Abstract
Externalities are believed to drive the productivity benefits of cities, and also of dense sub-parts within cities, e.g. the central business district (CBD). Recent research claims that density externalities accrue mostly to non-routine activities, and that their effects, e.g. human capital spillovers, attenuate sharply with distance. Consistent with these claims, I demonstrate strong clustering tendencies in non-routine professions as evidenced by job-switching patterns, specifically switchers’ distances moved between employers. Individual-level geo-coded data for switchers within Sweden’s metropolitan areas are used to illustrate that employees hired to non-routine occupations tend to switch to jobs close to the previous work establishment, while blue collar workers show dispersion. The differences are chiefly explained by (1) non-routine activities concentrate in the CBD (the strongest effect) and local employment centres, (2) non-routine activities cluster also outside of centres, and (3) industry-specific effects. The patterns are consistent with the importance of sharply attenuating non-market interactions (e.g. knowledge spillovers) in the production of non-routine products and services.
Introduction
Cities are increasingly understood as arenas for reducing interaction costs, as opposed to units economising on the transportation costs of goods (Glaeser and Kohlhase, 2003). One suggested reason for this development is the alleged increased role of human capital externalities in production – a role that emphasises the benefits of productive interaction, the extent of which is often proxied by local density levels.
The city has public-good-like characteristics in the mediation of knowledge, but the magnitude varies across the urban landscape, and also matters differently depending on whether the task to be performed is intensive in its use of knowledge as an input. The empirical literature on human capital externalities has established two main stylised facts: first, spillovers attenuate sharply with distance, and second, local density matters more in so-called non-routine – or interactive – occupations (Andersson et al., 2014; Bacolod et al., 2009; Rosenthal and Strange, 2008).
These findings imply a spatial equilibrium where knowledge-intensive 1 occupations cluster to each other and to the CBDs to take advantage of productivity gains that depreciate with space, and hence imply increasing returns to local density. Such clustering is consistent with an internal city structure where e.g. financial districts almost invariably are located in the centre of cities, while capital-intensive production is generally located in the outskirts of urban areas. If these arguments are correct, there is a built-in mechanism favouring stronger spatial concentration and CBD orientation in non-routine occupations, owing to the spatial concentration of the spillovers. Hence, theories of human capital externalities and their attenuation with space predict within-region clustering of functions at the neighbourhood level.
If central land is a commodity that contributes to the production function of workers in non-routine professions, then within-city clustering of such work tasks is predicted by a simple bid-rent framework. The original bid-rent observation that capital-intensive production requires space and is pushed to the outskirts by rents (cf. Alonso, 1960) is observationally equivalent with knowledge-intensive production being pushed to the centre by increasing returns to density. If non-routine workers and activities are employed in tighter within-city clusters, then workers switching jobs should find prospective employers closer to their current employer, all else equal. The empirical section of this paper tests this prediction using data on job-switching distances from the old to the new employer, by occupation group, in Sweden’s metropolitan areas: Gothenburg, Malmö and Stockholm. Switchers’ distances are used as dependent variable, under the assumption that such flows inform about clustering tendencies by occupation. I also include variables to investigate how the between-occupation differences relate to the CBDs and local density, local employment centres, clustering outside the CBD, and industry-specific effects, among other determinants in the pertinent metropolitan landscape.
Why is this question important? First, occupation clustering at the neighbourhood level has implications for planning of the built environment, zoning laws, infrastructure projects and so on. Second, the issue bears directly on the productivity of cities. Job-switching overlaps two micro foundations of agglomeration economies (e.g. Duranton and Puga, 2004): (1) learning: the knowledge flows that individual workers bring with them to their new employer, and (2) matching: increased productivity through specialisation in thick markets, implying a feedback mechanism. To the extent that job-switching is a neighbourhood phenomenon (the best matches are nearby) then so is, at least in part, externalities contingent on job-switching, including human capital spillovers. If sharp attenuation of human capital spillovers induces clustering and localisation of the labour force, then localised job-switching rates follow and could in fact further contribute to the within-city clustering. We may thus understand localised job-switching as part of a circular causation, driving clustering of interactive activities and workers.
A job-switch in a database registers the flow that over time makes up the city’s geography of jobs. Why not simply use the resulting distribution? First, there is inertia in full distributions, resulting from e.g. costs of relocation, meaning that powerful posterior distributions may conceal interesting year-by-year dynamics. Second, job switchers may serve as an indication of optimisation decisions (for all matched parties) based on current market conditions. From this point of view, the creation of some number of new matches says more about an area than does knowledge of the existence of the same number of pre-existing matches. Cluster development (e.g. formation) related issues are generally better analysed in a flow, compared with a stock, framework.
Background and motivation
The 20th century has seen a relative decrease of blue collar jobs performed in cities, and a corresponding increase of professions involving thought, communication, and analytical tasks (Michaels et al., 2013). Such job tasks are sometimes referred to as ‘non-routine’, meaning that they require induction to be solved, and cannot in general be routinised by if-then type algorithms (Autor et al., 2003; Levy and Murnane, 2004). Consistent with interdependence in the development of non-routine job tasks on the one hand and agglomeration gains on the other, Glaeser and Kahn (2001) show that the introduction of the automobile to the broad masses tended to decentralise American employment, with the exception of knowledge-intensive industries. Common for tasks carried out in such industries is that they tend to be non-routine, intensive in the degree to which they benefit human capital externalities, and that they consistently exhibit higher returns to density (Andersson et al., 2014; Bacolod et al., 2009) These tasks are aptly performed in dense environments, because such environments facilitate knowledge accumulation, implying that fewer inputs have to be replicated in production. If an input (knowledge) can be replicated without cost, homogeneity of degree one is no longer a binding assumption about the production function, and constant returns to scale no longer a constraint on growth (cf. Romer, 1986).
Early estimates that indicate depreciation with space of human capital externalities are found in Rosenthal and Strange (2001), who compare estimates on the zip code and state levels. Subsequent approaches have moved to using individual data and towards abandoning the idea of an administrative region as a homogenous unit of observation (e.g. Rosenthal and Strange, 2008). The main conclusion derived from these results is that there is spatial friction of knowledge diffusion, since the value of non-market interactions depreciates with space. The diffusion is limited by the extent of the interaction arena, which has increasingly proven to be quite small in spatial terms. Included in the rather vague concept of non-market interactions is any unpriced interaction of productive value, including observation and imitation (cf. Durlauf, 2004); the interaction arena can be thought about as naturally constrained by what can be heard, seen and felt (Glaeser, 2000). This narrative begs the question of whether workers who benefit productively from hearing, seeing and feeling have higher returns to density (cf. Bacolod et al., 2009), and it also explains why the effect’s attenuation is particularly sharp for such workers.
A growing body of evidence suggests that the attenuation of human capital externalities is sharp. Using Swedish data, Andersson et al. (2016) document spillovers that dissipate after less than 1 km for university-educated workers. Larsson (2014) estimates that an average Swedish worker may increase his or her wage by almost 10% by moving from an average-density neighbourhood to the densest neighbourhoods. Arzaghi and Henderson (2008) analyse the advertising industry on Manhattan and conclude that spillovers dissipate already after a few blocks. Koster et al. (2013) document evidence consistent with substantial within-building agglomeration gains.
Needless to say, firm and employment misallocation can prove quite costly. If the depreciation is sharp enough, location differences of only a few blocks may have profound consequences in terms of production. If the externalities are important for productivity, then agents locating in close proximity to each other will be more productive. Since such a process will render switchers spatially sticky, this will also be true of embodied knowledge flows. If the embodied flows are localised, say, to the CBD, then learning will be more pronounced there relative to other areas (Glaeser, 1999).
Figure 1 displays an employment index (1991 = 100) plotting growth in terms of employment less than 5 km from the Stockholm, Gothenburg and Malmö metropolitan CBDs, defined as the neighbourhood housing the largest number of 5-digit industries in the central local municipality. The initial decline is due to the crisis that hit the Swedish economy hard in the early 1990s.

CBD employment growth (1991 = 100), 5 km radii around the CBDs in Malmö, Gothenburg and Malmö.
Though this picture does not inform about the nature of the jobs created, a few things stand out. First, employment growth in Stockholm has been substantially higher than in Gothenburg, where it has been substantially higher than in Malmö, even in relative numbers and disregarding the fact that Stockholm’s CBD was substantially denser in 1991. Second there are marked differences in the speed of recovery after the crisis. The Stockholm CBD was back at its pre-crisis level in 1998, Gothenburg in 2001, Malmö not until 2006. Third, the absolute numbers are impressive, too. Together these three CBDs with a combined area of a little over 230 km2 added 1000 net jobs per km2 on average, between the crisis low in 1993 and 2011. This figure represents 40% of net job growth in the metropolitan regions over the time period on less than 2% of the metro surface area. It also represents almost 25% of national net job growth during the same time period.
Data, variables and estimation
The data source is a publicly audited matched employer–employee data set, maintained by Statistics Sweden. Excluded from the data are workers in the public sector, mining and agriculture. The full population of private-sector employees in the metropolitan areas of Sweden who were registered as employed full-time in November of each year between 2003 and 2010 are included, subject to data availability.
Sweden has 61 metropolitan area local municipalities, aggregated into three metropolitan regions, each of which is integrated in terms of commuting flows (Johansson et al., 2003). The empirical analysis focuses strictly on within-variation in these three areas. The metropolitan regions are: Gothenburg (16 local municipalities), Malmö (15 local municipalities) and Stockholm (30 local municipalities). Associated with each metropolitan region is one regional centre (CBD) and several (non-principal city) municipal centres, referred to as local employment centres. The data do not cover individuals who changed their place of residence between regions during the period, although they do include individuals who lived in different local municipalities within regions. This operation has three main motivations: first, distance as the crow flies is not a good measure of functional distance across regions that are located in different parts of the country. Second, switches between regions are not indicative of cluster formation within regions. Third, the average of the dependent variable would be dominated by outliers. In total, cutting out between-region switchers excludes 304,428 individual-year observations. The resulting unbalanced panel contains 3,539,713 individual-year observations and tracks 970,994 metropolitan employees. In 2010, the number of individual observations is 568,447, and out of this population 77,227 changed jobs within their metropolitan area during that year (in total, the data track 685,219 within-metro switches over 8 years). The overall probability that a random employee in the population becomes a within-metro job-switcher during an average year is about 15%, although varying slightly with the business cycle and reaching its maximum at 17% in 2007.
Each worker is registered as employed in a certain work establishment, all of which belong to a firm, but each firm may of course have many work establishments (a worker must be associated with a new firm to be considered a switcher). Each worker and each work establishment are geo-coded down to a point in the southwest corner of a square (referred to as a neighbourhood), within a uniform, exogenously determined grid of squares, all of which are sized 0.25 km × 0.25 km (a base of about 0.16 miles).
Identification using geocodes and occupations
An issue in the literature concerns how to separate between workers’ skills or, crudely speaking, how to drive wedges between classes of workers who by some logic are considered different (in this case routine versus non-routine). Andersson et al. (2016) analyse university-educated workers, and conclude that the returns to density are higher for such workers, and that the attenuation effect is sharper. But education level is a crude proxy if the aim is to differentiate between skill sets in the domain specific sense. The underlying identifying assumption in this paper is that skills are rewarded differently in cities, and that a worker’s occupation says something about his or her skill set. 2 A rigorous empirical treatment of this question is found in Bacolod et al. (2009), who conclude that market pricing of cognitive skills – as proxied by occupation – is increasing with local population in all of their specifications, and that social (or people) skills are also being rewarded more highly by density. This finding is further evidence suggesting marginally increasing returns to density for interactive and non-routine employees.
Why analyse occupations and not industries? First, the relatively crude industry classification in the Swedish registry data is based on the SIC code of the majority of the firm’s output, meaning that a large fraction of the employees can be – and often are – assigned to an inappropriate industry. Second, many firms registered as manufacturing firms in reality produce a considerable amount of services, and many service firms produce a wide array of different services, and so on. Being able to differentiate functions within firms (and even within establishments within firms) is therefore desirable. The variables included in the empirical part do however include industry-specific effects.
In this paper, I represent 3 non-routine workers by managers, legislators, and senior officials (ISCO-88, 1), and by professionals, including engineers and scientists (ISCO-88, 2). Both of these groups are also highly educated (see Table 1). These workers are contrasted with employees in blue collar professions, who are more dependent on strength and motoric skills, and who have been shown to reap smaller, or no, benefits from density (Andersson et al., 2014; Bacolod et al., 2009). These are found in crafts and trades related occupations (ISCO-88, 7) and in occupations relating to operating of machinery and assembly (ISCO-88, 8). Table 1 summarises these occupations, including a cautious estimate of the extent of non-routine work tasks associated with each broad occupation group (adapted from Nilsson Hakkala et al., 2014).
Included occupation types, and average share of non-routine work tasks.
Note: the ‘average non-routine work tasks’ column is the unweighted average of two-digit occupations under each category, adapted from Nilsson Hakkala et al. (2014). The average years of schooling column is a population-weighted mean derived from Statistics Sweden’s registry data.
The empirical relevance of this division is illustrated in Figure 2, which shows new matches by neighbourhood in the Stockholm metropolitan region. The top panel shows matches for non-routine workers, while the bottom panel shows matches for workers in blue collar professions.

Labour market matches, by neighbourhood (250 m × 250 m), in the Stockholm metropolitan area. The maps show matches for non-routine workers (including engineers, scientists, and managers, top) and for blue collar workers (including crafts, and machine operation, bottom).
First, note how concentrated both pictures are relative to the surrounding lands. Second, note how non-routine workers find jobs chiefly in close proximity to the CBD and also to a corridor to the north, heading to the industrial districts and research hubs of Solna and Kista. Notably, most of the densest neighbourhoods in terms of matches are within the borders of the central local municipality. The blue collar workers’ picture is more dispersed, with local clusters scattered in the surrounding local municipalities (such as Södertälje to the southwest and Järfälla to the north), but still with the bulk of matches taking place in proximity to the region centre of Stockholm. Taking the perspective of new matches as optimisation decisions, those decisions do appear to favour the CBD, and most notably so in non-routine professions.
Certainly, these patterns are consistent with the bid-rent predictions reviewed above, where firms that derive benefits (and profits) from density pay higher rents for access to density. This notion is investigated in the regression analysis part of this paper, where I make some attempts at discriminating between occurrences in the metropolitan landscape. In addition to basic observables, the empirical section contains a number of key variables sorted into three categories: bid-rent, clustering and industry-specific effects.
Bid-rent variables
One reason why we a priori would expect patterns such as in Figure 2 is simply that non-routine workers have their productivity determined as an increasing function of local density. If we expect the city centre to be a hub for recruitment of such workers, then we should expect them to move short distances as they shop around for jobs near the city centre. This class of variables include distances to the CBD, defined as the neighbourhood in the central municipality that is home to the largest number of 5-digit industry codes, distance to local employment centres, and local neighbourhood employment density.
Clustering variables
There is also the possibility of clustering of functions across the geography unrelated to the CBDs. In fact, powerful non-market interactions predict multiple equilibria across space, as clusters grow powerful by internalising knowledge. Such a case could be driven e.g. by large within-occupation specialisation gains, and the clusters may be based on natural advantages, historical accidents, local cultures (see e.g. Andersson and Larsson, 2016, who demonstrate local feedback effects in entrepreneurship) and so on. This step includes controls for ‘hotspots’ (places in space where the largest number of similar workers are employed, a measure independent of CBDs and local employment centres) and also overall clustering measures (the number of neighbourhoods employing similar workers relative to the total number of active neighbourhoods per region, and local municipality).
Industry-specific effects
There are also likely to exist industry-specific linkages that are not picked up by occupation-based variables. There is potentially an industry-specific effect localising workers, where the industry is bound to stay close to input suppliers, complementary services and so on, meaning that the average length of a move may be an effect of the industry in which the worker is employed. These variables include industry dummies at the 2-digit level and controls for the book value of capital stocks per firm.
All variables are defined and discussed in further detail in section ‘Variables’ below.
Dependent variable: Job-switching distances by occupation
An overview picture of the dependent variable for all private-sector workers is displayed in Figure 3 (switches from defunct firms and workplaces are excluded), where switchers’ distances between current and previous work establishments are plotted along the x axis.

Distance decay in within-region job-switching: distance in kilometres between new establishment and establishment in year t−1. N = 293,965.
There is a clear tendency towards ‘distance decay’ of matching in these metropolitan labour markets. In fact, in almost 20% of the cases, the workers locate within about 1 km from where they were employed in year t−1 (do note that this figure excludes those employees that left the region).
The graphs in Figures 1–3 do not speak of the sources of this behaviour, nor are they informative in terms of between-occupation differences. In Figure 4, the same data are reproduced for professionals (e.g. engineers and scientists, upper panel), and crafts and trades workers (lower panel). These occupation-specific graphs are contrasted against all other workers, represented by the outline-only bars in the foreground. Following the discussion above about matches as optimisations decisions by both agents, the occupation code used is simply the one at time t, i.e. the occupation switched to.

Distance decay in within-region job-switching: distance in kilometres between new establishment and establishment in year t−1.
The fraction of professionals who find a new employer in the same neighbourhood is nearly twice as large as the fraction of crafts and trades workers who locate to the same neighbourhood. The fraction of professionals is larger than all other workers for each distance-bar approximately up until the 10 km mark, after which the pattern is reversed. For crafts and trades workers, the pattern is quite different (in fact, it is close to the inverse), as such workers consistently appear to move farther than the average. The average distance between the old and the new workplace is just over 7 km, but there is substantial between-occupation variation as depicted in Table 2.
Average distances for within metropolitan regions private-sector job switchers, in km, between workplace in time t, and various points in the pertinent urban landscape.
Note: The occupations correspond to the ISCO-88 standard in parentheses. The statistics exclude workers who live more than 200 km from their workplace. Workers in mining and related industries are excluded from category (7).
The two first rows describe non-routine professions, while the second two rows describe blue collar professions. Non-routine workers move shorter distances between jobs, and their tendencies towards the central areas are stronger. The workers that move the shortest distances between workplaces also tend to be the workers that are employed close to the CBDs or close to the local employment centres. With professionals (including engineers and scientists), the tendency is even stronger than for managers. This may be seen in the marked difference between the previous and the latter group in terms of proximity to the CBD.
At first glance, the differences may seem immaterial, but two things should be noted. First, numerous studies using disaggregated data sources have concluded that a city’s public-good-like characteristics depreciate quickly with space, where some of those characteristics seem to be internal to a few blocks, or even a building. Second, when thinking about the differences as proxies for search ranges, a circle with radius 10.5 is actually about three and a half times as large in terms of surface area, compared with a circle with radius 5.6.
Estimation issues
This section describes the estimation procedure used to discriminate between the effects listed above. The most straightforward strategy may be to simply run an OLS regression on the workers that switched jobs. This practice would however assume that job switchers are a random subset of the population. Selection could be a concern if certain workers are more prone to switching jobs, while simultaneously exhibiting correlation with the variables of interest. To account for such selection, I use a Heckman selection model, which in a first step estimates the probability of an individual switching jobs. The variables of the selection equation and their coefficients are presented in Appendix 1. The full set of variables included in the selection equation are displayed in Table A1.1. The table also contains variable definitions, descriptions, and also summary statistics for each variable. The corresponding probit coefficients obtained by estimating (1) are displayed in Table A1.2. Formally the selection equation is defined as:
where Ni,t is a dummy, indicating 1 if individual i switched jobs between t−1 and t. Further, xi,t−1 contains lagged individual control variables, and Φ denotes the cumulative density function. Previous studies provide evidence that individuals of certain attributes (e.g. short tenure and young age) are more likely to become job switchers, providing a general idea of standard control variables (see e.g. Andersson and Thulin, 2013). Agents in employer–employee pairs that are productively matched in period t−1 are less inclined to look for a new match in period t. An important variable in this regard is tenure, 4 i.e. number of consecutive years spent with the same employer (Farber, 1994). Here, tenure is measured from 1991 (i.e. for up to 19 years).
The baseline outcome equation to be estimated is then specified as:
where yi,t is the number of kilometres that individual i moved between employers between periods t−1 and t, β contains the coefficients of the main variables of interest, relating to individual i’s occupation type (p) at time t. Further, Z is a matrix of control variables, as outlined below. The D variables are dummies, relating to year, labour market region, and region–year pairs, respectively. Finally, λi is the inverse Mills ratio obtained through (1), and εi,t is a white noise error term.
Variables
The variables in the outcome equation are defined in Table 3.
Variables used in in the outcome equation. The selection equation variables are displayed in Table A1.1.
Note: Region dummies are included in all regressions. Year dummies are also included in to control for time trends, as well as region*year dummies, to control for region-specific, time-variant shocks.
The variables are separated by five levels, and the regressions are run for each level separately to assess the contribution from each set of variables. Summary statistics for each variable are presented in Table 4.
Descriptive statistics for variables used in in the outcome equation.
Note: N = 685,219 (total number of job switches). The selection equation variables are displayed in Table A1.1.
The main variables of interest are the occupation dummies in the top part of the table. The estimation process advances in five steps, and the main idea here is to study the change in the dummy variables, as more controls are gradually included. To address concerns that the variables we add introduce collinearity issues to the model, steps 3–5 (containing the variables of interest), below are also estimated individually in Appendix 2, Table A2.1.
Step 1: Baseline
First, baseline occupational differences are estimated without controls. This step contains the ‘main and baseline variables’ from Tables 3 and 4, including dummies for year, labour market region, and region-year pairs.
The baseline model also includes indicators reflecting whether the previous work establishment was discontinued (slightly more than half of switches), which includes simple exit (e.g. through bankruptcy), but also whether the firm was bought up and continued as a ‘new’ firm under different ownership. The occupation dummies used are for ISCO-88 1-digit occupation levels 1, 2, 7 and 8 at time t, as outlined above (see e.g. Table 2). As can be seen from Table 4, all categories contain slightly less than 10% of the total number of observations, except for professionals, which contains 20%. The implication is that the base category is all other private-sector job switches included in the data set, stemming from any other category. This step finally includes dummy variables indicating the metropolitan region. More than half of the switches took place in Stockholm, a little over one-fourth in Gothenburg, and 15% in Malmö.
Step 2: Basic observables
The next set of coefficients are obtained controlling for sorting on basic observables. This equation includes variables describing processes that could conceivably constrain or encourage a worker’s movement across space, but also variables that are likely to correlate with such processes, such as immigration status, and sex. This category includes the yearly percent wage change between t−1 and t, since a higher yearly wage may act as a compensating differential and induce workers to move farther distances. Further, this step contains controls for distance to the home location, 5 as well as the establishment’s number of employees, since location in space is in part a product of residence location, and of sheer workplace size.
The second step also includes controls for observable human capital: schooling and experience, as suggested by Mincer (1974). This step allows us to observe the differences, keeping a common measure of human capital constant, allowing subsequent steps to analyse what is not captured by such a measure, including non-routineness of work tasks.
Step 3: Bid-rent variables
The next step includes controls for bid-rent positioning, wherein profitable firms are hypothesised to outbid their competition for attractive locations. The variables include distances to CBDs, and to local employment centres. These spots are defined as distance from the new employer to the point that is home to the largest number of 5*digit industries in each metropolitan region and in each local municipality, respectively. This is an empirical definition. It may seem most intuitive to simply use the densest points in terms of employment. A problem is that some workplaces outside the cities are so massive that they employ workers over a larger area than the squares (while still having all workers registered at the same point). This operationalisation would give nonsensical locations of some CBDs. Such is the case e.g. in Gothenburg, where the scale of the automotive industry would move the ‘CBD’ to the outskirts of the city.
The distance variables are complemented by local density measures, corresponding to employment density in the 1 km2 square that each workplace belongs to (the fact that each workplace belongs to a 0.25 km × 0.25 km square implies that 16 such squares make up 1 km2).
Step 4: Clustering variables
The fourth step includes clustering variables, controlling for the geographic concentration of employment for workers with the same 1-digit ISCO-88 standard classification, but these measures are independent of CBDs, local employment centres, and neighbourhood density.
The occupation clustering measure is defined as the total number of 0.25 km × 0.25 km squares where similar workers are employed, relative to the total number of squares with economic activity, per local municipality and metropolitan region, respectively. The variable ranges from a theoretical high of 1 (implying that the occupation is represented literally everywhere where there is economic activity) to a low of almost 0 (meaning that the occupation is only available in few spots). Table 4 informs that in practice this variable stretches from (close to) 0 to 29% locally, and 17% regionally.
Finally, the step includes distances to local and regional hotspots, defined as the neighbourhood where most workers sharing the same 1-digit ISCO-88 standard classification are employed. The hotspots are sometimes – but not necessarily – identical to the CBDs and local employment centres for some occupation groups.
Step 5: Industry-specific effects
Industry-related effects, such as capital intensity, accessibility to input suppliers and customers are likely to play a role. This phenomenon is controlled for using 2-digit NACE dummies corresponding to the industry that employed the worker at time t. Additionally, this step controls for the natural logarithm of the book value of capital per worker in the firm.
Results
The coefficients from the outcome equation (2) for the five steps described above are displayed in Table 5. The first step presents baseline estimates, conditioned only on selection. The second step adds basic observables (personal characteristics), while the three subsequent estimations investigate the occupation parameters as variables describing bid-rent, clustering, and industry-specific factors are sequentially added as covariates. The base category among the occupations is all private-sector job switchers who do not belong to any of the four groups discussed in the previous section.
Occupation differences in job-switching distances.
Note: Robust standard errors are clustered at the level of neighbourhoods (0.25 km × 0.25 km square). All variables are defined in section ‘Data, variables and estimation’. **p < 0.01, *p < 0.05. The parameters are estimated using a Heckman selection model (selection equation in Appendix A1). Period: 2003–2010. Dependent variable: job switchers’ distances in km between work establishment in year t, and year t−1. The last two lines offer selectivity estimates, investigating correlations between the residuals from the first stage, with the residuals from the second stage. Regressions adding all sets of variables individually are presented in Table A2.1.
The baseline regression essentially confirms the patterns from the descriptive Table 2: the difference between non-routine (the managers, and professionals categories) and blue collar (the crafts, and plant categories) workers is considerable. The difference in average distance between previous and current employer is about 4 km between someone hired as a professional and someone hired as a crafts worker. The picture confirms that non-routine workers switch jobs over much more concentrated areas of land, compared with blue collar workers.
In the second step, Mincer’s (1974) classic determinants of human capital, schooling and experience, are included together with an array of variables describing each worker’s characteristics, e.g. in terms of sex and immigration status. This step also includes an estimate of establishment size in terms of number of workers, and the distance in kilometres to the location of residence. The most notable result in this step is how little these characteristics contribute to the between-occupation differences, considering that schooling is controlled for at this stage. However, this set of variables do indeed explain some of the distance covered, foremost by crafts and trades workers.
Step three adds controls for bid-rent behaviour, in terms of distances to the CBD, and to the local employment centres, respectively. This step also contains a control for employment density of the 1 km2 neighbourhood to which the work establishment belongs. Bid-rent theory predicts that more profitable firms will be located closer to the city centre, since they will outbid less profitable competition for attractive locations. Knowledge-intensive firms are thought to become more profitable because they locate close to city centre, as discussed above.
Adding these controls pushes all variables of interest closer to the zero bound, and reveals that a substantial part of the differences are produced by non-routine workers finding new jobs primarily in and around the CBDs (cf. Glaeser, 1999) and by blue collar workers primarily finding new jobs outside of the city centres.
The lion’s share of the effect here is produced by distance to the CBD and by density of the new neighbourhood. Distance to the local employment centre is statistically speaking less important, as is evidenced by the lower level of significance. The coefficient becomes statistically indistinguishable from zero in subsequent steps. This result may be appreciated as further evidence of the importance of the CBD and local density in providing knowledge-intensive employment. In addition, the unstandardised coefficients actually understate the relative differences since a one standard deviation change is more than three times as large for the distance to CBD variable (see Table 4).
On average, moving to a workplace an additional kilometre away from the CBD is associated with about a 0.15 km longer move between employers. Just outside of the 5 km cutoff used in Figure 1, moves would on average be close to 1 km farther on average, more than enough for much of human capital externalities to dissipate in non-routine industries, such as the advertising industry described in Arzaghi and Henderson (2008). 5 km corresponds to slightly more than one standard deviation in Table 4.
Step four further includes controls for clustering, owing to the possibility of cluster formation outside of the metropolitan cores. The occupation clustering variables are statistically significant, both locally and regionally, albeit with coefficients of different signs. The results indicate that there is indeed some clustering outside of the centres, associated (at least statistically) with the tendency for non-routine workers to be more spatially concentrated (note the decreases in all coefficients of interest). The shifting signs of the clustering coefficients indicate that occupations clustered in the local municipality are associated with shorter moves, while regional clustering exhibits an inverse relationship. Hence: if similar jobs are clustered in the local municipality, the moves are shorter. Overall, a one standard deviation (0.02 in Table 4) increase in clustering is associated with about 0.5 km longer moves on the level of metropolitan regions, and about 0.45 km shorter moves on the level of local municipalities.
The fifth and final step adds controls relating to industry. In this step, dummies at the 2-digit NACE industry level are introduced, as is a variable indicating the size of the capital stock per employee (size in terms of the number of employees are controlled for since the second step). On average, a 1% increase in capital per employee, is associated with 0.1 km longer moves on average, indicating that capital-intensive production is dispersed around the city, and often concentrated to the outskirts. This step further pushes the between-occupation differences closer to zero (except for in management occupations). After this step and after accounting for all control variables, the only occupation group that moves more than 1 km farther than other groups is plant and machine operators and assemblers. It may also be noted that the selectivity statistics (the last two rows of the table) are close to zero at this stage, indicating low degrees of non-random sorting in the fully specified model.
Comparing steps 1–5 in Table 5, reveals that about three-fourths of the span in between-occupation differences are explained by the variables introduced, where bid-rent related variables – foremost distances to the CBDs and employment density – appear most important.
To assess the relative contribution of each category of variables, as well as issues of collinearity between them, the model is estimated for all categories individually in Appendix 2. Columns 1–2 are identical for reference and steps 3–5 estimates separately the effect of the bid-rent, clustering and industry categories. By studying the change in the coefficients in columns 4 and 5, it is revealed that clustering and industry-specific factors do contribute less than the bid-rent variables (additionally, some of the effect of clustering here is driven by the fact that these clusters are sometimes located close to, or in, the CBDs). It appears, then, that the largest share of the clustering behaviour observed is driven by a high proportion of non-routine workers finding new jobs close to their old place of work in the metropolitan areas’ main CBDs, local employment centres and other dense areas.
Conclusion
This paper demonstrates that there is substantial variation across occupations in terms of how they cluster within cities. Using job-switching data for Sweden’s three metropolitan regions, I analyse clusters of different economic activities as evidenced by clustering of occupations, by analysing job switchers’ length of moves between previous and current employers. I show that workers in non-routine type occupations move substantially shorter distances than other workers when they switch jobs, indicating that they on average find new employment in localised clusters, while blue collar, non-interactive, jobs do not show nearly as high levels of concentration. This finding corroborates theoretical arguments about clustering behaviour as a function of attenuating human capital externalities. It also has the power to explain the increasingly important role of the CBD in economic geography. The sharper is the depreciation with space of knowledge spillovers, the stronger are the incentives for strong clustering in space. The more intensive an occupation utilises knowledge inputs in production, the more likely we are to observe that occupation in the CBD. Further, with more knowledge-intensive production and more non-routine work tasks comes more powerful incentives to form parts of clusters.
Using regression analyses with switchers’ distances between employers on the left hand side, I show that – even though general clustering and industry-specific effects do play a role – the most powerful predictor of this behaviour is clustering towards the centres and the locally dense parts of the metropolitan regions. The effect is nowhere as strong as with non-routine type professions, such as engineering, science and upper-management occupations.
As the share of knowledge in production increases, the role of human capital externalities is gaining in importance, and so is the tendency towards powerful CBDs. The density-driven employment growth documented in this paper has clear policy implications. This phenomenon e.g. means that there is a direct link between zoning laws and building height regulations on the one hand, and the productivity and future growth prospects of cities on the other. Some occupations may indeed require a certain frequency of interaction in order to stay competitive. Measures to keep down building height, for instance, may have indirect consequences on employment growth, specifically in knowledge-intensive industries.
Since non-routine occupations seem to have close proximity to human capital externalities (and therefore economic density) as an important input in the production process, this research lines up with a growing body of evidence suggesting a strong link between the density of cities – as well as neighbourhoods within cities– and long-run growth prospects.
Footnotes
Appendix 1
Determinants of job-switching. Heckman selection equation, probit coefficients.
| Variable | Coefficient |
|---|---|
| Schoolingt−1 | −0.00858** (0.000742) |
| Experience t−1 | −0.0156** (0.000128) |
| Wage t−1 (ln) | −0.0782** (0.00217) |
| Tenure t−1 | −0.0346** (0.000321) |
| Market potential t−1 (ln) | 0.0311** (0.00210) |
| Discontinued firm | 3.297** (0.00534) |
| Discontinued workplace | 0.231** (0.00891) |
| Firm size t−1 (ln) | 0.0131** (0.000527) |
| Capital/employee t−1 (ln) | −0.0050** (0.000421) |
| Malmö metropolitan region (dummy) | −0.069** (0.00358) |
| Gothenburg metropolitan region (dummy) | −0.089** (0.0028) |
| Female (dummy) | −0.0379** (0.0027) |
| Immigrant (dummy) | 0.00646* (0.0030) |
| Married t−1 (dummy) | −0.00206 (0.00260) |
| Children staying at residence t−1 (dummy) | −0.00218 (0.00239) |
| Managers, legislators, senior officials (dummy) | 0.110** (0.00580) |
| Professionals, incl. engineers and scientists (dummy) | 0.0581** (0.00521) |
| Craft and related trades workers (dummy) | −0.00963 (0.00584) |
| Plant and machine operators and assemblers (dummy) | −0.000541 (0.00598) |
| Pseudo R-squared | 0.44 |
| Observations | 3,539,716 |
| Individuals | 970,994 |
Note: Standard errors in brackets. The regressions include year dummies to control for season effects, education type dummies, industry dummies at the two-digit level, and an additional six occupation dummies (1-digit ISCO-88). All variables are defined in Table A1.1. **p < 0.01, *p < 0.05.
Appendix 2
Occupation differences in job switching distances.
| Variable / Step (see section Data variables and estimation) | (1) | (2) | (3) | (4) | (5) |
|---|---|---|---|---|---|
| Baseline | Observables | Bid-rent | Clustering | Industry | |
| Managers, legislators, senior officials | −0.621** | −0.474** | −0.362** | −0.661** | −0.555** |
| (0.0834) | (0.0790) | (0.0783) | (0.123) | (0.0774) | |
| Professionals, including engineers and scientists | −1.690** | −1.424** | −0.870** | −1.320** | −1.041** |
| (0.113) | (0.0932) | (0.0912) | (0.183) | (0.0939) | |
| Crafts and related trades workers | 2.975** | 2.096** | 0.876** | 1.196** | 1.076** |
| (0.174) | (0.160) | (0.153) | (0.212) | (0.153) | |
| Plant and machine operators and assemblers | 2.769** | 2.642** | 1.131** | 1.280** | 2.094** |
| (0.260) | (0.226) | (0.221) | (0.239) | (0.218) | |
| Region: Malmö | −0.161 | −0.300 | −1.613** | −0.607 | −0.464 |
| (0.523) | (0.483) | (0.472) | (0.440) | (0.445) | |
| Region: Gothenburg | −1.158** | −1.221** | −1.739** | −1.750** | −1.289** |
| (0.423) | (0.402) | (0.526) | (0.522) | (0.383) | |
| Discontinued firm | −6.842** | −5.911** | −6.566** | −6.406** | −6.544** |
| (0.274) | (0.295) | (0.227) | (0.230) | (.209) | |
| Discontinued workplace | 3.117** | 3.080** | 3.126** | 3.204** | 3.013** |
| (0.224) | (0.211) | (0.224) | (0.220) | (0.210) | |
| Years of schooling | −0.137** | −0.0526** | −0.0854** | −0.154** | |
| (0.0179) | (0.0184) | (0.0187) | (0.0166) | ||
| Experience | −0.0723** | −0.0839** | −0.0794** | −0.0752** | |
| (0.00405) | (0.00360) | (0.00364) | (0.00336) | ||
| Wage change from previous year (%) | 0.313** | 0.274** | 0.295** | 0.221** | |
| (0.0398) | (0.0391) | (0.0392) | (0.0394) | ||
| Distance to home (km) | 0.0982** | 0.0874** | 0.0913** | 0.0957** | |
| (0.00359) | (0.00318) | (0.00318) | (0.00344) | ||
| Female (dummy) | −0.567** | −0.341** | −0.618** | −0.456** | |
| (0.0663) | (0.0672) | (0.0674) | (0.0531) | ||
| Immigrant (dummy) | −0.252** | −0.0108 | 0.0231 | −0.116* | |
| (0.0587) | (0.0583) | (0.0594) | (0.0566) | ||
| Workplace employee size (ln) | −0.209** | −0.151** | −0.152** | −0.276** | |
| (0.0347) | (0.0327) | (0.0333) | (0.0307) | ||
| Distance to CBD (km) | 0.175** | ||||
| (0.0124) | |||||
| Distance to local employment centre | 0.0407* | ||||
| (0.0183) | |||||
| Employment density (1km2 neighbourhood, ln) | −0.897** | ||||
| (0.117) | |||||
| Distance hotspot (km, region) | 0.212** | ||||
| (0.0103) | |||||
| Distance to hotspot (km, local) | −0.113** | ||||
| (0.0140) | |||||
| Occupation clustering (region) | 14.61* | ||||
| (6.844) | |||||
| Occupation clustering (local) | −2.28** | ||||
| (7.325) | |||||
| Capital/employee (ln) | 0.175** | ||||
| (0.0169) | |||||
| 2-digit SIC dummies | No | No | No | No | Yes |
| Year dummies | Yes | Yes | Yes | Yes | Yes |
| Region-year dummies | Yes | Yes | Yes | Yes | Yes |
| Constant | 10.24** | 12.10** | 19.35** | 11.54** | 11.54** |
| (0.332) | (0.494) | (1.413) | (0.634) | (.634) | |
| Observations | 3,539,713 | 3,539,713 | 3,539,713 | 3,539,713 | 3,539,713 |
| Individuals | 970,994 | 970,994 | 970,994 | 970,994 | 970,994 |
| R-squared | 0.11 | 0.14 | 0.22 | 0.16 | 0.20 |
|
|
0.0338* | 0.0728** | 0.0268* | 0.0420** | 0.0250** |
| (0.0137) | (0.0150) | (0.0112) | (0.0104) | (0.0108) | |
|
|
0.3710** | 0.7838** | 0.2766* | 0.4369** | 0.2665* |
| (0.1518) | (0.1651) | (0.1163) | (0.1092) | (0.1158) |
Note: Robust standard errors are clustered at the level of neighborhoods (0.25 km × 0.25 km). All variables are defined in section ‘Data variables and estimation’. **p < 0.01, *p < 0.05. The parameters are estimated using a Heckman selection model (selection equation in Appendix A1). Period: 2003–2010. Dependent variable: job switchers’ distances in km between work establishment in year t, and year t−1. The last two lines offer selectivity estimates, investigating correlations between the residuals from the first stage, with the residuals from the second stage.
Appendix 3
Acknowledgements
I am grateful to the editors of Urban Studies and to anonymous referees for valuable comments on the manuscript. Their recommendations have substantially improved all aspects of this article. The manuscript has also been substantially improved after comments by Professor Martin Andersson, Professor Henri de Groot and Dr. Özge Öner. I am indebted to Therese Norman for kindly producing the maps in this article.
Funding
I acknowledge funding from Jan Wallander’s and Tom Hedelius’ foundation, as well as from The Swedish Research Council FORMAS (project code 2014-5968-28390-32).
