Raising the Bar for Models of Turnover

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Raising the Bar for Models of Turnover Erwan Quintin and John J. Stevens 2005-23 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Raising the Bar for Models of Turnover Erwan Quintin John J. Stevens (cid:3) Federal Reserve Bank of Dallas Federal Reserve Board April 26, 2005 Abstract Itiswellknownthatturnoverratesfallwithemployeetenureandemployersize. We documentanewempiricalfactaboutturnover: Amongsurvivingemployers,separation rates are positively related to industry-level exit rates, even after controlling for tenure and size. Speci(cid:12)cally, in a dataset with over 13 million matched employee-employer observations for France, we (cid:12)nd that, all else equal, a 1 percentage point increase in exit rates raises separation rates by 1/2 percentage point on average. Among currentyear hires, the average e(cid:11)ect is twice as large. This relationship between exit rates and separation rates is robust to a host of data and statistical considerations. We review several standard models of worker turnover and argue that a model with (cid:12)rm-speci(cid:12)c human capital accumulation most easily accounts for this new empirical fact. Keywords: Firm survival; Employee turnover; Human capital. JEL classification: J24; J31; J63. ∗E-mail: erwan.quintin@dal.frb.organd john.j.stevens@frb.gov (corresponding author). We thank David Byrne, Paul Lengermann, and seminar participants at the Federal Reserve Bank of Atlanta, Southern Methodist University, Arizona State University and the Federal Reserve Bank of Minneapolis for their comments on earlierdrafts ofthis paper. We alsothank Olga Zograffor her computationalsupport and the Federal Reserve Bank of Kansas City for the use of their Beowulf computing cluster. The views expressed herein are those of the authors and may not reflect the views of the Federal Reserve Bank of Dallas, the Federal Reserve Board, or the Federal Reserve System.

1 Introduction The economies of industrialized nations are characterized by a tremendous amount of churning. Establishments die, workers quit or lose their jobs, industries contract, and occupations decline inimportance. This churning hasimportant implications forworkers’ decisions about human capital accumulation, job search, and the nature of employment contracts workers write with their employers. Of course, the reverse is also true: workers’ decisions have important implications for turnover and establishment survival. Regardless of the direction of causality, credible models of turnover need to be consistent with a set of well-established empirical facts: \(1) long-term employment relationships are common, (2) most new jobs end early, and (3) the probability of a job ending declines with tenure" (Farber, 1999, page 2441) and (4) turnover rates tend to decrease with employer size (see, for example, Anderson and Meyer, 1994). In this paper, we argue that theoretical models of turnover should be consistent with an additional fact: Among surviving establishments, turnover is positively related to industry-level exit rates even after controlling for di(cid:11)erences in the distribution of employee tenure and establishment size. Figure 1 illustrates the positive relationship between turnover and exit rates at the industry level. This relationship reflects, among other things, the fact that industries with lower exit rates tend to have larger (cid:12)rms and workers with more tenure. We (cid:12)nd, however, that exit rates and turnover rates remain strongly correlated even after controlling for differences in employer size and employee tenure. Using a dataset of over 13 million matched employee-employer observations for France, we estimate that, all else equal, a 1 percentage point increase in exit rates implies a 1/2 percentage point increase in the likelihood of the worker separating from his employer. Among current-year hires, a 1 percentage point increase in exit rates is associated with a 1 percentage point increase in the likelihood of a separation. This new empirical (cid:12)nding raises the bar for theoretical models of turnover. We review several standard approaches to modeling worker turnover and discuss the assumptions each theory needs in order to be consistent with the fact that workers are more likely to separate when establishment exit rates are higher. The approaches we consider are a model with 1

(cid:12)rm-speci(cid:12)c human capital (as in Oi, 1962, Becker, 1962, or Jovanovic, 1979b), a model with learning-by-doing (as in Jovanovic and Nyarko, 1995), and a model with job matching (as in Jovanovic, 1979a, or Mortensen and Pissarides, 1994). Models with (cid:12)rm-speci(cid:12)c human capital immediately predict the desired correlation. Learning-by-doing and matching models require stronger assumptions. In the learning-by-doing case for instance, the exogenous pace of learning needs to di(cid:11)er across employers with di(cid:11)erent propensities to exit. Naturally, the pace of learning does in fact vary across industries.1 The point of these theoretical exercises is not to suggest that matching models and models in which learning is exogenous are flawed or inconsistent with the evidence on turnover. Rather, we want to illustrate how the empirical regularity we document should help us build better models of turnover by stressing the importance of certain assumptions. 2 Data and institutional background Our principal source of data is the D(cid:19)eclaration Annuelle des Donn(cid:19)ees Sociales (DADS) database, which is distributed by France’s national statistical institute, Institut National de la Statistique et des Etudes Economiques (INSEE). These data include detailed information on workers and their employers. French law requires all employers to provide information on all employment spells during the year by January 31 of the following year. For each employment spell, employers provide a beginning and an end date, the total number of hours worked, the employee’s age, gender, whether the employee works on a full-time basis,2 and some occupation information (in 34 categories). The employer information includes the establishment’s industry (in 36 categories), geographical location, employment size on December 31, and gross payroll. The 2002 \Postes Exhaustif" version of the database we used excludes all payroll and earnings information.3 The DADSdata aremore comprehensive than other sources of data on French employees. 1Whether ornotlearningvariessystematicallywithexitratesisanempiricalquestionthatisbeyondthe scope of this paper. 2Thesurveyclassi(cid:12)esemployeesinfourcategories: full-time,part-time,\intermittents,"andhome-based. Intermittent workers are employees with a long-term contract who do not work year-round. 3Thesedataareacross-sectionalversionofthepaneldatausedbyAbowdetal(1999). Thepanelversion of these data is no longer available from INSEE. 2

In particular, France’s D(cid:19)eclaration Mensuelle de Main-d’Oeuvre (DMMO), which Abowd et al (1996) and Nagyp(cid:19)al (2002) have used for other purposes, only covers establishments with at least 50 employees.4 Furthermore, to obtain a matched employee-employer data set, the DMMO must be linked with sources of information on employers, for instance France’s Enqu^ete sur l’Emploi. Many observations are lost in the matching process, and the resulting data set emphasizes employment relationships less prone to separations, which is a problematic bias given our purposes (see Nagypa(cid:19)l, 2002). But unlike DMMO, the DADS data we use do not contain any information on what causes employment spells to end. Possible causes include 1) transfers within the (cid:12)rm, 2) military service, retirement, sickness, leave, or death, 3) end of apprenticeship or internship, 4) end of a short-term contract (Contrat (cid:18)a Dur(cid:19)ee D(cid:19)etermin(cid:19)ee, or CDD), 5) end of trial hire, 6) dismissal, or 7) quit. Standard models of turnover do not make any prediction for the (cid:12)rst three types of separations. The (cid:12)rst two account for less than 10 percent of all separations, and the bulk of these separations reflect retirements (Abowd et al, 1996, table 4). In order to minimize the number of retirement episodes in our sample, we drop all employees 55 years of age or older. We also exclude apprentices and interns from the analysis. More generally, we want to concentrate our attention on potentially durable employment relationships. To that end, we only include full-time employees in our analysis. We also exclude temporary workers by dropping all observations from the \Operational Services" category in France’s NES36 classi(cid:12)cation.5 We exclude most summer jobs by dropping employees under 26 years of age6 and the declarations (cid:12)lled by not-for-pro(cid:12)t associations and employers in the recreative activity industry. The details of our sampling decisions are provided in appendix B.1. The resulting sample contains 13,068,665 observations and is referred to in the following sections as the \all-employees sample." 4In 2000, over 90 percent of French (cid:12)rms outside of the Agriculture and Finance sectors counted fewer than50employees. Intheindustrialsector(miningandmanufacturing),firms withfewerthan50employees represented one-third of all employment. See INSEE, 2000/2001,p137. In our sample, establishments with fewer than 50 employees represent approximately 45 percent of all employment. 5That industry includes all temporary work agencies but also includes establishments that rent vehicles, machinery, equipment and appliances to businesses and individuals, establishments that provide investigation, security and cleaning services, and establishments involved in researchand development activities. 6The age restriction also reduces the likelihood of a separation due to military service. In any event, separations due to military servicein the 2002DADS arelikely very few. Since 1997,males born after 1978 are exempt from the conscription. 3

The empirical question we ask is whether employment spells with similar observable employer-employee characteristics are more likely to end in industries with high exit rates than in industries with low exit rates, even at similar tenure levels and comparable employer sizes. We calculated two proxies for industry-level exit rates from two distinct data sources. Our 2002 DADS dataset reports employer size as of December 31, 2002. Establishments with positive employment at some point during 2002 but no employment on December 31 are assigned size zero. Our (cid:12)rst proxy for exit rates is the fraction of employees in each industry whose employer reported size zero in the 2002 DADS. Note that by construction this proxy for exit rates is employment weighted. Our second proxy for exit rates comes from the OECD (cid:12)rm-level data project (see Bartelsman, et al., 2003, for a detailed description of the project). The French section of the OECD (cid:12)rm-level database reports for each year between 1991 and 1996 the number of continuing (cid:12)rms in 61 industries (constructed from two-digit ISIC categories) at the beginning of the year, the number of entering (cid:12)rms during the year, the number of (cid:12)rms that exited theindustry by the end of the year, and the number of employees in each of these categories of (cid:12)rms. In particular, it is possible to calculate the fraction of employees in exiting (cid:12)rms in each industry. To make this measure compatible withourDADSdatabase, itisnecessary tomaptheISIC-basedcategoriesusedbytheOECD into France’s NES36 classi(cid:12)cation. The details of this concordance are in appendix B.3. The resulting proxy for exit rates di(cid:11)ers from the DADS-based proxy for two main reasons: it is (cid:12)rm based rather than establishment based, and it is based on data for the 1991-96 time period rather than for 2002. Table 1 shows the sample means of the variables we use in our analysis. In particular, note that like in the U.S. (e.g., Anderson and Meyer, 1995, table 2), separation rates and establishment exit rates are higher in the trade sector (wholesale trade, retail trade and transportation) than in the service sector, higher in the service sector than in the manufacturing sector, and lowest in the public sector. Exit and separation rates also fall with the establishment’s employment size. Table 2 reports means for those workers hired after January 1; we refer to this sample in subsequent sections as the sample of \current-year hires." As in the U.S., separation rates in France are uniformally and markedly higher in the (cid:12)rst year of employment than in subsequent years. 4

When comparing these statistics to their U.S. counterparts however, it is important to bear in mind key distinguishing features of French labor markets. Legal restrictions on hiring and (cid:12)ring are more stringent in France than in the U.S., and, not surprisingly, hiring and (cid:12)ring rates are lower in France than in the U.S. In fact, the annual separation rates shown in table 1 are of the same order of magnitude as the quarterly separation rates reported by Anderson and Meyer (1995, table 2). Another key feature of French labor markets is the importance of short-term contracts (CDDs). Abowd et al. (1996) calculate that roughly onehalf of all terminations are ends of short-term contracts, and that roughly 70 percent of new employees are hired under CDDs. Under French law, employers may use CDDs for youth employment programs, completing temporary tasks, and testing new hires. These contracts can be renewed once for a total length of up to 18 months. Employers use CDDs for the majority of hirings, because CDDs carry fewer mandated costs (e.g., severance payments) than contracts of inde(cid:12)nite length (CDIs). Upon expiration, one-third of CDDs become long-term contracts, while the remainder are either renewed or terminated. Our data do not permit us to examine whether our main result|that separations are more frequent in industries with high exit rates|holds for both types of contracts. 3 Empirical Results At the industry level, a higher rate of establishment exits is positively related to the rate of worker separations ((cid:12)gure 1). This relationship suggests that workers in industries with high exit rates are more weakly attached to their employers than employees in industries where establishments are more likely to survive. However, this conclusion does not control for systematic di(cid:11)erences in the industry-level characteristics of employees and employers. For example, older workers are known to have lower rates of separation, therefore, a highsurvival industry will exhibit lower separation rates if it tends to employ older workers. This section investigates whether the positive correlation between exit rates and separation rates persists after controlling for the employer-employee characteristics we observe in the data described in the previous section. Webeginbymodeling theworker’s separationoutcomeusing aprobitmodelthatcontrols 5

for the worker and establishment characteristics available to us, as well as the estimated industry-level exit rates. As shown in the all-employees columns of table 3, coe(cid:14)cients have the expected signs.7 Older workers are less likely to separate, and the marginal decrement to theprobability ofseparationdeclines withage. Generally, employees whoseoccupationranks higher in the French occupational classi(cid:12)cation system (see appendix B.2) are less likely to separate. For instance, employees in occupational category 1 are less likely to separate than those in occupational category 4; however, occupational category 2, which includes top-level managers and other high-level professionals, appears to be an exception. The likelihood of a separation also declines monotonically with the employment size of the establishment. Workers in establishments with 100 to 199 employees are considerably less likely to separate thanemployees inestablishments with1to19employees (theexcluded category). Employees in the Paris agglomeration are more likely to separate. Separations are more likely in the services sector than in the manufacturing sector and more likely in the private sector than in the public sector. Most signi(cid:12)cantly for our purpose, exit rates continue to exhibit a strong positive correlation with the likelihood of a separation even after controlling for the employee and employer characteristics just described. Finally, a dummy variable that equals one if the employee is not a current-year hire indicates that workers with more tenure are, unsurprisingly, much less likely to separate from their employer. In order to better control for the e(cid:11)ects of tenure on the likelihood of a separation, for the remainder of this study we restrict our attention to current-year hires. All employees in this sample have a year of tenure or less. Within that sample, it is in fact possible to control for tenure more (cid:12)nely as we know the exact date of hire; in the next section we make use of this information to better control for seasonal employment. This tenure category is also of particular importance for understanding the determinants of turnover as, in France like in the U.S., separations are much more likely to occur within the (cid:12)rst year of employment than at higher tenure.8 The results for this sample of current-year hires, shown in table 3, are qualitatively similar to the results for the all-employees sample. Among the di(cid:11)erences, separation rates 7Statistical signi(cid:12)cance is not an issue given the size of our sample. 8Separation rates exceed 40 percent among current-year hires, compared to about 11 percent among returning employees. Half of all separation episodes in our sample occur among current-year hires. 6

now decline monotonically across all of the occupation codes (employment category 2 is no longer an exception). Also, the sign of the coe(cid:14)cient of the geographical location dummies changes, and suggests that among current-year hires, separation rates are lower in urban areas, particularly outside of Paris. Importantly, the coe(cid:14)cient on exit rates is larger in the sample of current-year hires than it was for the all-employees sample. Workers hired during the current calendar year are more sensitive to the industry-level exit rates. A possible explanation for this (cid:12)nding is that tenure dilutes the e(cid:11)ect of exit rates, as workers have time to learn about any idiosyncratic di(cid:11)erences between the establishment’s likelihood of exit and the overall likelihood suggested by the industry exit rates. Table 4 translates the coe(cid:14)cients on the exit rate into marginal e(cid:11)ects.9 For example, a 1 percentage point increase in the industry exit rate (from the mean) raises predicted separation rates by 1.115 percentage points (from the mean) in the sample of current-year hires, as compared to 0.447 percentage point in the all-employees sample. This di(cid:11)erence is not terribly surprising given the much smaller rate of worker separations in the all-employees sample. The two rightmost columns of table 4 present some relative e(cid:11)ects. A 1 standard deviation increase in the exit rate (which varies between the two samples) is associated with a 0.14 standard deviation increase in predicted separation rates in the all-employees sample and a much larger 0.34 standard deviation increase inpredicted separation rates for the sample of current-year hires. Similarly, a 1 standard deviation increase in exit rates is associated with an 8.5 percent increase in predicted separation rates in the all-employees sample and a 9.8 percent increase in predicted separation rates for the current-year hires. Thus, the relationship in (cid:12)gure 1 appears robust to controlling for the characteristics of employees and employers. A shortcoming of the probit model, however, is that it is a model of worker separations conditional on the establishment having survived. In general, whether a termination occurs is the result of two binary outcomes in each period: whether or not the establishment survived, and whether or not the employment relationship was terminated for reasons other than establishment failure. We are primarily interested in this latter outcome, but it is only partiallyobservable. Speci(cid:12)cally, weonlyobserve theseparationdecision iftheestablishment does not exit the industry. Our data do not permit us to identify whether a separation would 9Wecomputedthemarginale(cid:11)ectforeachobservation. Thee(cid:11)ectsshownintable4aresampleaverages. 7

have occured had the establishment survived. The probit model sidesteps the partial observability problem by discarding, in the case of thesampleofcurrent-yearhires, theroughly7-1/2percentofobservationsinwhichtheestablishment exits. Intuitively, we would expect separation rates to be higher in establishments that are more prone to exit, therefore, discarding these observations should downwardly bias the coe(cid:14)cient on exit rates. One way to control for this potential bias is to use a bivariate probit model (see Poirier, 1980; Meng and Schmidt, 1985; Greene, 1998). The bivariate probit is the discrete-choice analog to the standard sample selection correction frequently used in regression-based models. Index workers by i and let S 2 f0,1g be an indicator variable for whether or not i worker i separates from their employer; likewise, let A 2 f0,1g be an indicator variable i for whether or not the worker’s establishment survives the year. The function of interest is Pr(S = 1jA = 1), that is, the probability that the worker separates conditional on the i i establishment surviving the year. The basicidea isthat therearecharacteristics ofemployers seen by workers but unobserved by the econometrician. For example, the worker may know that the establishment is exiting and quit prior to its actual demise. However, if both events occur within the same year, our data do not reveal the fact that the worker quit. Formally, the model consists of a worker separation equation, an establishment survival equation, and a selectivity relationship that links the erros in the other two equations. The separation equation is S∗ = β(cid:48)x +(cid:15) i i i S = 1 if and only if S∗ > 0 and 0 otherwise, i i where x includes the industry exit rate and controls for worker and establishment chari acteristics. These controls include worker age, gender, occupational and broad industry dummies, establishment size dummies, dummies for location in Paris or other urban areas, and a dummy variable for being in the private sector. The establishment survival equation 8

is A∗ = α(cid:48)z +u i i i A = 1 if and only if A∗ > 0 and 0 otherwise. i i Note that S is observed only if A = 1 whereas A , x , and most elements of z are observed i i i i i forallworkers. Inourbenchmarkmodel, z containsthesamecovariatesasx except thatthe i i establishment sizedummiesareexcludedbecausewedonotobservethesizeofestablishments that exit. Other exclusion restrictions are introduced and motivated in the next section. The selectivity relationship is governed by the correlation ρ between the errors in the worker (cid:15)u and establishment equations, where:       (cid:15) 0 1 ρ  i (cid:24) N  , (cid:15)u u 0 ρ 1 i (cid:15)u If ρ = 0, there is no selection e(cid:11)ect: the (cid:12)rm’s exit outcome and the worker’s propensity (cid:15)u to separate are independent given observable employer and worker characteristics. Column 1 in table 5 displays the results of the bivariate probit estimation. As shown in the last row, ρ < 0, which implies that establishments more likely to fail because of (cid:15)u characteristics not spanned by our control variables (unobserved characteristics) also tend to have lower separation rates, when they survive. This should bias downward the coe(cid:14)cient on exit rates in the probit model. Indeed, establishments more likely to fail for unobserved reasons are less likely to survive in high failure rate industries. All else equal then, in industries with high failure rates, our data will emphasize employers more likely to survive for unobserved reasons. Not surprisingly then, the estimated coe(cid:14)cient on exit rates in the separationequationofthebivariatemodelislargerthanintheprobitmodel. The coe(cid:14)cients on establishment size are smaller in the bivariate probit model. This likely reflects the fact that exit rates are negatively related to establishment size, so that explicitly modeling exit partiallymitigatestheroleofestablishment size. Thesignsandmagnitudesofothervariables are otherwise little changed. The results for the establishment survival equation may be of independent interest. Naturally, those working in industries with high exit rates are more 9

likely to see their establishment exit. Older workers, women, and workers in urban areas are also more likely to see their establishment exit. There are many marginal e(cid:11)ects one can compute in a bivariate probit model such as ours, but the one most directly comparable to the single equation case is the sample average of the partial derivative of Pr(S = 1jA = 1) with respect to exit rates.10 The third row of i i table 4 shows that this marginal e(cid:11)ect is a bit larger in the bivariate probit case than in the single equation case, both in absolute and in relative terms. 4 Robustness In this section we explore the sensitivity of our (cid:12)ndings to features of French labor markets and various statistical considerations. The last four columns of table 5 include results from several of these robustness checks, and table 4 has the corresponding marginal e(cid:11)ects for exit rates. Exclusion restrictions. Our bivariatemodelisformallyidenti(cid:12)ed. However, Keane(1992) points out (in a di(cid:11)erent context, that of multinomial probit models) that identi(cid:12)cation can be tenuous even in formally identi(cid:12)ed models.11 For that reason, it is now common practice to use exclusion restrictions to aid with identi(cid:12)cation in multi-equation models. Given the data available to us, we chose to include statistics from the empirical distribution of surviving establishments in the employer’s industry in the survival equation.12 Speci(cid:12)cally, we included for each industry the fraction of surviving establishments in the four following categories: 0 to 49 employees, 50 to 99 employees, 100 to 199 employees, and 200 employees 10Formally, let β x and α x be the coe(cid:14)cients of the exit rate variable in the separation and the survival equations, respectively. Then the marginal e(cid:11)ect for workeri is (cid:34) (cid:35) (cid:8)(α 1 (cid:48)z i) β x φ(β(cid:48)x i)(cid:8)( α(cid:48)z (cid:112)i − 1− ρ (cid:15) ρ u 2 (cid:15) β u (cid:48)x i )+α x φ(α(cid:48)z i)(cid:8)( β(cid:48)x (cid:112)i 1 − − ρ (cid:15) ρ u 2 (cid:15) α u (cid:48)z i )−α x(cid:8)2(α(cid:48)z i ,β(cid:48)x i ,ρ (cid:15)u) (cid:8) φ( ( α α (cid:48) (cid:48) z z i i ) ) whereφ, (cid:8) denotethe density functionandthe cumulativedistributionofthe standardnormaldistribution, respectively,while,(cid:8)2 isthecumulativedistributionofthebivariatenormaldistributionwithunitstandard deviations. See Greene (1997, p.910). 11We did not experience any of the classic symptoms of fragile identi(cid:12)cation with any of our speci(cid:12)cations: in all cases, our maximum likelihood algorithm converges to the same solution from any set of initial conditions, and estimated standard errors are very small for all coe(cid:14)cients. 12Recall that due to data limitations, we cannot include individual establishment size information in the survival equation. Establishments that do not survive all report size 0 in the survey. 10

or more. Given that we know the actual size of the worker’s establishment, these statistics should contain little new information about the worker’s separation decision. But, from an establishment survival perspective, the size distribution of surviving establishments is likely correlated with the industry’s overall size distribution. Therefore, it should provide some information on the survival prospects of individual establishments. We do (cid:12)nd that these new variables signi(cid:12)cantly a(cid:11)ect survival outcomes. But as shown in table 5, these exclusion restrictions have almost no e(cid:11)ect on the coe(cid:14)cients of the separation equation. Obviously, the marginal e(cid:11)ects change little as well. Alternative exit rates. We repeated our analysis using the (cid:12)rm-level exit rates we obtained from the OECD. The rates we used previously were e(cid:11)ectively employee-weighted establishment-level exit rates, therefore we chose to weight the OECD (cid:12)rm-level rates by employment to make them more comparable.13 Although the correlation between the two sets of exit rates is only 9 percent, some of the industries that deviate from the regression line between the rates are fairly small industries that have little e(cid:11)ect on the aggregate. For example, removing just the realty industry boosts this correlation to 26 percent. Realty is dominated by large, multi-establishment companies; at the corporate-level survival rates are high, despite considerable churning at the establishment level. The bivariate probit results we obtain with these alternative exit rates are qualitatively unchanged, as can be seen in column 3 of table 5.14 The fourth row of table 4 shows that although the absolute marginal e(cid:11)ect is nearly 3 times larger in this case, the relative marginal e(cid:11)ects are qualitatively similar and only a bit larger quantitatively than what we obtain with the sample of current-year hires using DADS-based establishment exit rates. The large marginal e(cid:11)ect has a considerably smaller impact on the relative e(cid:11)ects because the (cid:12)rm-level exit rates we calculated with OECD data are less volatile than the DADS-based rates. Another virtue of using OECD exit rates is that they alleviate the concern that cyclical factors could drive our results. A negative industry-speci(cid:12)c demand shock might result in 13The results we obtain using the unweighted (cid:12)rm-level rates (not shown), are qualitatively similar. 14Estimating ρ (cid:15)u precisely proved di(cid:14)cult in this case. This likely owes to the fact that we now have fewer industries (21 vs. 34), which raises the collinearity between exit rates and our set of broad industry dummies. Infact,droppinganyofthe3industrydummiesfromthesurvivalequationsolvestheconvergence problems we experienced with our initial speci(cid:12)cation. The results we show in column 3 of table 5 are for the case where the trade dummy is excluded from the survival equation. Coe(cid:14)cients change little when another industry dummy is excluded. 11

the exit of some establishments and layo(cid:11)s at the remaining establishments, which would give rise to a positive relationship between exit rates and separation rates. The OECD exit rates we use are 6-year averages, which mitigates the possible e(cid:11)ect of cyclical shocks. In addition, demand shocks that occured in the (cid:12)rst half of the 1990’s are less likely to a(cid:11)ect separation rates in 2002. Non-seasonal workers. Despite our sampling restrictions, our data may still contain some seasonal workers. To test whether our results are sensitive to seasonal employment, we further restrict our data by removing all workers with less than 6 months of tenure. More precisely, the sample of non-seasonal workers only includes workers hired during the (cid:12)rst 3 months of the year whose employment spell lasts at least 6 months. We then consider a separation to have occurred if the spell ends between month 6 and month 9 of employment.15 This procedure removes all workers hired for short-lived tasks and reduces the number of observations among current year hires by about 85 percent. Nonetheless, the qualitative relationships we found in the base case are largely unchanged (column 4). The marginal e(cid:11)ect of exit rates is smaller than in the base case, but that is because separations occur at a lower rate in this sample as they can only occur in a three-month window. In fact, the relative e(cid:11)ects are of a magnitude similar to what we found with the sample of current-year hires. Industry restrictions. Substantial government involvement in some industries (both directly and via preferential treatment) likely a(cid:11)ects the relationship between establishment exit and worker separation. Therefore, we repeated our analysis excluding the roughly 331,000 establishment found in industries where government involvement is often signi(cid:12)cant; these industries include agriculture, postal services and telecommunications, education, social work, public administration, and associations. The bulk|about 80 percent|of the excluded establishments were attributable to the exclusion of social work and public administration. The bivariate probit results with these industry restriction (column 5) are qualitatively little changed from the base model and the marginal e(cid:11)ects (table 4) are quantitatively quite similar to the model with the alternative exit rates. Looking within size classes and broad industry categories. The size class and broad in- 15This de(cid:12)nition gives all workers in the sample the same window for a separation to occur. 12

dustry dummies we include as controls in our model speci(cid:12)cations are correlated with exit rates. For instance, small establishments are more likely to fail than large establishments, and exit rates are higher in the trade sector than in the manufacturing sector. In principle, thestatistical signi(cid:12)cance of exit ratescouldreflect thefact thatsize classes andbroadindustry categories are associated with separation rates in complex, nonlinear ways that dummy variables cannot capture. To assuage this concern, we re-estimated our models separately for each size class and broad industry category. In all cases, the coe(cid:14)cients and marginal e(cid:11)ects of exit rates remain positive, large, and signi(cid:12)cant. These results are available upon request. Occupation-specific differences. Occupational di(cid:11)erences may matter in ways not captured by our (cid:12)ve broad occupational controls. In order to assess this possibility we ran separate probits for 26 occupational categories (the number of categories with which our sampling restrictions leave us). In 23 cases, the coe(cid:14)cient on exit rates was positive and statistically signi(cid:12)cant; in the remaining three cases|public service executives, intermediate public service executives, and farm hands|the coe(cid:14)cient on exit rates was insigni(cid:12)cantly di(cid:11)erent from zero.16 The occupational probits also help to minimize concerns that our results reflect a correlation between occupations and particular worker unobservables. For example, some occupations tend to be characterized by workers that have a greater degree of \footlooseness," such as truck drivers or farm hands. Footloose workers may care less about the risk of establishment failures as they anticipate separating from their employer anyway. Therefore the data might show a positive correlation between establishment exit rates and worker separation rates due to this particular unobservable characteristic of workers. The occupational probits suggest that our results are either robust to this concern, or that these worker unobservables have little correlation with occupations. 5 Three models of turnover This section discusses the consistency of standard models of employee turnover with the empirical regularity we document in this paper. We (cid:12)rst argue that models of turnover 16These results are included in a separate appendix available from the authors. 13

founded on speci(cid:12)c human capital predict that separation rates should be higher in (cid:12)rms (or establishments) more likely to exit, even at equal levels of employee tenure. Exogenous learning-by-doing models, on the other hand, only predict a positive correlation between exit rates and separation rates at equal tenure provided the rate of learning di(cid:11)ers across employers with di(cid:11)erent likelihoods of survival. For matching models to predict the desired correlation, the process that governs the evolution of worker productivity must di(cid:11)er across employers, or search must be endogenous. We emphasize once again that our objective when comparing simple models of turnover is not to suggest that matching models or exogenous learning models are flawed, but rather to stress the importance of certain features of these models. We make these points in the context of a stylized general equilibrium environment in which time is discrete and in(cid:12)nite. Each period, constant measures of (cid:12)rms and workers are born. Firms belong either to industry H or to industry L.17 The fraction of (cid:12)rms born in each industry does not vary over time, and (cid:12)rms remain in the industry to which they are born until they die. In industry H, (cid:12)rms survive from one period to the next with likelihood p while in industry L they survive with likelihood p < p .18 Industries are H L H otherwise identical. In each period, (cid:12)rms in both industries can transform labor n (cid:21) 0 into quantity f(n) of the consumption good, where f is strictly concave, strictly increasing, and lim f(cid:48)(n) = +1. n(cid:55)→0 Workers are risk neutral and endowed with the same ownership share of all (cid:12)rms. They die with likelihood 1 − β 2 (0,1) at the end of each period. In each period, workers are employed by exactly one employer. The maximum quantity x of labor a worker can deliver to their employer is drawn from X, a (cid:12)nite set. We refer to x 2 X as a worker’s productivity. The three models we study di(cid:11)er only in terms in the stochastic process that governs the evolution of each worker’s productivity. 17While we refer to these two sets of (cid:12)rms as industries throughout this section, our arguments apply immediately when comparingemployerswhose perceivedlikelihoodof survivaldi(cid:11)ers for any reason,including size or geographicallocation. 18Dunne et al. (1988) show that exit rates di(cid:11)er markedly across industries in the U.S. and that those di(cid:11)erences tend to persist over time. Our assumption that exit rates are exogenously (cid:12)xed is strong, but it can be relaxed without altering our results. 14

5.1 Firm-specific human capital Our (cid:12)rst model is a two-industry version of Jovanovic’s (1979a) (cid:12)rm-speci(cid:12)c human capital model.19 Let X = f0,1,2g. In the (cid:12)rst period of employment with a given (cid:12)rm, x = 1 with probability 1. The evolution of a worker’s productivity in subsequent periods depends on the time s she devotes to training. With probability h(s), x moves up to 2, and with probability δ(s), x falls to 0, where δ is decreasing in s while h is increasing and strictly concave in s with h(0) > 0 and h(cid:48)(1) = 0. For simplicity, we assume that x = 0 and x = 2 are absorbing states. At the beginning of any period, workers can choose to separate from their current employer and begin providing labor to a new employer. The process governing the evolution of x may be correlated across workers in a given subset of (cid:12)rms, but we assume no aggregate uncertainty. Speci(cid:12)cally, in each industry, a fraction h(s) of the workers of productivity x = 1 who devote time s to training see their productivity rise, while a fraction δ(s) of these workers see their productivity fall. Productivity is(cid:12)rmspeci(cid:12)c inthismodelsinceworkers ofproductivity x = 2whoseemployer dies revert to x = 1 when they (cid:12)nd a new employer. In practice, much of human capital is general in nature, or occupation speci(cid:12)c (e.g., Kambourov and Manovskii, 2002) rather than (cid:12)rmspeci(cid:12)c. We make theassumption that allhuman capitalis (cid:12)rmspeci(cid:12)c for convenience. The result we establish below only requires that human capital be (cid:12)rm speci(cid:12)c in part, or, in a model where all human capital is occupation speci(cid:12)c, that workers whose employer die run the risk of not (cid:12)nding employment in the same occupation. We assume that (cid:12)rms behave competitively and pay workers the value of their marginal product in each period.20 We consider steady state equilibria, i.e. equilibria in which the price w of labor (the wage rate per e(cid:14)ciency unit of labor) in industry i 2 fH,Lg is i constant. A steady state equilibrium is a pair of wage rates such that labor markets clear in both industries in all periods. Quintin and Stevens (2003) show that a unique steady state equilibrium exists in this environment provided that βh(1) < δ(1). Under this assumption, all (cid:12)rms must hire new workers in every period in steady state. Therefore, wages adjust so 19A more general version of this model is analyzed in detail in Quintin and Stevens (2003). 20Di(cid:11)erent bargaining power assumptions would lead to di(cid:11)erent equilibria, but would not change our basic result: all else equal, workers in (cid:12)rms more likely to survive devote more time to human capital accumulation. 15

thatnewworkersareindi(cid:11)erent betweenthetwoindustries. Thisalsoimpliesthatincumbent workers at productivity level x = 1 are indi(cid:11)erent between quitting or staying with their (cid:12)rm.21 We assume, for concreteness, that they stay. With this convention, workers quit if, and only if, their productivity level falls to x = 0. Denote by si be the fraction of time workers with x = 1 devote to training in industry i 2 fH,Lg.22 Let Si(t) denote the fraction of workers of tenure t who separate from their employer in industry i in any given period. By tenure we mean the number of periods a worker has worked for their current employer at the beginning of a given period. In particular, all workers have tenure 0 in their (cid:12)rst period of employment, and workers of tenure 0 do not separate from their employer by construction. We can now state: Proposition 1. A unique steady state equilibrium exists. Furthermore, in steady state, 1. w < w , H L 2. sL (cid:20) sH with a strict inequality if and only if sH > 0, 3. Si(t) < Si(t+1) for all t (cid:21) 0 and i 2 fH,Lg, 4. SH(t) (cid:20) SL(t) for all t > 0, with a strict inequality if and only if sH > 0. The third item of the proposition says that this model, like all reasonable models of turnover, predictsanegativecorrelationbetweentenureandseparationrates. Thisisbecause workers with more tenure tend to have more (cid:12)rm-speci(cid:12)c human capital. But the model also predicts that when sH > 0, which occurs for instance when lim h(cid:48)(s) = +1 , workers in s(cid:55)→0 the high-survival industry should accumulate more (cid:12)rm-speci(cid:12)c capital than workers in lowsurvivalindustries. Thatisbecauseworkersemployed in(cid:12)rmsmorelikely tosurvivearemore likely to reap the bene(cid:12)ts from their investment in (cid:12)rm-speci(cid:12)c human capital. However, formalizing this simple intuition is complicated by general equilibrium considerations: the wage rate is higher in (cid:12)rms less likely to survive (a compensating di(cid:11)erential) which makes theopportunitycostofhumancapitalinvestment higher inthose(cid:12)rms. Theproofweprovide 21In this model, as in Jovanovic’s (1979a) model, all separations are initiated by the worker. Di(cid:11)erent rent-sharing arrangements would lead to di(cid:11)erent outcomes (see Jovanovic,1979a for a discussion.) 22Workers with x=2 devote no time to training because, by assumption, that state is absorbing. 16

in appendix A consists of arguing that this general equilibrium e(cid:11)ect is dominated by the direct impact of survival rates on returns to human capital investments. Importantly, the fact that the wage rate (the price of e(cid:11)ective labor) is higher does not imply that earnings are higher in (cid:12)rms more likely to fail. Earnings per period depend on workers’ productivity and the time they spend in training. Because they devote more time to human capital accumulation, workers in (cid:12)rms more likely to survive are more productive. When workers do not discount future consumption flows, it is in fact easy to argue that average earnings must be the same in the two industries. With discounting, average earnings comparisons depend on parameter selections.23 Note also that the compensating di(cid:11)erential stems solely from general equilibrium considerations. Whether human capital is speci(cid:12)c or not is irrelevant for this feature: the three models we consider all predict it. If we strictly interpret s as time devoted to training, then we can, in principle, test the endogenous view of (cid:12)rm-speci(cid:12)c capital by comparing training intensity across industries. However, in our view, the state of existing evidence on training, particularly informal training, does not allow for a convincing test along those lines (see Barron, 1997). But the last item of proposition 1 can be tested, as we have shown in this paper. Models where (cid:12)rmspeci(cid:12)c capital is in part endogenous correctly predict that turnover rates should be higher in high survival industries, even at equal employee tenure. 5.2 Learning-by-doing In the second model (as in the learning-by doing model of Jovanovic and Njarko, 1995) the accumulation of (cid:12)rm-speci(cid:12)c capital is exogenous. It di(cid:11)ers from the (cid:12)rm-speci(cid:12)c human capital model in one respect only: function h satis(cid:12)es h(s) = h(0) > 0 for all s 2 [0,1]. As a result of this assumption, workers always choose sH = sL = 0, and the average rate of growth of productivity with tenure is exogenous, and the same in the two industries. Proposition 1 continues to hold, except that separation rates are now the same at each tenure level in the two industries. Remark 1. If h(s) = h(0) 8s (cid:21) 0, then, in steady state, SH(t) = SL(t) 8t > 0. 23See Quintin and Stevens (2003) for more details. 17

Although separation rates are the same across industries at each tenure level, average separation rates are higher in the low-survival industry than in the high-survival industry even if sH = sL = 0. That’s because workers have more tenure on average in industries with high survival rates, and are therefore more productive on average as h(0) > 0. Unlike in the (cid:12)rm-speci(cid:12)c capital model, the correlation between exit and separation rates disappears in this model after controlling for tenure. Remark1impliesthatmakingthelearning-by-doingmodelconsistentwiththecorrelation betweenseparationandexitratesatequaltenurerequiresintroducingemployerheterogeneity beyond di(cid:11)ering survival probabilities. It su(cid:14)ces, for instance, to assume that the learning function (h) is steeper in the high-survival industry. 5.3 Matching Inmatching models (e.g. Jovanovic, 1979b), aworker’s productivity depends onhowwell she ismatchedwithheremployer. Shelearnsabouthermatchqualityovertime,andaseparation occurs when expected match quality falls below a certain threshold. Our goal in this section is to highlight the features matching models need in order to imply that separation rates should be higher in low-survival industries among observably similar workers. Assume that X = fx ,x g with x < x and that a worker-employer match is either 1 2 1 2 good, or bad. In any given period, well-matched employees draw productivity x = x with 2 probability p , while poorly matched employees draw x = x with probability p < p . G 2 B G Match quality is unknown to both the employer and the employee, but they know that in the (cid:12)rst period of employment, half the matches are good while the other matches are bad. Productivity draws are independent across periods, and workers update their prior probability of being in a good match according to Bayes’ rule as they observe more draws. The optimal separation rule in this environment is simple: a worker separates from their employer at the beginning of period t given her productivity history if the corresponding prior probability that she is in a good match falls below 1/2. Indeed, in that event, their expected future income falls below what they would expect if they started a new job. Since the separation rule is the same across industries, the distribution of beliefs is also the same in the two industries at each tenure level. It now follows that: 18

Remark 2. In steady state, SH(t) = SL(t) for all t > 0. As in the learning-by-doing model, separation rates are the same in the two industries at each tenure level. This is because in this most simple of matching models all workers have the same outside option, namely separate from their current employer and start a new job where the likelihood of a good match is 1/2. One way to make separation thresholds di(cid:11)er across industries is to introduce unemployment risk. Intuitively, this could make the opportunity cost of separating from one’s employer lower in low-survival industries as jobs are more likely to end for exogenous reasons in those industries. On the other hand, workers in low-survival industries would forgomore current-period income by separating asthey have higher wage rates. In an appendix available upon request we show that at least for certain parameter values, matching models with unemployment predict that separation rates should be higher in high-survival industries. This suggests that for matching models to predict a positive correlation between separation rates and (cid:12)rm death rates, features such as on-the-job search or heterogenous matching processes across employer are necessary. As in the (cid:12)rm-speci(cid:12)c human capital model, and for similar reasons, returns to on-the-job search are related to (cid:12)rm survival. Intuitively, time devoted to on-the-job search is more likely to pay o(cid:11) in establishments less likely to survive. On the other hand, as in the (cid:12)rm-speci(cid:12)c human capital model, general equilibrium considerations imply that the wage rate is higher in establishments less likely to survive, which makes the opportunity cost of time devoted to on-the-job search higher. Establishing that the direct e(cid:11)ect of exit rates on on-the-job search dominates may be di(cid:14)cult. But the discussion above shows that without such a feature, matching models make ambiguous predictions, at best, for turnover rates across industries. 6 Conclusion Using a French data set with matched employer-employee data, this paper provides new evidence that employee turnover is higher in industries where (cid:12)rms or establishments are more likely to exit. We (cid:12)nd that tenure alone cannot explain the higher separation rates in low-survival industries. Models of turnover founded on (cid:12)rm-speci(cid:12)c human capital models 19

are strongly consistent with this empirical regularity. Other standard models of turnover require additional assumptions. Our empirical result also has implications for the literature on industry rents. Speci(cid:12)cally, it suggests that industry rents could reflect di(cid:11)erences in job security across industries. Workers may accumulate less human capital in low survival industries, than their counterparts in industries where jobs are more secure. 20

A Proof of Proposition 1 Existence and uniqueness can be established with standard arguments (see Quintin and Stevens, 2003). In order to prove the other items of the proposition, we will (cid:12)rst describe the problem solved by workers in equilibrium. Let V (x) denote the expected income of i a worker of productivity level x in industry i 2 fH,Lg. In steady state, new workers are indi(cid:11)erent between workinginthetwoindustries, sowemust haveV (1) = V (1). Therefore, H L the expected income of a worker of productivity x in industry i 2 fH,Lg can be written as V (x) = max(1−s)xw +βfV (1)+p [(1−h(s))V (x)+h(s)V (2)]g (A.1) i i i i i i s∈[0,1] where (1 − s)xw is current period earnings if a fraction s of time is devoted to training. i Because the worker can change employers if necessary, her productivity is at least 1 in the next period. If her employer survives, which occurs with likelihood p , and if her training i e(cid:11)orts pays o(cid:11), which occurs with likelihood h(s), she moves up to productivity x = 2. Now assume by way of contradiction that w > w in steady state. Then, since the H L expected income associated with any training policy is strictly higher in industry H, we have V (1) > V (1). In that case, no worker would take a job in industry L which can’t be H L in steady state since labor demand is positive at all wages in both industries. Consider now the second item in the proposition. Given (A.1), workers in industry i 2 f1,2g choose a level si of training such that w (cid:20) h(cid:48) (si )βp [V (2)−V (1)] i i i i with a strict equality if and only if si > 0. (Recall that h(cid:48)(1) = 0.) We only need to argue that p [V (2) − V (1)] > p [V (2) − V (1)]. Assume that this inequality does not hold. H H H L L L Then, because w > w , inspection of (A.1) shows that V (1) > V (1), an inequality which L H L H cannot hold in equilibrium as no worker would join (cid:12)rms of industry H. This result yields the second item of the proposition. Consider (cid:12)nallythelast two itemsoftheproposition. Inindustryi 2 f1,2g,theevolution ofaworker’sproductivitywithtenurefollowsaMarkovchainwithstatesf1,2gandtransition Pi(x = 2jx = 1) = h(si). Because h(0) > 0 the chain’s expected value rises strictly with tenureinbothindustries, whichisitem3oftheproposition. Inaddition,sH (cid:21) sL impliesthat the Markov process governing a worker’s productivity in industry H (cid:12)rst order stochastically dominates the process governing the evolution of a worker’s productivity in industry L. This yields the last item of the proposition, and completes the proof. 21

B Description of the data B.1 Sample restrictions Our main source of data is the \Postes Exhaustif" version of the 2002 France’s D(cid:19)eclaration Annuelle des Donn(cid:19)ees Sociales. The database contains 32,951,110 observations. Below is a list of our sampling restrictions, with the size of the sample shown in parenthesis after each step: Drop observation if employee’s age is 55 years of more (30,933,580) Drop observation if employee’s age is 25 years or less (22,599,270) Drop observation if job status is not full-time (15,404,185) Drop observation if employer is in \Operational Services" industry (14,384,621) Drop observation if employer is an association (13,469,154) Drop observation if occupation is intern or apprentice (13,344,781) Drop observation if occupation is clergy (13,344,371) Drop observation if industry is recreative activities, or missing (13,068,663) These restrictions produce the \all-employees sample" which we use to estimate the (cid:12)rst probit speci(cid:12)cation of table 3. In the other probit speci(cid:12)cation, as well as the baseline bivariate probit speci(cid:12)cation, we further restrict the sample to employees hired during the current calendar year. This sample of \current-year hires" has 2,970,323 observations. 22

B.2 Definition of variables The following table de(cid:12)nes the variables we use in our analysis. When appropriate, we express in parenthesis each variable’s de(cid:12)ning criteria in terms of the variable names used by INSEE. Occupation category 5 and the smallest size category (1-19 employees) were the excluded dummy variables. Age Employee age in years; Male Employee gender (1 if male, 0 if female); Occupation 1 1 if occupation is CS=1 (e.g., heads of (cid:12)rms, store-owners and artisans), 0 otherwise; Occupation 2 1 if occupation is CS=2 (e.g., top-level managers and high-level professionals), 0 otherwise; Occupation 3 1 if occupation is CS=3 (e.g., intermediate-level managers, technicians and foremen), 0 otherwise; Occupation 4 1 if occupation is CS=4 (e.g., other employees, with the exception of factory and farm workers), 0 otherwise; Occupation 5 1 if occupation is CS=5 (e.g., factory and farm workers), 0 otherwise; Size, 1-19 1 if establishment size has 1 to 19 employees (TREFF=2), 0 otherwise; Size, 20-49 1 if establishment size has 20 to 49 employees (TREFF=3), 0 otherwise; Size, 50-99 1 if establishment size has 50 to 99 employees (TREFF=4), 0 otherwise; Size, 100-199 1 if establishment size has 100 to 199 employees (TREFF=5), 0 otherwise; Size, 200+ 1 if establishment size is 200 employees or more (TREFF=6), 0 otherwise; Private 1 if employer is in private sector (DOMEMPL=1 or 9), 0 otherwise; Paris agglomeration 1 if region of employment is Paris and its area (REG=01), 0 otherwise; Other urban 1 if department of employment had more than 1 million inhabitants in 2002 (DEP 2f06,13,31,33,38,44,57,59,68,76g),0 otherwise; Manufacturing 1 if industry is manufacturing and mining (NES5=ET), 0 otherwise; Services 1 if industry is services (NES5=EX), 0 otherwise; Trade 1 if industry is distribution (NES5=EW), 0 otherwise; Exit rate Fraction of employees in each industry (NES36) whose employer reports size 0 (TREFF=00) on December 31, 2002; Alive 1 if employer reports positive size on December 31 (TREFF>00), 0 otherwise; Tenure 1 if employee is a returning employee (DEBREMU=1), 0 otherwise; Separation 1 if job terminated before 12/31/02(FINREMU<360) and if Alive=1, 0 if job did not terminate before 12/31/02(FINREMU=360) and Alive=1, missing if Alive=0; 23

B.3 Construction of industry exit rates with OECD data Our second proxy for exit rates by industry in France comes from the OECD’s \Firmlevel study."24 The French section of the data is constructed using the Fiscal database (‘BRN’ (cid:12)le) and the Enterprise survey (‘EAE’ (cid:12)le). The unit of reference is the (cid:12)rm. For years 1990-96, the database covers all sectors and all (cid:12)rms whose sales exceed FF3.8mn in manufacturing and FF1.1mn in the service sector. Firms are generally classi(cid:12)ed by 2digit ISIC category, although some categories are merged. Our DADS database classi(cid:12)es establishments using France’s NES36 categories. To make the two classi(cid:12)cations compatible, some NES36 categories must be merged. We then use the following mapping: NES36 categories 2-digit ISIC categories A0 01, 02, 03 , 04, 05 B0 15 ,16 C1, F2 17, 18 , 19 C2, F3 20, 22 C3, F4 24, 25 C4, E2, E3, F5, F6 28, 29, 33 D0 34 E1 35 F1, G1 10, 11, 12, 13, 14, 23, 26 G2 40, 41 H0 45 H1, J1, J2, J3 50, 51, 52 J0, K0 65, 67 L0 65, 66, 67 M0 70 N1 64 N4 73 P1 55 Q1 80 Q2 85 R1 75 In each industry, (cid:12)rms present at the beginning of a given year are split in two categories: continuing (cid:12)rms((cid:12)rmsalso present intheregister thefollowingyear) andexiting (cid:12)rms(other (cid:12)rms). Firms that enter the industry in a given year are classi(cid:12)ed as one-year (cid:12)rms ((cid:12)rms that exit the industry before the end of the year) or entrants. For t 2 f1990,...1996g, we de(cid:12)ne empstock as the number of employees in continuing (cid:12)rms, exit (cid:12)rms and one year t (cid:12)rms in year t, and empexit as the number of employees in one-year (cid:12)rms and exiting (cid:12)rms. t For each of the 21 categories shown in the table above, we calculate the average fraction of employees in exiting (cid:12)rms in year as : 1 (cid:88)1996 empexit OECD exit rate = t . 7 empstock t t=1990 24A detailed description of the study is available on the OECD website at http://www.oecd.org/dataoecd/22/7/2767629.htm. 24

References Abowd, John M., Kramarz, Francis , and Margolis, David N. \High Wage Workers and High Wage Firms." Econometrica 67 (March 1999): 251-333. Abowd, John M., Patrick Corbel, and Francis Kramarz. \The Entry and Exit of Workers and the Growth of Employment: An Analysis of French Establishments." NBER Working Paper Series, #5551 (April 1996). Anderson,PatriciaM.,andMeyer,BruceD.\TheExtentandConsequencesofJobTurnover." Brookings Papers on Economic Activity: Microeconomics (1994): 177-236. Barron, John M., Berger, Mark C., and Black, Dan A. \How Well Do We Measure Training?" Journal of Labor Economics 15 (July 1997): 507-28. Bartelsman, Eric, Stefano Scarpetta, and Fabiano Schivardi. \Comparative Analysis of Firm Demographics and Survival: Micro-level Evidence for the OECD Countries." OECD Economics Department Working Paper No. 348 (January 2003). Becker, Gary, S. \Investment in Human Capital: A Theoretical Analysis." Journal of Political Economy 70 (1962): 9-49. Dunne, Timothy, Mark J. Roberts, and Samuelson, Larry. \Patterns of Firm Entry and Exit in U.S. Manufacturing Industries." The RAND Journal of Economics 19 (Winter 1988): 495-515. Farber, Henry S. \Mobility and Stability: The Dynamics of Job Change in Labor Market." In Handbook of Labor Economics, Volume 3b, edited by Orley Ashenfelter and David Card, 1999, 2439-83. Greene, William H. \Econometric Analysis." 3rd edition, Prentice Hall (1997). Greene, William H. \Sample Selection in Credit-Scoring Models." Japan and the World Economy 10 (1998): 299-316. Institut National de la Statistique et des Etudes Economiques (INSEE), D(cid:19)eclaration Annuelle des Donn(cid:19)ees Sociales, Fichier d(cid:19)etail exhaustif national (version postes) (2002), Paris, France. Institut National de la Statistique et des Etudes Economiques (INSEE), \Tableaux de l’Economie Fran(cid:24)caise (2000/2001)", INSEE, Paris, France. Jovanovic, Boyan. \Firm-speci(cid:12)cCapitalandTurnover." The Journal of Political Economy 87 (December 1979a): 1246-1260. Jovanovic, Boyan. \Job Matching and the Theory of Turnover." The Journal of Political Economy 87 (October 1979b): 972-990. Jovanovic, Boyan and Yaw Nyarko. \The Transfer of Human Capital." Journal of Economic Dynamics and Control 19 (September 1995): 1033-64. 25

Kambourov, Gueorgui, and Manovskii, Iourii. \Occupational Speci(cid:12)city of Human Capital." mimeo (2002). Keane, Michael P. \A Note on Identi(cid:12)cation in the Multinomial Probit Model." Journal of Business and Economic Statistics 10 (April 1992): 193-200. Meng, Chun-Lo, and Schmitt, Peter. \On the Cost of Partial Observability in the Bivariate Probit Model." International Economic Review 26 (1985): 71-85. Mortensen, Dale, T., and Pissarides, Christopher, A. \Job Creation and Job Destruction in the Theory of Unemployment." Review of Economic Studies 61 (1994) 397-415. Nagyp(cid:19)al, E(cid:19)va. \Learning-by-Doing Versus Learning About Match Quality: Can We Tell Them Apart?" mimeo (2002). Oi, W., Y. \Labor as a Quasi-Fixed Factor." Journal of Political Economy 70 (1962): 538-55. Poirier, Dale J. \Partial Observability in Bivariate Probit Models." Journal of Econometrics 12 (1980): 209-17. Quintin, Erwan, andStevens, JohnJ.\GrowingOldTogether: FirmSurvivalandEmployee Turnover." mimeo (2003). 26

Table 1: Sample Means for All Employees Occupation Category Exit Separation Obs. Age Tenure Male 1 2 3 4 rate rate All-employees sample 13,068,663 39.034 0.773 0.637 0.017 0.134 0.249 0.265 0.051 0.178 By industry Manufacturing 3,585,564 39.658 0.815 0.732 0.011 0.121 0.218 0.087 0.041 0.152 Services 6,197,297 39.161 0.778 0.533 0.012 0.169 0.284 0.378 0.045 0.175 Trade 2,239,950 37.699 0.714 0.631 0.035 0.100 0.261 0.338 0.084 0.218 Other 1,045,854 38.997 0.718 0.940 0.028 0.043 0.125 0.048 0.057 0.207 By geographical location Paris agglomeration 3,394,233 38.560 0.731 0.600 0.017 0.247 0.286 0.264 0.053 0.213 Other urban 2,953,830 39.150 0.780 0.664 0.016 0.113 0.254 0.264 0.051 0.168 Non-urban 6,720,600 39.22 0.790 0.644 0.017 0.086 0.229 0.266 0.051 0.165 By establishment size 1-19 employees 4,297,673 38.137 0.677 0.666 0.043 0.089 0.215 0.277 0.069 0.281 20-49 employees 1,811,327 38.399 0.744 0.684 0.012 0.110 0.231 0.228 0.057 0.205 50-99 employees 1,204,828 38.745 0.773 0.657 0.005 0.133 0.240 0.225 0.053 0.174 100-199employees 1,253,367 38.979 0.794 0.630 0.003 0.147 0.246 0.223 0.050 0.153 200+ employees 4,501,470 40.238 0.869 0.587 0.001 0.183 0.292 0.291 0.032 0.092 By sector Public 2,480,995 40.891 0.910 0.489 0.000 0.118 0.308 0.472 0.020 0.073 Private 10,587,668 38.598 0.740 0.672 0.021 0.138 0.235 0.217 0.059 0.204 Note. Tenure, male, and occupation categories are dummy variables. Table 2: Sample Means for Current-Year Hires Occupation Category Exit Separation Obs. Age Male 1 2 3 4 rate rate Current-year hires 2,970,323 36.220 0.639 0.011 0.137 0.215 0.278 0.062 0.404 By industry Manufacturing 661,566 36.910 0.681 0.006 0.135 0.183 0.097 0.043 0.379 Services 1,373,109 35.809 0.571 0.010 0.183 0.242 0.366 0.060 0.422 Trade 640,448 35.880 0.601 0.020 0.088 0.244 0.386 0.087 0.398 Other 295,202 37.326 0.947 0.012 0.031 0.097 0.042 0.057 0.393 By geographical location Paris agglomeration 910,769 36.079 0.622 0.011 0.248 0.256 0.275 0.062 0.379 Other urban 648,655 36.292 0.665 0.011 0.106 0.218 0.273 0.061 0.396 Non-urban 1,410,889 36.279 0.640 0.012 0.079 0.186 0.283 0.062 0.425 By establishment size 1-19 employees 1,386,096 36.580 0.658 0.022 0.082 0.193 0.313 0.073 0.491 20-49 employees 463,936 36.124 0.671 0.005 0.111 0.217 0.247 0.061 0.407 50-99 employees 273,549 35.972 0.642 0.003 0.143 0.226 0.238 0.057 0.368 100-199employees 258,209 35.848 0.618 0.002 0.166 0.228 0.234 0.055 0.350 200+ employees 588,535 35.728 0.580 0.001 0.269 0.252 0.257 0.042 0.271 By sector Public 222,120 35.839 0.458 0.000 0.206 0.275 0.457 0.022 0.318 Private 2,748,203 36.251 0.654 0.012 0.131 0.210 0.264 0.065 0.412 Note. Male and occupation categories are dummy variables.

Table 3: Estimated Coe(cid:14)cients in Probit Models of Worker Separations All employees Current-year hires Intercept 1.147 0.279 Age/100 -6.781 -2.112 (Age/100)2 6.215 2.508 Male 0.004 -0.073 Occupation 1 -0.573 -0.834 Occupation 2 0.075 -0.376 Occupation 3 -0.053 -0.325 Occupation 4 -0.007 -0.016 Size, 20-49 -0.207 -0.192 Size, 50-99 -0.298 -0.286 Size, 100-199 -0.371 -0.330 Size, 200+ -0.546 -0.521 Private 0.184 -0.076 Paris agglomeration 0.156 -0.013 Other urban 0.015 -0.041 Manufacturing 0.093 0.211 Service 0.131 0.255 Trade -0.007 0.026 Exit rate 2.005 3.022 Tenure dummy -0.789 n.a. No. observations 12,446,132 2,744,502 Notes. See the data appendix for a discussion of these variables. All coe(cid:14)cients are statistically signi(cid:12)cant at conventional levels as a result of the large sample size.

Table 4: The E(cid:11)ect of Exit Rates on Separation Rates Mean & standard deviation Relative e(cid:11)ects Marginal Predicted Exit Number of %-change Model e(cid:11)ect Separation rate rate std. dev.’s from mean Probit 1. All employees 0.447 0.184 (0.110) 0.051 (0.035) 0.14 8.5 2. Current-year hires 1.115 0.411 (0.119) 0.062 (0.036) 0.34 9.8 Bivariate probit 3. All current-year hires 1.153 0.411 (0.119) 0.062 (0.036) 0.35 10.1 4. Exclusion restrictions 1.126 0.411 (0.119) 0.062 (0.036) 0.34 9.9 5. Alternative exit rates 3.354 0.458 (0.123) 0.054 (0.016) 0.43 11.7 6. Worker restrictions 0.520 0.159 (0.059) 0.058 (0.033) 0.29 10.8 7. Industry restrictions 1.575 0.454 (0.131) 0.067 (0.034) 0.41 11.8 Notes. The marginal-e(cid:11)ect column shows the marginal e(cid:11)ect of exit rates on predicted separation rates. For the bivariate-probit models, this column shows the marginal e(cid:11)ect of exit rates on predicted separation rates conditional on the establishment surviving. The (cid:12)rst of the two relative-e(cid:11)ects columns shows the number of standard deviations by which predicted separation rates increase for a 1 standard deviation increase in exit rates, and the second shows the percent change in predicted separation rates due to a 1 standard deviation increase in exit rates.

Table 5: Estimated Coe(cid:14)cients in Bivariate Probit Models of Worker Separations (1) (2) (3) (4) (5) Baseline Exclusion Alternative Non-seasonal Industry Model Restrictions Exit Rates Workers Restrictions Separation equation Intercept 0.159 0.159 -0.337 -0.692 -0.492 Age/100 -1.642 -1.641 -2.181 -2.066 -1.246 (Age/100)2 2.110 2.108 2.659 1.922 1.507 Male -0.072 -0.072 -0.090 -0.074 -0.057 Occupation 1 -0.511 -0.511 -0.414 -0.336 -0.544 Occupation 2 -0.353 -0.353 -0.235 -0.232 -0.412 Occupation 3 -0.304 -0.304 -0.252 -0.179 -0.328 Occupation 4 -0.005 -0.005 0.059 -0.005 -0.024 Size, 20-49 -0.177 -0.177 -0.198 -0.152 -0.179 Size, 50-99 -0.264 -0.263 -0.282 -0.234 -0.271 Size, 100-199 -0.305 -0.304 -0.320 -0.222 -0.313 Size, 200+ -0.477 -0.477 -0.499 -0.319 -0.482 Private -0.022 -0.022 -0.068 0.166 0.482 Paris agglomeration -0.059 -0.059 -0.047 -0.040 -0.011 Other urban -0.057 -0.057 0.189 -0.076 -0.051 Manufacturing 0.178 0.178 0.238 0.106 0.221 Service 0.241 0.241 0.464 0.246 0.196 Trade 0.039 0.039 0.315 0.017 0.020 Exit rate 4.010 4.014 9.038 2.669 5.134 Establishment survival equation Intercept 2.630 2.728 2.726 2.910 2.713 Age/100 -1.839 -1.827 -2.206 -2.549 -2.033 (Age/100)2 1.343 1.333 1.764 2.444 1.770 Male 0.030 0.029 0.050 0.054 0.022 Occupation 1 -0.612 -0.617 -0.840 -0.380 -0.629 Occupation 2 0.016 0.014 0.006 0.148 0.020 Occupation 3 0.027 0.025 -0.065 0.071 0.014 Occupation 4 -0.043 -0.051 -0.222 -0.032 -0.056 Private -0.410 -0.398 -0.727 -0.631 -0.462 Paris agglomeration 0.255 0.255 0.305 0.103 0.242 Other urban 0.097 0.098 0.116 0.091 0.091 Manufacturing 0.119 0.151 0.339 0.182 0.113 Service -0.029 -0.017 0.064 -0.022 -0.015 Trade -0.046 -0.029 n.a. 0.189 -0.037 Exit rate -6.001 -6.388 -3.220 -4.311 -6.128 Correlation between errors ρ -0.805 -0.805 -0.725 -0.363 -0.763 (cid:15)u No. observations 2,970,323 2,970,323 2,590,387 488,874 2,639,413 Notes. See the data appendix for a discussion of these variables. All coe(cid:14)cients are statistically signi(cid:12)cant at conventional levels as a result of the large sample size. The model in column 2 also included the share of surviving establishments in the four size categories 20-49, 50-99, 100-199, 200+; the coe(cid:14)cients were -0.300, 0.025, -0.559, and -0.037 respectively.

Figure 1: Industry exit rates and separation rates 0.5 0.4 0.3 0.2 0.1 0 0.05 0.1 0.15 0.2 0.25 Exit rate etar noitarapeS Source: D(cid:19)eclarationAnnuelle desDonn(cid:19)eesSociales, \PostesExhaustif" version, andauthors’ calculations. Industries are de(cid:12)ned using France’s NES36 classi(cid:12)cation. Separation and exit rates are de(cid:12)ned in appendix B.1. Importantly, separation rates exclude terminations due to establishments exiting the industry.