“Universities and the Entry-Level Job Market: Evidence from Japanese
Panel Data,” by Yukiko Abe
Data Appendix
1. The Sources of the Selectivity Variables and the Normalization Procedure
Kawai-Juku’s selectivity data are obtained from “Daigaku Nyushi Ranking Hyo” (University Entrance Exam Ranking Table) from 1980 to 1991. Kawai-Juku’s standardized test scores are calculated on the basis of its nationwide model exam (Zenkoku Toitu Moshi, Nationwide Standard Model Exam) and the results of entrance exams in the following winter.
Obunsha’s selectivity variable is obtained from the following issues of Keisetsu Jidai published by Obunsha. As explained in the text, the selectivity variable is calculated from analyzing the model exam scores and outcomes of real entrance exams. The numbers in the square brackets indicate the corresponding years of entrance exams on which the calculation of the published selectivity is based.
1981-April [1980], 1982-April [1981], 1983-April [1982], 1984-April [1983], 1985-April [1984], 1986-April [1985], 1988-July Dai-1 Furoku (attachment 1) [1988], 1989-July Dai-1 Furoku (attachment 1) [1989], 1990-July Dai-1 Furoku (attachment 1) [1990], 1991-July Dai-1 Furoku (attachment 1) [1991]
The selectivity variable in the sample is based on the standardized test score (hensachi, in Japanese). For the analysis here, the second stage examination score is used when more than one scores (like the score for the first stage examination) are reported. The standardized test score is calculated by normalizing the distribution of raw test scores by a distribution of mean 50 and standard deviation 10. The normalization makes it possible to know the individual’s location in the test score distribution (Kawai-Juku’s scores are reported by 2.5 point interval, and the lowest portion is truncated at 37.4. I assign 32.62 to those cases, which is the truncated mean of a normal distribution of mean 50 and variance 100). The scores at individual schools are based on scores of the subjects in the entrance exams, so the same level of test score at schools with different exam subjects does not necessarily imply the same level of ability. However, it is possible to apply to most private schools by taking exams in three subjects (usually English, Japanese and one from math or social studies), so in most cases, applicants for private schools usually study 3 subjects for the entrance exams. Many schools provide multiple admission categories, and the borderline standardized score for each category may differ. In the sample used in the analysis, I take the midpoint of the standardized test score in case the borderline scores differ within university-department-year.
To make the standardized test scores comparable across years, I make the following adjustments. The basic idea is to assume ability distribution are the same across cohorts, while the movement in enrollment rates makes the ability distribution of college students (more precisely, that of individuals who apply to college) to change over time. Since standardized test scores are based on individuals who apply to colleges and scores are normalized to have mean 50 and variance 100 every year, one point of test score in a year and another year may reflect different degrees of ability difference. Therefore, actual test scores have to be adjusted to account for enrollment rates. I take the procedure explained below to make the adjustments.
I
assume population ability distribution follows a normal distribution of mean μ and variance σ2, which is stable across cohorts. I further assume that, for each cohort, those who are at the
upper tail of the ability distribution go to colleges. The ability distribution of college
applicants is obtained by truncating the population ability distribution at the
point of college enrollment rate.
The mean and variance of the truncated normal distribution is 
and 
, where αt is determined by the relation
and λis the hazard function of the normal distribution. I solve for μ and σ2 for the cohort enrolled in 1980, by
fitting the truncated mean and variance to be 50 and 100, respectively, at the α that corresponds to the enrollment rate in 1980. The raw scores from other years are
deflated to 1980 value by using μ, σ2 and enrollment rate for each year. Enrollment rate is obtained from Digest of Education
Statistics (Ministry of Education).
2. Company Ranking Data
The data for company rankings are from the following issues of Diamond Weekly (Shukan Diamond), published by Diamond:
1983/4/16, 1984/4/14, 1985/4/6, 1986/4/5, 1987/4/4, 1988/4/2, 1989/7/29, 1990/7/7, 1991/7/6, 1992/7/4, 1993/7/3, 1994/6/25. The company ranking for the cohort graduating in year t is published in the edition in year t-1.
3. School Characteristics Data
The number of graduates and the number of female students among graduates for each university-department-cohort are obtained from the printed version of “Daigaku-betu Shushokusaki Shirabe” (Employment Outcomes by Universities) by Recruit Research Corporation, for cohorts graduating 1984-1992. They are obtained from 1994, 1995 and 1996 volumes of “Daigaku Soran Data Book” (University Information Data Book) by Recruit Research Corporation, for cohorts graduating 1993-1995. These are used to construct female ratio variable, cohort size dummy variables and student-teacher ratio.
The number of full-time faculty members is obtained from “Zenkoku Daigaku Shokuin Roku” (University Personnel Directory), published by Kojunsha. I gathered information in years 1980, 1985, 1990 and 1994 (the year here corresponds to the academic year, which starts in April and ends in March. So the cohort graduating in March of 1984 after 4 years of college attend academic year 1980-1983). Since the number of teachers is gathered for only four years, the exact number of teachers for each cohort at each university-department is imputed. Specifically, for each year other than 1980, 1985, 1990 and 1994, I assign weighted average of these as the number of teachers: for example, the number of teachers in 1981 is imputed as 0.2*(the number of teachers in 1980)+0.8*(the number of teachers in 1985). Then, I assign the number of teachers for each cohort as the four-year average of the number of teachers during each cohort’s college attendance period. The student-teacher ratio variable is constructed as (the number of graduates of the cohort)*4/(the average number of teachers for the cohort). The numerator is multiplied by 4 to make the variable comparable to the number of enrollment divided by the number of teachers in four-year universities.
The metropolitan dummy takes value 1 if a school is located in one of the following prefectures: Saitama, Chiba, Tokyo, Kanagawa, Aichi, Kyoto, Osaka, Hyogo. It takes value zero otherwise.
4. BSWS Sample Selection (this part is for the results based on industry wage differentials, which is not included in the final version of the article)
Sample used for estimating wage regression is taken from the Table 2 of Volume 1 and 2 (Dai-2 Hyo of Dai-1 kai and Dai-2 kan,) of the Basic Survey of Wage Structure in 1994 (Ministry of Labor). In this table, the unit of observation is the average wage by gender-education-industry-age-tenure group (aggregated over firm size and occupation category). In order to estimate a wage regression, I select the sample male four-year college graduates. The sample is restricted to males since for many medium-level industry codes, wage data for women are unavailable. The wage equation is estimated by taking the log of monthly wage (defined as (12*(scheduled monthly cash earnings)+bonus)/12) and regress it on 25 industry dummies, age, age squared, tenure and tenure squared. The coefficients of industry dummies are used as diffi in equation (2) of the text.
5. Matching the Data Sets
(1) Identifying company names that appears in the company ranking
In the employment data by the Recruit Research Corporation, there are observations for which company names are missing. However, for most companies that appear on the company rankings, I was able to identify the corresponding observations in the employment data. This check was done by using both the company names and the company codes. Specifically, I checked the possibility that some of the companies in the ranking list did not hire anyone or hired very few people from the universities the Recruit Research Corporation surveyed. The number of companies in the ranking list I was not able to identify in the employment data is 7 in 1984, but it is less than 3 for years 1985-1995 (these numbers include Takashimaya, the company explained (3) below).[1] This enables me to obtain the probability of employment in the top ranking companies as accurately as possible.
(2) Identifying Public Sector Employees
In the analysis, the number of people who started to work in the public sector is counted as the number of graduates who started in the industry code “public sector (Kanko-cho)”. This category includes the central and the local governments, as well as governmentally-funded organizations (such as the Japan External Trade Organization). Also, companies that used to be publicly-owned but are now privatized were included in the public sector when they were publicly-owned. Some of the governmentally-funded organizations appear in the top ranking companies list, even though the central and the local governments are not included in the list. Therefore, the numerator of the PROB_200jt (including the public sector) is counted as the sum of the number of students who were employed in the top 200 companies and the public sector employees minus those who were employed by governmentally-funded organizations that appears in the company ranking list. This category is omitted in calculating industry-based wage differentials, since public employees are not surveyed in the BSWS.
(3) Missing Data Problems
In the magnetic tape of the employment data, the observations for the Takashimaya Corporation (a large department store) are missing for all universities. This company appeared on the top 200 list many times. Because the employment information is missing, the PROB_200jt measure does not reflect the information of Takashimaya. However, it is unlikely that one company has large effects on PROB_200jt.
(4) Reclassification of Industry Codes (this part is for the results based on industry wage differentials)
In order to match the industry wage information with the employment data, it is necessary to reassign the industry codes of the employment data. The procedure I use is explained below. The employment data have the industry name and the industry code variables. Basically, I match the industry name coded in the employment data to the corresponding industry in the BSWS. There are three problems in matching the two data sets based on industry name, however. First, the industry classification used in the employment data differs from that used in the BSWS wage data. Second, the industry classification system in the employment data changed in 1994. Finally, the published version of the BSWS wage data does not report wage information for some of the detailed industries. Therefore, I reclassify the observations in the employment data in the following way.
(4-1) I use major industry codes in the BSWS for the following industries: mining, construction, estate, wholesale/retail, service, transportation and utility. For mining, construction, estate and utility, the classification used in the employment data and the BSWS data are the same. For the remaining three, the two data sets use more detailed, but different classifications. Since it is hard to assign industry codes used in the BSWS data to the employment data, I use major industry codes for service, transportation and wholesale/retail industries. This simplification may have produced inaccurate estimates for diffi, since these sectors include a diverse set of companies. Furthermore, the correspondence is not exact for transportation and utility industries, since the BSWS wage data for these are calculated by including both the public sector and the private sector, whereas observations of the employment data used here are those in the private sector.
(4-2) Detailed industry code is used for manufacturing and finance industries. For manufacturing, the detailed industry codes in the employment data and the BSWS data are close, except for publishing industry. For finance industry, it is relatively easy to identify detailed industry code from the company names, so I reclassify the industry code of the employment data. In the original employment data, banks, trust banks and some other financial institutions are included in one category, whereas in the BSWS, these companies are classified into “banks and trust companies” and “financial and institutions for small sized enterprises, personal and housing credit agencies”. I reclassify the banks in the employment data so that they correspond to the BSWS. In doing so, I exclude some of the financial institutions in the employment data, since the wage data are unavailable for them. The excluded companies are in the following categories; financial institutions for agriculture, forestry and fisheries finances (norin-suisan kinyu gyo), financial auxiliaries (hojo-teki kinyu gyo) and investment institutions (tohshi-gyo).
(4-3) To make the data sets before 1994 and after 1994 to have consistent industry classification, I reassign industry codes for the 1984-1993 employment data. Before 1993, there had been a category called “mass communications and publishing”, but the companies in this category are classified into "manufacturing related to publishing" and "service" after 1994. The coding after 1994 is the one used in the BSWS. Therefore, I reassign the industry code for the employment data of 1984-1993, to make it comparable to the one after 1994. It is necessary to distinguish mass-communications or advertising companies (which are classified to be in the service industry) from publishers (which are included in the manufacturing industry). I used “List of Publishers in Japan (Nihon no Shuppan-Sha)” by Shuppan-News-sha (1994 and 1996 edition), to identify publishers.
(4-4) The information-related service industry is added as a new industry code in the employment data after 1994. Since it is hard to reclassify the industry code before 1993 to be comparable to the one after 1994, I include the observations coded as “information-related service” after 1994 into the service industry.
(4-5) The above procedure creates a sample of the following 25 industries (names are sometimes abbreviated): mining, construction, chemical products, electronics, textile and apparel, precision instruments, food, glass and rubber, machinery, fabricated metal, metal, nonferrous metal, pulp and paper, publishing and printing, steel, transportation equipment, other manufacturing, banks and trust companies, financial institutions for small sized enterprises, securities, insurance, wholesale or retail, service, estate, transportation and communication and utility.