When everyone goes to college: The causal effect of college expansion on earnings

Social Science Research 50 (2015) 229–245

Contents lists available at ScienceDirect

Social Science Research journal homepage: www.elsevier.com/locate/ssresearch

When everyone goes to college: The causal effect of college expansion on earnings Seongsoo Choi ⇑ Department of Sociology, Yale University, 493 College Street, New Haven, CT 06511-8933, USA

a r t i c l e

i n f o

Article history: Received 25 October 2013 Revised 23 November 2014 Accepted 25 November 2014 Available online 3 December 2014 Keywords: Returns to education College expansion Difference-in-differences South Korea Counterfactual

a b s t r a c t In this study, we estimate the causal effect of college expansion on earnings using the example of South Korea in the 1990s where the college enrollment rate increased from just over thirty percent to over eighty percent over a fifteen years period. We compare the preexpansion cohort and the post-expansion cohort in order to identify those who would attend college because of the expansion but would not attend otherwise (compliers). We, then, estimate compliers’ earnings gain from the college expansion relative to the earnings changes of two control groups: those who either would or would not go to college regardless of college expansion (always-takers and never-takers). We find a striking gendered pattern; for men, the earnings return to college expansion is moderate and mostly driven by the increasing skill price, whereas, for women, the return is significantly large even net of the skill price change. Ó 2014 Elsevier Inc. All rights reserved.

1. Introduction In most modern societies, completing more education results in increased earnings for individuals (Psacharopoulos and Patrinos, 2004). However, matters are less clear at the aggregate level. Does expanding access to higher education have economic benefits for those who can now attend college thanks to the expansion in the same way as an individual’s additional investment in higher education does? Does an increased share of college graduates improve the earnings of new college graduates and their relative positions compared to traditional college graduates? Would this be true even in the case where an expansion makes college attenders the majority in a cohort, making higher education close to universal? Answering these questions is important both theoretically and for policy. A mass expansion of education is not a simple sum of individual decisions for more educational investment but rather the transformation of institutions (Shavit et al., 2007). Thus, the consequences of educational expansion, as a newly emerging or modified social structure, demonstrate how a joint behavior at a collective level cannot be reduced to a simple collection of individual behaviors. In this way, the question about the social and economic rewards of educational expansion touches a core theme of sociology: the dynamic process of how individuals behaving at micro-levels influence and are influenced by macro-level phenomena. Furthermore, these questions have direct implications for policies aimed at expanding access to higher education. For example, President Obama proposed to make the United States the world leader in college graduation rates by 2020.1 Evaluating this kind of policy needs empirical evidence for the answers to the questions posed above. ⇑ Fax: +1 203 432 6976. E-mail address: [email protected] ‘‘Topping College Graduate Rates, Is It Worth It?’’ National Public Radio, January 29, 2013, available at http://www.npr.org/2013/01/29/170563090/ topping-college-graduate-rates-is-it-worth-it. 1

http://dx.doi.org/10.1016/j.ssresearch.2014.11.014 0049-089X/Ó 2014 Elsevier Inc. All rights reserved.

230

S. Choi / Social Science Research 50 (2015) 229–245

A crucial issue is that educational expansion is not applied to the entire population; only a group of individuals respond to or comply with expanded access to education. There are people who do not change their behaviors either because they would go to college anyway or because they would not attend regardless of a change in policy. Although many previous studies examine the societal-wide impact of educational expansion from various perspectives (Breen, 2010; Hannum and Buchmann, 2005; Lange and Topel, 2006; Shavit et al., 2007), the primary criterion that should be considered when evaluating the policies of educational expansion is how they impact their compliers. Distinguishing those who would comply with college expansion from the entire cohort of high school graduates is a challenging task methodologically because it requires us to think counterfactually. Researchers have tackled this issue directly or indirectly by adopting various research designs and methodological models. Some utilize varying propensities for college in a single cohort and infer whether expansion policies will be effective (e.g., Brand and Xie, 2010; Carneiro et al., 2011). Others more directly compare cohorts that are exposed differently to expansion (e.g., Devereux and Fan, 2011; Maurin and McNally, 2008), mainly capitalizing on the exogenous feature of the cohort membership as an instrumental variable (IV) and identifying compliers with the local average treatment effect (LATE) framework (Imbens and Angrist, 1994). In this paper, we examine a unique empirical case that offers an ideal setting for estimating the causal effect of educational expansion on earnings: the dramatic growth of college education in South Korea, where the college enrollment rate more than doubled from 30% to 80% during the 1990s. The South Korean college expansion offers a rare example because the expansion policy changed the status of college attenders from the minority to the majority within a short period of time. Adopting a counterfactual causal framework (Morgan and Winship, 2007), we identify a pre-expansion cohort and a postexpansion cohort and then estimate the change in earnings due to expansion for three groups: ‘‘always-takers’’ (who would be predicted to go to college before and after expansion), ‘‘compliers’’ (who would not have attended before the expansion but would do so after the expansion), and ‘‘never-takers’’ (who would not go to college in either period). Because the expansion only affected the educational outcomes of compliers, comparing the change in earnings for this group to the changes for always-takers and never-takers allows us to infer the effect of expansion on earnings. 2. Previous literature 2.1. College premium and expansion as increasing skill supply The mainstream explanation in economics about the role of educational expansion in the process of wage or earnings determination is the supply-demand model. In this framework, the increasing supply of skilled workers that occurs with educational expansion, along with the demand for skilled labor, determines the level of earnings. It is well documented that the temporal dynamics of the earnings returns to education over the last century in the US are largely explained by this supply and demand mechanism (Card and Lemieux, 2001; Goldin and Katz, 2008; Katz and Murphy, 1992). For example, the socalled college wage premium rose sharply in the US from the 1980s up to its record high level in the mid-2000s (Goldin and Katz, 2008). Considering a secular increase in demand factors (largely, skill-biased technological change, or SBTC), the widening earnings gap between college graduates and non-graduates can be attributed to a stagnated supply of college graduates (Goldin and Katz, 2008). The title of Goldin and Katz’s (2008) book The Race between Education and Technology aptly summarizes this history. The supply and demand framework has also been applied to other countries (e.g., Harmon et al., 2003; Walker and Zhu, 2008 for the UK and Gebel and Pfeiffer, 2010 for Germany) including South Korea. In South Korea, the increase in collegeeducated labor in the 1980s drove a slight decline in the college premium, but during the 1990s, when college education expanded dramatically, the college premium started to rise thanks to a greater demand for high-skilled labor (Choi and Jeong, 2005; Choi, 1996; Choi et al., 2005). This body of research only offers limited implications for examining the causal effect of educational expansion, however. A change in the college premium alone reveals little about whether expanded educational opportunities benefit those who attended college because of the expansion. An increased college earnings premium following educational expansion could be driven by increased earnings for those ‘‘compliers’’ who would not have attended before expansion. Yet the premium increase could also result from a change in the premium for the ‘‘always-takers’’, those who would also have attended even before expansion; the always-takers may lose their premium because of the increasing supply of college graduates (compliers) and also may earn more because of a shift in demand for skilled workers in the labor market. Such heterogeneity among college graduates addresses two important issues that need to be considered when examining the effect of college expansion. First, making claims about the causal effects of expansion requires distinguishing the returns of always-takers (who consistently have higher propensities for college) from the returns of compliers (whose lower propensities for college fall below the threshold for enrollment in the pre-expansion period but above it in the post-expansion period). Second, it is necessary to compare the returns of compliers with those of always-takers in order to control for any changes in the labor market that are induced by the dynamics of supply and demand in college-educated labor. 2.2. Heterogeneous earnings returns to college education and expansion policy Empirical studies of the heterogeneous returns to college show mixed results (Hout, 2012). For example, Brand and Xie (2010), using data from the National Longitudinal Study of Youth 1979 (NLSY79) and the Wisconsin Longitudinal Study, find

S. Choi / Social Science Research 50 (2015) 229–245

231

higher returns among those less likely to attend college. However, Carneiro et al. (2011) estimated marginal treatment effects (MTE) using the same NLSY 79 data and concluded that the marginal effect of college attendance is greater for those who are more likely to be college-educated than those with a lower likelihood. A similar inconsistency is found for Taiwan (Tsai and Xie, 2011, 2008). It is possible to extend the findings of positive selection (Carneiro et al., 2011; Tsai and Xie, 2011; Willis and Rosen, 1979) and negative selection (Brand and Xie, 2010; Tsai and Xie, 2008) to the implications of expansion policies. For example, the theory and findings of positive selection can be interpreted as a skeptical prediction of expansion policies because they imply that current college-related choices among high school graduates are made on the basis of the principle of comparative advantage (Willis and Rosen, 1979) and, therefore, there is no point in inducing non-college graduates to college against their current comparative advantage. In contrast, the pattern of negative selection suggests that the expansion of college education would be beneficial to new college graduates, who would otherwise have been unlikely to attend college (Brand and Xie, 2010). But those studies of the heterogeneous returns to education seem to have an inevitable limitation as long as their empirical analyses are based on a single cohort. Since they infer policy implications from a variation in propensity for college in a single cohort (Brand and Xie, 2010; Carneiro et al., 2011), they do not estimate the effect of an actual expansion policy. Instead, they tend to produce generalized estimates under implicit or explicit hypothetical scenarios of college expansion after making corresponding assumptions. Moreover, deriving a general policy implication from the propensities for college education of one peculiar cohort (e.g., the cohort of NLSY79, most members of which went to college in the mid-1980s) does not seem to be the fit for the consideration of policies at present. Those studies seem to be of limited use especially when we consider expansion policies that lead to a rapid influx of new students rather than a gradual increase. Since our interest is primarily in the economic consequence of a substantial growth in new college graduates, the single-cohort studies provide only limited information. 2.3. Cohort-based studies on educational expansion An easy way to overcome the limitation of studies with the single-cohort design is to compare the earnings of a preexpansion cohort with those of a post-expansion cohort. Given that a policy implementation is mostly an exogenous event to individual decisions for college entry and, therefore, can be considered as quasi-experiment (Card, 1999), a direct comparison, among the compliers of the policy, between pre- and post-policy periods can give a more direct and reliable effect estimate as long as an appropriate case and data are available. There are a few studies that compare cohorts differentially exposed to an expansion of college education. Devereux and Fan (2011) used cohort membership as an instrument, taking advantage of variation in postsecondary enrollment in the UK in the 1990s for cohorts born between 1970 and 1975. They found significant earnings gains among post-expansion cohorts, suggesting that those who complied with the expansion earned significantly more due to their postsecondary education. Gurgand and Maurin (2007) examine the impact of educational expansion in France by comparing the wages of different birth cohorts. They argue that, because the share of top-tier school graduates in France has remained stable while the overall share of the population attending postsecondary institutions has expanded, change in relative wages shows gains among compliers for the post-expansion cohorts. Attewell and Lavin (2007) and Maurin and McNally (2008) utilize suddenly-lowered criteria determining postsecondary eligibility in the US and France to identify students who would not have been accepted before the higher standards were relaxed. Both studies found positive (or at least non-negative) economic benefits of college, relative to always-takers, for those on the margin of college enrollment. Most of these cohort-based studies feature the use of cohort membership as an instrument for identifying the compliers of the educational expansion. Birth cohort membership and the varying extent that the cohorts are exposed to the expanded access to college are usually exogenous to earnings in the labor market. Under the assumption of a monotonic expansion of education, then, the IV estimate of economic returns to education can be understood as returns to education for the compliers of the expansion (Devereux and Fan, 2011; Imbens and Angrist, 1994). In this study, we extend this line of research with two additions. First, as mentioned earlier, the South Korean college expansion is much more substantial than the cases in other countries especially in that the expansion changed the status of college graduates from a minority to the majority within a short period. This setting enables us to examine what the economic benefit of college education would be if college education becomes almost universal. Second, we demonstrate a novel method that overcomes a limitation in the LATE estimation. In the IV approach, the LATE tells only about the earnings gain for compliers and informs nothing about the gains for always-takers and never-takers (Devereux and Fan, 2011). For a more appropriate assessment about the impact of college expansion, however, the gain of compliers should be examined in comparison with those of two control groups, especially always-takers. If compliers share a large proportion of their earnings gain with always-takers, we cannot attribute the gain solely to the pure effect of compliers’ following the expansion policy because always-takers’ college-going decisions are not affected by the expansion. 3. Economic returns to college expansion: the human capital effect and the skill price effect The economic benefit from college expansion for compliers can be understood as a joint product of two separate mechanisms, which conceptually correspond to the composition effect and the price effect of education (Juhn et al., 1993). Com-

232

S. Choi / Social Science Research 50 (2015) 229–245

pliers’ increased quantity of higher education can create more human capital to the compliers, change them into skilled workers, and affect compliers’ economic returns (human capital effect). At the same time, compliers can experience a gain or loss from a change in the price of higher education or skill in labor markets (skill price effect). This change in skill price is due to market changes that are not directed related to college expansion, but compliers would enjoy or suffer from the skill price effect because they joined higher education. The effect of college expansion on earnings, therefore, are decomposed into these two effects. The human capital effect is unique for compliers whereas the skill price effect holds for both compliers and always-takers. 3.1. Human capital effect Social science theories largely offer two major views on the human capital effect of college expansion on the economic outcome: human capital theory and positional good theory. Human capital theory is built on the micro-behavioral assumption that education is one of the major investments through which productive skills can be obtained. As a form of capital, an investment in education can raise marginal productivity with subsequent rewards recognized by the market (Becker, 1993). Such rewards, whether wages or earnings, reflect the prices that are determined by the balance of the supply of skilled workers and the demand for their skills. Holding the demand for skilled labor constant, college expansion as an upgrade in human capital for compliers is expected to raise economic returns for compliers, while it is also expected to lower the average economic return for all college graduates (alwaystakers and compliers). On the other hand, there are alternative views holding education as a relative good rather than something with absolute value as human capital. Signaling/sorting theory, for example, contends that educational attainment is of value not for the skills acquired but for the unmeasured ability it represents (Spence, 1973; Weiss, 1995); the amount of schooling an individual receives is the best observed indicator of his or her unobserved ability, which is the true carrier of market productivity (Galor and Moav, 2000; Taber, 2001). Therefore, educational returns are returns to unmeasured ability, not returns to skills. Another prominent sociological theory grounding a positional good perspective is credential theory, wherein education matters as a social and cultural construct that stems from social conflict between competing groups (Brown, 2001; Collins, 1979). Credential theory focuses primarily on group-level dynamics; how occupational groups – firms or professional organizations – use education credentials to keep their prestige through restricted entry, which is allowed only for those with more or less selective credentials. Educational certificates, therefore, carry cultural value beyond their contributions to marginal productivity (Bills and Brown, 2011). When we predict the impact of college expansion on labor market outcomes, signaling/sorting theory offers an individual-level explanation. Those with higher ability are more successful than those with lower ability at using education to signal their ability to employers; thus, compliers, despite investing more in education during the expansion, still earn less than always-takers due to their lower ability. Credential theory, on the other hand, offers a more sociological and organizational story. Occupations or jobs with higher socioeconomic status require more selective credentials as more people come to apply with postsecondary degrees (Bills and Brown, 2011). As a result, credentials from more selective universities or graduate degrees become necessary to gain access to positions in more selective jobs, leaving compliers with little chance of entering these occupations that remain more readily available to always-takers. The persistent gap between compliers and always-takers can be seen as a product of social or organizational dynamics rather than individual differences in ability. Whether the underlying mechanism is individual or organizational, the theories that see education as a positional good commonly suggest that a college expansion does not necessarily have a positive human capital effect for its compliers. Drawing on the theoretical discussion so far, we can construct two competing hypotheses on the human capital effect in compliers’ economic returns to college expansion: Holding demand-side factors constant, (1) Human capital hypothesis: The expansion of higher education will bring more earnings to its compliers. (2) Positional good hypothesis: The expansion of higher education will not bring significantly more earnings to its compliers. 3.2. Skill price effect The economic reward of higher education in the labor market or college premium, in short, changes over time. The change in college premium that compliers share with always-takers is the skill price effect of college expansion. This skill price effect is a product of a change in supply of college graduates (or skilled labor) and a change in demand for them (Card and Lemieux, 2001; Goldin and Katz, 2008). On the one hand, college expansion as a large influx of college graduates induces a negative externality. As a result of such an increasing supply of college graduates, the college premium of both always-takers and

Percent

S. Choi / Social Science Research 50 (2015) 229–245

233

100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%

Year

Fig. 1. College enrollment rates in South Korea. Source: Yearbook of Education Statistics (each year), available at Korea Education Statistics Service (httpe:// cesi.kedi.re.kr).

compliers diminishes, although always-takers did not experience any change in their educational attainment.2 On the other hand, the skill price can shift as the demand for skilled labor rises. The literature documents that such a shift occurred in many advanced countries (Choi and Jeong, 2005; Goldin and Katz, 2008; Psacharopoulos, 2009; Walker and Zhu, 2008), although it is controversial whether the shift in college premium is truly due to the rising returns to skill (Berman and Machin, 2000; Berman et al., 1998) or the rising returns to other unmeasured factors that are related to productivity (Choi and Jeong, 2007; Taber, 2001). Theoretically, whether the skill price effect is positive, negative or ignorable is difficult to predict a priori. If the demand for higher education from the market does not follow the increasing supply resulting from a college expansion, the skill price effect might be negative. On the contrary, if the market demands skilled or educated workers more even after absorbing all the compliers of college expansion, the skill price would be positive. What is important is that the skill price effect is part of the economic gain for compliers because they would not be able to take advantage of the gain if they did not attend college by complying with an expansion policy.

4. Empirical context: the expansion of college education in South Korea To estimate the causal effect of massive educational expansion on economic returns, our empirical study must meet a few conditions. First, the expansion of the educational system should be monotonic. This means that no individual will experience a decrease in the propensity for college education throughout the expansion period. The monotonicity assumption is necessary to link micro-outcomes (earnings) to macro-treatment (expansion) so that their relationship becomes identifiable; such monotonicity is a standard assumption for identification in the literature exploring policy effects (Angrist and Pischke, 2009; Imbens and Angrist, 1994). The second condition is that the educational expansion needs to be more than a minor change in enrollment since a large increase in the share of students at a given level of schooling is a necessary condition for estimating its effect. Finally, to satisfy the first two conditions, the educational expansion would, ideally, take place rapidly. If an educational system has expanded gradually over decades, many external forces may intervene and influence the overall level of educational attainment. It would then be difficult to identify the pure effect of educational expansion. The South Korean experience meets these conditions reasonably well. Fig. 1 shows how rapidly the South Korean college system expanded: before 1992, fewer than 35% of high school graduates went to college, but by 2004, 80% did. The number of students enrolled in college almost doubled and the number of postsecondary institutions also increased (from 233 to 329) over this fifteen-year period. This data shows that the expansion was substantial not only because enrollments almost doubled but also because the share of compliers was more than a third of the population. Before the expansion, college graduates were the minority; after the expansion, they became the majority. As a result, the expansion of higher education in Korea enables us to examine the impact of the rapid change by comparing earnings in a period where only a small proportion of individuals in a cohort hold a college degree with earnings in a period where almost everybody in a cohort does. What has driven college expansion in South Korea? The answer seems quite straightforward. The government initiated a new educational reform in May 1995, which is now known as the 5.31 Education Policy. This reform declared that it would promote competitiveness in higher education by introducing a more market-oriented system (Chang, 2009). The reforms relaxed and simplified the central government’s strict standards for the establishment of new institutions, thereby lowering the barriers for newcomers. It also included a policy allowing existing universities and colleges to manage student openings, 2 There may be a positive externality that results from the social returns to college expansion. For example, an upgraded skill level in an industrial sector, owing to the influx of compliers, may promote productivity and innovation and, consequently, can bring higher earnings for non-complier groups, including always-takers in the same sector. The presence of non-ignorable social return or spillover effect of higher education has been well documented (Hout, 2012; Kim and Lim, 2012; Moretti, 2004; Psacharopoulos and Patrinos, 2004), but most of evidence is found for high school graduates. Previous studies rather suggest that the social return diminishes for those with higher levels of education (Kim and Lim, 2012; Moretti, 2004), implying the minimal social return for alwaystakers.

234

S. Choi / Social Science Research 50 (2015) 229–245

which had been previously under the government’s control. Given that higher educational aspirations had already been widespread in South Korea, the reform triggered a rapid expansion of the college system. In summary, individuals’ beliefs about college did not change, but the government’s changed policies allowed more individuals to complete the education they wanted. 5. Data and methods 5.1. Data For this analysis, we use data from the Korea Labor and Income Panel Study (KLIPS). The KLIPS started in 1998 with individuals aged 15 and above in 5000 households which were sampled in a nationally representative way. The total number of individuals in the first wave sample is 13,321. The most recent available wave is the 12th, which was collected in 2009. We take a cohort-based approach. We compare the 1965–1971 birth cohort with the 1976–1982 birth cohort. Given that most people finish high school at age 19 in South Korea, those in the earlier cohort would have completed high school between 1984 and 1990 and those in the later cohort between 1995 and 2001. The cohorts were chosen because they faced the choice of whether or not to enter college before and during the expansion, respectively. The choice of the second cohort was also determined by data availability. For those who were born after 1982, relatively little information on their earnings is available. In this analysis, birth cohort membership is equivalent to treatment status in the expansion of the college system; the earlier cohort is the control group and the later cohort is the treatment group. As Fig. 2 illustrates, the two birth cohorts have strikingly different distributions of college attendance, consistent with the official statistics shown in Fig. 1. For the outcome, we use monthly labor earnings from ages 27 to 33, measured in 1998 for the earlier cohort and in 2009 for the later cohort. 5.2. Methodological strategy 5.2.1. The difference-in-difference (DID) model The DID model for our main analysis can be expressed as follows: 0 AT CP AT logðyi Þ ¼ a þ b1 t CP i þ b2 t i þ hT i þ d1 ðt i  T i Þ þ d2 ðt i  T i Þ þ b X i þ ei

ð1Þ

AT where tCP i and t i represent compliers and always-takers respectively, T denotes the post-expansion period (or cohort), and X is a vector of control variables that are likely to affect earnings between the pre- and post-expansion cohorts differentially for reasons other than the college expansion (e.g., age, labor market experience). b1 and b2 show the pre-expansion earnings levels of compliers and always-takers (CP and AT, hereafter) relative to the reference group, never-takers (NT, hereafter), and h captures the growth in earnings for NT during the period of college expansion. In this model, the key parameters of interest are d1 and d2. The returns to college expansion for CP is identified as the interaction between the indicator of the expansion and the indicator of being compliers, d1, which captures the total economic gain compliers earn due to the expanded access. As discussed earlier, this total gain is the sum of what the compliers could gain due to their added schooling (human capital effect) and what they could gain or lose due to the change in the labor market rewards of higher education (skill price effect). This skill price effect is captured by d2, the change in the college

100% 90% 80% 70% 60%

4-Year College, 28.5% 2-Year College, 11.5%

4-Year College, 48.5%

50% 40% 30%

No College, 60.0%

2-Year College, 28.4%

20%

No College, 23.1%

10% 0%

Pre-Expansion Cohort

Post-Expansion Cohort

(1965-1971)

(1976-1982)

Fig. 2. The distribution of college attendance in the KLIPS samples.

235

S. Choi / Social Science Research 50 (2015) 229–245 Table 1 Construction of never-takers, compliers, and always-takers.

Pre-expansion Cohort Pre-expansion Post-expansion decision decision (actual) (counterfactual)

Post-expansion Cohort Pre-expansion Post-expansion decision decision (counterfactual) (actual)

Never-takers (NT)

Compliers (CP)

0

↗

0

0

← Monotonicity

↘ Simulation (Ignorability)

1

0

↖

→ Monotonicity

1

1

↙ Simulation (Ignorability)

Always-takers (AT) 1

0

1

premium of AT during the years of the college expansion, if we assume that there is no quality difference in education between compliers and always-takers.3 Since d1 captures the sum of the human capital effect and the skill price effect, the human capital effect can be identified by d1  d2. This parameter can be estimated directly from the equivalent DID model in which AT is omitted as a reference group. The key identifying assumption of the DID model is that the trends of the change in the outcome are parallel between the group cases in the absence of treatment (Angrist and Pischke, 2009). In this double-treatments setup of our DID model, the assumption of a parallel trend states that NT, CP, and AT would experience the same changes in earnings in the absence of both the college expansion and the education-related market changes. The assumption seems to hold true, almost by the definitions of NT, CP and AT. To summarize, our DID model not only estimates the total returns to college expansion but also disentangles the human capital effect and the skill price effect of college expansion. 5.2.2. Defining never-takers, compliers, and always-takers The core of our effort in this analysis is to identify NT, CP, and AT. Since individuals are observed as a member of only one cohort – either pre- or post-expansion – it is impossible to distinguish three groups based only on the observable information. Therefore, we need to construct counterfactual decisions that individuals of one cohort would have made if they had belonged to the other cohort. With those counterfactuals, we are then able to classify each individual into either the NT, CP, or AT group using the rule described in Table 1. To generate counterfactual college-going decisions, we invoke two assumptions: ignorability and monotonicity. The ignorability assumption means that we are able to predict the unbiased propensity for college attendance from the set of observed information. Therefore, invoking the ignorability assumption enables us to reconstruct a college-going decision with just a set of observable variables.4 The assumption of monotonicity means that no individual would choose a lower level of education after the expansion than the level they chose or would have chosen before the expansion.5 With the monotonicity assumption, the counterfactual college-going decisions of two groups are easily figured out. Those in the NT group did not go to college after the expansion (about 29% of the post-expansion cohort) and would not have gone to college in the pre-expansion years. In contrast, those in the AT group went to college before the expansion (about 40% of the pre-expansion cohort) and also would have gone to college after the expansion. It is less clear how to distinguish CP from NT among those in the pre-expansion cohort who did not go to college, and from AT among those who did go to college in the post-expansion cohort. We take a simulation approach with the following steps. First, we compute a counterfactual propensity score for college attendance by multiplying the given individual’s observed variables by the estimated corresponding coefficients of the other cohort. For example, for an individual in the pre-expansion cohort, we obtain her propensity score for college-going that she would have had if she had belonged to the post-expansion cohort. Likewise, all the individuals of the post-expansion cohort can have their own counterfactual probabilities of going to college by this procedure of extrapolation.6 3 It is reasonable to believe that there is a quality difference between higher education in the pre-expansion period and that in the post-expansion period. If an expansion occurs in a sudden and drastic manner as in South Korea, many key educational standards can be downgraded. In such a case, the quality of education that compliers attain may be lower than that for always-takers, and, then, the skill price effect for compliers estimated from d2 can be upwardly biased. We address this issue in a later discussion. 4 The ignorability assumption is not immune to the possibility of being violated especially when we do not have a good measure of academic ability to control. We address this concerns in the robustness analyses section. 5 Although we cannot test whether the monotonicity assumption is true or not, we argue that the assumption seems to be convincing in the South Korean context. It is hard to think of any peculiar demographic group experiencing a decline in the college enrollment during the period of large-scale college expansion. This is confirmed by the official statistics of the South Korean educational authority (see http://cesi.kedi.re.kr/ for the Educational Statistics of South Korea). 6 This extrapolation can be justified only when the distributions of the observed predictors of college education between two cohorts are not different. However, as the descriptive statistics in Table 2 shows, they are somewhat different. To sort this out, we take an additional procedure to adjust for this imbalance as is discussed later.

236

S. Choi / Social Science Research 50 (2015) 229–245

Second, for each individual, we draw a random binary variable from the Bernoulli function where the probability parameter is set to the propensity score, so that those with higher propensity scores are more likely to get 1 (=going to college) and those with lower propensity scores are more likely to end up with 0 (=no college). By such imputation, we assign counterfactual college-decisions to those without college education in the pre-expansion period and to those with college education in the post-expansion period. Third, with the combination of the pre- and the post-expansion decision, one of which is actual and the other is counterfactual, we assign each individual to one group of NT (0, 0), CP (0, 1), and AT (1, 1). All of the steps described so far are summarized in Table 1. Since the probabilities are realized only in the long run, we repeat the random imputation and subsequent steps many times such that we have many sets of the counterfactual college decision for each individual. After estimating the DID model (Eq. (1)) with each set of randomly imputed counterfactual college decisions, we average the estimated coefficients and standard errors using the formulas given in Eq. (2). This procedure of iteration is analogous to the multiple imputation method (Allison, 2002) where b is an estimated coefficient, SE(b) is its standard error, and the number of iterations is 500: 500 X ¼ 1 b br 500 r¼1 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u   500 500 2 u 1 X 1 X ðbr  bÞ 2 SEðbÞ ¼ t SEðbr Þ þ 1 þ 500 r¼1 500 r¼1 R  1

ð2Þ

A binary decision of whether someone attends any college probably oversimplifies the college-going decision because the economic returns to postsecondary education might differ by type of degree, which is also closely related to quality of skill that students acquire or employers expect. In the South Korean higher education system, there are largely two types of postsecondary institutions: junior (2-year) colleges, and 4-year colleges or universities. To address this difference, we additionally analyze the DID models with alternative outcomes such as attending a junior college and a 4-year college. These outcomes identify two different complier groups: (1) the compliers of the 2-year college expansion: those who would not go to any college before the expansion but would go to a 2-year college after the expansion, (2) the compliers of the 4-year college expansion: those who would not go to any college or would go to a 2-year college before the expansion but would go to a 4-year college after the expansion.7 5.3. Variables Table 2 lists the variables we account for in the analysis including the precollege variables used to predict propensity scores for college attendance, the post-college variables used as controls in the main earnings analysis and the outcome variables. Most of the precollege measures are common predictors of educational attainment in sociological research: years of schooling for both mother and father, parental occupational status (of the main breadwinner) at age 14 measured by the International Socioeconomic Index for Occupational Status (Ganzeboom and Treiman, 1996), and number of siblings. In addition, we include variables indicating high school track (general versus vocational) and residential area at age 14 (the greater Seoul area, other major cities, and other small towns and rural areas). We also include a series of birth year dummies to account for possible heterogeneity within each of cohorts. And, finally, we include a dummy for sex. Ideally one would fit models separately for each sex, but, in order to keep the sample size sufficiently large, we estimated models for both sexes together. Instead, we include several interaction effects of sex and other significant variables. To minimize the loss of sample size due to missing data, especially of parental education, occupational status, number of siblings, and high school track, we use multiple imputation (Allison, 2002).8 The final sample sizes are 2028 for the 1965–71 cohort and 1584 for the 1995–82 cohort. For the main analysis, we include dummy variables for age and years of labor force experience in 1998 for the earlier cohort and in 2009 for the later cohort along with the main treatment variables. 6. Analysis and findings 6.1. Decisions for college education in pre- and post-expansion We analyze models predicting college outcomes by precollege covariates using logit models. The estimated results are generally in line with the previous literature (see Table A in the supplementary document for the result tables presenting 7 Here, we additionally assume the monotonicity of college expansion based on the strict hierarchy between junior (2-year) colleges and 4-year colleges: no one prefers 2-year colleges over 4-year colleges when a 4-year college becomes available due to the expansion. This assumption is reasonably accepted in the South Korean context (Park, 2007). 8 Those five imputed variables contain non-ignorable proportions of missing observations: father’s education (4.6%), mother’s education (25.7%), number of siblings (25.3%), high school track (25.4%), and parental occupational status (12.7%). We imputed these variable using the method of imputation with chained equation (ICE), which assumes that the missing patterns of the imputed variables are interrelated as to be multivariate but not necessarily in a monotonic way. To predict missing values, we regress each imputed variable on college destination, residential locations at birth and at age 14, birth year, gender, parent’s employment status at 14, as well as other imputed variables.

237

S. Choi / Social Science Research 50 (2015) 229–245 Table 2 Descriptive statistics. Pre-expansion cohort (1965–1971)

Post-expansion cohort (1976–1982)

Female (vs. male) Mother’s education (year) Father’s education (year) Parental occupational status Number of siblings General high school (vs. vocational)

.494 5.314 (4.445) 8.167 (4.275) 33.751 (13.826) 3.294 (1.754) .612

.479 9.007 (4.324) 10.600 (4.451) 37.502 (18.294) 1.745 (1.473) .662

Residential area at age 14 Seoul (the capital city) Other major cities Other small towns and rural areas

.235 .183 .582

.296 .262 .443

Birth year 1965/1976 1966/1977 1967/1978 1968/1979 1969/1980 1970/1981 1971/1982

.146 .143 .127 .165 .145 .135 .140

.139 .144 .132 .143 .176 .137 .130

College decision No college Junior college (2–3 year) 4-Year college

.600 .115 .285

.231 .284 .485

Monthly labor earningsa,b Socioeconomic index of occupationb Labor market experience (year)b

1586 (806) 42.944 (11.986) 7.993 (3.509)

1830 (869) 45.502 (12.901) 7.830 (2.838)

N

2028

1587

Note: 1. Numbers in parentheses are standard deviations (in case of continuous variables). 2. The sampling weights offered by the KLIPS are applied. a Earnings are measured in 1000 Korean won (KW), which is roughly equivalent to 1 USD. Earnings in 1998 are adjusted to 2009 CPI. To minimize the effect of outliers, I top-coded using 10,000,000 KW as an upper bound. b For those who were employed at the time of the interview (1998 and 2009) only (905 and 909 respectively; 1814 in total).

the average partial effect estimates on attending any colleges and attending 4-year colleges). A male advantage in the preexpansion cohort disappears in the post-expansion cohort. The education of parents (mothers, in particular) and their occupational status when the respondent was 14 are significant predictors of college destination. High school track is a very strong predictor of college, and residential location at age 14 also matters. Table 3 shows the distribution of individuals in the NT, CP and AT groups after the 500 random drawing simulations described in an earlier section. In the case of attendance of any postsecondary institutions, 34% of the pre-expansion cohort belong to NT, another 26% to CP, and the remaining 40% to AT. Similarly, 37%, 25% and 39% of the post-expansion cohort are NT, CP and AT respectively. When it comes to 4-year college attendance, about 18% of the pre-expansion cohort and 16% of the post-expansion cohort are CP, those who would not have gone to a 4-year college (but possibly have gone to a 2-year college) before the expansion but would have gone to a 4-year college after the expansion. Regarding the distributions of the precollege covariates, there are notable pre–post differences in some covariates such as parental schooling, parental occupational status, and sibling size as shown in Table 3. Part of the sharp shift in college enrollment might be explained by this pre–post difference in the distribution of the family background factors, not by the expansion itself. This is mainly because parental generations also experienced rapid upgrades in educational attainment and a large shift in the occupational structure as a consequence of rapid industrialization in Korea. To address this concern, we adjust the distributions of parental education, parental occupational status, and number of siblings for the post-expansion cohort to match those of the pre-cohort by the weighting method proposed by Iacus et al. (2012). We then apply the generated weight to the main analysis of the DID models (see Appendix B in the online supplement for the details of the weighting procedure). The goal of this weighting procedure is to balance the distributions of potential propensity scores for college attendance between the pre-expansion cohort and the post-expansion cohort. In other words, we expect that, with weighing, there should be no meaningful difference in the distribution of propensity for college attendance between the pre- and the post-expansion cohorts. Fig. 3 shows that significant cohort differences in propensities both for pre-expansion college attendance and for post-expansion attendance disappear after the weighting procedure applies. This means that, with the weighting adjustment, a concern about possible influences of observed precollege factors in our main analysis can be ruled out.

238

S. Choi / Social Science Research 50 (2015) 229–245

Table 3 The distribution of college attendance groups. Attending any colleges

Never-taker Compliers Always-takers

Attending 4-year colleges

Pre-expansion cohort

Post-expansion cohort

Pre-expansion cohort

Post-expansion cohort

.339 .263 .398

.285 .280 .435

.539 .179 .282

.577 .197 .226

Note: The proportions of the college entry groups are the averages from 500 simulations of random drawings.

Pre-Expansion, Adjusted 20

20

Pre-Expansion, Unadjusted

Mean Difference=-.025

0

0

5

5

Percent 10 15

Percent 10 15

Mean Difference=.116*

0

.2

.4 .6 Propensity Score

.8

1

0

.4 .6 Propensity Score

.8

1

Post-Expansion, Adjusted

30

30

Post-Expansion, Unadjusted

.2

Mean Difference=.030

0

0

Percent 10 20

Percent 10 20

Mean Difference=.178*

0

.2

.4 .6 Propensity Score

.8

1

Pre-Expansion Cohort

0

.2

.4 .6 Propensity Score

.8

1

Post-Expansion Cohort

Fig. 3. Distributions of propensity scores for attending college in pre- and post-expansion periods before and after matching adjustment. Note: The cohort bars show the distributions of propensity scores that are predicted from the pre-expansion model and the post-expansion model, with and without adjustment, respectively; ⁄ denotes statistical significance at the .05 level.

6.2. Earnings returns to college expansion Table 4 reports the results from the DID models estimating earnings returns to the expansion of the overall higher education. The result of the model with both genders reveals that, before the expansion, both NT and CP had significantly lower levels of earnings than AT, showing 11–12% of college premium. This is not very surprising because NT and CP did not differ by the observed college outcome in the pre-expansion period. The small and insignificant coefficient of college expansion (.024) shows that individuals in the NT group experience an ignorable increase in earnings between 1998 and 2009. It suggests that there was little economic progress among high school graduates with no college education. The interaction of CP and the expansion captures the total effect of the college expansion. The estimate suggests that CP experienced about an 18 percentage point increase in earnings more relative to what NT experienced, which is statistically significant. The estimated coefficient of the interaction between AT and the expansion identifies the skill price effect of college expansion. The result documents an economic gain of about 11 percentage points of the earnings for workers with skilled labor or higher education. Therefore, out of the 18 percentage point earnings gain, CP would have gained only 7 percentage points (=18–11) if there had been no education-favoring demand-side changes in the labor market between 1998 and 2009. This 7 percentage point difference comes short of statistical significance. When we consider the results of men and women separately, however, we find a very different picture. For men, CP earned about 12% in total from college expansion but almost all of the total return is attributed to the increasing returns for skilled labor (the skill price effect). Net of the skill price change, the male compliers would have benefited very little from their additional educational investment. For women, however, the pattern differs strikingly. The pre-expansion college premium is as large as 36–41%. The CP group earned considerably more than NT and AT: 31 percentage points of significant earnings returns to college expansion, net of the skill price effect. The skill price effect is slightly negative and statistically ignorable (.072), suggesting a modest decline in college premium among women likely due to the dominant impact of the sharp increase in college graduates and the relatively small benefit from skill-favoring market changes. The total return to the expansion for female compliers amounts to 24%, although it fails to reach statistical significance possibly, in part, due

239

S. Choi / Social Science Research 50 (2015) 229–245 Table 4 Estimates from the earnings returns to college expansion. Men and women

Men

Women

CP (vs. NT) AT (vs. NT) Expansion CP  Expansion (vs. NT  Expansion) AT  Expansion (vs. NT  Expansion) CP  Expansion (vs. AT  Expansion)

.011 .123 .024 .179 .106 .073

.002 (.048) .041 (.043) .033 (.050) .115 (.071) .115 (.058)* .000 (.059)

.042 (.140) .406 (.107)*** .136 (.111) .241 (.155) -.072 (.125) .312 (.123)*

N

1814

1202

612

(.048) (.041)** (.046) (.063)** (.054)* (.051)

Note: 1. The estimates are computed from 500 simulations of random drawings. 2. Matching weights for balancing family background covariates between two cohorts are applied. 3. Age dummies and residential areas at age 14 are included. Their estimates are omitted but available upon request.   p < .1. * p < .05. ** p < .01. *** p < .001.

Table 5 Estimates of the earnings returns to college expansion, by type of colleges. Attending 2-year colleges

CP (vs. NT) AT (vs. NT) Expansion CP  Expansion (vs. NT  Expansion) AT  Expansion (vs. NT  Expansion) CP  Expansion (vs. AT  Expansion) N

Attending 4-year colleges

Men and women

Men

Women

Men and women

Men

Women

.021 (.043) .003 (.046) .070  (.039) .107  (.058) .085 (.069) .022 (.073)

.014 (.041) .029 (.044) .052 (.040) .084 (.059) .114 (.071) .030 (.075)

.065 (.146) .111 (.142) .131 (.099) .112 (.164) .014 (.178) .126 (.198)

.019 (.042) .170*** (.033) .094** (.033) .198*** (.057) .077 (.048) .121* (.060)

.007 (.041) .087* (.035) .088* (.035) .120  (.069) .095  (.051) .025 (.071)

.090 (.127) .435*** (.085) .208* (.081) .206 (.140) .073 (.111) .279* (.137)

1100

749

351

1814

1202

612

Note: 1. The samples for 2-year college attendance include high school graduates and 2-year college attendees only (714 4-year college graduates are excluded). 2. The estimates are computed from 500 simulations of random drawings. 3. Matching weights for balancing family background covariates between two cohorts are applied. 4. Age dummies and residential areas at age 14 are included. Their estimates are omitted but available upon request.   p < .1. * p < .05. ** p < .01. *** p < .001.

to a relatively small sample size (N = 612). This divergent pattern between men and women suggests that the impact of college expansion on earnings differs by gender, with women benefiting the most. Table 5 presents the results from the analyses in which we consider the postsecondary institutions separately by type: junior colleges and 4-year colleges. The results show that the previous pattern for attending any postsecondary institutions is largely driven by 4-year colleges rather than junior colleges, for both men and women. The returns for junior colleges are much smaller than the returns for 4-year colleges, approximately by half (11% versus 20%). More specifically, male compliers of 2-year colleges gained 8% in total owing to their shift from no college to a junior college, compared with 11% of their AT counterparts’ gain between 1998 and 2009. The human capital effect is, therefore, insignificantly negative. Female compliers, however, earned about 11% of earnings gain from their added education (human capital effect) with an ignorable skill price effect, benefiting 13% in total from the expansion. Such a gendered pattern in coefficient estimates is consistent with the previous finding from attending any college, just with a relatively smaller scale and statistical insignificance. We suspect that relatively small sizes of these 2-year analysis subsamples (749 men and 351 women) are partly responsible for such indifferences in statistical significance. The economic returns for 4-year college compliers also show a consistent pattern with the results from overall and junior colleges. Male CP’s gain is about 12%, which is roughly similar to the gain of AT, 10%. Therefore, 12% of the total gain from the expansion of 4-year institutions is decomposed into a 10 percentage point skill price effect with marginal significance at the

S. Choi / Social Science Research 50 (2015) 229–245

Economic Returns to College Expansion

240

0.35 0.30

0.31

CP (Total Effect)

0.28

AT (Skill Price Effect) 0.25

0.24

CP-AT (Human Capital Effect)

0.21

0.20 0.15

0.12

0.12 0.12

0.10 0.05

0.10 0.03

0.00

0.00 -0.05 -0.10

Any College

4-Year College Men

-0.07

-0.07

Any College

4-Year College

Women

Fig. 4. Economic returns to college expansion. Note: All the estimates are earnings returns to college expansion relative to returns among the NT group. The bold italic numbers indicate statistical significance at the .05 level and plain italic numbers indicate statistical significance at the .1 level.

.1 level, and a 2 percentage point human capital effect that is far short of statistical significance. Female compliers, however, gained a 21% total earnings return that consists of a moderately negative skill price effect component (7 percentage point) and a significantly sizable human capital effect component (28 percentage point). Fig. 4 visually presents the results from the analyses of attending any college and attending 4-year college, in which the full samples are analyzed. We summarize our findings in the light of our theoretical discussions. First, the resulting pattern of the human capital effect suggests that the positional good hypothesis is supported for men whereas, for women, the human capital hypothesis better predicts. For men, a moderate total gain of compliers fails to exceed significantly the gain of alwaystakers. The human capital effect is very small. That is, added education for compliers failed to promote their earnings, which suggests education worked as a positional good. On the contrary, female compliers gained exceedingly much, compared with always-takers who underwent a small loss in earnings. Such a large and significant difference between compliers and always-takers evinces a substantial human capital effect. Thanks to such a large human capital effect, female compliers could narrow the gap from always-takers. Second, the estimates of the skill price effect provide evidence that the college premium in earnings rose for men and stagnated for women. Indeed, our auxiliary estimation of the changes in college premium (not shown) shows about 10 (4-year college) to 12 (any college) percent of a significant increase in premium for men and about 2 (4-year college) to 1 (any college) percent of a statistically negligible change in premium for women. This result suggests that men enjoyed the skill-favored labor market changes, but women did not. Rather, for women, the large gain resulting from the compliers’ human capital was offset by always-takers’ loss probably resulting from the increasing supply of college graduates. 6.3. Robustness analyses There are two major sources inducing bias for our estimates; first, the estimates can be biased when there are unobserved confounders affecting both college decision-making and earnings. This is a traditional selection bias, which emerges when the ignorability assumption – one of the key assumptions we make – does not hold. In the context of the estimation of the returns to education, it can be interpreted as ability bias (Card, 1999). The possibility of the presence of ability bias cannot be ruled out particularly in our case because we were not able to include any indicators of academic ability such as grades and test scores due to data unavailability. To address the concern of this source of bias, we reanalyze the main DID models after adding a hypothetical confounder that is correlated with college attendance (attending any college and attending a 4-year college) and earnings. The confounder may be considered a measure of omitted ability. We vary the correlation of the confounder with college outcome and earnings and see whether doing so yields significantly different estimates of CP’s earnings gain. To deliberately consider more realistic correlations, we take advantage of the fact that the KLIPS collected information on the score of the Scholastic Aptitude Test (SAT) from a small subsample. We use that information to calibrate the more realistic correlations of SAT score with college outcomes and with earnings in the subsample. The correlations are .15 to .2 (with earnings), .16 to .24 (attending any college), and .42 to .47 (attending a 4-year college).9 The results of this robustness analysis with the hypothetical confounder are reported in Table D of the online supplementary document. To summarize the result of this robustness analysis, our estimates are largely robust to the unmeasured ability confounder on the regions of those empirically calibrated correlations. 9 We residualized SAT score level on birth year to make the score invariant across birth cohorts. The estimated correlations provide upper bounds because the SAT score is correlated with other family SES covariates and, if the test score is residualized on the family SES covariates, the correlations would be lower than the currently estimated ones.

241

S. Choi / Social Science Research 50 (2015) 229–245 Table 6 The estimates of earnings returns to expansion alternative scenarios. Main estimate (Table 4)

1997 Earnings (A)

Selection into employment (B)

(A) + (B)

Any college CP  Expansion (vs. NT  Expansion) CP  Expansion (vs. AT  Expansion)

.179 (.063)⁄⁄ .073 (.051)

.151 (.063)⁄ .042 (.052)

.164 (.064)⁄ .079 (.052)

.148 (.064)⁄ .059 (.052)

4-Year college CP  Expansion (vs. NT  Expansion) CP  Expansion (vs. AT  Expansion)

.198 (.057)⁄⁄⁄ .121 (.060)⁄

.196 (.058)⁄⁄ .095 (.060)

.201 (.058)⁄⁄⁄ .124 (.060)⁄

.193 (.058)⁄⁄ .106 (.061) 

The second possible mechanism is that the South Korean economic crisis in 1998 may induce a certain bias that consequently affects our estimates. The South Korean labor market was hit hard by the economic crisis in 1998, which is the year when the earnings of the pre-expansion cohort were measured. Therefore, the earnings level for that year might have been unusually low and the impact of the crisis differed by education.10 The unemployment rate also soared drastically in 1998. If those unusual conditions occurred differentially across educational groups (which is quite likely), our estimates could be biased. To address those possibilities, we conduct two additional analyses. To start, we estimate the main DID model with the 1998 earnings adjusted to the 1997 distribution. We create a weight from the official Korean labor statistics of earnings in 1997 and 1998 by sex, industry, and education such that the applied weight deflates the 1998 earnings to the 1997 earnings level for each sex-industry-education group. We expect that this adjustment removes, or at least attenuates the impact of the economic crisis on the earnings distribution among the pre-expansion cohort, and, consequently, minimizes the bias. Following that, we adjust for the possible different selection into employment due to the economic crisis, applying a weight that balances the pre- and the post-expansion cohorts with regard to sex, age, and employment status.11 This analytic weight is expected to balance a possible different selection into employment between earnings measured in 1998 and 2009. Table 6 shows the estimated earnings gains of the complier groups under the different scenarios of bias. Most of those estimates do not yield significantly different results from the original findings. One minor difference is that the size of the earnings return to 4-year college expansion net of the skill price effect diminishes slightly below the level of statistical significance when applying the 1997 weight. The result, however, suggests that our estimates are generally robust to certain plausible impacts of the economic crisis. Finally, we conduct a supplementary analysis of the main models using a score from the socioeconomic status index of occupations (SEI) (Ganzeboom and Treiman, 1996) as our outcome replacing earnings. Occupational status is a good alternative measure of attained socioeconomic position. Unlike earnings, which is an attribute of an individual worker, SEI characterizes the status of an occupation. Thus, we can expect to disentangle different mechanisms, working at different levels (individual or occupational) by comparing the results from our earnings analyses with the results from the analyses of occupational status. The results from the occupation status models are presented in Fig. 5 (see Table E in the online supplement for the estimation table). The result can be summarized as two major patterns. First, for both genders, the human capital effect is greater for 2-year college CP than for 4-year college CP. Both male and female compliers of 2-year college expansion experienced significantly positive gains, net of the skill price effect (9 and 11 points in SEI), but, among 4-year college compliers, only men earned a significantly positive gain in occupational status (5.2 points). Second, for both men and women, the AT of 2-year colleges experienced a decline in their occupational status during the period of college expansion (a negative skill price effect). The AT of 4-year colleges, however, upgraded their occupation. In the case of men, such an upgrade was large enough to be statistically significant (4.5 points). These results are quite different from those of the earnings models in many aspects. They provide a clue for a better understanding of a gendered pattern in earnings returns. Male compliers experienced large significant growth in occupational status but with little growth in earnings. This discrepancy suggests a shift from a large between-occupation earnings inequality to an increasing inequality in earnings within the group of occupations sharing a similarly high level of skill requirement. Indeed, such decoupling implies that the earnings return of male compliers might be best understood at an individual level rather than at an occupation-level. Out of the positional good perspectives, signaling theory, which offers an individual-level explanation, might be at play particularly for 4-year college compliers12; in post-expansion, employers

10 According to the labor statistics of Korea, for men aged 25–34, high school graduates without any college degree underwent a 3.4% decrease in monthly wages while the wages of college graduates decreased only by .3%. For women with the same age range, the trend is opposite; high school graduates experienced an increase in wages by 1.5% but college graduates suffered a 0.9% decrease in wages. 11 We account for sex, age, and employment status for two main reasons. First, the economic crisis forced many college graduates to postpone starting their working careers voluntarily or involuntarily and selection into employment among the young graduates under the crisis, in 1998, could differ from that in more normal labor market condition such as in 2008. Second, in South Korea, for women staying in the labor force becomes harder, especially among women aged over thirty because of family formation. But this difficulty works differently not only across educational groups but also between the two periods. The result from matching shows that employment by education becomes balanced after matching. 12 The results from the male samples in Table 5 show that the pre-expansion college premium in earnings was significantly large for 4-year institutions (.087, p < .05), not for junior colleges (.029). This suggests that we can apply the signaling/sorting explanation only to the expansion of 4-year colleges.

S. Choi / Social Science Research 50 (2015) 229–245

Socioeconomic Index of Occupation

242

CP (Total Effect) AT (Skill Price Effect) CP-AT (Human Capital Effect)

15.0

10.0 7.2

6.9

8.5

11.0 9.0

7.8

6.5

4.5

4.0

5.0

9.7

5.2 4.1 2.7 1.4

0.6

0.3

0.0

-4.5

-5.0

-5.0

Men

Women

Any College

Men

Women

2-Year College

Men

Women

4-Year College

Fig. 5. Occupational status returns to college expansion. Note: All the estimates are earnings returns to college expansion relative to returns among the NT group. The bold italic numbers indicate statistical significance at the .05 level and plain italic numbers indicate statistical significance at the .1 level.

offering high-skill occupations hire compliers but find that they are less productive than always-takers because of ability difference. In response, they offer lower earnings to compliers than to always-takers by allocating them to less rewarding work. Female compliers, in contrast, experienced a large earnings increase without a comparable occupational upgrade. That is, for women, the college expansion as augmented human capital was effective on earnings (human capital hypothesis), but not on occupation-level socioeconomic status (positional good hypothesis). This type of discrepancy requires more occupational-level explanations than an individualistic account. First, the earnings gain for female compliers is more likely to be explained by an overall improvement in earnings of female-dominated jobs or occupations likely thanks to promoted productivity or prestige (e.g., service sector). Second, in a more general occupational context, women could be more successful than before due to their upgraded human capital or skill, although staying in occupations with similar level of status or prestige. The positional good theories suggest that female compliers’ failure to upgrade occupationally might be the consequence of occupation-level barriers such as credentialism or social closure, whereas male compliers’ failure to earn more may be attributed to ability sorting mechanisms at the individual level. Indeed, male compliers did not suffer from social or cultural barriers their female counterparts faced. Some female compliers seem to be the major beneficiaries of the college expansion, but our results also suggest that they could have had higher earnings returns to college expansion if they had not undergone any practice of credentialism. 7. Discussion In this study, we estimated the causal effect of college expansion on earnings using the example of the 1990s South Korean policy experience. We found evidence that the effect strikingly differs by gender. For men, the total economic benefit from college expansion is moderate, and most of the gain is due to the skill price effect, which is driven by increasing demand for skilled labor. Female compliers, however, earned considerably more in total from college expansion, and the gain is mostly attributable to the human capital effect. For both men and women, such patterns are largely driven by the benefit from 4-year college expansion, while the returns to the expansion of junior colleges was quite limited. Various checks for robustness of the results suggest that our findings hold with altering assumptions. There are a couple caveats. The first is about the ages at which we measured earnings. The age range is from 27 to 33. It only covers the earliest years of a labor market career. It is meaningful to see individuals’ earnings at this early stage because part of an educational decision may depend on short-term foreseeable expectations and there is also a high level of uncertainty both for employers and job-seekers at this stage. Nevertheless, earnings trajectories are known to diverge between different educational groups over their careers. Thus, future research investigating whether our current finding holds over the later career stages is expected to follow. Second, our identification of the skill price effect is based on the assumption that compliers received the same quality of higher education as always-takers did, therefore, compliers could become a perfect substitute for always-takers. This assumption may not be true especially in the South Korean context. It is reasonable to expect that postsecondary institutions newly established during the expansion provide less qualitative education than existing institutions. Considering the highly stratified system of higher education in South Korea (Park, 2007), new institutions are also unlikely to belong to selective colleges. If compliers could only get lower quality education compared with always-takers, the skill price effect based on the return for always-takers is overestimated, and the human capital effect is underestimated. However, available empirical

S. Choi / Social Science Research 50 (2015) 229–245

243

evidence does not strongly support such a concern. First, according to the time-series statistics on postsecondary institutions (Ahn, 2011), in terms of several major indicators measuring college quality, such as student-faculty ratio, the number of library books per student, and the overall expense invested per student, new institutions are not particularly lower than existing ones.13 Second, our supplementary analysis, where we estimate the returns to college expansion excluding the graduates of selective 4-year colleges, does not yield a different result (see Table C in the supplementary document for the estimated result).14 This suggests that selective colleges, which tend to provide a higher quality service than non-selective competitors, do not result in particularly skillful workers gaining higher earnings. Considering a relatively stable proportion of those who went to selective colleges and, in contrast, a substantially increasing number of non-selective college attenders due to college expansion, it is suggestive that a quality difference between institutions compliers attended and those always-takers attended might not be a critical issue. Our finding of a remarkable gender difference, especially a contrasting pattern of decoupling of earnings returns from occupational status returns, deserves more discussion. The finding suggests a narrowing gender gap in earnings and a persistent gap in occupational status during the period of college expansion. Previous research on gender inequality in the Korean labor market generally confirms such a trend in gender differences. Indeed, a gender wage gap has been narrowing measurably over the past few decades, although the gap remains considerable (Kim, 2009; Lee and Kim, 2010). Researchers pay attention to two reasons; first, an increasing share of women and increasing productivity in service sector contributed to a decreasing difference in wages between men and women (Ihm, 2010; Lee and Kim, 2010); second, the persistent occupational gender segregation slows down the narrowing trend of a gender gap (Lee and Kim, 2010). Our analysis adds to this body of research a finding that college expansion played a significant role in this process. That is, female compliers’ increased human capital seems to have contributed to women’s higher proportion in the service sector and higher productivity and, ultimately, a decline in the gender gap in earnings. However, it does not mean that women underwent an economic advancement enough to be comparable to men due to their upgraded higher education; rather it might reflect the low level of earnings among low-skilled female workers before the expansion. On the other hand, a relatively low occupational status return for female compliers, than for their male counterparts, suggests that college expansion in South Korea was not successful in alleviating gender segregation at the occupational level. More generally, we believe that our findings link the literatures of occupational gender segregation (e.g., Charles and Grusky, 2004) and the recent studies on female-favoring gender gap in education in advanced societies (e.g., DiPrete and Buchmann, 2013). It is also of great importance to think about the generalizable implications of our findings, especially from a policy perspective; do our major findings suggest that a policy of massive expansion in higher education is worth for reducing inequality and promoting economic mobility? Our two competing hypotheses correspond to the competing predictions about whether pre-expansion heterogeneity between those who were near the margin of the college-going propensity and those who were considerably above the margin will persist (positive selection) or decline (negative selection) after the expansion. In spite of recent findings suggesting a negative selection (Brand and Xie, 2010; Hout, 2012), empirical evidence is still inconclusive. Our study, based on a direct pre–post cohort comparison, offers evidence that both predictions could be true simultaneously. As the KLIPS men showed, in a case where a massive expansion of higher education, from a minority to the majority, the returns for people at the margin may not be greater than for those with higher propensities. However, for historically disadvantaged groups, such as South Korean women, a sharp increase in participation in higher education can lead to a dramatic increase in human capital for those at the margin and provide an opportunity to catch up with the always-taker competitors. This leads us to two policy implications. Firstly, educational expansion per se may fail to narrow the socioeconomic gap in the labor market even when it successfully reduces inequality in higher education. Indeed, it may widen the gap because compliers pay additional costs for the educational investment that they otherwise would not have to spend, especially in societies with highly privatized postsecondary education system such as South Korea (Kirk, 2011) and the US (Archibald and Feldman, 2010). The large gap between the estimated total return to the expansion and the return net of the skill price effect confirms such a concern. A college expansion may help its compliers earn more, but, only with proper changes in skill demand in the labor market. In this vein, the major lesson of our study is that an expansion policy should be designed with a view to labor market conditions. A simple shift in educational attainment at the aggregate level, without an accompanying increase in labor market demand, does not always promise high economic mobility and growth. Second, the impact of educational expansion is contingent on historical and social contexts. More specifically, policies aiming at expanding education may be effective particularly to certain groups that have been disadvantaged in the labor market for social and historical reasons even if those policies are not rewarding for general compliers. The evidence from the South Korean women demonstrates such a case. A large and significant economic benefit from joining higher education can be fully understood with the phenomenon of ‘‘the rise of women’’ in South Korea, and, more generally, other advanced countries including the US (DiPrete and Buchmann, 2013). Our finding implies that educational expansion can be an effective instrument for promoting pre-existing social disadvantages among social minorities in the labor market such as South Korean women. 13 We compared the average values of several college quality measures for institutions that were established after 1992 with those of all postsecondary institutions. As of 2000, the number of students per faculty was slightly higher for new institutions than the overall average (33.2 versus 24.5). So were the number of library books per student (83.5 versus 44.9) and the overall expense invested per student (5,735,000 versus 4,758,000 in Korean won). 14 For selective institutions, I consider fifteen top-ranked colleges and universities with regard to scholastic aptitude test score.

244

S. Choi / Social Science Research 50 (2015) 229–245

Acknowledgments I thank Richard Breen, Vida Maralani, Olav Sorenson, Sara Goldrick-Rab, Matt Lawrence, and Laura Rangner for their insightful and encouraging comments. I am also grateful to two anonymous reviewers for their thoughtful and productive feedback. The earlier version of this paper was presented at RC28 meeting at the University of Virginia in August 2012. I am also thankful to the participants of the conference for their helpful comments.

Appendix A. Supplementary material Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.ssresearch.2014.11.014. References Ahn, M.-S., 2011. The Statistics of Changes in Higher Education: From 5.31 Educational Policy Reform, 1995 to 2010. Korea Research Institute of Higher Education, Seoul, South Korea (in Korean). Allison, P.D., 2002. Missing Data. Sage Publications, Thousand Oaks, CA. Angrist, J.D., Pischke, J.-S., 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press, Princeton, NJ. Archibald, R.B., Feldman, D.H., 2010. Why Does College Cost So Much? Oxford University Press. Attewell, P.A., Lavin, D.E., 2007. Passing the Torch: Does Higher Education for the Disadvantaged Pay Off across the Generations? Russell Sage Foundation, New York. Becker, G.S., 1993. Human Capital: A Theoretical and Empirical Analysis, with Special Reference to Education, third ed. The University of Chicago Press, Chicago, IL. Berman, E., Machin, S., 2000. Skill-biased technology transfer around the world. Oxf. Rev. Econ. Policy 16, 12–22. Berman, E., Bound, J., Machin, S., 1998. Implications of skill-biased technological change: international evidence. Quart. J. Econ. 113, 1245–1279. Bills, D.B., Brown, D.K., 2011. New directions in educational credentialism. Res. Soc. Stratif. Mobil. 29, 1–4. Brand, J.E., Xie, Y., 2010. Who benefits most from college? Evidence for negative selection in heterogeneous economic returns to higher education. Am. Sociol. Rev. 75, 273–302. Breen, R., 2010. Educational expansion and social mobility in the 20th century. Soc. Forces 89, 365–388. Brown, D.K., 2001. The social sources of educational credentialism: status cultures, labor markets, and organizations. Sociol. Educ. 74, 19–34. Card, D., 1999. The causal effect of education on earnings. In: Ashenfelter, O.C., Card, D. (Eds.), Handbook of Labor Economics. Elsevier, Amsterdam, pp. 1801–1863. Card, D., Lemieux, T., 2001. Can falling supply explain the rising return to college for younger men? A cohort-based analysis. Quart. J. Econ. 116, 705–746. Carneiro, P., Heckman, J.J., Vytlacil, E.J., 2011. Estimating marginal returns to education. Am. Econ. Rev. 101, 2754–2781. Chang, S., 2009. The analysis of 5.31 university policy: focusing on deregulation. Trend Prospect 77, 9–49 (in Korean). Charles, M., Grusky, D.B. (Eds.), 2004. Occupational Ghettos: The Worldwide Segregation of Women and Men. Stanford University Press, Stanford, CA. Choi, K.-S., 1996. The impact of shifts in supply of college graduates: repercussion of educational reform in Korea. Econ. Educ. Rev. 15, 1–9. Choi, K.-S., Jeong, J., 2005. Technological change and wage premium in a small open economy: the case of Korea. Appl. Econ. 37, 119–131. Choi, K.-S., Jeong, J., 2007. Does unmeasured ability explain the wage premium associated with technological change?: quantile regression analysis. Appl. Econ. 39, 1163–1171. Choi, K.-S., Jeong, J.-H., Jung, H.J., 2005. The rising supply of college graduates and declining returns for young cohort: the case of Korea. Glob. Econ. Rev. 34, 167–180. Collins, R., 1979. The Credential Society: An Historical Sociology of Education and Stratification. Academic Press, New York, NY. Devereux, P.J., Fan, W., 2011. Earnings returns to the British education expansion. Econ. Educ. Rev. 30, 1153–1166. DiPrete, T.A., Buchmann, C., 2013. The Rise of Women: The Growing Gender Gap in Education and What It Means for American Schools. Cambridge University Press. Galor, O., Moav, O., 2000. Ability-biased technological transition, wage inequality, and economic growth. Quart. J. Econ. 115, 469–497. Ganzeboom, H.B.G., Treiman, D.J., 1996. Internationally comparable measures of occupational status for the 1988 international standard classification of occupations. Soc. Sci. Res. 25, 201–239. Gebel, M., Pfeiffer, F., 2010. Educational expansion and its heterogeneous returns for wage workers. Schmollers Jahrb. 130, 19–42. Goldin, C.D., Katz, L.F., 2008. The Race between Education and Technology. Harvard University Press, Cambridge, MA. Gurgand, M., Maurin, E., 2007. A Large Scale Experiment: Wages and Educational Expansion in France (Working Paper No. 2007-31). Centre de Recherche en Economie et Statistique. Hannum, E., Buchmann, C., 2005. Global educational expansion and socio-economic development: an assessment of findings from the social sciences. World Dev. 33, 333–354. Harmon, C., Hogan, V., Walker, I., 2003. Dispersion in the economic return to schooling. Labour Econ. 10, 205–214. Hout, M., 2012. Social and economic returns to college education in the United States. Annu. Rev. Sociol. 38, 379–400. Iacus, S.M., King, G., Porro, G., 2012. Causal inference without balance checking: coarsened exact matching. Polit. Anal. 20, 1–24. Ihm, J., 2010. Analysis of gender wage discrimination by educational level and occupational type. J. Korean Womens Stud. 26, 39–61 (in Korean). Imbens, G.W., Angrist, J.D., 1994. Identification and estimation of local average treatment effects. Econometrica 62, 467–475. Juhn, C., Murphy, K.M., Pierce, B., 1993. Wage inequality and the rise in returns to skill. J. Polit. Econ. 101, 410–442. Katz, L.F., Murphy, K.M., 1992. Changes in relative wages, 1963–1987: supply and demand factors. Quart. J. Econ. 107, 35–78. Kim, Y.-M., 2009. A distributional approach to the changes in gender wage gap in Korea, 1982–2004. Econ. Soc., 206–229 (in Korean). Kim, C.-U., Lim, G., 2012. Social returns to college education: evidence from South Korean college education. Appl. Econ. Lett. 19, 1537–1541. Kirk, D., 2011. South Korean students protest rising college tuition. Christ. Sci. Monit.. Lange, F., Topel, R., 2006. The social value of education and human capital. In: Hanushek, E., Welch, F. (Eds.), Handbook of the Economics of Education. Elsevier, Amsterdam, pp. 459–509. Lee, S., Kim, Y.-M., 2010. The effect of expanding service industry on gender gap in wages in South Korea. Korean J. Sociol. 44, 1–25 (in Korean). Maurin, E., McNally, S., 2008. Vive la Révolution! Long-term educational returns of 1968 to the angry students. J. Labor Econ. 26, 1–33. Moretti, E., 2004. Estimating the social return to higher education: evidence from longitudinal and repeated cross-sectional data. J. Econom. 121, 175–212. Morgan, S.L., Winship, C., 2007. Counterfactuals and Causal Inference: Methods and Principles for Social Research. Cambridge University Press, New York, NY. Park, H., 2007. South Korea: Educational Expansion and Inequality of Opportunity for Higher Education. In: Stratification in Higher Education: A Comparative Study. Stanford University Press, Stanford, CA, pp. 87–112.

S. Choi / Social Science Research 50 (2015) 229–245

245

Psacharopoulos, G., 2009. Returns to Investment in Higher Education: A European Survey (No. Report for the Higher Education Funding Reform Project, CHEPS-led Consortium for the European Commission). European Commission, Brussel, Belgium. Psacharopoulos, G., Patrinos, H.A., 2004. Returns to investment in education: a further update. Educ. Econ. 12, 111–134. Shavit, Y., Arum, R., Gamoran, A., Menahem, G. (Eds.), 2007. Stratification in Higher Education: A Comparative Study. Stanford University Press, Stanford, CA. Spence, M., 1973. Job market signaling. Quart. J. Econ. 87, 355–374. Taber, C., 2001. The rising college premium in the eighties: return to college or return to unobserved ability? Rev. Econ. Stud. 68, 665–691. Tsai, S.-L., Xie, Y., 2008. Changes in earnings returns to higher education in Taiwan since the 1990s. Popul. Rev. 47, 1–20. Tsai, S.-L., Xie, Y., 2011. Heterogeneity in returns to college education: selection bias in contemporary Taiwan. Soc. Sci. Res. 40, 796–810. Walker, I., Zhu, Y., 2008. The college wage premium and the expansion of higher education in the UK. Scand. J. Econ. 110, 695–709. Weiss, A., 1995. Human capital vs. signalling explanations of wages. J. Econ. Perspect. 9, 133–154. Willis, R.J., Rosen, S., 1979. Education and self-selection. J. Polit. Econ. 87, S7–S36.