China’s College Expansion and the Timing of the College-to-Work Transition: A Natural Experiment

Abstract

This article examines the effect of China’s 1999 acceleration of higher education expansion on when college graduates find their first skilled job. We use a natural experiment to test our hypotheses and exploit the unique education and work history data of a nationally representative survey, as well as estimate a causal inference model. We find that the 1999 education expansion caused a delay in the landing of a skilled job among graduates from technical colleges, while graduates from four-year colleges were not affected in job acquisition. We also find that family origins and individual social positions are significant determinants of who entered college both before and after the education expansion. These findings shed new light on the workings of early adulthood and on social inequality in China.

Keywords

China college graduate higher educational expansion skilled employment

The turn of the new millennium witnessed an accelerated higher education expansion in China. In 1998, the number of new college admissions was 1.08 million; that number reached 6.08 million in 2008, about a six-fold growth (China Ministry of Education 1999–2015). Scholars consider a key motivation for this higher education expansion to be the high unemployment rates among high school graduates who were competing for jobs with those who had been laid off from state-owned enterprises in the late 1990s (Knight, Deng, and Li 2017). China’s college-educated millennials face a complex array of opportunities and vulnerabilities after this higher education expansion that reshaped the transition to work for college graduates in the 2000s compared to those who completed college in the 1990s. We conceive the macro-factor-driven difference between two successive cohorts’ experience as a natural experiment of the 1999 college expansion, an exogenous shock, to China’s young adult population. We evaluate the causal effect of this college expansion on job acquisition of college graduates. This study may shed light on our understanding of universal early adulthood in newly developed countries in Asia and around the world (Yeung and Alipio 2013).

The multifold increase in college enrollment rates in the 2000s offered opportunities for Chinese youth to accumulate human capital. Simultaneously, the acceleration of higher education expansion in a short period caused a series of repercussions. First, demand for skilled labor in the then labor-intensive, export-oriented economy did not grow at the same speed as the supply of skilled labor. Second, universities were not ready to serve a massive student body, contributing to graduates of lower quality. Third, university curricula and training programs did not prepare college graduates to execute work responsibilities with additional training. These consequences spurred an increase in unemployment rates among college graduates (Bai 2006; China Statistics Bureau 1990–2015).

The unemployment problem among college graduates may be viewed within the context of the time it takes graduates to find an appropriate job. The college education rate in China’s labor force is low, at 17.4 percent (China Statistics Bureau 1990–2015). And a sudden increase in labor supply may decrease one’s success in finding a job. Therefore, rather than asking whether college graduates are able to find skilled jobs, we ask how long it takes college graduates to find a skilled job. To answer this question of timing, we exploit the unique education and work history data of a nationally representative survey (China Labor-Force Dynamics Survey) to estimate a causal inference model for the effect of the 1999 expansion on the timing of the college-to-work transition.

College Expansion and Macro Conditions for the College-to-Work Transition

We consider China’s college expansion to be a plausible explanation for the emerging “early adulthood” in the course of China’s transforming economy. Furstenberg (2010) identifies factors that cause later transitions to adulthood, including the expansion of education, consistent with China’s situation in the early 2000s. China’s 1999 college expansion targeted both four-year universities and three-year technical colleges nationwide. While mass college education adds three to four years to early adulthood, determining whether the expansion of education causes a delay in the acquisition of skilled jobs requires rigorous testing, which is the primary objective of this article.

The event of interest is mass college expansion with its inception in 1999, as well as the resultant increase in the skilled labor supply in the initial five years of expansion from 2003 to 2007. For comparison, we examine the skilled labor supply at college graduation from 1995 to 1999—the last five years before mass college education.¹ Below we introduce the 1999 expansion and describe macro-level conditions that were stable over the two periods and societal events that may or may not have changed the condition between the two periods.

The 1999 higher education expansion

In 1999, the Chinese government began to accelerate the pace of higher education expansion. As a result, the number of new college admissions in 2008 reached 6.08 million, about six-fold of the 1.08 million in 1998; this expansion claimed to be the fastest in the world (China Ministry of Education 1999–2015). A closer look into the admission periods corresponding to the two periods of college graduation of 1995 to 1999 and 2003 to 2007 shows that college admissions increased slowly (6.2, 7.5, 9.2, 9.0, 9.3) from 1991 to 1995 and rapidly (16.0, 22.1, 26.8, 32.1, and 38.2) from 1999 to 2003. Thus, we take the 1999 higher education expansion as a macro factor driving the oversupply of skilled labor in the second period.

Higher education curricula

China’s higher education system in the early expansion years lacked diversity. The concept of the job market did not have a foothold in colleges and universities because college graduates were assigned to state-sector jobs before the 1990s (Maurer-Fazio 1999). During the expansion, a shortage of professors, outdated curricula, and inefficient practice programs remained or worsened (He and Mai 2015). The lack of competitiveness, applied abilities, specialty, and flexibility remained the same over both periods.

Labor-intensive, export-oriented economy

The continuous labor-intensive, export-oriented economy in the 1990s and 2000s did not demand a rapid growth of skilled workers. The structure of China’s industries created a limited capacity to absorb a rapid increase of university graduates. It was projected that more than one million graduates would be unlikely to find employment (Qian 2000). The official employment rates in 2003 to 2004 were 84 percent for four-year college graduates and 61 percent for those with fewer than four years of college education (China Ministry of Education 1990–2015).

In summary, the period of 1995 to 1999 before the college expansion and the period of 2003 to 2007 since the first class of graduates under this expansion shared stable macro conditions of rigid higher education curricula on the supply side and a labor-intensive economy on the demand side. The two periods differ markedly in the skilled labor supply due to the accelerated expansion of higher education in 1999. Using a natural experiment framework, we define higher education expansion as the treatment and the timing of the transition from college to the first skilled job as the outcome. The cohort of youth who graduated from college during the second period was exposed to the treatment, and the cohort of youth who graduated from college during the first period was not exposed to the treatment. The two cohorts of graduates were not randomly assigned but were not self-selected. In other words, the timing of landing a skilled job after college graduation should be independent of the treatment assignment after controlling for individual characteristics. We hypothesize that higher education expansion would increase the waiting time for college graduates who graduated during the accelerated expansion to find a skilled job (hypothesis 1).

To make this natural experiment credible, we had to control for individual characteristics: the structure-driven micro attributes of individuals that influence their chances of attending college. Controlling for these attributes enabled us to ensure a similar marginal distribution of individuals in the two cohorts, thus making the cohorts comparable. In the next section, we identify those attributes.

Micro Attributes of Sorting into College

We applied theories of educational stratification and inequality and cumulative advantage to the China case and identified structure-driven micro attributes in sorting youth into college going in China. In China, family origins are first and foremost captured by household registration (hukou) status—rural versus urban. Individuals acquire this household registration status at birth, and a person’s hukou status in large measure determines their educational and employment opportunities and has a significant influence over his or her longer-term life chances (Hao, Hu, and Lo 2014; Wu and Treiman 2004). In addition, parental education indicates class position, and parental party membership indicates political capital (Li 2003). In the context of accelerated higher education expansion from 1999, we consider youth born to parents under the restrictive one-child policy that dramatically reduced the number of siblings and increased familial resources allocation to children (Ye and Wu 2011).

In China, rural and urban secondary education systems are segregated, and rural education is inferior to urban education. At the college entrance exams, rural students are less competitive than urban students are, as a result of cumulative effects from the inferior educational quality throughout primary and secondary school. In addition, educational quality differs by region. Children living in the East, Central, and West regions experience different educational quality throughout childhood. The Han ethnic majority has a long history of having better educational opportunities than ethnic minorities. Thus, rural hukou status at birth, Central and West regions of residence during childhood, and being a non-Han minority mean persistently lower educational quality that lowers the probability of college attendance.

To understand how persistent lower educational quality affects the types of college attended, the cumulative advantage thesis reviewed in DiPrete and Eirich (2006) offers further rationales. The cumulative nature of educational attainment (i.e., a higher-level attainment is built on a lower-level attainment) suggests that the initial advantage of the privileged social groups at lower levels of education will persist throughout educational transitions. The mechanisms include compound returns to earlier human capital stock and access to increasingly greater resources brought about by earlier higher achievement, including social capital and network connections. In China, the cumulative advantage of privileged students (e.g., urban hukou at birth, higher parental socioeconomic status, parental party membership, fewer or no siblings, Han majority, living in the East) leads to a greater likelihood of admission to a four-year college, whereas disadvantaged students end up in three-year technical colleges.

In addition to family origins, individual attributes also matter in making the two cohorts comparable. In China, gender and party membership stand out as important individual characteristics. Women’s increased educational attainment, a trend in developed countries, has also occurred in China (Hannum and Xie 1994). We expect that being male would be less important in the probability of attending college for the postcohort than the precohort. The political capital accrued to party members applies to both parental and own party membership (Li 2003) and remains important after the acceleration of education expansion.

These micro attributes provide additional insight into our primary question regarding the accelerated expansion of higher education and its effect on waiting time before landing a skilled job after college graduation. The “battle” in college entrance exams yield a trichotomous outcome: four-year college admission, three-year college admission, and no college going. In China, technical college education receives less investment and development opportunities than four-year college education, and access to technical versus four-year college education differs by structural exclusion (Tam and Jiang 2015). Combined with the sorting mechanism discussed above, we anticipate that the student body of the two types of colleges differs in family origins and individual characteristics. Both the lower educational quality in technical colleges and the lower social status of technical college students imply that the prospect of in-time transition from college to skilled employment may diverge by the type of college. Thus, we hypothesize that the causal effect of the 1999 accelerated expansion of higher education on the time it takes graduates to find their first job is concentrated among graduates from technical colleges (hypothesis 2).

Data, Samples, and Variables

Data source

Our data source is the China Labor-Force Dynamics Survey (CLDS), a nationally representative, longitudinal survey of labor force in China (excluding Hainan, Tibet, Hong Kong, and Macao). Every two years since 2012, the CLDS follows individuals, families, and communities to update the cross-sectional representation of the population. The sampling design is multistage, stratified, and proportional to size with rotation groups. This study draws data from the 2014 CLDS containing four rotation groups, including 23,594 individuals aged 16 to 64 or older if working.

The CLDS provides unique data needed for the current study on both education history and work history.² The large sample size also guarantees a sufficient size of the two college graduate cohorts defined by the current study. We use retrospective data on education and work history rather than the panel data because the panel data were collected later than the periods of interest.³

Analytic samples

This study entails two analytic samples from the 2014 data. Using the auxiliary sample, we examine family origins and individual characteristics in affecting the probability of college going before and after the expansion. We consider a typical college graduation age range of 21 to 23 for four-year college graduates and 20 to 22 for three-year college graduates (with a one-year margin of error). The auxiliary sample includes 6,246 individuals in total: the precohort sample includes individuals born between 1972 and 1979 who could potentially graduate between 1995 and 1999 (n = 3,363 with 421 college graduates); the postcohort sample includes individuals born between 1980 and 1987 who could potentially graduate from college between 2003 and 2007 (n = 2,783 with 626 college graduates).

The primary sample includes only college graduates, with a purpose of estimating the causal effect of the 1999 expansion. The treatment group in the natural experiment includes 626 individuals who graduated from college during the period of 2003 to 2007; the control group of the natural experiment includes 421 individuals who graduated from college during the period of 1995 to 1999.

Multiple imputation

Missing data are ubiquitous in survey data. While the absence of all substantive variables in our two types of samples is moderate (18 percent and 17 percent, respectively), we use a multiple imputation method to maintain the population multivariate structure of the data and of the analytic sample size. Under the missing-at-random assumption conditional on the variables in the analysis model, we apply the imputation method with chained equations, including not only substantive variables but also variables describing the missing mechanism (such as interviewer quality and respondent quality) to reduce the dependence on only substantive variables in imputation (Rubin 1987). We create twenty complete datasets for the auxiliary and primary sample respectively.⁴ The results reported in the article are a summary of parameters obtained from twenty completes using Rubin’s rule (taking the mean point estimates as the final point estimates and using the squared-root of the total variance to combine the between and within variance as the standard error), except those specifically mentioned.

Dependent variables

To model college attendance, we define the dependent variable as college attendance in full-time four-year or three-year colleges leading to a degree conferred during 1995 to 1999 or 2003 to 2007. Individuals with postgraduate degrees are included; individuals who obtained continuous education degrees (while holding full-time employment) are not considered in the same population, given the different curricula of continuous education and their complicated work histories.

Two dependent variables of our primary analysis of the causal effect of the 1999 accelerated expansion of higher education are finding a skilled job within half a year of graduation and within 1.5 years of graduation, conditional on not yet finding a skilled job. Data required to construct these two dependent variables include birth year, educational attainment, and both education history and work history. First, we treat the reported birth year and educational attainment as precise information. Second, the graduation year is based on the self-reported graduation year within the typical three-year range of graduation, and, if missing, a randomly chosen year from the three-year range of graduation. Third, the year starting a skilled job comes from work history data. The CLDS 2014 collected the total number of jobs with information on the starting and ending and occupation of three specific jobs: the current or recent job, the first job if having more than one job, and the job right before the current/recent job if having more than two jobs.⁵ The majority (93 percent) of individuals in our primary sample (aged 27–42 in 2014) had no more than three jobs by 2014 with consistent work histories. Fourth, skilled job acquisition for four-year college graduates takes either 1 if the occupation code indicates managerial, technical, and administrative jobs in an office or 0 otherwise. The skilled jobs for three-year college graduates include administrative occupations in security and communication. Fifth, using the notion of hazard (i.e., the conditional probability of finding a skilled job given that the event has not occurred), we consider skilled job acquisition during the graduation year⁶ and the year following the graduation year.

Treatment

The treatment of our natural experiment is the 1999 accelerated expansion of higher education. The treatment assignment is a dichotomous variable: 1 indicates those who graduated from college in 2003 to 2007 (postcohort herein) and 0 indicates those who graduated from college in 1995 to 1999 (precohort herein).

Control variables

The control variables are of two sets. The first set is to model the probability of college education. Sorting into college depends on family origins, variables capturing childhood experience, and individual characteristics. Family origins are measured with hukou status at birth (1 denotes rural hukou, 0 denotes urban hukou), parental education (the highest years of parents’ schooling), number of siblings (top numbers are coded at 10), and parental party membership (1 denotes either parent is a party member, 0 denotes no party membership of parents). Childhood experience is captured by region of residence at age 14 (Central and West with East as the reference) and Han majority (1 = yes, 0 = no). Individual characteristics include gender (1 = male, 0 = female) and party membership (1 = yes, 0 = no). We also include a measure of conscientiousness, a tendency to show self-discipline, act dutifully, and aspirations for achievement, at age 14, a composite based on three items (Cronbach’s alpha = .82). The second set is to make the treatment and control groups of the natural experiment comparable. We add on to the first set the rank of college (3 = national, 2 = provincial, and 1 = local), which is available for college graduates only.

Analytic Strategies

Exploiting natural experiments in causal inference with observational data has gained increasing popularity. Natural experiments refer to shocks that create exogenous variations in the phenomenon of interest (Angrist, Imbens, and Rubin 1996). In our context, the natural experiment is the government-directed, top-down command of accelerated expansion of higher education nation-wide that increased admissions to college. Knight, Deng, and Li (2017) argue that the increased supply of college graduates may be endogenous to past and expected future demand, and yet the suddenness, unexpectedness, speed, and size of the supply shock means that it might create a natural experiment for analyzing short-term labor market consequences.

We place the higher education expansion in a counterfactual framework: the treatment is the accelerated expansion of higher education that has increased the number of admissions and graduates since 1999. The outcome of interest is the timing of landing a skilled job. With constructed event history data, we investigate the conditional probability (hazard) of skilled employment within 0.5 and 1.5 years after graduation, respectively, conditional on the risk set that treats post–college study as a competing risk.

In our study of college graduates, the threat of selection bias is of specific concern. We consider how the two sources of selection bias are eliminated with observed covariates. The first source is sample selection of those receiving college education among age-appropriate individuals. From Heckman’s correction for a sample selection model (Heckman 1979), we create “inverse Mill’s ratio,” a variable in a nonlinear function of the observed and unobserved variables in the sample selection model. We estimate the inverse Mill’s ratio for the precohort and postcohort separately to take into account different higher education policies and practices in the preexpansion and postexpansion eras in addition to individual characteristics. The estimated inverse Mill’s ratios for individuals is entered into the causal model for the accelerated expansion to reduce the potential bias in estimates due to sample selection of college graduates.

The second source of selection refers to the process of assigning individuals to the treatment versus control groups, that is, the selection into the precohort versus the postcohort. One approach to remedy the threat of this selection bias is propensity score matching (Rosenbaum and Rubin 1983) and an extension using inverse probability weighted regression adjustment (Wooldridge 2007, 2010). These methods use many characteristics of individuals and fully specify statistical models with those measures.

The third source of bias from unobserved variables is still probable, even with the extensive use of statistical controls. The extensive array of control variables including family origin, past experience, and current characteristics may reduce this probability.

We perform the inverse probability weighted logistic regression adjustment for the hazard of finding a skilled job within the graduation year or by the end of the year after the graduation year for the primary sample and then separately for bachelor’s degrees and technical degrees with the primary sample. Keeping all and only observations at risk of finding a skilled job at t = 0.5 or t ≤ 1.5 (collapsed to one record per person), we estimate logistic regression for the hazard of finding a skilled job:

\log (\frac{h_{i t}}{1 - h_{i t}}) = β D_{i} + δ ’ z_{i}

(1)

To test hypothesis 1, we estimate the average treatment effect for the treated controlling for covariates, $\hat{β}$ , for the primary sample. If this estimate is negative and significant, it will support hypothesis 1.

To test hypothesis 2, we obtain the same estimate of the average treatment effects for the treated separately for the bachelor’s degrees and the technical degrees. If the average treatment effect for the treated is concentrated in the technical degree, then it supports hypothesis 2.

Results

Sample descriptive distributions

Table 1 shows descriptive distributions for the auxiliary sample by cohorts and those for the primary sample of college graduates by cohorts and types of degrees. Examining the cohort differences in the auxiliary sample, we see that, compared to the precohort, the postcohort is characterized by a larger proportion of college graduates, higher parental education, lower parental party membership, a smaller number of siblings, and more from the East. These trends are consistent with China’s development in general, as well as the expansion of higher education, in particular, over the turn of the new millennium.

Table 1

Sample Descriptive Distribution

			Primary Sample
	Auxiliary Sample		Total		Bachelor’s		Technical
Variable	Precohort	Postcohort	Precohort	Postcohort	Precohort	Postcohort	Precohort	Postcohort
College education	0.13	0.23	—	—	—	—	—	—
Skilled job within graduation year	—	—	0.30	0.25	0.35	0.30	0.25	0.20
Skilled job within 1.5 years	—	—	0.37	0.34	0.42	0.38	0.30	0.29
Technical degree	—	—	0.45	0.45	—	—	—	—
Rural hukou at birth	0.82	0.84	0.44	0.56	0.43	0.46	0.45	0.67
Parental years of schooling	5.93	7.52	9.47	10.00	9.93	10.59	8.90	9.27
Parental party membership	0.15	0.12	0.34	0.24	0.33	0.28	0.35	0.19
Number of siblings	2.73	1.88	1.80	1.03	1.69	0.89	1.93	1.20
East region at age 14	0.37	0.43	0.39	0.49	0.39	0.42	0.38	0.58
Central region at age 14	0.29	0.29	0.29	0.30	0.31	0.35	0.27	0.23
West region at age 14	0.33	0.28	0.32	0.21	0.30	0.23	0.35	0.19
Han majority	0.88	0.86	0.91	0.95	0.92	0.94	0.89	0.95
Male	0.47	0.45	0.55	0.49	0.57	0.50	0.52	0.48
Own party membership	0.08	0.07	0.36	0.24	0.44	0.35	0.26	0.10
Conscientious scale at 14	–0.01	–0.04	0.22	0.17	0.29	0.19	0.14	0.14
High college rank	—	—	0.23	0.20	0.33	0.30	0.09	0.07
Mid college rank	—	—	0.40	0.39	0.43	0.50	0.35	0.25
Low college rank	—	—	0.38	0.42	0.23	0.20	0.56	0.68
Inverse Mill’s ratio	—	—	1.11	0.91	1.00	0.74	1.25	1.13
n	3,363	2,783	421	626	232	343	189	282

NOTE: Estimates are based on multiply imputed complete datasets using Rubin’s rule.

The next panel of Table 1 shows the distribution of variables among college graduates in the primary sample. For the total sample of college graduates (columns 3 and 4), the postcohort shows a lower probability of finding a skilled job within half a year or 1.5 years following graduation, consistent with our expectation. The proportion of those with a technical degree remains similar. The primary sample and the auxiliary sample show a similar trend in the cohort comparison in parental party membership, number of siblings, and proportion of eastern region residence at age 14. Notable differences for the postcohort of the college graduate sample include more graduates with at-birth rural hukou, more from the Han majority, fewer men, and lower party membership. These postcohort features for college graduates reflect the changing selection to college education after the accelerated expansion of higher education. For unique covariates of college graduates, the greater proportion of low-ranked local universities suggests where the expansion was mostly located.

When we disaggregate further by types of degree (bachelor’s and technical in columns 4–8), the precohort and postcohort comparisons are more telling. In particular, the differential distributions for at-birth rural hukou and low college rank reveal further information on where and for whom the accelerated expansion of higher education was concentrated. While the percentage of at-birth rural-hukou increases by 3 percentage points among those with bachelor’s degrees, the increase is 21 percentage points among those with technical degrees—almost all students with origin of rural hukou went to three-year colleges. Furthermore, the proportion in low-rank colleges actually decreases from .23 to .20 for the bachelor’s degree holders; in contrast, the number increases from .56 to .68 for the technical degree holders. These stark contrasts by hukou and college rank strongly suggest that the causal effect of education expansion may diverge between the two types of degrees.

Selection into college education

The purpose of the analysis of selection into college education, separately for the precohort and the postcohort, serves two purposes. First, the selection model creates inverse Mill’s ratios to be included in the causal analysis to correct for the sample selection bias. Second, the estimation helps to check the validity of the control variables for meeting the conditional independence assumption of the treatment assignment and the outcome. (See the estimates from the probit model of college education in appendix Table A1.) To gauge the relative importance of the control variables (see appendix Figure A1), we find the importance of (1) at-birth rural hukou, number of siblings, parental education, and parental party membership; (2) for the postcohort exclusively, residing in the Central region at age 14 and Han majority; and (3) own party membership and conscientiousness. It is interesting to note that gender no longer plays a role in determining college education for the postcohort.

Inverse probability weighted distribution

The inverse probability weighting method starts by matching propensity scores, followed by obtaining the inverse probability to weight the data such that the assignment to the two cohorts is independent of the outcome. To show the inverse probability weighting, we randomly chose one version of the twenty complete datasets from multiple-imputation. We estimate a logit model for the assignment to the precohort and postcohort as a function of the control variables discussed in the analytic strategy section. The empirical distribution suggests that bachelor’s degree holders and technical degree holders are of two different subpopulations and should be analyzed separately.

Appendix Figure A2 shows the propensity score distribution of the postcohort (cohort = 1) superimposed on that of the precohort (cohort = 0) for the technical degree and bachelor’s degree. The two plots are similar in that the masses fall in the overlapped region (0.05–0.85 for the technical and 0.1–0.95 for the bachelor’s). The inverse probability weighting is to make the distributions of all covariates similar between the two cohorts (balanced). In the after-weighting overlap plot for the number of siblings in appendix Figure A3, we find a nice overlap for the technical degree, but the result appears not as satisfactory for the bachelor’s degree.

Covariate balance statistics include the difference in a standardized covariate mean before and after inverse probability weighting and the variance ratio. The differences in the standardized means are reduced to 0, and the variance ratios are converged to 1 for both the technical degree and the bachelor’s degree, as shown in Table 2. Overall, the covariates in the two cohorts are balanced after weighting, as the overidentification test does not reject the null that all covariates are balanced between the two cohorts after inverse probability weighting for both types of degrees.

Table 2

Covariate Balance Statistics of the Inverse Probability Weighting

	Technical Degree				Bachelor’s Degree
	Standardized Differences		Variance Ratio		Standardized Differences		Variance Ratio
Variable	Raw	Weighted	Raw	Weighted	Raw	Weighted	Raw	Weighted
Rural hukou at birth	0.45	−0.05	0.89	1.04	0.05	−0.06	1.01	0.99
Parental years of schooling	0.10	−0.01	0.68	0.78	0.20	−0.05	0.67	0.82
Parental party membership	−0.37	0.00	0.68	1.00	−0.10	−0.03	0.92	0.97
Number of siblings	−0.53	0.04	0.52	1.15	−0.60	0.01	0.56	1.17
Central region at age 14	−0.11	0.04	0.88	1.05	0.09	0.06	1.06	1.05
West region at age 14	−0.36	0.05	0.68	1.08	−0.15	−0.07	0.85	0.93
Han majority	0.23	−0.02	0.48	1.08	0.08	0.06	0.76	0.81
Male	−0.09	0.04	1.00	1.00	−0.14	0.05	1.02	1.00
Own party membership	−0.40	0.01	0.49	1.01	−0.19	0.04	0.92	1.03
Conscientious scale at 14	0.02	0.08	1.13	0.78	−0.15	−0.01	0.99	0.86
High college rank	−0.11	−0.04	0.73	0.88	−0.07	0.04	0.94	1.04
Mid college rank	−0.18	0.09	0.85	1.12	0.12	−0.06	1.02	1.00

NOTE: Analysis is based on one of the twenty multiply imputed complete datasets.

Results of causal analysis

Table 3 compiles the estimates of inverse probability weighted logistic regression adjustment for the causal effect of the 1999 accelerated expansion of higher education on timing of finding a skilled job within half a year or 1.5 years of college graduation. The top panel records estimates for technical degrees, while the bottom panel displays estimates for bachelor’s degrees. Focusing on the average treatment effect for the treated (ATT), substantial negative coefficients are found for the technical degree, which is significant at the .05 level for the half a year timing and at the .10 level for the 1.5 years timing, controlling for the additional influence of control variables (hence regression adjustment). The magnitude of the negative effect is substantial with the odds ratio at e^–0.17 = 0.844 and e–^0.142 = 0.868. That is, the odds of finding a skilled job for the postcohort individuals with a technical degree is 84 percent of the odds for the precohort counterparts within half a year of graduation. The odds for postcohort individuals only improve a little at 87 percent within 1.5 years of graduation. This result supports hypothesis 1 in that the causal effect is negative. In contrast, the ATT for the bachelor’s degree subsample is close to zero, statistically nonsignificant, and positive. This striking contrast in causal effect provides evidence in support of hypothesis 2, which states that the causal effect diverges between the two types of degrees. This is the core finding of our casual analysis.

Table 3

Estimates of Inverse Probability Weighted Logistic Regression Adjustment: Two Timings of Finding a Skilled Job

Variable	Within 0.5 Years of Graduation		Within 1.5 Years of Graduation
	Coef.	SE	Coef.	SE
Technical degree
ATT	−0.170	0.086**	−0.142	0.084*
Potential outcome	0.370	0.082	0.432	0.079
D = 0
Rural hukou at birth	1.561	0.866*	1.682	0.771**
Number of siblings	−0.049	0.215	0.012	0.197
Male	−0.881	0.574	−0.502	0.522
Own party membership	−0.814	0.964	−0.661	0.851
High college rank	0.350	1.010	−0.072	0.921
Mid college rank	−0.268	0.614	−0.150	0.601
Inverse Mill’s ratio	−1.719	1.071	−1.538	0.902*
Constant	0.879	1.143	0.639	1.006
D = 1
Rural hukou at birth	1.475	0.527**	1.541	0.493**
Number of siblings	−0.202	0.205	−0.174	0.165
Male	−0.203	0.361	−0.345	0.312
Own party membership	−0.176	0.765	−0.476	0.716
High college rank	0.915	0.671	1.118	0.564**
Mid college rank	0.966	0.396**	0.616	0.389
Inverse Mill’s ratio	−0.681	0.590	−0.668	0.528
Constant	−1.679	0.557**	−1.067	0.516
Treatment model (postcohort = 1)
Rural hukou at birth	1.083	0.237***	1.083	0.237***
Parental years of schooling	0.044	0.034	0.044	0.034
Parental party membership	−0.707	0.256**	−0.707	0.256**
Number of siblings	−0.412	0.091***	−0.412	0.091***
Central region at age 14	−0.226	0.268	−0.226	0.268
West region at age 14	−0.422	0.273	−0.422	0.273
Han majority	−0.111	0.411	−0.111	0.411
Male	−0.317	0.213	−0.317	0.213
Own party membership	−0.739	0.289**	−0.739	0.289**
Conscientious scale at 14	−0.035	0.145	−0.035	0.145
High college rank	−0.217	0.405	−0.217	0.405
Mid college rank	−0.035	0.255	−0.035	0.255
Constant	0.774	0.568	0.774	0.568
Bachelor’s degree
ATT	0.004	0.056	0.047	0.057
Potential outcome	0.295	0.051	0.330	0.051
D = 0
Rural hukou at birth	−0.282	0.556	−0.590	0.544
Number of siblings	0.243	0.171	0.180	0.166
Male	−0.455	0.350	−0.255	0.345
Own party membership	0.729	0.678	0.868	0.629
High college rank	0.438	0.513	0.596	0.460
Mid college rank	0.457	0.456	0.345	0.431
Inverse Mill’s ratio	0.237	0.817	0.746	0.723
Constant	−1.572	0.940	−1.712	0.859
D = 1
Rural hukou at birth	0.072	0.373	0.384	0.359
Number of siblings	0.108	0.121	0.023	0.119
Male	−0.179	0.243	−0.383	0.231
Own party membership	0.244	0.423	−0.181	0.402
High college rank	0.286	0.367	0.128	0.342
Mid college rank	0.253	0.333	0.283	0.306
Inverse Mill’s ratio	0.285	0.482	−0.025	0.462
Constant	−1.416	0.476**	−0.617	0.439
Treatment model (postcohort = 1)
Rural hukou at birth	0.698	0.208**	0.698	0.208**
Parental years of schooling	0.042	0.030	0.042	0.030
Parental party membership	−0.283	0.214	−0.283	0.214
Number of siblings	−0.524	0.084***	−0.524	0.084***
Central region at age 14	0.209	0.212	0.209	0.212
West region at age 14	−0.039	0.241	−0.039	0.241
Han majority	−0.210	0.409	−0.210	0.409
Male	−0.336	0.185*	−0.336	0.185*
Own party membership	−0.206	0.187	−0.206	0.187
Conscientious scale at 14	−0.150	0.135	−0.150	0.135
High college rank	−0.116	0.255	−0.116	0.255
Mid college rank	0.187	0.239	0.187	0.239
Constant	0.768	0.576	0.768	0.576

NOTE: Estimates are based on multiply imputed complete datasets using Rubin’s rule.

p < .10. **p < .05. ***p < .001.

After balancing the covariates through inverse probability weighting, are there any additional effects on the timing of skilled job acquisition in the logistic regression adjustment? Because the covariates are balanced, multicollinearity and nonconvergence would occur if all covariates were to be included in the regression adjustment. Substantively, we keep the social position variables (at-birth rural hukou, party membership, male gender, college rank); inverse Mill’s ratio for correcting the sample selection bias; and number of siblings, which captures a potential shift in worldviews and work values among individuals, many of whom are the only child. We examine the results for the technical and bachelor’s degree subsamples in turn. First, at-birth rural hukou stands out with its significant positive effect on the timing of finding a skilled job. In contrast to the strong negative effect of this same variable in sorting youth into college education (see appendix Table A1), the positive effect of at-birth rural hukou on finding a skilled job suggests that rural-hukou college students after successfully graduating from college are more competitive in the job market. Second, sporadic positive coefficients are found for the high-rank and midrank college levels as expected. Third, other social position variables (party membership and gender) have nonsignificant coefficients. Fourth, the nonsignificance of number of siblings suggests that the postcohort, which was predominantly without siblings, shows little shift in worldview and work value with respect to finding a skilled job. Finally, the inverse Mill’s ratio is significant in only one of four situations, perhaps due to the balanced covariates. Turning to the estimates for the bachelor’s degree, no additional covariate effects are statistically significant in the regression adjustment.

Finally, we examine the treatment assignment model results. For the technical degrees, at-birth rural hukou, parental party and own party membership, and number of siblings must be controlled to balance the covariates. For the bachelor’s degrees, only at-birth rural hukou and parental party membership must be stringently controlled, but gender becomes marginally significantly negative, indicating that there is a larger percentage of women in the postcohort compared to the precohort.

Sensitivity analysis

We report one sensitivity analysis to address the common problem of a lack of precise timing of event histories when data are drawn from surveys. As discussed in the data section, we have taken systematic measures to make the results robust. For example, we executed the following measures: (1) focusing on full-time college students to avoid complicated work histories of part-time students, (2) using birth years to define cohort membership taking into account the typical three-year range of graduation year, (3) treating self-reported graduation year as the precise timing if it falls in the typical graduation range, (4) applying randomness when assigning the graduation year within the three-year range when precise graduation year is unavailable, (5) allowing the timing of the first skilled job to be shortly before the graduation year when comparing the graduation year with the job starting year, (6) comparing the graduation timing and job start timing one year at a time, and (7) applying the same rule for both cohorts.

Our sensitivity analysis imparts the following question: if we delay the typical graduation range by one year (i.e., ages 22–24 for bachelor’s degrees and ages 21–23 for technical degrees), will the results remain predominantly unchanged? This sensitivity analysis shows similar patterns as the current analysis, including the negative, substantial effects for the technical degrees and close-to-zero, nonsignificant effects for the bachelor’s degree. Specifically for the technical degrees, the ATT for the timing within a half a year is −.127 (e^−.127 = 0.763) with t = −1.608, and the timing within 1.5 years is −.180 (e^−.127 = 0.835) with t = −2.462. This sensitivity analysis suggests the robustness of our current findings.

Conclusion

This article, motivated by a need for macro-level explanations for the emerging delayed transition from college to skilled employment in rapidly developing societies, uses a natural experiment in China to examine the causal effect of the country’s 1999 expansion of higher education on the timing of skilled job acquisition among college graduates. Applying development and social stratification theories to China at the turn of the new millennium, we provide testable hypotheses that are both general to theory and specific to China’s situation. To test these hypotheses, we set up a rigorous approach to a natural experiment, including defining the treatment and assigning the treatment and control groups, identifying covariates for control to meet the conditional independence of the treatment assignment and the outcome, and measuring the outcomes in terms of the timing of the college-to-work transition. The analysis exploits the unique education and work history data of a nationally representative survey and estimates a causal inference model. The results provide evidence to support the posed hypotheses.

The current analysis is limited in the precise college graduation time and job starting time, a common issue with using survey event history data due to self-report and recall errors, especially when the educational history timing and the work history timing are compared. We have developed systematic ways to maximize the use of the data while minimizing its limitation. Sensitivity analysis with a slight change in the typical graduation age range provides similar findings, supporting the robustness of our current findings.

With this caveat, we offer three major findings. First, the 1999 accelerated expansion of higher education caused a delay in timing for landing a skilled job among graduates from technical colleges, while graduates from four-year colleges were not affected. Second, with the exception of gender, family origins, social positions affecting childhood experience, and individual social positions remain significant in the selection into college going both before and after the acceleration of education expansion. Third, above and beyond its negative effect on college education sorting, at-birth rural hukou status exerts a positive effect on the timing of finding a skilled job among technical college graduates.

These findings shed new light on early adulthood and social inequality. Similar to other developing societies, China witnessed emerging early adulthood as a consequence of the accelerated expansion of higher education. Unlike in other societies, however, China has seen an emerging divide in early adulthood dictated by types of degrees in the college-educated population. A source of this divide are the institutions themselves—the historically lower quality of instruction in technical colleges in comparison to that found in four-year colleges, as well as the further widening of the quality gap during the expansion, accounts for emerging early adulthood. Another related influence is rapid innovation in technology in today’s digital age. Technical colleges have not fully adapted to these updated technologies. Further, students of lower family origins, lower social positions during childhood, and lower individual social positions are often sorted into technical colleges. These social, structural disadvantages continue to operate in the college-to-work transition, placing technical college graduates at a disadvantage.

Our findings suggest two policy recommendations. First, given the need for quality higher education, increases in government investment in three-year and four-year colleges should be made to advance both the quality of education and increase the capacity to instruct an influx of students. Second, college admission policy, particularly the quota allocations to provinces, alongside other structural barriers that reproduce and exacerbate educational inequality, should be eliminated to maximize the human capital and talent of China’s millennials.

Footnotes

Appendix

Table A1

Estimates from Probit Model of College Education

Variable	Precohort		Postcohort
	Coef.	SE	Coef.	SE
Rural hukou at birth	−0.86	0.07***	−0.882	0.081***
Parental years of schooling	0.09	0.01***	0.084	0.010***
Parental party membership	0.15	0.08*	0.267	0.094***
Number of siblings	−0.14	0.02***	−0.204	0.027***
Central region age 14	−0.03	0.08	−0.181	0.075**
West region at age 14	0.13	0.09	−0.094	0.084
Han majority	0.02	0.12	0.268	0.127**
Male	0.12	0.07*	0.006	0.064
Own party membership	1.18	0.09***	1.339	0.114***
Conscientious scale at 14	0.21	0.04***	0.269	0.042***
Constant	−1.20	0.17***	−0.776	0.179***
n	3,363		2,783

NOTE: Estimates are based on multiply imputed complete data sets using Rubin’s rule.

p < .10. **p < .05. ***p < .001.

NOTE:

This study was supported by the Hopkins Population Center under a small grant (PI: Lingxin Hao; R24 HD042854). The findings and conclusions in this article are those of the authors and do not necessarily represent the official position of the National Institutes of Health.

Notes

Lingxin Hao is a professor of sociology at Johns Hopkins University. Her specialties include social inequality, sociology of education, migration, family and public policy, and quantitative methodology. Her research has appeared in the American Journal of Sociology, Demography, Social Forces, Sociology of Education, and Child Development, among others.

Dong Zhang is a lecturer in sociology at Chongqing Technology and Business University in China.

References

Angrist

Joshua D.

Imbens

Guido W.

Rubin

Donald B.

1996. Identification of causal effects using instrumental variables. Journal of the American Statistical Association 91 (434): 444–55.

Bai

Limin

. 2006. Graduate unemployment: Dilemmas and challenges in China’s move to mass higher education. China Quarterly 185:128–44.

China Ministry of Education. 1990–2015. Educational statistical yearbook of China. Available from http://edu.tom.com (accessed 15 May 2018).

China Statistics Bureau. 1990–2015. China statistical yearbook, chain labor statistical yearbook. Beijing: China Statistics Press.

DiPrete

Thomas A.

Eirich

Gregory M.

2006. Cumulative advantage as a mechanism for inequality: A review of theoretical and empirical developments. Annual Review of Sociology 32:271–97.

Furstenberg

Frank F.

Jr . 2010. On a new schedule: Transitions to adulthood and family change. The Future of Children 20 (1): 67–87.

Hannum

Emily

Xie

Yue

. 1994. Trends in educational gender inequality in China: 1949–1985. Research in Social Stratification and Mobility 13:73–98.

Hao

Lingxin

Alfred

Jamie

. 2014. Two aspects of the rural-urban divide and educational stratification in China: A trajectory analysis. Comparative Education Review 58 (3): 509–36.

Mai

Yinhua

. 2015. Higher education expansion in China and the “ant tribe” problem. Higher Education Policy 28 (3): 333–52.

10.

Heckman

James J.

1979. Sample selection bias as a specification error. Econometrica 47 (1): 153–61.

11.

Knight

John

Deng

Quheng

Shi

. 2017. China’s expansion of higher education: The labor market consequences of a supply shock. China Economic Review 43:127–41.

12.

Chunling.

2003. Socio-political changes and inequality of educational opportunity. Social Sciences in China. Available from http://en.cnki.com.cn/Article_en/CJFDTOTAL-ZSHK200303007.htm (accessed 15 May 2018).

13.

Maurer-Fazio

Margaret

. 1999. Earnings and education in China’s transition to a market economy: Survey evidence from 1989 and 1992. China Economic Review 10 (1): 17–40.

14.

Qian

Xianying

. 2000. The idea of using education to stimulate economy is too subjective. In On China’s educational policy, ed. Zhenguo

Yuan

, 190–91. China: Educational Science Press.

15.

Rosenbaum

Paul R.

Rubin

Donald B.

1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70 (1): 41–55.

16.

Rubin

Donald B.

1987. Multiple imputation for nonresponse in surveys. New York, NY: Wiley.

17.

Tam

Tony

Jiang

Jin

. 2015. Divergent urban-rural trends in college attendance: State policy bias and structural exclusion in China. Sociology of Education 88 (2): 160–80.

18.

Wooldridge

Jeffrey M.

2007. Inverse probability weighted estimation for general missing data problems. Journal of Econometrics 141 (2): 1281–1301.

19.

Wooldridge

Jeffrey M.

2010. Econometric analysis of cross section and panel data. 2nd ed. Cambridge, MA: MIT Press.

20.

Xiaogang

Treiman

Donald J.

2004. The household registration system and social stratification in China: 1955–1996. Demography 41 (2): 363–84.

21.

Hua

Xiaogang

. 2011. Fertility decline and the trend in educational gender inequality in China. Sociological Study 155 (5): 153–77.

22.

Yeung

Wei-Jun Jean

Alipio

Cheryll

. 2013. Transitioning to adulthood in Asia: School, work, and family life. The ANNALS of the Academy of Political and Social Science 646 (1): 6–27.