + { \in _i}\end{equation}$$(1) where |${Y_i}$| is the outcome of interest for individual i, |${Z_i}$| is the treatment indicator (1 for those offered treatment and 0 otherwise). |${X_i}$| are baseline control variables, and |${ \in _i}$| is the error term. Because there are as many experimental groups as the number of days the random allocation lasted, equation (1) also includes fixed effects by day of random assignment, τi.9 The effective treatment-on-the-treated (TOT) parameter, on the other hand, is estimated by the 2SLS estimator following an instrumental-variable approach in which actual participation in the training program (T) is instrumented by treatment status (Z): $$\begin{equation}\left\{ {\begin{array}{@{}*{1}{l}@{}} {{Y_i} = \alpha + \beta {T_i} + X_i^\prime \gamma + {\tau _i} + { \in _i}}\\ {{T_i} = \delta {Z_i} + X_i^\prime \gamma + {\tau _i} + {\varepsilon _i}} \end{array}} \right.\end{equation}$$(2) Table S1.5 in the supplemental online appendix shows the first-stage estimation results. The coefficient associated with the instrumental variable Z is statistically significant at the 1 percent level, and the resulting F-statistics confirm the relevance and strength of the instrument. The main results for the vocational training intervention are presented in table 4. Four outcomes of interest are examined: employment in the week prior to the survey, monthly earnings, skills matching, and self-employment. Skills matching is defined by a dummy variable that takes the value of 1 if an individual self-reported using trained vocational skills at current work, 0 otherwise.10 Table S1.6 in the supplemental online appendix provides a full description of all outcome variables. The upper panel shows short-term (six-month) treatment effects, while the lower panel shows medium-term (eighteen-month) mean effects. Odd- and even-numbered columns show results for specifications without and with control variables. Robust standard errors are reported in parentheses. Table 4. Average Impacts on Labor-Market Outcomes, Mongolian VT Program . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Short-term impacts: 6 months ITT 0.037 0.054 42,030 53,255* 0.048* 0.060** 0.040** 0.036* (0.034) (0.035) (27,745) (27,430) (0.029) (0.029) (0.018) (0.019) TOT 0.076 0.126 85,931 123,467** 0.099** 0.139** 0.083** 0.085* (0.067) (0.077) (61,244) (57,104) (0.057) (0.065) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.455 0.455 234,326 234,326 0.212 0.212 0.068 0.068 N 1043 1043 1043 1043 1043 1043 1043 1043 Medium-term impacts: 18 months ITT −0.022 0.010 33,563 62,988* 0.028 0.038 0.040** 0.040** (0.035) (0.036) (29,283) (33,662) (0.031) (0.032) (0.018) (0.019) TOT −0.045 0.023 68,289 145,384* 0.058 0.088 0.081** 0.092** (0.068) (0.079) (57,104) (74,676) (0.062) (0.072) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.550 0.550 303,516 303,516 0.250 0.250 0.064 0.064 N 974 974 974 974 974 974 974 974 . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Short-term impacts: 6 months ITT 0.037 0.054 42,030 53,255* 0.048* 0.060** 0.040** 0.036* (0.034) (0.035) (27,745) (27,430) (0.029) (0.029) (0.018) (0.019) TOT 0.076 0.126 85,931 123,467** 0.099** 0.139** 0.083** 0.085* (0.067) (0.077) (61,244) (57,104) (0.057) (0.065) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.455 0.455 234,326 234,326 0.212 0.212 0.068 0.068 N 1043 1043 1043 1043 1043 1043 1043 1043 Medium-term impacts: 18 months ITT −0.022 0.010 33,563 62,988* 0.028 0.038 0.040** 0.040** (0.035) (0.036) (29,283) (33,662) (0.031) (0.032) (0.018) (0.019) TOT −0.045 0.023 68,289 145,384* 0.058 0.088 0.081** 0.092** (0.068) (0.079) (57,104) (74,676) (0.062) (0.072) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.550 0.550 303,516 303,516 0.250 0.250 0.064 0.064 N 974 974 974 974 974 974 974 974 Source: Author's calculations based on First and Second Follow up data. Note: Robust standard errors in parentheses. Fixed effects by day of random assignment are included in all models. Even-numbered columns include control variables selected by “post-double-selection” LASSO method of Belloni, Chernozhukov, and Hansen (2014). A total of 35 baseline variables are included in the selection algorithm from which the following variables were selected as control variables: gender, whether ever worked, whether required to sign quadrilateral contract, and residence in Nalaik district for employment and earnings outcomes; gender, whether ever worked, and residence in Nalaik district for skills matching outcome; and gender, whether required to sign quadrilateral contract, and residence in Nalaik district for self-employment outcome. Intent-to-treat (ITT) parameters estimated by OLS models. Treatment on the Treated parameters (TOT) estimated by 2SLS that instruments the treatment (T) by the randomly assigned treated status (Z) of participants. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 4. Average Impacts on Labor-Market Outcomes, Mongolian VT Program . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Short-term impacts: 6 months ITT 0.037 0.054 42,030 53,255* 0.048* 0.060** 0.040** 0.036* (0.034) (0.035) (27,745) (27,430) (0.029) (0.029) (0.018) (0.019) TOT 0.076 0.126 85,931 123,467** 0.099** 0.139** 0.083** 0.085* (0.067) (0.077) (61,244) (57,104) (0.057) (0.065) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.455 0.455 234,326 234,326 0.212 0.212 0.068 0.068 N 1043 1043 1043 1043 1043 1043 1043 1043 Medium-term impacts: 18 months ITT −0.022 0.010 33,563 62,988* 0.028 0.038 0.040** 0.040** (0.035) (0.036) (29,283) (33,662) (0.031) (0.032) (0.018) (0.019) TOT −0.045 0.023 68,289 145,384* 0.058 0.088 0.081** 0.092** (0.068) (0.079) (57,104) (74,676) (0.062) (0.072) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.550 0.550 303,516 303,516 0.250 0.250 0.064 0.064 N 974 974 974 974 974 974 974 974 . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Short-term impacts: 6 months ITT 0.037 0.054 42,030 53,255* 0.048* 0.060** 0.040** 0.036* (0.034) (0.035) (27,745) (27,430) (0.029) (0.029) (0.018) (0.019) TOT 0.076 0.126 85,931 123,467** 0.099** 0.139** 0.083** 0.085* (0.067) (0.077) (61,244) (57,104) (0.057) (0.065) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.455 0.455 234,326 234,326 0.212 0.212 0.068 0.068 N 1043 1043 1043 1043 1043 1043 1043 1043 Medium-term impacts: 18 months ITT −0.022 0.010 33,563 62,988* 0.028 0.038 0.040** 0.040** (0.035) (0.036) (29,283) (33,662) (0.031) (0.032) (0.018) (0.019) TOT −0.045 0.023 68,289 145,384* 0.058 0.088 0.081** 0.092** (0.068) (0.079) (57,104) (74,676) (0.062) (0.072) (0.036) (0.044) Controls included No Yes No Yes No Yes No Yes Mean control group 0.550 0.550 303,516 303,516 0.250 0.250 0.064 0.064 N 974 974 974 974 974 974 974 974 Source: Author's calculations based on First and Second Follow up data. Note: Robust standard errors in parentheses. Fixed effects by day of random assignment are included in all models. Even-numbered columns include control variables selected by “post-double-selection” LASSO method of Belloni, Chernozhukov, and Hansen (2014). A total of 35 baseline variables are included in the selection algorithm from which the following variables were selected as control variables: gender, whether ever worked, whether required to sign quadrilateral contract, and residence in Nalaik district for employment and earnings outcomes; gender, whether ever worked, and residence in Nalaik district for skills matching outcome; and gender, whether required to sign quadrilateral contract, and residence in Nalaik district for self-employment outcome. Intent-to-treat (ITT) parameters estimated by OLS models. Treatment on the Treated parameters (TOT) estimated by 2SLS that instruments the treatment (T) by the randomly assigned treated status (Z) of participants. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Short-term point estimates (six months) show positive and statistically significant effects for self-employment and skills matching regardless of the econometric specification and parameter of interest. Relative to the mean of the control group, the magnitude of the treatment effects is somewhat large, particularly for the TOT point estimates (from 0.099 to 0.139 for skills matching and from 0.083 to 0.085 for self-employment). These results suggest that the vocational skills of MVTP participants are better aligned to occupations as compared to the skills of individuals in the control group. Similarly, and although the incidence of self-employment remained very low in comparison to what is commonly observed in other developing countries, these point estimates indicate that self-employment more than doubled among MVTP participants relative to nonparticipants. For the employment outcome, on the other hand, positive but imprecisely measured treatment effects are observed, ranging from 0.037 to 0.054 (ITT) and from 0.076 to 0.127 (TOT).11 Finally, earnings effects are positive across all specifications, but their statistical relevance depends upon both the use of control variables and alternative specifications for the outcome itself. In this regard, significant and sizable impacts arise whenever LASSO control variables are included in the specification models, while imprecisely measured earnings effects emerge when the econometric specifications do not incorporate control variables (see table 4). The results of sensitivity tests that use the inverse hyperbolic sine transformation, trimming, and winsoring for earnings outcome are shown in table S1.7 in the supplemental online appendix.12 While statistically significant results hold under trimming and winsoring approaches, imprecise point estimates emerge when the distribution of the outcome is smoothed out. All in all, the magnitude of the mean effects for monthly earnings is somewhat large (as high as 23 percent for the ITT) from a mean of the control group of approximately USD 110.13 Mean effects eighteen months after completion of training are presented table 4. Relative to short-term effects, a decay in the magnitude and statistical significance of treatment effects emerges for all outcome variables except self-employment. The effects on skills matching dissipate and become statistically not significant, while effects for the employment outcome become either negligible or negative. For monthly earnings, the point estimates from both ITT and TOT parameters are still in line with short-term findings because smaller but positive statistically significant effects emerge depending on the specification used. Self-employment, on the other hand, is the only outcome that shows steady positive and statistically relevant impacts over time, with sizable point estimates ranging from 0.04 to 0.092. Sizable effects can be observed when comparing MVTP confidence intervals for earnings ITT treatment effects, which are measured in terms of percent gain relative to the control group and reported in footnote 13, against the corresponding findings that emerged from 11 RCT vocational-training programs that targeted youth worldwide (see table 1), somewhat sizable effects can be observed. This is particularly interesting given the short duration and low cost of this training initiative. As a result, and although there is uncertainty associated with the gains in earnings, a straightforward cost-benefit calculation reveals that it requires six to eight months of treatment effects to recoup the costs of the courses. This is plausible given the short-term impacts depicted in table 4.14 These findings, which are also in line with other studies that measured effects over a longer time span, found that treatment effects fell further over time. (e.g., Alzúa, Cruces, and Lopez 2016; Ibarrarán et al. 2019). Overall, results suggest that training helped trainees move to tailored jobs right after treatment ended, but over time the MVTP fell somewhat short of achieving its promises. This is plausibly explained by the short duration of employment that is likely secured at business firms that offered internships. However, extending these job contracts could have not been a priority for many business firms over time. Unlike most developing countries, employment in Mongolia is mostly formal, and thus extending job contracts to disadvantaged youth that participated in short vocational training courses might not be a priority for many business firms. This, in turn, might explain the consistency of positive impacts for self-employment.15 Heterogeneous Effects A solid pattern in the existing training literature is the substantial heterogeneity in the impacts of training. To account for the heterogeneity of effects across subgroups of participants, the estimation framework and model specification given in equations (1) and (2) are applied after interacting the treatment status variable with baseline covariates of interest: gender (men vs. women), age (15–21 vs. 22–30), poverty status (poorest vs. less poor), and educational attainment (less than high school vs. high school vs. technical and university education).16 These policy variables, commonly used in the assessment of vocational-training programs worldwide, are related to the efficiency of the targeting approach. As before, the same four outcomes of interest are monitored, at 6 (table 5a) and 18 months (table 5b) after the intervention. Table 5a. Heterogenous Impacts for VT Program, 6 months Mongolian VT Program, 2014–2016 . Employment 6-month . Earnings 6-month . Skills match 6-month . Self-employment 6-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT Program (0.056) (0.132) (46,984) (110,833*) (0.048) (0.114) (0.016) (0.041) (0.043) (0.096) (29,500) (65,876) (0.036) (0.079) (0.021) (0.047) Males 0.157*** 0.156** 213,390*** 209,985*** 0.064 0.059 0.066** 0.058 (0.059) (0.063) (49,039) (52,741) (0.050) (0.053) (0.033) (0.037) VT * Males −0.004 −0.014 18,046 34,185 0.027 0.054 0.057 0.118 (0.073) (0.148) (68,758) (140,051) (0.062) (0.127) (0.045) (0.093) VT program 0.104** 0.218** 82,763*** 176,738*** 0.094** 0.201*** 0.046* 0.102** (0.042) (0.087) (31,592) (66,497) (0.037) (0.076) (0.024) (0.051) Age 15–21 0.007 0.022 −26,901 −20,498 −0.002 0.004 −0.017 −0.017 (0.062) (0.067) (43,369) (112,393) (0.049) (0.054) (0.029) (0.032) VT * age 15 –21 −0.143** −0.293** −83,909 −20,498 −0.099* −0.196* −0.028 −0.050 (0.072) (0.148) (54,541) (47,534) (0.058) (0.119) (0.038) (0.078) VT program 0.096** 0.211** 61,903* 138,853** 0.082** 0.182** 0.054** 0.121** (0.040) (0.086) (32,936) (70,851) (0.034) (0.075) (0.024) (0.052) Poor 0.073 0.112 −30,004 −15,889 0.004 0.029 0.004 0.023 (0.065) (0.074) (37,829) (45,280) (0.051) (0.059) (0.027) (0.033) VT * poor −0.158** −0.397** −32,749 −71,324 −0.083 −0.201 −0.068* −0.167* (0.079) (0.198) (54,680) (134,722) (0.063) (0.158) (0.035) (0.086) VT Program 0.202*** 0.386*** 116,158** 227,330** 0.130** 0.255** 0.042 0.088 (0.063) (0.118) (54,105) (101,169) (0.057) (0.105) (0.034) (0.063) Less than high school 0.056 0.095 93 22,748 0.015 0.044 0.005 0.004 (0.082) (0.098) (61,177) (75,358) (0.070) (0.082) (0.044) (0.053) High school 0.016 0.032 −17,922 −15,594 −0.049 −0.049 −0.021 −0.024 (0.063) (0.069) (44,683) (48,544) (0.075) (0.059) (0.033) (0.035) VT * less high school −0.274*** −0.597** −157,077* −351,078 −0.189** −0.439** −0.030 −0.047 (0.098) (0.263) (87,597) (228,775) (0.083) (0.213) (0.053) (0.141) VT * high school −0.174** −0.313** −57,243 −94,337 −0.055 −0.087 0.001 0.009 (0.078) (0.143) (65,074) (119,056) (0.068) (0.124) (0.042) (0.077) VT Program −0.065 −0.160 −17,368 −37,280 0.049 0.131 0.017 0.049 (0.063) (0.154) (47,692) (116,469) (0.049) (0.122) (0.033) (0.082) Optimism −0.064 −0.092 4,070 −12,401 0.065 0.004 −0.010 −0.014 (0.062) (0.069) (50,163) (55,242) (0.051) (0.139) (0.032) (0.036) VT * optimism 0.177** 0.386** 103,921* 216,632* 0.013 0.064 0.027 0.047 (0.075) (0.066) (56,070) (125,034) (0.063) (0.077) (0.040) (0.089) N 1043 1043 1043 1043 1043 1043 1043 1043 . Employment 6-month . Earnings 6-month . Skills match 6-month . Self-employment 6-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT Program (0.056) (0.132) (46,984) (110,833*) (0.048) (0.114) (0.016) (0.041) (0.043) (0.096) (29,500) (65,876) (0.036) (0.079) (0.021) (0.047) Males 0.157*** 0.156** 213,390*** 209,985*** 0.064 0.059 0.066** 0.058 (0.059) (0.063) (49,039) (52,741) (0.050) (0.053) (0.033) (0.037) VT * Males −0.004 −0.014 18,046 34,185 0.027 0.054 0.057 0.118 (0.073) (0.148) (68,758) (140,051) (0.062) (0.127) (0.045) (0.093) VT program 0.104** 0.218** 82,763*** 176,738*** 0.094** 0.201*** 0.046* 0.102** (0.042) (0.087) (31,592) (66,497) (0.037) (0.076) (0.024) (0.051) Age 15–21 0.007 0.022 −26,901 −20,498 −0.002 0.004 −0.017 −0.017 (0.062) (0.067) (43,369) (112,393) (0.049) (0.054) (0.029) (0.032) VT * age 15 –21 −0.143** −0.293** −83,909 −20,498 −0.099* −0.196* −0.028 −0.050 (0.072) (0.148) (54,541) (47,534) (0.058) (0.119) (0.038) (0.078) VT program 0.096** 0.211** 61,903* 138,853** 0.082** 0.182** 0.054** 0.121** (0.040) (0.086) (32,936) (70,851) (0.034) (0.075) (0.024) (0.052) Poor 0.073 0.112 −30,004 −15,889 0.004 0.029 0.004 0.023 (0.065) (0.074) (37,829) (45,280) (0.051) (0.059) (0.027) (0.033) VT * poor −0.158** −0.397** −32,749 −71,324 −0.083 −0.201 −0.068* −0.167* (0.079) (0.198) (54,680) (134,722) (0.063) (0.158) (0.035) (0.086) VT Program 0.202*** 0.386*** 116,158** 227,330** 0.130** 0.255** 0.042 0.088 (0.063) (0.118) (54,105) (101,169) (0.057) (0.105) (0.034) (0.063) Less than high school 0.056 0.095 93 22,748 0.015 0.044 0.005 0.004 (0.082) (0.098) (61,177) (75,358) (0.070) (0.082) (0.044) (0.053) High school 0.016 0.032 −17,922 −15,594 −0.049 −0.049 −0.021 −0.024 (0.063) (0.069) (44,683) (48,544) (0.075) (0.059) (0.033) (0.035) VT * less high school −0.274*** −0.597** −157,077* −351,078 −0.189** −0.439** −0.030 −0.047 (0.098) (0.263) (87,597) (228,775) (0.083) (0.213) (0.053) (0.141) VT * high school −0.174** −0.313** −57,243 −94,337 −0.055 −0.087 0.001 0.009 (0.078) (0.143) (65,074) (119,056) (0.068) (0.124) (0.042) (0.077) VT Program −0.065 −0.160 −17,368 −37,280 0.049 0.131 0.017 0.049 (0.063) (0.154) (47,692) (116,469) (0.049) (0.122) (0.033) (0.082) Optimism −0.064 −0.092 4,070 −12,401 0.065 0.004 −0.010 −0.014 (0.062) (0.069) (50,163) (55,242) (0.051) (0.139) (0.032) (0.036) VT * optimism 0.177** 0.386** 103,921* 216,632* 0.013 0.064 0.027 0.047 (0.075) (0.066) (56,070) (125,034) (0.063) (0.077) (0.040) (0.089) N 1043 1043 1043 1043 1043 1043 1043 1043 Source: Author's calculations based on First Follow up data. Note: Robust standard errors in parentheses. ITT and TOT parameters estimated by multivariate OLS models with control variables selected by the post-double LASSO selection approach. See notes in table 4 for full specification details. Poor is defined as 1 for those in the bottom quantile of the household wealth assets index. Index is estimated by PCA and includes indicators for whether unit lives in a slum (Ger), has car, motorcycle, computer at home, washing machine, vacuum cleaner, TV, and refrigerator. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 5a. Heterogenous Impacts for VT Program, 6 months Mongolian VT Program, 2014–2016 . Employment 6-month . Earnings 6-month . Skills match 6-month . Self-employment 6-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT Program (0.056) (0.132) (46,984) (110,833*) (0.048) (0.114) (0.016) (0.041) (0.043) (0.096) (29,500) (65,876) (0.036) (0.079) (0.021) (0.047) Males 0.157*** 0.156** 213,390*** 209,985*** 0.064 0.059 0.066** 0.058 (0.059) (0.063) (49,039) (52,741) (0.050) (0.053) (0.033) (0.037) VT * Males −0.004 −0.014 18,046 34,185 0.027 0.054 0.057 0.118 (0.073) (0.148) (68,758) (140,051) (0.062) (0.127) (0.045) (0.093) VT program 0.104** 0.218** 82,763*** 176,738*** 0.094** 0.201*** 0.046* 0.102** (0.042) (0.087) (31,592) (66,497) (0.037) (0.076) (0.024) (0.051) Age 15–21 0.007 0.022 −26,901 −20,498 −0.002 0.004 −0.017 −0.017 (0.062) (0.067) (43,369) (112,393) (0.049) (0.054) (0.029) (0.032) VT * age 15 –21 −0.143** −0.293** −83,909 −20,498 −0.099* −0.196* −0.028 −0.050 (0.072) (0.148) (54,541) (47,534) (0.058) (0.119) (0.038) (0.078) VT program 0.096** 0.211** 61,903* 138,853** 0.082** 0.182** 0.054** 0.121** (0.040) (0.086) (32,936) (70,851) (0.034) (0.075) (0.024) (0.052) Poor 0.073 0.112 −30,004 −15,889 0.004 0.029 0.004 0.023 (0.065) (0.074) (37,829) (45,280) (0.051) (0.059) (0.027) (0.033) VT * poor −0.158** −0.397** −32,749 −71,324 −0.083 −0.201 −0.068* −0.167* (0.079) (0.198) (54,680) (134,722) (0.063) (0.158) (0.035) (0.086) VT Program 0.202*** 0.386*** 116,158** 227,330** 0.130** 0.255** 0.042 0.088 (0.063) (0.118) (54,105) (101,169) (0.057) (0.105) (0.034) (0.063) Less than high school 0.056 0.095 93 22,748 0.015 0.044 0.005 0.004 (0.082) (0.098) (61,177) (75,358) (0.070) (0.082) (0.044) (0.053) High school 0.016 0.032 −17,922 −15,594 −0.049 −0.049 −0.021 −0.024 (0.063) (0.069) (44,683) (48,544) (0.075) (0.059) (0.033) (0.035) VT * less high school −0.274*** −0.597** −157,077* −351,078 −0.189** −0.439** −0.030 −0.047 (0.098) (0.263) (87,597) (228,775) (0.083) (0.213) (0.053) (0.141) VT * high school −0.174** −0.313** −57,243 −94,337 −0.055 −0.087 0.001 0.009 (0.078) (0.143) (65,074) (119,056) (0.068) (0.124) (0.042) (0.077) VT Program −0.065 −0.160 −17,368 −37,280 0.049 0.131 0.017 0.049 (0.063) (0.154) (47,692) (116,469) (0.049) (0.122) (0.033) (0.082) Optimism −0.064 −0.092 4,070 −12,401 0.065 0.004 −0.010 −0.014 (0.062) (0.069) (50,163) (55,242) (0.051) (0.139) (0.032) (0.036) VT * optimism 0.177** 0.386** 103,921* 216,632* 0.013 0.064 0.027 0.047 (0.075) (0.066) (56,070) (125,034) (0.063) (0.077) (0.040) (0.089) N 1043 1043 1043 1043 1043 1043 1043 1043 . Employment 6-month . Earnings 6-month . Skills match 6-month . Self-employment 6-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT Program (0.056) (0.132) (46,984) (110,833*) (0.048) (0.114) (0.016) (0.041) (0.043) (0.096) (29,500) (65,876) (0.036) (0.079) (0.021) (0.047) Males 0.157*** 0.156** 213,390*** 209,985*** 0.064 0.059 0.066** 0.058 (0.059) (0.063) (49,039) (52,741) (0.050) (0.053) (0.033) (0.037) VT * Males −0.004 −0.014 18,046 34,185 0.027 0.054 0.057 0.118 (0.073) (0.148) (68,758) (140,051) (0.062) (0.127) (0.045) (0.093) VT program 0.104** 0.218** 82,763*** 176,738*** 0.094** 0.201*** 0.046* 0.102** (0.042) (0.087) (31,592) (66,497) (0.037) (0.076) (0.024) (0.051) Age 15–21 0.007 0.022 −26,901 −20,498 −0.002 0.004 −0.017 −0.017 (0.062) (0.067) (43,369) (112,393) (0.049) (0.054) (0.029) (0.032) VT * age 15 –21 −0.143** −0.293** −83,909 −20,498 −0.099* −0.196* −0.028 −0.050 (0.072) (0.148) (54,541) (47,534) (0.058) (0.119) (0.038) (0.078) VT program 0.096** 0.211** 61,903* 138,853** 0.082** 0.182** 0.054** 0.121** (0.040) (0.086) (32,936) (70,851) (0.034) (0.075) (0.024) (0.052) Poor 0.073 0.112 −30,004 −15,889 0.004 0.029 0.004 0.023 (0.065) (0.074) (37,829) (45,280) (0.051) (0.059) (0.027) (0.033) VT * poor −0.158** −0.397** −32,749 −71,324 −0.083 −0.201 −0.068* −0.167* (0.079) (0.198) (54,680) (134,722) (0.063) (0.158) (0.035) (0.086) VT Program 0.202*** 0.386*** 116,158** 227,330** 0.130** 0.255** 0.042 0.088 (0.063) (0.118) (54,105) (101,169) (0.057) (0.105) (0.034) (0.063) Less than high school 0.056 0.095 93 22,748 0.015 0.044 0.005 0.004 (0.082) (0.098) (61,177) (75,358) (0.070) (0.082) (0.044) (0.053) High school 0.016 0.032 −17,922 −15,594 −0.049 −0.049 −0.021 −0.024 (0.063) (0.069) (44,683) (48,544) (0.075) (0.059) (0.033) (0.035) VT * less high school −0.274*** −0.597** −157,077* −351,078 −0.189** −0.439** −0.030 −0.047 (0.098) (0.263) (87,597) (228,775) (0.083) (0.213) (0.053) (0.141) VT * high school −0.174** −0.313** −57,243 −94,337 −0.055 −0.087 0.001 0.009 (0.078) (0.143) (65,074) (119,056) (0.068) (0.124) (0.042) (0.077) VT Program −0.065 −0.160 −17,368 −37,280 0.049 0.131 0.017 0.049 (0.063) (0.154) (47,692) (116,469) (0.049) (0.122) (0.033) (0.082) Optimism −0.064 −0.092 4,070 −12,401 0.065 0.004 −0.010 −0.014 (0.062) (0.069) (50,163) (55,242) (0.051) (0.139) (0.032) (0.036) VT * optimism 0.177** 0.386** 103,921* 216,632* 0.013 0.064 0.027 0.047 (0.075) (0.066) (56,070) (125,034) (0.063) (0.077) (0.040) (0.089) N 1043 1043 1043 1043 1043 1043 1043 1043 Source: Author's calculations based on First Follow up data. Note: Robust standard errors in parentheses. ITT and TOT parameters estimated by multivariate OLS models with control variables selected by the post-double LASSO selection approach. See notes in table 4 for full specification details. Poor is defined as 1 for those in the bottom quantile of the household wealth assets index. Index is estimated by PCA and includes indicators for whether unit lives in a slum (Ger), has car, motorcycle, computer at home, washing machine, vacuum cleaner, TV, and refrigerator. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 5b. Heterogenous Impacts for VT Program, 18 Months Mongolian VT Program, 2014–2016 . Employment 18-month . Wages 18-month . Skills match 18-month . Self-employment 18-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT program 0.006 0.015 37,131 88,097 0.013 0.034 0.030 0.071 (0.044) (0.096) (31,575) (69,239) (0.039) (0.085) (0.021) (0.047) Males 0.189*** 0.188*** 208,986*** 197,430*** 0.006 −0.003 0.100** 0.095** (0.059) (0.062) (45,506) (46,640) (0.056) (0.061) (0.036) (0.038) VT * males 0.006 0.014 72,961 153,954 0.078 0.164 0.028 0.060 (0.073) (0.146) (63,228) (127,315) (0.069) (0.141) (0.049) (0.097) VT program 0.036 0.071 80,319 173,267 0.055 0.119 0.049** 0.107** (0.042) (0.086) (41,881) (86,083) (0.042) (0.086) (0.025) (0.051) Age 15–21 −0.046 −0.035 −59,935 −58,676 −0.081 −0.080 −0.028 −0.029 (0.063) (0.070) (41,402) (45,768) (0.051) (0.056) (0.033) (0.036) VT * age 15–21 −0.078 −0.166 −53,152 −97,530 −0.040 −0.075 −0.028 −0.049 (0.074) (0.153) (53,456) (108,470) (0.060) (0.123) (0.039) (0.079) VT program 0.028 0.060 88,402** 193,679** 0.067* 0.147* 0.064*** 0.139*** (0.040) (0.086) (35,986) (77,223) (0.037) (0.080) (0.023) (0.049) Poor 0.023 0.039 21,989 50,425 0.046 0.071 0.036 0.059 (0.065) (0.075) (46,888) (52,371) (0.059) (0.068) (0.035) (0.041) VT * poor −0.076 −0.193 −102,305* −245,803* −0.102 −0.251 −0.091** −0.223** (0.081) (0.198) 954,387) (13,425) (0.070) (0.174) (0.045) (0.112) VT program 0.159** 0.288** 134,650*** 259,468*** 0.063 0.126 0.076** 0.148** (0.066) (0.118) (46,134) (86,262) (0.062) (0.112) (0.032) (0.057) Less than high school 0.040 0.111 34,269 85,383 −0.072 −0.068 0.055 0.078 (0.079) (0.102) (53,773) (67,606) (0.073) (0.089) (0.048) (0.064) High school 0.019 0.039 −9549 −6645 −0.082 −0.083 0.005 0.005 (0.067) (0.071) (46,151) (49,473) (0.061) (0.065) (0.031) (0.033) VT * less high school −0.309*** −0.770*** −233,955*** −586,534*** −0.063 −0.119 −0.118 ** −0.284* (0.097) (0.277) (66,861) (185,504) (0.088) (0.227) (0.057) (0.168) VT * high school −0.166** −0.295** −46,103 −67,002 −0.015 −0.018 −0.023 −0.032 (0.081) (0.143) (61,595) (111,954) (0.074) (0.130) (0.042) (0.074) VT Program −0.046 −0.114 35,569 101,563 −0.019 −0.041 0.042 0.114 (0.063) (0.154) (62,510) (154,355) (0.056) (0.138) (0.034) (0.086) Optimism −0.073 −0.086 −17,218 −20,298 −0.046 −0.058 −0.020 −0.017 (0.061) (0.068) (39,735) (44,478) (0.056) (0.063) (0.031) (0.034) VT * optimism 0.081 0.181 39,074 55,746 0.088 0.180 −0.003 −0.029 (0.076) (0.167) (64,236) (146,627) (0.068) (0.150) (0.041) (0.092) N 974 974 974 974 974 974 974 974 . Employment 18-month . Wages 18-month . Skills match 18-month . Self-employment 18-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT program 0.006 0.015 37,131 88,097 0.013 0.034 0.030 0.071 (0.044) (0.096) (31,575) (69,239) (0.039) (0.085) (0.021) (0.047) Males 0.189*** 0.188*** 208,986*** 197,430*** 0.006 −0.003 0.100** 0.095** (0.059) (0.062) (45,506) (46,640) (0.056) (0.061) (0.036) (0.038) VT * males 0.006 0.014 72,961 153,954 0.078 0.164 0.028 0.060 (0.073) (0.146) (63,228) (127,315) (0.069) (0.141) (0.049) (0.097) VT program 0.036 0.071 80,319 173,267 0.055 0.119 0.049** 0.107** (0.042) (0.086) (41,881) (86,083) (0.042) (0.086) (0.025) (0.051) Age 15–21 −0.046 −0.035 −59,935 −58,676 −0.081 −0.080 −0.028 −0.029 (0.063) (0.070) (41,402) (45,768) (0.051) (0.056) (0.033) (0.036) VT * age 15–21 −0.078 −0.166 −53,152 −97,530 −0.040 −0.075 −0.028 −0.049 (0.074) (0.153) (53,456) (108,470) (0.060) (0.123) (0.039) (0.079) VT program 0.028 0.060 88,402** 193,679** 0.067* 0.147* 0.064*** 0.139*** (0.040) (0.086) (35,986) (77,223) (0.037) (0.080) (0.023) (0.049) Poor 0.023 0.039 21,989 50,425 0.046 0.071 0.036 0.059 (0.065) (0.075) (46,888) (52,371) (0.059) (0.068) (0.035) (0.041) VT * poor −0.076 −0.193 −102,305* −245,803* −0.102 −0.251 −0.091** −0.223** (0.081) (0.198) 954,387) (13,425) (0.070) (0.174) (0.045) (0.112) VT program 0.159** 0.288** 134,650*** 259,468*** 0.063 0.126 0.076** 0.148** (0.066) (0.118) (46,134) (86,262) (0.062) (0.112) (0.032) (0.057) Less than high school 0.040 0.111 34,269 85,383 −0.072 −0.068 0.055 0.078 (0.079) (0.102) (53,773) (67,606) (0.073) (0.089) (0.048) (0.064) High school 0.019 0.039 −9549 −6645 −0.082 −0.083 0.005 0.005 (0.067) (0.071) (46,151) (49,473) (0.061) (0.065) (0.031) (0.033) VT * less high school −0.309*** −0.770*** −233,955*** −586,534*** −0.063 −0.119 −0.118 ** −0.284* (0.097) (0.277) (66,861) (185,504) (0.088) (0.227) (0.057) (0.168) VT * high school −0.166** −0.295** −46,103 −67,002 −0.015 −0.018 −0.023 −0.032 (0.081) (0.143) (61,595) (111,954) (0.074) (0.130) (0.042) (0.074) VT Program −0.046 −0.114 35,569 101,563 −0.019 −0.041 0.042 0.114 (0.063) (0.154) (62,510) (154,355) (0.056) (0.138) (0.034) (0.086) Optimism −0.073 −0.086 −17,218 −20,298 −0.046 −0.058 −0.020 −0.017 (0.061) (0.068) (39,735) (44,478) (0.056) (0.063) (0.031) (0.034) VT * optimism 0.081 0.181 39,074 55,746 0.088 0.180 −0.003 −0.029 (0.076) (0.167) (64,236) (146,627) (0.068) (0.150) (0.041) (0.092) N 974 974 974 974 974 974 974 974 Source: Author's calculations based on Second Follow up data. Notes: Robust standard errors in parentheses. ITT and TOT parameters estimated by multivariate OLS models with control variables selected by the post-double lasso selection approach. See notes in Table 4 for full specification details. Poor is defined as 1 for those in the bottom quantile of the household wealth assets index. Index is estimated by PCA and includes indicators for whether unit lives in a slum (Ger), has car, motorcycle, computer at home, washing machine, vacuum cleaner, TV and refrigerator. *** p < 0.01, ** p <0.05, * p < 0.1. Page 47 of 51 Open in new tab Table 5b. Heterogenous Impacts for VT Program, 18 Months Mongolian VT Program, 2014–2016 . Employment 18-month . Wages 18-month . Skills match 18-month . Self-employment 18-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT program 0.006 0.015 37,131 88,097 0.013 0.034 0.030 0.071 (0.044) (0.096) (31,575) (69,239) (0.039) (0.085) (0.021) (0.047) Males 0.189*** 0.188*** 208,986*** 197,430*** 0.006 −0.003 0.100** 0.095** (0.059) (0.062) (45,506) (46,640) (0.056) (0.061) (0.036) (0.038) VT * males 0.006 0.014 72,961 153,954 0.078 0.164 0.028 0.060 (0.073) (0.146) (63,228) (127,315) (0.069) (0.141) (0.049) (0.097) VT program 0.036 0.071 80,319 173,267 0.055 0.119 0.049** 0.107** (0.042) (0.086) (41,881) (86,083) (0.042) (0.086) (0.025) (0.051) Age 15–21 −0.046 −0.035 −59,935 −58,676 −0.081 −0.080 −0.028 −0.029 (0.063) (0.070) (41,402) (45,768) (0.051) (0.056) (0.033) (0.036) VT * age 15–21 −0.078 −0.166 −53,152 −97,530 −0.040 −0.075 −0.028 −0.049 (0.074) (0.153) (53,456) (108,470) (0.060) (0.123) (0.039) (0.079) VT program 0.028 0.060 88,402** 193,679** 0.067* 0.147* 0.064*** 0.139*** (0.040) (0.086) (35,986) (77,223) (0.037) (0.080) (0.023) (0.049) Poor 0.023 0.039 21,989 50,425 0.046 0.071 0.036 0.059 (0.065) (0.075) (46,888) (52,371) (0.059) (0.068) (0.035) (0.041) VT * poor −0.076 −0.193 −102,305* −245,803* −0.102 −0.251 −0.091** −0.223** (0.081) (0.198) 954,387) (13,425) (0.070) (0.174) (0.045) (0.112) VT program 0.159** 0.288** 134,650*** 259,468*** 0.063 0.126 0.076** 0.148** (0.066) (0.118) (46,134) (86,262) (0.062) (0.112) (0.032) (0.057) Less than high school 0.040 0.111 34,269 85,383 −0.072 −0.068 0.055 0.078 (0.079) (0.102) (53,773) (67,606) (0.073) (0.089) (0.048) (0.064) High school 0.019 0.039 −9549 −6645 −0.082 −0.083 0.005 0.005 (0.067) (0.071) (46,151) (49,473) (0.061) (0.065) (0.031) (0.033) VT * less high school −0.309*** −0.770*** −233,955*** −586,534*** −0.063 −0.119 −0.118 ** −0.284* (0.097) (0.277) (66,861) (185,504) (0.088) (0.227) (0.057) (0.168) VT * high school −0.166** −0.295** −46,103 −67,002 −0.015 −0.018 −0.023 −0.032 (0.081) (0.143) (61,595) (111,954) (0.074) (0.130) (0.042) (0.074) VT Program −0.046 −0.114 35,569 101,563 −0.019 −0.041 0.042 0.114 (0.063) (0.154) (62,510) (154,355) (0.056) (0.138) (0.034) (0.086) Optimism −0.073 −0.086 −17,218 −20,298 −0.046 −0.058 −0.020 −0.017 (0.061) (0.068) (39,735) (44,478) (0.056) (0.063) (0.031) (0.034) VT * optimism 0.081 0.181 39,074 55,746 0.088 0.180 −0.003 −0.029 (0.076) (0.167) (64,236) (146,627) (0.068) (0.150) (0.041) (0.092) N 974 974 974 974 974 974 974 974 . Employment 18-month . Wages 18-month . Skills match 18-month . Self-employment 18-month . . ITT . TOT . ITT . TOT . ITT . TOT . ITT . TOT . VT program 0.006 0.015 37,131 88,097 0.013 0.034 0.030 0.071 (0.044) (0.096) (31,575) (69,239) (0.039) (0.085) (0.021) (0.047) Males 0.189*** 0.188*** 208,986*** 197,430*** 0.006 −0.003 0.100** 0.095** (0.059) (0.062) (45,506) (46,640) (0.056) (0.061) (0.036) (0.038) VT * males 0.006 0.014 72,961 153,954 0.078 0.164 0.028 0.060 (0.073) (0.146) (63,228) (127,315) (0.069) (0.141) (0.049) (0.097) VT program 0.036 0.071 80,319 173,267 0.055 0.119 0.049** 0.107** (0.042) (0.086) (41,881) (86,083) (0.042) (0.086) (0.025) (0.051) Age 15–21 −0.046 −0.035 −59,935 −58,676 −0.081 −0.080 −0.028 −0.029 (0.063) (0.070) (41,402) (45,768) (0.051) (0.056) (0.033) (0.036) VT * age 15–21 −0.078 −0.166 −53,152 −97,530 −0.040 −0.075 −0.028 −0.049 (0.074) (0.153) (53,456) (108,470) (0.060) (0.123) (0.039) (0.079) VT program 0.028 0.060 88,402** 193,679** 0.067* 0.147* 0.064*** 0.139*** (0.040) (0.086) (35,986) (77,223) (0.037) (0.080) (0.023) (0.049) Poor 0.023 0.039 21,989 50,425 0.046 0.071 0.036 0.059 (0.065) (0.075) (46,888) (52,371) (0.059) (0.068) (0.035) (0.041) VT * poor −0.076 −0.193 −102,305* −245,803* −0.102 −0.251 −0.091** −0.223** (0.081) (0.198) 954,387) (13,425) (0.070) (0.174) (0.045) (0.112) VT program 0.159** 0.288** 134,650*** 259,468*** 0.063 0.126 0.076** 0.148** (0.066) (0.118) (46,134) (86,262) (0.062) (0.112) (0.032) (0.057) Less than high school 0.040 0.111 34,269 85,383 −0.072 −0.068 0.055 0.078 (0.079) (0.102) (53,773) (67,606) (0.073) (0.089) (0.048) (0.064) High school 0.019 0.039 −9549 −6645 −0.082 −0.083 0.005 0.005 (0.067) (0.071) (46,151) (49,473) (0.061) (0.065) (0.031) (0.033) VT * less high school −0.309*** −0.770*** −233,955*** −586,534*** −0.063 −0.119 −0.118 ** −0.284* (0.097) (0.277) (66,861) (185,504) (0.088) (0.227) (0.057) (0.168) VT * high school −0.166** −0.295** −46,103 −67,002 −0.015 −0.018 −0.023 −0.032 (0.081) (0.143) (61,595) (111,954) (0.074) (0.130) (0.042) (0.074) VT Program −0.046 −0.114 35,569 101,563 −0.019 −0.041 0.042 0.114 (0.063) (0.154) (62,510) (154,355) (0.056) (0.138) (0.034) (0.086) Optimism −0.073 −0.086 −17,218 −20,298 −0.046 −0.058 −0.020 −0.017 (0.061) (0.068) (39,735) (44,478) (0.056) (0.063) (0.031) (0.034) VT * optimism 0.081 0.181 39,074 55,746 0.088 0.180 −0.003 −0.029 (0.076) (0.167) (64,236) (146,627) (0.068) (0.150) (0.041) (0.092) N 974 974 974 974 974 974 974 974 Source: Author's calculations based on Second Follow up data. Notes: Robust standard errors in parentheses. ITT and TOT parameters estimated by multivariate OLS models with control variables selected by the post-double lasso selection approach. See notes in Table 4 for full specification details. Poor is defined as 1 for those in the bottom quantile of the household wealth assets index. Index is estimated by PCA and includes indicators for whether unit lives in a slum (Ger), has car, motorcycle, computer at home, washing machine, vacuum cleaner, TV and refrigerator. *** p < 0.01, ** p <0.05, * p < 0.1. Page 47 of 51 Open in new tab The results highlight the heterogeneity of vocational-training effects across demographic groups. Individuals aged 15–21, normally the demographic group at the highest risk of unemployment, benefit least from the program 6 months after treatment (see table 5a). According to TOT point estimates at 6 months after training, the likelihood of employment, self-employment, and skills matching is, respectively, 29, 19, and 5 percentage points lower for the youngest group relative to the 22–30-year-old cohort. Eighteen months after the intervention, however, these differences are no longer statistically significant. This suggests a rapid deterioration of MVTP effects over time for all age groups rather than an improvement in labor-market outcomes for the youngest vs. the older cohorts. Consistent (negative) differential effects for monthly earnings also emerged for the youngest group 6 and 18 months after the intervention. Interestingly, men and women benefit equally across all outcomes of interest, as shown in the second panel of tables 5a and 5b. This important result is contrary to what has been observed in similar demand-driven approaches in other countries, particularly in Latin America, as shown by Attanasio, Kugler, and Meghir (2011) for Colombia; Card et al. (2011) for the Dominican Republic; Alzúa, Cruces, and Lopez (2016) for Argentina; and Díaz and Rosas (2016) for Peru. All these studies showed that young women benefited more from these types of vocational training initiatives than did young men. Nonetheless, the point estimates for the independent gender variable indicate that men have much higher employment rates and earnings than do women, reflecting reported large gender gaps in Ulaanbaatar's labor markets (Pastore 2010; Shatz et al. 2015). While the MVTP was originally designed to target youth from poor households, findings show that people in the bottom quartile of the household asset index benefit less from the program (see tables 5a and 5b). In comparing those at the bottom of the poverty index to those at the middle and upper end of the distribution, the ITT and TOT parameters show statistically significant differential coefficients for employment (minus 18 and minus 39 percentage points) and self-employment (minus 7 and minus 16 percentage points). These sizeable differences in treatment effects hold 18 months later, although statistical significance is only observed for the self-employment outcome. Monthly earnings for the poorest among the poor, on the other hand, show negative differential effects at 6 and 18 months after the intervention. Statistical significance, however, is uncertain as it depends upon the specific definition of the earnings outcome (levels vs. inverse sin transformation) and time span (6 vs. 18 months). Finally, sizable heterogeneous effects emerge based on participants’ formal level of schooling. Individuals with less than a high-school education and those who had completed high school show negative differential effects in comparison to participants with technical or university education. These differences are monotonic with respect to education level. In particular, the magnitude of these differential effects is striking for those at the bottom of the schooling ladder and for all outcomes of interest. The TOT point estimates at 6 and 18 months after the intervention show, for instance, that the likelihood of employment is 59 and 77 percentage points lower for individuals with the least amount of formal schooling in comparison to those at the upper end of the schooling distribution. These substantial differential effects hold 18 months after the intervention. Importantly, all subgroups for each variable of interest contain at least 20 percent of the sample, and these sizable heterogeneous effects are therefore not driven by small sample sizes of subgroups. Neither is the assessment fundamentally affected when correcting for multiple-hypothesis testing by the Hochberg's false-discovery-rate method.17 Rather, these heterogeneous results reflect the contextual reality of Ulaanbaatar, where substantial gaps in labor-market outcomes across age, education, and poverty ranks have been similarly observed for the youth population. The 2012 Mongolian Labor Survey shows, after controlling for a rich set of observable covariates, that young individuals with university education earned on average 50 to 60 percent more than individuals with secondary education, while individuals with a primary education earned between 18 and 23 percent less than people with secondary education. Likewise, individuals aged 25–29 reported 53 percent more earnings than individuals aged 15–19, while the NEET rates by poverty status ranged from 40 percent (poor) to 11 percent (nonpoor) (Shatz et al. 2015). It would have been extremely difficult for a short-term program like the one examined in this study to change these structural labor-market gaps. Furthermore, because bias toward optimism has been reported to be particularly striking among the unemployed in other settings (e.g., Spinnewijn 2015), treatment-effect heterogeneity by subjective beliefs regarding job prospects is also assessed. Measuring at baseline a subjective belief in finding a job within the successive 6 months, a median of 80 points is recorded in the sample (on a scale of 0 to 100), while around 70 percent of the sample self-reported being “very optimistic” about getting a job when answering a categorical-type similar baseline question. Compared to the actual number of employed individuals (46 percent) six months following the baseline, the sample of youth unemployed seems to show indeed over-optimistic beliefs regarding their job prospects. Point estimates for the vocational treatment variable interacted with the dummy variable “very optimistic beliefs” is depicted in the bottom of tables 5a and 5b. In the short run, positive differential effects emerge for individuals who are optimistic about their labor market prospects relative to individuals with lower expectations. These differential effects are statistically significant for employment and earnings. In the medium term, positive differential effects are still observed for those with optimistic beliefs regarding their labor prospects, but statistical precision is lost. All in all, these results suggest that subjective beliefs regarding job prospects may hinder the effectiveness of vocational training among unemployed youth who hold pessimistic beliefs about their job prospects. However, this evidence should be taken with caution and only as suggestive because subjective beliefs may have been correlated with unobserved factors that are themselves associated with the outcomes of interest. The effect of treatment on those who would have attained (un)favorable outcomes in the absence of training is also assessed. Because random assignment does not identify the joint distribution of potential outcomes with and without training, the endogenous-stratification method proposed in Abadie, Chingos, and West (2018) is used. Separate analyses are conducted for three subgroups of participants, defined in terms of intervals of the predicted outcome without training. Two alternative sets of baseline variables are used to estimate an index that reflects each participant's predicted outcome in the absence of treatment and implement the leave-one-out estimator to minimize bias in the estimation of the ITT parameter as a result of the use of in-sample information on the outcome variable to stratify the sample.18 ITT parameters are displayed in table 6, together with bootstrap standard errors for the adjusted endogenous stratification estimator. Results show heterogeneous intent-to-treat gains across three predicted-outcome groups. Regardless of the outcome of interest and time length for impact analysis, large and positive impacts emerge for the high-predicted outcome group, while negative or close to zero effects are observed for individuals with low-predicted outcomes. Statistical significance, however, varies according to the outcome of interest, set of information used in the prediction of outcomes, and time span. The clearest pattern emerges six months following treatment as one observes positive and statistically significant effects for the high-predicted outcome group. These estimates suggest that the MVTP improved labor-market outcomes largely for individuals who would have obtained favorable outcomes in the absence of treatment. Table 6. Endogenous Stratification Estimation Results Mongolian VT Program, 2014–2016 . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Panel A: Six-months impacts ITT 0.054 53,255* 0.060** 0.036* (0.035) (27430) (0.029) (0.019) Low −0.029 −0.048 −16,375 23,361 −0.061 −0.016 0.020 0.002 (0.084) (0.092) (51,496) (52,760) (0.067) (0.066) (0.032) (0.033) Medium −0.013 0.101 66,869 32,884 0.050 0.049 −0.021 0.014 (0.086) (0.089) (61,676) (68,088) (0.073) (0.071) (0.047) (0.045) High 0.149** 0.060 45,841 71,061 0.134** 0.122 0.117** 0.124* (0.078) (0.081) (76,516) (76,244) (0.076) (0.078) (0.057) (0.066) Panel B: 18-months impacts ITT 0.010 62,988* 0.038 0.040** (0.036) (33,662) (0.032) (0.019) Low −0.053 −0.030 12,212 7195 −0.013 0.029 0.051 −0.025 (0.098) (0.096) (53,573) (54,063) (0.065) (0.065) (0.039) (0.039) Medium 0.002 −0.055 23,233 −83,328 −0.066 0.027 0.042 −0.012 (0.090) (0.090) (58,851) (60,298) (0.083) (0.084) (0.049) (0.047) High −0.030 0.025 67,131 136,042 0.048 0.045 0.064 0.120** (0.082) (0.084) (86,479) (88,084) (0.088) (0.095) (0.061) (0.057) . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Panel A: Six-months impacts ITT 0.054 53,255* 0.060** 0.036* (0.035) (27430) (0.029) (0.019) Low −0.029 −0.048 −16,375 23,361 −0.061 −0.016 0.020 0.002 (0.084) (0.092) (51,496) (52,760) (0.067) (0.066) (0.032) (0.033) Medium −0.013 0.101 66,869 32,884 0.050 0.049 −0.021 0.014 (0.086) (0.089) (61,676) (68,088) (0.073) (0.071) (0.047) (0.045) High 0.149** 0.060 45,841 71,061 0.134** 0.122 0.117** 0.124* (0.078) (0.081) (76,516) (76,244) (0.076) (0.078) (0.057) (0.066) Panel B: 18-months impacts ITT 0.010 62,988* 0.038 0.040** (0.036) (33,662) (0.032) (0.019) Low −0.053 −0.030 12,212 7195 −0.013 0.029 0.051 −0.025 (0.098) (0.096) (53,573) (54,063) (0.065) (0.065) (0.039) (0.039) Medium 0.002 −0.055 23,233 −83,328 −0.066 0.027 0.042 −0.012 (0.090) (0.090) (58,851) (60,298) (0.083) (0.084) (0.049) (0.047) High −0.030 0.025 67,131 136,042 0.048 0.045 0.064 0.120** (0.082) (0.084) (86,479) (88,084) (0.088) (0.095) (0.061) (0.057) Source: Author's calculations based on First and Second Follow up data. Note: Bootstrap standard errors based on 500 bootstrap repetitions are reported in parenthesis. Adjusted endogenous stratification method follows leave-one-out estimator (Abadie, Chingos, and West et al. 2018). Odd-numbered columns use a set of standard socio-demographic baseline variables to compute predicted outcomes: age, gender, poverty index, schooling, married, whether has children, household size, whether living in Gers, disability status, and district of residence. Even-numbered columns use a larger set of baseline covariates by including 35 socio-demographic, labor and subjective expectations on jobs prospects variables. The treatment effects for each one of the three subgroups is estimated by a linear regression of the outcome variable on the treatment indicator, LASSO covariates, and fixed effects by day of random assignment. See notes in table 4 for further details. Open in new tab Table 6. Endogenous Stratification Estimation Results Mongolian VT Program, 2014–2016 . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Panel A: Six-months impacts ITT 0.054 53,255* 0.060** 0.036* (0.035) (27430) (0.029) (0.019) Low −0.029 −0.048 −16,375 23,361 −0.061 −0.016 0.020 0.002 (0.084) (0.092) (51,496) (52,760) (0.067) (0.066) (0.032) (0.033) Medium −0.013 0.101 66,869 32,884 0.050 0.049 −0.021 0.014 (0.086) (0.089) (61,676) (68,088) (0.073) (0.071) (0.047) (0.045) High 0.149** 0.060 45,841 71,061 0.134** 0.122 0.117** 0.124* (0.078) (0.081) (76,516) (76,244) (0.076) (0.078) (0.057) (0.066) Panel B: 18-months impacts ITT 0.010 62,988* 0.038 0.040** (0.036) (33,662) (0.032) (0.019) Low −0.053 −0.030 12,212 7195 −0.013 0.029 0.051 −0.025 (0.098) (0.096) (53,573) (54,063) (0.065) (0.065) (0.039) (0.039) Medium 0.002 −0.055 23,233 −83,328 −0.066 0.027 0.042 −0.012 (0.090) (0.090) (58,851) (60,298) (0.083) (0.084) (0.049) (0.047) High −0.030 0.025 67,131 136,042 0.048 0.045 0.064 0.120** (0.082) (0.084) (86,479) (88,084) (0.088) (0.095) (0.061) (0.057) . Employment . Monthly earnings . Skills match . Self-employment . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Panel A: Six-months impacts ITT 0.054 53,255* 0.060** 0.036* (0.035) (27430) (0.029) (0.019) Low −0.029 −0.048 −16,375 23,361 −0.061 −0.016 0.020 0.002 (0.084) (0.092) (51,496) (52,760) (0.067) (0.066) (0.032) (0.033) Medium −0.013 0.101 66,869 32,884 0.050 0.049 −0.021 0.014 (0.086) (0.089) (61,676) (68,088) (0.073) (0.071) (0.047) (0.045) High 0.149** 0.060 45,841 71,061 0.134** 0.122 0.117** 0.124* (0.078) (0.081) (76,516) (76,244) (0.076) (0.078) (0.057) (0.066) Panel B: 18-months impacts ITT 0.010 62,988* 0.038 0.040** (0.036) (33,662) (0.032) (0.019) Low −0.053 −0.030 12,212 7195 −0.013 0.029 0.051 −0.025 (0.098) (0.096) (53,573) (54,063) (0.065) (0.065) (0.039) (0.039) Medium 0.002 −0.055 23,233 −83,328 −0.066 0.027 0.042 −0.012 (0.090) (0.090) (58,851) (60,298) (0.083) (0.084) (0.049) (0.047) High −0.030 0.025 67,131 136,042 0.048 0.045 0.064 0.120** (0.082) (0.084) (86,479) (88,084) (0.088) (0.095) (0.061) (0.057) Source: Author's calculations based on First and Second Follow up data. Note: Bootstrap standard errors based on 500 bootstrap repetitions are reported in parenthesis. Adjusted endogenous stratification method follows leave-one-out estimator (Abadie, Chingos, and West et al. 2018). Odd-numbered columns use a set of standard socio-demographic baseline variables to compute predicted outcomes: age, gender, poverty index, schooling, married, whether has children, household size, whether living in Gers, disability status, and district of residence. Even-numbered columns use a larger set of baseline covariates by including 35 socio-demographic, labor and subjective expectations on jobs prospects variables. The treatment effects for each one of the three subgroups is estimated by a linear regression of the outcome variable on the treatment indicator, LASSO covariates, and fixed effects by day of random assignment. See notes in table 4 for further details. Open in new tab Effect by Field of Training Knowing whether mean effects vary according to the field of study could be important for policy because it may signal the importance of providing participants with training in certain fields rather than in others. Intent-to-treat point estimates according to field of study are presented in table S1.8 in the supplemental online appendix. Rather than considering only one treatment indicator, multiple treatment variables in equation (1) were incorporated, according to the field of study. Each of these treatment indicators takes the value of 1 if it refers to the specific “X” field of study and is 0 otherwise. The following categories were included: mechanical/machinery, hairdressing and beauty services, craftsmanship, agriculture and gardening, cooking and baking, and a dummy for all other vocations, which together accounted for all training courses. These point estimates should be taken as merely indicative and be assessed with caution, however, because sorting or self-selection of trainees into specific fields of study may have been driven by unobserved factors (e.g., personal traits) that are, in turn, correlated with the outcomes of interest. Results indicate some impact heterogeneity for the MVTP according to the chosen field of study in the short run, although most of the parameters lack statistical significance. “Mechanical/machinery” and “hairdressing and beauty services” show positive and significant effects, with the latter showing the largest effects across all outcomes of interest six months after the intervention. These heterogeneous differences by field of study tend to dissipate across all fields of study at 18 months following treatment, however. Still, “hairdressing and beauty services” is the only field of study that shows positive and significant mean gains for the self-employment outcome 18 months after the intervention, while “agriculture and gardening” and “craftsmanship” show negative differential effects for earnings, skills match, and self-employment outcomes. This is consistent with the sectoral earnings distribution in Mongolia, which shows that average earnings in agriculture and related occupations are on average 50 percent lower than in services. Newsletter Treatment Out of the 410 trainees who attended 141 training courses, 291 were assigned to the newsletter-treatment group (101 courses) and 119 to the control group (40 courses). Not all 291 individuals received the newsletters because some of them never showed up at training centers. However, the focus is placed on the estimation of ITT treatment effects without any formal distinction between ITT and TOT parameters because the rate of no-shows was relatively low for the newsletter treatment. Four intermediate outcomes of interest are assessed: days the respondent attended MVTP, whether the respondent received MVTP qualification (passed the exam), whether the respondent received an MVTP certificate (formal graduation from the program), and respondents’ dropout rates. All outcome variables were self-reported with the exception of the last because administrative information was not recorded in a systematic way. The results show positive (3.7) and statistically significant impacts for “days attended training,” as depicted in the upper panel of table 7. Clustered standard errors by class level are shown in parentheses, and, similar to the estimation of MVTP main effects, a post-double-selection LASSO estimator is implemented for control variables.19 Relative to the average days attended by trainees in the newsletter control group (22), this represents an increase of 17 percent. Consistently, the average effects on MVTP qualification (the respondent passed the exam) are positive (0.045), but they lack statistical precision. This suggests that trainees updated their behavior in response to new information on market returns to training and, thus, that the new information might have had some value to them. This finding is in line with the relatively recent stream of literature in development economics that has successfully used information policies to improve economic outcomes in developing settings (e.g., Chong 2011). For the other two intermediate outcomes – certification and dropout rates – unexpected signs and uninformative treatment effects are observed. The dropout rate is rather low, and the sample size provides too small a level of power for detection of meaningful effects to occur. Table 7. Information Letters ITT Impacts Mongolian VT Program, 2014–2016 . Days attended VT training . Got VT qualification . Got VT certificate . Dropout (adm. variable) . Newsletter impacts 3.723** 0.045 −0.048 0.039 (1.774) (0.073) (0.078) (0.036) Heterogenous impacts Newsletter 2.415 0.030 −0.079 0.054 (2.084) (0.082) (0.098) (0.045) Newsletter * males 1.976 0.066 0.098 −0.063 (5.717) (0.129) (0.145) (0.088) Newsletter 2.845 0.081 −0.081 0.064 (1.625) (0.085) (0.082) (0.040) Newsletter * age 15–21 2.565 −0.109 0.096 −0.076 (2.929) (0.110) (0.113) (0.067) Newsletter 2.991 0.043 −0.054 0.076** (1.948) (0.078) (0.084) (0.035) Newsletter * poor 3.7123 0.011 0.017 −0.204** (5.529) (0.132) (0.128) (0.098) Newsletter 4.742 0.156 0.024 0.095 (3.514) (0.114) (0.142) (0.050) Newsletter * less high school 2.093 −0.088 −0.020 −0.130 (4.743) (0.142) (0.186) (0.096) Newsletter * high school −2.565 −0.177 −0.131 −0.073 (4.564) (0.128) (0.140) (0.096) Newsletter 5.405 0.091 −0.005 0.108 (3.711) (0.092) (0.132) (0.063) Newsletter * very optimistic −2.264 −0.057 −0.056 −0.095 (4.307) (0.011) (0.135) (0.070) Mean control group 22 0.29 0.73 0.05 N 359 359 359 359 . Days attended VT training . Got VT qualification . Got VT certificate . Dropout (adm. variable) . Newsletter impacts 3.723** 0.045 −0.048 0.039 (1.774) (0.073) (0.078) (0.036) Heterogenous impacts Newsletter 2.415 0.030 −0.079 0.054 (2.084) (0.082) (0.098) (0.045) Newsletter * males 1.976 0.066 0.098 −0.063 (5.717) (0.129) (0.145) (0.088) Newsletter 2.845 0.081 −0.081 0.064 (1.625) (0.085) (0.082) (0.040) Newsletter * age 15–21 2.565 −0.109 0.096 −0.076 (2.929) (0.110) (0.113) (0.067) Newsletter 2.991 0.043 −0.054 0.076** (1.948) (0.078) (0.084) (0.035) Newsletter * poor 3.7123 0.011 0.017 −0.204** (5.529) (0.132) (0.128) (0.098) Newsletter 4.742 0.156 0.024 0.095 (3.514) (0.114) (0.142) (0.050) Newsletter * less high school 2.093 −0.088 −0.020 −0.130 (4.743) (0.142) (0.186) (0.096) Newsletter * high school −2.565 −0.177 −0.131 −0.073 (4.564) (0.128) (0.140) (0.096) Newsletter 5.405 0.091 −0.005 0.108 (3.711) (0.092) (0.132) (0.063) Newsletter * very optimistic −2.264 −0.057 −0.056 −0.095 (4.307) (0.011) (0.135) (0.070) Mean control group 22 0.29 0.73 0.05 N 359 359 359 359 Source: Authors' calculations. Note: Clustered standard errors by vocational class. ITT parameters estimated with control covariates selected by post-double-selection LASSO (Belloni, Chernozhukov, and Hansen 2014) among 35 baseline variables. For the administrative dropout outcome, the selected control variables are two: whether respondent was student at baseline and whether respondent was enrolled in technical vocational training education at baseline. For all other outcomes, no covariate was selected by the LASSO approach. Open in new tab Table 7. Information Letters ITT Impacts Mongolian VT Program, 2014–2016 . Days attended VT training . Got VT qualification . Got VT certificate . Dropout (adm. variable) . Newsletter impacts 3.723** 0.045 −0.048 0.039 (1.774) (0.073) (0.078) (0.036) Heterogenous impacts Newsletter 2.415 0.030 −0.079 0.054 (2.084) (0.082) (0.098) (0.045) Newsletter * males 1.976 0.066 0.098 −0.063 (5.717) (0.129) (0.145) (0.088) Newsletter 2.845 0.081 −0.081 0.064 (1.625) (0.085) (0.082) (0.040) Newsletter * age 15–21 2.565 −0.109 0.096 −0.076 (2.929) (0.110) (0.113) (0.067) Newsletter 2.991 0.043 −0.054 0.076** (1.948) (0.078) (0.084) (0.035) Newsletter * poor 3.7123 0.011 0.017 −0.204** (5.529) (0.132) (0.128) (0.098) Newsletter 4.742 0.156 0.024 0.095 (3.514) (0.114) (0.142) (0.050) Newsletter * less high school 2.093 −0.088 −0.020 −0.130 (4.743) (0.142) (0.186) (0.096) Newsletter * high school −2.565 −0.177 −0.131 −0.073 (4.564) (0.128) (0.140) (0.096) Newsletter 5.405 0.091 −0.005 0.108 (3.711) (0.092) (0.132) (0.063) Newsletter * very optimistic −2.264 −0.057 −0.056 −0.095 (4.307) (0.011) (0.135) (0.070) Mean control group 22 0.29 0.73 0.05 N 359 359 359 359 . Days attended VT training . Got VT qualification . Got VT certificate . Dropout (adm. variable) . Newsletter impacts 3.723** 0.045 −0.048 0.039 (1.774) (0.073) (0.078) (0.036) Heterogenous impacts Newsletter 2.415 0.030 −0.079 0.054 (2.084) (0.082) (0.098) (0.045) Newsletter * males 1.976 0.066 0.098 −0.063 (5.717) (0.129) (0.145) (0.088) Newsletter 2.845 0.081 −0.081 0.064 (1.625) (0.085) (0.082) (0.040) Newsletter * age 15–21 2.565 −0.109 0.096 −0.076 (2.929) (0.110) (0.113) (0.067) Newsletter 2.991 0.043 −0.054 0.076** (1.948) (0.078) (0.084) (0.035) Newsletter * poor 3.7123 0.011 0.017 −0.204** (5.529) (0.132) (0.128) (0.098) Newsletter 4.742 0.156 0.024 0.095 (3.514) (0.114) (0.142) (0.050) Newsletter * less high school 2.093 −0.088 −0.020 −0.130 (4.743) (0.142) (0.186) (0.096) Newsletter * high school −2.565 −0.177 −0.131 −0.073 (4.564) (0.128) (0.140) (0.096) Newsletter 5.405 0.091 −0.005 0.108 (3.711) (0.092) (0.132) (0.063) Newsletter * very optimistic −2.264 −0.057 −0.056 −0.095 (4.307) (0.011) (0.135) (0.070) Mean control group 22 0.29 0.73 0.05 N 359 359 359 359 Source: Authors' calculations. Note: Clustered standard errors by vocational class. ITT parameters estimated with control covariates selected by post-double-selection LASSO (Belloni, Chernozhukov, and Hansen 2014) among 35 baseline variables. For the administrative dropout outcome, the selected control variables are two: whether respondent was student at baseline and whether respondent was enrolled in technical vocational training education at baseline. For all other outcomes, no covariate was selected by the LASSO approach. Open in new tab Heterogeneous effects for the newsletter intervention are reported in the lower panel of table 7. Unlike the impact of vocational training, which was of benefit to relatively better-off young participants, the newsletter-related intervention shows no statistical differences in average gains across demographic groups or over time. Indeed, no statistically significant heterogeneous effects by age, gender, poverty, or level of schooling arise. The lack of statistical significance is likely related to small sample sizes across all subgroups. A related policy question is whether the combination of vocational training with newsletters about returns to training leads to higher earnings and employment for MVTP participants. Because both interventions are orthogonal to each other, two independent variables are constructed, one that takes the value 1 for individuals who were offered both treatments, 0 otherwise, and a second dummy variable that takes the value of 1 for individuals who were offered only the vocational training, 0 otherwise. Point estimates and p-values that test the null of equality between these independent variables are shown in table 8. The estimation model follows the same specification used in the computation of the main treatment effects given in table 4. While the magnitude of the coefficient associated with both treatments is higher than that for vocational training alone across most outcomes of interest, no statistically meaningful differential results are observed at 6 and 18 months after the intervention. The p-values shown at the bottom of table 8 do not reject the null of equal coefficients across all four outcome variables of interest. This result is arguably related to the short length of training offered by the MVTP, since attending three extra days of training might be marginal to change the employment and earnings odds of those receiving the newsletter treatment. Table 8. ITT Impacts of VT and Information Letters Mongolian VT Program, 2014–2016 . Employment . Labor income . Skills match . Self-employment . . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . Training + Newsletter (T1 + T2) 0.061 −0.024 31,117 41,463 0.088** 0.019 0.046* 0.019 (0.045) (0.045) (32,762) (47,352) (0.039) (0.042) (0.027) (0.027) Training (T1) 0.047 0.022 63,712** 70,872** 0.047 0.054 0.033 0.050** (0.038) (0.038) (30,555) (31,840) (0.032) (0.035) (0.021) (0.022) p-value: T1+T2 = T1 0.759 0.284 0.315 0.432 0.287 0.386 0.630 0.272 mean control group 0.455 0.550 234,326 303,516 0.212 0.250 0.068 0.064 R2 0.11 0.14 0.13 0.17 0.12 0.12 0.14 0.12 N 1043 970 1043 970 1043 970 1043 970 . Employment . Labor income . Skills match . Self-employment . . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . Training + Newsletter (T1 + T2) 0.061 −0.024 31,117 41,463 0.088** 0.019 0.046* 0.019 (0.045) (0.045) (32,762) (47,352) (0.039) (0.042) (0.027) (0.027) Training (T1) 0.047 0.022 63,712** 70,872** 0.047 0.054 0.033 0.050** (0.038) (0.038) (30,555) (31,840) (0.032) (0.035) (0.021) (0.022) p-value: T1+T2 = T1 0.759 0.284 0.315 0.432 0.287 0.386 0.630 0.272 mean control group 0.455 0.550 234,326 303,516 0.212 0.250 0.068 0.064 R2 0.11 0.14 0.13 0.17 0.12 0.12 0.14 0.12 N 1043 970 1043 970 1043 970 1043 970 Source: Author's calculations based on First and Second Follow up data. Note: Robust standard errors in parenthesis. Intent-to-treat (ITT) parameters estimated by multivariate OLS models with post-double-selection LASSO approach for selection of control variables. See notes in table 4 for full details of selected control covariates. All specifications include fixed effects by day of random assignment. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 8. ITT Impacts of VT and Information Letters Mongolian VT Program, 2014–2016 . Employment . Labor income . Skills match . Self-employment . . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . Training + Newsletter (T1 + T2) 0.061 −0.024 31,117 41,463 0.088** 0.019 0.046* 0.019 (0.045) (0.045) (32,762) (47,352) (0.039) (0.042) (0.027) (0.027) Training (T1) 0.047 0.022 63,712** 70,872** 0.047 0.054 0.033 0.050** (0.038) (0.038) (30,555) (31,840) (0.032) (0.035) (0.021) (0.022) p-value: T1+T2 = T1 0.759 0.284 0.315 0.432 0.287 0.386 0.630 0.272 mean control group 0.455 0.550 234,326 303,516 0.212 0.250 0.068 0.064 R2 0.11 0.14 0.13 0.17 0.12 0.12 0.14 0.12 N 1043 970 1043 970 1043 970 1043 970 . Employment . Labor income . Skills match . Self-employment . . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . 6-month . 18-month . Training + Newsletter (T1 + T2) 0.061 −0.024 31,117 41,463 0.088** 0.019 0.046* 0.019 (0.045) (0.045) (32,762) (47,352) (0.039) (0.042) (0.027) (0.027) Training (T1) 0.047 0.022 63,712** 70,872** 0.047 0.054 0.033 0.050** (0.038) (0.038) (30,555) (31,840) (0.032) (0.035) (0.021) (0.022) p-value: T1+T2 = T1 0.759 0.284 0.315 0.432 0.287 0.386 0.630 0.272 mean control group 0.455 0.550 234,326 303,516 0.212 0.250 0.068 0.064 R2 0.11 0.14 0.13 0.17 0.12 0.12 0.14 0.12 N 1043 970 1043 970 1043 970 1043 970 Source: Author's calculations based on First and Second Follow up data. Note: Robust standard errors in parenthesis. Intent-to-treat (ITT) parameters estimated by multivariate OLS models with post-double-selection LASSO approach for selection of control variables. See notes in table 4 for full details of selected control covariates. All specifications include fixed effects by day of random assignment. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab 7. Conclusions This study examined the effectiveness of the demand-driven Mongolian Vocational Training Program in Ulaanbaatar, the capital of Mongolia. The setting of this evaluation is new to the literature because little is known about Mongolia's labor markets, considering the country's relatively recent transition to a market economy from a centrally planned economy, in which unemployment was set to zero by law. Like other demand-driven vocational-training programs implemented since the 1990s, the MVTP aimed to counteract high levels of youth unemployment and idleness by responding to actual labor market needs through a mix of traditional classroom courses and on-the-job training (internships). This study implemented a complementary random allocation design to assess the role of information about market returns to training on the length of exposure to training. One striking result is the low take-up rate for the intervention. Around 42 percent of individuals randomly assigned to training did not attend the courses. Analysis of the determinants of take-up showed that institutional constraints, notably the requirement that quadrilateral contracts be signed prior to the start of the course, as well as the chosen field of study and some demographic variables (gender, education), played an important role in take-up decisions. In particular, the quadrilateral-contract requirement, which seeks to secure jobs for trainees after the completion of the program, seemed to be the major barrier to participation, and its role cannot be understated. Scaling-up this vocational training initiative would be problematic under the current contractual framework. Overall, positive average effects of the training intervention were observed on outcomes of interest. However, statistical significance varies across variables, model specification, and over time. Informative results are mainly observed in the short run, while a distinctive decay in the magnitude of effects is observed 18 months following the intervention. Given the short duration and thus low cost of this MVTP initiative, the magnitude of the treatment effects are somewhat sizable, and thus can inform policy about the relative success of implementing market-based ALMP approaches in settings where, prior to an economic transition, the state had been the sole provider of worker training, and individuals had little choice regarding the training to which they were assigned. As in the case of other ALMPs implemented worldwide, results showed that not everyone benefited equally from the program, which highlights the importance of policies that target both participants and the design of training content. In fact, heterogeneous effects emerged because relatively better-off, older, and better-educated students benefitted disproportionately from the intervention. Such results indicate that the MVTP should adjust its targeting design if the idea is to help those most in need. Finally, this study found that providing information to young participants on market returns to training had positive, statistically significant effects on the length of exposure to training (days trainees attended the program). Given the low cost of providing newsletters to young participants who are arguably information-constrained, such an approach could become a standard component of vocational-training programs, or the scope of such programs could be expanded to include information on returns to specific trades before prospective trainees make choices. Footnotes 1 " McKenzie (2017) assessed the results of 12 RCT vocational training programs from 8 different developing countries and reported an increase of 2.3 percentage points in employment and a 17 percent increase in earnings on average, although less than one-third of these 12 studies have reported statistically significant effects. Kluve et al. (2019) focused exclusively on vocational training programs that targeted youth following both experimental and non-experimental designs. Based on 113 studies from 31 developed and developing countries, they reported statistically significant impacts with effect sizes measured by 0.04 and 0.05 standard deviation for employment and earnings, respectively. Effects sizes for developing countries are reported to be higher than those for developed countries. 2 " The annual rate of return for an additional year of schooling ranges from 7 to 10 percent (Shatz et al. 2015; Batchuluun and Dalkhjav 2019). 3 " The MVTP is different from “technical vocational training education,” which is considered a form of formal secondary education and lasts for years. This type of technical education has led to gains in employment and wages (Field et al. 2019). 4 " In the context of formal schooling in the Dominican Republic, Jensen (2010) showed that students tended to underestimate the returns to formal schooling. When they were correctly informed, however, both enrollment in the subsequent year and the average length of formal schooling increased. 5 " Although data on oversubscribed courses is not available, unavailable trade courses refer fundamentally to courses not being offered in the corresponding call rather than to courses with high demand. 6 " The sample size was originally set to 2100 individuals corresponding to 1400 in the treatment group and 700 in the control group to be able to detect a 3 percentage-point increase in employment with a power of 80 percent and a dropout rate of 30 percent. Unfortunately, and as a result of budget revisions, the Mongolian government slashed the number of potential beneficiaries for the 2013 call. 7 " Table 1, column 9, shows attrition rates that range from 18.5 (Card et al. 2011) to 46 percent (Cho et al. 2013). Alzua, Cruces, and Lopez et al. reported no attrition because their reliance on administrative data. 8 " Gers are traditional Mongolian dwellings. They are portable, round tents covered with skins or felt that are used by several distinct nomadic groups in Central Asia. Mongolia is divided into Ulaanbaatar, the capital, and twenty-one provinces (aimags). Ulaanbaatar is divided into six main and three small suburban districts (khoroos), which are further divided into smaller administrative units called khesegs. Poor khesegs are called gers, or impoverished neighborhoods, which, in most cases, lack water, drainage, and electric infrastructure. 9 " Estimations were carried out for three different specifications of the outcome model. First, a specification with no controls. Second, a specification in which control covariates selected by post-double-selection LASSO were added (Belloni, Chernozhukov, and Hansen 2014). After considering a large set of baseline variables (+35), only four control variables were selected: gender, whether the respondent had ever worked, whether the respondent signed a “quadrilateral” contract, and whether the respondent resided in Nalaik district. A third specification added standard variables highlighted in human capital models of wage regressions (age, gender, and schooling) to the selected LASSO controls. The main table of results reports only the first and second specification because the last one yields quantitative results similar to those of the second specification. 10 " This included any previous training that the respondent had completed in any public or private training institution, including technical vocational education and training (TVET) education. 11 " In terms of percent gain on the control employment rates, the 95 percent confidence interval for employment ITT effects (column 2 in table 4) is [−3, 27] and [−11, 15] six- and eighteen-months following treatment, respectively. 12 " Winsorization and trimming used the top 95 percentile of earnings as the cut-off value. 13 " In terms of percent gain on the control income, the 95 percent confidence interval for earnings ITT effects (column 4 in table 4) is [−2, 45] and [−1, 42] 6- and 18 months following treatment. 14 " This calculation considers the explicit costs of training that range from USD 90 to USD 140 plus implicit (forgone) costs that are computed as one month of control group mean earnings times the probability of being employed. Monthly benefits are computed using the statistically significant ITT monthly earnings depicted in column 4 in table 4 (Tughrik $53255, or equivalent to approximately USD 27). A conservative skills depreciation rate (20 percent per year) is assumed for this calculation. 15 " In Ulaanbaatar, informal employment for youth is estimated to be lower than 15 percent, way below the observed informal employment rate in Latin America or African countries. 16 " An issue to consider when comparing the ITT and TOT impacts is that take-up rates vary with some of these demographic variables. Thus, while ITT estimates capture the combined effect of any differences in take-up across groups with differences in the effectiveness of training by group, TOT estimates are subject to the concern that complier status differs across demographic groups. 17 " The Hochberg's test, which is implemented at level δ = 0.10, controls the proportion of false positives within the set of rejected hypotheses. After implementing this test for the heterogenous results, it still rejects the null hypothesis for the point estimates that are reported in tables 5a and 5b, holding (unadjusted) statistical significance at the 1 percent or 5 percent level. However, this test does not reject the null hypothesis for the point estimates that are reported in tables 5a and 5b holding, (unadjusted) statistical significance at the 10 percent level. In this regard, correcting for multiple-hypothesis testing leads to the loss of statistical significance for the differential effects of both age on skills matching and of poverty status on self-employment six months following the intervention. For the medium-term effects, statistical significance is lost for the differential impacts of poverty status on earnings. 18 " In the first step, the outcome is predicted in the absence of the treatment using two sets of information to assess the sensitivity of results to the selection of baseline variables: a basic set of socio-demographic variables and a rich set of +35 baseline variables. In the second step, an adjusted-outcome specification is implemented by using the same specification employed in the computation of the main treatment effects – i.e., fixed effects by day of random assignment and LASSO selected control covariates. 19 " For the administrative dropout outcome, the LASSO approach selected two control variables: whether the respondent was a student at baseline and whether the respondent was enrolled in technical vocational education at baseline. No control variable was selected for the other outcomes of interest. Notes María Laura Alzúa (corresponding author) is Deputy Director at CEDLAS, Universidad Nacional de La Plata, Buenos Aires, Argentina and Research Director at the Partnership for Economic Development; her email address is [email protected]. Soyolmaa Batbekh is an associate professor at the National University of Mongolia, Ulaanbaatar, Mongolia; her email address is [email protected]. Altantsetseg Batchuluun is an associate professor at the National University of Mongolia, Ulaanbaatar, Mongolia; her email address is [email protected]. Bayarmaa Dalkhjav is a lecturer at the National University of Mongolia, Ulaanbaatar, Mongolia; her email address is [email protected]. José Galdo is an associate professor at Carleton University, Ottawa, Canada; his email address is [email protected]. The authors would like to thank the comments from the editor and two anonymous referees, Brian Feld, Natalia Cantet, participants at the LACEA 25th meeting, the 3rd IZA workshop in skills and preferences in post-transition and emerging economies (St. Petersburg, Russia), and the DECON seminar series at the Universidad de la República, Uruguay. This work was carried out with financial and scientific support from the Partnership for Economic Policy (PEP) and with funding from the Department for International Development (DFID) of the United Kingdom (or UK Aid) and the Government of Canada through the International Development Research Centre (IDRC). The views and opinions expressed in this publication are those of the authors and do not necessarily reflect those of PEP. This study is registered at the AEA RCT with reference number AEARCTR-0002461. A supplementary online appendix is available with this article at the World Bank Economic Review website. References Abadie A. , Chingos M. M., West M. R.. 2018 . “ Endogenous Stratification in Randomized Experiments .” Review of Economics and Statistics 100 ( 4 ): 567 – 80 . Google Scholar Crossref Search ADS WorldCat Acevedo P. , Cruces G., Gertler P., Martinez S.. 2017 . “Living Up to Expectations: How Job Training Made Women Better Off and Men Worse Off.” NBER Working Paper No. 23264 . National Bureau of Economic Research . Cambridge, MA, USA . Adoho F. , Chakravarty S., Korkoyah D. T., Lundberg M., Tasneem A.. 2014 . “ The Impact of an Adolescent Girls Employment Program: The EPAG Project in Liberia .” Policy Research Working Papers . The World Bank . OpenURL Placeholder Text WorldCat Alzúa M. L. , Cruces G., Lopez C.. 2016 . “ Long-Run Effects of Youth Training Programs: Experimental Evidence from Argentina .” Economic Inquiry 54 ( 4 ): 1839 – 59 . Google Scholar Crossref Search ADS WorldCat Attanasio O. , Kugler A., Meghir C.. 2011 . “ Subsidizing Vocational Training for Disadvantaged Youth in Colombia: Evidence from a Randomized Trial .” American Economic Journal: Applied Economics 3 ( 3 ): 188 – 220 . OpenURL Placeholder Text WorldCat Bandiera O. , Buehren N., Burgess R., Goldstein M., Gulesci S., Rasul I., Sulaiman M.. 2020 . “ Women's Empowerment in Action: Evidence from a Randomized Control Trial in Africa .” American Economic Journal: Applied Economics 12 ( 1 ): 210 – 59 . Google Scholar Crossref Search ADS WorldCat Belloni A. , Chernozhukov V., Hansen C.. 2014 . “ Inference on Treatment Effects after Selection among High-Dimensional Controls .” Review of Economic Studies 81 ( 2 ): 608 – 50 . Google Scholar Crossref Search ADS WorldCat Batchuluun A. , Dalkhjav B.. 2019 . “ Supply Analysis of Labor Market in Mongolia .” Economics: Theory and Practice 33 : 87 – 120 . OpenURL Placeholder Text WorldCat Betcherman G. , Godfrey M., Puerto S., Rother F., Stavreska A.. 2007 . “ A Review of Interventions to Support Young Workers: Findings of the Youth Employment Inventory .” Social Protection Discussion Paper Series, No. 715. The World Bank. Washington, DC, USA . OpenURL Placeholder Text WorldCat Calero C. , Leite C., Carlos H., Gonzales V., Kluve J., Soares Y.. 2014 . “ Can Arts-Based Interventions Enhance Labor Market Outcomes Among Youth? Evidence from a Randomized Trial in Rio De Janeiro .” IZA Discussion Paper No. 8210 . OpenURL Placeholder Text WorldCat Card D. , Ibarrarán P., Regalia F., Rosas-Shady D., Soares Y.. 2011 . “ The Labor Market Impacts of Youth Training in the Dominican Republic .” Journal of Labor Economics 29 ( 2 ): 267 – 300 . Google Scholar Crossref Search ADS WorldCat Card D. , Kluve J., Weber A.. 2018 . “ What Works? A Meta-Analysis of Recent Active Labor Market Program Evaluations .” Journal of the European Economic Association 16 ( 3 ): 894 – 931 . Google Scholar Crossref Search ADS WorldCat Cho Y. , Kalomba D., Mobarak A. M., Orozco-Olvera V. H.. 2013 . “ Gender Differences in the Effects of Vocational Training: Constraints on Women and Drop-Out Behavior .” IZA Discussion Paper No. 7408 . OpenURL Placeholder Text WorldCat Choe C. , Flores-Lagunes A., Lee S.-J.. 2015 . “ Do Dropouts with Longer Training Exposure Benefit from Training Programs? Korean Evidence Employing Methods for Continuous Treatments .” Empirical Economics 48 : 849 – 81 . Google Scholar Crossref Search ADS WorldCat Chong A. 2011 . Development Connections. Unveiling the Impact of New Information Technologies . London, UK : Palgrave McMillan . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Dammert A. , Galdo J.. 2013 . “ Program Quality and Treatment Completion for Youth Training Programs .” Economics Letters 119 ( 3 ): 243 – 6 . Google Scholar Crossref Search ADS WorldCat Díaz J. J. , Rosas D.. 2016 . “ Impact Evaluation of the Job Youth Training Program .” IDB Working Paper Series No. IDP-WP-693. Inter-American Development Bank. Washington, DC, USA . OpenURL Placeholder Text WorldCat Dinkelman T. , Martínez A. C.. 2013 . “ Investing in Schooling in Chile: The Role of Information about Financial Aid for Higher Education .” Review of Economics and Statistics 96 ( 2 ): 244 – 57 . Google Scholar Crossref Search ADS WorldCat Field E. , Linden L., Malamud O., Rubenson D., Wang S.-Y.. 2019 . “ Does Vocational Education Work? Evidence from a Randomized Experiment in Mongolia .” NBER Working Paper No. 26092 . National Bureau of Economic Research . Cambridge, MA, USA . OpenURL Placeholder Text WorldCat Gabaix X. , Laibson D.. 2006 . “ Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive Markets .” Quarterly Journal of Economics 121 ( 2 ): 505 – 40 . Google Scholar Crossref Search ADS WorldCat Galdo J. , Chong A.. 2012 . “ Does the Quality of Public-Sponsored Training Programs Matter? Evidence from Bidding Processes Data .” Labour Economics 19 ( 6 ): 970 – 86 . Google Scholar Crossref Search ADS WorldCat Hirshleifer J. , Riley J. G.. 1992 . The Analytics of Uncertainty and Information . Cambridge, UK : Cambridge University Press . Google Scholar Crossref Search ADS Google Scholar Google Preview WorldCat COPAC Honorati M. 2015 . “ The Impact of Private Sector Internship and Training on Urban Youth in Kenya .” Policy Research Working Papers 7404. The World Bank . Washington, DC, USA . OpenURL Placeholder Text WorldCat Ibarrarán P. , Kluve J., Ripani L., Shady D. R.. 2019 . “ Experimental Evidence on the Long-Term Effects of a Youth Training Program .” ILR Review 72 ( 1 ): 185 – 222 . Google Scholar Crossref Search ADS WorldCat Ibarrarán P. , Ripani L., Taboada B., Villa J. M., Garcia B.. 2014 . “ Life Skills, Employability and Training for Disadvantaged Youth: Evidence from a Randomized Evaluation Design .” IZA Journal of Labor & Development 3 ( 1 ): 10 . Google Scholar Crossref Search ADS WorldCat Ibarrarán P. , Rosas D.. 2009 . “ Evaluating the Impact of Job Training Programmes in Latin America: Evidence from IDB Funded Operations .” Journal of Development Effectiveness 1 ( 2 ): 195 – 216 . Google Scholar Crossref Search ADS WorldCat International Labour Organization (ILO) . 2015 . Youth and Informality Promoting Formal Employment Among Youth: Innovative Experiences in Latin America and the Caribbean (Lima, Regional Office for Latin America and the Caribbean) . Geneva : ILO . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC International Labour Organization (ILO) . 2017 . Global Employment Trends for Youth 2017 : Paths to a Better Working Future. ILO. Geneva, Swtizerland . OpenURL Placeholder Text WorldCat Jensen R. 2010 . “ The (Perceived) Returns to Education and the Demand for Schooling .” Quarterly Journal of Economics 125 ( 2 ): 515 – 48 . Google Scholar Crossref Search ADS WorldCat Kluve J. , Puerto S., Robalino D., Romero J. M., Rother F., Stöterau J., Weidenkaff F., Witte M.. 2019 . “ Do Youth Employment Programs Improve Labor Market Outcomes? A Quantitative Review .” World Development 114 : 237 – 53 . Google Scholar Crossref Search ADS WorldCat Lehmann H. , Muravyev A.. 2012 . “ Labor Market Institutions and Labor Market Performance: What Can We Learn from Transition Countries? ” Economics of Transition and Institutional Change 20 ( 2 ): 235 – 69 . Google Scholar Crossref Search ADS WorldCat McKenzie D. 2017 . “ How Effective Are Active Labor Market Policies in Developing Countries? A Critical Review of Recent Evidence .” World Bank Research Observer 32 ( 2 ): 127 – 54 . Google Scholar Crossref Search ADS WorldCat Pastore F. 2010 . “ The Gender Gap in Early Career in Mongolia .” International Journal of Manpower 31 ( 2 ): 188 – 207 . Google Scholar Crossref Search ADS WorldCat Research Institute of Labor and Social Protection . 2017 . “ Mongolia's Labor Markets .” http://rilsp.gov.mn/tasudalgaa.php?count = 50?page=1&action=more&sudalgaa_id=1 . OpenURL Placeholder Text WorldCat Schmillen A. , Sandig N. W.. 2018 . Perceptions of Precariousness: A Qualitative Study of Constraints Underlying Gender Disparities in Mongolia's Labor Market . Washington, DC : World Bank . Google Scholar Crossref Search ADS Google Scholar Google Preview WorldCat COPAC Shatz H. J. , Constant L., Perez-Arce F., Robinson E., Beckman R. L., Huang H. C., Glick P., Ghosh-Dastidar B.. 2015 . Improving the Mongolian Labor Market and Enhancing Opportunities for Youth . Santa Monica, CA : RAND Corporation . https://www.rand.org/pubs/research_reports/RR1092.html . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Spinnewijn J. 2015 . “ Unemployed but Optimistic: Optimal Insurance Design with Biased Beliefs .” Journal of the European Economic Association 13 ( 1 ): 130 – 67 . Google Scholar Crossref Search ADS WorldCat World Bank . 2013 . “ Mongolia Enterprise Survey (ES). Ref. MNG_2013_ES_v01_M .” World Bank . Washington, DC, USA . https://datacatalog.worldbank.org/dataset/mongolia-enterprise-survey-2013 . OpenURL Placeholder Text WorldCat © The Author(s) 2020. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)