Pathways from School to Work in the Developing World

This paper uses micro data from the ILO-STWT surveys to provide novel evidence on the duration, end point and determinants of the transition from school to work in a sample of 23 low and middle-income countries around the world. The negative effects of low levels of human capital and high levels of population growth on job finding rates, seems to be at least in part offset by widespread poverty and lack of unemployment insurance, leading to overall faster transitions in low income economies compared to middle income economies. By lowering reservation wages and speeding transitions these latter forces lead overall to worse matches, as measured by the probability of attaining stable employment in the long-run.


Introduction
In this paper we present novel evidence on transition durations from school to work and on the probability of ever finding employment and stable employment in a sample of 23 low and middle-income countries around the world. To do so we exploit unique and yet unutilized micro data from the ILO School to Work Transition (STWT) surveys that collect retrospective information on work histories on a sample of around 35,000 individuals aged 15 to 29 around the world.
Young individuals are at particular risk of unemployment and joblessness. This is true both in developed as well as in the developing world (ILO 2013(ILO , 2014. In addition to joblessness, underemployment, informality, low wages and lack of quality employment affect youth labor markets, especially in low-income countries.
A critical stage in an individual's working life is the transition leading from school to the first employment spell. Lengthy transitions impose an array of individual and possibly social costs and they might potentially lead to worse job matches, either due to the effects of prolonged unemployment spells on human capital stock or to employers perceiving long unemployment spells as a signal of low productivity. Long transitions might even translate into worse labor market outcomes in the long term, i.e. lower employability, lower wages and/or poor quality jobs.
One should not necessarily assume though that fast transitions are an indication of good matches or are individually or socially desirable. Short durations might be the result of low reservation wages or poor prospects of finding decent employment and in turn lead to long-term underemployment and poor quality employment. Some evidence exists on the length of transitions from school to work in more advanced and, to a minor extent, in emerging economies (see for example Cunningham and Salvagno 2011and OECD 1998, Quintini and Martin 2007, Ryan 2001 for high income countries). Much of this evidence points to the role of lack of adequate skills as an obstacle to timely and successful transitions.
Considerably less is known, however, about the characteristics and key correlates of transition from school to work in developing countries. As the majority of youth worldwide live in low and middle income countries, characterizing these transitions and understanding their determinants is of primary importance.
There is convincing evidence from more advanced economies on the negative effect of protracted job search and unemployment, especially during youth. Long-3 term unemployment has itself adverse consequences on the probability of finding work through negative state dependence (Machin and Manning 1999) and postunemployment wages appears to be negatively affected by the length of the unemployment period (see for example Card, Chetty and Weber 2007, Lalive 2007, Schmieder, von Wachter and Bender 2014. A number of related studies also show evidence of the scarring effects of joblessness, which can persist in an individual's life (see Ellwood 1982, Gregg and Tominey 2001, Mroz and Savage, 2006, von Wachter and Bender 2006. Economic theory rooted in the canonical search and matching model (Pissarrides 2000) suggests that the demographic bulges and the low arrival rate of high quality wage offers that characterize many developing economies will negatively affect employment prospects in the long run, although their effects on transition durations are ambiguous. While demographic pressure and the paucity of decent job opportunities are likely to mechanically lead to lower job offers arrival rates and longer transition durations, these will also push workers to accept low wages and low quality jobs. As better opportunities than the current ones are unlikely to materialize in the future, there are disincentives to wait when job-to-job mobility is costly, fostering circles of poor quality employment and underemployment.
Similarly, widespread poverty and lack of unemployment insurance that also characterize low income economies, will likely lower reservation wages and speed the transition process, as well as possibly directly lead to worse labor market outcomes in the log-run.
In order to systematically investigate the length and end-point of the transition process from school to the labor market and its determinants in low and middle income countries, in this work we use retrospective data on work histories since the time of leaving school on a sample of around 35,000 individuals from Africa, Asia, Latin America and the Caribbean and Eastern Europe. These data are clearly rightcensored, as not all individuals will have transited to employment by the time of the survey. More importantly, not all of those who are observed not having transited at the time of the survey will necessarily ever transit to employment, let alone to high quality employment. In order to account for this, we fit to the data a split cure model (Schmidt and Witte, 1989) where we simultaneously parameterize the probability of ever transiting to employment (or stable employment) and the duration of 4 unemployment. This allows us to obtain estimates of baseline hazard rates net of differences in socioeconomic characteristics of the population across countries.
In our sample of countries, and similar to high-income countries, hazard rates display pronounced negative duration dependence. While transition durations are on average not very dissimilar from those observed in high-income countries, this masks substantial heterogeneity across regions. Transition durations are the lowest in low income countries in Sub-Saharan Africa and Asia and the Pacific and the highest in middle income countries in Latin America and the Caribbean and in particular in the Middle East and North Africa. We also find that in some low and middle-income countries a substantial fraction of individuals are predicted to never transit to employment.
We next investigate the determinants of the heterogeneity in the transition duration and the probability of ever transiting to employment or to stable employment across individuals and countries. Those with higher levels of education (proxied by higher school leaving age) and those with work experience while in school transit faster and are more likely to ever find employment than early school leavers, possibly pointing to the role of early human capital accumulation and education in shaping transitions. Women are less likely to transit, and, if they transit, generally do so at a slower rate than men, although we reaming agnostic on whether this reflects preferences or constraints.
We also show that, despite the extreme pressure put on labor markets by demographic bulges that negatively affect job finding rates especially in Sub-Saharan Africa, transition durations are markedly lower in low income countries compared to middle and high income countries due to a combination of widespread poverty and lack of unemployment insurance, both of which negatively affect reservation wages.
These same forces appear to lead to a lower probability of finding employment and stable employment in the long run. Although one has to be cautious in drawing any causal inference from this evidence, these results suggest that shorter durations driven by low reservation wages in low-income countries have potentially long-run welfare costs in terms of a reduced probability of achieving high quality employment.
The rest of the paper is organized as follows. Section 2 presents the data. Section 3 presents simple descriptive statistics on youths' labor market status across 28 countries for which the ILO survey data are available. Section 4 presents estimates of the split sample model for 23 (of the 28) countries for which work histories can be 5 reliably derived from the surveys and investigates the individual and macro correlates of such phenomena. Section 5 concludes.

Data: The ILO School to work transition surveys
Through the Work4Youth partnership with The MasterCard Foundation, the International Labour Organization has recently embarked on an unprecedented data collection effort on youths' labor market outcomes in a sample of 28 low and middleincome countries around the world. This effort is in response to the paucity of data, in particular of systematically comparable data, on labor market outcomes and transitions from many non-high income countries. The list of countries, including sample sizes, geographical coverage and sample period is presented in Table A1. The The surveys collect a large set of information on current labor market status plus selected retrospective information on labor market experiences for a nationally representative sample of individuals aged 15-29. Retrospective information is collected since the time the individuals left school or since the first labor market experience for those who never attended school. This is integrated with information on current individual and household-level characteristics (such as age, gender, highest education level completed, age left education, area of residence, etc.).
Retrospective labor market information covers all past spells of employment and job search. For each spell the data report the start and end months and years. Note that the surveys only collect information on past employment spells among individuals not currently in education. Information on work during school for those not currently in 6 education is limited to a variable for whether or not an individual worked while attending school; there is no information on labor market experience for those currently in school.
No information is available on wages or earnings other than for the current employment spell. There is also little information on individual and household characteristics in the past or on how they evolved over time. 2 As the ILO-STWT surveys only collect information on labor market spells from the time the individual left school, we cannot identify employment spells that happened and concluded before leaving school (although, as said, we have an indicator on whether individuals worked or not while attending school). More importantly, for each employment spell that started before leaving school but that was ongoing at the time the individual left school, the survey reports as a start date the month following the one of leaving school. 3 In practice, this means that one cannot distinguish genuine direct transitions from school to work from apparent transitions, i.e., transitions that happened before leaving school and that continued after the individual left school. For this reason, one needs to be very cautious in interpreting spells recorded as starting just after leaving school as direct transitions from school to work.
In the following section we present descriptive statistics on current employment status. In Section 4 we turn to an analysis of the transition from school to work. Section 5 concludes.

Labor market and education outcomes at the time of the survey
Labor market and education outcomes for young individuals aged 15-29 in each of the 28 countries are reported in Table 1. In the remainder of the analysis we weigh observations for each country by sampling weights. This allows us to obtain estimates of the population parameters in each of the countries analyzed.
2 For each past employment spell, the data also report the type of employment (whether wage-worker, unpaid family worker or self-employed), and, for wage workers, the existence and characteristics of the work contract (whether written or oral and whether temporary or stable) but as said no information on wages. 3 The only exception is Brazil for which employment spells prior to the time of leaving school are recorded. For consistency, we artificially left-censor the data for Brazil at the time of leaving school.

7
The average (population weighted) unemployment rate across all countries is 13 percent (column 1). This is in contrast to an average unemployment rate for 15-29 years olds across the OECD of 16.2 percent in 2012 (ILO, 2013). Unemployment is the lowest in low income countries, in particular SSA and AP (11 and 8 percent, respectively) and the highest in MENA (18 percent) and LAC (16 percent). 4 Unemployment is only one indicator of young persons' fortunes in the labor market. Table 2 reports data on the type of jobs held by employed youth. Again, some regional patterns are apparent: unpaid family work and self-employment are more prevalent in low income countries (46 and 25 percent, respectively, in SSA relative to a sample mean of 25 and 18 percent, respectively); and wage employment is less likely in low income countries (28 percent in SSA relative to a sample mean of 56 percent) compared to high income countries. In contrast, the fraction of young workers working for a wage is as high as 88 percent in EECA.
In order to complement information on the quality of jobs held, the last column of Table 2 reports information on the fraction of working youth in stable employment, defined as wage work with an indefinite contract or with a contract of at least 12 months' duration. 5 Again, not surprisingly, the fraction of youth in stable employment tends to be higher in middle-income countries than in low-income countries, reflecting patterns of stable and wage employment in the adult population. EECA countries in particular show the highest rates of stable employment (79 percent), followed by countries in LAC (62 percent).
In sum, youth unemployment is widespread in middle-income countries and so are, comparatively speaking, high quality jobs, while the reverse is true in low-income countries. These results point to the direction of lack of high quality jobs, widespread poverty and lack of social insurance possibly driving both low unemployment and low quality employment among youth in low-income countries. We try to assess the 8 effect of these variables more formally in the next section through a focus on transitions from school to work.

Transition from school to work
Higher incidence of youth unemployment in middle income countries -and in particular in MENA countries -compared to low income countries can be due to lower inflows into employment or higher outflows out of employment and higher turnover. In the rest of the paper we focus on accession rates and in particular on the transition to the first employment spell and to the first spell of stable employment. We focus on individuals who left education; we disregard those who never attended school, for whom the concept of school to work transition clearly does not apply. 6 Among those who left education, a sizeable fraction (28 percent, i.e. 45 percent of those who ever attended) left education before age 15 (see column 2 of Table 3).
As expected, this fraction is particularly high in SSA (65 percent of those who ever attended). Table 3 presents also information on completed transitions. On average, 38 percent of youth aged 15-29 report having attended school and having had one work experience since the time they left education (column 5). This fraction is the lowest in MENA (27 percent). A substantial fraction of individuals -around 20 percent -also worked while in school (column 4).
Note that if around 61 percent of those out of school had a job since leaving education (this is 37.6 in column 5 divided by 61.2 in column 1), only about 31 per cent of these individuals (=19.2/61.2) had at least one spell of stable employment; this fraction is unsurprisingly the lowest in SSA (14 percent).
Finally, a substantial share of individuals had no work experience after leaving school (16.5 percent, this is the sum of columns 7 and 8). 19 percent (=11.8/61.2) of individuals have been continuously out of the labor force since the time of leaving school while a small fraction (8 percent = 3.1/65.5) had at least one spell of job search or training. As illustrated below, youth continuously out of the labor force are disproportionately women.  Table 3 are right-censored as individuals who have not transited to employment by the time of the survey might do so in the future. To circumvent this problem, we fit a duration model to the data (see Lancaster 1992 and Jenkins 2005), separately by country. This allows us to compute duration to employment -overall and by sub-groups defined on observable characteristics -and to derive baseline transition probabilities net of compositional effects, i.e. differences in observable characteristics of the youth population across countries. We can also explore the determinants of transition duration across countries and we turn to this in the next section.

Data in
As illustrated in Table 3, a relatively large number of out of school youth have been continuously inactive since the time they left education. This is possibly an indication that some of these individuals will never transit to employment. Treating these individuals as if they had extremely long transition durations might lead one to erroneously conclude that mean transition durations are very high. In order to account for this possibility, we employ a split-population (also sometimes referred to as spiltcure) model (Schmidt and Witte, 1989). This model provides a simultaneous estimate of the duration of the transition to employment among those who are expected to transit as well of the probability of never transiting (also sometimes referred to as the probability of being cured).
Before proceeding to the estimation of the model, we have applied a variety of sample selection criteria to the data (see Table A1 in the Appendix). The information provided by the surveys as well as sample sizes vary considerably across countries. In some countries the data quality is poor or the information provided is scant, making it necessary to exclude them from the analysis. This applies to Bangladesh, Colombia, Liberia, Malawi and Zambia, leaving us with a sample of 23 countries out of the 28 for which ILO-STWT surveys are available. 7 Note that for Egypt we drop more than 40 percent of the sample while for Macedonia and the OPT we drop about 20 percent 10 of the observations. We retain these countries in the analysis, but some care should be exerted in interpreting the estimates, especially in the case of Egypt.
On average we have around 1,500 observations per country. Given the limited number of observations, highly parameterized models often fail to converge. For this reason, and after some experimentation, we have decided to parameterize the probability of never experiencing a failure (i.e., of never transiting to a job) as a loglog distribution and to constrain the hazard function to follow a proportional hazard Weibull distribution. The Weibull parameterization constraints the hazard function to be monotonically increasing or decreasing in duration while the proportional hazard specification restricts the hazard rate to be parallel across groups with different values of the covariates. 8 Model estimates of the transition to the first employment spell for the 23 countries for which we have data are reported in Appendix. Table A3 contains the estimates of the probability of never transiting to a job and Table A4 the estimates of the duration model for individuals expected to eventually transit to a job. The splitpopulation model fails to converge for Brazil due to a very small number of censored observations in the data. For this country we decided, therefore, to estimate a standard proportional hazard Weibull model and hence, while we report model estimates for the duration of the transition from school to the first job (in Table A4), we do not report model estimates for the probability of never transiting (Table A3).
We include a number of covariates in the model. Given the limited number of observations and characteristics available in the survey, and since most of the characteristics are observed at the time of the survey rather than at the time of leaving school (i.e., at the onset of risk) we are constrained in the number of variables that we can include in the model. The model includes a gender dummy, a dummy for urban/rural location, dummies for father's educational level (up to primary, up to secondary and higher), dummies for three school leaving age groups (less than 16, between 16 and 18 and greater than 18), 9 a dummy for whether or not the individual 8 Although this parameterization is necessarily restrictive, as it does not allow the hazard function for example to first decrease and then increase in duration, its advantage is parsimony, as the hazard function only depends on two parameters. Given the small number of observations available for each country and the varying quality of the data, we have found that split-population models that allow for a larger number of parameters fail to converge in most of the countries. 9 We prefer to use age left education rather than highest education level achieved because the latter is defined only coarsely and the categories of completed education vary across countries. Some degree of caution is needed here though, as, due to late entry, intermittent attendance and widespread grade ever worked while in school and (a 3-year average of) per capita GDP growth at the time of leaving school, in order to account for the possible impact of initial labor market conditions on the subsequent probability of finding employment. Finally, we include dummies for missing values of all included variables. Note that not all variables (or not all values of these variables) are available for all the countries in the sample, so we report associated coefficients for the variables (or the categories of variables) for which data are available. As for the descriptive statistics, estimates are weighted by sampling weights.

Probability of never transiting into employment after leaving school
In order to characterize the transition from school to work across countries we focus on the predicted probabilities of never transiting and mean durations (among those expected to transit) based on the estimated split-cure model. These figures are easier to interpret than the estimates of the model parameters. We also present mean durations and mean probabilities separately by groups defined based on observable characteristics in order to characterize heterogeneity across groups. Table 4, column (1), reports the predicted probability of never transiting to work in each of the countries analyzed. This fraction ranges from 1 percent in Vietnam to 50 percent in Samoa. On average, across all our sample of countries, around 10 percent of youth out of school are expected to never transit to employment. The predicted fraction of youth never expected to transit is the highest in the MENA region, on the order of 10 percent, and it is the lowest in SSA and AP, with a probability of transition to employment of more than 90 percent. 10 Estimates of the transition to stable employment are reported in column (2) of Table 4. Some caution is needed here, as stable employment is rare is many countries in our sample and the number of individuals in stable employment in the data is often very limited, implying that estimates of the model parameters are often imprecise.
The share of youth expected to never transit to a stable job is the lowest in LAC (on the order of 45 percent) and the highest in MENA (at around 78 percent). retention individuals with the same age left education might have different levels of completed education.

12
In order to quantify the role of observable characteristics in explaining differences in transition probabilities across countries, we have recomputed these probabilities while keeping individual characteristics fixed. To do so, we have computed predictions from each country-specific model over the entire sample of data (i.e., for all the individuals in all the countries in our dataset). Figure A1 plots these composition-free estimates versus in-sample predictions.
The figure shows that these two series line up remarkably well, implying that observable characteristics play little role in explaining differences across countries in the probabilities of transition. These differences are most likely due to differences in structural or cyclical economic factors, although clearly one cannot rule out the possibility that unobserved individual characteristics, which we cannot account for, are also responsible for these differences.

Duration of transitions from school to work
Estimates of the baseline hazard function, i.e., of the probability of finding a job in the current period conditional on not yet having found one up to the previous period, are reported in Figure A2 in the appendix. There is evidence across all countries of negative duration dependence: the probability of finding a job among the survivors falls as duration increases, in line with findings from the unemployment literature in more developed economies (e.g. Machin and Manning 1999).
Column (1) of Table 5 reports the average estimated duration of the transition to the first employment spell among those who are ever expected to transit. Durations are top-coded at 150 months. Average duration to first employment across all countries considered is just above two years (i.e., 27 months). This number again masks substantial heterogeneity across regions and countries. Transitions are the longest in MENA, with an average duration of 52 months, i.e., four and half years, and the lowest in EECA, at 17 months, followed by LAC, SSA and AP.
Given that the distribution of durations is highly skewed to the right, median durations are possibly a better indication of central tendency. Median durations are reported in column (3) of Table 5. This is effectively the time by which 50 percent of individuals in each country are predicted to have transited to employment after leaving school. As expected, median durations are systematically below average 13 durations, especially in countries with high durations (and on average 17 months, i.e., almost one and a half year), but the ranking across regions is preserved.
An alternative way to characterize the distribution of durations is to compute the predicted fraction of youth expected to find their first job within 6, 12 or 36 months of leaving school. These are reported in Appendix Table A7. In most of the countries at least 50 percent or more of the youth (among those eventually expected to transit to a job) are expected to be in employment within six months after leaving school, the exception being countries in the MENA region and a few others such as, for example, Armenia and Togo. After three years 90 percent of youth have transited to a job in most countries. In countries from the MENA region and a few from SSA, however, there are between 20 percent and 25 percent of youth still in transition three years after leaving school.
As in the case of the probability of ever transiting, we have attempted to assess to what extent cross-country differences in mean durations can be attributed to differences in the composition of the population. As shown in Figure A3, there is no evidence of compositional effects being responsible for a significant share of the differences in average transition durations across countries.
Column (2) of Table 5  It is instructive to compare the durations in the low and middle-income countries under analysis to those in high-income countries. The bottom panel of Table 6 reports the estimated average transition duration in sample of European countries estimated by Quintini et al (2007) based on the European Community Household Panel (waves 2 to 8). Despite differences in methodology and data collection instruments, average transition time to the first job across the EU countries in the sample (23 months) is very close to the average in our sample of low and middle income countries (27 months). Again, there is variability across EU countries but this is not as pronounced as the one in our sample.
In sum, we find that transition durations in low and middle-income countries, with the notable exception of the MENA region, are slightly shorter than in advanced economies. This overall conclusion is tempered by the observation that a substantial fraction of youth in the low and middle income countries under analysis -on average 10 percent -are expected to never transit to employment, let alone to stable employment. Not only are transitions to stable employment unlikely to occur but also, when they do in fact occur, they are lengthy.

Heterogeneity analysis
In this section we investigate differences in the probability of never transiting to work and in the duration to employment across groups defined based on observable characteristics.

Gender
Gender appears to be a very significant predictor of young persons' transition status after leaving education. The first two columns of Table 6 show that the probability of never transiting to employment is disproportionately high for females (on average 25 percent compared to 9 percent for males); this is particularly true in middle-income countries in MENA, LAC and EECA. In contrast, there are not very pronounced gender differences in AP and SSA 12 .
Not only are females less likely to transit to employment than males, but also those who do in fact eventually transit are expected to experience substantially longer transition durations than men in several countries. This is shown in columns (1) and (2) of Table 8 that report average median duration separately for boys and girls. Even restricting to those eventually expected to transit, females suffer a disadvantage in terms of duration to employment compared to males of around 13 months, i.e., over a year (25 months versus 12 for boys).

School leaving age
Column (2) of Table 3 illustrates that in many countries in our sample, especially from SSA and AP, a substantial fraction (on average 28 percent) of youth leave school by age 15.
Columns (3) to (5) of Table 6 show that in almost all regions early school leavers are disadvantaged in terms of the probability of ever transiting to employment relative to those who stay in school longer (the probability of ever transiting is 86 percent for early school leavers compared to 92 percent among those who leave education after age 18). With the exception of MENA, where those with high levels of education tend to transit at a slower rate, the disadvantage among early school leavers is also evident in the duration of transition to employment, in columns (3) to (5) of Table 8. While median transition duration is on the order of almost two years (23 months) for those who left education by age 15, among those who leave school after age 18 is less than one year (11 months). If school-leaving age precedes the legal age of employment this could contribute to explain longer transition durations among early school leavers compared to those who stay on longer. The circumstance that the probability of ever transiting to a job is lower for early school leavers suggests some returns from attending school longer (although unobserved heterogeneity is clearly an alternative explanation).

Work prior to leaving education
A question that naturally arises in analyzing the transition from school to work is the impact of involvement in work prior to leaving education. As illustrated in column (4) of Table 3 the share of youth that worked while school is far from negligible (on average 20 percent); this fraction is particularly high in low-income countries in SSA and AP. Although the data do not provide information on the characteristics of work 16 performed (whether continuous or seasonal, its duration, the sector of employment, whether market work or unpaid occasional work in the family farm/enterprise etc.), in many cases those combining school and work began doing so prior to the minimum working age, and therefore were child laborers according to national laws and international standards.
While early labor market involvement might harm subsequent employability due to its negative effect on schooling, it could also operate in the opposite direction: learning-by-doing associated with work early in life might also provide valuable human capital (in the form of both cognitive and non cognitive skills) and speed the transition into the labor market. However, one has to be cautious in attaching a causal interpretation to these estimates, as those with early work experiences might have persistently low reservation wages, implying that they display persistently high labor market attachment (see Hotz et al. 2002).
The results in Tables 6 and 7 (columns 6 and 7) show that working while in school is associated to a greater probability of working later in life and typically shorter durations. As in fact we are conditioning on school leaving age -hence, albeit imperfectly, controlling for the detrimental effect of early work on school attainmentthis suggests that either low reservation wages or the experience associated with early work involvement lead to greater employment in youth. This clearly does not mean that these youth are better off during their life cycle, as the jobs that they eventually attain are likely of worse quality relative to the jobs of those who did not work as children.

Additional covariates
We fail to find consistently signed effects of the other covariates on the duration to employment and on the probability of ever transiting to employment across the countries in the sample. While, for example, higher GDP growth at the time of entering the labor market appears to reduce the length of transition from school to work in many of the countries analyzed, there are exceptions to this (e.g. in Tanzania). Possibly higher economic growth, while being associated with stronger labor demand and hence faster transitions, might also affect reservation wages or the expectation of finding work in the future and hence increase duration transitions.
Mixed results are also found for the rural/urban dummy and for indicators for father's education. Again, while higher levels of education among fathers might positively affect the speed of transition or the probability of ever finding employment through, for example, valuable labor market contacts or intergenerational persistence in ability and human capital, higher father's education this might also negatively affect durations and the overall probability of employment through an increase in reservation wages.

Exploring the macro determinants of transition from school to work
In this section we attempt to gauge some evidence about the determinants of the cross-country dispersion in the fraction of individuals predicted to never transit to the labor market and in the duration from school to work.
To do so, we regress the estimated mean transition durations and the probabilities of never transiting to employment or to stable employment (from Tables 4 and 5 Due to the limited number of observations, we are severely restricted in the number of variables that we are able to include in the model. After some experimentation we have decided to retain only the variables that have typically statistically significant effects on the outcomes of interest (see Table A8). First, we include log per capita GDP as a proxy indicator for country economic development.
As increased supply of labor is likely to affect youth employment at fixed labor demand, we also include in the model the growth rate of the population. One would expect higher population growth to increase both the duration of transition to work and the probability of never transiting.
We also include the national poverty rate, as measured by the fraction of the population living on less than 1.25 US dollars per day. Greater incidence of poverty is likely to be associated with lower reservation wages and hence shorter durations.
Alongside these variables we include two variables capturing labor market policies. First, we include a dummy for the availability of unemployment insurance.
Unemployment insurance schemes might increase durations via their positive effect 18 on the reservation wage, although they might eventually lead to better matches.
Second, in order to control for labor market regulations, we include a measure of minimum wage bite (the ratio between the minimum wage and value added per worker). This should capture higher labor costs that might in turn reduce firms' hiring. 13 All these variables are measured in the most recent year in which they are available (typically between 2012 and 2014). Summary statistics for these variables are reported in Appendix Table A9. These statistics refer to 22 countries out of the 23 for which duration models can be estimated (excluding the Occupied Palestinian Territory, for which no information is available from published statistics).
The table clearly shows that population growth is much higher in SSA than in the rest of the world (although MENA countries also have high population growth) as is the poverty rate (followed by AP). Perhaps surprisingly, the minimum wage bite is the highest in AP followed by SSA countries, although clearly this only typically applies to wage workers in urban areas, i.e., a small share of the labor force. Finally, unemployment insurance is typically more widespread in middle-income countries in EECA, MENA and LAC compared to low-income countries in AP and SSA.
Regression results are reported in the top panel of Table 8. Each column refers to a separate dependent variable and each row to a different regressor. In the model we include dummies for missing values of the included variables. For ease of interpretation we express all variables in terms of their standard deviation. We also weight observations by population weights using GLS.
We start by focusing on average duration to the first employment spell in column (1). 14 Although national economic development, as measured by per capita GDP, appears to speed the rate of transition, the coefficient is not statistically significant at conventional levels.
Row 2 illustrates that a one standard deviation increase in the rate of population growth (around 1, i.e. half of the difference between EECA and SSA) leads to an increase in average duration of around 17 months. The effect of the poverty rate is the 13 The threat posed by minimum wages to the employment of youths, for whom the statutory minimum wage is more likely to be binding than for adults, has long been recognized, as employers in competitive markets are predicted to respond to an increased minimum wage by reducing labor demand. This prediction however has not found broad empirical support (see seminal work by Card and Krueger 1994). 14 Results not reported are very similar if we use median as opposed to mean durations. opposite: a one standard deviation increase the poverty rate (20 percentage points, again not too far from half of the difference between SSA and EECA) leads to a reduction in transition durations of around 17 months.
We next turn to the policy variables. Both an increase in minimum wage bite and the availability of unemployment insurance lead to an increase in the length of transitions, although only the latter displays a statistically significant effect. An increase in the fraction of the population covered by unemployment insurance of one standard deviation (0.40, again roughly the difference between SSA and EECA) leads to a rise in unemployment insurance of around six months.
Column (2) reports results for the average transition duration to the first stable employment spell. Coefficients have typically the same sign as those for the duration to the first employment spell (whether stable or not), although they are also typically larger in magnitude. Overall, it appears that the same factors that explain differentials in transition durations to employment across countries also explain differentials in transition durations to stable employment.
Column (3) focuses on the probability of ever transiting to employment.
Regression coefficients are all statistically insignificant, although mostly of the expected sign. Column (4), that focuses on the probability of attaining stable employment, is perhaps the most informative. It appears in particular that greater incidence of poverty hampers young peoples' probability of finding stable employment (with a one standard deviation increase in the poverty rate leading to a rise in the probability of never attaining employment of 14 percentage points).
Unemployment insurance (UI) acts in the opposite direction, with a one standard deviation increase in the fraction of the population covered by UI leading to a rise in the probability of finding stable employment over one's life of six percentage points.
Results in the bottom panel of Table 9 include in addition region fixed effects. In practice we only exploit the variation across countries in the same region for identification, easing some concerns that the estimates are driven by unobserved differences across countries with different unobservable characteristics. Although results are less precise, they are qualitatively similar to those obtained without the inclusion of region fixed effects in the top panel, lending some credibility to the estimates discussed above.
Overall, results in Table 8 show that demographic and structural economic factors are major determinants of the transition from school to work in low and 20 middle-income countries. While greater pressure on the labor markets induced by higher population growth leads to longer durations in low income countries compared to middle income countries, these effects are largely compensated by much lower reservation wages driven by widespread poverty and lack of unemployment insurance that together speed the transition from school to work. As for MENA, where durations are remarkably high, it appears that the combination of high population growth and relatively generous unemployment insurance are largely responsible for long durations.
Although it appears that unemployment insurance lengthens transition durations, there is also evidence that the support to the unemployed offered by unemployment insurance schemes leads to an overall increase in the probability of ever finding employment and in particular stable employment. Although one has be cautious in interpreting these estimates as causal, this might point to the negative effects of short durations on the probability of finding stable employment, highlighting the trade-off that policy makers face in these countries.

Summary and conclusions
This paper uses ILO STWT-survey data to provide novel systematic evidence on duration and end point of the transition from school to work in a sample of 23 low and middle income countries around the world, and to investigate some of the determinants of the differential lengths of transition across these countries.
Although transition durations are on average not very dissimilar from those in high-income countries, notably Europe, on the order of two years, there is substantial heterogeneity across the countries in the sample. In particular, countries in the Middle East and Northern Africa region display markedly longer transition duration than all other countries, in line with high rates of youth unemployment. Moreover a substantial fraction of youth in the low and middle-income countries under analysison average 10 percent -are expected to never find employment, let aside quality employment, over their life cycle.
Durations as well as the risk of never transiting to employment over one's life cycle are considerably higher among women compared to men, although we remain agnostic on whether these differentials reflect preferences or constraints. There is also a clear positive association between levels of human capital, notably school-leaving 21 age and work prior to leaving school, on labor market success, measured by a greater probability of finding employment and a shorter search spell.
In closing we have assessed the determinants of the durations and end points of the school-to-work transition across the countries analyzed. Consistent with predictions from theory, we find that demographic and structural economic factors are major determinants of the transition from school to work in low and middle-income countries. While higher population growth leads to longer durations in low income countries compared to middle income countries, this effect is more than compensated by widespread poverty and lack of unemployment insurance that together depress reservation wages, speed the transition from school to work and reduce the probability of finding quality employment over one's life cycle, especially in Sub -Saharan Africa and Asia and the Pacific regions.
Although, given the limitation of the data, we are unable to carry out a full welfare analysis, this evidence is suggestive of the circumstance that in low income countries short search durations induced by low reservation wages and lack of social protection lead to poor labor market matches that harm young workers' probability of finding employment, and in particular stable employment, in the long-run, in turn contributing to persistent social and economic exclusion.

Figure A1. Correlation between in-sample and composition-free predictions -Probability of never transiting
Notes. The figure reports the correlation between composition-free estimates of the probability of never transiting to employment and in-sample predictions. A 45-degree line also reported. See text for details.

Figure A2. Estimates of the hazard function: duration to first employment spell
Notes. The figure reports estimated hazard functions based on estimates in Table A4.  Figure A3.