Compulsory Schooling and Early Labor Market Outcomes in a Middle-Income Country

Abstract

The 1997 reform in Turkey which extended compulsory schooling from 5 to 8 years provides an opportunity to estimate the returns to schooling in a middle-income country. The availability of a rich set of early labor market variables also provides an opportunity to assess mechanisms through which returns to schooling occur. I find quite small effects of compulsory schooling on earnings of men but large positive effects on earnings of women who work, without raising their overall low rate of labor force participation. In terms of mechanisms, I find that women who worked moved into higher skill and formal sector jobs, which involved more complicated tasks on average.

Appendices

Appendix 1: Tables

Table 9 The cohorts affected by the law

Table 10 The estimated effect of compulsory schooling reform on college degree

Table 11 Returns to compulsory schooling (1st stage results)

Table 12 Returns to compulsory schooling estimates for log monthly earnings (OLS and IV estimates)

Table 13 Main Tasks and duties of the person in the workplace (ISCO 88)

Table 14 Economic activity of the unit in which the person works (regrouped categories of NACE1.1)

Table 15 The effect of the reform on occupational choice and economic activities - men

Table 16 Tasks performed in occupations - Dictionary of Occupational Titles-1991

Appendix 2: Data

I use Household Labor Force Survey micro data releases from 2004 to 2011. After merging the original data sets with the single age data and year of birth data, I have the following numbers of observations in each year respectively: 472823, 490,008, 497,132, 481,600, 481,149, 503,329, 522,171, and 517,076 (3,965,288 observations in total). After appending data from 8 years, I drop those who are younger than 20 or older than 29. This leaves me with 600,000 observations. Then I drop observations with inconsistent age and year of birth information (I drop if year of birth is less than (survey year-age-one) or year of birth is greater than (survey year-age)). This leaves me with 599,485 observations. 25,477 (4%) observations among the remaining ones do not have year of birth information, and I drop them as well. Thus I am left with 574,008 observations aged 20–29 from survey years 2004–2011.

The regressions that investigate the effect of the reform on educational attainment use this sample. The sample of non-college bound individuals excludes those who attended college or had graduated from college when surveyed. When I estimate the wage returns to schooling, I drop individuals with monthly wages less than 100 Turkish Liras or more than 4000 Turkish Liras. This corresponds to less than 1 percent of total wage earners with wage information. As I mentioned in fourth section, only 5% of the salary workers have missing wage information.

Key Variables

Schooling

The survey does not provide the exact years of education. Instead it provides the information on the highest completed degree. Therefore I categorize individuals into four: Less than middle school, Middle School Graduates, High School Graduates and College Graduates. In the HLF micro data, those who are in categories 0, 1, 2 are considered in “Less than Middle School” category. Although this category may technically involve who do not have any education at all, the share of this group should be ignorable. Vocational middle school and general middle school are considered in the same category in the survey. I follow the survey design and call them Middle School Graduates. Similarly, vocational and general high school graduates are grouped as High School Graduates. Finally, those who have 2-year college degree or a higher degree are grouped as College Graduates.

The completed years of education is calculated through assignment of usual years of education to each degree. I assign 5 years of education to those who have less than middle school degree, 8 years to those with middle school degree, 11 years to those with high school degree and 15 years to the college graduates.

Age and Year of Birth

As mentioned in the text, I obtained additional data on single age and year of birth of individuals from the Turkish Statistical Institute. I observe the age and year of birth information of 574,008 individuals aged 20–29 from survey years 2004–2011.

Employment Status

The surveys from 2004 to 2008 categorize employment status into five: regular, casual, employer, self-employed and unpaid family worker, whereas surveys 2009–2011 combine the first two into the same category of regular or casual workers. To be consistent, I combine the first two into the same category in the early surveys as well. The category of regular or casual workers is used as the sample of salary workers.

Status of Workplace

Only in the 2009–2011 releases I observe the status of workplace. There are three categories: private, public and other. While investigating the effect of the reform on sectors, the binary indicator for public sector takes the value one for those who are in the second category, and zero for those in the first and the third category.

Type of Residence

The HLS considers the settlements with population of 20,000 and less as rural, and settlements with bigger population as urban. I follow this categorization in my analysis.

Marital Status

I categorize “single that is never married” and “living together without getting married” into the category of “single”, and categorize “married”, “married but not living together”, “divorced” and “widowed” into the category of “ever-married”. So I generate a binary indicator for being ever-married.

Computing DOT Task Means for ISCO-88 Categories

Each DOT category has an assigned number for the task requirement of the occupation in relation to data, people and things. My ultimate target is to obtain assessments of these requirements for ISCO occupation categories in the HLF survey. I start by matching more than 12,000 thousand DOT occupation categories to around 800 SOC-2000 occupation categories. Then I calculate the simple average of three assigned values for each SOC category so that I obtain the task requirement of around 800 SOC occupations regarding data, people and things. Then I match SOC occupation categories to 26 ISCO-88 categories in my survey data, and calculate the simple average of the task values once more. As a result, I obtain three values for each ISCO-88 occupation category regarding the task requirements in relation to data, people and things. Occupation categories that require more complicated tasks have lower values, and those that require less complicated tasks have higher values.

Appendix 3: Theory

Card (1999) assumes that individuals maximize the following utility function by choosing the optimal level of schooling S,

$$ U\left(S,y\right)= logy(S)-h(S) $$

(3)

where y (S) denotes the average level of earnings an individual with schooling level S receives, and h(S) is a convex function representing the individual cost of schooling such as forgone earnings, tuition, fees and psychological costs. Including heterogeneity in costs and benefits of schooling yields heterogeneity in optimal schooling levels. Following Card (1999), let the marginal costs and benefits of schooling be characterized by

$$ \frac{y^{\hbox{'}}(S)}{y(S)}={b}_i-{k}_1S\kern1.50em {h}^{\hbox{'}}(S)={r}_i+{k}_2S $$

(4)

where b_i and r_i are random variables with means $\overline{b}$ and $\overline{r}$, ${b}_i=\overline{b}+{\varphi}_i$, ${r}_i=\overline{r}+{\eta}_i$ where b_i corresponds to abilities, r_i corresponds to psychological and monetary costs of schooling. k₁ and k₂ are non-negative constants, representing the concavity of the earning function and convexity of the cost function respectively. The optimal schooling level that solves the problem is

$$ {S}_i^{\ast }=\left({b}_i-{r}_i\right)/k $$

(5)

where k = k₁ + k₂. This shows how individual schooling decisions depend on individual returns and costs and why we expect different returns to schooling at different schooling levels.

In this framework with heterogeneous returns, Card shows that the Ordinary Least Squares (OLS) estimate of the effect of schooling on log wages is different from the average effect because individuals with higher returns acquire more schooling. Moreover, because compulsory schooling laws affect some subgroups more than others, IV estimates do not provide an average return.

If there is variable treatment intensity and treatment effect heterogeneity, then the IV estimate can be interpreted as a weighted average of causal responses to a unit change in treatment (Angrist and Imbens 1995). Suppose that the population can be categorized into groups of individuals with common values for ability and taste terms, and the policy intervention leads to a change ΔS_g in the mean schooling of group g, and β_g denotes the marginal return to schooling for group g in the absence of intervention. Then, following Card’s notation, the IV estimator of the return to schooling has the probability limit

$$ {plimb}_{iv}=\frac{E\left[{\beta}_g{\Delta S}_g\right]}{E\left[{\Delta S}_g\right]} $$

(6)

where expectations are taken with respect to the probability distribution of the population across cells. The IV estimator is consistent with the average marginal returns to schooling if there are identical marginal returns across subgroups (${\beta}_g=\overline{\beta}\Big)$ or change in schooling is homogenous across subgroups (Card 1999). Within this framework, my estimates revealing the LATE for people induced to attain more schooling after the reform.

Remember, before the reform, the average schooling level among women was significantly lower than among men: 55% of women aged 20–29 and 35% of men aged 20–29 did not have a middle school degree before the reform. Since the ratio of women with less than middle school degree was higher than men in the first place, and since the social and economic barriers were more prominent among women, the compulsory schooling reform that mandates 8 years of schooling is expected to lead to bigger changes in schooling in the sample of women. Moreover, assuming that individuals who have higher cost of schooling also have lower returns to schooling, the average returns to schooling among women who did not have a middle school degree and are enforced to finish eighth grade is higher than those among men. Therefore, an estimation that uses the compulsory schooling reform as an instrumental variable is expected to lead to higher estimated effects among women than among men.

To clarify the case, I propose the following example which models the reform as a complete enforcement. There are four subgroups of individuals with common cost and benefit terms ( b_g, η_g) within each subgroup, and log earnings are linear in schooling, i.e. k₁ = 0. Let σ_bη < 0, and b₁ < b₂ < b₃ < b₄ and η₁ > η₂ > η₃ > η₄ , so those with higher marginal costs of schooling also have lower marginal returns to schooling. Men and women from the same group g have the same characteristics except for the mean marginal cost, ${\overline{r}}_w>{\overline{r}}_m.$

Using Eq. (5), the optimal schooling level would be higher in the sample of men as ${\overline{\ r}}_w>{\overline{r}}_m$. For simplicity, suppose the subgroups of women choose 1, 2, 3, and 4 units of schooling respectively, and the subgroups of men choose 2, 3, 4, and 5 units of schooling. Note that women who have 1 unit of schooling have the same innate characteristics except for $\overline{r}$ as men who have 2 units of schooling. Now, the compulsory schooling reform makes it mandatory to finish at least 3 units of schooling. The IV estimate will reflect the marginal returns of the first and second group in the sample of women, but the marginal returns of the first group in the sample of men. By assumption b₁ < b₂, so the LATE in the sample of women will be bigger than the LATE in the sample of men.