Education in transition : regime change impact on income by using regression kink esign

(1)

Education in Transition: Regime Change

Impact on Income by Using Regression Kink

Design

Beatriˇ

c˙e Leiput˙e

University of Amsterdam, Faculty of Economics and Business

MSc Economics, Development track

July 14, 2016

Abstract

We analyse the causal effect of a change in the education on the change in individual income in Lithuania. By using regression kink design, we explore the impact on income for individuals with an endowment of only the old type of education (acquired in socialistic regime in Lithuania before 1990) compared to

individuals that have some new education. We find that this substitution in

education has no statistically significant impact on income. Linear and quadratic model fits are explored on the data used from Survey on Income and Living Conditions (EU-SILC) 2005–2014. In addition, we perform a placebo test for kink assignment and control for macroeconomic changes in economy.

Key words: socialistic education impact; regime change; transition

econ-omy; regression kink design.

Supervisor: Dr. Adam Booij

Second reader: Prof. Dr. Erik Plug Student number: 11089598

(2)

Statement of originality

This document is written by Beatriˇc˙e Leiput˙e who declares to take full re-sponsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creat-ing it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1 Introduction

The period of 1989-1990 symbolize rapid changes in the economic and social systems: the end of the Cold War, the fall of the Berlin Wall and Soviet Union collapse. These events brought many countries into a transition from planned to market economies. The transition period resulted in high declines in output after opening the walls. Simai (2006) calculated that total loss of output and income in the Commonwealth of Independent States (CIS) and Central and Eastern European (CEE) countries was roughly equal to about 3 years of GDP of the former Soviet Union during the first 10 years of the transition. Job security and egalitarian wage distribution which was ensured under socialistic regimes collapsed. Consequently, wage inequality increased and a private sector emerged. With these rapid changes, the perception of skills, ability and, in general, human capital changed. People with an educational endowment acquired under socialism were suddenly exposed to a dynamic labour market system. This analysis reviews long-lasting effects of the socialistic education on people’s labour market performance in Lithuania in 2005–2014. To what extent is their current labour performance affected by the education acquired under socialism compared to individuals that followed new education?

There is no clear consensus on when these economies can be considered as completed with their transition. One natural suggestion is clear: when the old or restructured enterprises are completely taken over by the new ones that are more productive than their descendants, the transition can be considered as over (World Bank, 2002). The differences in productivity should be no longer explained by historical categories of old, new and restructured. In this paper, we take the human capital approach. Brunello et al. (2010) point out that since the first generations educated under the new system most likely entered the labour market during the recession, this comparison must be done only in late 2000s when alternative transitional effects fade away. In addition, large part of the current labour forces of the countries that transitioned around 1990 are endowed with old type of education. Consider Lithuania, where the workers of age 45–60 make up almost 40% of the total labour force in 2015 (Statistics Lithuania, 2016). This leads us to the motivation for this paper, which is implied by the facts stated above.

In this paper we use the EU–SILC (Statistics on Income and Living Con-ditions) survey data collected for 2005–2014 in Lithuania. We inspect the causal effect of a change in the educational system from only years of old (ac-quired before 1990) to new (ac(ac-quired in 1990 and later) education on income. The null hypothesis is that this substitution in education has no causal effect on individual’s income. An alternative hypothesis is that there is a causal

(5)

effect on income caused by this substitution in education. We predict that a detrimental impact of socialistic education can be seen in individual wages in 2005–2014. This may be supported by the fact that individuals with an endowment of only old type of education lack skills that are useful in the new market economy.

We use a Regression Kink Design (RKD) which tries to see if the kink (change in slope) of old years of education is accompanied by a kink in the average annual gross income earned at the time of the survey. By comparing observations laying just below and above the cut-off value we assume that the only difference in individuals is the treatment (new type of education) assignment. This design is closely related to the Regression Discontinuity design, where the effect is reflected in levels rather than slopes. We closely follow Card et al. (2012), who provide a theoretical RKD background and Nielsen et al. (2010), whom were the first ones to identify it as a separate design.

Little research was carried out on the human capital in the three Baltic States mainly due to the availability of data. To the extent of the literature review that was done, there is no extended research which explores the causal effect of change in education on income in modern Lithuania. Choice of design contributes to the literature on transitional economies, which mostly uses investment in human capital models.

We find that the substitution from only old to new years of education is statistically significant for 41–49 year old individuals. However, this substi-tution has no accompanying change in slope of income variable and, hence, no causal effect. Our linear model estimates a statistically insignificant local average treatment effect (LATE) of -22%, whereas the quadratic fit yields a coefficient of -16%. Thus, we accept the null hypothesis that, on average, there is no impact of change in education on labour performance. In addition, we acknowledge that the second stage estimation is most likely underpow-ered. We perform a couple of robustness tests: 1) we shift the cut-off value from 0 to ±1-3 years and find that our main model estimate is in the range of [-0.481, 0.248]. 2) By adding macroeconomic controls such as GDP per capita and harmonized consumer price index (HICP) we find that the linear model coefficient of interest decreases to -27%. Overall, our results may imply that adult individuals of age 41-49 adapt to rapid changes in economy and labour market and do not face any consequences due to underlying differences in their educational endowment.

Most of the previous literature relies on investigating time trends in re-turns to education before and during the transition and uses versions of the Mincer (1974) equation. Campos et al. (2007) find that returns to secondary and post-secondary education has increased in the period of 1986–2004 in

(6)

Hungary but decreased for the vocational education. Age cohort compar-ison yields higher returns for older age cohorts, whom acquired education before the fall of the Berlin Wall. These findings support the hypothesis that pre-transitional education is not outdated for the market economy. Similar analysis was carried by Munich et al. (2005), focusing on the educational re-turns in 1991–1996 Czech Republic. Authors find that rere-turns to education were lower in pre-socialist era. In contrast, Orlowski et al. (2009) find that returns to education are lower in Eastern Germany compared to the Western counterpart twenty years after the merger. The authors argue that educa-tional skills formed to serve planned economy are of little value in the new economy. Perhaps the most comprehensive recent research is carried out by Brunello et al. (2010). These authors use data for 22 economies in Eastern and Western Europe. From the analysis they conclude that junior workers earn higher returns than their senior counterparts in East.

The paper is organized as follows. The next section provides a short overview of an existing empirical literature that focuses on human capital changes in transitioned economies. In Section 3 the context of the setting and theoretical channels of transition are described, followed by an empirical design in Section 4. Data and Results are presented in Sections 5 and 6. Robustness tests are performed in Section 7.

2 Literature review

This section gives an overview of past research done on human capital in transitioned economies. We focus on the literature that analyses the impact of education on labour market performance. There is an extensive literature covering the topic, thus, given the size of existing studies, this review may be selective. This section is divided into two parts: first one dedicated to the literature investigating the changes in returns to education before, during and after the transition period and the second one focuses on age cohort effects.

2.1 Before, during and after the transition

One difficulty in investigating time trends is that most of the household surveys in the transitioned economies were started to conduct only after 1990. Data which is available for the period before 1990 is most likely different in terms of methodologies, can be non-representative, or deceiving (Simai, 2006). Thus, empirical evidence on returns to education before the transition is scarce, mainly accompanied by theoretical articles. In addition, most of the

(7)

articles analyse situations in countries that are rich in data, such as Hungary, Czech Republic, Poland and Germany.

Most of the empirical literature base their findings on two alternative hypotheses. The first one predicts that returns to education will increase after the transition. This hypothesis is based on the fact that in centrally planned economies wages were set by grids, which implicitly under-evaluated specific educational levels and were based on labour market experience. The wage distribution was set to reflect egalitarianism and the dispersion within an industry was modest, leading to a very low skill premium (Munich et al. 2005). The most common method used is a standard and/or extended Mincer (1974) equation. Endogeneity caused by ability bias is noticed mainly by the articles written after 2000, with some of the authors acknowledging the problem without taking measures to solve it.

An extended analysis on returns to education, covering around 4 million wage earners in Hungary was carried by Campos and Jolliffe (2002, 2007). Their analysis covers periods before the transition, just after, and later on in 2001 and 2004. By using a standard and extended Mincer equation authors acknowledge that they have no possible instrumental variable that could correct for ability bias. They find that returns to education in Hungary increased from 6% in 1986 to 10.7% in 2004. However, a decline is seen in returns to vocational and primary education. This is interesting because large part of educational finances was invested in vocational education due to focus on heavy industry before the transition (The World Bank, 2000). This may imply that wage earners with vocational education did not adjust to the new labour market. Authors of the article conclude that labour market experience acquired before the reforms is not valuable. There has not been a lot of research done in the area of labour experience, acquired in planned economy, and it’s validity in market economy, marking the contribution of Campos and Jolliffe.

An improvement on the methodology is carried out by Munich et al. (2005) who focus on men in the Czech Republic before (1948–1989) and after (1991–1996) the transition. The authors find that returns to education rose from 2.7% in 1989 to 5.8% in 1996. In contrast to Campos and Jolliffe (2002, 2007), Munich finds that returns to vocational educational increase between 1989 and 1996. An interesting finding is that annual growth of returns to education does not statistically significantly differ between public, private, and newly established firms. In addition, the most rapid growth of returns to vocational education as well as higher education can be seen within the public sector. However, a university degree is the least valuable at the state sector compared to the private and newly established firms. This may imply slow adjustment of the wage grids and skill premium in the public

(8)

sector. Another aspect of the study is that authors decompose the variance of an individual wage related to their wage premium. Commonly used cross-sectional data does not allow this kind of analysis, however, panel data can be decomposed to observed and unobserved determinants of wages. Even though this estimation method excludes individuals who worked only in one of the two (before and after) periods, it provides evidence on the time change in wage determinants.

Evidence on decreasing returns over time can be found in analyses carried by Flanagan (1998) and Chase (1998). The authors hypothesize that returns to education decrease after the rapid change in economy because skills ac-quired under the old regime are not suited for market economy. Munich et al. (2005) point out that contradicting results found by the above mentioned analyses could be due to the fact that those studies chose to analyse early periods of transition (1991 and 1993 respectively) when newly introduced firms were not yet eminent. Thus, their results provide a somewhat flatter wage-experience profile. This is as well supported by Munich et al.’s (2005) empirical findings who use later time periods.

2.2 Cohort effects

An interesting focus point for the cohort analysis is the fact that younger individuals in transitioned societies acquired their education under different circumstances whereas older cohorts attended centrally provided education. Exact changes in education and labour market will be discussed into details in Section 3.

Brunello et al. (2010) inspect returns to education and experience in 22 economies in the late 2000s, when most of these countries became mem-bers of the European Union. The chosen treatment group consists of wage earners whom were of age 25–39 in 1989, have received their education, and currently work in East Europe. Two control groups were chosen. The first one includes people of the same age as the treatment group but studied and reside in selected Western European countries. The second control group consists of East Europeans that were of age 6–13 in 1989, meaning that they were partly educated before and after the change in the system. Method-ologically Brunello et al. (2010) consider that men and women of different ages are imperfect substitutes of labour. In order to address ability bias, the authors difference out gender-invariant components by comparing males and females by age by country by time. In their final model, the authors include country by year–and–country dummies to control for time invariant and variant unobservables. They find that younger males in East European countries receive higher returns to their education than the older workers,

(9)

implying that education acquired under the old regime is punished in the new market system. However, this evidence does not imply an increasing school quality but rather an increased demand for skilled labour.

In his article Jurajda (2005) focuses on returns to different educational levels in the Czech Republic in 2002. The returns to years of schooling and different educational levels vary very little over different age groups and is not significantly different when compared to each other. In contrast to previously described evidence found by Brunello et al. (2010), Jurajda (2005) find that older wage earners receive similar returns to education as their younger counterparts. The most plausible explanation for this difference may be the fact that Brunello et al. (2010) carefully consider and control for ability bias, country and time effects, whereas Jurajda (2005) does not take these components into account.

Orlowski et al. (2009) point out that returns to experience in transitioned economies lag behind whereas returns to education signal individual ability which immediately catches up after the new reforms. By comparing East and West Germany wage structures in 2002–2006 and controlling for possible biases, authors find that wage-experience profiles are significantly different between East and West. The authors emphasize that experience may be correlated with an error term because the longer an individual is prepared to look for a job, the better his chances are to find a match. Following previous attempts to control for this, authors take up an approach which provides unbiased OLS coefficients. In sum, even after 20 years of unification, returns to experience and tenure are different in East and West. Another article, written by Fuchs and Marsella (2013), uses this unique setting of East and West Germany and accommodates a difference-in-difference design. The authors explore birth cohorts of 1971–1977 from both East (treated) and West (non-treated). They identify that within the same age cohort, one year older individuals hold one more year of old type education. The main conclusion is that an additional year of old type education decreases the probability of college degree obtainment. In addition, there is a significant detrimental effect on wages and managerial or professional job obtainment for older males.

Previously mentioned articles have also carried cohort analyses as part of their papers. In their log-wage equation, Munich et al. (2005) add separate regressors for years of new and old education and find that years of new ed-ucation yield significantly higher returns than years of old. By allowing the coefficient on years of education to vary with age, authors conclude that re-turns for younger men are significantly different than for older wage-earners. This result is mainly driven by vocational education. Campos and Jolliffe (2007) find that the difference in returns to schooling between the people

(10)

that were younger than 20 and those that were older in 1986 was as high as 61%. Their analysis shows that this gap narrowed down to 17% in 1992. However, after that returns for the young earners stagnated and returns for the older counterparts started increasing until 1998. This is in line with the hypothesis of declining school quality after the rapid changes in the system, but may be also influenced by other channels such as more able people within the younger cohort, increasing share of educated people in the older cohort, and a change in the returns to experience.

This review leads us to a conclusion that there is mixed evidence on the effect of pre-transitional education varying in terms of age, education, sector and time period chosen. Most of the articles investigating time trends find that returns to education increased when economies shift from planned to market system. Only a small amount of research analyses current income by age cohorts of people, who perform in the same labour market but have endowment of skills acquired under different regimes. Another contribution to the literature is the use of data from 2005–2014 since the existing research relies on early-transition data. In addition, we deviate from the analyses by using a quasi-experimental design, which is used less in the field of transi-tioned economies. Finally, the design chosen contributes to literature that tackles endogeneity issues.

3 Context: regime change

Lithuania declared its Restoration of Independence in 1990. Since then, there have been many drastic reforms in educational system and labour market that gives us an idea on possible differences between the periods before and after the shift. We will briefly describe the main changes in education and labour market. Later on, we will continue with the theoretical channels of the economic transition.

3.1 Key labour market reforms

To understand the key labour market and educational reforms, it is important to enlist the main aspects of the economic transition in the country. Most of the reforms brought, went hand in hand with changes in social, economic, and political life. Cumulative GDP loss was around 63% in 1989–1993 (UNECE, 1998). Annual hyperinflation in Lithuania jumped from 383% in 1991 to around 1163% in 1992 (Kuodis, 2008). From 1995 onwards GDP started to have a positive annual growth. An important estimate is that at the end of 1980s almost one quarter of the whole Soviet Union GDP was spent on heavy industry and military capital (Eslund, 2001). A big part of this

(11)

capital became useless after 1990 and was reorganised. These and other signals required immediate reforms that would stabilise the market and lay market economy foundations. Big reforms such as the privatisation process, liberalization of prices and trade, agriculture decollectivization, and fiscal and monetary policies required parallel changes in labour market and educational system that would ensure full transition.

The main reforms in the labour market can be reflected by investigating differences between the Code of Labour Laws of Lithuanian Soviet Socialist Republic (LSSR CLL) adopted in 1972 and the new Labour Contract Law adopted in 1991, updated in 1994 and supplemented by the Labour Code in 2002. The main reason for the establishment of the Labour Code in 2002 was to prepare for the membership at the European Union and ratify International Labour Organization (ILO) and other EU laws and conventions. The main differences between these documents can be seen in Table 1.

(12)

Table 1: Key labour market changes by aspect before and after 1990.

Aspect Before After

Employment

contract Verbal or written Written

Employment termination

A list of 8 specific basis reasons

List of wide types of basis reasons

Collective agreement

Trade union committee, integral part of the Com-munist party, is a collec-tive agreement party

Trade union, authorized by workers conference, is a collective agreement party

Wages

Wage determined by the state. State has the power to fix tariffs, schemes and wage grids

Wage and qualification re-quirements are set by the employer. The govern-ment sets a minimum hourly wage

Annual holidays Min 15 calendar days

Min 28 calendar days with other options based on disabilities, high-risk jobs and other

Other leave Pregnancy and delivery leave

Pregnancy,

maternity/paternity and educational leave

Damages

Pecuniary damage respon-sibility applicable to work-ers only

Hedonic damage was added, applicable for employer, worker and the state

(13)

3.2 Key educational system reforms

Education was uniformly designed and centrally managed in Lithuanian So-viet Socialist Republic (LSSR) since 1940. The main focus was given to mem-orisable, factual knowledge, which was highly useful for the well-predictive planned economy. Education was propped by vocational education and tech-nical studies in order to serve the heavy industry. Flanagan (1998) points out that “. . . overinvestments in vocational-apprenticeship training and un-derinvestment in university education appear to be the major distortions in human capital formation under central planning”. In addition, there was a known lack of focus on humanitarian and social science.

The main educational reforms were set by a General Concept of Educa-tion in Lithuania in 1992. It was dedicated to shape dynamic self-sufficient society, which would be able to adapt their knowledge and skills to fast changes in the market (OECD, 2002). More precisely, big reforms in fields such as compulsory education, accessibility and national curriculum were made. Changes in curriculum reflected elimination of ideologically oriented elements and new teaching materials. New standards were introduced such as profiling during the last grades of high school and grade 12 examinations. State enterprises, for which vocational students were trained, ceased to ex-ist, thus many educational institutions were decentralised and deregulated. An important step was to conform degrees as defined by the Bologna Joint Declaration. As for the higher education, new legal framework providing university autonomy and new research infrastructure was created. Quality assurance mechanisms, lifelong learning and social partnership – these and many more concepts were introduced. The summary of the main educational reforms is given in the Table 2.

(14)

Table 2: Key educational system changes by aspect before and after 1990

Aspect Before After

Number of public/private higher education institu-tions 13/0 universities 66/0 colleges 108 vocational institutions (1989) 14/12 universities 13/8 colleges (2016) Investment in vocational education (% of total in-vestment in education)

10% (1989) 2.6% (1995), 4% (2014)

Compulsory education

(number of years) 9 10

HE tuition fees Free and

scholarships offered

Voucher system (since 2009). Tuition differs by study field

HE enrollment HE enrollment increased by 45% from 1995 to 1999

Share of students contin-uing studies at secondary school after completing basic education 55.6% (1992) 77.3% (2014) Expenditure on education (% of GDP) 2% of total national budget expenditure (1988–1990) 5.9% of GDP (2011)

Sources: OECD (2002), Statistics Lithuania (2016).

3.3 Theoretical channels of an economic transition

We will close this section with economic theories behind the economic tran-sition and its effects. In addition, we will discuss how our hypotheses fit in the theory and what outcomes of the analysis we expect.

Orazem and Vodopivec (1997) describe three channels that influence wage structure and labour mobility during transition. Authors emphasize that despite of their simultaneous effect, they all work towards increase in em-ployment and wages for skilled labour. The first channel describes forces that are associated with corrections of distortions caused by constraints on the labour market before the transition. Svejnar (1996) shows that employ-ment decisions of firms become responsive to wages. The effect on unskilled labour is not clear but changes in skilled employment are due to the shift of relative demand towards skilled labour. Another channel exploits changes in demand for goods and services. Due to a switched emphasis from

(15)

manufac-turing (low skilled) to retail and service sectors (high skilled), it is expected that demand for labour in low skilled sectors will decrease. In addition, the main short-term shift in demand is associated with the collapse of traditional trade linkages, which as well focused on low-skill intensive goods.

As Orazem and Vodopivec (1997) describe, the third channel is associated with a disequilibrium and uncertainty caused by transition itself. Authors argue that socialistic educational plan did not focus on entrepreneurship skills, which would be in short supply during and after the transition. Due to this fact, individuals with desired skills would be rewarded with an in-crease in relative employment. In addition, if this skill is complementary – returns to entrepreneurship should rise. Lamo et al. (2011) add that spe-cialized education acquired under socialism reduce worker’s mobility and his ability to adapt to changes in economy. Authors point out that the speed of labour market adjustment, unemployment duration spells, and likelihood of market exit depend on the skills accumulated. Wasmer (2006) specifies that economies which are rich in specific skills in steady-states, face high transitional costs during the shift.

Oreoupoulos (2008) investigates long-lasting effects of adverse initial labour market conditions on the earnings of college graduates. These ef-fects can be especially harsh on lower ability graduates or those attended lower quality institutions. Aspects such as equality and availability of jobs, wage adjustments, and human capital accumulation tend to decrease during recessions. Authors provide evidence that wages of new labour market en-trants respond more to shocks in the economy than those already employed. They find that average ability graduates recover and catch up in ten years, while the effects in terms of wages are longer lasting for low ability graduates. Furthermore, Oreoupoulos (2008) provides information on another class of models that links career progression to human capital accumulation at firm or industry level. One model assumption is that graduates entering the market during a recession face higher job mobility which results in less firm or industry specific skills acquired. Thus, recovery within firms can be explained when individuals with a concave learning profile catch up with the lucky ones. A similar pattern can be provided by the models of long-term wage contracting with renegotiation. Recovery occurs when the wage is renegotiated based on higher opportunity costs. Long-lasting effects can occur if there is no renegotiation or if it is not perfect.

Based on the model implications described above, we formulate a null hypothesis of no causal impact with an alternative two-sided hypothesis of a causal effect of change in education on change in income. We predict that causal effect of socialistic education can be seen in individual wages in 2005–2014. This may be supported by the fact that individuals with

(16)

an endowment of old type education lack skills that are needed in the new market economy. However, this effect may be different for skilled labour, that may have experienced an increase in demand after the transition. While the effect on unskilled is not clear, we cannot differentiate between types of labour in our analysis. In contrast, able individuals may have accumulated labour experience that compensates for the difference in their educational endowment.

4 Design

The aim of this work is to estimate the substitution of years from old to new education effect on the change in individual income. We set up a model which accommodates the identification strategy for a Regression Kink Design.

4.1 Regression Kink Design

In this section we introduce a general idea behind the Regression Kink Design (RKD). The name of RKD was first introduced by Nielsen et al. (2010), though previously used by other authors. Card et al. (2012) provide an elaborate theoretical RKD background. Hereinafter, we will reference to these and other articles.

To start with, RKD is an extension of a quasi-experimental Regression Discontinuity (RD) design, which isolates a causal effect of an intervention. In case of RD, there is a visible cut-off value in intervention assignment. Comparison of observations which lie just below and above the threshold yields an unbiased estimator. This design serves cases in which randomiza-tion is impossible and intervenrandomiza-tion depends on the forcing variable such as test scores, certain income level, class size and others. Observations just be-low and above the cut-off value are assumed to hold similar characteristics and are only different by eligibility for the treatment. RD can be classified into so called Sharp and Fuzzy RD. The first one describes treatment as a deterministic and discontinuous function of covariate x, whereas Fuzzy RD stands for discontinuities (jump) in the probability of treatment conditional on x. In case of Fuzzy RD, this jump becomes an instrumental variable for the treatment. In case of homogeneous effects, both designs give an unbiased estimator if there is no discontinuous jump in unobservables at the cut-off value. In case of heterogeneous effects, estimator represents a local aver-age treatment effect (LATE). In both cases, design varies depending on the functional specifications on both sides of the discontinuity.

The main difference between RKD and designs described above is the fact that instead of a change in levels, there is a change in the slope of the

(17)

outcome function, where the variable of interest is continuous and most of the time depends on another forcing (running) variable. This can be illus-trated by the changes in policy, reforms and regimes which are determined by endogenous forcing variable. An issue here is to find a strong instrument, which would be relevant and satisfy the exclusion restriction. As Card et al. (2012) emphasize, it is almost impossible to find individual characteristics that would determine the level of a certain policy variable. However, in most cases, we can observe a kink in the policy variable formula without being plausibly unrelated to other things. This allows to isolate the causal effect even when a suitable instrument cannot be found. The main idea is to look for a kink in the outcome variable and relate it to the change (kink) in the policy. In example, Nielsen et al. (2010) describe the case when the student aid depends on parental income, which has a causal effect on college enrol-ment. Card et al. (2012) explore the effect of unemployment benefits, where size depends on previous earnings, and on the unemployment duration.

Following Nielsen et al. (2010), consider the model

Y = τ B + g(V ) + ε, (1)

where Y is the outcome variable, B = b(V ) is the continuous function of forcing variable V and v = v0 is a kink point. In order to identify the causal

effect of B on Y , the identification strategy is that if there is a kink between

B and V at v = v0 then a kink should be expected between Y and V at

the same point v0. We can define the main RKD estimator as the change in

slope of the outcome variable Y at the kink point divided by the change in slope of the policy variable B:

τRK =

limv↓v0dE(Y |V = v)/dv − limv↑ v0dE(Y |V = v)/dv

limv↓v0db(v)/dv − limv↑v0db(v)/dv

. (2)

The denominator of this expression is analogous to the first stage expression in Fuzzy RD.

Card et al. (2012) set out necessary and sufficient conditions for a valid RK design, pointing out one of the key conditions (proposition 3): conditional expectations of any baseline covariates have to be continuously differentiable at the kink point. This means that there is no significant change in the slope of covariate at the kink point, which may alternatively affect the outcome. Thus, the kink in the outcome is only explained by the kink in the policy variable. Nielsen et al. (2010) show that then g(·) and E(ε|V = v) have derivatives that are continuous in v at the kink point. By satisfying this condition, the problem of individuals bunching around the kink point is not binding anymore.

(18)

4.2 Econometric specification

To start with, if years of old education have no impact on income, we would not expect to see any discontinuous changes in the slope of income at the kink point. If education is outdated and people do not adjust to the new market, we expect to see discontinuous changes in the slope of income. In this case, we can estimate by how much the slope of the income changes at kink point.

Define a control group comprised of individuals whose total years of edu-cation can be classified as old eduedu-cation only, acquired before 1990. Define a treatment group, which acquired their education during or after 1990, mean-ing that these individuals hold some years of new education. We assume that the bunching problem is not relevant since individuals cannot choose their date of birth as well as predict the change in regime. Define the assignment formula for type of education:

EDU_inew =              0, if gradi < 1990 EDUtotal

i , if gradi− EDUitotal≥ 1990

gradi− 1990, if gradi ≥ 1990

and gradi− EDUitotal< 1990

(3) EDU_iold =             

0, if gradi− EDUitotal ≥ 1990

EDUtotal

i , if gradi < 1990

EDUtotal

i − EDUinew, if gradi ≥ 1990

and gradi− EDUitotal < 1990 (4)

where gradi is the year of graduation, EDUitotal, EDUioldand EDUinewstands for number of years of total, old and new education, i represents individual observation.

For each age profile we can define a cut-off value v0 expressed as an exact

year value, where the cut-off in average EDUold_{can be seen. Our final setting} can be described as a Fuzzy RK design using two-stage least squares model. We treat the change in education slope as unknown and estimate it by the 1st stage regression (5). The final RK estimator β1 is estimated by the second

(19)

EDU_iold = π0+ π1(zit· yeari) + π2yearit+ (5) + N =49 X i=41 γ3iDi+ N =49 X i=41 γ4iDi · yeari+ γ0Zi+ η2it,

ln Yit = β0+ β1EDU\iold+ β2yeari (6) + N =49 X i=41 β3iDi+ N =49 X i=41 β4iDi· yeari+ β0Zi+ η1it,

where ln Yit measure individual’s i annual gross income at year t. EDUiold is the endogenous variable of years of old education, acquired before 1990 by an individual i and is constant over time. zit· yeari is an instrumental variable equal to a value in range of [2005, 2014] if year ≥ v0 (above the cut-off)

given age as indicated by z, otherwise zero. The running continuous variable

year is interacted with age dummies in order to define different cut-off values

for each age profile. 41–49 age range dummies are as well included. Zi is a vector of exogenous individual controls such as gender and urban/rural.

The coefficient of interest β1 captures the effect of interest. It can be

interpreted as the change in slope (kink) of ln Yit before and after the cut-off year value divided by the estimated kink in EDUold

i . In addition, E(η1it|zit·

yeari) = 0.

5 Data

5.1 Survey design

The data are drawn from the EU Statistics on Income and Living Conditions (EU-SILC) randomized household survey, conducted by the Department of Statistics in Lithuania. The survey for Lithuania was started in 2005 and the last year available is 2014. This survey is based on a rotational principle, where every year one quarter of a previous sample is dropped and new rep-resentative subset added. Households and selected individuals are followed for 4 years at most (see Figure 7, Appendix B). A detailed methodological explanation, which slightly varies across the countries, is available on Euro-stat website (see Reference [7]). Each annual survey contains two datasets: household and individual level. We will focus on individual level data since our interest is to explore income levels based on changes in education. The annual individual level dataset for Lithuania contains information on ap-proximately 11000 selected individuals. Data contains variables on personal

(20)

characteristics, health conditions, educational status and annual income lev-els. By combining annual records, we design a long type unbalanced panel dataset where individual ID represents an entity and year of the record rep-resents time variable.

EU-SILC survey holds detailed information on individual income. In order to construct our final dependent variable we have to sum several com-ponents of income that are recorded separately. Income variable of interest Y components in gross amounts: annual wage; non-cash income (company car and other monetized benefits received from employer); annual benefits from self-employment; pension from individual plans; social benefits such as unemployment, old age, survivors, sickness and educational benefits; em-ployer‘s social insurance contribution. An advantage of this survey is that information for income levels is primarily achieved through national social insurance, tax and other authorities. In case of missing data the record will reflect a self-reported value. However, in case of our dataset, a distinction between self-reported and imputed values is impossible. In addition, outlier values are double-checked by comparing them to national averages and after detailed investigations outlier values may be treated as missing. Missing val-ues are double-checked at the Statistics Department and imputed by either using previous year inputs (if available) and (or) national laws for income components.

5.2 Limitations

There are several limitations of the dataset. First of all, reported incomes do not take into account possible differences in tenures of contracts. More suitable variable of interest would be hourly wage, which is not available. The second shortcoming is the educational experience which is reported as the highest level of educational level (standardized by ISCED–97 and ISCED– 11) acquired. Since we are interested in the total number of years taken to study, we will have to rely on implied standards for each level rather than average number of years of schooling. As well, we create two variables to separate number of years of old (before 1990) and new education which may not take into account the fact that person may have had gaps in his educational path. In addition, the survey does not provide information on graduation location or institution, which would help sort individuals by local and foreign education. However, since we are interested in the effect of old to new education, it is known that individuals from socialistic countries were less likely to study abroad and self-organised individual movements were not available. Individuals directed abroad were most likely to study within the Republics of Soviet Union (known as “propiska” rule). As well, we see

(21)

new education as more general term of education formed to serve market economy rather than specific educational system of Lithuania. We argue that education acquired in market economy abroad may lead to similar effect as the new education system in Lithuania.

In case of panel data with rotational subsample attrition is not a concern due to several reasons. First to mention is that the variable of education represents number of years of total/old/new education and is fixed per in-dividual per time since we eliminate those actively in education from our sample. Another reason is that income variable reflects annual observations of a different cross-section each survey year, with a number of individuals that have been in the sample before. However, each of the quarter of the whole sample is representative itself, thus we have an advantage of analysing all four quarters (full cross-section sample) each year.

5.3 Summary statistics

In order to form a sample of active labour force, we have selected individuals that: i) do not study ii) are older than 16 iii) graduated younger than 30 iv) are not permanently disabled v) does not participate in a military service vi) are not retired. It is important to elaborate more on these conditions. The first one is applied due to the fact that those active in education may not have permanent income which would reflect their educational level and in case of working, individuals most likely have a part-time job to support their study related expenses. The second condition is applied to eliminate new-borns and children from the households. The third condition can be debatable, however, it incorporates the fact that even at older age there are people who acquire education under the new system. In order to investigate the system effect on individuals who followed the most common educational path, we eliminate individuals that were older than 30 when graduated. Arguably, these individuals exist in both regimes and are not affected by the type of system. More precisely, 8% of the observations are eliminated based on this condition. Permanently disabled individuals are dropped from the sample due to restricted employment opportunities. By applying all six prerequisites we scale down our sample size from initial 124088 to 50085 observations.

Table 3 represents descriptives of the full sample (after applying condi-tions described above) and the selected sample (age 41–49) statistics. We include statistics for individuals from the selected sample that are employed – have a positive income input and we define this as our final sample. More details on the final selected sample can be found in Section 6.1. From the Table 3 below we can see that total average years of education is approx-imately 14 years for both full and final sample. The main difference is in

(22)

average years of old education since we selected age cohort where the switch between old and new education can be seen. Annual average gross income varies between 6706–7048 (EUR) for both final and selected sample. Individ-uals of age [41, 49] that have positive income yield 13669 observations which is 27% of full sample and 11% of initial survey sample.

(23)

T able 3: Summary statistics Statistics Sample size A: Characteristics Range mean s.d. y ear F ull sample, N Age 41-49, N Age 41-49 emplo y ed, N F ul l sample (N=50085) 2005 4957 1547 1449 Age [16, 83] 44.603 11.474 2006 4832 1579 1505 Male { 0, 1 } 0.506 0.499 2007 5138 1602 1557 Urban { 0, 1 } 0.699 0.458 2008 4891 1494 1453 A ge 41-49 (N=13669) 2009 4487 1355 1337 Age [41, 49] 45.271 2.555 2010 5553 1577 1488 Male { 0, 1 } 0.467 0.499 2011 5215 1428 1317 Urban { 0, 1 } 0.696 0.459 2012 5298 1402 1313 2013 4872 1247 1177 B: Education 2014 4842 1135 1073 F ul l sample (N=50085) T otal y ears of education [4, 22] 13.846 2.433 Y ears of old education [0, 22] 11.457 4.957 Y ears of new education [0, 22] 2.389 4.709 A ge 41-49 (N=13669) T otal y ears of education [4, 22] 14.001 2.059 Y ears of old education [2, 22] 13.834 2.048 Y ears of new education [0, 12] 0.167 0.835 C: Income F ul l sample Ann ual gross inc ome (EUR) (N=50085) [0, 103815] 6604.265 6274.948 Ann ual gross inc ome (EUR), emplo y ed (N=47240) [3.07, 103815] 7002.003 6241.902 A ge 41-49 Ann ual gross inc ome (EUR), (N=14366) [0, 99439] 6706.087 6295.069 Ann ual gross inc ome (EUR), emplo y ed (N=13669) [15.613, 99439] 7048.039 6264.061 Num b er of individuals 50085 14366 13669 Note: Panel A summarizes individual char acteristics for both ful l and sele cte d final sample. Panel B denotes educ at ional endowment and Panel C summarizes inc ome variables use d. On the right side of the T able, sample size for each survey ye ar is describ ed.

(24)

6 Empirical results

6.1 Graphical analysis

The regression kink design can be visually demonstrated by plotting means of years of old education and the outcome income variable over year in bins, which denote distance to the kink. The size of bin, in this case, is 1 year. The goal is to look for potential kinks in education that correspond and may explain kinks in the outcome.

Average years of old education, pursued before 1990 vary by age. As seen in the Figure1, individuals younger than 20 do not have any experience in old education. We argue that at younger age income most likely reflects the start of the career and is tend to be volatile. Years of old education start increasing after age 20 up until 40. Records for individuals of age 31–39 might as well represent permanent income, however they pursued a mix of education, having attended institutions during both old and new regimes. At around age 40–45 we can see a change in the slope, which defines the change between old and new education. Educational endowment is volatile at an older age, possibly due to different economic, social and historical circumstances. We select age cohort 41–49 as our scope of analysis. More precisely, at given age, we observe a year during our data window [2005, 2014] when individuals start having years of new education. The main selection yields individuals of age [41, 49] with cut-off values at [2005, 2014] respectively. We identify that one year older individuals experience a kink one year later in their average years of old education.

We plot average years of old education and average annual gross income for previously selected age cohort. We normalize time so that at each given age in [41, 49], the cut-off value represents 0 and decreases/increases with year – represents distance to the threshold. For example, for age 45 the cut-off value is at year 2009, thus all values, given age and year, are coded as 0. Individuals that are 45 year old and were surveyed in 2008 will be coded with -1, 2010 – 1, 2011 – 2 and so on. From the Figure 2we can see, that on average, the last observation where individuals hold only old years of education (equal to total) is directly at the kink point 0 and decreases as distance increase, meaning that individuals with some years of new education enter each age category in [41, 49]. Due to the increasing trend, the kink point is less visible for the income variable (Figure3). Plotted figures provide us some visual evidence for the causal impact of substitution in education on income, however we need to inspect this statistically by estimating a model described in Section 4.2.

(25)

Figure 1: Average years of old education by age, 2005–2014.

Note: X-axis represents normalized time values for data window [2005,

2014]. Each observation is coded so that the kink (year) is at 0 for given age.

(26)

Figure 3: Average annual log gross income (EUR) for age cohorts 41–49.

6.2 Results

Table 4 contains the main empirical results. Firstly, as we do not observe wage offers for individuals who do not work, we run into a possible sam-ple selection problem. To test this we run a probit model on our selected regressors, which defines a probability of working. The dependent variable contains a dummy which is equal to 1 if individual has positive income and 0 otherwise. The estimated model results can be seen in Column 1. We scale the magnitude of coefficients by 100 without changing the interpretation of their significance or sign. The coefficient on kink variable is equal to -0.003 and is statistically insignificant, which means that there is no difference in individual decision of working for both individuals before and after the cut off. We conclude that we do not run into a sample selection problem in this case and continue with the main model estimation.

The estimated first stage effects are given in Column 2. An estimate of 0.01 indicates that there is a statistically significant kink point in old educa-tion for individuals aged 41–49. However, the coefficient has an unexpected positive sign. Based on the Figure2described above, we would expect to see a negative change in the slope before and after the kink point. By the rule of thumb, first stage F − stat = 37.61 > 10, which implies strong instruments.

(27)

By adding control variables, the estimation remains the same and is statis-tically significant (Column 4). It increases model’s explanatory power from 5% to 10% and F − stat = 80.50. The model fit with controls can be seen in Figure4, described as a linear fit. These results yield that the visible kink in years of old education is statistically significant at the chosen cut-off value and implies that the predicted substitution from only old type of education to some new education exists. This supports our design and allows us to inspect whether this significant substitution has an impact on the outcome variable chosen.

The second stage results can be found in Column 3. Overall, the local average treatment effect (LATE) of the substitution in education on income is approximately -21% and is not statistically significant at any conventional significance level. However, the coefficient is of an expected sign which de-notes that the impact may be negative. By adding controls, the coefficient of interest decreases to -23% but remains insignificant. We find that the effect is not statistically significantly different for female and male individuals. In contrast, the effect for individuals from rural location differs from urban by -5% and this difference is significant at 5% level. The overall results can be seen in Figure5as denoted by linear fit. We notice that it does not perfectly fit our data, especially at observations far from the threshold. We find that the overall average effect of substitution in education from old to new has no impact on change in individual income. Based on the results described above, we can conclude that the linear fit does not describe our data well enough, which is a motivation for higher order polynomial fits.

In addition, we inspect and run the linear model by age. Results for the second stage regressions can be found in the Table 6, Appendix B. The coef-ficient of interest is in the range of [0.222, 0.648] for age 42–45. In addition, the change in education has a statistically significant impact on income for age 42–44. However, estimations yield a positive coefficient in contrast to our main model and may imply that old education has no detrimental effect on income. For ages 46–49 the estimated coefficient is not robust, with large standard errors.

Results in Columns 6–7 describe the second order (quadratic) polynomial fit, which includes squared interactions of kink dummy and year at each side of the cut-off. From the Column 6 we can see that substitution from only old to new education yields a negative coefficient of -12%, which is statis-tically significant. Quadratic fit explains data better and has an expected negative first stage sign. The estimate in Column 7 denotes the second stage coefficient, which is approximately equal to -16% but is statistically insignifi-cant. In addition, final quadratic model has the highest explanatory power as compared to all other fits (13%). The quadratic fit for both first and second

(28)

stage estimations can be seen in Figures 4 and 5. From the visual evidence, we notice that this fit suits our data better than the linear fit and, again, estimates an expected negative change in slope in the old education.

An issue we acknowledge is that our analysis is most likely underpowered. The second stage results in Column 7 yield a standard error of 0.120. By choosing a conventional level of 80% power, we can calculate a minimum detectable effect (MDE) size. In order to obtain the chosen power for a 95% confidence interval, the true effect size must be at least 2.8 standard errors from the comparison point (assuming a normal distribution for estimation error). Given the estimated standard error, with 80% we can identify 0.12 · 2.8 = 0.336 = ±33.6% effect. This implies the smallest true effect, that can be found statistically significant. We believe this is a dramatically large effect. In addition, considering that the true effect is equal to the coefficient found −0.162, we can then calculate the power of analysis, which is equal to 27%. It may be that the variation in income variable is too small to have enough statistical power to detect changes in slopes because the bins are too small. In addition, our estimate lies within the confidence interval of [−0.362, 0.037], which contains values that could have been meaningful.

Furthermore, there may be other underlying issues and possible reasons for the results found. First to mention is year to year variation in old edu-cation that can be seen in Figure 2. We notice that there is a slight increase in average years of old education between points 0 and 1, however the last observation where individuals have old education only can be seen at point 0 not 1. In addition, functional form of the model chosen may be to simple to explain data on both sides of the cut-off. Secondly, labour market mech-anism rewards individuals for their ability and experience. It is likely that individuals at age 41–49 hold enough labour experience to compensate for their old education endowment. Adding labour experience variable in further research could possibly enrich the model. On the other hand, the design cho-sen analyses impact of only years of old education substitution to one (some) years of new education. This does not allow us to draw any conclusions on possible impact of only old education as compared to only new education.

Estimated results are in line with conclusions found by Jurajda (2005). This author finds that there is no statistically significant difference in returns to schooling and different educational levels by age groups. Jurajda provides empirical evidence by using data of old and young wage earners in Czech Republic in 2002. However, our design choice is not directly comparable to this article, thus, the consensus can only be drawn for conclusions and taken care with caution. Other articles described in Section 2.2. (Brunello et al. (2010), Orlowski et al. (2009), Fuchs and Marsella (2013)) find a significant difference in returns to education when compared old to young age groups.

(29)

Campos and Jolliffe (2007) find that the gap in returns between previously mentioned age groups is decreasing overtime, however, the last point in time analysed is 1998. This may signal that the gap, if followed the same trend, is non-existent in 2014, which is our last data point in time.

To conclude, our chosen design is unique in the topic of transitioned economies and to the extent of literature review carried, has never been used before. Further work with data from other countries or longer time periods could solve power issues and provide empirical evidence to back the reliability of design chosen. In addition, it would help explaining mixed evidence found by previous research.

(30)

Figure 4: First stage fits, age cohorts 41-49.

(31)

T able 4: Empirical results (1) (2) (3) (4) (5) (6) (7) Sample _selection First stage Second stage First _stage, con trols

Second stage, _con

trols Quadratic fit, first stage Quadratic fit, second stage E D Uol d --0.207 _(0.160) --0.223 _(0.167) --0.162 _(0.120) z · y ear − 0 .003 ◦ (0.000) 0 .01 ◦ (0.000)*** -0 .01 ◦ (0.000)*** --0.121 (0.029)*** -Ag e dummies X X X X X X X Ag e · y ear X X X X X X X y ear X X X X X X X C ontr ol s X -X X X X p − v al 0.325 0.001 0.194 0. 001 0.181 0.000 0.112 F − stat -37.61 -80.50 -76.73 -R 2 0.01 0.0465 -0.102 -0.110 0.135 N 14366 13669 13669 13669 13669 13669 13669 Note: Column (1) repres en ts selection equation with a de p enden t v ari able di , whic h is equal to on e if individ-ual’s inc ome is p os itiv e. Columns (2)–(6) presen t first stage (dep enden t v ariable E D U ol d i ) and second stage (dep end en t v ariable ln Yit ) regressions. Columns (2)–(3) represen t basic mo del, (4) adds con trols and (5)–(6) represen t e stimates of quadr atic fit with con trols. ◦ indicates that co efficien ts w ere scaled up b y 100. Con trols include gender and urban/rural d ummies. Age dummies and age and y ear in teractions are alw a ys included. Robust stand ard errors rep orted in paren theses. ∗ / ∗ ∗ / ∗ ∗∗ denote signi ficance at a 10/5/1 p ercen t confidence lev el. Al l results are estimated on a selected sample for age 41–49.

(32)

7 Robustness tests

In order to test the robustness of our results, we apply several robustness tests. First, we apply a placebo test in order to inspect how the final effect changes by shifting the cut-off value. The initial model assumed cut-off value at 2005 for age 41, 2006 for age 42,. . . , 2013 for 49. We start by shifting cut-off value one year forward, so that it equals 2006 for age 41, 2007 for 42,. . . , 2014 for 49. Analogous transformations performed by shifting cut-offs by ±1–3 years for given age.

Estimated coefficients of final causal effect and 95% confidence intervals are plotted in Figure 6. We can see that by shifting the cut-off value by +1/2 years, regressions yield coefficients in the range of [-0.481, -0.002]. Our main model coefficient is as well in this range. These results may imply that by shifting the cut-off value forward we still observe individuals that have only old type of education and individuals that start having new type of education. We notice that a shift by 3 years forward yields a high coefficient of 3.013 with large confidence interval. This may be explained by a prediction that 3 year shift is too far forward, meaning that we do not observe people with only old education anymore and compare those with a mix of education both before and after the cut-off.

By shifting the cut-off backwards by -1/2/3 years we get an interval of positive coefficients [0.013, 0.248]. This may be explained by the fact that by switching cut-off values we go back in time and assign kinks at years, were there were only people with old type of education and individuals just above the cut-off do not attend new education yet. Thus, we argue that this placebo test shows us that our main model is stable with a correctly chosen kink point and would as well yield insignificant values by shifting the cut-off with 1-2 forwards. However, since all of the estimated coefficients are statistically insignificant, we cannot rely on the results and conclude that the causal effect of change in education on income is negative as found by the main model.

Another test we use adds additional controls to our model to control for the macroeconomic changes that may have affected income observations over time. More specifically, Harmonized Index of Consumer Prices (HICP) and Gross Domestic Product (GDP) per capita in given year prices are added as additional control variables in the main model. By adding controls, the linear estimate reflecting causal effect on income decreased by five percentage point from -22% to -27%. In addition, only HICP is significant at 5% level, which indicates that an increase in HICP index by 1 unit leads to an increase in income by 5%. However, the coefficient of interest is still statistically insignificant. We conclude that controlling for macroeconomic shocks does

(33)

not significantly change the results of model used. This may mean that the magnitude and sign of coefficient is stable and does not reflect other channels that may affect income variable. Exact estimates for both placebo test and macroeconomic controls can be find in Table 6.

Note: Plotted coefficients represent estimated causal effect by shifting

cut-off values by -1, -2, -3 and 1, 2, 3 years for given age from the initial cut-off selection as indicated by 0 (main model).

(34)

T able 5: Robustness tests Placeb o tests Macro economic c on trols (1) (2) (3) (4) (5) (6) (7) (8) +1 +2 +3 -1 -2 -3 Main _mo del Macro _con trols E D Uol d -0.002 _(0.073) -0.481 _(0.430) 3.012 _(4.668) 0.248 (0.068)*** 1.093 (0.456)** 0.013 _(0.039) -0.223 _(0.167) -0.273 _(0.186) Ag e dummies X X X X X X X X Ag e · y ear X X X X X X X X y ear X X X X X X X X C ontr ol s X -X X X X X GD P -X H I C P -X p − v al 0.976 0.264 0.519 0.000 0.017 0.754 0.181 0.142 R 2 0.120 -0.160 -0.139 -N 13669 13669 13669 13669 13669 13669 13669 13669 Note: Only se cond stage results rep orte d. A ge dummies ar e always include d. R obust standar d err ors rep orte d in p ar entheses. ∗ / ∗ ∗ / ∗ ∗∗ denote signific anc e at a 10/5/1 p er cent confidenc e level. A ll results ar e estimate d on a final sele cte d sample for age 41–49 .

(35)

8 Discussion and conclusion

The majority of current active labour forces in transitioned societies contain individuals that have acquired education in a system dedicated to a well planned and predictable economy. We analyse the causal effect of a substi-tution in education from years of old, acquired before 1990, to some years of new on a change in income. We examine earnings of active labour force, aged 41–49, in 2005–2014 in Lithuania. By comparing individuals that only hold years of old education to those that have some new education, we look for a kink in annual gross income.

We use a Fuzzy Regression Kink design expressed as a two-stage least square estimator. In the first stage, we find a significant kink in the substi-tution from old to new years of education. A quadratic second stage fit yields a coefficient of -16%, which implies the size by which change in education af-fects change in slope of income before and after the kink point. However, our coefficient of interest is not significant at any conventional statistical level. Separate age regressions yield significant results for ages 42–44, however, the expected impact is positive rather than negative. The effect for individuals from rural location significantly differs from urban by -5%. We find no differ-ences in results by gender. Furthermore, performed placebo cut-off tests and macroeconomic controls prove a stable magnitude and sign of the coefficient of interest.

Labour market structure and effort mechanisms may suggest why our final effect found is insignificant. It may be that individuals with only old education have adapted to rapid changes in economy and are rewarded to their labour market experience and ability rather than official education. However, the result found may as well imply an existing diploma effect, which is a signal for an employer that individual has working credentials despite the actual skills or type of the education acquired. Other reasons such as year to year variation in endogenous variable and functional form chosen most likely contribute to the significance level of results. Conclusions drawn are consistent with previous research done by Jurajda (2005). Due to differences in data, design and specifications chosen, this comparison must be treated with caution. The findings reported are important because they may imply that the transition in the country is over and differences in wages can no longer be explained by differences in educational endowments but rather ability and effort.

The main analysis prospect would be to solve the power issues that oc-curred in the second stage estimations. One of the possible solutions could be to analyse data with longer time periods. Other performance outcomes such as length or probability of unemployment and likelihood of market exit

(36)

would lead to an interesting further research. In addition, it would be im-portant to note differences in labour market experience and scale it down to differences in contracts and occupational industry. This latter comparison may allow to explore differences in skill and industry premiums. There is a broad scope for further analysis on different educational levels such as higher education or vocational education only.

To conclude, the analysis carried out emphasizes that a market economy and its labour market can absorb frictions and large differences in human capital. By the evidence found, individuals with only years of old education are rewarded as much as their counterparts that have as well studied in the new system.

(37)

References

[1] Brunello, G.; Crivellaro, E.; Rocco, L., 2010. The Returns to education acquired under communism 15 years after the fall of the Berlin wall. In:

Lost in transition?, IZA Discussion Paper No. 5409.

[2] Campos, N.; Jolliffe A., 2002, 2007. Earnings, Schooling and Economic Reform: Econometric Evidence from Hungary (1986–2004). In: IZA

Discussion Papers 2678, Institute for the Study of Labor (IZA).

[3] Card, D.; Lee, David S.; Pei, Zhuan; Weber, A., 2012. Nonlinear Pol-icy Rules and the Identification and Estimation of Causal Effects in a Generalized Regression Kink Design. In: NBER Working Paper, No. w18564. SSRN 2179402.

[4] Chase, R. S., 1998. Markets for Communist Human Capital: Returns to Education and Experience in the Czech Republic and Slovakia. In:

Industrial and Labor Relations Review, vol. 51, 1998, no. 3, pp. 401–423.

[5] Economic Commission for Europe (UNECE), 1998. In: The Economic

Survey of Europe 3.

[6] Eslund, A., 2001. The Myth of Output Collapse After Communism.

Voprosy economiki, 7 issue.

[7] Eurostat. EU statistics on income and living conditions (EU-SILC) methodology - private households. Available at:

http://ec.europa.eu/eurostat/statistics-explained/index. php/EU_statistics_on_income_and_living_conditions_(EU-SILC) _methodology_-_private_households

[8] EU-SILC 2010–2014 data source in CSV format. Available at: http: //goo.gl/c1I18k

[9] Flanagan, R. J., 1995. Wage Structure in the Transition of the Czech Economy. In: IMF Working Paper, no. 95/36.

[10] Fuchs-Schundeln, N.; Masella, P., 2016. Long-Lasting Effects of Socialist Education. In: IZA Discussion Paper No. 9678.

[11] IMF, 2014. 25 Years of Transition Post-Communist Europe and the IMF. In: Regional Economic Issues, Special Report, Washington DC p.10–20. [12] Karvelien˙e L., 2007. Darbo teis˙es reforma Lietuvoje po nepriklau-somyb˙es atkurimo (The Reform of Labour Law in Lithuania after the Reestablishment of Independence) In: Magistro darbas.

Education in transition : regime change impact on income by using regression kink esign