Job hopping, earnings dynamics, and industrial agglomeration in the software publishing industry

Journal of Urban Economics 64 (2008) 590–600

Contents lists available at ScienceDirect

Journal of Urban Economics www.elsevier.com/locate/jue

Job hopping, earnings dynamics, and industrial agglomeration in the software publishing industry Matthew L. Freedman ∗ Cornell University, ILR School, Ithaca, NY 14850, USA

a r t i c l e

i n f o

a b s t r a c t

Article history: Received 13 February 2008 Revised 12 June 2008 Available online 16 July 2008

This paper investigates the implications of industrial clustering for labor mobility and earnings dynamics in one large and increasingly important high-technology sector. Taking advantage of longitudinal employee-employer matched data, I exploit establishment-level variation in agglomeration to explore how clustering in the software publishing industry affects labor market outcomes. The results show that clustering makes it easier for workers to job hop within the sector. Higher earnings levels in more agglomerated areas are partly attributable to sorting across locations among workers and firms in the industry on the basis of observable and unobservable characteristics. Controlling for this heterogeneity, workers in clusters have relatively steep earnings-tenure profiles, accepting lower wages early in their careers in exchange for stronger earnings growth and higher wages later. These findings are consistent with theoretical models in which agglomeration improves labor market coordination and facilitates greater learning and human capital formation. © 2008 Elsevier Inc. All rights reserved.

JEL classification: R12 J60 J31 Keywords: Agglomeration Clustering Labor mobility Wage level and structure

1. Introduction Over a quarter of the nation’s workers in the software publishing industry are located in one state, and nearly a third of that state’s software publishing workers are employed in a single county.1 Though one of the most prominent examples of industrial clustering, software is not the only sector in which it occurs; evidence suggests that firms in a number of industries, from automobile manufacturing to biotechnology, concentrate in particular locations to an extent over and above what we would expect given the distribution of economic activity more generally (Porter, 1990; Krugman, 1991; Kim, 1995; Ellison and Glaeser, 1997, 1999; Holmes and Stevens, 2004). While a large literature addresses the potential sources of agglomeration economies, only a small number of studies investigate how geographic clustering by firms in particular industries interacts with local labor market dynamics. Using longitudinal employee-employer matched data, this paper examines the nature and extent of industrial clustering and explores the relationship between agglomeration among establishments and labor mobility,

*

Address for correspondence: Cornell University, ILR School, 359 Ives Hall East, Ithaca, NY, USA. E-mail address: [email protected]. 1 Author’s calculations based on publicly available County Business Patterns data for 2005. 0094-1190/$ – see front matter doi:10.1016/j.jue.2008.07.002

© 2008 Elsevier Inc.

All rights reserved.

earnings levels, and earnings growth rates in one large and dynamic high-technology sector. The empirical analysis reveals that clustering among establishments in the software publishing industry is associated with shorter job durations and greater job-hopping among individuals within the sector. Starting and average wages are higher in clusters, but this is due in part to sorting among workers and firms across locations. After addressing biases arising from self-selection, the results indicate that, relative to those employed by more isolated firms, workers in more clustered firms have steeper earningstenure profiles, accepting lower salaries at the start of their careers in exchange for stronger earnings growth and higher salaries later. These findings suggest that, in understanding the relationship between clustering and labor market outcomes, we must look beyond traditional models of agglomeration that predict that clustering should affect wage levels, but not necessarily wage growth rates or labor mobility. To the extent that workers exhibit more job mobility in clusters and that a fraction of the wage premium that arises in clusters is the result of a wage growth effect as opposed to a level effect, the results are consistent with theoretical models in which clustering improves labor market coordination and fosters greater learning and human capital formation. The paper proceeds as follows. The next section reviews the literature on agglomeration and its implications for labor market outcomes. Section 3 describes the data, discusses the methodology I employ to measure clustering, and presents basic descriptive statistics. Section 4 turns to the empirical analysis and discusses

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

the results in light of theoretical models with predictions for the impact of industrial clustering on local labor market dynamics. Section 5 concludes. 2. Literature A substantial body of evidence suggests there is a large and persistent urban wage premium (Glaeser and Maré, 2001; Rosenthal and Strange, in press). For many of the same reasons we might expect wage levels to be higher in cities, we might also expect wage levels to be higher in industrial clusters (Duranton and Puga, 2004; Rosenthal and Strange, 2004). Firms that cluster with others in the same industry may enjoy higher worker productivity due to greater local demand or input-output linkages (Krugman, 1991; Ciccone and Hall, 1996). Alternatively, information externalities or knowledge spillovers that increase the productivity of firms could contribute to higher wages within clusters, as Lucas (1988) and Rauch (1993) have argued might occur more generally in cities. It could also simply be that industrial clusters attract relatively highquality firms or high-ability workers, the former perhaps more capable of capitalizing on agglomeration’s benefits and the latter possibly deriving more utility from other local amenities (Combes et al., 2008). In any of these cases, we would expect wage levels, but not necessarily wage growth rates or job mobility, to be higher in industrial clusters. On the other hand, if clustering improves labor market coordination or promotes greater human capital accumulation among workers, wage growth rates as well as job mobility could be higher within clusters. For example, if agglomeration reduces job search frictions, not only could working in an industry cluster induce more job hopping as workers and firms seek out better matches, but it also might reduce risks associated with industry-specific human capital investment. These lower risks to investment might motivate workers to specialize, which in turn could lead to stronger wage growth (Becker and Murphy, 1992; Rotemberg and Saloner, 2000). Also, if workers in clusters more readily exchange information regarding production techniques, investment in human capital might be less expensive, which could give rise to a relatively greater amount of human capital accumulation and steeper wage-tenure profiles. Just as Glaeser (1999) and Glaeser and Maré (2001) suggest occurs in cities, industrial clustering could speed the rate of interactions and thus facilitate more rapid learning. Hence, while some theories suggest that wages among workers in clusters should be uniformly higher than wages among those outside clusters, others imply that higher wages may only come with time. Initial wage levels among workers starting their careers in clusters might even be lower than those among workers starting outside clusters if clustering is conducive to on-the-job search and promotes more aggressive bidding among firms over skilled labor. In this case, agglomeration could induce greater job mobility as well as steepen wage-tenure profiles, with workers in clusters accepting lower wages initially in anticipation of higher wages later in their careers owing to their ability to claim a greater share of the economic rents from competing firms (Burdett and Mortensen, 1998; Postel-Vinay and Robin, 2002a, 2002b). Together with differences in the production processes and organizational structures of individual firms as well as variation in the costs of doing business across locations (such as congestion and rents), balancing the downside of heightened competition over skilled labor against any benefits from agglomeration could help to explain why some firms might choose to locate inside and outside clusters in equilibrium (Holmes, 1999; Combes and Duranton, 2006). Thus, studying patterns of job and income mobility across different locations over time can shed light on the nature of any interaction between clustering and labor market dynamics. Al-

591

though a few studies have investigated job mobility patterns in high-technology industries in the context of possible knowledge spillovers (Saxenian, 1994; Almeida and Kogut, 1999; Fallick et al., 2006), the broad ramifications of clustering in particular industries for earnings dynamics are not well explored, in large part due to a lack of appropriate data. Using a large, longitudinal employeeemployer matched dataset, this paper documents job mobility and earnings dynamics among workers at clustered and dispersed software firms and considers these patterns in light of models with different implications for clustering and labor market outcomes. 3. Data 3.1. Sources To capture job and income mobility of individuals as they move within and between firms over time, I require a data set that combines information about workers and their employers and that permits me to track each over a long period. Due to incomplete information about individuals’ employment and earnings histories, small sample sizes, and reporting problems, traditional survey data render it difficult to measure job mobility or to evaluate the temporal pattern of earnings changes among workers (Bound et al., 2001). I study the relationship between industrial agglomeration and local labor market dynamics using an employee-employer matched data set constructed and maintained by the U.S. Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) Program. LEHD integrates quarterly administrative earnings information for workers derived from U.S. state unemployment insurance records with internal Census Bureau censuses and surveys.2 The result is a database that is particularly well suited to examining job mobility and earnings dynamics and that provides an opportunity to explore how clustering interacts with local labor markets more extensively than have past studies. LEHD data boast several advantages over household and business based survey data. The data are current and relatively accurate because businesses face financial penalties for misreporting their workers’ employment and earnings information. Since the scope of the quarterly longitudinal data is nearly the full universe of firms and workers, I can follow individuals over time as they move across the earnings distribution and across employers. Additionally, the integrated records contain information on workers’ demographic characteristics, including date of birth, race, sex, and education. Though sparse relative to the information on individuals in surveys such as the Current Population Survey and Panel Study of Income Dynamics, the worker characteristics on the LEHD data permit some flexibility in investigating variation across demographic groups and serve as important controls in the empirical analysis. Critically for this study, LEHD data also contain a detailed industry classification code (six-digit NAICS) and a unique address, including latitude/longitude coordinates, for nearly all establishments. The LEHD data have several limitations. First, the data are not available for all U.S. states, and the amount of historical data varies by state.3 Second, coverage is limited for workers and firms in some sectors, including agriculture, non-profits, and public administration.4 Third, LEHD data lack information on hours worked,

2 More extensive descriptions of LEHD data appear in Abowd et al. (2006) and Haltiwanger et al. (2007). 3 As of early 2008, 48 states (including the District of Columbia) are participating in the LEHD Program. This is an ongoing project, and additional states are expected to join. 4 See Stevens (2002) for a more detailed description of the LEHD database coverage issues.

592

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

making it impossible to calculate an hourly wage rate and requiring one to infer full-time employment status (as I discuss in greater detail below). Finally, business identifiers in the LEHD data are State Employer Identification Numbers (SEINs), which are used for state tax collection purposes and are potentially more aggregated business entities than establishments. While this aggregation requires the imputation of some measures of workforce composition and earnings for the small number of establishments that are part of larger multi-unit operations,5 it is nevertheless possible to pinpoint individual establishments within multi-unit SEINs geographically without resorting to imputation using LEHD data.6 3.2. Sample In this paper, I focus on establishments and workers in the software publishing industry (NAICS 5112).7 This segment of the high-technology sector is a natural candidate for studying the relationship between labor market dynamics and agglomeration for several reasons. Products in the software industry are generally sold in national or international markets, minimizing the importance of product market considerations in driving firms’ location decisions (U.S. Government Accountability Office, 2006). Further, proximity to natural resources such as bodies of water is relatively unimportant, as is access to upstream suppliers of capital goods. Meanwhile, innovation in the industry over the past decade has been rapid and human capital investment is important in the sector. The potential for accelerated human capital formation and knowledge spillovers in clusters coupled with the heightened competition over skilled labor that comes with agglomeration make the software industry ripe for research on how job and income dynamics interact with geographic clustering. For this study, I use data for one large U.S. state (its name withheld for confidentiality reasons) for the third quarter of 1991 through the third quarter of 2003. I selected the sample state based on its size (in terms of both population and geography), the relatively long time span of its data, and the quality of the geographic coding of its establishments.8 In addition, the sociodemographic characteristics of workers in this state are generally representative of the U.S. working population as a whole. I extract from the statewide data the complete employment and earnings histories of all individuals observed to work at least one full quarter in software publishing, where being full quarter employed at

5 Fewer than 10% of establishments in the sample are part of multi-unit operations, though close to one-fourth of workers in the sample are employed in establishments that are part of multi-unit operations. 6 While SEINs are potentially more aggregated business entities than establishments, LEHD data provide breakouts of establishments for multi-unit SEINs, which are termed SEIN units. Only for a subset of multi-unit SEINs does LEHD have information on precisely which individuals are employed at each SEIN unit, though the geographic location of each SEIN unit and its total employment are known. When the unit of work is unknown for a particular worker attached to a multi-unit SEIN, LEHD imputes that workplace based on the worker’s place of residence and the distribution of employment across establishments within the SEIN. See Abowd et al. (2006) for details on the imputation procedure. 7 The Census Bureau defines NAICS 5112 as consisting of “establishments primarily engaged in computer software publishing or publishing and reproduction. Establishments in this industry carry out operations necessary for producing and distributing computer software, such as designing, providing documentation, assisting in installation, and providing support services to software purchasers. These establishments may design, develop, and publish, or publish only.” See http://www.census.gov/naics for details. 8 Geographic coding in my sample state, while better than in many other states, is not perfect. Roughly 88% of sample establishments have rooftop-confident geographic coordinates (latitude and longitude), the most precise coordinates possible. About 95% of establishments have coordinates that are accurate at least at the Census tract level, and 98% have coordinates that are accurate at least at the county level. Throughout the analysis, I use the sample of establishments for which we have accurate rooftop address information so as to minimize measurement error in the clustering variables.

time t requires a worker have positive earnings at a given establishment in periods t − 1, t, and t + 1. This largely eliminates from the sample workers employed only part of a quarter, and hence whose reported earnings represent compensation for an indeterminate amount of time (anywhere from one to 90 days). Over 2400 unique software establishments, 153,000 software workers, and 170,000 software jobs (i.e., worker-firm matches) appear in the data over the entire sample period. 3.3. Descriptive statistics In this section, I present basic descriptive statistics on the nature and extent of agglomeration as well as on firm and workforce characteristics in the software publishing industry. I use a modified version of the conventional location quotient (LQ) measure to examine the extent of clustering at the establishment level.9 The LQ is a measure of an industry’s level of concentration in a particular location over and above what one would expect in light of the general spatial distribution of economic activity. Typically, the LQ is computed as the ratio of an industry’s share of total establishments or employment in a “local” area, such as a county, relative to its share of total establishments or employment in a larger area, such as the nation. As a refinement on the conventional measure, I take advantage of the rich geographic information in the LEHD data and construct establishment-specific LQs by drawing concentric circles with radii of five, ten, 25, and 50 miles around each establishment in the sample, and computing the industry’s share of total establishments or total employment within those rings relative to its share of total establishments or total employment in the state. For establishment j in industry k, the LQ for a circle with radius r is LQ r , jk = ( E r , jk / E r , j )/( E k / E ) where E r , jk is the number of establishments or employment within the circle of radius r around establishment j in industry k (excluding establishment j), E r , j is the number of establishments or employment in all industries within the circle of radius r around establishment j, E k is the number of establishments or employment in the entire state in industry k, and E is the number of establishments or employment in the state across all industries. Values of the LQ exceeding one reflect higher than average concentration at a particular location; values less than one indicate less than average concentration. This methodology provides for each establishment a single measure of agglomeration that reflects the extent to which that establishment is more or less clustered than is typical for all businesses. For example, a software establishment with a large number of other software establishments or workers nearby will not have a LQ greater than one unless it is the case that software is overrepresented in the immediate vicinity compared to economic activity more generally. Importantly, and in contrast with the more traditional LQ measure, the measure I construct can be applied at a variety of different spatial scales that need not be dictated by administrative boundaries.10 Table 1 presents clustering statistics for software establishments in the sample over different distances for selected years. Establishment-specific LQs, whether measured using neighboring

9 Holmes and Stevens (2002) develop a similar establishment-specific measure of industrial clustering to explore the relationship between firm size and agglomeration. 10 I calculate clustering statistics for the sample state alone, so the value of the LQ for each establishment approaches unity as the radius rises to include the entire state. While the measure I construct respects state boundaries, it does not depend on any administrative boundaries within the state and can be applied more generally to a more geographically expansive dataset.

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

Table 1 Establishment-specific LQs for the software publishing industry in selected periods, means and standard deviations (a) Establishment-based LQs

Number of establishments 5-Mile radius (LQ5 ) 10-Mile radius (LQ10 ) 25-Mile radius (LQ25 ) 50-Mile radius (LQ50 )

1992Q2

1998Q2

2003Q2

822 Mean

1020 Mean

SD

882 Mean

SD

1.66 1.41 1.04 0.75

2.07 1.90 1.60 1.40

1.61 1.35 0.97 0.69

SD

2.29 1.95 2.15 2.02 1.49 1.98 1.53 0.89 1.67 1.38 0.65 1.47 (b) Establishment-based LQs 1992Q2

1998Q2

2003Q2

Number of establishments

822 Mean

SD

1020 Mean

SD

882 Mean

SD

5-Mile radius (LQ5 ) 10-Mile radius (LQ10 ) 25-Mile radius (LQ25 ) 50-Mile radius (LQ50 )

2.22 2.03 1.63 1.46

1.98 1.71 1.24 0.89

2.52 2.23 1.93 1.66

4.19 2.81 1.59 1.16

2.92 2.51 2.06 1.63

5.09 3.58 1.83 1.18

593

employment-based LQ (again measured with a radius of 25 miles) greater or less than one on average while we observe them in the data.11 Consistent with the findings of Holmes and Stevens (2002), larger establishments tend to be more clustered, though the distributions of establishment size are similar across different areas. Annualized earnings (in 1997 dollars, deflated by the U.S. consumer price index), including starting earnings (i.e., annualized earnings in the first full quarter at a job) and ending earnings (i.e., annualized earnings in the last full quarter), are higher in clustered establishments, as is within-job wage growth.12 These figures do not control for other worker characteristics, which also vary to some degree across clustered and dispersed establishments. For example, workers in clustered establishments tend to be slightly younger and more highly educated, and they are more likely to be non-white. In the empirical analysis that follows, I examine whether job mobility and earnings dynamics among workers at more and less clustered establishments vary after controlling for these characteristics as well as other, unobservable heterogeneity. 4. Empirical analysis

Table 2 Establishment-specific LQs for selected industries, 2003Q2 Employment-based LQ25 Industry (NAICS)

Mean

SD

Nonmetallic Mineral Mining (2123) Apparel Knitting Mills (3131) Aerospace Product and Parts Manufacturing (3364) Communications Equipment Manufacturing (3342) Scientific Research and Development Services (5417) Steel Product Manufacturing (3312) Pharmaceutical and Medicine Manufacturing (3254) Radio and Television Broadcasting (5151) Securities & Commodity Contracts Brokerage (5231) Department Stores (4521) Junior Colleges (6112)

3.18 1.82 1.66 1.65 1.55 1.34 1.28 1.18 1.16 1.06 0.83

22.64 0.98 1.03 1.34 1.23 1.38 0.76 0.69 1.04 0.46 0.39

establishments (panel (a)) or employment (panel (b)), are on average well above one in the software industry for circles with radii of up to at least 50 miles. While average LQs have declined modestly over time on an establishment basis in the software industry, employment-weighted LQs rose on average between the early 1990s and early 2000s. In other words, employment in the industry has grown increasingly concentrated, though establishments have not. Compared to other four-digit NAICS industries, economic activity in the software sector is quite clustered taking into account to the broader spatial distribution of economic activity. Table 2 presents employment-based LQs using a 25-mile radius for several industries in the second quarter of 2003. I chose some of these industries, such as nonmetallic mineral mining (NAICS 2123) and department stores (NAICS 4521), based on their similarity to the broader set of industries within their general industry category, while I selected others, such as apparel knitting mills (NAICS 3131) and pharmaceutical and medicine manufacturing (NAICS 3254), because they have served as the subject of prior research into clustering. Although there is a clear tendency toward clustering in the software industry, it is not the case that all activity in the sector is concentrated in a few locations. In fact, at a 25-mile radius, only 59% of sample establishments had an establishment-based LQ greater than one in 2003; for the employment-based measure, this figure was 50%. I exploit this variation in the extent of clustering across establishments as well as changes over time in clustering within establishments (as other businesses enter and exit nearby) to identify relationships between agglomeration and labor market outcomes in the next section. Table 3 presents basic descriptive statistics for establishments and jobs in the sample, broken out by whether they have an

In this section, I turn to an empirical analysis of how establishment clustering interacts with job mobility and earnings patterns in the software publishing industry. First, I examine the relationship between industrial clustering and job mobility, and in particular job durations and job-hopping. Next, I evaluate earnings patterns in the industry, investigating the extent to which worker and firm sorting affects wages and exploring how starting salaries and subsequent earnings growth vary with the extent of clustering. I consider the results in light of different models with predictions for the implications of clustering for local labor market dynamics. I find evidence of heightened job mobility as well as steeper wage-tenure profiles in agglomerated areas, evidence consistent with improved labor market coordination and greater learning and knowledge accumulation in clusters.13 4.1. Job mobility I begin my analysis of job mobility by identifying individual separations from software industry jobs, which I do by comparing worker and establishment matches in consecutive quarters in the data. A separation in quarter t occurs when a worker is employed

11 The descriptive statistics as well as the empirical results are similar using LQs with radii of up to 50 miles, and in the analysis that follows, I continue to rely on a LQ defined at 25 miles. In the sample, 88% of workers who job-hop within the industry transition to a new establishment that is fewer than 25 miles away from their old establishment. By comparison, 38% job hop to a new establishment that is fewer than five miles away from their old establishment, and 62% hop to one that is fewer than ten miles away. 12 Average end-of-spell earnings and within-job wage growth could be inflated if workers leaving their jobs receive taxable severance pay or exercise stock options. Using earnings one year before separation instead of end-of-spell earnings in calculating average annual earnings growth, the means for clustered and non-clustered establishments are uniformly lower (in the case of within job wage growth, by 1–2 percentage points). I take this as evidence that, while non-salary forms compensation may be important in the industry, clustered and non-clustered establishments do not differ substantially in the degree to which they use such compensation. 13 The LQ that I use to measure industrial clustering is a relative measure of density. To check whether absolute density might also matter, I experimented in each regression in the empirical analysis with replacing the LQ with the number of software workers within 25 miles. The coefficients on this absolute measure of density (as well as on its higher order terms and interactions) had the same signs as the LQ, but were generally less significant. The impact of measurement error might be more pronounced using absolute measures. If an establishment is located near a state border, for example, the absolute measure could understate nearby software employment. However, the LQ would mitigate the problem by using as a denominator measured overall employment nearby, which would also be affected by the boundary.

594

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

Table 3 Descriptive statistics for establishments and jobs All

Clustered (LQ25 > 1)

Non-clustered (LQ25  1)

Number of establishments Average employment Share with <5 employees Share with 5–9 employees Share with 10–19 employees Share with 20–49 employees Share with 50+ employees

2414 27 0.50 0.15 0.12 0.12 0.11

1337 32 0.48 0.15 0.12 0.13 0.13

1077 20 0.53 0.16 0.13 0.11 0.08

Number of jobs Average starting earnings ($1997) Average ending earnings ($1997) Average annual earnings growth Average spell length (quarters) Average starting age Share male Share white Share education <12 years Share education 12–15 years Share education 16+ years

170,196 63,929 89,823 17% 9 34 0.63 0.65 0.08 0.46 0.46

113,515 68,899 102,693 18% 9 34 0.63 0.63 0.08 0.45 0.47

56,681 53,974 64,049 13% 9 35 0.63 0.69 0.09 0.48 0.43

Notes. Includes left-censored job spells. Average annual earnings growth calculated as the percentage change in earnings on the job divided by years on the job.

for a full quarter at an establishment in period t but not in period t + 1. The quarterly separation rate in the data averages 7%, slightly lower than estimates for the broader economy of close to 11% (Bjelland et al., 2007), but not surprising given the age, gender, and educational profile of workers in software. The restriction that individuals in the sample be employed for a full quarter at each employer also eliminates workers with very transient spells in the industry.14 For job spells that are not right censored, I can distinguish whether workers in the software publishing industry move to new establishments in the same industry, establishments in different industries, or out of the sample. The majority of separated workers in the sample ultimately transition into new jobs as opposed to out of the sample, though most take positions at establishments outside the software industry. Of the 130,127 separations from software establishments that I observe, about 63% ultimately result in a transition to a job outside the software industry, while about 9% result in a transition to another software job.15 The remainder of the separations result in exit from the sample, which could be due to a right-censored spell of unemployment, withdrawal from the labor force, or a move to another state.16 I first estimate the probability of exit from a software industry job in a given quarter as a function of worker and establishment characteristics. I assume the hazard is proportional and of the following form:

where X is a vector of worker characteristics, Z is a vector of establishment characteristics, and β and Φ are vectors of parameters to be estimated. The baseline hazard λ0 (t ) is allowed to be nonparametric and is estimated using the Cox partial likelihood technique (Kalbfleish and Prentice, 1980). I use the Peto-Breslow

approximation to deal with ties, which occur when more than one job spell ends in a given period.17 I include in the vector X individual characteristics including quadratics in end-of-spell age and in log real annualized earnings (averaged over the course the spell), as well as dummies for gender, race, and education. In Z , I include dummies for establishment size (based on average employment over the course of the job spell) and the geographic variables of interest. In particular, I include a quadratic in the establishment’s employment-based LQ (calculated using a 25-mile radius and averaged across the course of the spell) as well as a quadratic in a measure of urban density.18 The latter, which is calculated as the average over all quarters of the job spell in the share of total sample employment within a circle with a radius of 25 miles around each establishment, is aimed at capturing the extent to which the immediate vicinity is built up. I also include in the regression time dummies (corresponding to the last quarter of the spell) to control for the influence of cyclical macroeconomic factors, while I include establishment county dummies to control for unmeasured, time-invariant differences across counties such as variation in infrastructure quality, taxes, zoning restrictions, and industry composition. The sample consists of all software job spells that are not left censored, and thus for which we have accurate job tenure information up to the point of a job separation (of which there approximately 111,000) or right censoring (41,000). Since individuals can have more than one job in the data and since the observations for each person are likely not independent over time, I correct the standard errors by clustering on worker; that is, the error term is assumed to be independent across workers but not necessarily within workers over time.19 I present estimated coefficients from the hazard-rate regression for separations of any type in column 1 of Table 4. The results reveal that, controlling for other worker and establishment characteristics, those individuals working at more clustered software

14 Firm exit could be responsible for observed worker separations. However, the worker separation rate is over twice the rate of firm exit in NAICS 5112. The average quarterly firm exit rate in the sample is 3.2%. 15 The destination job for a move is defined as the first subsequent full-quarter job. Of the workers who leave software publishing for another industry, 59% stay within NAICS 5 (28% of all private-sector workers in the U.S. economy are employed in NAICS 5, according to the 2002 Economic Census), 16% go to NAICS 4 (22%), 14% to NAICS 3 (13%), and 4% to NAICS 6 (14%). The remaining 7% of separating workers transition to jobs in other industries. 16 The data do not permit me to distinguish between involuntary separations and voluntary separations (i.e., between layoffs and quits).

17 The global test statistic based on scaled Schoenfeld residuals suggests that the data do not violate the proportional-hazards assumption (χ 2 = 39.26, p = 0.9964). 18 Using LQs defined using radii of up to 50 miles yields similar results. 19 Clustering standard errors at the individual level allows for arbitrary correlation between errors within individuals, but imposes zero correlation between errors across workers. As a robustness check, I experimented with clustering at more aggregate levels (e.g., the establishment or county level); doing so tends to decrease the significance primarily of the urban density variable’s estimated coefficient, leaving largely intact the significance of other estimated coefficients in the regression (and, in fact, raising the significance of the estimated effects of most of the demographic control variables).

λ(t ; X , Z ) = λ0 (t ) exp( X  β + Z  Φ),

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

595

Table 4 Competing risks model (1) Exit LQ25 LQ25 Squared Urban Density25 Urban Density25 Squared Log Annualized Earnings Log Annualized Earnings Squared Employment 5–10 Employment 10–50 Employment >50

(2) Exit to software job

(3) Exit to other industry

(4) Exit to nonemployment

−0.1207***

0.4451***

−0.2112***

−0.1344***

(0.0202)

(0.0825)

(0.0265)

(0.0451)

0.0307***

−0.0644***

0.0486***

0.0297***

(0.0033)

(0.0138)

(0.0044)

(0.0074)

2.6848***

7.2455***

1.4317***

4.6032***

(0.3208)

(1.3918)

(0.4191)

(0.7014)

−8.5820***

−15.0087***

−5.7380***

−14.2048***

(0.9470)

(3.6011)

(1.2310)

(2.0531)

−0.3116***

2.5629***

−0.2799***

−0.5904***

(0.0445)

(0.4216)

(0.0550)

(0.0775)

0.0138***

−0.1022***

0.0106***

0.0262***

(0.0021)

(0.0192)

(0.0027)

(0.0037)

−0.0062

−0.0243

0.0246

−0.0591

(0.0187)

(0.0979)

(0.0238)

(0.0381)

−0.0431***

0.3361***

0.0045

−0.2205***

(0.0154)

(0.0773)

(0.0198)

(0.0311)

−0.1058***

0.5101***

−0.0716***

−0.3336***

(0.0151)

(0.0761)

(0.0196)

(0.0302)

−0.0499***

−0.0322***

−0.0191***

−0.0951***

(0.0019)

(0.0077)

(0.0026)

(0.0039)

Age Squared

0.0005***

0.0001

0.0001*

0.0011***

(0.0000)

(0.0001)

(0.0000)

(0.0000)

Male

0.0040

0.1150***

0.0233***

−0.0803***

(0.0056)

(0.0229)

(0.0073)

(0.0119)

−0.0071

−0.0146

−0.0505***

0.0962***

(0.0057)

(0.0225)

(0.0073)

(0.0122)

Age (years)

White Some college

−0.0276***

−0.0732*

−0.0665***

0.0675***

(0.0096)

(0.0385)

(0.0123)

(0.0213)

−0.0309***

−0.0967**

−0.0971***

0.1303***

(0.0100)

(0.0396)

(0.0128)

(0.0220)

Spells

151,896

151,896

151,896

151,896

Exits

111,053

9795

71,103

30,155

College degree

Notes. Excluded categories include establishment employment <5, female, non-white, and <12 years of education. Additional controls include county dummies and year and quarter dummies. Standard errors in parentheses allow for arbitrary correlation over time within the same person. * ** ***

Significant at 10%. Significant at 5%. Significant at 1%.

establishments tend to have modestly longer tenures overall. A 10% increase in the LQ from its mean is associated with an approximately 1% decline in the hazard rate. That compares to the 2% decrease in the hazard rate associated with a 10% increase in earnings, and the 2% increase associated with a 10% increase in urban density. While gender and race have little bearing on the hazard, each additional year of age decreases it by about 5%. Firm size also figures heavily in the likelihood of separation, with workers at larger software firms substantially less likely to separate in any given period than workers at smaller firms. While informative, the results in column 1 of Table 4 do not distinguish between different types of separations and constrain the baseline hazard to be the same for all software jobs. Of greater interest is whether separating workers in more clustered establishments show a greater propensity to remain within the industry and how hazards might differ across different possible destination states. If agglomeration improves labor market coordination within the sector or encourages greater investment in industry-specific human capital, we would expect clustering to be associated with greater job hopping within the sector. I investigate this in a competing risks framework, which allows us to evaluate how clustering is related to the time to exit from a software industry job to another job in the same industry, to a job in a different industry, or

to non-employment.20 The observed duration of a software industry job spell in the data can end in one of these destinations or be right censored. I present in columns 2 through 4 in Table 4 results for each of the three exit types.21 The extent to which a worker’s establishment is clustered clearly has different implications for the likelihood of each type of exit. A higher LQ is associated with a greater likelihood that a worker will leave his current software job for another software job, while it is associated with a lower likelihood of leaving for a job in a different industry or leaving the sample. Indeed, a 10% increase in the LQ from its mean is associated with an over 7% increase in the hazard for transitioning into a new software job and a 2% decrease in the hazard for transitioning into a job in a different industry. I plot in Fig. 1 estimated survival functions for transitions within the industry for LQs one standard deviation above and below the mean in the sample. The functions are based on the re-

20 See Kalbfleish and Prentice (1980) and Lancaster (1990) for discussions of competing risks models. 21 Results breaking exit types into categories that include having spans of intervening non-employment between jobs produced similar results for the exit types of interest and yielded few additional insights beyond those presented.

596

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

Fig. 1. Survival functions based on results from a Cox proportional hazards model for workers transitioning within the software industry. The graphs assume 35 year-old white, male workers with 16+ years of education earning $60,000, mean urban density (0.14), and an establishment with 10–49 employees.

gression results in column 2 of Table 4 and assume a 35-year old white, male worker with a college degree earning $60,000 and employed at an establishment with 10–49 employees that is located in an area with average urban density (0.14). The graph shows a distinctly lower survival rate in more clustered establishments controlling for other worker and firm characteristics. After five years, the likelihood that a worker at a more clustered establishment will leave for another software job is about five percentage points greater than that for an observationally equivalent worker earning the same amount at a less clustered, but otherwise similar establishment. The difference grows to ten percentage points after ten years. As the results in Table 4 suggest, taking into account firm and worker heterogeneity along dimensions such as establishment size as well as worker gender and educational attainment is important in determining how clustering is related to the probability of job hopping within the industry. Indeed, in regressions that only include time and county dummy variables (results not shown), clustering appears to have a much smaller impact on the likelihood of transitioning between jobs within the industry; in that case, the effect of changes in the LQ on the hazard for moving to a new software job upon separating are only about half the magnitude as in the full model. Thus, understanding the relationship between clustering and job hopping requires taking sorting along certain observable dimensions into consideration. Notably, as Table 4 reveals, while a higher wage at a software job is associated with a lower likelihood of transitioning to other industries or out of the sample, the opposite is true for transitions within the industry, indicating perhaps that accumulated industry-specific knowledge (reflected in higher earnings) might make workers more attractive to other software firms. Also, being at a larger firm is associated with a higher hazard rate for transitioning into jobs in the same industry, but has little relationship with hazard for transitioning into other industries. This pattern would be consistent with larger firms offering more training opportunities, thereby facilitating more industry-specific human capital investment, or with jobs at larger firms serving as springboards for talented software workers who can take advantage of their ex-

perience (or trade secrets to the extent that non-compete clauses are not used or enforced) to go on to acquire better positions. Meanwhile, in line with expectations, older individuals and those with more education are less likely to change jobs, all else equal. The results in this section lend some preliminary support to models that suggest that clustering within industries may reduce job search frictions and encourage greater industry-specific human capital investment. A closer examination of earnings dynamics among workers in the software industry can shed additional light on the extent to which different theories of agglomeration can help us understand how clustering interacts with labor market outcomes. 4.2. Earnings dynamics The previous results highlighted the importance of clustering in facilitating job hopping within the software industry. This section builds upon those results first by shedding light on the extent of sorting across locations among workers and firms in the software industry, and then by examining how wage-tenure profiles differ among workers at clustered versus isolated firms controlling for observable and unobservable heterogeneity. As Table 3 suggests, some of the differences in earnings levels and growth rates across locations may be attributable to differences in observable worker and firm characteristics, such as educational attainment and firm size. However, even controlling for these characteristics, there still may be sorting among workers and firms along unobservable dimensions, which could in turn have implications for wages in more or less clustered areas. Exploiting the longitudinal structure of the data and using job fixed effects, I can evaluate the potential impact of such sorting on observed wage patterns across locations and, in turn, control for it in order to isolate clustering’s relationship with earnings levels and growth rates. In particular, I run panel regressions with and without a jobspecific earnings intercept, γi j . The general model is specified as follows: ln(earningsi jt ) = X it β + Z jt Φ + M i jt Ψ + γi j + εi jt ,

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

where, in the fixed-effect specification, ln(earningsi jt ) is the natural log of real annualized earnings for worker i in establishment j at time t, X it includes a cubic in tenure for worker i at t, Z jt includes four establishment size dummies as well as quadratics in the LQ and urban density variables for firm j at t, and M i jt is a vector that includes interactions between the tenure variables and the geographic variables for i and j at t. For comparison purposes, I run an ordinary least squares (OLS) regression in which I constrain γi j to be the same for all jobs (γi j = γ ), and in which X it and Z jt also include observable time-invariant characteristics that would otherwise be subsumed by the job fixed effect (including worker i’s gender, race, educational attainment, and age as well as establishment j’s county). In both the OLS and fixed-effect specifications, I include year and quarter dummies to eliminate bias due to any correlation between clustering and earnings resulting from shocks that vary over time but that are constant across jobs. Since the observations for each person are likely not independent over time, and since individuals can have multiple jobs in the data, I correct standard errors by clustering on each worker. Comparing OLS and fixed-effect results will provide some indication of the extent to which biases in estimated relationships between agglomeration and earnings might arise in crosssectional analyses because workers or firms self-select into clusters. If more innately talented individuals select into clustered establishments, estimates of the earnings effects of clustering would be biased upward. Using job fixed effects, which are defined for each unique worker-establishment match, also mitigates potential problems stemming from firms’ endogenous location decisions. To the extent that unobservable but time-invariant factors, such as firms’ organizational structures, rent-sharing arrangements, or particular product lines, might affect firms’ location decisions, job fixed effects will resolve the endogeneity problem associated with firm location choice.22 Table 5 presents the results of the OLS and job fixed-effect regressions. Since the coefficients on higher order polynomials are difficult to interpret, I also graph in Fig. 2 estimated earningstenure profiles based on the fixed-effect results for workers in establishments with LQs one standard deviation above and below the mean in the sample, but that are otherwise similar. In both the OLS and job fixed-effect regressions, earnings are increasing at a diminishing rate in tenure, as expected. However, the OLS and job fixed-effect regression results differ with respect to the estimated coefficients on the LQ, which capture the immediate effect of starting a job in more clustered software establishment (when tenure equals zero). The OLS results show a positive but diminishing relationship between clustering and earnings, while the fixed-effect results show a negative but diminishing relationship. Critically, in the latter specification, the source of identification is changes over time in clustering that occur in the external environment of a software establishment as businesses enter and exit and employment expands and contracts after an employment relationship is formed. That is, we are identifying the relationship between clustering and earnings based on variation over time in the LQ within jobs, and thereby eliminating observable and unobservable time-invariant factors associated with each job that might otherwise bias estimates. Comparing the OLS and fixed-effect results, then, the job effects are picking up some unmeasured characteristics of workers and

22 While the job fixed effects help to address the traditional ability bias, bias still could be present if selection is on the basis of learning capacity. If it is the case that better learners self-select into clusters, wage-tenure profiles would be steeper in more agglomerated areas even in the absence of any competitive effects or knowledge externalities. Addressing this issue would require finding an instrument correlated with an individual’s propensity to obtain a job in a cluster but not related to his wage growth, a challenging task that is the subject of future work.

597

firms that otherwise translate into higher starting salaries in more clustered establishments. In other words, sorting on the basis of unobserved characteristics among workers and firms contributes to the premium in starting salaries observed in clusters. In principle, a job fixed effect could also capture some notion of match quality. To the extent that greater clustering enhances the ability of workers to find suitable jobs and improves the ability of firms to find workers with needed skills, it could improve the quality of matches and thereby raise starting salaries. This effect would not be consistent with the finding that clustering is associated with greater job hopping; if workers searching for jobs or firms looking for employees can find good matches more quickly in clusters, we would expect the likelihood of separation for another job in the industry to be lower in more agglomerated areas (Wheeler, 2008).23 Still, one could argue that improved match quality in the spirit of Jovanovic (1979) or Helsley and Strange (1990) could be responsible for overall differences in earnings levels. To get a sense of the importance of match effects relative to sorting effects, I estimate a regression that includes separate worker and firm fixed effects instead of job fixed effects (results not shown). Though the high dimensionality of the model necessitates the use of special algorithms in estimation, additively separable person and firm effects will not capture match-specific factors that might be correlated with clustering and independently affect earnings.24 The estimated coefficients for the variables of interest from this regression are nearly identical to those from the regression with job fixed effects; for example, the estimated coefficients on the LQ and its square are −0.0554 and 0.0074, respectively.25 To the extent that match effects bear on earnings, any impact they might have is dwarfed by sorting effects. Once we control for unobserved heterogeneity using job fixed effects, the results suggest that individuals in more clustered establishments accept lower initial wages relative to their counterparts in less clustered establishments. According to the results in column 2 of Table 5, a 10% increase in the LQ from its mean is associated with a roughly 1% lower starting salary. If workers who begin their careers in software clusters expect to be compensated with higher earnings later in their careers, either because of faster learning in clusters or a better outlook for future job opportunities, they might accept such a discount. Consistent with this idea, the initial discount workers in clusters accept eventually turns into a premium. As the coefficients on the interactions between the LQ and tenure reveal, and as the figure illustrates, earnings rise faster for those workers in more clustered establishments. Controlling for observable and unobservable heterogeneity among workers and firms, an individual in an establishment with a LQ one standard deviation above the mean is earning more than an individual in an establishment with a LQ one standard deviation below the mean after about five years on the job, with the gap growing steadily thereafter.

23 Non-compete clauses in employment contracts, which might limit the ability of workers solicit or accept offers from other firms once employed at a particular software firm, could raise reservation wages of workers searching for jobs. 24 Given the large sample size, estimating a wage equation with separate worker and firm fixed effects is computationally challenging. Therefore, in estimating the regression, I apply the conjugate gradient algorithm developed by Abowd et al. (2002). This methodology arranges the observations into connected groups of workers and firms, with each group comprising all the workers that any firm in the group ever employed and all the firms for which any worker in the group ever worked. This allows for identification of person effects and firm effects within groups, with all but one person effect and one firm effect identified for each group. After estimation, group means of person and firm effects are set equal to zero. 25 Andersson et al. (2007) examine the interplay between worker and firm fixed effects estimated separately in a single regression and draw implications for the sorting and matching of workers and firms across more and less urban areas; whether their conclusions regarding cities extend to industry clusters is the subject of future research.

598

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

Table 5 Earnings regressions

Tenure Tenure Squared Tenure Cubed

LQ25 LQ25 Squared LQ25 ×Tenure LQ25 ×Tenure Squared LQ25 ×Tenure Cubed LQ25 Squared×Tenure LQ25 Squared×Tenure Squared LQ25 Squared×Tenure Cubed

Urban Density25 Urban Density25 Squared Urban Density25 ×Tenure Urban Density25 ×Tenure Squared Urban Density25 ×Tenure Cubed Urban Density25 Squared×Tenure Urban Density25 Squared×Tenure Squared Urban Density25 Squared×Tenure Cubed

Employment 5–10 Employment 10–50 Employment >50

Age Age Squared Male White Education 12–15 years (0/1) Education 16+ years (0/1)

Constant County dummies Observations R-squared

(1) OLS

(2) Job fixed effects

0.0197*** (0.0029) −0.0006** (0.0002) 0.0000 (0.0000)

0.0105*** (0.0023) −0.0001 (0.0002) −0.0000 (0.0000)

0.0404*** (0.0126) −0.0023 (0.0021) 0.0077*** (0.0020) −0.0006*** (0.0002) 0.0000*** (0.0000) −0.0019*** (0.0004) 0.0001*** (0.0000) −0.0000*** (0.0000)

−0.0552*** (0.0188) 0.0061** (0.0030) 0.0102*** (0.0014) −0.0005*** (0.0001) 0.0000*** (0.0000) −0.0018*** (0.0003) 0.0001*** (0.0000) −0.0000*** (0.0000)

2.1348*** (0.2175) −5.8708*** (0.6377) 0.0155 (0.0448) −0.0022 (0.0036) 0.0001 (0.0001) −0.0903 (0.1455) 0.0116 (0.0117) −0.0003 (0.0002)

4.6332*** (0.8905) −8.7642*** (2.3779) −0.0573* (0.0297) 0.0021 (0.0022) −0.0000 (0.0000) 0.3078*** (0.0964) −0.0146** (0.0073) 0.0001 (0.0001)

0.1375*** (0.0142) 0.3230*** (0.0138) 0.4324*** (0.0137)

0.0743*** (0.0093) 0.1483*** (0.0110) 0.1740*** (0.0116)

0.1167*** (0.0020) −0.0013*** (0.0000) 0.3202*** (0.0037) 0.1073*** (0.0038) 0.1003*** (0.0069) 0.2561*** (0.0070) 7.2989*** (0.0446) Yes 1,522,806 0.31

10.1232*** (0.0757) N/A 1,522,806 0.81

Notes. Excluded categories include establishment employment <5 (for both regressions), female, non-white, and <12 years of education (for OLS). Additional controls include year and quarter dummies. Standard errors in parentheses allow for arbitrary correlation over time within the same person. * ** ***

Significant at 10%. Significant at 5%. Significant at 1%.

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

599

Fig. 2. Based on job fixed-effect regression with a cubic in tenure fully interacted with quadratics in the LQ and in urban density. Additional controls include four firm size dummies and year and quarter effects. The graphs assume mean urban density (0.14), an establishment with 10–49 employees, and business cycle conditions as in the second quarter of 1998.

Turning to the relationships between other covariates and earnings, the results suggest that, while wages also tend to grow faster for workers with jobs in more urban establishments, those jobs come with a significant initial wage premium as well. While the job fixed effects (and, to a more limited extent, county dummies in the OLS regression) absorb any time-invariant location characteristics and may capture some differences in costs of living and doing business across areas, the urban density variable may be picking up the influence of changes over time in costs associated with increasing density, such as changes in property rents. The estimated coefficients on the remaining covariates are in line with expectations. As the OLS results in column 1 of Table 5 show, earnings are higher for males, whites, older individuals, and those with higher educational attainment. Also consistent with past literature, larger establishments pay more on average than smaller establishments (Davis and Haltiwanger, 1996). However, the firm-size wage premium diminishes greatly once we include job fixed effects, pointing again to the importance of sorting across locations among establishments on the basis of unobservable characteristics. Without controlling for these time-invariant characteristics, some of their effects on earnings could be attributed to firm size alone. Thus, job fixed effects absorb the impacts of time invariant but unobservable worker and firm characteristics whose omission from the regression might otherwise cloud estimates of the effect of clustering on earnings patterns. As previously discussed, the fixed effects furthermore help to resolve selection and endogeneity problems that could arise in cross-sectional analyses of the relationship of agglomeration and earnings dynamics. After addressing these sources of bias and controlling for observed and unobserved worker and firm heterogeneity, what is left is the key result that establishment clustering in the software industry interacts positively with worker tenure, tilting the earnings-tenure profile in a way consistent with theories that emphasize agglomeration’s potential roles in reducing job search frictions and facilitating faster human capital accumulation.

5. Conclusion This paper takes advantage of new employee-employer matched micro-data from the U.S. Census Bureau’s Longitudinal EmployerHousehold Dynamics (LEHD) Program to examine the implications of industrial clustering for labor mobility and earnings dynamics. I find evidence that workers in clustered establishments in the software publishing industry are more likely to job hop to other software establishments. Furthermore, controlling for worker and establishment heterogeneity, workers in more clustered software firms have steeper earnings-tenure profiles than workers in more isolated establishments. Models of agglomeration that predict only a level effect of clustering on wages, such as those that emphasize demand externalities or input–output linkages, cannot fully account for the results. Moreover, while worker and firm sorting across locations is partly responsible for differences in labor market outcomes across locations, the findings indicate that sorting alone cannot entirely explain variation in earnings patterns and in the degree of job hopping in different areas. Rather, the results suggest that clustering in the software industry may smooth the functioning of the labor market and facilitate greater human capital formation. Acknowledgments This document reports the results of research and analysis undertaken by the U.S. Census Bureau staff. It has undergone a Census Bureau review more limited in scope than that given to official Census Bureau publications. This research is a part of the U.S. Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) Program, which is partially supported by the National Science Foundation Grants SES-9978093 and SES-0427889 to Cornell University (Cornell Institute for Social and Economic Research), the National Institute on Aging Grant R01∼AG018854-02, and the Alfred P. Sloan Foundation. The views expressed on statistical, methodological, and technical issues are those of the author(s) and not necessarily those of the U.S. Census Bureau, its program spon-

600

M.L. Freedman / Journal of Urban Economics 64 (2008) 590–600

sors, or its data providers. Some of the data used in this research are confidential data from the LEHD Program. The U.S. Census Bureau supports external researchers’ use of these data through the Research Data Centers (http://www.ces.census.gov). For questions regarding the data, please contact the U.S. Census Bureau, LEHD Program, Room 6H136C, 4600 Silver Hill Road, Suitland, MD 20746, USA (http://lehd.did.census.gov). The author gratefully acknowledges guidance and helpful comments from Dr. John Haltiwanger, Dr. John Shea, Dr. Michael Pries, Dr. Fredrik Andersson, and Dr. Stuart Rosenthal. The author also thanks staff at the U.S. Census Bureau’s LEHD Program and CES as well as anonymous referees for comments on earlier versions of this paper. All errors are the author’s. References Abowd, J., Creecy, R., Kramarz, F., 2002. Computing person and firm effects using linked longitudinal employer-employee data. LEHD Technical Paper TP-2002-06. U.S. Census Bureau. Abowd, J., Andersson, F., McKinney, K., Roemer, M., Stephens, B., Vilhuber, B., Woodcock, S., 2006. The LEHD infrastructure files and the creation of the Quarterly Workforce Indicators. LEHD Technical Paper TP-2006-01. U.S. Census Bureau. Almeida, P., Kogut, B., 1999. Localization of knowledge and the mobility of engineers in regional networks. Management Science 45 (7), 905–917. Andersson, F., Burgess, S., Lane, J., 2007. Cities, matching and the productivity gains from agglomeration. Journal of Urban Economics 61 (1), 112–128. Becker, G., Murphy, K., 1992. The division of labor, coordination costs, and knowledge. The Quarterly Journal of Economics 107 (4), 1137–1160. Bjelland, M., Fallick, B., Haltiwanger, J., McEntarfer, E., 2007. Employer-to-employer flows in the United States: Estimates using linked employer-employee data. Working Paper 2007-30. Federal Reserve Board. Bound, J., Brown, C., Mathiowetz, N., 2001. Measurement error in survey data. In: Heckman, J., Leamer, E. (Eds.), Handbook of Econometrics, vol. 5. North-Holland, Amsterdam. Burdett, K., Mortensen, D., 1998. Wage differentials, employer size and unemployment. International Economic Review 39 (2), 257–273. Ciccone, A., Hall, R., 1996. Productivity and the density of economic activity. American Economic Review 86 (1), 54–70. Combes, P., Duranton, G., 2006. Labor pooling, labor poaching, and spatial clustering. Regional Science and Urban Economics 36 (1), 1–28. Combes, P., Duranton, G., Gobillon, L., 2008. Spatial wage disparities: Sorting matters! Journal of Urban Economics 63 (2), 723–742. Davis, S., Haltiwanger, J., 1996. Job Creation and Destruction. MIT UP, Cambridge, MA. Duranton, G., Puga, D., 2004. Micro-foundations of urban agglomeration economies. In: Henderson, J.V., Thisse, J. (Eds.), Handbook of Regional and Urban Economics, vol. 4. North-Holland, Amsterdam. Ellison, G., Glaeser, E., 1997. Geographic concentration in U.S. manufacturing industries: A dartboard approach. Journal of Political Economy 105 (5), 889–927. Ellison, G., Glaeser, E., 1999. The geographic concentration of industry: Does natural advantage explain agglomeration? American Economic Review 89 (2), 311–316.

Fallick, B., Fleischman, C., Rebitzer, J., 2006. Job-hopping in silicon valley: Some evidence concerning the micro-foundations of a high-technology cluster. Review of Economics and Statistics 88 (3), 472–481. Glaeser, E., 1999. Learning in cities. Journal of Urban Economics 46 (2), 254–277. Glaeser, E., Maré, D., 2001. Cities and skills. Journal of Labor Economics 19 (2), 316– 342. Haltiwanger, J., Lane, J., Spletzer, J., 2007. Wages, productivity, and the dynamic interaction of businesses and workers. Labour Economics 14 (3), 575–602. Helsley, R., Strange, W., 1990. Matching and agglomeration economies in a system of cities. Regional Science and Urban Economics 20 (2), 189–212. Holmes, T., 1999. Localization of industry and vertical disintegration. Review of Economics and Statistics 81 (2), 314–325. Holmes, T., Stevens, J., 2002. Geographic concentration and establishment scale. Review of Economics and Statistics 84 (4), 682–690. Holmes, T., Stevens, J., 2004. Spatial distribution of economic activities in North America. In: Henderson, J.V., Thisse, J. (Eds.), Handbook of Regional and Urban Economics, vol. 4. North-Holland, Amsterdam. Jovanovic, B., 1979. Job matching and the theory of turnover. Journal of Political Economy 87 (5), 972–990. Kalbfleish, J., Prentice, R., 1980. The Statistical Analysis of Failure Time Data. Wiley, New York, NY. Kim, S., 1995. Expansion of markets and the geographic distribution of economic activities: The trends in U.S. regional manufacturing structure, 1860–1987. Quarterly Journal of Economics 110 (4), 881–908. Krugman, P., 1991. Increasing returns and economic geography. Journal of Political Economy 99 (3), 483–499. Lancaster, T., 1990. The Econometric Analysis of Transition Data. Cambridge UP, Cambridge, UK. Lucas, R., 1988. On the mechanics of economic development. Journal of Monetary Economics 22 (1), 3–42. Porter, M., 1990. The Competitive Advantage of Nations. Macmillan, New York. Postel-Vinay, F., Robin, J., 2002a. Equilibrium wage dispersion with worker and employer heterogeneity. Econometrica 70 (6), 2295–2350. Postel-Vinay, F., Robin, J., 2002b. The distribution of earnings in an equilibrium search model with state-dependent offers and counteroffers. International Economic Review 43 (4), 989–1016. Rauch, J., 1993. Productivity gains from geographic concentration of human capital: Evidence from the cities. Journal of Urban Economics 34 (3), 380–400. Rosenthal, S., Strange, W., 2004. The micro-empirics of agglomeration economies. In: Henderson, J.V., Thisse, J. (Eds.), Handbook of Regional and Urban Economics, vol. 4. North-Holland, Amsterdam. Rosenthal, S., Strange, W., in press. The attenuation of human capital spillovers. Journal of Urban Economics. Rotemberg, J., Saloner, G., 2000. Competition and human capital accumulation: A theory of interregional specialization and trade. Regional Science and Urban Economics 30 (4), 373–404. Saxenian, A., 1994. Regional Advantage: Culture and Competition in Silicon Valley and Route 128. Harvard UP, Cambridge, MA. Stevens, D., 2002. Employment that is not covered by state unemployment insurance laws. LEHD Technical Paper TP-2002-16. U.S. Census Bureau. U.S. Government Accountability Office, 2006. Offshoring: U.S. semiconductor and software industries increasingly produce in China and India. GAO-06-423. Wheeler, C., 2008. Local market scale and the pattern of job changes among young men. Regional Science and Urban Economics 38 (2), 101–118.