The Hazards of Unemployment: A Macroeconomic Model of Job Search and Résumé Dynamics ∗ Ross Doppelt November 20, 2015 Abstract I introduce a dynamic general-equilibrium model to investigate the relationship between the duration of unemployment and the probability of nding a job. In particular, I analyze the hypothesis that a long unemployment spell sends a negative signal about a worker's quality, which in turn aects the worker's probability of being hired. Each worker has a permanent skill level, either high or low. At every stage of an employment relationship, high-skill workers draw from a more favorable productivity distribution. Consequently, high-skill workers are more likely to be hired, conditional on being seen by a rm. Skills are unobservable, but a worker's history of unemployment and on-the-job productivity provide observable signals of her underlying type. I refer to a worker's posterior probability of being highly skilled, conditional on these observables, as the worker's résumé. Spending time in unemployment damages a worker's résumé: Because high-skill workers are more likely to form matches, prospective employers infer that unmatched workers are less likely to be highly skilled. Consequently, a worker's job-nding rate changes over the course of an unemployment spell as her résumé declines. The match surplus incorporates the fact that hiring a worker makes her résumé look better, and not hiring her makes her résumé look worse. Wages therefore include a compensating dierential for the résumé value of employment. I calibrate model to match data on job-nding rates as a function of duration. The average worker experiences signicant negative duration dependence, and incomplete information also generates considerable heterogeneity in job-nding rates. I extend the model to discuss how informational concerns interact with human capital decay as a source of duration dependence. Finally, I discuss the theory's empirical predictions and econometric implications. ∗ Department of Economics, Penn State University. Contact: lated under the shorter title, The Hazards of Unemployment. ross.doppelt@psu.edu. This paper previously circu- Please download the most recent version on my website: https://sites.google.com/site/rossdoppelt/. This work is based on the third chapter of my doctoral thesis at New York University; I am grateful to my thesis advisor, Tom Sargent, and committee members, Ricardo Lagos and Gianluca Violante, for their feedback and criticisms. I also beneted from discussions with Gadi Barlevy, Dhruva Bhaskar, Saki Bigio, Katka Borovi£ková, Joe Briggs, Je Campbell, Chris Flinn, Dan Greenwald, Victoria Gregory, Gregor Jarosch, Boyan Jovanovic, Nic Kozeniauskas, Elliot Lipnowski, Lars Ljungqvist, Simon Mongey, Jo Mullins, Keith O'Hara, Laura Pilossoph, João Ramos, Edouard Schaal, Shouyong Shi, and numerous seminar participants. Any imperfections are purely my own. 1 1 Introduction In this paper, I build a theoretical macroeconomic model to investigate the relationship between the duration of unemployment and the probability of nding a job. Figure A.1, constructed with data from the Current Population Survey, shows that the probability of nding a job in the coming month is about 30 percentage points lower for someone who has been looking for a year, relative to a newly unemployed worker. 1 Recent evidence from the empirical microeconomic literature, which I will discuss below, suggests that some of this correlation is causal: Fewer rms are interested in hiring applicants who have been unemployed for a long time. Economists have oered several hypotheses to explain this relationship; my focus will be on information stigma, the notion that the length of an unemployment spell sends a signal about worker quality. After searching for a certain amount of time, a job seeker may fail to be hired because no rms saw her application, or she was a poor t with all of the rms that did see her application. Prospective employers will never know exactly why an applicant has struggled to nd work, but rms will draw inferences. If good workers are more likely to be successful in making matches, then there is reason to think that someone who fails to make a match is less likely to be a good worker. In turn, an individual worker's job-nding rate will change as she spends time in unemployment because rms' demand for her services will change as the length of her unemployment spell reveals information about her. Jobless spells therefore inict a kind of résumé damage, which aects a worker's future employment and earnings prospects. I explore this hypothesis by introducing a dynamic general-equilibrium model with heterogeneous agents, incomplete information, and frictional labor markets. Workers in the model can have one of two skill levels, either high or low, and this true type is not observed by anyone in the economy. Whereas skill is a permanent feature of workers, productivity is stochastic and match-specic. The highly skilled are not necessarily more productive, but they do have a more advantageous productivity distribution. Consequently, people's labormarket outcomes reveal information about their skills at each stage of their working lives. For the unemployed to be hired, they not only need to be seen by a rm; they need to realize a productivity draw good enough to justify forming a match. After being hired, a worker's productivity continues to evolve, with the highly skilled being more likely to enjoy productivity increases, less likely to suer productivity losses, and less likely to experience separations. Of central importance is a worker's probability of being highly skilled, conditional on her history of unemployment and on-the-job performance. I will refer to this probability as a worker's résumé: Each period, beliefs about a worker's skill level get updated according to Bayes's rule, so the résumé summarizes a worker's background and experience, as they pertain to her future productivity. 1 Data details are in Appendix D. 2 Search markets are segmented by expected skill level. When rms post vacancies, they solicit applications from workers with a particular résumé, and the unemployed apply to the openings commensurate with their résumés. In the market for résumé-r workers, a Pissarides-style matching function determines how many applications land in the hands of rms. Prospective employers have access to a screening technology, which allows them to see the applicant's productivity in her rst period on the job. If the productivity draw is good, then the rm contacts the worker, and they begin production in the following period. If the draw is bad, the rm never contacts the worker, who remains unemployed with a résumé updated to reect her failure to nd a job. The job-nding hazard therefore depends on two things: (1) the relative demand for workers with dierent résumés, and (2) the speed with which a worker's résumé deteriorates during unemployment. Moreover, the amount of stigma associated with joblessness depends on how easy it is to be hired, which underscores the importance of analyzing this problem in general equilibrium. 1.1 Main Results The model allows me to make progress on three fronts. I begin by developing a novel theoretical framework for analyzing informational dynamics in frictional labor markets. Then, I use a calibrated version of the model to provide a quantitative exploration of the economic forces at work. Finally, I ush out the theory's empirical predictions and econometric implications. This approach allows me to address several questions. Informational concerns play an important role in determining match surpluses and, by extension, wages and job-nding rates. The total surplus from a match equals the expected discounted value of production, minus the worker's value of search. Both of these values depend on the worker's résumé. By construction, a worker's expected skill level provides a sucient statistic for forecasting her output. Moreover, the worker's résumé will determine her employment prospects the next time she has to search for a new job. The match surplus incorporates the fact that hiring a worker makes her résumé look better, and not hiring her makes her résumé look worse. An employee's pay is determined by a linear surplus-splitting condition. Consequently, wages include a compensating dierential for the résumé value of employment. In a sense, rms pay workers for their time and output, while workers pay for the résumé improvement they get from being employed. Using the model, I derive a wage decomposition that quanties the contribution of information to a worker's take-home compensation. Understanding the theory of how information is priced into wages is necessary for understanding the association between résumés and job-nding rates. The model suggests that a worker's probability of being hired is a hump-shaped function of her résumé. With full information, the most productive workers are in the highest demand: The market for workers who denitely have high skills is tighter than the market for 3 workers who denitely have low skills. With incomplete information, however, the tightest markets are for workers with good, but not pristine, résumés. Workers of uncertain quality are willing to take substantial pay cuts to improve their résumés, whereas workers who are known to have high skills need to be paid high wages. For rms posting vacancies, recruiting workers with mid-level résumés may result in lower expected productivity, but higher expected prots. Thus, market tightness is a non-monotonic function of résumé. After expositing the theoretical framework, I use the calibrated model to analyze the relationship between job-nding rates and unemployment durations, on both the aggregate and individual levels. Despite the clear pattern in Figure A.1, the drop in average job-nding rates does not directly imply that an individual's probability of being hired actually changes over the course of an unemployment spell. An alternative explanation is that people have heterogeneous but time-invariant job-nding rates: In that case, fast job nders would be hired after a short amount of time, so the long-term unemployed would comprise the people who had the lowest job-nding rates all along. Disentangling the relative importance of dynamic selection, due to unobserved heterogeneity, and true duration dependence, due to time-varying job-nding rates, has been a contentious issue in the literature. My model features both of these mechanisms. Not only is duration dependence endogenous, so is the distribution of résumés across workers, which determines the amount of heterogeneity in job-nding rates. For the key parameters governing stigma eects and skill heterogeneity, I calibrate the model to match the moments of the aggregate data that are of primary interest, namely, average job-nding rates, as a function of duration. Because I do not target the hazard curves of individual workers, this approach is fairly agnostic about what causes the pattern in Figure A.1. On the individual level, the results paint a rich picture of true duration dependence. A worker's résumé is a decreasing function of unemployment duration, but the job-nding rate is a hump-shaped function of the worker's résumé. Individuals will have heterogeneously shaped hazards, depending on the résumés they had when they entered unemployment. The baseline calibration suggests that the tightest market is for workers with probability .80 of being highly skilled. The vast majority of unemployed workers have résumés below this level, and for them, duration dependence is negative. That is, their job-nding rates decline as their unemployment spells progress. For a worker entering unemployment with the average résumé, the job-nding probability drops by nearly half after 52 weeks. As I discuss below, some of the most recent and persuasive empirical studies nd negative duration dependence; in this respect, the model is broadly consistent with the microeconomic evidence. But for the workers with the highest résumés, duration dependence is positive: Their job-nding rates actually go up from one period of search to the next, even though expectations about their skills are revised downward. This result makes sense in light of how résumé dynamics inuence wages. As high-résumé workers spend more time searching, they are eectively willing to take larger pay cuts in order to stanch the stigma from joblessness. Nevertheless, after a suciently long unemployment 4 spell, beliefs about a worker's productivity will eventually deteriorate to the point that her job-nding rate declines. In that sense, all workers will eventually experience negative duration dependence. On the aggregate level, information stigma is necessary for the calibrated model's ability to account for the observed correlation between unemployment durations and job-nding rates. Partly, this is a consequence of the fact that most workers have downward-sloping hazards. The model allows us to look separately at the eects of true duration dependence and dynamic selection. Mechanically, the model features a lot of dispersion in job nding rates, and this heterogeneity can generate much of the drop o in average job-nding rates, even in the absence of true duration dependence. Economically, though, stigma and heterogeneity are not orthogonal phenomena. Instead, I argue that information eects are responsible for creating much of the dispersion in job-nding rates. The dierence in job-nding rates between workers with known skill levels is much smaller than the range of job-nding rates across workers with unknown skill levels. When workers have heterogeneous skills, but there is full information about those skills, the aggregate correlation between job-nding rates and durations is much less pronounced. Human capital decay is the leading alternative to information stigma for explaining negative duration dependence, and the model can be extended to incorporate skill loss during unemployment and skill gain during employment. job-nding rates. Relative to the baseline model, this modication can actually attenuate the drop in When skills change over time, information about an unemployed worker's current skill level is less valuable. Employees are therefore less willing to accept low wages in order to improve their résumés. Ultimately, the value of forming a match depends less on the résumé of the worker when she is hired. In turn, job-nding rates become more uniform across the markets for workers with dierent expected skill levels, so individual hazards become atter. Finally, the model has useful implications for empirical research. Because information is revealed about a worker's skill level every period, a worker's type will eventually be revealed after a suciently long time in the marketplace. In the model, the résumé distribution will never degenerate because workers die and are replaced by new-born agents, whose skills are unknown. However, the model does suggest that, if duration dependence were a byproduct of learning about worker quality, then older workers would have less information revealed about them during their most recent unemployment spell. That being said, workers of dierent ages will be subject to dierent amounts of true duration dependence, as well as dierent amounts of heterogeneity in job-nding rates. Thus, it's still necessary to have a careful econometric strategy for disentangling true duration dependence from unobserved heterogeneity. Fortunately, the theory also provides guidance for the specication and identication of econometric models, in particular the mixed proportional hazard model. This framework provides a theoretical justication for tting a mixed proportional hazard model, but only when looking at workers who enter unemployment with comparable résumés. Moreover, to 5 obtain identication of the proportional hazard model, the theory places restrictions on the information sets of econometricians, relative to agents in the economy. 1.2 Advancement of the Theoretical Literature Other papers have looked at information as a source of duration dependence from a theoretical macroeconomic perspective. Examples include Jarosch and Pilossoph [2014], Fernandez-Blanco and Preugschat [2014], Acemo§lu [1995], and Lockwood [1991]. 2 Those models unanimously conclude that information stigma leads to negative duration dependence, although the calibration exercises by Jarosch and Pilossoph [2014] and Fernandez-Blanco and Preugschat [2014] suggest that dynamic selection eects are quantitatively more important. A shared limitation of these theories is that they break the connection between a worker's current job search and her future career outcomes value of information. and this modeling choice is crucial for the way agents assess the In each of the above papers, a worker's type is perfectly revealed after a match is made and production commences. This feature is coupled with one of two assumptions: Either a worker remains employed with the same job until she dies, or all information about a worker is reset once she 3 Such simplications allow authors to make progress on other fronts. However, reenters unemployment. these assumptions restrict the value of information: If a worker's planning horizon does not extend beyond her next job, then her résumé becomes irrelevant for her subsequent employment and wage prospects. In contrast, my results suggest that the résumé value of employment can be large when workers face the specter of repeat unemployment spells, so preventing the stigma of unemployment contributes signicantly to the surplus generated by forming a match. This fact is responsible for the non-monotonic relationship between expected skills and job-nding rates, which is not found in other models. To the best of my knowledge, Gonzalez and Shi [2010] provide the only other model in which learning takes place across multiple unemployment spells, but those authors nd that individuals have unambiguously upward-sloping hazards. Gonzalez and Shi construct a directed-search model in which workers learn about their search ability, which determines how good they are at nding jobs when unemployed, rather than their productive ability, which determines how good they are at creating output when employed. Qualitatively, 2 The earliest general-equilibrium treatment of this problem appears in Lockwood [1991], who proposes a random-matching model in which rms can test candidates' abilities; because these tests are imperfect, search duration provides an additional signal of worker quality. Acemo§lu [1995] combines information asymmetry with human capital depreciation by positing that workers have to exert eort, which is unobservable, to keep their skills from decaying during unemployment. More recently, Jarosch and Pilossoph [2014] propose a model with two-sided heterogeneity and positive assortative matching between workers and rms. All workers are assumed to meet rms at a constant and exogenous rate, but as a worker's search duration increases, fewer rms are willing to pay the cost of screening her. Fernandez-Blanco and Preugschat [2014] construct a contract-posting model in which rms oer wages contingent on unemployment duration. 3 Mathematically, these are almost identical. Lockwood [1991] and Acemo§lu [1995] make the former assumption; Jarosch and Pilossoph [2014] and Fernandez-Blanco and Preugschat [2014] make the latter. 6 those authors tell a story about discouragement during unemployment: As spells drag on, workers infer that their search ability is poor, so they target lower-wage jobs in order to increase the probability of being hired. However, as I discuss below, the most convincing microeconomic evidence suggests that true duration dependence is largely negative, which is dicult to reconcile with Gonzalez and Shi's prediction of upward- 4 Nevertheless, on a theoretical level, Gonzalez and Shi's result demonstrates sloping individual hazards. that negative duration dependence is not a trivial outcome in models where information about workers is revealed during unemployment spells. I also nd that some workers will have increasing job-nding rates because they are willing to accept lower wages; however, this outcome manifests in only a minority of job seekers. Relative to existing theories, my model incorporates richer dynamics for individuals' skills, as well information about those skills. One novelty of my model is the analysis of how skill change interacts with information stigma. Of the above papers, only Acemo§lu [1995] looks at both skill loss and incomplete information. In his model, skill change is necessary for the existence of duration dependence; in mine, skill change can make duration dependence less pronounced. Finally, all of the above models assume that learning takes place exclusively in unemployment. The framework that I propose allows beliefs to evolve following hirings, separations, and changes in productivity. These features of the model respect the large body of evidence suggesting that information about a worker is revealed throughout her career. 5 It also becomes easier to interpret the posterior beliefs about a worker's skill as a résumé when those beliefs reect information from previous jobs. 6 1.3 Relation to the Empirical Literature This model speaks to the large empirical literature that tries to measure duration dependence using microeconomic data. The complete set of observational studies is too large to catalogue here, but the results are, in the words of Ljungqvist and Sargent [1998], mixed and controversial. See Machin and Manning [1999] and Van den Berg [2001] for reviews of the labor and econometrics literatures. One reason for this discord is that true duration dependence cannot be separated from dynamic selection eects in observational data without imposing identifying assumptions, which are usually driven more by statistical practicality than economic theory. Nevertheless, there exists good evidence for negative duration dependence. Recently, authors have used eld experiments to circumvent the identication issues inherent in using 4 To be fair, much of this evidence came out after Gonzalez and Shi [2010] was published. Also, because of dynamic selection eects, Gonzalez and Shi's model can generate a negative correlation between job-nding rates and unemployment durations in the aggregate, even though true duration dependence is positive. 5 Seminal papers along these lines include Altonji and Pierret [2001] and Gibbons and Katz [1991]. 6 Morchio [2015], who cites an earlier version of my paper, adopts the rhetorical device of referring skill level as her résumé. However, his model only has learning on the job, not during unemployment. 7 to a worker's expected observational data, and the results are compelling. Oberholzer-Gee [2008], Kroft et al. [2013], and Eriksson and Rooth [2014] sent out fake applications to real help-wanted ads, and the applications with the longest gap since the last job were the least likely to elicit responses from rms. Moreover, some of these authors' results point to information stigma as one of the economic mechanisms for duration dependence. In addition to conducting a eld experiment, Oberholzer-Gee [2008] conducted a survey of managers in charge of hiring decisions at Swiss rms. For managers who prefer not to hire the long-term unemployed, the most important reason is the belief that a productive worker would have already found a job; skill decay is cited as the second most important reason. Exploiting the variation in labor-market conditions across localities, Kroft et al. [2013] nd that duration dependence is more strongly negative in places that have low unemployment. 7 Those authors interpret their results through the lens of the information-stigma hypothesis: Failing to nd a job in a tight market sends a stronger signal about an individual's inability to get hired. There is also direct evidence to support the search protocol in my model: Unemployment duration is one hiring criterion that rms mention in help-wanted ads, and workers systematically apply to dierent openings as their spells become longer. The New York Times, The Wall Street Journal, and other media outlets have documented that some rms advertise vacancies while stating an explicit preference for applicants who are 8 These stories conjecture that businesses expect the long-term unemployed currently or recently employed. to be of a lower average quality. However, not all rms that discriminate on the basis of unemployment duration are exclusionary. Although they draw less attention from the popular press, there are also helpwanted ads that specically encourage applications from the unemployed and underemployed. 9 By welcoming applicants who have had little success nding jobs elsewhere, some rms are likely trying to nd people who can be recruited easily or paid cheaply. I am unaware of any systematic evidence of which hiring practice is more common, but it's clear that dierent rms target dierent types of workers, with dierent histories of joblessness. Of course, current unemployment status is not the only thing that employers consider; most job postings also specify desired levels of education and experience. Analogously, a rm in the theoretical model solicits applications from people with a specic résumé, which is a composite of all past experiences of both employment and unemployment. Furthermore, Kudlyak et al. [2013] provide evidence that workers apply for lower-skill jobs as their unemployment spells grow longer. Their study exploits longitudinal data from a job-posting website to see what kinds of people apply to what kinds of jobs. They nd that after several weeks of search, well-educated workers eventually apply to the same jobs to which less-educated workers apply at the outset. As those authors point out, their results are consistent with the hypothesis that agents 7 Using U.S. data, Imbens and Lynch [2006] also nd that hazard curves are steeper in tight labor markets, although Biewen and Stees [2010] nd no such correlation in German data. Both of these studies use observational data, so they are subject to the same qualications mentioned above. 8 See, e.g., Rampell [2011] and Banjo [2012]. 9 Searching for the words unemployed or unemployment on a job-search website will turn up many such postings. 8 use unemployment duration to learn about worker quality. 10 I will proceed as follows. Section 2 describes the model environment. Section 3 denes an equilibrium and discusses aggregation. In Section 4, I calibrate the model and choose parameters to match the aggregate data on job-nding rates and search durations. Section 5 contains the results, including decompositions of hazards and wages to quantify the importance of information stigma. Sensitivity analysis is in Section 6. Empirical implications are discussed in Section 7. Figures are in Appendix A, computations are discussed in Appendix B, proofs are in Appendix C, and data details are in Appendix D. 2 Model Environment 2.1 Preferences and Demographics There is a unit mass of risk-neutral workers. During periods of unemployment, workers get leisure value A worker also has a constant probability µ of dying from one period to the next. Each period, a measure of workers die and are replaced by a measure the eective discount factor is β≡ µ of new workers. Agents discount the future at rate ρ; λ. µ hence, 1−µ 1+ρ . Firms seek to maximize the present discounted value of prots, and the measure of rms is determined by free entry. 2.2 Skills and Productivity There are two types of workers, those with high skills and those with low skills. Being skilled does not necessarily confer a higher level of productivity, but skills do put workers in a more advantageous position to make high-quality matches and master jobs more quickly. A worker's output is given by aggregate productivity, and x is idiosyncratic. Match-specic productivity process that depends on the worker's type. In particular, states of the world, the process governing workers. x x has support x zx, where z evolves according to a Markov X ≡ {0, x1 , . . . , xnx }, and in all for high-skill workers stochastically dominates that for low-skill Upon encountering a rm, a high-skill worker realizes a value of match-specic productivity with probability Ω̄ (x | 1); for low-skill workers, this probability is probability of a high-skill worker transitioning from probability is given by 10 Kudlyak Ω (x, x0 | 0). is x to x0 Ω̄ (x | 0). is given by x For the duration of a job, the Ω (x, x0 | 1); for low-skill workers, this The distinction between high- and low-skill workers is captured by the et al. [2013] cite Gonzalez and Shi [2010] as one theory consistent with their results; my model is also consistent with these ndings. 9 stochastic-dominance assumptions: X Ω̄ (x | 1) X < x≤X X Ω̄ (x | 0) . Ω (x, x0 | 1) X < x0 ≤X Ω (x, x0 | 0) . (2.2) x = 0, which is an absorbing state. The x0 ≤X I will assume that there is always positive probability of realizing fact that 0 < Ω̄ (0 | 1) < Ω̄ (0 | 0) (2.1) x≤X means that no one can perform all possible jobs, so even highly skilled workers may fail to be hired. Similarly, the fact that 0 < Ω (x, 0 | 1) < Ω (x, 0 | 0) means that everyone has some risk of entering unemployment. It will be convenient to dene for r ∈ [0, 1]: Ω̄ (x | r) ≡ Ω (x, x0 | r) ≡ rΩ̄ (x | 1) + (1 − r) Ω̄ (x | 0) . (2.3) rΩ (x, x0 | 1) + (1 − r) Ω (x, x0 | 0) . (2.4) Notice that each of the above quantities is a conditional probability. For instance, suppose that a rm encounters an unemployed worker who has probability r of being highly skilled; then, Ω̄ (x | r) is the probability of realizing an initial value of match-specic productivity equal to of transitioning from x to x0 for an employee who has probability x. r Likewise, Ω (x, x0 | r) is the probability of being highly skilled. 2.3 Information and Filtering In reality, rms cannot perfectly predict how fruitful an employment relationship will be; instead, prospective employers draw inferences based on a candidate's history. Likewise, in the model, a worker's skill level will not be directly observable, but rms will engage in a ltering problem to discern her ability using Bayes's rule. To this end, I will adopt the following denitions. Denition. For each worker, public information is the complete history of when the worker was unem- ployed, when the worker was employed, and all realizations of x in jobs where the worker was employed. Public information is available to workers and all prospective employers. Although information is incomplete, it is symmetric: Workers are also learning about their skills through the progression of their careers. Notice that public information includes realizations of realizations of 11 It's x x in jobs where a worker has been employed that result in an unemployed worker's application being rejected. not 11 Also, if a worker holds fair to ask whether an employee's on-the-job productivity is publicly observable. The evidence from empirical micro studies is somewhat mixed: Schönberg [2007] and Kim and Usui [2013] suggest that this assumption is realistic, and Kahn [2013] suggests that it is not. Nevertheless, the employer-learning literature commonly assumes that someone's idiosyncratic 10 a job with productivity x, and a transition to x0 induces her to enter unemployment, then the realization x0 is included in public information. That is, a prospective employer can see why an applicant left her last job. For each newborn worker, there is a prior probability of that worker's being highly skilled. This probability is drawn from an exogenous distribution born with prior probability Denition. A worker's r F (·). I will maintain the assumption that, amongst the workers of being highly skilled, a fraction r is, in fact, highly skilled. résumé, denoted r ∈ [0, 1], is the posterior probability of that worker's being highly skilled, conditional on public information. A worker's initial résumé is just the prior probability of her being highly skilled. This value corresponds to 12 someone's qualications upon entering the labor market, based on features such as academic performance. Each subsequent period, a worker's résumé gets updated, depending on her labor-market experience. From one period to the next, there are three things that can happen to a worker: 1. An unemployed worker can remain unemployed. 2. An unemployed worker can be hired by a rm with match-specic productivity 3. An employed worker's productivity can transition from x to x0 , x. possibly resulting in separation. Each of these transitions sends a dierent signal. Suppose a worker is unemployed with résumé worker is hired with productivity will be denoted résumé B u (r). x, then her posterior probability of being highly skilled B h (r | x); x to x0 , If the i.e., her updated if the worker remains unemployed, then this value will be denoted Now, suppose a worker is employed with match-specic productivity productivity transitions from r. then her résumé is updated to B e (r | x, x0 ).13 B u (r) ≡ P [highly skilled | unemployed → B h (r | x) ≡ P [highly skilled | unemployed → x, B e (r | x, x0 ) ≡ P [highly skilled | x → x0 , résumé x and résumé r. Concisely: unemployed, résumé résumé r] r] . If the worker's r] (2.5) (2.6) (2.7) The updating of beliefs will depend on the endogenous probability of making an employment transition. Consequently, I will defer presenting the exact expressions for these functions until later, in Section 2.9. productivity is observable to the entire marketplace, so current and prospective employers are using the same noisy signals to form inferences about a worker's skill. For example, Altonji and Pierret [2001], Lange [2007], and Kahn and Lange [2014] take this approach. 12 If employers make inferences about worker quality on the basis of time-invariant characteristics, such as race or sex, these features would also be subsumed into the prior. 13 Recall that a transition from updated according to B e (r | x, x0 ) x to x0 may or may not result in the continuation of a match. because the transition from x 11 to x0 is assumed to be observable. Regardless, beliefs will be When appropriate, I will denote the inverses of these functions by lowercase letters; e.g., the function bh (r | x) satises r = B h bh (r | x) | x 14 . 2.4 Matching The economy features a continuum of labor markets, indexed by r ∈ [0, 1]. There is one search market for each résumé value. When a rm posts a vacancy, it solicits applications from workers with a specic résumé r; in turn, workers with résumé r respond to the rms seeking to hire employees like them. With some probability, a prospective rm sees a worker's application; below, I will detail the matching technology for determining this probability. Firms are assumed to have some screening ability, which allows them to discern whether a particular worker will be a good match. Specically, the prospective employer can observe the worker's initial value of x. If the employer decides the match is worthwhile, they contact the worker, and they begin production in the following period. The timing of the matching process ensures that information between employers and employees will be symmetric: realization of bad x x A worker will only be contacted by a rm if her makes for a viable match, so a worker who remains unemployed never nds out if she had a draw, or if her application was simply never seen. This search protocol has both conceptual and technical justications. Segmenting the search market along the dimension of expected skill is realistic: Help-wanted ads typically specify that a job is intended for people with a certain level of experience, set of credentials, or history of unemployment, as I have already documented. The setup also implies that workers with dierent résumés do not congest one another's search. In a number of industries, rms have separate recruitment processes for junior and senior positions, so this assumption provides a reasonable approximation of the hiring process. From a model-making perspective, I am assuming that search markets segmented. 15 In addition to delivering tractability, the search protocol in the present model is clearly a normal-form Nash equilibrium: Workers with résumé of searching in market r, and rms posting vacancies in market applicants who have résumé r. r r play the strategy play a strategy of considering only those Given the recruitment strategies of rms, there's no reason for workers to deviate from their search strategies, and given the search strategies of workers, there's no reason for rms to deviate from their recruitment strategies. In each market r ∈ [0, 1], the number of meetings is determined by a constant-returns-to-scale function of vacancies and searchers. Denote by workers, in market r. θ (r) the market tightness, or the ratio of vacancies to unemployed If a worker is looking for a job in a market with tightness θ, then her application will 14 It will become apparent in Section 2.9 that B h (· | x) and B e (· | x, x0 ) are guaranteed to be invertible in their rst arguments. In the numerical exercises, 15 Elsewhere, B u (·) is invertible as well. in the directed-search literature, authors focus on the endogenous segmentation of markets. Examples include Gonzalez and Shi [2010] and Menzio and Shi [2010, 2011]. 12 min ζθ1− , 1 . be seen by a rm with probability If a worker is searching in market r, then the probability of her being seen by a rm is: n o 1− p (r) ≡ min ζθ (r) ,1 . The probability of a rm encountering a worker in market r (2.8) is: n o − q (r) ≡ min ζθ (r) , 1 . (2.9) 2.5 The Firm's Problem As will become clear in Section 2.7, the surplus of a match will depend not only on the worker's résumé, but also on the résumé the worker would have if she were not employed. This counterfactual résumé, and therefore wages, will depend on whether the worker was employed in the previous period. Let denote a worker's tenure status: s ∈ {e, h} s = h signies that a worker is newly hired, and s = e signies that a worker is in a continuing employment relationship. Let Gs (r, x) be the value associated with owning a rm matched with a worker in state (r, x, s). Let ws (r, x) be the wage paid to a worker in state (r, x, s); I will detail the wage determination process in Section 2.7. Firm owners have an outside option of zero. An active rm's Bellman equation is: Gs (r, x) = zx − ws (r, x) + β s.t.: Let X Ω (x, x0 | r) max {Ge (r0 , x0 ) , 0} r0 = B e (r | x, x ) . V (r) be the value associated with posting a vacancy in market r. cost κ (2.10) x0 0 For a potential rm, there is a constant to maintaining a vacancy in any market in a given period. With probability q (r) the rm encounters a worker. The potential rm's Bellman equation is: ! V (r) = −κ + β [1 − q (r)] V (r) + q (r) X Ω̄ (x | r) max {Gh (r0 , x) , V (r)} (2.11) x s.t.: r0 = B h (r | x) . Free entry requires that: V (r) = 0, ∀r. 13 (2.12) Hence, the value of posting a vacancy in any market must satisfy: κ = βq (r) X Ω̄ (x | r) max Gh B h (r | x) , x , 0 . (2.13) x 2.6 The Worker's Problem Let Hs (r, x) be the value associated with being an employed worker in state associated with being an unemployed worker with résumé Hs (r, x) = ws (r, x) + β X r. (r, x, s); let U (r) be the value The employed worker's Bellman equation is: Ω (x, x0 | r) max {He (r0 , x0 ) , U (r0 )} (2.14) x0 s.t.: r0 = B e (r | x, x0 ) . The unemployed worker's Bellman equation is: ! U (r) = λ+β s.t.: [1 − p (r)] U rh0 = B h (r | x) (ru0 ) + p (r) and X Ω̄ (x | r) max {Hh (rh0 , x) , U (ru0 )} (2.15) x u ru0 = B (r) . 2.7 Wage Determination I will assume that wages are determined each period by a linear surplus-splitting rule, with a fraction η of the total match surplus going to the worker. This value will depend on the counterfactual résumé that the worker would have if a new match were not formed or if an extant match were not maintained. Consequently, there are separate surplus-spitting conditions for new hires (s = h) First, consider an unemployed worker searching in market updated to updated to market r B u (r), and she receives value B h (r | x), gets value U (B u (r)). and she receives value Gh B h (r | x) , x ; Hh r. and continuing employees If she is not hired, then her résumé is If she is hired with productivity B h (r | x) , x . (s = e). x, then her résumé is A rm that successfully hires a worker in free entry assures that rm owners get nothing if they fail to make a hire. Hence, the surplus-splitting condition for newly hired workers is: Hh B h (r | x) , x − U (B u (r)) = η Gh B h (r | x) , x + Hh B h (r | x) , x − U (B u (r)) . (2.16) Now, consider a worker in an existing employment relationship who, in the previous period, produced x− with résumé r− . If productivity transitions to Because the complete history of x x, then the worker's résumé becomes r = B e (r− | x− , x). is assumed to be observable in any consummated match, the worker's 14 résumé is unaected by staying on the job for an additional period. Hence, the surplus-splitting condition for continuing employees is: He (r, x) − U (r) = η [Ge (r, x) + He (r, x) − U (r)] . More succinctly, we can evaluate (2.16) at résumé ηGs (r, x) = rc ≡ bh (r | x) and consolidate it with equation (2.17) to obtain: (1 − η) [Hs (r, x) − U (rc )] B u bh (r | x) if s = h r In the above, rc (2.17) if (2.18) s = e. represents the counterfactual résumé that the worker would have if unmatched. When interpreting this wage-determination process, some caveats are in order. In many models, linear surplus-splitting conditions can be supported in two dierent ways: (1) as the axiomatic solution to a cooperative game, or (2) as the subgame-perfect equilibrium of a non-cooperative game. associated with Nash [1950]; the latter, with Rubinstein [1982]. directly assert a surplus-splitting rule as a primitive assumption. The former is Some authors, such as Pissarides [1994], 16 In the present model, any of the standard justications are valid for the surplus-splitting condition in continuing matches (2.17). For new matches, however, Rubinstein's alternating-oers protocol does not apply: If a worker were contacted by a rm, but (inexplicably) declined to accept any oer, then she would continue searching with the knowledge that she was worth hiring. In that case, the worker would have a private signal of her own skill, relative to her next employer. Non-cooperative bargaining under asymmetric information is an interesting problem, but it's a 17 non-trivial one that falls outside the scope of the current endeavor. Alternatively, the Nash [1950] solution species a fair bargain (p. 158), dened as an allocation that satises a collection of axioms. This approach takes as parameters the feasible payos when agents cooperate, along with utility oors that agents obtain without cooperating. unemployed worker's payo from not forming a match is U (B u (r)), Equation (2.16) species that an the value of continuing unemployment with public information. This modeling choice sidesteps the complications that would arise with information asymmetries. As a plausible defense, one can imagine that workers and rms send their applications and openings to employment agencies. In the event of a compatible match, the agency applies the Nash axioms 16 In the environment developed by Pissarides [1994], the space of payos is non-convex, which creates complications for either Nash [1950] or Rubinstein [1982] bargaining. See Shimer [2006] for a discussion. 17 Firms have no reason to contact workers unless they expect to reach a bargain. Nevertheless, solving the non-cooperative bargaining game requires computing the payos associated with deviations from equilibrium outcomes, including deviations that lead to information asymmetries. 15 to determine a fair bargain and connects the trading parties. contacted with any contract worth more than U (B u (r)), An unemployed worker would want to be and the rm would want to be contacted with any contract worth more than zero. These conditions will coincide if wages are determined according to equation (2.18). 2.8 Policies Agents decide which matches are worth forming and maintaining. In light of the surplus-splitting rules, a match is worthwhile for a worker if, and only if, the match is worthwhile for the rm. The set of match-specic productivities that will result in a worker with résumé r being hired is: X h (r) = x ∈ X | Gh (rh0 , x) + Hh (rh0 , x) ≥ U (ru0 ) , rh0 = B h (r | x) , ru0 = B u (r) . (2.19) In subsequent periods, agents decide whether to continue or allow the employment relationship to dissolve. The set of match-specic productivities that will perpetuate a match with a worker who has résumé current productivity x r and is: X e (r | x) = {x0 ∈ X | Ge (re0 , x0 ) + He (re0 , x0 ) ≥ U (re0 ) , re0 = B e (r | x, x0 )} . (2.20) 2.9 Beliefs Having dened the meeting rates and the policy functions, it is now possible to compute the belief-updating rules. Applying Bayes's rule yields: u B (r) = B h (r | x) = B e (r | x, x0 ) = 1 − p (r) P 1 − p (r) P x∈X h (r) x∈X h (r) Ω̄ (x | 1) Ω̄ (x | r) r (2.21) Ω̄ (x | 1) r Ω̄ (x | r) Ω (x, x0 | 1) r. Ω (x, x0 | r) (2.22) (2.23) Notice that the amount of information revealed by an unemployment spell is endogenous because depends on the equilibrium job-nding rate. All workers searching in market r, B u (r) regardless of their actual p (r) of being seen and screened by a rm. A genuinely high-skill worker P rate of p (r) x∈X h (r) Ω̄ (x | 1), whereas the job-nding rate for the average skill level, have the same probability with résumé r has a job-nding worker with résumé r is p (r) P x∈X h (r) Ω̄ (x | r). Provided that X h (r) consists of all values of x meeting or exceeding a reservation productivity, high-skill workers will have a greater probability of being hired than 16 low-skill workers with the same résumé because two implications. First, B u (r) < r, Ω̄ (· | 1) stochastically dominates Ω̄ (· | 0). This feature has so a worker's résumé will, in fact, deteriorate during unemployment. Second, the information-updating process will occur more rapidly for workers searching in tight markets: Holding X h (r) constant, B u (r) is decreasing in p (r). Although unemployment will always send a bad signal, the strength of that signal depends on aggregate conditions. Failing to nd a job in a slack market is more indicative of rms' tepid appetite for hiring, whereas failing to nd a job in a tight market is more indicative of a worker's ability to make a match. 3 Equilibrium 3.1 Denition We are now prepared to dene an equilibrium. Denition. A recursive search equilibrium comprises: 1. Value functions 2. Policies X h (r) 3. Contact rates Gs (r, x), V (r), Hs (r, x), and p (r) 6. Belief functions U (r) X e (r | x) and q (r) 4. A market-tightness function 5. A wage function and θ (r) ws (r, x) B u (r), B h (r | x), and B e (r | x, x0 ) subject to: 1. Bellman equations: (2.10), (2.11), (2.14), and (2.15) 2. Optimal match formation and dissolution: (2.19) and (2.20) 3. Matching technology: (2.8) and (2.9) 4. Free entry (2.12) 5. Surplus splitting (2.18) 6. Bayesian updating: (2.21), (2.22), and (2.23). Several comments are in order. I deliberately constructed the equilibrium so that it has a tractable solution. In particular, the assumption of segmented search markets ensures that policies, wages, and worker ows do not depend on the distribution of workers over states. 18 However, with the equilibrium objects in hand, the distributional dynamics become a matter of accounting, which I will lay out in Section 3.2. The model's structure makes it manageable to solve using the method described in Appendix B. That appendix 18 Shi [2009] refers to this property as block recursivity. 17 reduces the equilibrium conditions to a system of functional equations for the market-tightness function, the worker's value of unemployment, and the joint value of employment to the worker and the rm. The equilibrium objects have to be the xed point of a functional operator, so this line of argument suggests an iterative algorithm for solving the model numerically. Even though such a computational approach only establishes the existence of an equilibrium, I conjecture that the equilibrium is unique. In other models of endogenous information revelation, such as Lockwood 19 In those models, as in mine, the probability [1991] and Acemo§lu [1995], there can be multiple equilibria. of nding a job determines the informational content of a worker's joblessness and the skill composition of the unemployed workforce. The crucial dierence is, in other models, workers with all dierent expected skill levels apply to the same vacancies, so prospective employers care about the distribution of information and skills in the unemployed population. In such environments, self-fullling equilibria are possible: Firms are reluctant to hire, which makes workers spend more time in unemployment, which makes the pool of unemployed applicants look worse, which makes rms reluctant to hire. That mechanism for multiplicity is absent in my model because a rm can choose to seek a worker with a specic expected skill level. Consequently, rms' hiring decisions alter the distribution of unemployed workers across markets, but the skill composition of workers within each market remains unchanged. Moreover, there are two markets in this economy that have full information, corresponding to r=0 and r = 1. Both of these markets are essentially discrete-time versions of the environment created by Mortensen and Pissarides [1994], which is known to have a unique solution. It would be somewhat surprising if equilibrium outcomes were uniquely determined in markets r ∈ {0, 1}, but not in markets r ∈ (0, 1). 3.2 Aggregation As demonstrated above, we can solve for the model's main ingredients independently of the distribution of workers across states. However, to compute any kind of aggregate statistics, it's necessary to know this distribution, which is determined endogenously. Let in continuing employment relationships with productivity F h (r | x) r. dF (r) x be the measure of employed workers whose résumé is less than or equal to be the measure of newly hired workers with productivity F u (r) Let F e (r | x) x r, and a fraction r Let whose résumé is less than or equal to be the measure of unemployed workers whose résumé is less than or equal to is the measure of workers born with résumé r. of these dF (r) r. Recall that workers were actually born with high skills. Because the prior is assumed to be rational and agents engage in Bayesian updating, r will be the fraction of résumé-r workers who are indeed highly skilled. 19 Like me, Gonzalez and Shi [2010], Fernandez-Blanco and Preugschat [2014], and Jarosch and Pilossoph [2014] cannot provide armative proof of uniqueness, but they cannot nd instances of more than one equilibrium either. 18 I will dene a functional operator that characterizes the law of motion for these measures. We want an F e, F h, F u operator that maps dened in Section 3.1. Let Dene the mapping today into F e, F h, and Fu F e, F h, F u (1 − µ) = Υ2 F , F , F 0 0 (r , x ) Ω (x, x0 | r) d F e (r | x) + F h (r | x) p (r) Ω̄ (x0 | r) dF u (r) (1 − µ) = (3.1) Be (r 0 |x,x0 ) ˆ u by: Xˆ x h tomorrow, given the components of the equilibrium be the respective spaces in which these measures must reside. Υ : Fe × Fh × Fu → Fe × Fh × Fu Υ1 F e , F h , F u (r0 , x0 ) e 0 (3.2) Bh (r 0 |x0 ) Υ3 F e , F h , F u (r0 ) (1 − µ) = Xˆ Ω (x, x0 | r) d F e (r | x) + F h (r | x) Bd (r 0 |x,x0 ) x,x0 ˆ + (1 − µ) Bu (r 0 ) X 1 − p (r) Ω̄ (x | r) dF u (r) + µF (r0 ) , (3.3) x∈X h (r) where I have dened: B e (r0 | x, x0 ) ≡ [0, be (r0 | x, x0 )] ∩ {r | x0 ∈ X e (r | x)} B h (r0 | x) ≡ 0, bh (r0 | x) ∩ r | x0 ∈ X h (r) (3.4) B d (r0 | x, x0 ) ≡ [0, be (r0 | x, x0 )] ∩ {r | x ∈ / X e (r | x)} (3.6) B u (r0 ) ≡ F e, F h, F u The measures {r | B u (r) ≤ r0 } . To understand the logic of (3.1), notice that r and productivity x. following period with productivity if x0 ∈ X e (r | x), integrating over x F e (r0 | x0 ), than or equal to 0 = Υ F e, F h, F u . (3.8) dF e (r | x) + dF h (r | x) is the measure of employed workers Of these workers, a fraction x0 and résumé B e (r | x, x0 ). (1 − µ) Ω (x, x0 | r) r0 . with respect to F e (r | x) + F h (r | x) r0 over if r ∈ [0, be (r0 | x, x0 )]. r ∈ B e (r0 | x, x0 ) the measure of workers in continuing jobs with productivity Concisely, 0 (F e ) = Υ1 F e , F h , F u will survive to the These workers will continue in their jobs and their updated résumés will be weakly less than (1 − µ) Ω (x, x0 | r) yields (3.7) then evolve according to: F e, F h, F u with résumé (3.5) x0 Thus, and summing and résumé less . Similar reasoning accounts for the components of (3.2) and (3.3). Aggregate employment is given by: e≡ X F e (1 | x) + F h (1 | x) . x 19 (3.9) Likewise, aggregate output is given by: y≡z X x F e (1 | x) + F h (1 | x) . (3.10) x Consequently, the evolution of unemployment and output will always depend on the full distribution of workers across states. 4 Calibration One time period in the model corresponds to one week. Some the model's parameters are common in search models, so I will take values used elsewhere in the literature. The rate of time preference is ρ= .04 52 , corresponding to a real interest rate of approximately four percent. The probability of an agent dying from one period to the next is 1 40×52 , implying that the average working life lasts forty years. Workers and rms have equal surplus shares, with leisure is λ = 1 2. η = 1 2 . Aggregate productivity is normalized as z = 1. The ow value of As will be clear in a moment, the consumption of unemployed workers will equal half the output of a newly hired worker; this value is in line with other papers in the literature (c.f. Shimer [2005] and Hall [2005]). Following much of the literature, I will set the elasticity of the matching function to = 1 2 . The matching function's scaling factor ζ will be chosen to match the steady-state unemployment rate using a procedure I will describe below. I will follow Hall and Milgrom [2008], who set the per-period cost of maintaining a vacancy to 43% of one period's productivity. As will become clear, a newly hired worker produces one unit of output, so I will set κ = .43. The details of the productivity processes are as follows. Recall that the support of I will assume that x is X = {0, x1 , . . . , xnx }. {log (x1 ) , . . . , log (xnx )} constitute a 21-point, equally spaced grid from -1 to 1. ers in continuing employment relationships, I will assume that log (x) For work- follows a discrete approximation to a random walk with drift, the parameters of which depend on the worker's skill level. There is a constant, type-specic probability of productivity dropping to zero. Conditional on remaining positive, x has a con- stant, type-specic probability of taking a step up; there is also a constant, type-specic probability of taking a step down. That is, for x 1 < k < nx : Ω (xk , x0 | i) = (1 − δi ) αi (1 − δi ) (1 − αi − χi ) (1 − δi ) χi δi 20 if x0 = xk+1 if x0 = xk (4.1) 0 if x = xk−1 if x0 = 0. At the endpoints of the positive support (i.e. x = x1 and x = x nx ) productivity can only move in one direction, so I will specify: Ω (x1 , x0 | i) = (1 − δi ) αi (1 − δi ) (1 − αi ) δi 0 if x = x2 if x0 = x1 if x0 = 0 Ω (xnx , x0 | i) = (1 − δi ) χi δi if x0 = xnx if x0 = xnx −1 if x0 = 0. (4.2) The stochastic dominance of Ω (x, x0 | 1) α0 , χ0 > χ1 , In other words, high-skill workers have a higher probability of experiencing and δ0 > δ 1 . over Ω (x, x0 | 0) (1 − δi ) (1 − χi ) is captured by the parameter restrictions α1 > productivity gains and a lower probability of experiencing productivity losses. A diculty in choosing specic values for {αi , χi , δi }i=0,1 is that these parameters quantify the dierence between types, and we're interested in a situation where a worker's type is not directly observable. In models with only one type of worker, or with worker types corresponding to observable categories in the data, authors typically use parameters that are considered reasonable in light of existing microeconomic estimates. My calibration strategy will be to choose two sets of parameters within the range of reason, but with one on the high side, and the other on the low side. In Section 6, I will conduct sensitivity analysis to assess the impact of various parameter choices. Conditional on x remaining positive (and not being at an endpoint of X ), the mean and variance of log productivity growth are: 0 x | x1 < x < xnx , 0 < x0 , E log x 0 x V log | x1 < x < xnx , 0 < x0 , x where dx ≡ log (xnx /x1 ) / (nx − 1) i i = (αi − χi ) dx . (4.3) = h i 2 (αi + χi ) − (αi − χi ) d2x , (4.4) is the increment size of the log productivity grid. For each productivity process i, the above two moments will guide my choice of the two parameters αi and χi . In the model, wages and productivity are not identical, yet the quantitative results will show that productivity is highly correlated with wages for workers who are in continuing employment relationships. Consequently, my calibration for {αi , χi }i=0,1 will be motivated by existing estimates for wage and earnings processes. I will set the conditional 0 x .20 .10 mean of log to x 52×10 for high-skill workers and 52×10 for low-skill workers. So, over ten years, matchspecic human capital is expected to grow by about 20% for a high-skill worker and about 10% for a low-skill worker. The rationale for these numbers comes from the empirical literature on the return to tenure, which labor economists typically interpret as the growth rate of match-specic human capital. The calibrated values fall within the range suggested by existing studies: Altonji and Shakotko [1987] nd a cumulative 21 20 For both ten-year growth rate of 6.6%, whereas Topel [1991] nds a ten-year growth rate in excess of 25%. skill types, I will set the variance of log 0 x x .0313 52 . The justication for this choice comes from Meghir to and Pistaferri [2004], who estimate the variance of permanent innovations to log earnings; they obtain an unconditional variance of δ0 = 1 52×2.25 and δ1 = .0313 when using annual data pooled across dierent education groups. I will set 1 52×2.75 . These values imply that the average low-skill worker lasts 2.25 years before 21 As a point suering a terminal productivity shock, whereas the average high-skill worker lasts 2.75 years. of reference, Shimer [2005] calibrates separation rates so that the average job lasts 2.5 years. For newly hired workers, I will assume that the initial value of match-specic productivity has the following distribution: Ω̄ (x | 1) to stochastically dominate is an absorbing state, we can interpret Ω̄ (x | 0), ωi if 1 − ωi if i Ω̄ (x | i) = For ω x=1 (4.5) x = 0. it's necessary and sucient that ω1 > ω0 . Because x=0 as the probability of a worker being able to do a job, conditional on her true type. An advantage of this parsimonious specication is that the dierence in job-nding rates between high- and low-skill workers can be summarized by the dierence between skill workers with résumé jobs with probability r nd jobs with probability ω0 p (r). The parameters ω0 and ω1 p (r), ω1 ω1 and ω0 ; that is, high- whereas low skill workers with résumé r nd will be important for determining the informational content of unemployment spells because the belief-updating function is: B u (r) = where I have dened ωr ≡ (1 − r) ω0 + rω1 . 1 − ω1 p (r) r, 1 − ωr p (r) (4.6) For a given meeting rate p (r), the dierence between governs how quickly a worker's résumé is updated while she is searching for a new job: in ω1 − ω0 , and B u (r) = r when F (r). and ω0 is decreasing ω1 = ω0 . It remains to to pick numerical values for for newly born agents B u (r) ω1 ω1 , ω0 , the matching eciency ζ, and the résumé distribution To accomplish this, I will select parameters to match the primary moments of interest, namely, the steady-state unemployment rate and the population job-nding rate as a function of duration. To economize on free parameters, I will assume that the fraction of high-skill workers in the economy is F (r) is Beta(αF , 10 − αF ). αF 10 . I will also adopt the normalization that This normalization comes with minimal loss of generality: When 20 For a more recent review of the 21 The model features endogenous desirable. Strictly speaking, δi This implies that p (r) ω0 + ω1 = 1. is strictly less than one, doubling ζ is literature and evidence, see Altonji and Williams [2005]. separation because x can gradually drift downward to the point that a match is no longer is not a type-specic separation rate, but a lower bound on the separation rate. In the numerical exercises I perform, the majority of separations will come from x dropping to zero immediately, as opposed to declining gradually. 22 observationally equivalent to cutting both conduct a grid search over values of ω0 ω1 − ω0 and and ω1 αF in half. What matters is the dierence ω1 − ω0 . I will in order to minimize the quadratic deviation between the population job-nding hazard implied by the model and the empirical job-nding hazard constructed using CPS data. 22 I exclude workers who report being on layo as their reason for unemployment because Fujita and Moscarini [2013] present evidence that many workers on layo are recalled by the employers from whom they were initially separated. Only 13% of unemployed workers in my sample were on layo; excluding them slightly diminishes the drop in job-nding rates at low durations. consist of an equally-spaced grid of 51 points between zero and 1 2 ; the candidate values of equally-spaced grid of 10 points between 1.5 and 8.5. For each pair results in a 6% steady-state unemployment rate. ω1 = .60, αF = 4.0, and The candidate values of (ω1 − ω0 , αF ), αF ω1 − ω0 consist of an I nd a value for 23 This procedure points to parameter values of ζ that ω0 = .40, ζ = .2612. Figure A.2 shows the duration-specic job-nding probabilities implied by the model, along with the duration-specic job-nding probabilities from the CPS data. Overall, the t is good. It's worth noting that this procedure does not place any ex ante restrictions on the shapes of individuals' hazard curves. Recall from Section 1.2 that most existing theories imply declining hazards, but Gonzalez and Shi [2010] nd increasing hazards. Negative duration dependence is not hardwired into the present model. Likewise, this approach allows either true duration dependence or dynamic selection eects to drive the negative correlation between search durations and job-nding rates. The results in the following section demonstrate that all of these forces are at work. The large majority of job-seekers have declining hazards, but a minority has increasing hazards. True duration dependence and dynamic selection eects are both important, but much of heterogeneity in job-nding rates is a consequence of informational concerns. 5 Results The market tightness function θ (r) summarizes the demand for workers, given their expected skill level, and determines job-nding rates for the unemployed. of a worker's résumé. Figure A.3 shows market tightness as a function Strikingly, market tightness is not monotonically related to r. In the absence of informational concerns, one would expect the highest-skill workers to be in the highest demand. Indeed, the 22 More precisely, the criterion being minimized takes the form (m − m0 )0 W (m − m0 ). Here, m is a 52 × 1 vector, the dth d weeks in their rst CPS survey and who element of which is the fraction of workers reporting an unemployment duration of report being employed in their second CPS survey. When computing these fractions, workers are weighted using the CPS nal weights. The vector m0 is the probability of a worker with an unemployment duration of in the subsequent four weeks. The weighting matrix W is diagonal, the dth d weeks in the model nding work element of which is proportional to the number of unemployed workers in the CPS who report an unemployment duration of d weeks. Details on the data are available in Appendix D. 23 Numerically, it appears that the steady-state unemployment rate is monotonic in condition was accomplished using a bisection method. 23 ζ, so nding a value that satised this markets r=0 and r=1 have full information, and θ (1) is 17% greater than θ (0). However, the dierence in market tightness between the two markets with full information is dwarfed by the the range of tightnesses for markets with incomplete information. Market tightness is highest for workers with résumé .80, and this market is 5.4 times tighter than the market for workers who are highly skilled with certainty. Because B u (r) < r, a worker's résumé unambiguously declines during unemployment, but this does not imply that a worker's job-nding probability declines as well. If a worker begins her search in a market r close to one, then she will actually move into tighter markets as she spends time in unemployment. Eventually, though, her résumé will deteriorate until she is to the left of the peak in Figure A.3; subsequently, her job-nding rate drops with each additional period of joblessness. In Section 5.1, I will show that the non-monotonicity of the market-tightness function is related to the value of information: Workers of uncertain quality are willing to take pay cuts to improve their résumés, whereas workers who are known to have high skills need to be paid high wages. Recruiting workers with mid-level résumés may result in lower expected productivity, but higher expected prots. Then, in Section 5.2, I will explore the model's quantitative predictions for the job-nding hazard in more detail. 5.1 Wages and the Value of Information To understand the economic forces behind job creation, it is essential to understand how the value of information is priced into wages. Holding down a job and staying out of unemployment improve a worker's résumé. Therefore, the equilibrium wage should reect the amount that a rm pays for a worker's services, minus the amount that the worker pays for the résumé value of being employed. The following decomposition makes this point clear. In a recursive search equilibrium, wages are given by the following function of the worker's résumé and productivity: Proposition 1. ws (r, x) = ηzx + (1 − η) λ + ηκθ (rc ) − (1 − η) β X x0 rc ≡ B u bh (r | x) if s = h r if s = e re0 ≡ B e (r | x, x0 ) ru0 ≡ B u (rc ) . Proof. See Appendix C. 24 Ω (x, x0 | r) [U (re0 ) − U (ru0 )] (5.1) We see that wages depend on three terms. The rst is ηzx + (1 − η) λ, a convex combination of the worker's output while employed and her ow value of leisure while unemployed. This expression is equal to the wage that would arise in a static model of Nash bargaining. The second term is ηκθ (rc ), which I will call the search wedge. This expression represents the worker's option value of search; when markets are tight, the worker can leverage the relative ease with which she can nd a new job. Recall from Section 2.7 that rc is the counterfactual résumé a worker would have if she were not employed. In a textbook model of search and bargaining, such as Pissarides [2000], bargained wages are equal to the static Nash outcome, plus a search wedge. In this environment, though, the search wedge depends on θ (rc ), the tightness of the market in which the worker would be searching if she were not employed in her current match. nal term in equation (5.1) is − (1 − η) βE [U (re0 ) − U (ru0 ) | r, x], The where the expectation is taken over x0 . I will call this term the information wedge because it represents the information value of one additional period of employment to the worker. To make this interpretation clear, consider a worker with résumé r who has the opportunity to work today, but will enter unemployment tomorrow. If she is employed today, then her résumé is updated to re0 tomorrow, and her value of searching will be today, then she will have counterfactual résumé ru0 = B u (rc ). rc , If she is unemployed and tomorrow she will be be unemployed with résumé In that case, her value of searching will be U (ru0 ). By reaching an agreement for one period of employment, the worker improves her future search prospects by equals the expected discounted value of U (re0 ). U (re0 ) − U (ru0 ), U (re0 ) − U (ru0 ). The information wedge scaled by the rm's surplus share (1 − η). Hiring someone makes that person look good in the eyes of future employers, so a rm extracts some of this value from the worker. Figure A.4 shows the wages of newly hired workers, and Figure A.5 shows the decomposition of these wages. Even though all workers have the same output in their rst period on the job, wages can vary substantially, depending on a worker's résumé. Many people will receive negative wages, with the lowest pay going to workers who are probably highly skilled, yet who have some uncertainty surrounding their true types. The reason for this result is that these workers have the information wedges with the largest magnitudes. For workers whose type is known, i.e. those with résumés r ∈ {0, 1}, a job is just a source of consumption; there's no information value from employment. Notice that the information wedges for these workers is zero: Because the belief-updating functions all map zero to zero and one to one, for people with (r = 0) r ∈ {0, 1}. Unsurprisingly, high-skill workers (r = 1) U (re0 ) = U (ru0 ) get paid more than low-skill workers because they have better outside options, reected by larger search wedges. However, the dierence in wages between the two types with full information is negligible compared to the range of compensations received by newly hired workers with Figure A.7 shows rc = B u bh (r | x) r ∈ (0, 1). Figure A.6 shows U (r), the value of unemployment, and , the counterfactual résumé a newly hired worker would have if she 25 were not employed, as a function of the worker's current résumé. The workers with the largest (absolute) information wedges are those for whom being hired sends the greatest signal (r U (r) the signal is of the greatest value (the slope of − rc is large), and for whom is large). 24 Figures A.8 and A.9 show the wages and wedges of workers in continuing employment relationships. Wages are only slightly increasing in the worker's résumé, but earnings scale up almost linearly with productivity. If it were possible to regress w on x and r, one might conclude productivity is the main determinant of wages, and informational concerns play only a minor role. Decomposing wages according to equation (5.1) suggests otherwise. Figure A.9 demonstrates that the search and information wedges are reasonably large, but for continuing employees, these terms have comparable magnitudes and opposing signs. In other words, there are two strong eects that come close to canceling out: The workers who get the most information value from employment are also those who would have the best job-nding prospects if they were to leave their current employers. The information wedge is smaller for workers in continuing employment relationships, relative to new matches. All employees enjoy the information value of holding down a job, but new employees also benet from the signal that they were good enough to hire out of unemployment. The behavior of wages sheds some light on the shape of the market-tightness function, which is the key to determining the job-nding hazard. For workers in continuing matches, the search wedge is proportional to market tightness, and it is nearly the mirror image of the information wedge. The tightest markets are the therefore the ones that provide the greatest résumé value to workers, not the ones that result in the highest expected output. This fact explains why high-résumé workers have upward-sloping hazard curves: As these people spend more time in unemployment, they are eectively willing to take a larger pay cut in order to halt the résumé damage from joblessness. Workers with good, but not pristine, résumés are in the highest demand because these matches are likely to be fruitful, and rms can hire labor at a relative bargain. In a full-information environment, the only thing an employer can oer a worker is some fraction of the goods she is producing. When there is uncertainty about workers' skills, rms can use the value of information to compensate workers without giving up output. Some of the model's predictions for wages are quantitatively drastic, but the qualitative predictions are clear and logical. The most extreme outcome is that many workers literally pay for the opportunity to work in their rst week on the job, and in subsequent periods, wages come close to tracking productivity. On a technical level, this lopsided stream of earnings over the life of a job is a symptom of period-by-period surplus 24 Figure A.9 show only a single line for the information wedge, even though equation (5.1) suggests that the information wedge is a function of current-period productivity x. In fact, under the chosen specication for wedge is identical for all workers whose productivity is in the interior of X: productivity increase and a one-increment productivity decrease, so the distribution over relatively few workers with the maximal productivity level xnx , the information re0 does not depend on x. There are and their information wedges are quantitatively very similar. In equilibrium, there are no workers with the minimal positive productivity 26 Ω (x, x0 | i), Everyone has the same probability of a one-increment x1 because there is no match surplus. 25 splitting with linear utility, which implies that workers are not bothered by sudden swings in consumption. Nevertheless, real-world wage-tenure proles are often increasing and concave. Theorists have proposed a number of mechanisms to explain this fact, such as workers' ability to acquire skills or rms' attempts to retain workers in the face of outside oers. The present model provides an alternative explanation, based on the informational rents that rms extract from new employees. Workers want to show future employers that they were worth hiring in their current job. The signal value of being hired, for which workers are willing to take a pay cut, is greater than the signal value of renewing an existing match. Entry-level internships t comfortably into this theory: They often have short durations and low pay, yet workers accept them to bolster their résumés. Finally, for the purposes of explaining market tightness, what matters is the present value of expected prots over the full lifetime of the match, not the timing of prots within the match. When wages are determined period-by-period, the information value that rms extract from workers is concentrated in the rst period of the match. Even if payments were smoothed out, workers with uncertain skill levels would be willing to accept a contract with a lower expected present value of wages; the markets for such workers would still be tight relative to the markets for workers with known skill levels. 5.2 Unemployment Durations and Job-Finding Rates Individual Hazards Now, I will investigate the shapes of individuals' true job-nding hazards. Consider a worker who enters unemployment with résumé denotes the function B u (·) r0 . After t periods of search, she will have résumé composed with itself t t ωi p (B u ) (r0 ) . t p (B u ) (r0 ) . ωi t (B u ) (·) Thus, conditional on her This representation resembles a multiplicative proportional hazard model: The probability of nding a job equals a function of duration times a scaling factor where times. Consequently, her probability of being seen and screened by a rm, as a function of unemployment duration, is true skill level, her job-nding hazard is t (B u ) (r0 ), that depends on an individual worker's underlying type. t p (B u ) (r0 ) , 26 Two things determine how the worker's job-nding rate evolve over the course of an unemployment spell: (1) how quickly her résumé deteriorates and (2) how her probability of being seen by a rm changes as a function of her résumé. To show how information is revealed during unemployment, Figure A.10 displays initial initial résumés r0 . t (B u ) (r0 ) for several Everyone's résumé declines in unemployment, but the drop is slower for people close to zero or one. This pattern comes from two sources. First, workers with résumés close to 1 2 have the most prior uncertainty about their skill levels, so an additional observation carries more weight. 27 with r0 25 See Burdett and Coles [2003] and Stevens [2004] for illustrations of how the curvature of the utility function aects wage- tenure proles. 26 In Section 7, I'll discuss the theory's implications for 27 Conditional on having résumé r , a worker's skill level the econometric implementation of proportional hazard models. follows a Bernoulli distribution with parameter 27 r; the variance of this Second, the markets for upper-middle-résumé workers are the tightest, and failing to be hired in a tight market does more damage to a worker's résumé, as seen from equation (2.21). To show how information aects contact rates, the left panel of Figure A.11 shows function of résumé; the right panel shows market r. ωr p (r), p (r), the probability of being seen by a rm as a the average job-nding probability of workers searching in Both functions inherit the hump shape of the market-tightness function Figure A.12 displays t p (B u ) (r0 ) ωi ) is proportional (by a factor of for various initial résumés to t p (B u ) (r0 ) r0 . θ (r), shown in Figure A.3. An individual's job-nding probability , so this gure illustrates the nature of true duration dependence. Evidently, dierent workers can have hazard curves with totally dierent shapes. Microeconomic studies often try to establish whether true duration dependence is positive or negative, but in the present model, the worker's initial résumé inuences both the level and the slope of the hazard. Of newly unemployed workers, 13% will exhibit positive duration dependence, in the sense that their job-nding probability will increase if they spend a second period in unemployment. Nevertheless, because the overwhelming majority of workers exhibit negative duration dependence, the model is broadly consistent with the experimental studies cited in the Introduction. Amongst the workers with downward-sloping hazards, those with the higher résumés will experience a faster drop in their job-nding probabilities. The average worker entering unemployment has a résumé of .37, and such a worker will see her job-nding probability drop by 29% after 13 weeks, by 40% after 26 weeks, and by 46% after 52 weeks. Aggregate Hazards A long-standing problem in the literature is how to account for the negative correlation between job-nding rates and unemployment durations that appears in aggregate data: Quantitatively, how much of this correlation is due to true duration dependence, and how much to unobserved heterogeneity? In the calibrated economy, both mechanisms are important. As one measure of true duration dependence, Jarosch and Pilossoph [2014] propose looking at the average change in job-nding rates experienced by workers who have been unemployed for t 28 Figure A.13 shows that this metric of true duration dependence exhibits periods. a 20% drop o over the course of a year of search. By contrast, Jarosch and Pilossoph [2014] nd that true duration dependence in their model accounts for virtually no change in the job-nding rate. We can perform other experiments to isolate the eects heterogeneity and true duration dependence. distribution is 28 That (1 − r) r, which is maximized at is, for each duration t, r= 1 . 2 compute: ˆ 100 × where bu (·) p (r) dFtu (r) , p (bu )t (r) B u (·), (bu )t (·) is bu (·) composed with itself t times, and Ftu (·) is the résumé distribution amongst t consecutive periods. Hence, a worker who has a job-nding rate of ωi p (r) in her tth t u job-nding rate of ωi p (b ) (r) at the outset of her unemployment spell. is the inverse of workers who have been unemployed for period of search had a 28 Figure A.14 contrasts three curves. The rst is the average job-nding rate in the unemployed population as a function of duration. The second is the population job-nding rate that we would observe if workers experienced duration dependence, but the skill composition of workers remained xed. 29 The drop in the job-nding rate coming from true duration dependence is appreciable: Holding the composition of workers xed, the average job-nding rate drops by 20% after 13 weeks, by 34% after 26 weeks, and by 43% after 52 weeks. The third curve in Figure A.14 is the population job-nding rate that we would observe if workers had heterogeneous but time-invariant job-nding rates; the distribution over job-nding rates is taken to 30 Despite the importance of be the distribution implied by the model for workers entering unemployment. true duration dependence, the amount of heterogeneity in the model is also capable of generating a strong negative correlation between unemployment duration and the probability of nding a job. More importantly, the heterogeneity amongst job seekers is not purely mechanical; dierences in jobnding rates across workers are endogenously determined economic outcomes. As evidenced by Figure A.11, the dierence in job nding rates between workers with known skills (i.e. dierence in job-nding rates between workers with unknown skills (i.e. r ∈ {0, 1}) r ∈ (0, 1)). is small relative to the Incomplete information provides an economic mechanism for the relationship between search durations and job-nding rates, and this mechanism operates through the channels of both heterogeneity and true duration dependence. This point is further illustrated in the next section, which examines an economy with full information. 5.3 A Full-Information Benchmark To appreciate the role of incomplete information, it's instructive to compare the economy of Section 2 to a benchmark environment where workers' skill levels are known, but the expected distribution over labor productivity is unchanged. distributions Suppose that a worker born with résumé r Ω̄ (x | r) and Ω (x, x0 | r) for her entire lifetime, with certainty. draws her productivity from the Instead of representing a worker's expected skill, the résumé in this context represents the worker's actual, observable skill level: If then Ω̄ (x | r1 ) 29 This stochastically dominates Ω̄ (x | r0 ), and Ω (x, x0 | r1 ) stochastically dominates r1 > r0 , Ω (x, x0 | r0 ). is the average job-nding rate as a function of duration, where the average is taken over skills and initial résumés, with respect to the distribution of workers entering unemployment: ˆ Hazard Due to True Duration Dependence where F0u (·) 30 To = [rω1 + (1 − r) ω0 ] p (B u )t (r) dF0u (r) , is the distribution of résumés amongst workers entering unemployment. F0u (·) denote the distribution over résumés amongst workers entering unemployment. A measure rdF0u (r) ω1 p (r), and a measure (1 − r) dF0u (r) has job-nding rate ω0 p (r). If everyone had a constant probability of nding a job, then the fraction of workers with job-nding rate ωi p (r) remaining in unemployment t after t periods is [1 − ωi p (r)] . The average hazard curve, holding all individual job-nding rates constant, is therefore: ´ [1 − ω1 p (r)]t ω1 p (r) r + [1 − ω0 p (r)]t ω1 p (r) (1 − r) dF0u (r) . Hazard Due to Heterogeneity = ´ [1 − ω1 p (r)]t r + [1 − ω0 p (r)]t (1 − r) dF0u (r) be precise, let of newly unemployed workers has job-nding rate 29 As before, there is a continuum of segmented search markets, indexed by r ∈ [0, 1], vacancy chooses to solicit applications from workers with a particular résumé r. and a rm posting a Agents' Bellman equations are the same as before, except that the belief-updating functions are replaced by identity functions. Such an environment is like having a continuum of Mortensen-Pissarides economies, each with its own idiosyncratic productivity process, indexed by r. This full-information benchmark diers from the model of Section 2 in two important respects. First, jobnding rates are not subject to duration dependence. A worker born with résumé consequently, a worker who begins searching in market of her spell. r will continue searching in market Second, there is no information value from employment. wedge disappears because re0 = ru0 = r. r will have résumé r forever; r for the duration In equation (5.1), the information Employers can no longer compensate their workers with the résumé value of holding a job, so the rm's cost of labor goes up. As a result, market tightness declines. Figure A.15 compares the job-nding rate these functions align at incomplete. ωr p (r) r = 0 and in the full- and incomplete-information economies. By construction, r = 1, but for all r ∈ (0, 1), markets are tighter when information is Furthermore, when types are known, market tightness is monotonically related to skill level; in contrast, the market-tightness function has a pronounced hump shape when types are unknown. Besides causing genuine duration dependence, information frictions also widen the range of job-nding rates amongst the unemployed. Figure A.16 compares the average aggregate job-nding rates, as a function of duration, between the baseline model and the full-information benchmark. The full-information economy falls far short of replicating the negative relationship observed in the data. In part, this is because individual workers have constant hazards, but this result also comes from the compression of job-nding rates across workers with dierent expected skill levels. Although the results in Section 5.2 suggest that dynamic selection eects are important, heterogeneity should not be interpreted as an alternative explanation for information stigma when accounting for the aggregate relationship between job-nding rates and search durations. Instead, the model suggests that information stigma contributes to heterogeneity in job-nding rates. 5.4 Skill Decay and Alternative Sources of Duration Dependence Thus far, information stigma has been the only source of duration dependence that I have considered, but other explanations are worth exploring. Amongst the competing hypotheses, skill decay during unemployment is the leading alternative. The basic idea is that low-skill workers are in lower demand, so workers will have declining job-nding rates as they lose human capital during unemployed. Naïvely, one might think that information stigma and skill decay can both generate negative duration dependence, so a model with both forces would have a steeper hazard curve than a model with only one. 30 In the environment I have Table 1: Parameterizing Skill Change Parametrization Prob. of Skill Gain On the Job (γ) Prob. of Skill Loss O the Job Baseline 0 0 SC1 0 SC2 1 2×52 1 26 1 26 1 26 1 26 SC3 (τ ) constructed, however, skill loss during unemployment can actually attenuate duration dependence. If skills evolve over time, then a signal about a worker's current type becomes less valuable. Consequently, the value of creating a match may deteriorate less over the course of an unemployment spell, even though someone's résumé may deteriorate more. The model allows us to explore these forces in more detail. The framework of Section 2 can be extended to allow workers' skills to change, depending on their employment status. being unemployed in period t + 1. t, high-skill workers have probability Likewise, conditional on being employed in period high-skill workers in period t + 1. t, τ Suppose that, conditional on of becoming low-skill workers in period low-skill workers have probability γ of becoming The worker's actual skill level remains unobservable. Mathematically, the only things that change are the belief-updating functions, which become: Ω (x, x0 | 1) r Ω (x, x0 | r) Ω̄ (x | 1) r B h (r | x) = (1 − τ ) Ω̄ (x | r) P 1 − p (r) x∈X h (r) Ω̄ (x | 1) u P B (r) = (1 − τ ) r. 1 − p (r) x∈X h (r) Ω̄ (x | r) B e (r | x, x0 ) Table 1 shows the parametrizations for = γ γ + (1 − γ) and τ (5.2) (5.3) (5.4) that I will consider. All other parameters are held constant. As seen in Figure A.1, the average job-nding rate, as a function of duration, appears to level o after about six months. With this in mind, I set τ= 1 26 , implying that it takes an average time of six months for someone to experience a drop in skills. Parametrization SC1 species that skills can be lost during unemployment, but not cannot be regained on the job. Parametrizations SC2 and SC3 allow for skill upgrading during employment; on average, an improvement takes two years under SC2 and six months under SC3. As a point of reference, Ljungqvist and Sargent [1998] specify that the average rate of human-capital depreciation 31 during unemployment is twice as large as the rate of human-capital accumulation during employment. For each parametrization, Figure A.17 shows the average job-nding rate as a function of résumé. The probability of being hired is less sensitive to changes in a worker's résumé. Also, for most markets, job-nding 31 In Ljungqvist and Sargent's model, human capital is synonymous with productivity. In my model, skill is a distribution over productivities, so the parameters of their model cannot be compared directly to mine. 31 rates are lower in the economies with skill change; as seen from equation (5.4), résumés deteriorate more quickly during unemployment when job-nding rates are high. The drops in most workers' hazard curves are therefore more acute in the economy that has information stigma as the only source of duration dependence. On the aggregate level, the average job-nding rate declines more rapidly as a function of duration under the baseline calibration, relative to SC1-SC3. The results suggest that, when types are permanent and unobservable, dierences in workers' job-nding rates have more to do with dierences in how workers value information, as opposed to dierences in expected productivity. In particular, under the baseline specication, rms are most eager to hire the workers most willing to give up wages in order to improve their résumés. Adding stochastic skill change to the model diminishes the marginal value of information. Recall that the information wedge in workers' wages is proportional to the expected value of U (re0 ) − U (ru0 ), which captures how the résumé improvement from being employed today increases the value of search tomorrow. The slope of U (r) therefore provides an indication of how much a worker values the marginal résumé improvement she gets from holding a job. Figure A.18 shows U (r), the value of being unemployed with résumé r, under each parametrization. Notice that U (r) is steeper under the baseline calibration than under SC1-SC3. If workers can lose their skills in unemployment, there's less benet to being a high-skill worker; if workers can gain skills on the job, there's less harm in being a low-skill worker. In either case, the possibility of skill change causes U (r) to atten out, implying that information about a worker's quality has less eect on her search prospects. There are other possible explanations for duration dependence, besides information stigma and skill decay. It's possible that workers become discouraged and expend less eort looking for jobs. One could extend the environment of Section 2 to allow for variable search intensity, modeled as in Chapter 5 of Pissarides [2000]. In general, though, the optimal search intensity selected by a worker will be constant within an unemployment spell unless some other feature of the environment is time-varying. Consequently, this notion of search intensity can amplify, but not cause, duration dependence. Yet another possibility is that employment transitions are determined by stock-ow matching. Under this conception of the matching process, a newly unemployed worker surveys the stock of all existing vacancies to see if there is a suitable match. If such a job is available, the worker is hired; if not, the worker surveys the ow of new vacancies as they are posted by rms. In that case, duration dependence occurs by assumption, as a result of the matching technology. Also, this explanation only applies to the drop in job-nding rates between the rst and second periods of search. A promising approach may be to merge stock-ow matching with one of the mechanisms discussed above. Doing so, however, is not a straightforward extension, so I defer the problem to future research. 32 6 Sensitivity Analysis To assess the sensitivity of the results, I will experiment with some alternative parameter values. In the model, information about workers is revealed both on and o the job. Although the model's main application is analyzing stigma eects from unemployment, the quantitative results depend in part on the on-the-job productivity dierences between employed workers, captured by dierence between and ω1 , Ω (x, x0 | 0) and Ω (x, x0 | 1) Ω (x, x0 | 0) and Ω (x, x0 | 1). Although the plays a non-trivial role, the results are most sensitive to ω0 which determine the faction of jobs each type of worker is capable of doing and, by extension, the probability of being hired conditional on being seen by a rm. To illustrate this point, I solve the model with Ω (x, x0 | 0) = Ω (x, x0 | 1). δ0 = δ1 = I set the conditional mean of log 0 x x , given by equation (4.3), to .15 52×10 , and I set 1 52×2.50 . All other parameters remain the same as before. Under this alternative parametrization, high- and low-skill workers are equally productive on a job, conditional on being able to do that job at all. The market-tightness function (not shown) remains hump-shaped, though the maximal value of and occurs when r = .69. θ (r) is lower Consequently, a larger fraction of newly unemployed workers, 32%, experience positive duration dependence. A worker entering unemployment with the average résumé will see her jobnding probability drop by 15% after 13 weeks, by 23% after 26 weeks, and by 29% after 52 weeks. This decline is not as pronounced as the one described in Section 5.2, but it is nevertheless substantial, especially given that the two skill levels are assumed to be equally productive once a match has formed. The main parameters that control the degree of information stigma from unemployment are ω1 which are the respective probabilities of high- and low-skill workers being able to do a given job. If ω0 and ω0 , ω1 and were the same, then the probability of making a match, conditional on being seen by a rm, would not be correlated with on-the-job productivity; in that case, no learning about worker quality would take place during unemployment. Conversely, if ω1 − ω0 increases, the job-nding probability for high-skill workers goes up relative to low-skill workers with the same résumé, so an additional period of unemployment makes it seem even more likely that a worker is unskilled. To demonstrate the importance of these parameters, I re-solve the model with dierent values of nding rate ω1 − ω0 , keeping ωr p (r), as a function of a worker's résumé. As ω1 +ω0 xed. Figure A.19 plots the average job2 ω1 −ω0 between high- and low-skill workers (i.e. the dierence between importantly, though, the hump in ωr p (r) gets more pronounced. goes up, the dierence in job-nding rates ω1 p (1) and ω0 p (0)) becomes larger. More For each of these parameter congurations, Figure A.20 shows the average job-nding rate as a function of duration. The aggregate relationship between the duration of search and the probability of being hired becomes stronger as (ω0 , ω1 ) = (.40, .60) ω1 − ω0 increases. The values used in the baseline calibration are the ones that come closest to replicating the data. 33 7 Implications for Empirical Research 7.1 The Precision of Beliefs over Time Because résumés are updated every period, the variance of posterior beliefs about a worker's skill level will decline stochastically, and with a sucient amount of labor-market experience, a worker's true type will become known with certainty. 32 If duration dependence is solely a consequence of learning, then a worker can only experience true duration dependence if there is uncertainty about her skill. Thus, the model implies that workers who have been in the market the longest are subject to the least amount of duration dependence in unemployment. 33 This suggests comparing the relationship between unemployment durations and job-nding rates for workers with dierent amounts of labor-market experience. To summarize these correlations, consider the following semiparametric regression, applied to workers who are surveyed by the CPS in consecutive months and who report being unemployed in the rst month: job foundi = h (durationi ) + x0i β + i . In the above, durationi is the number of weeks that worker rst survey month; h (·) i (7.1) claims to have been looking for a job in the is a smooth function; job foundi is an indicator variable that is equal to one if the worker reports being employed in the second survey month; xi is a vector of covariates; and i is a residual. Data details are in Appendix D. I divide the sample by potential labor-market experience, dened as age minus years of education, and I t equation (7.1) separately for each quartile of the potential experience 34 distribution. Figure A.21 shows the estimated values of distribution. 100 × h (d) /h (0) for each quartile of the potential-experience Indeed, it appears that the drop in average job-nding rates, as a function of duration, is strongest for those in the bottom quartile of the potential experience distribution and weakest for those in the top quartile. 32 To see this, let ri,t 35 The decline is slightly larger for the second quartile than for the third, but taken denote the résumé of worker level. Beliefs are a martingale: Et [ri,t+1 ] = ri,t . i at time t. Then, (1 − ri,t ) ri,t is the posterior variance of a worker's skill Combining this fact with Jensen's inequality yields: Et [(1 − ri,t+1 ) ri,t+1 ] < (1 − Et [ri,t+1 ]) Et [ri,t+1 ] = (1 − ri,t ) ri,t , where the inequality is strict because convergence theorem implies 33 Under (1 − r) r is strictly concave in r, and ri,t+1 a.s. (1 − ri,t ) ri,t → 0. In other words, ri,t has positive variance. Thus, the supermartingale converges to zero or one. the baseline calibration, the model also predicts that experienced workers have less dispersion in their job-nding rates. However, the prediction that highly experienced workers have at hazards is a general consequence of Bayesian learning, not a byproduct of a particular choice of parameter values. 34 More specically, I t a local-polynomial regression, with additional linear covariates, following Robinson [1988]. How- ever, when performing the local-polynomial regressions, I employ the estimator described by Breidt and Opsomer [2000] for incorporating survey weights (in this case, the CPS nal weights). 35 As an aside, younger workers tend to have higher job-nding rates than older workers. This would just represent a vertical shift of each of the population hazards depicted in Figure A.21, which are normalized by the job-nding rates of newly unemployed workers. 34 together, those in the middle two quartiles show a steeper decline than the most-experienced workers, and a more moderate decline than the least-experienced workers. Figure A.22 shows the results when the sample excludes workers who are seeking part-time employment. In that case, the relationship between job-nding rates and durations is the basically same across the bottom three quartiles of the potential-experience distribution. However, it remains true that average job-nding rates fall less with duration for the most experienced workers. As a point of comparison, Figure A.23 shows the how the average job-nding rate changes with duration for dierent experience groups, using articial data generated by the theoretical 36 Qualitatively, the model-generated data shows that there correlation between duration and job- model. nding rates declines somewhat with experience, although the magnitudes dier from those constructed using the CPS. Of course, one would ideally want to distinguish between true duration dependence and dynamic selection eects, and the regression equation (7.1) can only capture correlations. Nevertheless, these results are at least suggestive of information stigma as a source of duration dependence. Field experiments, which send fake résumés to real vacancies, provide a possible avenue for measuring how duration dependence stems from the diusion of information about a worker's skill. When fabricating résumés, one would want to vary two things independently: the length of the current unemployment spell and the precision of beliefs at the outset of the unemployment spell. The theory of Bayesian learning suggests that prospective employers will have more precise beliefs about applicants who are perceived as having spent more time in the labor market. To vary the amount of information available on an articial résumé, one could adjust graduation dates and the number of previous jobs held. If applications containing more information about worker quality exhibited strong negative duration dependence, then that would suggest the importance of human capital decay, rather than stigma eects. 7.2 Specication and Identication of Proportional Hazards Models The theoretical results place restrictions on the proper use of the multiplicative proportional hazard model, which is the most popular tool in the econometric literature. probability of nding a job after disturbance and hT (t), xi t periods of search is That specication asserts that worker vi hX (xi ) hT (t), where vi i's is a mean-one idiosyncratic is a vector of person-specic predictors. In applications, the main object of interest is which is called the baseline hazard. In light of the economic theory, a proportional hazard model raises two concerns: specication and identication. For a typical cross-sectional dataset, the proportional hazard model is misspecied. More subtlety, if a prospective employer and an econometrician have the same information about workers, then the proportional hazard model is correctly specied 36 In but not identied. To the theoretical model, age is geometrically distributed, so the quartiles in Figure A.23 correspond to the quartiles of the potential experience distribution from the CPS. Also, the job-nding rates shown in Figure A.23 are computed directly from the model-generated data, rather than from estimating the semiparametric regression in equation (7.1). 35 obtain identication of a correctly specied model, it's necessary for an econometrician to have an indicator of workers' skills that falls outside the information set of rms. First, consider a random sample of unemployment spells from the model economy. Recall from Section 5.2 that an individual's job-nding hazard is t ωi p (B u ) (r0 ) , where r0 denotes the worker's résumé at the 37 There is heterogeneity amongst job seekers in both r0 , so idiosyncratic eects cannot be multiplicatively separated from the time-varying portion of the hazard. Graphically, it's clear outset of the spell. ωi and that a proportional hazard model will not be correctly specied for a sample of workers with dierent initial résumés: Figure A.12 shows the shapes of some individuals' true hazard curves, and they are obviously not in constant proportion to one another. In fact, this may be one of the reasons for the diversity of results obtained using proportional hazard models. Now, suppose that an econometrician could observe unemployment spells for a collection of workers who entered unemployment with identical résumés. vi hX (xi ) hT (t), where vi ≡ ωi /ωr0 , hX (xi ) ≡ 1, In that case, the job-nding rate can be written as and t hT (t) ≡ ωr0 p (B u ) (r0 ) . r0 , So, for a xed the true data-generating process is consistent with a multiplicative proportional hazard model. Unfortunately, the proportional hazard model is not identied with a dataset that has no variation in r0 . In the case where the data contain only one unemployment spell per worker, Elbers and Ridder [1982] prove that identiability of the proportional hazard model hinges on variance. in xi , hX (xi ) being a non-constant function and xi having non-zero 38 Even if an econometrician could observe aspects of a worker's publicly observed history to include they would be relevant only insofar as they determine r0 . In other words, Elbers and Ridder [1982] show that identiability of the proportional hazard model requires observable dierences in workers with the same baseline hazard, but the theory of information stigma suggests that the only workers who share the same baseline hazard are observationally identical, in the sense of having the same résumé. Identication requires the econometrician to have access to some indicator of workers' skills that is not included in the publicly observable history that is available to prospective employers. For example, a longitudinal dataset may allow an econometrician to see a worker's wage on her fth job, which would not be visible during her third unemployment spell; nevertheless, these wages will be correlated with the worker's skills. Given covariates xi that are outside rms' information set, one could specify hX (xi ) ≡ E [ωi | xi , r0 ]; it would also be necessary to assume that the expectational error in the econometrician's beliefs, given by vi ≡ ωi /hX (xi ), 37 If is independent of xi , conditional on r0 . there were more than two skill levels, the hazard would still assume this form, but r would be vector-valued. The following arguments will still apply. 38 Observing multiple spells per worker oers little help in this context. Honoré [1993] proves that variation in xi is unnecessary for identication when we can observe multiple spells per person, but this result assumes a person has the same vi and hT (t) across spells. Even if a person's skill level remains constant between spells, her résumé almost surely will not. Consequently, one person's multiple spells are not informative of a single baseline hazard function. 36 8 Conclusion Summing up, this model generates several novel theoretical results. First, there is a non-monotonic relationship between expected skill and job-nding rates. The total match surplus depends not only on the worker's productivity when employed, but also on the résumé damage the worker would have experienced when unemployed. Consequently, the tightest markets are generally not those for the most productive workers, but for those who most value the résumé improvement from being hired. Second, the model illustrates how information is priced into wages, which is essential to understanding the patterns in job-nding rates. The model provides a decomposition of wages that includes a compensating dierential for the value of information. Third, human capital decay, which is an alternative explanation for negative duration dependence, can actually attenuate the drop in job-nding rates. When skills change over time, information about a worker's current skill level is less valuable, thus blunting the above mechanism. Besides the theoretical results, the quantitative results from the calibrated model suggest that informational concerns can play a large role in shaping job-nding rates. The model can replicate the aggregate relationship between search durations and job-nding rates reasonably well. For individual workers, hazards can change substantially over an unemployment spell. Although heterogeneity in job-nding rates can generate a strong aggregate correlation between search durations and job-nding rates, incomplete information is responsible for much of the heterogeneity in job-nding rates, in addition to causing true duration dependence for individuals. The overwhelming majority of job-seekers experience negative duration dependence, which is qualitatively consistent with the micro evidence. In addition to the implications for empirical research discussed in Section 7, the model also suggests some directions for future theoretical research. There have been relatively few macroeconomic models of information stigma. And, as I have discussed, most of those models do not consider how the information value of employment inuences future unemployment experiences. Whereas I have embellished the structure of workers' skills and information, other authors have taken a closer look at the characteristics and behaviors of rms. In particular, Jarosch and Pilossoph [2014] argue that rm heterogeneity is important, because the rms that do not bother interviewing the long-term unemployed would not have been compatible with those workers anyway. An environment that incorporates the mechanisms in each of our models would provide a fruitful avenue for future research. Another extension would be an examination of business cycles. Although I have focused on the steady state, the environment I have developed can accommodate aggregate shocks. This aspect makes the model a fairly general framework for analyzing information dynamics in labor markets. 37 References Daron Acemo§lu. Public policy in a model of long-term unemployment. Economica, 62(246):16178, 1995. Joseph G. Altonji and Charles R. Pierret. Employer learning and statistical discrimination. The Quarterly Journal of Economics, 116(1):313350, 2001. Joseph G. Altonji and Robert A. Shakotko. Do wages rise with job seniority? The Review of Economic Studies, 54(3):437459, 1987. Joseph G. Altonji and Nicolas Williams. Do wages rise with job seniority? A reassessment. Industrial and Labor Relations Review, pages 370397, 2005. Shelly Banjo. Measures aim to end bias against long-term jobless. URL The Wall Street Journal, February 24, 2012. http://online.wsj.com/news/articles/SB10001424052970204778604577241693309654990. Martin Biewen and Susanne Stees. Unemployment persistence: Is there evidence for stigma eects? Eco- nomics Letters, 106(3):188190, 2010. F. Jay Breidt and Jean D. Opsomer. Local polynomial regression estimators in survey sampling. Annals of Statistics, pages 10261053, 2000. Ken Burdett and Melvyn Coles. Equilibrium wage-tenure contracts. Econometrica, 71(5):13771404, 2003. Chris Elbers and Geert Ridder. True and spurious duration dependence: The identiability of the proportional hazard model. The Review of Economic Studies, 49(3):403409, 1982. Stefan Eriksson and Dan-Olof Rooth. Do employers use unemployment as a sorting criterion when hiring? Evidence from a eld experiment. American Economic Review, 104(3):101439, 2014. Javier Fernandez-Blanco and Edgar Preugschat. On the eects of ranking by unemployment duration. 2014. Shigeru Fujita and Giuseppe Moscarini. Recall and unemployment. 2013. Robert Gibbons and Lawrence F. Katz. Layos and lemons. Journal of Labor Economics, 9(4):35180, October 1991. Francisco M. Gonzalez and Shouyong Shi. An equilibrium theory of learning, search, and wages. Economet- rica, 78(2):509537, 2010. Robert E. Hall. Employment uctuations with equilibrium wage stickiness. (1):5065, 2005. 38 American Economic Review, 95 Robert E. Hall and Paul R. Milgrom. The limited inuence of unemployment on the wage bargain. American Economic Review, 98(4):165374, 2008. Bo E. Honoré. Identication results for duration models with multiple spells. The Review of Economic Studies, 60(1):241246, 1993. Guido W. Imbens and Lisa M. Lynch. Re-employment probabilities over the business cycle. Portuguese Economic Journal, 5(2):111134, 2006. Gregor Jarosch and Laura Pilossoph. Statistical discrimination and duration dependence in the job nding rate. 2014. Lisa B. Kahn. Asymmetric information between employers. American Economic Journal: Applied Eco- nomics, 5(4):165205, 2013. Lisa B. Kahn and Fabian Lange. Employer learning, productivity and the earnings distribution: Evidence from performance measures. Review of Economic Studies, 2014. Seik Kim and Emiko Usui. Employer learning, job mobility, and wage dynamics. 2013. Kory Kroft, Fabian Lange, and Matthew J. Notowidigdo. Duration dependence and labor market conditions: Evidence from a eld experiment. The Quarterly Journal of Economics, 128(3):11231167, 2013. Marianna Kudlyak, Damba Lkhagvasuren, and Roman Sysuyev. Systematic job search: New evidence from individual job application data. 2013. Fabian Lange. The speed of employer learning. Lars Ljungqvist and Thomas J. Sargent. Journal of Labor Economics, 25(1):135, 2007. The European unemployment dilemma. Journal of Political Economy, 106(3):514550, 1998. Ben Lockwood. Information externalities in the labour market and the duration of unemployment. The Review of Economic Studies, 58(4):733753, 1991. Stephen Machin and Alan Manning. The causes and consequences of longterm unemployment in Europe. Handbook of Labor Economics, 3:30853139, 1999. Costas Meghir and Luigi Pistaferri. Income variance dynamics and heterogeneity. Econometrica, 72(1):132, 2004. Guido Menzio and Shouyong Shi. Block recursive equilibria for stochastic models of search on the job. Journal of Economic Theory, 145(4):14531494, 2010. 39 Guido Menzio and Shouyong Shi. Ecient search on the job and the business cycle. Journal of Political Economy, 119(3), 2011. Iacopo Morchio. Information frictions, match quality and lifetime unemployment. 2015. Dale T. Mortensen and Christopher A. Pissarides. Job creation and job destruction in the theory of unemployment. The Review of Economic Studies, 61(3):397415, 1994. John F. Nash, Jr. The bargaining problem. Econometrica, pages 155162, 1950. Felix Oberholzer-Gee. Nonemployment stigma as rational herding: A eld experiment. Journal of Economic Behavior & Organization, 65(1):3040, 2008. Christopher A. Pissarides. Search unemployment with on-the-job search. The Review of Economic Studies, 61(3):457475, 1994. Christopher A. Pissarides. Catherine Rampell. York Times, July Equilibrium Unemployment Theory. The MIT Press, 2000. The 26, help-wanted 2011. sign comes URL with a frustrating The New asterisk. http://www.nytimes.com/2011/07/26/business/ help-wanted-ads-exclude-the-long-term-jobless.html. Peter M. Robinson. Root-n-consistent semiparametric regression. Ariel Rubinstein. Perfect equilibrium in a bargaining model. Uta Schönberg. Econometrica, pages 931954, 1988. Econometrica, pages 97109, 1982. Testing for asymmetric employer learning. Journal of Labor Economics, 25(4):651691, 2007. Shouyong Shi. Directed search for equilibrium wage-tenure contracts. Robert Shimer. Econometrica, 77(2):561584, 2009. The cyclical behavior of equilibrium unemployment and vacancies. American Economic Review, 95(1):2549, 2005. Robert Shimer. On-the-job search and strategic bargaining. European Economic Review, 50(4):811830, 2006. Robert Shimer. The probability of nding a job. American Economic Review, 98(2):26873, 2008. Margaret Stevens. Wage-tenure contracts in a frictional labour market: Firms' strategies for recruitment and retention. The Review of Economic Studies, 71(2):535551, 2004. 40 Robert Topel. Specic capital, mobility, and wages: Wages rise with job seniority. Journal of Political Economy, 99(1):14576, 1991. Gerard J. Van den Berg. Duration models: Specication, identication and multiple durations. of Econometrics, 5:33813460, 2001. 41 Handbook A Figures Figure A.1: Job-Finding Rates as a Function of Duration 0.7 P[ Job Found Next Month ] 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0 10 20 30 Duration (Weeks) 40 50 Each x in the scatter plot represents the fraction of unemployed workers of a given duration who found a job by the time they were surveyed in the subsequent sample month. The number of workers who found a job and the number of workers with a given unemployment duration are weighted with CPS nal weights. The size of each x is proportional to the weighted number of workers who report that unemployment duration. The solid line is a local-polynomial kernel regression estimate. Data details are in Appendix D. 42 Figure A.2: Fitting the Population Hazard 0.7 One−Month Job−Finding Prob. 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0 10 20 30 Duration 40 50 The scatter plot is the same as the one in Figure A.1, except that the above excludes workers on layo. The solid line is the relationship implied by the model. Figure A.3: Market Tightness 8 7 6 θ(r) 5 4 3 2 1 0 0.2 0.4 0.6 r 43 0.8 1 Figure A.4: Wages of New Hires 1.5 1 0.5 wh(r|1) 0 −0.5 −1 −1.5 −2 −2.5 −3 0 0.2 0.4 0.6 0.8 1 r Figure A.5: Wedges in the Wage Equation for New Hires 2 1 0 Wedge −1 −2 −3 −4 −5 −6 0 Search Wedge Information Wedge 0.2 0.4 0.6 r 44 0.8 1 Figure A.6: Value of Unemployment: U (r) 1.12 1.1 ( 1 − β ) U(r) 1.08 1.06 1.04 1.02 1 0 0.2 0.4 0.6 0.8 1 r Figure A.7: Signal from Hiring: Counterfactual Résumés rc 1 0.9 Bu(bh(r)) o 45 0.8 0.7 rc 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 r 45 0.8 1 Figure A.8: Wages of Continuing Employees 2 1.8 we(r,x) 1.6 1.4 1.2 1 1 0.8 3 2 0.5 1 0 0 r x Figure A.9: Wedges in the Wage Equation for Continuing Employees 2 1.5 Search Wedge Information Wedge Wedge 1 0.5 0 −0.5 −1 −1.5 0 0.2 0.4 0.6 r 46 0.8 1 Figure A.10: Résumé Updating During Unemployment 1 Updated Résumé 0.8 0.6 0.4 0.2 0 0 1 20 0.5 40 60 0 Initial Résumé Unemp. Duration (Wks) Figure A.11: Meeting and Matching Probabilities as a Function of Résumé Matching Rate Meeting Rate 0.75 0.5 0.7 0.45 0.65 0.4 0.6 0.35 ω p(r) 0.5 r p(r) 0.55 0.45 0.3 0.25 0.4 0.2 0.35 0.15 0.3 0.25 0 0.5 r 0.1 0 1 47 0.5 r 1 Figure A.12: Contact Rates as a Function of Unemployment Duration 1 0.9 Probability: p(r) 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0 1 20 0.5 40 60 0 Initial Résumé Unemp. Duration (Wks) Figure A.13: Average Drop in Individual Hazards, by Duration Relative Job-Finding Rate 100 95 90 85 80 75 0 10 20 30 Unemp. Duration (Wks) 48 40 50 Figure A.14: The Role of Heterogeneity in the Population Hazard 0.25 Job−Finding Rate (Wkly) Average Hazard Hazard due to Heterogeneity Hazard due to Stigma 0.2 0.15 0.1 10 20 30 40 Unemp. Duration (Wks) 50 60 Figure A.15: Job-Finding Rates in the Full-Information Economy 0.5 0.45 Incomplete Info. Economy Full Info. Economy 0.4 ωr p(r) 0.35 0.3 0.25 0.2 0.15 0.1 0 0.2 0.4 0.6 r 49 0.8 1 Figure A.16: Aggregate Average Job-Finding Rates in the Full-Information Economy 0.25 Job−Finding Prob. (Wkly.) Incomplete Info. Economy Full Info. Economy 0.2 0.15 0.1 0 10 20 30 40 Duration (Weeks) 50 Figure A.17: Skill Change and the Job-Finding Rate 60 ωr p (r) 0.5 Avg. Job−Finding Prob. ωr p(r) 0.45 0.4 Baseline SC1 SC2 SC3 0.35 0.3 0.25 0.2 0.15 0.1 0 0.2 0.4 0.6 Résumé 50 0.8 1 Figure A.18: Skill Change and the Value of Unemployment U (r) 1.12 1.1 ( 1 − β ) U(r) 1.08 1.06 1.04 Baseline SC1 SC2 SC3 1.02 1 0 0.2 0.4 0.6 0.8 1 Résumé Figure A.19: Job-Finding Rates by Résumé, Alternative ω1 − ω0 0.7 0.6 ωr p(r) 0.5 0.4 0.3 0.2 0.1 0 0.3 0.2 ω1 − ω0 1 0.1 0.5 0 0 Résumé 51 Figure A.20: Average Aggregate Job-Finding Rates by Duration, Alternative ω1 − ω0 0.4 0.35 Probability 0.3 0.25 0.2 0.15 0.1 0.05 0 0.3 0.2 20 0.1 40 60 0 Duration ω1 − ω0 Figure A.21: Job-Finding Rates by Duration and Potential Experience 100 1st Quartile of Potential Experience 2nd Quartile of Potential Experience 3rd Quartile of Potential Experience 4th Quartile of Potential Experience Normalized Job-Finding Rate 95 90 85 80 75 70 65 60 0 10 20 30 Duration (Weeks) 52 40 50 Figure A.22: Job-Finding Rates by Duration and Potential Experience, Excluding Part Time Normalized Job-Finding Rate 100 1st Quartile of Potential Experience 2nd Quartile of Potential Experience 3rd Quartile of Potential Experience 4th Quartile of Potential Experience 95 90 85 80 75 70 65 0 10 20 30 40 50 Duration (Weeks) Figure A.23: Job-Finding Rates by Duration and Potential Experience, Model-Generated Data 100 1st Quartile of Potential Experience 2nd Quartile of Potential Experience 3rd Quartile of Potential Experience 4th Quartile of Potential Experience Normalized Job-Finding Prob. 95 90 85 80 75 70 65 0 5 10 15 20 25 30 35 Unemp. Duration (Weeks) 53 40 45 50 B Computational Outline I will begin by combining expressions to arrive at a concise system of functional equations that must be satised in equilibrium. These equilibrium functions must be the xed point of an operator, which I will dene below, so the computational strategy will be to iterate on this operator. I then provide some details about how this iterative procedure is executed in practice. Dene: Cs (r, , x) ≡ Gs (r, x) + Hs (r, x) . It will be convenient to evaluate the surplus-splitting condition (2.18) at Doing so, and using the denition of Cs (r, x), (B.1) (B e (r | x, x0 ) , x0 , e) and B h (r | x) , x0 , h . yields: Ge (B e (r | x, x0 ) , x0 ) = (1 − η) [Ce (B e (r | x, x0 ) , x0 ) − U (B e (r | x, x0 ))] Gh B h (r | x) , x0 = (1 − η) Ch B h (r | x) , x0 − U (B u (r)) . (B.2) (B.3) Combining the above with (2.10) and (2.14) gives us: Cs (r, , x) = zx + β X Ω (x, x0 | r) max {Ce (B e (r | x, x0 ) , x0 ) , U (B e (r | x, x0 ))} . (B.4) x0 Notice that the right-hand side does not depend on write C (r, x). s, so neither does the left-hand side. Hence, I will simply We can write the value of unemployment (2.15) as: U (r) = λ + β u U (B (r)) + ηζθ (r) 1− X h 0 u Ω̄ (x | r) max C B (r | x) , x − U (B (r)) , 0 ! . (B.5) x Using the surplus-splitting condition for new hires allows us to write the free-entry condition as: θ (r) = β X 1−η 1− ζθ (r) Ω̄ (x | r) max C B h (r | x) , x0 − U (B u (r)) , 0 . κ x (B.6) The above allows us to reduce further the value of unemployment to: U (r) = λ + ηκ θ (r) + βU (B u (r)) . 1−η 54 (B.7) Given C (r, x), U (r), and θ (r), the policy functions can be written as: X e (r | x) = X h (r) = Similarly, once we know θ (r), {x0 | C (re0 , x0 ) ≥ U (re0 ) , re0 = B e (r | x, x0 )} x | C (re0 , x) ≥ U (ru0 ) , rh0 = B h (r | x) , ru0 = B u (r) . (B.8) (B.9) we know the belief functions. Thus, solving for an equilibrium amounts to nding a solution to the following system of functional equations: C (x, r) = zx + β X Ω (x, x0 | r) max {C (B e (r | x, x0 ) , x0 ) , U (B e (r | x, x0 ))} (B.10) x0 U (r) θ (r) Dene T (C, U, θ) guesses for ηκ θ (r) + βU (B u (r)) 1−η X 1−η 1− Ω̄ (x | r) max C B h (r | x) , x0 − U (B u (r)) , 0 . ζθ (r) = β κ x = λ+ as the operator mapping (C, U, θ), I iterate on the T (C, U, θ) (B.11) (B.12) to the right-hand side of (B.10)-(B.12). Given initial operator until I attain practical convergence. Practically, I perform these iterations as follows. I will solve for the functions of interest on a grid of values 0 = r1 ≤ · · · ≤ rnr = 1. Dene: 0 r ≡ x ≡ r1 ··· rnr x0 x1 ··· (B.13) 0 where x0 nr is understood to be zero. Dene the nr × (nx + 1) xnx , matrix C (B.14) and the vectors C (r1 , x0 ) · · · C (r1 , xnx ) . . .. . . C ≡ . . . C (rnr , x0 ) · · · C (rnr , xnx ) 0 U ≡ U (r1 ) · · · U (rnr ) and Θ by: (B.15) (B.16) 0 Θ ≡ U θ (r1 ) · · · 55 θ (rnr ) . (B.17) nr × (nx + 1) Dene the 2 matrix: e e B (r1 | x0 , x0 ) · · · . .. . B ≡ . . B e (rnr | x0 , x0 ) · · · B (r1 | x0 , xnx ) e ··· B (r1 | xnx , x0 ) ··· . . . .. . . . e B e (rnr | x0 , xnx ) · · · . B e (rnr | xnx , x0 ) · · · e B (r1 | xnx , xnx ) . , . . B e (rnr | xnx , xnx ) (B.18) where elements of interpolation of Be are imputed to be zero for infeasible 11×(nx +1) ⊗ C along the grid given by Ce = where each Ceh is nr × (nx + 1), Ξei where Ω (xh , · | i) Dene Ue Ce be the column-wise Ce0 Ce1 Cenx ··· element of Ceh , (B.19) C (B e (rj | xh , xk ) , xk ). is 1(nx +1)×1 ⊗ Ω0i Inx +1 ⊗ 1(nx +1)×1 0 Ω (x0 , · | i) 0(nx +1)×1 ··· 0(nx +1)×1 . .. . 0(n +1)×1 Ω (x1 , · | i)0 . . x = . .. .. . . . . 0(nx +1)×1 0 0(nx +1)×1 ··· 0(nx +1)×1 Ω (xnx , · | i) Dene: ≡ represents the (h + 1) st row of as the column-wise interpolation of U = (j, k + 1) element of Ueh is Ωi , and , (B.20) represents element-by-element multiplication. U ⊗ 11×(nx +1)2 along the grid given by e where the transitions. Let That is: (j, k + 1) and the Be . (x, x0 ) Be . That is: Ue0 Ue1 ··· U (B e (rj | xh , xk )). Uenx , (B.21) Observe that column (h + 1) of max {Ce , Ue } Ξei is 0 max {Ceh , Ueh } Ω (xh , · | i) , which gives the expected value (over x0 ) of max {C (B e (r | xh , x0 ) , x0 ) , U (B e (r | xh , x0 ))} under the distribution Ωi , conditional on starting with TC (C, U, Θ) x = xh . The updated value of = z1nr ×1 x0 + β (max {Ce , Ue } Ξe1 ) r11×(nx +1) C is given by: +β (max {Ce , Ue } Ξe0 ) 1nr ×(nx +1) − r11×(nx +1) . Dene the nr × 1 (B.22) vector: u B ≡ 0 u B (r1 ) · · · 56 u B (rnr ) . (B.23) Dene Uu as the interpolation of U along the grid given by TU (C, U, Θ) = λ1nr ×1 + Dene the nr × 1 Bu . The updated value of U is given by: ηκ Θ + βUu . 1−η (B.24) vector: Bh ≡ 0 h B (r1 ) · · · h B (rnr ) . (B.25) The above incorporates the specication assumed in the calibration exercise; namely, all newly hired workers have the same productivity the grid given by Bh . x = 1. Ch Dene as the column-wise interpolation of C ( : , I [x0 = 1] ) Dene: P ≡ min ζΘ1− , 1 value of Θ (B.27) min {·} operator and the power (1 − ) are understood to apply element-by-element. The updated is given by: TΘ (C, U, Θ) = β C (B.26) ≡ rω1 + (1nr ×1 − r) ω0 , ω ~r where the along 1−η κ Pω ~ r max Ch − Uu , 0 . (B.28) Proof of Proposition 1 Evaluating equation (B.11) at Hs (r, x) − U (rc ) = rc and subtracting it from ws (r, x) + β X Hs (r, x), as dened in equation (2.14), yields: Ω (x, x0 | r) max {He (B e (r | x, x0 ) , x0 ) − U (B e (r | x, x0 )) , 0} x0 −λ + X Ω (x, x0 | r) [U (B e (r | x, x0 )) − U (B u (rc ))] − β x0 κη θ (rc ) 1−η X η β Ω (x, x0 | r) max {Ge (B e (r | x, x0 ) , x0 ) , 0} 1−η 0 x X κη 0 −λ + β Ω (x, x | r) [U (B e (r | x, x0 )) − U (B u (rc ))] − θ (rc ) , 1−η 0 = ws (r, x) + (C.1) x where the second line uses the surplus-splitting condition (2.18) evaluated at (r, x, s). Multiplying the above by 1 − η, multiplying equation (2.10) by η, (B e (r | x, x0 ) , x0 , e) equating the results using the surplus-splitting condition (2.18), and rearranging terms completes the proof. 57 instead of D Data I use the basic monthly CPS data from 1994 to 2013, available from the NBER's website, maintained by 39 Although earlier data are available, the CPS underwent a redesign in 1994, and it's known Jean Roth. that this change altered the measured distribution of unemployment durations. The CPS is a rotating panel, where the same household is interviewed for 4 months, not interviewed for 8 months, and then interviewed for 4 months. Questions include labor force status and, if unemployed, the duration of search. The goal is to identify unemployed people who have been unemployed for a given duration, and we want to know what fraction of these people found jobs by the following interview date. structure of the CPS, it's necessary to match people across months. However, to exploit the longitudinal To do this, I use a method similar to the one used by Shimer [2008]. I merge the monthly les on the basis of household identier (hrhhid), sample identier (hrsample), serial sux (hrsersuf ), household number (hrhhnum), and person line number (pulineno). When I nd observations that match across consecutive months, I drop observations in which the person's race, sex, or indication of Hispanic origin changes. changes by two or more years between months. I also drop observations in which age The unemployment duration data are heavily imputed, except for workers in their rst or fth months in sample. To minimize the eects of imputation, I conne my sample to workers who were unemployed in their rst or fth interview month. Ultimately, I obtain 194,715 observations where a worker was unemployed in one month, and I can observe the same worker in the following month. In Section 7.1, I explore the data by tting a semiparametric regression, as specied by equation (7.1). The non-linear component h (·) is assumed to be locally cubic. The covariates in xi include race, sex, Hispanic origin, reason for unemployment, an indicator for being a high-school dropout, an indicator for being a highschool graduate, an indicator for whether the worker was seeking part-time employment, the state-level unemployment rate, and a quadratic function of age. The sample excludes people on temporary layo because many of them are recalled by their previous employers (Fujita and Moscarini [2013]). I also conne the sample to workers ages 18 to 64. With these restrictions, there are 155,638 observations in the sample; of these, 13% indicate that they are seeking part-time employment. Potential labor-market experience is dened as age minus years of education. Because the CPS asks people about their level of schooling, not their years of schooling, this requires imputing a certain number of years for each level of schooling. I impute that those who did not reach 9th grade received 13 years of education; those who reached a maximum of 9th grade (10th grade, 11th grade) received 14 years (15 years, 16 years) of education; those who reached 12th grade but did not graduate from high school received 17 years of education; those who attended some 39 See: http://www.nber.org/data/cps_basic.html. 58 college but did not graduate received 19 years of education; and those whose highest degree is a high-school diploma (associate's degree, bachelor's degree, master's degree, professional degree, doctorate) received 18 years (20 years, 22 years, 23 years, 24 years, 26 years) of education. 59