Regulating the Timing of Job Search: Evidence from New College Graduates in Japan ∗ Hiroko Okudaira† please click here for the latest version February, 2017

Abstract Students search their post-graduation positions, sometimes quite in advance to the actual date of work, prompting debates to regulate the timing of job search. This paper examines a unique case of new college graduates labor market in Japan, where the guideline revision has successfully delayed the timing of job search and forced market participants to search under the shorter horizon. Based on differential exposures to the guideline revision across regions, I found that the revision significantly increased employment rate at the time of graduation. No positive effect is observed on students’ human capital investment. Timing regulation primarily altered job search behaviors.

Keywords: entry-level labor market; job search with deadline; federal court clerkship. JEL Classification: J40; K31.

∗

The author is grateful to Toru Kitagawa, Yoichi Arai, Richard Blundell, Christian Dustmann, Fumio Hayashi, Yoichiro Higashi, Kristiina Huttunen, Shin Kanaya, Keisuke Kawata, Shin Kimura, Ayako Kondo, Attila Lindner, Koyo Miyoshi, Hyejin Ku, David Neumark, Souichiro Ohta, Fumio Ohtake, Suphanit Piyapromdee, Ponpoje Porapakkarm, Fabien Postel-Vinay, Imran Rasul, Kei Sakata, Masaru Sasaki, Mariko Suzuki, Tatiana Surovtseva, Michela Tincani, Ryota Yabe, Jeff Wooldridge as well as seminar participants at UCL, GRIPS, Hiroshima University, EALE, ESPE, Japan West Labor Workshop, and JEA for their helpful comments. Yuki Umeoka provided excellent research assistance. The author thanks Higher Education Bureau, the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT), for providing the questionnaire data set. The analysis using School Basic Survey was conducted by following an agreement approved by MEXT. University Coop Survey (44th to 48th waves) was provided by the Social Science Japan Data Archive, Center for Social Research and Data Archives, Institute of Social Science, The University of Tokyo. The author acknowledges research grants from the Japan Society for the Promotion of Science (Grant-in-aid for Scientific Research 15K03434, JSPS Overseas Research Fellowship), the Kikawada Foundation and Okayama University. † Department of Economics, Okayama University, 3-1-1, Tsushima-naka, Kita-ku, Okayama city, Okayama 700-8530, Japan ([email protected]).

1

1

Introduction

In entry-level job markets, students tend to seek for their post-graduation positions, quite in advance to the actual date of work, even more than a year prior to the completion of school program. For example, in the market for US federal law clerks, judges make offers sometimes as early as the beginning of the second year, almost two years prior to the graduation (Avery et al., 2001, 2007). In the US gastroenterologist fellowship market, students had obtained offers nearly one year prior to the graduation until a new centralized match was introduced in 2006 (Niederle and Roth, 2009). Japan also has experienced an intense case of job market unraveling among new college graduates. According to School Basic Survey (Ministry of Education, Technology and Sciences), more than 350,000 students enter the market at the same time every year, equally seeking positions starting on the 1st of April, the very next day they officially leave the school. Similar to the entry-level professional labor markets in the U.S., the first job substantially matters to the new graduates’ career in Japan. A failure to obtain full-time regular positions at the time of graduation is shown to have persistent detrimental effect on the students employment status and earnings ( Kondo, 2007; Genda et al., 2010), making students desperate for securing position offers long before a job will start.1 By 2010, the timing of job search has been advanced to the middle of the junior year, nearly 18 months prior to the graduation. 2 The unorganized hiring process has invited debates to regulate the timing of job search, by setting up uniform dates to start interviewing or make offers. 3 According to the proponents of such a regulation, the timing of job search matters. Questionnaire surveys on federal appellate judges reveal their concerns that the information on students’ ability has not become fully available at an early stage of their school program (Avery et al., 2001, 2007). Others also concern that an early and dispersed matching would deprive students of their time to devote to classes at school, thereby discouraging the investment in their own human capital (in an official request by three of the Japanese ministers, October 8, 2010). Despite the numerous attempts to directly regulate the timing, however, 1 Knowing this stigma effect, students even take a strategy to postpone a graduation and wait for the next job hunting season if they do not obtain any full-time positions, as will clearly be shown in the analysis. 2 Studies have highlighted mechanisms which can induce the advancement of the timing. In principle, any factors that can dilute the quality of the applicant pool, or market thickness, at later stages of job search trigger the unraveling of the market. These factors include: connections between firms and candidates when firms can only observe noisy signals of worker productivity at earlier stage (Fainmesser, 2013); continuous informational gains in relation to staying in the market for later stage (Ambuehl and Groves, 2015). 3 Another important approach taken in some entry-level labor markets is to implement centralized matches, including some medical residency programs. See Roth and Xing (1994) for more details on this approach.

2

timing regulation has barely become successful, often led market participants deviate from the rule. 4 More critically, little is known about what the behavioral consequences would be, if such a timing regulation had been implemented successfully. This paper aims to identify the consequences of regulating the timing of job search, by examining a unique case of new college graduates job market in Japan, where guideline revision successfully delayed the timing of job search and forced students to search under the shorter horizon. The guideline was originally introduced in 1996 by the biggest business association in Japan, and was announced to be revised in 2010 to set up the date to start job search, for the first time since its introduction. Similar to the previous attempts to directly regulate the timing, the revision was not legally binding and market participants could always deviate from the rule. Unlike the previous cases, however, the revision successfully moved the overall timing of market backward, due to a closure of the popular online platform until the first date specified in the revision. Because college students start communicating with firms via these online platforms by registering for the first-step seminars, and also because these online platforms had been do dominant, the market became substantially diluted in that much less firms and students were available in the market prior to the first date. Based on an administrative survey data covering nearly entire population of colleges in Japan, I first report that the revision indeed delayed the timings at which students start their job search and also declined the search duration, at least during the sample period analysed in this paper. In particular, the beginning of job search has been moved back by eight weeks on average. Interestingly, the timing at which firms offer positions to students remained relatively stable even after the guideline revision. As a result, students were forced to search under the shorter search horizon. In order to identify the impact of the delayed timing and the shorter search horizon, this paper also exploits the differential program intensity across regions. Because the impact of the revision was severer in regions which had initially experienced a large advancement in the timing of job search, the differential extent of timing changes can be used to identify the consequence of delaying the timing of job search. To adopt this identification variation, this paper estimates difference-in-differences (DID) models, which is also complemented with the instrumental variable estimation strategies. The identification of estimating model relies on some important assumptions, and they are also examined. 4

In case of federal law clerkships in the U.S., Judicial Conference made six attempts between 1978 to 1998 ( Haruvy et al., 2006), and most recent Federal Law Clerk Hiring Plan initiated in 2005 was formally abolished in 2013. In case of new college graduates market, there have been about four major waves to set up or revise such recruitment agreements (see Appendix A).

3

One important aspect of this policy experiment is that the market participants were forced to search under the shorter search horizon. When students and firms can spend much less time in search, several consequence can arise. First, market participants have an incentive to exert higher search intensity to increase the probability to be matched with firms. Second, reservation values of market participants are reduced at a given point in time, as the perspective that they would have less chances to meet potential partners in the next round drives down the marginal benefit of continuing search. Consistent with both of the two potential mechanisms, I found a positive and relatively large effect on job placement rate, an employment rate at the time of graduation. The result is robust against several potential threats to the underlying identification assumption. For instance, DID with program intensity approach such as the one in this paper entails a fuzziness in treatment definition thereby masking the true causal effect (Chaisemartin and Haultfoeuille, 2015). By looking at heterogeneous impacts across college types, this paper also confirmed that the impact is concentrated and stronger among universities which are less likely to deviate from the treatment status, and thus less likely to be switchers. While a rather fuzzy design prevents us from obtaining a precise point estimate, conservative specifications suggest that the revision increased the job placement rate by at least 4 percentage point (mean of job placement rate = 0.568). On the other hand, I found little evidence that students significantly increased their human capital investment, despite that they had more time to concentrate on their studies at college. In particular, the third (i.e., junior) year students who were about to enter the job market did not increase the number of days when they were on campus, the number of books purchased, hours of reading, and their perspectives to study abroad. Moreover, the second year (i.e., sophomore) students are less likely to be on campus, reflecting students’ strategy to postpone the timing to take classes, in the expectation that they would have more time in the second semester of the junior year, than students piror to the revision did. The positive impact on job placement rate can be interpreted in several ways. If the shorter search horizon induced students and firms to reduce their reservation values, the quality of the match would be lowered since a reduction in reservation value deters sorting in the labor market. In contrast, if the shorter search horizon encouraged market participants to exert higher search efforts, the quality of the match could be improved. Unfortunately, the current paper provides only limited evidence differentiating one mechanism from the other. However, some side evidence found

4

no significant change in the proxies for students’ search effort, pointing to a possibility that market participants reduced their reservation values in response to the declined search horizon. Thus, the timing regulation has altered job search behavior and potentially the later labor market outcome, without significantly improving students’ dedication to on-campus studies, which is in contrast to the original intentions of the guideline revision. Few studies have quantified the consequence of the timing regulation so far. Most related evidence is Fr´echette et al. (2007), where they estimated the efficiency gain of postponing the early matches between college football teams and post-season play in 1990s. While their findings rest on the informational gains after the end of season records became available, thereby enabling bowls to generate championship matches that attracted more television viewership, this paper sheds light on the resulting consequences on search behaviors and student’s human capital investment. In fact, the information gains due to the guideline revision were limited, as will be discussed later. Another related evidence, Niederle and Roth (2009), has found in a case of new gastroenterology fellows that the use of centralized match has increased a mobility of the fellows between hospitals or cities, implying that the early contracting is inefficient in the sense that the labor market is decentralized. However, their evidence compared the states with and without centralized match maker, and thus less suggestive for markets which has traditionally failed to establish strong market intermediaries, such as a market for the federal court clerkships in the U.S. This paper provides the first evidence on the timing regulation in the context of labor market, in the absence of any match makers. Moreover, this paper also provides an interesting case study in which labor market participants were forced to search under the shorter search horizon. A simple finite-horizon job search model predicts that the shorter search horizon reduces the reservation wage of workers at a given round, which have been tested in a controlled experiment (Cox and Oaxaca, 1989a). However, such a situation has been rare in the actual labor market. Although a similar idea has been investigated in studies looking at the unemployment benefit durations (van Ours and Vodopivec, 2008; Caliendo et al., 2013; Farber et al., 2015; Farber and Valletta, 2015), the main underlying mechanism in these studies is a direct reduction in reservation wages via reduced outside options, and not a reduction in the participants’ perspectives on their deadline. In fact, a discussion in this paper clarifies a possibility that a reduction in search horizon accelerates a reductions in reservation values one another in the entry-level labor markets, since the reservation values are partly determined by the quality of potential

5

partner’s pool in later rounds. Faced with a perspective that their returns to continue theirs search is lower due to the more deteriorated applicant’s pool in the next rounds, market participants are less likely to remain in the market, driving down their reservation values further. The remainder of the paper proceeds as follows. Section 2 summarizes background information and introduces the timing data. Section 3 discusses the potential mechanisms. Section 4 describes the identification strategy. Section 5 presents results along with some robustness checks. Section 6 concludes.

2

Background

2.1

The Market for New College Graduates in Japan

Japanese job market for college graduates has some distinctive features. First, the market is quite large, and becomes extremely congested at some point every year. According to School Basic Survey, at least 388,500 students entered the market in 2014-15 school year. Students are equally seeking for a full-time position starting on the 1st of April, the very next day they officially leave the university. It is not common for students to take gap periods or to advance their graduation by skipping a year. Second, the first job matters substantially in Japan. A failure to obtain full-time regular positions at the time of graduation is shown to have persistent negative effect on the labor market outcomes later in career (Kondo, 2007; Genda et al., 2010). Third, wages are barely negotiated. The variation in the first-year salary is extremely small, although students know that they will be compensated by differential wage profiles later in the career. 5 Fourth, students simultaneously apply for multiple firms. Unlike in job market for federal law clerks in the U.S., some students hold several offers, although firms threaten students not to. 6 Finally, grades information is not an essential part of firms screening process.7 Instead, firms screen students in a lengthy selection process. Job hunting is indeed time-consuming to students. At the early stage of their job search, students spend a lot of time in attending seminars held by an individual firm or jointly by a group of companies. 5 Among 2862 firms, which constitute about 80% of those listed in Tokyo Stock Exchange, average first-year monthly salary is 207,450 JPY or 1852 USD (?). The distribution is quite concentrated around the mean: only 144 firms responded that they pay more than 240,000 JPY or 2143 USD (Ibid.). 6 Among students who obtained any full-time position, an average student received 1.6 offers by the end of their job search in 2006 (JILPT, 2006) However, about 20%of students experienced exploding offers, where firms threaten students to terminate their job search on other firms before making their own offers. ( Cabinet Office, 2016) 7 Nearly 70% of the students responded that they were never asked about their college grades during interviews (MEXT 2015). More than 50% responded that they were never asked to submit the transcripts (Ibid).

6

In many cases, students have to attend the seminars for firms they would apply to, in order to get admitted to the screening process by the firm. After browsing numbers of firms, students apply to firms by writing an application form which often takes some time to complete. Then, firms review the applications and decide whether to invite the student to an interview. Students usually have to go through several interviews within the same company until they are offered a position. In 2014, an average student attended 34.6 seminars, sent applications to 19.4 firms, got invitations to interviews from 10.1 firms, and obtained 2.02 position offers (Recruit, 2015). As completing this process takes times and efforts, students can barely attend classes during the job hunting season. 8 Similar to some entry-level professional labor markets in the U.S., the market has always been subject to the advancement in the hiring schedule. As laid out in Appendix A, students in the late 1990s engaged in their job hunting at the beginning of the senior year, less than 12 months before the graduation. However, the timing started to unravel further in the 2000s, and by the end of 2009, it has moved to the beginning of second semester or October of junior year, eighteen months piror to the graduation. Similar to the other market, there have also been numerous attempts to enforce the uniform dates in the hiring process, although they often turned out to be unsuccessful. 9 The first recruitment agreement was initiated as an official agreement among employers, universities, and government, and specified the first dates for firm visits/recommendations and interviews. After a long history of struggles for revisions, the recruitment agreement was abolished in December 1996, and replaced with a soft guideline set voluntarily among employers only. The soft guideline, or Rinri Kensho by Japanese Business Federation (JBF), no longer specified the uniform dates in hiring, until it was announced to be revised in 2010. 10

2.2

Guideline Revision

This paper exploits a unique case of new college graduates in Japan, where the guideline revision has successfully delayed the timings to start job search and forced students to search under the shorter 8

According to annual survey by Council Board for Recruitment Issues on New College Graduates ( Shusyoku Mondai Kondan Kai ), 87.8% of universities responded that students attend less classes, 35.4% said that professors have more difficulty to supervise their theses in 2011. 9 See Appendix A1 for more details on background history. Roth and Xing (1994) also has detailed information about unraveling observed in Japan between 1970 to 1990. 10 Although the soft guideline did not specify the first dates for firm visits and interviews, they did mention the first date of formal job offer (naitei ) at October 1st of the students’ final year (six months prior to the graduation). However, the date of formal offer has been well recognized as a dead letter, since firms provide informal offers ( nai-naitei ) well in advance to the formal offer date. For a long time up to today, October 1st is known for the senior students’ travel day to participate in the company’ ceremony (naitei-shiki ) to confirm their final intention to work for them.

7

horizon. The guideline was originally introduced by JBF in 1997 (Rinri Kensho), and was announced to be revised in 2010 to set up the date to start job search, for the first time since 1996 when the previous recruitment agreement was officially abolished. The revision intended to calm down the situation where a substantial number of students engaged in job search quite in advance, instead of focusing on their college studies. The announcement was unexpected. It initiated right before the junior year students in 2010 (e.g., graduating class of 2012 cohorts) were about to start job search by attending the firm seminars. In September 10, 2010, Japanese cabinet office approved a new policy package containing employment programs for new college graduates, soon followed by the ministers’ official request to “redress the early and lengthy hiring schedule (October 8, 2010)”, which was sent to major employers associations. One of the associations, Japan Foreign Trade Council (JFTC), first responded to the official commitment in November 17, 2010, by announcing to delay their hiring schedule. The biggest and most influential association, JBF, then followed and announced the revised guideline ( Rinri Kensyo), “refraining firms from releasing any classified adds until December 1st on the junior year (16 months before graduation).” Despite that the guideline revision was not a legal requirement, and that the employers’ associations do not cover the entire population of companies in Japan, the revision successfully imposed the first date to post classified ads, December 1st in the student’s junior year. It continued to work effectively, because major online intermediary services postponed the registrations to the platforms until the 1st of December. 11 The online platforms became extremely popular among students since the early 2000s, covering almost all students and substantial number of companies 12 . They have played a critical role throughout the screening process, in getting firms recruiting information and in booking slots in firm seminars. Unlike cohorts prior to the revision had done, students could no longer book firm seminars until December 1st of their junior year. As the registration was restricted to start after the December 1st, the platforms has become quite congested on the first day of the registration. National medias reported that the most dominant platform at the time, Rikunabi, had to shut down for four hours on the 1st of December 2011, due to the excessive number of students trying to access to their web pages to register (e.g., Nikkei Newspaper, December 2, 2011). 11

Ricruit Co. Ltd, which runs the most dominant platform at that time, Rikunabi, has been a member of JFB. According to JILPT (2014), about 95% of career support divisions at universities responded that students obtained job vacancy information through these platforms. Other informal survey run by Human Resource Research Institute also found that at least 95% of students in class of 2017 were actively using online platforms as of the last week in March 2016. 12

8

Of course, it is still possible that students and firms preempt from the online platforms and start forming the match, however, the informal direct hires had not been dominant and remained relatively stable before and after the revision. According to Surveys on Hiring Perspective (Mainichi Communications Co.), prior to the revision, the ratio of firms hired students through direct headhunting by recruiters was 10.5% and 9.6% for students in classes of 2010 and 2011, whereas it was 8.0% and 7.9% for students in classes of 2012 and 2013 after the revision. Hiring through the internship was also rare and stable during the same period, ranging from 2.1% to 2.4%. Thus, there was no obvious change in firms’ hiring methods. The restricted access to the platforms had an exclusive effect in diluting the market, in that much less students and firms actively engaged in search activity before the 1st of December. In the following analysis, I consider that students in classes of 2012 and later were affected by the guideline revision. Although the revision was initially targeted to students in classes of 2013 and later, the employers’ associations responded to the government’s request, just around the time when the class of 2012 students started their job search in the fall of their junior year in 2010. As will be shown in the next section, the revision indeed postponed the actual dates of job search also for the students in classes of 2012. Thus, I consider that the announcement of the revision affected students in classes of 2012, and continued to work effectively in the following cohort.

2.3

Timing Data

To construct the information on the timing of job search, this paper draw on the annual files from the Questionnaire Survey on Students Job Search Activity (Job Search Survey). This survey is conducted annually by the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT), as a part of projects at Council Board for Recruitment Issues for New College Graduates (Shushoku Mondai Kondankai ).13 The questionnaire was sent to the career service divisions at postsecondary educational institutions, including technical colleges and both two- and four-year colleges, at the end of June every year, asking about job search activities of their students in the final school year. Response rates are quite high and ranges from 92.1% to 97.8%, indicating that the survey covers a nearly entire population of all tertiary educational institutions in Japan. Although not all students rely on the services provided by career service division, a majority of them are monitored by 13

The author is grateful to Higher Education Bureau, the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT), for providing the questionnaire data set.

9

the division.14 Thus, I consider that Job Search Survey represents the actual timings of job search among representative students at each college. To measure the timing of job search activity, I use the responses by a career division at each college to a question, “when did students start participating in interviews and screenings?”. To measure the length of the job search, I take a difference in responses between this question and another one, “when was the job offers to the students were concentrated?” Since the responses were recorded as categorical variables to indicate the timing, in either of the beginning/middle/end of specific month, they were transformed into a numeric variable, a number of weeks to graduation. Appendix B contains further details on the coding procedure. In the following, note that the Japanese school year starts in April 1st and ends in March 31st. Panel A in Table 1 shows summary statistics for the timings in terms of the number of weeks to graduation. Figure 1 shows corresponding cumulative distributions, where horizontal axis measures a number of weeks to graduation. The classes of year indicate cohorts of students who graduated from college in March of that year. As a reference, 70 weeks to graduation roughly corresponds to the first week of December in the year prior to graduation (i.e., December in the third year at college). Table 1 and Figure 1 show that a guideline revision indeed delayed the timing of job search for students exposed to the revised guideline. Average timing when students started their job search reported by colleges has declined from 63.67 weeks for those graduated in 2010 to 55.71 weeks for those in 2013, indicating that the timing has moved from the beginning of January to the beginning of March by eight weeks. Importantly, the proportion of those colleges in the market as of the first week of December (70 weeks to graduation) started to decline dramatically for class of 2012 cohort, as shown in panel A of Figure 1. It continued to decline and reached to almost zero for class of 2013 students. Thus, although the guideline was not a mandatory statute, the announcement of the revision indeed discouraged firms and students from participating in the market, at least for the two cohorts. Interestingly, the timing of job offers remained mostly stable, despite the delayed start in hiring process after the announcement of guideline revision. Panel A of Table 1 confirms that the timing of job offer ranged around 40 weeks to the graduation, the middle of June every year. 15 As the timing 14 About 80% of the students on average were covered by the initial registrations and within institution surveys conducted by the divisions (JILPT 2006). 15 One potential reason for this is the exam and interview schedules to hire public servants, especially those prestigious positions such as those in the central government. Students considering these positions have to involved in the exams and interviews intensively in June and July. Private firms have an incentive to make offers in advance and prevent them from taking the exams. Footnote 35 in Roth and Xing (1994) also explains this fact.

10

of job offer remaind stable, the delayed start was mostly reflected into the shorter duration of job search. According to Panel A of Table 1, the average duration of job search reported by colleges has been reduced by about 6.5 weeks for those students in the class of 2013, when compared against those in the class of 2010. While Job Search Survey contains unique information on the timing of job search observed by career support divisions at each college, the ministry unfortunately does not provide the school identification codes in the data set. For this reason, I collapsed the university-level timing data into prefecture-level information, and matched them with the department or student-level outcome variables at prefecture level. Rationale behind this procedure is that an individual decision when to start job search depends crucially on when the other participants in the same local market enter the market, as it could affect their match outcome. To the extent that students base their job search in their own region, the collapsed data should represent the timings that most students engage in their job search in the local market. 16 Of course, there may be some universities deviating from the representative timing in the local labor market. Later in the analysis, this paper also conducts robustness tests against this fuzziness in the design, by examining the heterogenous effects across the types of college. To better understand the identification variations exploited in this paper, note that universities in a region that initially experienced early start of job search were also the ones more likely to be affected by the guideline revision. Figure 2 shows this point, by drawing the cumulative distributions of the timings to start job search for students in classes 2010 and 2013. Cdf’s are shown separately by initial status, which was defined according to the extent to which job search was advanced prior to the announcement of revision. In particular, “in affected prefectures” include universities in prefectures which initially experienced early start of job search, namely, prefectures where at least 40% of universities responded that students in class of 2010 had already started their job search as of the first week of December in their third grade. 17 “Others” include the rest of the universities in other prefectures. Dashed line indicates the first week of December in the students’ junior year, or 70 months piror to the graduation. As so was defined, the gap in the proportion of those colleges in 16 According to Employment Status Survey conducted by Ministry of Health, Labour, and Welfare, 63.8% of the newly employed graduates stayed in the same prefecture in 2010 where they had lived before hiring, while 57.3% stayed in the same prefecture in 2013. Note that the newly employed graduates here include high school and junior high school graduate, in addition to college graduates. 17 “Affected” prefectures include Akita, Fukushima, Saitama, Tokyo, Toyama, Fukui, Shiga, Nara, Wakayama, Okayama, Tokushima, and Saga.

11

the market between affected and other prefectures is quite large for class of 2010 cohort. However, it shrinks significantly after the revision, to almost zero for the class of 2013. In fact, Wilcoxon ranksum test shows a significant differences in the two cdf’s in 2010 panel (p < 0.0001), while they are not significantly different in 2013 panel (p < 0.374). Among the affected prefectures, the proportion of colleges in the market as of the first week of December has declined by 44 percentage point from 47.8 % in 2010 to 3.8% in 2013, while among other prefectures it has declined by 23 percentage point from 26.2% in 2010 to 3.1% in 2013. As will be extensively examined later, the differential exposures to the revision do not reflect any pre-existing local trends or some observable prefecture traits. This paper uses this differential exposure to the guideline revision across prefectures to identify the impact of the guideline revision.

3

Mechanism and Hypothesis

It is not necessarily clear whether the revision announcement can increase students’ investment on their human capital, despite the original intention of the government to request the revision (Official Request, October 8, 2010). Indeed, a change in the timing of job search may have simply altered the timings to engage in job search or college studies. Unless the earlier investment on skills yields the larger returns, the timing of job search itself would not affect the total amount of human capital accumulation by students. However, if the investment dynamically complements later investments, and/or if students have an ambiguous perspective on the timing at which they will obtain a job offer, the delayed timing may indeed increase students’ investment at college. Especially when a job market is much less dense in that many employers are not on the market, it discourages a decision to participate in the market for those students who face a lower probability to obtain job offers in such a diluted market. These students are better off by investing their time on college studies, rather than spending their time on job search at an early stage. 18 On the other hand, the guideline revision can also induce the unexpected outcomes on search behavior. Since market participants were forced to search under the shorter search horizon, they face a decline in a discounted sum of the marginal benefit to continue job search at a given round. 18

Although the identification variation in the timing of job search is not large as shown in the previous section, the delayed timing may have potentially increased investments in their human capital, since it allowed students more time to take exams in the second semester of their third year. Second semester starts in September/October and ends in January/February. Students usually take exams from January to the first week of February. Table 1 confirms that the average first week job search was postponed from the beginning of January to the beginning of March.

12

This changes optimal search behaviors of agents mainly by two mechanisms. First, due to a reduced expected gains of continuing job search in later rounds, the market participants face larger incentives to increase their search intensity at a current round, thereby increasing the probability to be matched with their partner at an earlier stage of their search. Second, since the marginal gain of continuing job search is declined, market participants are also tempted to reduce their reservation values at a given point in search so as to avoid remaining in the market. 19 Indeed, Cox and Oaxaca (1989b) directly observed in a simple lab experiment that the shorter search horizon lowers subjects’ reservation values and their search durations. Importantly, a reduction in search horizon accelerates further reductions in reservation values in the entry level labor market. The acceleration process arises, because entry level labor market has a deadline for job search and market participants concern about the quality of potential partners in later stage of the search. Consider a two-sided, finite-horizon search and matching model with heterogenous types and complementarity between types, as analysed in Damiano et al. (2005). Suppose that a search horizon is exogenously reduced. Then, the discounted sum of expected marginal gains to continue job search decreases at each round, reducing the reservation values at a given round. Once the reservation value is reduced, the quality of participant pool in later rounds deteriorates, because only those at the lower tail of type distributions remain in the market. This triggers further reductions in reservation value at each round: with a fear of ending up with vacancy or being matched with the lower types in later rounds, participants further reduces the reservation values and try to form a match as soon as possible. A further reduction in reservation values again worsens the quality of participant pool in next rounds in turn. An exogenous reduction in the search horizon, therefore, drives the reservation values go down further and further, eventually generating random matches in the first round of job search. 20 To test these insights on job search behaviors, this paper first looks at the overall effect on job placement rate at an individual department. Job placement rate is an “employment rate” at the time of graduation (see section 4.3 for more details on outcome variables). Because a reduction 19

Although students sometimes hold multiple offers JILPT (2006), this paper ignores a possibility that students search with recall, because firms threaten students not to search for other positions once they accept their offers, limiting student’s ability to hold recall options in reality (Cabinet Office, 2016). 20 The reservation value in the first round equals to outside options available for market participants. For an individual student, the outside option equals to the expected value of staying another year at college and participating a search in the next job hunting season. As long as the student believes that the revision will continue to work in the next year, the expected value of job search in the next year is also reduced by the revision announcement, since researvation values are expected to be lower also in the next year.

13

in search horizon drives both an increase in search intensity and a decrease in reservation values, job placement rate is expected to rise due to the revision announcement. How much it is affected depends on a distribution of students’ ability type at each department, which will be discussed in heterogeneity analysis. Of course, the impact on job placement rate is still a compound, and also entails alternative possibilities such as overall shifts in reservation values due to more human capital investment by students. This paper further examines the impacts on proxies for students’ investment on their human capital, along with job placement rate. Interestingly, the two mechanisms on search behavior predict opposite consequence on the quality of the match. If the shorter search horizon induced market participants to reduce their reservation value, the quality of match is lowered and deters the sorting in the labor market. If the shorter search horizon encouraged market participants to exert higher search efforts, the quality of match could be improved. Some studies have examined the impact of unemployment benefit reforms on a match quality by looking at post-unemployment outcomes, such as tenure length or wage rate ( van Ours and Vodopivec, 2008; Petrongolo, 2009; Caliendo et al., 2013). Unfortunately, such information is not available in a credible dataset in Japan. 21 Instead of analyzing the impacts on a match quality, therefore, this paper examines the effects on some proxies on student’s search intensity, although these proxies are admittedly crude. Finally, it should be noted that this paper does not focus on informational gains arising from the delayed timing although it can potentially interact search and matching behavior mentioned above. When market participants engage in early job searches, some employers are obliged to make offers even if information on students’ ability, such as transcript records or instructor’s evaluation, is not fully available. By waiting more, the quality of match can be improved. In fact, this is one premise that a centralized match has been introduced in some residency programs (Niederle and Roth, 2009) Although not in the context of labor market, Fr´echette et al. (2007) also examined that the postponing the college football match significantly increased the television viewership. Nevertheless, such informational gain is much less relevant in a case analyzed in this paper. One central reason is that employers do not take student’s grades into account during hiring process in Japan. As a matter of fact, a majority of students are not asked about their grades during interviews, nor even 21

In particular, to apply the estimation strategy introduced in the next section, it is necessarily to match a prefecture where a university locates/students had engaged in job search, with post-search labor market outcomes. To the best of author’s efforts and knowledge, there is no such administrative or credible data in Japan which allows a matching with university locations. One option is designing an online survey, although its external validity may be limited.

14

asked to submit their transcripts (see footnote 8). This made the information gain by delaying the timing much less limited, although the delayed timing made it possible to employers to access to the extra information in the transcript available at the end of the second semester of students’ third year.22 This is in contrast to a market for the federal law clerkships or college job markets in some other countries, where the transcript plays a non-trivial role in hiring decisions. 23 Thus, this paper considers that information gains in the Japanese case hardly affects the interpretation of the mechanisms behind results reported below. 24

4

Identification Strategy

4.1

Model

Difficulty in identifying the impact of the timing and duration change arises because the timing at which students start their job search can reflect some unobserved but important confounders specific to the prefecture. To identify the impact, this paper exploits the differential exposures from a revision announcement across regions in Figure 2. Provided that a control group (e.g. other prefectures in Figure 2) constitutes a valid counterfactual, this paper examines how the observations in a treatment group (e.g., the affected prefectures in Figure 2) would have been, had the timing not been affected that much by the revision announcement. The idea is very similar to the program intensity approach in the previous literature, such as Duflo (2001). The baseline model begins with a simple differencein-differences (DID) estimation with fixed effects. Let Yit denote a department-level outcome variable, namely, an employment rate at the time of graduation (job placement rate) among class of year t students in department i:

22

h i Yit = α + γ Dt ∗ I(Sp10 ≥ 0.4) + xpt β + θi + λt + it

(1)

In the second semester, students take exams from January to the beginning of February. Transcripts usually become available by the end of February. 23 Ambuehl and Groves (2015) analyse a situation where the information on students’ ability gradually arrive in two-sided one-to-one matching model. The presumption of their model is that the grade information is essential to the quality of the match, as was argued in an example of the federal appellate judge clerkships in the U.S. 24 Another possibility is the informational gain in type distribution of wage or employer types. When the type distribution is uncertain to students in the sense of Knight, the delayed timing would increase the expected marginal gains in continuing the search, thereby decreasing the reservation value. This would result in a decrease in employment rate. This paper ignores this possibility, because the identification variation in the timing to start job search is only several weeks at most (see Figure 2), which is too short to expect any substantial information gains in employers’ type distribution.

15

A department in this model indicates an individual department, separately coded for each university. I(S10 p ≥ 0.4) is a treatment indicator to denote that a department i is in a prefecture most likely to be affected by the guideline change. More specifically, I(Sp10 ≥ 0.4) is defined as a dummy that takes one when prefecture p saw at least 40% of universities responded that students in class of 2010 started their job search by the first week of December in their third year. Dt is a dummy variable to indicate the cohorts affected by the announcement of guideline revision, which takes a value of one if t = 2012, and zero otherwise. pt is the error term. The model also controls for department level fixed effects along with year dummies, thereby absorbing the effects of individual linear terms I(Sp10 ≥ 0.4) and Dt . xpt is a set of prefecture-level control variables, including logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2; average monthly wage for new college graduates at year t − 2; and university advancement rate at year t − 4.25 Standard errors are clustered at the department level to allow for arbitrary serial correlation within the university. If this model is correctly identified, γ should capture the impact of the delayed timing as well as the reduced search horizon due to the revision announcement, specific only to observations in a treatment group or affected prefectures.26 To identify this model, it is essential that both treatment and control observations must follow the same outcome path, had they not been largely exposed to the guideline revision. In the context of this paper, there are two potential threats to this assumption. First, the affected prefecture dummy, I(Sp10 ≥ 0.4), may not be randomly given across prefectures. If the initial advancement in the timing of job search reflects some time-variant regional traits, the common trend assumption can be violated. Second, the treatment status is assumed to be sharp in the above DID model, while it may be fuzzy in reality. In particular, I(Sp10 ≥ 0.4) was defined by those prefectures with no less than 40% of colleges reporting that students had already started job search as of the first week in December. Thus, some departments may have been included in the control group, when they were in fact heavily affected by the revision announcement, thus would be more likely to be treated as time goes by. In such a fuzzy design, the estimated DID estimator should be adjusted according to the extent that these switchers are included in a control group (Chaisemartin and Haultfoeuille, 2015). 25

Lags in the first three variables are taken to account for the local labor market conditions right before job hunting season. A lag in the last variable is taken to partially proxy for the quality of the cohort graduating at year t at the time of entering college. 26 As the model is department-level, this paper also estimated the weighted least squared model with average number of students in a final year within each department as an analytical weight. Results are quite similar to those reported in this paper.

16

To examine if there are any pre-existing differences between affected and other prefectures, Table 2 compares some observed characteristics in 2010 by treatment status. Any of the predetermined traits are statistically different between the two groups, although control prefectures have slightly higher college advancement rates. Thus, treatment and control prefectures are comparable, at least in terms of these observed traits. In the following analysis, this paper conducts further robustness tests, including a standard falsification test with a placebo revision dummy to check any pre-existing differential trends across treatment status. To check the validity of the results under a fuzzy design, it is necessarily to find a control group in which treatment is stable over time (Chaisemartin and Haultfoeuille, 2015). In the context of this paper, we need to identify changes in the timing of job search for each observed unit. Unfortunately, the timing information from Job Search Survey contains only prefecture code and college ID is not available. This paper, therefore, takes various alternative approaches. For instance, a later analysis avoids a sharp DID definition of treatment status, by directly exploiting continuous variations in a policy variable (FD-IV estimations). I also estimate the effects only specific to those observations which are less likely to be switchers, by looking at heterogeneous effects across college types. In a similar vein, some control prefectures can be spatially related to the adjacent treatment prefecture and thus subject to spillovers from the treatment units. Since the spillovers to control observations may again scale up/down the true effect, later analysis obtain the estimate by limiting to those sample which are geographically unlikely to experience any spillovers. Finally, the robustness against the definition of the treatment variable is also tested, by varying the threshold values in I(Sp10 ≥ 0.4).

4.2

Data for Outcome variables

The information on the outcome variables are taken from two sources. Job placement rate (an employment rate at the time of graduation) is constructed from the first source of information. The dataset is drawn from a college-department level panel in tertiary educational institution files in School Basic Survey, for the years 2010 to 2013. The surveys are conducted annually by the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT), by sampling a population of all tertiary educational institutions in Japan. Job placement rate is calculated by dividing the number of students who obtained full-time positions by the total number of students in

17

the final year.27 I restrict the sample to non-STEM departments in four year universities and exclude any science and engineering departments, because STEM students face broader variations in career paths, such as referrals by professors in graduate course, etc. 28 For the same reason, observations from evening courses as well as two-year and technical colleges are also excluded from the sample, although two-year and technical colleges were included when constructing the timing information from Job Search Survey.29 Panel B of Table 1 presents the summary statistics. Job placement rate ranges from 54% to 61% in our sample period. The second dataset contains the information on students’ investment on their human capital. The dataset is drawn from University Coop Survey, conducted between October and November every year by university cooperation. 30 The data consists of repeated cross-sections of student-level information on their daily activities, consumption behavior, and their sources of expenses. The questionnaires were sent to students belonging to universities with any on-campus university cooperation shops. In contrast to School Basic Survey in which a population of all universities are covered, University Coop Survey over-represents students from national universities, indicating that the survey allows us to estimate the effect mainly on those students at the upper tail of ability distribution. 31 To measure the students’ commitment to their college study, this paper looks at the impacts on the following items; (1) number of days they were on campus last week; (2) average hours of reading per day; (3) number of books purchased for the purpose of their study or research in a past month (including electric books); and (4) a dummy variable to indicate whether they plan to spend their money to study abroad in the next six months. The observations are limited to the second (i.e., sophomore) and third (i.e., junior) year students in non-STEM departments. Those students who had already determined to go to graduate school at the time of survey are excluded from the sample. Panel C of Table 1 presents the summary statistics. It is noteworthy that the hours of reading is surprisingly low at about 30 minutes per day on average. 32 27

The total number of students in the final year is calculated by summing the number of students graduated in March and those who repeat a year(s) and stay at the university more than four years as of 1st of May. 28 The included departments are: law, economics, education, social sciences, humanities and arts. 29 Since the ministry (MEXT) does not provide the school identification code, it is not possible to identify two-year and technical colleges in Job Search Survey. Among all the tertiary educational institutions, 29% consist of two-year colleges and 5% consists of technical colleges (School Basic Survey). 30 The data set was provided by the Social Science Japan Data Archive, Center for Social Research and Data Archives, Institute of Social Science, The University of Tokyo. 31 According to School Basic Survey, 17.5% of students belong to national universities in 2012, whereas 53.4% of students belong to national universities in University Coop Survey. 32 Non-trivial proportion of students in Japan tends to avoid demanding classes and seeks for off-campus experiences (Kaji, 2015).

18

The impact on student-level outcomes are estimated in a very similar DID framework to departmentlevel model in equation (1). Let Yjt be the outcome of student j in graduation class of year t:

h i Yjt = α + γ1 Dt ∗ I(Sp10 ≥ 0.4) + xpt β1 + ψs + λt + jt

(2)

The model controls for the same set of prefecture covariates, along with a dummy to indicate the cohorts affected by the guideline revision, Dt , interacted with the program intensity. The model also controls for sets of dummies to indicate the cohort and college that a student j belongs to. Standard errors are clustered at university level. As the University Coop Survey is conducted every October to November, the estimated γ reveals the impact of guideline revision on the investment right before and one year before the job search started.

5

Results

5.1

Baseline results

Table 3 presents baseline results obtained by directly estimating the main DID model. Panels A and B show estimates for the impact on job placement rate and human capital investment by students, respectively. Note that results in panel A is obtained from college-department level panel estimation in equation (1), whereas results in panel B come from student-level estimation in equation (2). Each cell represents an estimate of the main DID interaction term from a separate model, where the model controls for observed prefecture traits. 33 To circumvent the possibility that any existing trend confounds the results, the model also controls for prefecture-specific linear trends. Panel A of Table 3 shows a significant treatment effect on job placement rate. Those departments in prefectures that initially observed more than 40% of colleges already active in the local markets, and thus more severely affected by the guideline revision, experienced an increase in job placement rate by about 2.1 percentage point. Since the affected prefectures experienced the larger decline in search horizon than those in other prefectures did, a positive impact on job placement rate imply that the revision changed search behaviors of agents, as discussed in section 3. Of course, the estimated 33

Although not shown for the purpose of brevity, results remain quite similar even when the model does not include the prefecture control variables. Thus, the estimated treatment effects in Table 3 are not driven by any potential correlations with variables likely to move with local market conditions.

19

effect is a compound, and can potentially reflects other possibilities. For example, due to the delayed start in job search, students can invest more in their human capital and increment the overall ability distribution. This could shift the reservation values of students upward, thereby increasing the job placement rate. However, panel B of Table 3 reveals that this was not the case. Despite the original intention of the government to make students concentrate on their college studies, I found little evidence that students significantly increased their commitment to college. In particular, the third year students who were about to enter the job market did not increase the number of days when they were on campus, the number of books purchased, hours of reading, and their perspectives to study abroad. 34 Although the estimate of number of days on campus (column 1 in panel vB) is relatively large compared to the standard error, the effect will get closer to zero after limiting the observations to those less likely to be switchers (see section 5.3?). Interestingly, the second year students are significantly less likely to be on campus. Specifically, the second year students in more affected region went to the college about 0.22 days less than those in other prefectures after the guideline revision. This result is consistent with the second students’ strategy to postpone the timing to take classes, in the expectation that they would have more time in the second semester of the third year, than students prior to the revision did. Thus, results in panel B show no evidence that the reduced length of job search encouraged the students’ investment on their own human capital. Importantly, the fact that I found no positive effect on students’ investment may not be a big surprise, as the estimated effects are associated with relatively small identification variation. A rough calculation suggests that the universities in affected prefectures experienced on average about 3.4 weeks of further delay in the timing to start job search. 35 Results here indicate that this 3.4-week gain in student’s time at college may not be sufficient to induce any positive change in students’ decisions to invest in their human capital. To put it in a different way, the three-week delay in the search timing is trivial in terms of the marginal gain in the additional investment by students, but it does matter to the actual screening process, because the 3.4 weeks of change can be relatively large in terms of search duration. Thus, it is unlikely that the positive treatment effect found on job 34 The impact on the perspective to study abroad is estimated in a linear probability model to avoid incidental parameter issues. 35 ˉ pt denote the average timing of job search at prefecture p among class of year t students, in terms of Let W ˉ t=2010 = 66.38 and W ˉ t=2012 = 56.01. For the other the number of weeks to graduation. For affected prefectures, W ˉ ˉ ˉ prefectures, Wt=2010 = 62.56 and Wt=2012 = 55.58. Difference in ΔWpt between affected and the other prefectures equals to 3.38 weeks.

20

placement rate reflects overall distributional change in students’ human capital. Rather, the baseline results suggest that the delayed timing altered job search behaviors of students and firms, as far as the identification strategy is valid.

5.2

Placebo Test

The key to the identification of the previous DID approach is a common trend assumption. For the control units to be valid counterfactuals for treated units, universities in the affected prefectures should follow the same outcome path as those in non-affected prefectures did, if they had not been affected that much. Although this assumption is not directly testable, this paper conducts several robustness tests. To begin with, this section conducts a standard placebo test, in order to check if significant results in Table 3 are not driven by the pre-policy trend. The idea here is to examine whether the affected prefectures shared the same trend with non-affected prefectures prior to the guideline revision. In this analysis, Dt ∗ I(Sp10 ≥ 0.4) in equation (1) or (2) is replaced by d11t ∗ I(Sp10 ≥ 0.4), where d11t is a dummy variable to indicate a placebo guideline revision at t = 2011. The sample in this analysis is limited to pre-revision cohorts, classes of 2010 and 2011. All models control for prefecture and year dummies as well as prefecture controls. Prefecture linear trends are not included as the pre-revision sample consists of only two cohorts. Table 4 presents the estimates from this placebo test. Similar to Table 3, panel A shows the estimate for job placement rate, and panel B shows the estimates for measures on human capital investment. In contrast to the previous results, the impact of placebo treatment is not significant in all specifications. In most relevant cases, column 1 in panel A and column 5 in panel B, the estimates are both close to zero. Although this placebo test is admittedly crude in having very short sample period, the placebo test here suggests that significant impacts estimated in Table 3 are not driven by pre-existing trend prior to the revision announcement.

5.3

Various Threshold Values

Another concern on the baseline results is a rather arbitrary choice for the threshold value in I(Sp10 ≥ 0.4). As a robustness check, this section varies a threshold value in the treatment definition. In particular, rather than taking one when Sp10 is greater than 0.4 for students in class of 2010, it is

21

now allowed to take other threshold values, i.e. n in Sp10 > n. If the estimated effect gets stronger as the threshold value gets larger, then a specification of the treatment is verified in the sense that it correctly captures those most likely affected by the guideline change. Panels A and B of Table 5 estimate exactly the same models in Table 3, but by varying the threshold values in I(Sp10 ≥ t). Columns 3 in each panel replicate the baseline estimates from Table 3. In panel A, the estimated effects on job placement rate are significant when the threshold values are equal to or greater than 40%. The magnitude of estimates become slightly larger, as the threshold values of the treatment gets larger. This monotone pattern is consistent with an intuition that the treatment effects are intensified when the threshold value is larger, supporting the validity of timing information used in the analysis. On the other hand, results on “Number of Days” in panel B exhibit seemingly a puzzling pattern on alternative threshold values. Although the estimates are negatively significant overall, the estimated magnitudes are not monotonically increasing: the estimate is relatively small when the threshold value is 40%. Interestingly, this mixed pattern disappears when the sample is limited to observations less likely to be switchers. A switcher is a control unit such that are more likely to be treated and thus affected more severely by the revision as time goes by. In the context of the analysis here, switchers are those observations which are assigned to control group or other prefectures but would deviate from a local job market and travel to urban prefectures such as Tokyo or Osaka to fare better opportunities. If these switchers are included in control group, a DID estimate would be contaminated, to the extent that they are sampled in a control group (Chaisemartin and Haultfoeuille, 2015). Although details are relegated to the heterogeneity analysis later, dataset used in panel B of Table 5 (University Coop Survey) disproportionately samples those public universities likely to attract switcher students, while job placement rate in panel A is taken from a population survey (School Basic Survey). Section 5.5 along with Table 8 will discuss and confirm the monotonicity of the estimates in threshold values in the selected sample.

5.4

IV Approaches and Evaluations

Another important concern on the main difference-in-differences (DID) approach is fuzziness in treatment definition. Under the approach of this paper, the treated or affected prefecture, I(Sp10 ≥ 0.4) was defined by those prefectures with no less than 40% of universities reporting that the third (ju-

22

nior) year students had already started job search as of the first week in December. Thus, some departments may have been included in the control group, when they were in fact heavily affected by the guideline revision, thus would be more likely to be treated as time goes by. In such a fuzzy design, the DID estimator represents the true treatment effects discounted by the extent of an increase in such switchers in control group, thereby being biased (Chaisemartin and Haultfoeuille, 2015). To recover the true treatment effect, Chaisemartin and Haultfoeuille (2015) proposes estimators under relatively weak assumptions, where the counterfactual time trends are constructed for both switchers and untreated units in control group. Unfortunately, data set used in this paper does not allow us to exploit the search timing information at student-level, posing an empirical challenge in directly defining a control group with stable treatment. Instead of recovering the point estimate, therefore, this and the next sections take several alternative approaches to check the consistency of the estimates across fuzzy and less-fuzzy designs. Although the limitation in the data prevents us from obtaining a precise estimate of the policy effect, the alternative approaches allows us to infer the direction of the bias and also conservative bounds of policy impact. This section employs a framework in which the continuous change in the timing of job search is exploited to identify the model under relatively weak assumptions which are likely to hold in the setting of this paper. Specifically, I estimate the first difference instrumental variable (FD-IV) model, with lagged timing variable as an instrument to allow for contemporaneous correlation between the error term and the variable of our interest. Instead of estimating a reduced form model such as equation (1), FD-IV model directly relates the proportion of colleges on the market as of the first week in December to the outcome variable Yit : Yit = α0 + γ0 Spt + xpt β0 + μi + εit

(3)

where Spt is the time-variant version of Sp10 in equation (1), namely, the the proportion of colleges at prefecture p responding students in a class of year t had already started their job search as of the first week of December. To support the hypothesis, γ0 is expected to take a negative value, as the early start of job search, represented by an increase in Spt , increases the number of stages to form the match and thus the reservation threshold for both parties, thereby reducing job placement rate at a graduation. To obtain a consistent estimate for γ0 , the paper proposes a model with relatively weak assump23

tions that are likely to hold under the setting of this paper. In particular, I employ the following FD-IV method, to allow the contemporaneous correlation between Spt and εit . This method takes the first difference between years t and t − 1 in equation (1): ΔYit = γ0 ΔSpt + Δxpt β + Δεit

(4)

ΔSpt is then instrumented with a two year lag of the timing variable,Spt−2 . The model is estimated by 2SLS. While FD-IV approach comes with a loss of efficiency, it provides consistent estimates under relatively weak and reasonable assumptions. To see this point, note that the identification of a model in equation (2) requires the following two assumptions: E(Δεit Spt−2 ) = 0

(5)

E(ΔSpt Spt−2 ) 6= 0

(6)

The first assumption may be relatively easily satisfied, since it only requires strict pre-determinedness in the variable of interest: E(εit Sps ) = 0 for ∀t > s.36 Thus, the model allows for contemporaneous correlations between the error term and the timing variable. The second assumption is justified, if the timing to start job search significantly affected the extent of timing change between the following two years. In our setting, the validity of instrument comes from the exogenous timing variation due to the guideline revision, since our sample is limited to classes of 2010 to 2013, two years before and after the revision (i.e., t = 2012 or 2013 and t − 2 = 2010 and 2011).37 Thus, in order Spt−2 to be a valid instrument, the guideline revision must have affected more severely on departments in those prefectures where students started job search relatively early prior to the revision. This is a similar identification variation exploited in DID approach. The first two columns of Table 6 present results from FD-IV estimation. First stage estimate shown in column (1) confirms that the lagged proportion, Spt−2 , is significantly related to ΔSpt . The estimate suggests that 10 percentage point increase in Spt−2 translates into a 2.7 percentage point reduction in ΔSpt , consistent with the idea that a prefecture with more schools attending the job 36

The error term εit is assumed to have no serial dependence. This is also due to the data availability:MEXT conducted the same survey prior to 2010; however, they measure the timing of job search on different categories; they did not conduct the survey for students in class of 2014. 37

24

market at an early stage saw a more intensified reduction in the proportion of those in the market within the following years. The explanatory power in the first stage is sufficiently high enough to pass the weak identification test. Second stage estimate in column (2) is consistent with findings in the baseline DID result in table 3. Ten percentage point decrease in changes in the proportion of those on the job market as of the first week of December, thus, a delay in the timing of job search, significantly increases the changes in job placement rate by 1.7 percentage point. Column (5) shows an OLS estimate in equation (3). The estimate is much smaller and insignificant, implying that the OLS estimate is upward biased. 38 To infer the extent of bias arises from the fuzzy definition of the treatment, Table 6 also compares the results from an IV version of DID model. Specifically, I estimate an equation (3) by instrumenting Spt with Dt ∗ I(Sp10 ≥ 0.4). Since the FD-IV model entails less fuzzy policy variable in exploiting the continuous information on the Spt , the difference in the estimates should suggest the sign and the magnitude of the treatment effect for the switchers in other prefectures, provided that the treatment increases over time in other prefectures (Chaisemartin and Haultfoeuille, 2015, see Theorem 3.1). Columns (3) and (4) show the estimates from DID-IV model. Both FD-IV and DID-IV estimations point to the same conclusion that the delayed timing of job search significantly increased job placement rate. Similar to the FD-IV case, the first stage estimate is significantly negative and has a high explanatory power. Consistent with the baseline DID strategy, the estimate suggests that prefectures that had initially experienced a severer advancement in the timing also saw a larger decline in the share of colleges in the local labor market as of the first week in December. More importantly, the second stage estimate in column (4), -0.152, is significant and comparable to the one obtained in FD-IV approach, -0.172 in column (2), although the FD-IV estimate is slightly larger in magnitude than the DID-IV estimate. Because FD-IV approach exploits more continuous variations in a policy variable than DID approach at prefecture-level, results in Table 6 suggest that the estimated DID impact in a fuzzy design is biased downward. Of course, FD-IV design still entails a fuzziness in that it does not account for the possibility that students within a control prefecture deviate from local labor market and will be affected by a treatment prefecture. However, Table 6 is still suggestive in that prefecture-level deviations from the DID treatment definition narrows down the actual comparison between treatment and control group, implying that the baseline DID approach provides one of 38 I also estimated the same FD-IV model where Spt is replaced by Wpt , the timing to start job search in terms of the average number of weeks to graduation within a prefecture p. The obtained implication is quite similar to those in Table 6.

25

the lower bounds for treatment effect. One useful aspect of the IV approaches is that it allows us to evaluate the actual impact of guideline revision by constructing the Wald estimate of the policy change. Back of the envelop calculation suggests that the revision increased job placement rate by at least 4 percentage point (mean of job placement rate = 0.568).39 Again, due to rather fuzzy design, this is merely a conservative evaluation, and the actual magnitude of the revision impact can be larger. Unfortunately, since University Coop Survey constitutes from repeated cross-sectional student-level data, it is not possible to conduct FD-IV analysis on students’ investment decision and infer the bias, unless the data is aggregated into university-level panel data. I chose not to take this strategy, as the aggregation shrinks a sample size and causes a serious power issue in the estimation. 40 Instead, the next section analyses heterogeneity in treatment effect and examine the robustness against a fuzzy design.

5.5

Heterogeneous Effects

The strategy in this section is to exclude those observations which are more likely to deviate from control status as time goes by and see if the revision impact is concentrated in particular groups. This is done through two types of analysis. First analysis excludes those who are more likely to be switchers, according to the characteristics of colleges. The college traits are matched with the original observations by school identifications available in outcome datasets. Second analysis is to detect a possibility that treatment effect spills over beyond prefecture borders. Similar to the first case, any spillover effect can potentially makes the treatment definition fuzzy, and masks the true causal effect. The second analysis therefore estimates the impact by excluding the observations in the control prefectures geographically adjacent to the treatment group or affected prefectures. The first analysis employs two measures to specify whether the observation is likely to belong to colleges with many switchers. One is a number of graduates from each university who are board members in companies listed in Tokyo Stock Exchange as of 2014. 41 Companies listed in Tokyo Stock Exchange often locates their headquarters in Tokyo, whereas Tokyo Prefecture was classified into a treatment group. Assuming that those universities who have many graduates who are currently board 39

Average shares of colleges on the market as of the first week of December in student’s junior year are 0.355 before revision announcement, and 0.091 after the announcement. From the DID-IV estimate in column (4) of Table 6, the effect of revision is equivalent to 4.011 percentage point change in job placement rate. 40 The number of university included in the analysis is 94 each year. As University Coop Survey comes only with a rough department code, STEM or Non-STEM, the department-level panel cannot be constructed. 41 The numbers of graduates are taken from the “Quarterly Company Handbook for Students (2014, Shushoku Shikiho, Toyo Keizai)”.

26

members have relatively large hiring networks, students in universities with many of such graduates are likely to be treated even when they are classified into a control group. The other measure classifies observations into private or public universities. Public universities in Japan constitute from elite schools in local areas, while a majority of private universities accept relatively lower ability students, with exceptions of some large-scale private universities in urban areas. Provided that students from public universities in non-urban regions often travel to urban areas such as Tokyo to seek for better job opportunities, and thus are likely to be affected by regions with most advanced timing, they are more likely to be switchers. Table 7 presents results from the same specification in Table 3, but by limiting observations according to the two measures. Column (1) of each panel replicates the baseline results in Table 3. Columns (2) to (5) restricts the observations, as are shown at the top of each column. N (board) in the table represents for the number of current board members in listed companies who graduated from the university. Results in Panel A show that the positive treatment effect on job placement rate are concentrated on observations less likely to be switchers. As shown column (2), the estimate becomes slightly larger than the baseline estimate, when observations are limited to those with no more than 100 alumni currently on the companies’ board. If the observations are further limited to those with N (board) ≤ 1 or to private universities, the estimates increases further. The impact is largest in column (4), 0.298, almost a half larger than the baseline. No significant effect was found when the sample is limited to departments in public universities in column (5). The results here again suggest that the baseline DID approach biases the true treatment effect downward due to its fuzzy treatment definition. The rest of the table repeats the same exercise for the number of days on campus and reading hours, separately for the third and the second year students. Panel B1 reports that the treatment effect on the number of days on campus per week for the third year students are not significant, in all specifications except in column (5). In column (5), public university students are found to sigfniciantly increase the number of days on campus. However, this effect is unlikely to be a treatment effect, since students in public universities often travel to urban areas and thus severely affected by the revision regardless of the treatment status. More important to the quality of the students’ investment at school, no positive treatment effect was found on the third year students’ hours of reading. In fact, as shown in column (4) of Panel B1, students in private universties significantly decreased the hours of

27

reading. This result may reflect the fact that students delayed the timing to read the guidance books which includes tips for interviews and information on firms. Results for the second year students are quite consistent with our baseline. As before, the second year students in affected region significantly decreases the number of days on campus after the revision announcement, and this effect becomes larger in columns (2) to (4) of Panel B2. No positive effect was observed when the sample is limited to public universities. Similar to the third year students, no evidence was found that the treatment increased reading hours for the second year students. Given these heterogeneity in treatment effect, Table 8 revisits the robustness test against various threshold values as was done in Table 5. Contrary to Table 5, the samples in Table 8 are limited to observations belonging to private universities. As Table 7 has already confirmed, results in Table 8 also present the larger impact of the treatment. In contrast to the previous threshold test, it is more evident that the negative impact on the number of days on campus is intensified as the threshold values increases (Panel B). Although the estimate in column (6) of Panel B is not significant (p = 0.113), the size of the estimates are mostly consistent with the monotonicity implied in treatment definition. Since dataset used in the investment analysis disproportionately samples students from public universities (see section 4.3), they are likely to include a relatively large group of switcher students. While the specification here is still not admittedly fuzzy-free, one potential interpretation for the mixed pattern in Table 5 is that switchers were distributed heavily and unevenly across the threshold values. The second analysis checks the robustness against potential spillover effect. Suppose students in a control prefecture observed that job hunting had already started in an adjacent prefecture which was included in a treatment group. These students could travel to the treatment prefecture and engage in job search quite in advance, thereby affected more severely by the guideline revision. If this was the case, similar to the fuzziness in treatment definition, we will need to adjust the baseline DID estimate up to the extent that such students were included in control groups. For this reason, I estimate the same model as before by excluding observation in those control prefectures which locate geographically adjacent to treatment prefectures. Results are shown in Table 9. By excluding the control units potentially exposed to the spillover effect, I found similarly significant but larger impacts of the treatment. In particular, Panel A shows that the positive treatment effect on job placement rate is 0.0363, substantially higher than the

28

baseline estimate. In column (5) of Panel B, the negative effect on the number of days on campus found among the second year students is also larger than the baseline. Interestingly, these negative effects are even larger than some of the heterogeneous effects observed in Table 7, implying that geographical location is relatively important in specifying the bias due to the switchers in control group. Contrary to the baseline results, columns (3) and (4) show significantly negative treatment effect on some proxies for the third year students’ investment decision. Again, no positive impact of the delayed timing was found on the proxies for human capital investment.

5.6

Discussion on Match Quality

The positive impact observed on job placement rate can be interpreted in several ways. Importantly, consequences on the match quality depends on which of these mechanisms is most dominant. As discussed in section three, if the shorter search horizon induced students and firms to reduce their reservation value, the quality of the match would be lowered since a reduction in reservation values deters sorting in the labor market. In contrast, if the shorter search horizon encouraged market participants to exert higher search efforts, the quality of the match could be improved. Unfortunately, the evidence so far is limited in explicitly differentiating one mechanism from the other. This section presents some side evidence on search effort and briefly discusses the potential impact on match quality. To complement the analyses of this paper, Appendix Table 1 applies the same DID analysis to some rough measures on search effort. The information is taken from University Coop Survey. Similar to the previous analysis on human capital investment, I estimate student-level DID model in equation (2). Proxies for search efforts are taken from the four questions in the survey; (1)do you behave consciously about your job hunting?; (2) have you taken any actions to obtain your desirable positions or jobs?; (3) do you have any plans to spend money on your job search in the next six months? ; (4)do you have any plans to spend money on training school in the next six months? The last question is intended to account for the possibility that some students in Japan take course in training schools to get ready for interviews and writing applications. The questionnaire survey unfortunately does not include the information on exact amount of search costs. Thus, I estimate the linear probability model version of DID effect, by replacing the dependent variable in equation (2) by a dummy variable indicating “Yes” for each of the four questions. While search intensity

29

is essentially unobservable, the analysis here aims to clarify any possibilities that the revision has induced students to exsert higher search efforts. All the estimates shown in Appendix Table 1, separately for the third and the second year students, are close to zero and statistically insignificant. Thus, both third and second year students did not increase their search efforts in response to the treatment, at least in terms of these rough measures. Although these proxies are admittedly crude, they point to a possibility that market participants reduced their reservation values in response to the declined search horizon.

6

Conclusion

Entry-level labor markets have observed numerous attempts to delay the timing of job search, despite a lack of empirical evidence. This paper offers the first evidence to examine the consequence of timing regulation. While timing regulations often found to fail, I exploit a unique case of new college graduates in Japan, where a guideline revision has successfully delayed the timing of job search and forced market participants to search under the shorter horizon. Based on an administrative survey, I first observed that the revision announcement indeed delayed the timing of job search and also declined the search duration, at least during the sample period analyzed in this paper. I then estimated a difference-in-differences model to measure the impact of the guideline change on the employment rate at the time of graduation, along with some proxies for human capital investment. The most conservative estimate suggests that the guideline revision have increased the employment rate by 4 percentage point (mean of job placement rate = 0.568). On the other hand, no positive effect is observed on students’ human capital investment. These results are robust against changing the threshold values in treatment definition, or excluding observations more likely to deviate from the treatment status. The placebo test detected no pre-existing trend. Although the analyses in this paper are not to specify the exact mechanisms behind the policy change, the results are highly indicative that the revision altered job search behavior of students and firms. Some caveats apply to interpreting the results in this paper. First, this paper did not provide any welfare analyses. The positive impacts on the employment rate can be associated with both increase or decrease in match quality. If the shorter search horizon induced students and firms to reduce their reservation values, the quality of the match would be lowered since a reduction in reservation value deters sorting in the labor market. In contrast, if the shorter search horizon 30

encouraged market participants to exert higher search efforts, the quality of the match could be improved. Unfortunately, the evidence in this paper does not allow us to conclude which dominates the other, although side evidence on search efforts points to the former possibility. Second, the identification variation exploited in this paper is relatively small. The revision delayed the timing of job search only for a couple of weeks on the margin of identification. Thus, it may not be a big surprise that no significant gains in human capital investment was observed. If the exploited change in the timing of job search is larger, students may well change their investment decision, depending on how their investment dynamically complements their investment later on. Nonetheless, it is still surprising to observe a significant and relatively large increase despite rather small identification variation. In fact, the discussion on mechanisms in this paper clarified the points that even a small decline in search horizon can induce a significant drop in reservation value.

References Ambuehl, Sandro and Vivienne Groves, “Unraveling Over Time,” 2015. Avery, Chirstopher, Christine Jolls, Richard Posner, and Alvin Roth, “The New Market for Federal Judicial Law Clerks,” The University of Chicago Law Review, 2007, 74 (1), 447–486. Avery, Christopher, Christine Jolls, Richard a. Posner, and Alvin E. Roth, “The Market for Federal Judicial Law Clerks,” The University of Chicago Law Review, 2001, 68 (2), 793–902. Cabinet Office, Reports on Students’ Job Search Activity (in Japanese) 2016. Caliendo, Marco, Konstantinos Tatsiramos, and Arne Uhlendorff, “BENEFIT DURATION, UNEMPLOYMENT DURATION AND JOB MATCH QUALITY: A REGRESSIONDISCONTINUITY APPROACH,” Journal of Applied Econometrics, 2013, 28, 604–627. Chaisemartin, Cl´ ement De and Xavier D Haultfoeuille, “Fuzzy Differences-in- Differences,” 2015. Cox, James C and Ronald L Oaxaca, “Laboratory Experiments with a Finite Horizon Job-Search Model,” Journal of Risk and Uncertainty, 1989, 2, 301–330. and

, “Laboratory Experiments with a Finite Horizon Job-Search Model,” Journal of Risk and

Uncertainty, 1989, 2, 301–330. 31

Damiano, Ettore, Hao Li, and Wing Suen, “Unravelling of dynamic sorting,” Review of Economic Studies, 2005, 72 (4), 1057–1076. Duflo, Esther, “Schooling and labor market consequences of school construction in Indonesia: Evidence from an unusual policy experiment,” American Economic Review, 2001, 91 (4), 795–813. Fainmesser, Itay P., “Social networks and unraveling in labor markets,” Journal of Economic Theory, jan 2013, 148 (1), 64–103. Farber, Henry S. and Rrobert G. Valletta, “Do extended unemployment benefits lengthen unemployment spells? Evidence from recent cycles in the U.S. labor market,” Journal of Human Resources, 2015, 50 (4), 873–909. , Jesse Rothstein, and Robert G. Valletta, “The effect of extended unemployment insurance benefits: Evidence from the 2012-2013 phase out,” American Economic Review, 2015, 105 (5), 171–176. ¨ Fr´ echette, Guillaume R, Alvin E Roth, and M Utku Unver, “Unraveling yields inefficient matchings: evidence from post-season college football bowls,” RAND Journal of Economics, 2007, 38 (4), 967–982. Genda, Yuji, Ayako Kondo, and Souichi Ohta, “Long-Term Effects of a Recession at Labor Market Entry in Japan and the United States,” Journal of Human Resources, 2010, 45 (1), 157–196. ¨ Haruvy, Ernan, Alvin E. Roth, and M. Utku Unver, “The dynamics of law clerk matching: An experimental and computational investigation of proposals for reform of the market,” Journal of Economic Dynamics and Control, mar 2006, 30 (3), 457–486. JILPT, “Survey Reports on College Students’ Job Search Activity (in Japanese),” Technical Report 17, The Japan Institute for Labour Policy and Training 2006. , “Surveys on Career Supports at Post-secondary institutions (in Japanese),” Technical Report, The Japan Institute for Labour Policy and Training 2014. Kaji, Sahoko, Japanese Universities Need Brighter Ideas, https://www.ft.com/content/bb4cb8107717-11e5-a95a: Financial Times, 2015.

32

Kondo, Ayako, “Does the first job really matter? State dependency in employment status in Japan,” Journal of the Japanese and International Economies, sep 2007, 21 (3), 379–402. Niederle, Muriel and Alvin E. Roth, “The Effects of a Centralized Clearinghouse on Job Placement, Wages, and Hiring Practices,” NBER Chapters, 2009, pp. 235 – 271. Petrongolo, Barbara, “The long-term effects of job search requirements: Evidence from the UK JSA reform,” Journal of Public Economics, 2009, 93 (11-12), 1234–1253. Recruit, “White Paper on Job Search Activities (in Japanese),” Technical Report, Recruit Co., Ltd. 2015. Roth, AE Alvin E and Xiaolin Xing, “Jumping the gun: Imperfections and institutions related to the timing of market transactions,” The American Economic Review, 1994, 84 (4), 992–1044. van Ours, Jan C. and Milan Vodopivec, “Does reducing unemployment insurance generosity reduce job match quality?,” Journal of Public Economics, 2008, 92 (3-4), 684–695.

33

Table 1: Summary Statistics Variable Panel A. Job Search Survey (college-level) First week of job search (in # of weeks to graduation) Class of 2010 Class of 2011 Class of 2012 Class of 2013 Duration of job search (in # of weeks) Class of 2010 Class of 2011 Class of 2012 Class of 2013 Week of job offers (in # of weeks to graduation) Class of 2010 Class of 2011 Class of 2012 Class of 2013 Panel B. School Basic Survey (department-level) Job placement rate (employment rate at graduation) Class of 2010 Class of 2011 Class of 2012 Class of 2013 Panel C. University Coop Survey (student-level) 3rd year students Number of days on campus last week Readings per day (hour) Number of books purchased in the past one month Plan to study abroad in the next six months (= 1) 2nd year students Number of days on campus last week Hours of readings per day Number of books purchased in a past month Plan to study abroad in the next six months (= 1)

34

Obs

Mean

Std. Dev.

Min

Max

920 1069 1047 1083

63.67 64.37 58.57 55.71

10.17 11.32 10.14 8.86

35 34 34 34

77 77 76 76

845 1021 1018 1071

22.00 23.98 19.08 15.43

9.07 10.03 9.85 8.89

0 0 0 0

42 43 42 42

862 1024 1026 1078

41.98 40.71 39.54 40.36

5.77 5.69 5.59 5.73

35 34 34 34

64 63 63 63

1070 1085 1116 1124

.54 .55 .57 .61

.16 .17 .16 .16

0 0 0 0

1 1 1 1

7523 5159 5117 5169

4.24 .49 1.26 .02

1.18 .7 1.96 .13

0 0 0 0

7 7 30 1

8733 7132 7064 7027

4.81 .49 1.26 .02

.86 .68 1.98 .13

0 0 0 0

7 7 30 1

0

.2

.4

cdf

.6

.8

1

(A) Timings to start job search

80

70

60 50 number of weeks to graduation class of 2010 class of 2012

40

30

class of 2011 class of 2013

0

.2

.4

cdf

.6

.8

1

(B) Job search duration (in weeks)

0

10

20 number of weeks class of 2010 class of 2012

30

40

class of 2011 class of 2013

Figure 1: Cumulative Distributions of Timing and Duration for Job Search Note: Figures show cumulative distributions of timing in job search and job search duration for each graduation year. The data comes from university-level responses in the Questionnaire Survey on Students’ Job Search Activity, conducted by the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT). Graduation year indicates that the students graduated from the universities in March of that year.

35

.8 .6 cdf .4 .2 0

0

.2

.4

cdf

.6

.8

1

Class of 2013

1

Class of 2010

80

70 others

60

50

40

30

80

affected prefectures

70 others

60

50

40

30

affected prefectures

Timing to start job search (in number of weeks to graduation) Figure 2: Cumulative Distributions of Timings in Job Search Note: Figures show separate cumulative distributions of timing in job search for unraveling and other prefectures. Horizontal axis indicates start timing in terms of number of weeks to graduation. Green dashed line indicates 70 weeks or one year prior to graduation. Treatment status is defined by the proportion of universities in each prefecture as of 2010. In particular, “in affected prefectures” include universities in prefectures where at least 40 % universities responded that students already started job search as of first week of December in 2010. “Others” include universities in prefectures less than 40% responded that students already started job search as of first week of December in 2010. “Affected” prefectures include: Akita, Fukushima, Saitama, Tokyo, Toyama, Fukui, Shiga, Nara, Wakayama, Okayama, Tokushima, and Saga. Classes of the graduation indicate years in which students graduated from universities, often in March.

36

Table 2: Predetermined Prefecture Traits (t = 2010) by Treatment Status Variable Log prefecture population Fraction of population aged 15-64 Average monthly wage for new college graduates, in 10000 JPY Jobs-to-application ratio Unemployment rate College advancement rate Number of colleges per 104 residents

Affected 14.53 (0.123) 0.627 (0.004) 20.76 (0.147) 0.626 (0.034) 0.065 (0.002) 0.483 (0.012) 0.086 (0.008)

Note: Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

37

Others 14.40 (0.253) 0.627 (0.007) 20.95 (0.197) 0.631 (0.053) 0.065 (0.003) 0.503 (0.019) 0.089 (0.009)

Diff. 0.126 (0.256) -0.0002 (0.007) -0.194 (0.277) -0.004 (0.066) 0.0007 (0.0004) -0.020 (0.026) -0.003 (0.010)

Table 3: Baseline Estimates

Panel A. Employment rate at graduation (university-institutional level) (1) job placement rate

Dt ∗ I(Sp10 ≥ 0.4) N R-squared

0.0208** (0.0104) 4,395 0.0934

Panel B. Human capital investment (student level) Sample = 3rd year students

(1) days on campus

(2) reading hours

(3) books purchased

(4) Prob(study abroad)

Dt ∗ I(Sp10 ≥ 0.4)

0.2716 (0.2024)

-0.0687 (0.1014)

-0.2951 (0.2954)

-0.0189 (0.0143)

N R-squared Sample = 2nd year students

7,523 0.0880 (5) days on campus

7,132 0.0391 (6) reading hours

7,064 0.0390 (7) books purchased

7,027 0.0340 (8) Prob(study abroad)

Dt ∗ I(Sp10 ≥ 0.4)

-0.2161** (0.1052)

-0.0327 (0.0808)

0.0477 (0.3003)

-0.0174 (0.0336)

8,733 0.0979

8,303 0.0502

8,236 0.0394

8,095 0.0285

N R-squared

Note: Robust standard errors clustered at university-department level (panel A) or at university level (panel B) in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All models control for university-department fixed effects (panel A) or university fixed effects (panel B), along with year dummies, prefecture controls and prefecture linear trends. Prefecture controls include; logarithm of prefecture population at year t−2, jobs-to-application ratio at year t−2; average monthly wage for new college graduates at year t − 2; and university advancement rate at year t − 4. Lags in the first three

variables are taken to account for the local labor market conditions right before job hunting season. Columns (4) and (8) were estimated in a linear probability model.

38

Table 4: Placebo Tests (sample cohorts = 2010-2011) Panel A. Employment rate at graduation (university-institutional level) (1) job placement rate

d11t ∗ I(Sp10 ≥ 0.4) N R-squared

-0.0029 (0.0101) 2,155 0.0237

Panel B. Human capital investment (student level) Sample = 3rd year students

(1) days on campus

(2) reading hours

(3) books purchased

(4) Prob(study abroad)

d11t ∗ I(Sp10 ≥ 0.4)

-0.1348 (0.1467)

0.0107 (0.0531)

-0.0772 (0.1604)

0.0022 (0.0102)

N R-squared Sample = 2nd year students

3,802 0.1015 (5) days on campus

3,759 0.0485 (6) reading hours

3,697 0.0497 (7) books purchased

3,535 0.0478 (8) Prob(study abroad)

d11t ∗ I(Sp10 ≥ 0.4)

0.0989 (0.0813)

0.0553 (0.0428)

0.1556 (0.1791)

0.0015 (0.0133)

4,498 0.0991

4,443 0.0639

4,360 0.0456

4,167 0.0409

N R-squared

Note: Observations are limited to those in classes of 2010 and 2011. Robust standard errors clustered at universitydepartment level (panel A) or at university level (panel B) in parentheses. *** p <0.01, ** p<0.05, * p<0.1. All models control for university-department fixed effects (panel A) or university fixed effects (panel B), along with year dummies, prefecture controls and prefecture linear trends. Prefecture controls include; logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2; average monthly wage for new college graduates at year t − 2; and

university advancement rate at year t − 4. Lags in the first three variables are taken to account for the local labor market conditions right before job hunting season.

39

Table 5: Alternative Definitions of Treatment Status (Sp10 )

Panel A. Job Placement Rate (department-level) Sp10 Dt ∗ I(Sp10 ≥ 0.4) N R-squared

(1) = 36%

(2) = 38%

(3) = 40%

(4) = 42%

(5) = 44%

(6) = 46%

0.0091 (0.0094)

0.0060 (0.0096)

0.0208** (0.0104)

0.0231** (0.0108)

0.0232** (0.0115)

0.0235* (0.0122)

4,395 0.0927

4,395 0.0926

4,395 0.0934

4,395 0.0935

4,395 0.0935

4,395 0.0935

Panel B. Number of Days on Campus (2nd year, student-level) Sp10 Dt ∗ I(Sp10 ≥ 0.4) N R-squared

(1) = 36%

(2) = 38%

(3) = 40%

(4) = 42%

(5) = 44%

(6) = 46%

-0.4054*** (0.1249)

-0.3259*** (0.1091)

-0.2161** (0.1052)

-0.2921*** (0.1068)

-0.3165*** (0.1092)

-0.2209* (0.1289)

8,354 0.1007

8,354 0.1000

8,354 0.0992

8,354 0.0995

8,354 0.0995

8,354 0.0990

Note: Robust standard errors clustered at university-department level (panel A) or at university level (panel B) in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All models control for university-department fixed effects (panel A) or university fixed effects (panel B), along with year dummies, prefecture controls and prefecture linear trends. Prefecture controls include; logarithm of prefecture population at year t−2, jobs-to-application ratio at year t−2; average monthly wage for new college graduates at year t − 2; and university advancement rate at year t − 4. Lags in the first three

variables are taken to account for the local labor market conditions right before job hunting season.

40

Table 6: IV Estimations on Job Placement Rate: DID and First-Difference (FD) Approaches (1)

(Dependent variable) Spt−2

(2) FD-IV 1st stage 2nd stage (ΔSpt ) (ΔYpt )

(3)

(4) DID-IV 1st stage 2nd stage (Spt ) (Ypt )

(6)

(ΔYpt )

(Ypt )

-0.272*** (0.0179)

ΔSpt

-0.176*** (0.0414)

-0.0062 (0.0194)

Dt ∗ I(Sp10 ≥ 0.4)

-0.0867*** (0.00798) -0.152** (0.0637)

Spt N First-stage F stat

(5)

2,142 231.04

2,142

4,395 239.79

4,395

-0.0041 (0.0245) 2,142

4,395

Note: Ypt = job placement rate. Robust standard errors clustered at institutions in parentheses. *** p <0.01, ** p<0.05, * p<0.1. First-stage F stat represents for Cragg-Donal Wald F-statistics for weak identification test. All models control for year dummies and prefecture controls. In columns (3), (4) and (6), prefecture controls include; logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2; average monthly wage for new

college graduates at year t − 2; and university advancement rate at year t − 4. In columns (1), (2) and (5), prefecture

controls include first-differenced items from the same variables. Lags in the first three variables are taken to account for the local labor market conditions right before job hunting season. Department dummies are included in columns (3), (4) and (6). They are differenced out in columns (1), (2) and (5).

41

Table 7: Heterogenous Effects

Panel A. Employment rate at the time of graduation (department-level)

Job placement rate

N

(1) Baseline

(2) N (board) ≤ 100

(3) N (board) ≤ 1

(4) private univ

(5) public univ

0.0208** (0.0104)

0.0218* (0.0114)

0.0230* (0.0120)

0.0298** (0.0125)

0.0051 (0.0162)

4,395

3,918

3,719

3,495

900

Panel B1. Human capital investment (student-level, 3rd year students)

Days on campus

N Reading hours

N

(1) Baseline

(2) N (board) ≤ 100

(3) N (board) ≤ 1

(4) private univ

(5) public univ

0.2716 (0.2024)

0.1638 (0.2537)

-0.0891 (0.2531)

0.1245 (0.2226)

0.5026** (0.2344)

7,523

4,618

3,980

3,432

4,091

-0.0687 (0.1014)

-0.1048 (0.1131)

-0.1656 (0.1386)

-0.3215*** (0.1153)

0.0225 (0.1310)

7,132

4,386

3,789

3,269

3,863

Panel B2. Human capital investment (student-level, 2nd year students)

Days on campus

N Reading hours

N

(1) Baseline

(2) N (board) ≤ 100

(3) N (board) ≤ 1

(4) private univ

(5) public univ

-0.2161** (0.1052)

-0.2673* (0.1396)

-0.3368** (0.1499)

-0.4583* (0.2271)

-0.1322 (0.1193)

8,733

5,333

4,611

4,196

4,537

-0.0327 (0.0808)

0.0978 (0.0837)

0.0739 (0.0869)

-0.0574 (0.1030)

-0.0099 (0.0931)

8,303

5,065

4,380

4,005

4,298

Note: Robust standard errors clustered at institutions in parentheses. *** p <0.01, ** p<0.05, * p<0.1. Each column represents for the coefficients estimate of Dt ∗ I(Sp10 ≥ 0.4). All models control for prefecture fixed effects, prefecture

linear trends, and prefecture controls. Prefecture controls include; logarithm of prefecture population at year t − 2,

jobs-to-application ratio at year t − 2; average monthly wage for new college graduates at year t − 2; and university

advancement rate at year t − 4.

42

Table 8: Alternative Definitions of Treatment Status, private universities only

Panel A. Job Placement Rate (department-level) Sp10 Dt ∗ I(Sp10 ≥ 0.4) N R-squared

(1) = 36%

(2) = 38%

(3) = 40%

(4) = 42%

(5) = 44%

(6) = 46%

0.0105 (0.0110)

0.0069 (0.0113)

0.0298** (0.0125)

0.0320** (0.0132)

0.0307** (0.0140)

0.0306** (0.0148)

4,395 0.1188

4,395 0.1186

4,395 0.1202

4,395 0.1204

4,395 0.1201

4,395 0.1201

Panel B. Number of Days on Campus (2nd year, student-level) Sp10 Dt ∗ I(Sp10 ≥ 0.4) N R-squared

(1) = 36%

(2) = 38%

(3) = 40%

(4) = 42%

(5) = 44%

(6) = 46%

-0.4024 (0.3590)

-0.4356* (0.2461)

-0.4583* (0.2274)

-0.5126** (0.2208)

-0.5126** (0.2208)

-0.4643 (0.2857)

3,495 0.0978

3,495 0.0983

3,495 0.0982

3,495 0.0985

3,495 0.0985

3,495 0.0977

Note: Robust standard errors clustered at institutions in parentheses. *** p <0.01, ** p<0.05, * p<0.1. All models control for prefecture fixed effects, prefecture linear trends, and prefecture controls. Prefecture controls include; logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2; average monthly wage for new college graduates at year t − 2; and university advancement rate at year t − 4.

43

Table 9: Spillover Effect

Panel A. Employment rate at graduation (university-institutional level) (1) job placement rate

Dt ∗ I(Sp10 ≥ 0.4) N R-squared

0.0363** (0.0169) 2,312 0.1033

Panel B. Human capital investment (student level) Sample = 3rd year students

(1) days on campus

(2) reading hours

(3) books purchased

(4) Prob(study abroad)

Dt ∗ I(Sp10 ≥ 0.4)

-0.0732 (0.2508)

-0.1827 (0.1369)

-0.6587** (0.3255)

-0.0220* (0.0130)

N R-squared Sample = 2nd year students

3,751 0.0888 (5) days on campus

3,590 0.0471 (6) reading hours

3,543 0.0440 (7) books purchased

3,518 0.0468 (8) Prob(study abroad)

Dt ∗ I(Sp10 ≥ 0.4)

-0.4925** (0.1954)

-0.0137 (0.0973)

-0.1127 (0.4021)

0.0153 (0.0504)

4,671 0.1246

4,431 0.0594

4,389 0.0341

4,342 0.0319

N R-squared

Note: Estimations exclude control observations in prefectures geographically adjacent to the treatment prefectures. Robust standard errors clustered at university-department level (panel A) or at university level (panel B) in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All models control for university-department fixed effects (panel A) or university fixed effects (panel B), along with year dummies, prefecture controls and prefecture linear trends. Prefecture controls include; logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2; average monthly wage

for new college graduates at year t − 2; and university advancement rate at year t − 4. Lags in the first three variables

are taken to account for the local labor market conditions right before job hunting season.

44

Appendix A. Further Background Information The market unraveling in college graduates job market has a long history in Japan. Similar to the other entry-level labor market such as the one for law clerks in the US, there have been numerous efforts to delay the timing, although they usually do not last too long. These agreements were mostly initiated by Japanese government, namely, Ministries of Education and Labour, requesting universities and employers associations to enforce the uniform dates in the hiring schedule. The first such attempt dates back to 1953, when the ministries, universities and business associations set up the recruitment agreement (Shusyoku Kyotei ) and specified the first dates of sending recommendations/firm visits and having interviews. As was the case in other markets, however, the agreements were no more than a gentlemen agreement and not legally enforced. They often found themselves reneged by some firms giving informal offers before the agreed date. Roughly counting, there have been at least four major waves to set up or review the agreement since 1953. 42 This paper makes use of the forth of such attempts where the revision indeed postponed the actual date of job search, at least temporarily after 2010. Just to visiualize an idea how job hunting seasons have shifted in the past decades, Appendix Figure 1 shows the number of newspaper articles related to students job search activities. The number of articles is obtained in a keyword search in major newspaper databases, Nikkei Telecom, by typing job search Activity for new graduates (Shusyoku Katsudo or Shu Katsu).43 As newspapers often post new hiring plans of major firms, seminar information, as well as columns and advices on job search activities, the obtained figures provide a rough measure when the market hits its peak in job search activities every year. The graphs show the enormous change in the timing of job hunting in the last several decades. In the early 1990s, the number of related articles peaked in July, eight months before graduation. However, the timing of job search moved into the winter of the third year in the late 1990s, especially from 1995 to 1997, corresponding to the fact that the recruit agreement ended up being abolished at the end of 1996. It continued to advance further, and in the 2000s, students were looking for their jobs, sometimes as early as in October, a beginning of the second semester of 42

Those include; (1) recruitment agreement initiated in 1953; (2) a revision of the agreement in 1972; (3) series of attempts to revise and enforce the recruitment agreement by Recruiting Compliance Association, starting in 1982; (4) abandonment of the recruitment agreement in December, 1996; and set up a new guideline ( Rinri Kensho), which was announced to be revised in 2010 to specify the uniform dates for the first time since 1997. 43 I exclude those articles unrelated to the job searches of new college graduates (e.g., job search for new high school graduates) and also those related to policy debates to regulate the date of job offers. Further details can be found on the footnote of the figures.

45

third year.44

Appendix B. Construction of Timing Variable This section briefly explains how the main variable in the analysis, the timings to start job search is constructed from Job Search Survey (MEXT). Since the timing data at hand does not contain university identification codes due to information discloser policy, the constructed variable is aggregated into prefecture-level information, either on the proportion of universities within a prefecture that responded that students had already started their job search as of the first week of December, 2010 (I(Start2010p ) in the main analysis), or on the time-variant version of the same variable (i.e., Spt in Appendix B). The timing information was drawn from responses by a career division at each school to a question, “when did students start participating screening/interviews?”. Since the responses were recorded as a categorical variable to indicate the timing fell in either of the beginning/middle/end of specific month, they were transformed into the number of weeks to graduation. More specifically, the responses were given by the following categories: (a) before October, (b1) beginning of November, (b2) middle of November, (b3) end of November, (c1) beginning of December, ...,(i1) beginning of June, (i2) middle of June, (i3) end of June, and (j) after July. Each of these categories were transformed into the number of weeks to graduation, by counting the number of weeks for the 10th of each month if it is the beginning of the month, for the 20th of each month if it is the middle of the month, and for the last day of the month if it is the end of the month. For the first and last category, I counted the number of weeks to graduation for the 10th of October for category (a) before October, and for the last day of July for category (j) after July. Since Japanese school year starts in the 1st of April, the number of weeks to graduation is 52 for category (f3) end of March. It should be noted that the variable is censored at the first and last categories. The censoring can be found in panel A of Figure 2, where cdf’s jump at the start of accumulation. In Table 1, I show the summary statistics for the length of job search. This variable is constructed by taking differences between the response to the above-mentioned question and responses to another one that asks, “when was the job offers to the students were concentrated?” The responses to this question were also transformed from categories into the number of weeks to graduation by following the same rule. 44 An example of articles include that a major department store advanced their schedule to launch a stall for recruitment suit that college students customarily wear (Nikkei Newspaper, October 29, 2004).

46

2001

1 2 3 4 5 6 7 8 9 101112

2007

2000

1 2 3 4 5 6 7 8 9 101112

2006

1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 101112

2013

1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 101112

2012

1995

1994

1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 101112

1989

1988

1 2 3 4 5 6 7 8 9 101112

2008

1 2 3 4 5 6 7 8 9 101112

2002

1 2 3 4 5 6 7 8 9 101112

1996

1 2 3 4 5 6 7 8 9 101112

1990

1 2 3 4 5 6 7 8 9 101112

2009

1 2 3 4 5 6 7 8 9 101112

2003

1 2 3 4 5 6 7 8 9 101112

1997

1 2 3 4 5 6 7 8 9 101112

1991

1 2 3 4 5 6 7 8 9 101112

2010

1 2 3 4 5 6 7 8 9 101112

2004

1 2 3 4 5 6 7 8 9 101112

1998

1 2 3 4 5 6 7 8 9 101112

1992

1993

1 2 3 4 5 6 7 8 9 101112

2011

1 2 3 4 5 6 7 8 9 101112

2005

1 2 3 4 5 6 7 8 9 101112

1999

1 2 3 4 5 6 7 8 9 101112

suits, etc. Major shifts took place in the late 1990s and the early 2000s.

of the included articles are: information on combined seminars, featured interviews to firm personnel regarding this year’s new hiring strategies, sales on recruitment

activities for high school students, or if they talk about policy debates on the timing regulations for the next season but not about current job hunting. Examples

Telecom Database by typing job search activity for new graduates (Shusyoku Katsudo or Shu Katsu). Some articles are excluded if they are related to job search

Note: Figures indicate the number of major newspaper articles related to job search activity for new university graduates in Japan. They are obtained in Nikkei

Appendix Figure 1: Long-term Shifts in Job Hunting Season, in terms of related newspaper articles

Number of Articles

0 10 20 30 40 50

0 10 20 30 40 50

0 10 20 30 40 50

0 10 20 30 40 50

0 10 20 30 40 50

47

48

0.0668 (0.0579)

7,316 0.0550 (5) attention 0.0293 (0.0580) 8,490 0.0282

Dt ∗ I(Sp10 ≥ 0.4)

N R-squared Sample = 2nd year students

Dt ∗ I(Sp10 ≥ 0.4)

N R-squared

8,495 0.0285

0.0359 (0.0559)

7,330 0.0350 (6) getting ready

0.0359 (0.0597)

(2) taking any actions

8,087 0.0695

-0.0085 (0.0246)

7,094 0.1181 (7) Prob(search cost)

-0.0094 (0.0492)

(3) Prob(search cost)

8,097 0.0223

-0.0075 (0.0323)

7,031 0.0304 (8) Prob(training class)

-0.0208 (0.0413)

(4) Prob(training class)

the local labor market conditions right before job hunting season.

average monthly wage for new college graduates at year t − 2; and university advancement rate at year t − 4. Lags in the first three variables are taken to account for

prefecture controls and prefecture linear trends. Prefecture controls include; logarithm of prefecture population at year t − 2, jobs-to-application ratio at year t − 2;

Note: Robust standard errors clustered at universities in parentheses. *** p <0.01, ** p<0.05, * p<0.1. All models control for university fixed effects, year dummies,

(1) act consciously

Sample = 3rd year students

Appendix Table 1: The Impacts on Proxies for Search Efforts