How do occupational relatedness and complexity condition employment dynamics in periods of growth and recession?

ABSTRACT

Related diversification has generated interest in policy (Smart Specialisation) and academic (regional branching) circles, linking path creation to regional capabilities and performance. We develop measures of occupational relatedness and complexity for local labour market areas in Sweden over the period 2002–12 to examine whether these constructs are helpful in explaining spatial and temporal variations in employment growth. The results indicate that increases in occupational relatedness are positively related to employment growth, while changes in occupational complexity have no significant impact. Separating the results either side of the global financial crisis shows that the influence of relatedness on employment is stronger pre-2008, while after the crisis building specialization in more complex occupations was positively associated with employment growth.

KEYWORDS:

• related branching
• occupation space
• Smart Specialisation
• employment growth
• occupational relatedness
• complexity

JEL:

• J21
• R58

INTRODUCTION

The process of economic transformation is currently high on the agenda in both academic and policy environments as many regional economies have struggled to renew their competitiveness following the financial recession of 2008, and as inequality threatens economic futures (Iammarino et al., 2019). Prompted by core arguments within evolutionary economic geography (Frenken & Boschma, 2007), considerable work has focused on how regions diversify into new products (Hidalgo et al., 2007), industries (Neffke et al., 2011) and technologies (Balland et al., 2015; Rigby, 2015), generating new specializations (Kogler et al., 2017) and alternative growth paths (Hassink et al., 2019). The Smart Specialization Agenda of the European Union (EU) (Foray et al., 2011) is heavily influenced by these ideas, though perhaps less so in practice (Gianelle et al., 2019; Marrocu et al., 2020).

How regions fare socio-economically depends to a great extent on the labour market’s ability to adapt to structural change and crises. As suggested by Moro et al. (2021), regional adaptability does not stand or fall with a single factor but depends on the internal composition of resources, and the relatedness and interaction among these components. Consistent with these claims, the regional diversification literature shows that regions do not evolve in a random fashion; rather, they tend to build new capabilities along trajectories that are strongly related to existing practice (Boschma et al., 2013; Rigby & Essletzbichler, 1997), while abandoning activities that are not as well embedded. New research is beginning to link patterns of related and unrelated diversification to regional economic resilience and other measures of performance (Balland et al., 2015; Frenken et al., 2007). At the same time, a growing interest in economic complexity (Hidalgo & Hausmann, 2009) seeks to understand which development trajectories may be most beneficial for regions to develop (Mewes & Broekel, 2020; Pintar & Scherngell, 2021; Rigby et al., 2019). For Rodríguez-Pose et al. (2014), the Smart Specialisation framework emphasizes that regions should not simply imitate their best performing neighbours (see also Foray et al., 2011), but should rather seek to identify and develop unique specializations that build on existing capabilities and which hold promise in terms of adding value. Balland et al. (2019) combine these two concepts of ‘relatedness’ and ‘complexity’ within a Smart Specialisation framework to explore the costs and benefits of regional diversification along competing pathways.

The primary aim of this paper is to use the framework of Balland et al. (2019) to explore the links between changes in occupational relatedness and complexity and the rate of growth of employment within Swedish labour markets. In so doing, we respond to the calls of Boschma (2017) to deploy the concept of relatedness in different domains and across new research questions. Muneepeerakul et al. (2013), Jara-Figueroa et al. (2018), Farinha et al. (2019), Davies and Maré (2021) and Lo Turco and Maggioni (2020) have already taken up this task, developing innovative variants of an occupation space to explore how changes in the occupational structure of places and firms impact economic performance. We also seek to inform policy. The measures of occupational relatedness that we construct provide useful information regarding occupational complementarities at the local level that might be used to enhance short-run cohesion or longer run resilience in labour markets (Moro et al., 2021). Further, our index of occupational complexity might be considered as a proxy of the importance of different occupations to the generation of urban competitive advantage. Indeed, Hausmann et al. (2021) report that complexity is the single largest determinant of the wage gap in parts of Mexico. Our results show that both relatedness and complexity are related to employment growth within Swedish labour markets, exerting effects that vary across periods of growth and recession, adding a temporal dimension to relatedness research that has recently been called out by Kuusk and Martynovich (2021).

The research presented adds value in several ways. First, many of the papers outlined above use regional co-location as a relatedness measure raising questions about what co-location captures. Here we follow Neffke and Henning (2013) and Neffke et al. (2017) and link occupations from observed worker flows between them, providing a micro-perspective of the mechanism by which occupations are related.

Second, most previous work on relatedness and regional diversification has tended to be output focused, often disregarding the labour skills and work tasks required as inputs in the development of new economic activities. This might bias our understanding of regional transformation since the know-how embodied within groups of workers, alongside the demands generated by new technologies and the geographies they foment, are critical to the long-run fortunes of workers and regions (Boschma & Martin, 2010). This is especially problematic in an age where regional economies are specializing in specific functions and work tasks rather than in industries (Baldwin, 2006; Wixe & Andersson, 2017).

Third, the relatedness of occupations is supplemented with an analysis of their complexity (Hidalgo & Hausmann, 2009). We follow the Smart Specialisation model of Balland et al. (2019), operationalizing their arguments with occupation data and thereby answering the call of Davies and Maré (2021) for more comparative work on relatedness and complexity in labour markets.

Fourth, we go further than most analyses of related diversification, seeking to show how changes in occupational relatedness and complexity at the regional level are related to employment growth. Linking the concepts of relatedness and complexity to economic performance adds to the regional policy-makers toolkit.

Fifth, Eriksson and Hane-Weijman (2017) report that the employment impacts of the financial recession of 2008 have been markedly uneven across Swedish regions, echoing similar findings elsewhere (Fratesi & Rodríguez-Pose, 2016; Marelli et al., 2012). In response to this work, we examine how the relationship between occupational relatedness, complexity, and employment growth varies before and after the financial crisis. In this respect, our analysis follows Frenken et al. (2007) who suggest that relatedness might work differently in periods of economic growth and decline.

The remainder of the paper is structured as follows. In the next section, a brief summary of employment growth theory is outlined along with motivation for focusing on occupations. The third section outlines the data and construction of measures of the relatedness and complexity of occupations and regions. The fourth section presents descriptive statistics for Sweden’s labour markets and results from an empirical model linking occupational relatedness and complexity to regional employment growth. The last section offers a brief conclusion.

THEORETICAL FRAMEWORK

There is mounting evidence that over the past decades, employment, earnings and career trajectories have become much more unstable, at least for those at the less-skilled end of the labour market (Farber, 2008; Hollister, 2011). The combined impacts of technological innovation and the offshoring of jobs within an increasingly integrated global economy have accelerated the fragmentation of commodity production and the geographical separation of work tasks that had previously tied industries and sets of skills to particular locations (Baldwin, 2006; Cooke et al., 2019; Jensen & Kletzer, 2010; Moretti, 2013). Sweden is a prime example of these tendencies since only 13 of 72 local labour markets have contributed to national job creation since 1990 (Eriksson & Hane-Weijman, 2017), and regional specialization has become less tied to specific sectors and more closely connected to the spatial division of occupations (Wixe & Andersson, 2017). On top of these longer running trends, the Great Recession exacerbated labour market precarity within the United States and much of the EU, especially for younger workers (Ayllon & Ramos, 2019; Lowe, 2018). The future of work and employment is generating increased debate across many parts of the world (Rani & Grimshaw, 2019), including Sweden (Woolfson et al., 2014), where Henning and Eriksson (2021) report that regions with traditional manufacturing cores have undergone intensified changes in occupational structures.

The volume of research on regional employment growth is vast and far beyond the scope of this paper to review extensively. A recent, broad overview of the field is provided by Criscuolo et al. (2014). In brief, three distinct streams of research on (regional) employment growth may be identified. The first investigates the relationship between firm dynamics (especially start-ups and the role of small and medium-sized enterprises) and employment growth and might be considered a reaction to the early work of Birch (1979) who claimed that small firms are the primary creators of jobs. Empirical evidence on the net contribution of small firms is contradictory, and factors such as firm age (e.g., Adelino et al., 2017; Haltiwanger et al., 2013), technology intensity (e.g., Hathaway, 2013; Anyadike-Danes et al., 2015) and institutional context (e.g., Li & Rama, 2015) greatly structure job generation. Further research suggests that small firms and new start-ups tend not to be a stable source of employment growth in general (Eriksson & Hane-Weijman, 2017; Essletzbichler, 2007), and even less so in times of recessions (Criscuolo et al., 2014).

A second branch of research on employment change focuses on the influence of economic structure, heavily influenced by the seminal work of Glaeser et al. (1992). The antecedents of this work lie in the well-known claims of Marshall (1920) and Jacobs (1969) regarding the respective roles of industrial (and other forms of) specialization and diversity on regional development. Duranton and Puga (2004) tighten these theoretical claims in relation to the economies of sharing, matching and learning that generate positive feedback mechanisms attracting workers and firms and expanding regional economies. The broad empirical overview of Beaudry and Schiffauerova (2009) suggests that empirical evidence supporting either specialization or diversity is far from conclusive. In a seminal paper that links research on economic structure to newer work on relatedness (see also Boschma & Frenken, 2007; Hidalgo et al., 2007), Frenken et al. (2007) distinguish related from unrelated variety and examine the influence of these different forms of diversity on regional employment and other measures of growth. For them, the distinction between regional economies as specialized or diversified is too crude, as they push to consider the relatedness between economic sectors. Thus, different sectors that share some underlying characteristics, such as a knowledge base, are considered more related than sectors that share little. Frenken et al. (2007) argue that the recombinatorial dynamics hinted at by Jacobs (1969) are more likely present in regions with greater related variety than unrelated variety, while unrelated variety appears to limit the impacts of negative shocks on regional economies (see also Balland et al., 2015). The authors report broad empirical support for these claims, which is bolstered by subsequent analyses on Italian (Boschma & Iammarino, 2009), Finnish (Hartog et al., 2012) and Swedish (Boschma et al., 2014) regions. In complementary research, Neffke and Henning (2013) and Boschma et al. (2014) argue that regions characterized by high shares of skill-related activities (i.e., sectors that rely on similar human capital resources) promote diversification and ease growth-enhancing inter-industry matching. Opportunities for exploiting existing place-specific assets in novel ways play a central role in new work on regional path creation and regional resilience (Boschma, 2015; Diodato & Weterings, 2015; Eriksson et al., 2016; Hane-Weijman, 2021; Pike et al., 2010).

A third element of research on employment change focuses on the characteristics of workers. While traditionally only having assessed the concentration of human capital in regions (e.g., Glaeser et al., 1992), Autor and colleagues offer new ways of measuring worker skills and tasks in a series of important empirical papers that focus on the performance of labour markets and workers in relation to various shocks (Acemoglu & Autor, 2011; Autor et al., 2003). Although occupation-oriented approaches have a long tradition in economic geography (e.g., Markusen 2004; Massey, 1984; Thompson & Thompson, 1985), research on the creative class placed the type of skills and occupations present in regions back on the regional growth agenda (Florida, 2004). A series of new papers combines this skill approach with the above-mentioned work on relatedness that is tied to employment and occupations. The lives and deaths of specific kinds of jobs are explored by Hasan et al. (2015) who report that technical interdependencies within clusters of work tasks generate job longevity. In this sense they capture aspects of the relatedness between occupations first outlined by Muneepeerakul et al. (2013). Building on the concept of related diversification introduced by Hidalgo et al. (2007) and developed within a regional setting by Neffke et al. (2011) and Boschma et al. (2013), Muneepeerakul et al. (2013) explore a bipartite network of the structure of employment by US urban areas and occupations. They link urban productivity variations to the occupational structure of cities and they report how occupational upgrading is dependent on that structure. Using similar techniques, Shutters et al. (2015, 2016) examine how US cities transition towards creative economies, providing additional detail to the broad claims of Florida et al. (2008), and showing how economic resilience can be quantified with occupational relatedness measures. Farinha et al. (2019) go further in unpacking the meaning of relatedness by combining employment and occupation data for US urban areas and distinguishing between occupational relatedness that is based on the similarity of skills, on complementarity along a value chain, and that which generates local synergies (cf. Duranton & Puga, 2004). An innovative recent paper by Davies and Maré (2021) refines measures of occupational relatedness and complexity using local area data from New Zealand and reports that complex employment practices experienced faster employment growth since 1981, especially in cities with a diverse range of complex activities, but that relatedness had little impact on employment growth.

We seek to build on components of the work above, especially the arguments of Frenken et al. (2007) and Davies and Maré (2021), to investigate the roles of relatedness and complexity on regional employment growth in Sweden. Following their work, we expect that regions generating net new jobs in more rather than less related occupations would enjoy faster employment growth due to the recombinatory potential and local synergies that arise (Frenken et al., 2007; Hane-Weijman, 2021). We also suspect that local labour markets that perform better in terms of adding jobs in more complex occupations will outperform those that add jobs in less complex occupations. One reason for this is that more complex, skill-intensive jobs generate broader multiplier effects in local labour markets (Moretti, 2013). A second reason is that complex activities tend to be less geographically mobile than less complex activities (Balland & Rigby, 2017; Sorenson et al., 2006) and thus jobs in complex occupations may be more embedded in local and regional economies, perhaps generating other forms of place-based competitive advantage. In terms of crisis, our theoretical priors are less precise because there is less published work that we can draw upon. However, Frenken et al. (2007) note that unrelated variety may dampen negative employment shocks in regions through acting as a portfolio effect. We do not measure unrelated variety directly, though the same effect might be captured in our analysis as regions building specializations in sectors where relatedness is lower than average. Whether the development of complexity raises or lowers employment growth in periods of recession is simply unknown. More work is required on this question, both theoretical and empirical.

DATA AND METHODS

To analyse how changes in the mix of occupations within regions influences employment growth, we use individual micro-data from Statistics Sweden. These data record the occupations of individual workers active in the Swedish labour market, their place of work (municipality) and industry affiliation. To define an occupation, the three-digit level of the Swedish SSYK96 occupation nomenclature is used (broadly consistent with the international ISCO-88). The period of analysis, 2002–12, is chosen for two main reasons. First, work by Henning and Eriksson (2021) identified the years after 2000 as a phase of significant polarization in Sweden, linked to changing occupational structures. Second, a revised occupational classification introduced in 2013 makes longer run comparative analysis difficult.

The data include all workers with an occupation code who receive their main income from paid employment. Our focus is on the regional dimension of labour market processes. The spatial units examined are based on functional economic areas (FA regions) developed by the Swedish Agency for Economic and Regional Growth (2011). These are created by aggregating municipalities into 72 FA regions based on a combination of observed commuting flows and labour market analyses. As shown in previous studies on Sweden (e.g., Boschma et al., 2014), these spatial units mitigate the risk of spatial autocorrelation as they describe labour market areas where workers both reside and work.

Regional trajectories: distance and direction of regional transformation

The regional branching literature rests upon the idea of change (or transformation) as being related or unrelated in different degrees to the existing economic structure of the region (Boschma, 2017). This framework is operationalized in the following way. First, regional transformation is captured by changes in the occupational specialization of a region. By convention, such changes are measured by temporal shifts in the pattern of regional comparative advantage (RCA). RCA is another term for a location quotient and rendered binary (0/1) depending on whether a region has a higher share of employment in an occupation than the Swedish average. A region (r) is said to be specialized in an occupation (i) at time (t) if: (1) ${\displaystyle \frac{{\mathrm{E}\mathrm{m}\mathrm{p}}_{r,i}^t/{\sum }_i{\mathrm{E}\mathrm{m}\mathrm{p}}_{r,i}^t}{{\sum }_r{\mathrm{E}\mathrm{m}\mathrm{p}}_{r,i}^t/{\sum }_r{\sum }_i{\mathrm{E}\mathrm{m}\mathrm{p}}_{r,i}^t}>1}$(1)

All 112 occupations identified in Sweden’s occupational classification (excluding military) are categorized as specialized or not for all functional labour markets each year.

In a second step, the distance or relatedness between occupations is measured using the notion of skill relatedness based on relative risks (Neffke et al., 2017), where the observed flow from occupation i to j (F_ij) for the whole study period is divided by the expected flow (the product of all inflows to occupation j and outflows from occupation i, divided by total flows): (2) $R_{ij}={\displaystyle \frac{F_{ij}}{({\sum }_{kj}{F}_{kj}\ast {\sum }_{ki}{F}_{ik})/F}}$(2)

The relatedness values between occupations range from 0.00001 to 0.82606, with a mean of 0.00314. High relatedness values indicate that two occupations are closely related or that the worker flows between these occupations are larger than expected. High relatedness values between occupations may be thought of as indicative of skill- and task-based similarities between them. Low relatedness values indicate that a pair of occupations has little overlap in terms of the nature of work. However, we are interested in the dynamics at the regional level. Therefore, we calculate (1) the average relatedness of the occupations in which the region is not specialized; (2) the average relatedness of all new specializations that they enter; and (3) we compare these two values and aggregate them. The resulting measure indicates whether a region is branching into new specializations that are more/less related to the region’s existing capabilities out of those available. The same is done for exiting specializations by comparing the relatedness of specializations that are abandoned in relation to those that previously existed in the region.

In a third step of the analysis, the complexity values of the occupations are generated. We use the complexity measure of reflections proposed by Hidalgo and Hausmann (2009) for products and adapted in matrix form by Balland and Rigby (2017) and Davies and Maré (2021). The complexity measure combines two components: (1) the diversity of the regional occupation mix; and (2) the ubiquity of occupations. An occupation is defined as complex if relatively few regions are specialized in that particular occupation and if this relatively rare occupation is typically found (RCA = 1) in regions that are diverse. It is constructed from a bipartite network of regions and occupations represented in matrix form in a binary adjacency matrix M that has dimension 72 × 112. After row standardizing matrix M along with its transpose ${\mathit{M}}^{\mathit{T}}$, we find the product $\mathit{D}={\mathit{M}}^{\mathit{T}}\mathbf{\ast }M$. The second eigenvector of the square-matrix D yields the complexity values for all 112 Swedish occupations.

There are competing measures of complexity (Fleming & Sorenson, 2001; Mewes & Broekel, 2020) that have different properties. It remains unclear if any one of these measures is clearly superior to others. He et al. (2016) show that the method of reflections generates a bipartite page rank and thus an eigenvector centrality-based ranking. If occupations are to be understood as bundles of job tasks that are highly correlated with the education and skills of workers (Autor & Handel, 2013), we see the complexity of occupations as embodying the value of knowledge that the workers within occupations add to the production process. Lo Turco and Maggioni (2020) provide empirical support for this view. For complex occupations to be conducted in particular regions requires a diverse range of activities distributed across other occupations, many of them also complex, to be enjoined in the same place. Less complex occupations rest upon a narrower assemblage of related activities and thus may be found in more numerous labour markets in Sweden as elsewhere. Imagined in this way, employment in more complex occupations generates spillovers that support employment elsewhere in the economy, in the way that we might think of an employment multiplier, but where employment in different occupations is weighted by the complexity of those occupations. Thus, we expect occupational complexity to be related to employment growth within labour market regions, notwithstanding the flows that clearly occur between such regions. For a list of the most complex and least complex occupations together with the occupations they are most related to, see Tables A1 (most complex) and A2 (least complex) in Appendix A in the supplemental data online.

Before turning to the results, a note on context is warranted. After a long period of regional convergence during the post-war period (Enflo & Henning, 2016), Sweden faced a period of increasing regional divergence following the major macroeconomic recession in the beginning of the 1990s, when gross domestic product (GDP) fell by 6% and unemployment rose from 1.5% (1989–90) to 8.2% in 1993 (Magnusson, 2002). In addition, there has been a significant change in the internal structure of regional economies (Henning & Eriksson, 2021). Following Metcalfe et al. (2006), Figure 1 shows the industry mix and occupation mix of all Swedish FA regions (based on the shares of employment in respective two-digit industry and three-digit occupation) in 2002, correlated with the same variables in each subsequent year from 2003 to 2013. If growth was proportional, these coefficients would be stable over time. However, this is not the case as the two lines are decreasing monotonically, meaning that new jobs are created in certain industries/occupations and destroyed in other industries/occupations. Of particular interest for this study is that the correlation of the industry mix is decreasing faster than the correlation of the occupation mix. This implies that where people work and what they are producing are changing over time, but what people are doing remains more stable.

Figure 1. Correlation of industry and occupation mix 2002 with each subsequent year.

THE OCCUPATION SPACE AND THE COMPLEXITY OF SWEDISH OCCUPATIONS

Figure 2 shows the Swedish occupation space. Nodes represent different occupations, node size indexes their relative complexity and the colours represent the one-digit aggregate occupation classes. Nodes that are relatively close to one another have vectors of relatedness values that are highly correlated. Nodes that are distant have relatedness vectors that are uncorrelated. Neighbouring nodes thus represent occupations that have many similarities in terms of skills and work tasks. Only the strongest links (the top quartile of edges) are shown. Looking across the regional data for Sweden (for examples of occupational spaces for different regional types, see Figures A1a–c in Appendix A in the supplemental data online), in general, the larger a region in terms of population, the more complex are the occupations in which the region is specialized (pairwise correlation between average complexity and regional size = 0.82). Moreover, it is clear from the comparisons that the occupational mixes of different regions vary in terms of the skills and competences employed in the production of goods and services.

Figure 2. Sweden’s occupation space.

As well as generating complexity scores for different occupations, we can do the same for each Swedish labour market. Table 1 illustrates the top of the complexity ranking for Swedish labour-market areas in 2002 and 2012. Unsurprisingly, the three metropolitan regions dominate the complexity ranks for both years. These three labour markets contain a relatively high share of workers in the most complex occupations, a factor that demands explicit attention in the analysis below. There is considerable stability in the labour-market complexity ranks over time. Only four labour markets that were in the top-10 ranking in 2002 fell out of the top-10 ranks in 2012, and three of these regions did not fall very far. In similar fashion, the four regions moving into the top-10 complexity ranks by 2012 did not move a great distance in the ranks.

Table 1. Complexity ranking for Swedish labour-market areas, 2002 and 2012.

Top ranks	Ranking 2002	Ranking 2012
1	Stockholm	Stockholm
2	Gothenburg	Gothenburg
3	Malmo	Malmo
4	Sundsvall	Almhult
5	Almhult	Sundsvall
6	Boras (12)	Haparanda (16)
7	Umea	Gotland (21)
8	Vaxjo (15)	Umea
9	Karlskoga (42)	Arjeplog (25)
10	Ostergotland (22)	Kiruna (18)

Note: Values in parentheses show ranks for the other period.

It is instructive to explore how Swedish labour markets have adjusted their occupation mix over time in terms of both relatedness and complexity. Do regions branch into occupational specializations that are similar to existing regional capabilities and/or into more complex occupations? These are questions at the core of the Smart Specialisation literature within the EU (Balland et al., 2019; Foray et al., 2011; McCann & Ortega-Argiles, 2013). The answers to these questions are revealed in Figure 3, which maps Swedish labour markets in a ‘Smart Specialisation space’ for occupational entry and exit during the study period (Rigby et al., 2019). The entry space of Figure 3a indicates whether regional labour markets have tended to add specialization in occupations that are above/below average in terms of relatedness and complexity relative to the region’s base-year occupational structure. The exit space of Figure 3b provides the same information for labour markets in terms of losing specialization in particular occupations.

Figure 3. (a) Smart Specialisation spaces (entry) for Swedish labour markets, 2002–12; and (b) Smart Specialisation spaces (exit) for Swedish labour markets, 2002–12.

The entry space in Figure 3a shows that the three large metropolitan areas of Stockholm, Gothenburg and Malmo developed specializations in occupations with complexity values greater than their respective regional averages over the period 2002–12. In Stockholm and Gothenburg, these new occupations were relatively unrelated to the existing skill- and work-task mix of the regions, while in Malmo entry into more complex occupations was consistent with the region’s existing occupational structure. Smaller labour market areas in Sweden are widely distributed over the four quadrants of the Smart Specialisation space. For Sweden’s larger regional centres, the three northernmost areas (Sundsvall, Umea and Lulea) have all been adding occupations with relatively high levels of complexity, while a region like Halmstad has moved in the opposite direction.

The situation is not that different when we turn to the loss of specialization across Swedish regions (Figure 3b). More than half of Sweden’s labour markets have been losing specialization in occupations that are more complex than the average (these are the locations above the dashed line on the complexity axis). The top-right quadrant of the exit space also reveals that a good number of these regions lost specializations in occupations that are closely related to their respective occupational cores. This pattern is inconsistent with a strategy of building competitive advantage around existing capabilities and we suspect that it reflects a set of dynamics driven, at least in part, by the recession in 2008 and the loss of regional core competencies that no longer are in demand. The three large metro areas have moved out of specializations with lower complexity than the average, and in the cases of Stockholm and Gothenburg, these specializations were also in occupations with lower relatedness to the regional occupational cores. Like Davies and Maré (2021) and Kuusk and Martynovich (2021), we show that labour markets of different scale and structure evolve in distinct ways in terms of the relatedness and complexity of economic activities.

HOW DOES THE OCCUPATION SPACE CONDITION REGIONAL DEVELOPMENT?

With the information from Figure 3, we now seek to understand whether Swedish regions that adjusted their occupational mix in a fashion consistent with the broad tenets of Smart Specialisation – entering more related and complex occupations and exiting less related and complex occupations – performed better or worse than average in terms of employment growth over the period 2003–12. Annual employment data for the 72 labour markets form a panel used to examine the question posed. The key independent variables are derived from Figure 3, representing annual shifts within each region of the relatedness (RLTDN) and complexity (CMPLXT) of occupations added to the region (ENTRY) and removed from the region (EXIT). Analysis focuses on whether occupational relatedness or complexity is the more important driver of employment growth, and on how the relatedness and complexity variables impact changes in employment in periods of growth (pre-2008) and periods of crisis (post-2007).

Additional covariates, which theory links to employment growth, are added to the panel model. These span the three broad approaches to employment dynamics discussed above. A first covariate tracks the employment share of small- and medium-size establishments (SME_SHARE), reflecting the importance of SMEs (fewer than 250 employees) to employment change. A second set of covariates captures the influence of economic structure, including the regional employment share in the public sector (PUBLIC_SHARE), and in manufacturing (MANUF_SHARE), as well as the number of occupational specializations in a labour market (N_RCA). The number of specializations (occupations with RCA = 1) captures the influence of occupational diversity and varies with labour-market size. Public sector employment is relatively high in some (mainly less prosperous) regions of Sweden and it is less prone to cyclical swings than other sectors. The opposite could be said about manufacturing, which dominates employment in many smaller regions. Given the focus on employment growth, the lagged level of employment (EMPLOYMENT) for each labour market area is added to the model, providing evidence of convergence or divergence in regional job growth. A final covariate (HIGHED_SHARE) captures the share of employed workers in each region with tertiary education. The variable RECESSION is a binary variable that takes the value 0 for all years before 2008, and 1 for 2008 onwards, denoting years of normal economic growth prior to the financial recession and years of slower growth during and afterwards. N_RCA and EMPLOYMENT are log-transformed due to skewness. See Table A3 in Appendix A in the supplemental data online for descriptives and a correlation matrix.

The fixed-effects model specification is particularly suitable for the task at hand, permitting investigation of within-region shifts in independent variables and their impact on employment growth. The primary advantage of the fixed-effects approach is that it negates concerns with unobserved regional attributes (e.g., local institutions), at least insofar as they may be considered constant over time. The Hausman test supports our use of fixed- over random-model specifications. Given that our observational units are regions we take explicit account of spatial dependence in the data. In a test of spatial dependence, the geography of Sweden’s 72 labour-market areas was captured using inverse distance weights between the full set of locations. These distance weights were built from coordinates mapping the centroid of each area. Using these distance weights to compute Moran’s index, no significant spatial autocorrelation in employment growth rates across the years investigated was found. Furthermore, employing the spatial linear model (splm) package in R, there was no evidence of significant spatial lag or error terms in spatial panel models of annual employment growth using a base model consistent with those reported in Table 2. Further, since most economic variables are correlated endogeneity is a concern. While we do not employ instrumental variables approaches, we do lag all continuous explanatory variables one year to dampen concerns with simultaneity bias (implying that entries and exits are regressed 2003–12 on employment 2004–13). Finally, year dummies are included in the model to control for time-specific heterogeneity that is not place specific and which is shorter in duration than the multi-year recession dummy.

Table 2. Fixed-effect (FE) models of the influence of occupational structure on the rate of growth of regional employment in Swedish regions, 2004–13.

	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
RELATEDNESS (RLTDN)	0.0060** (0.0027)			0.0089* (0.0045)
COMPLEXITY (CMPLXT)	0.0036 (0.0046)			−0.0030 (0.0075)
ENTRY RLTDN (RLTDNE)		−0.0002 (0.0033)	−0.0014 (0.0030)		−0.0024 (0.0082)	−0.0070 (0.0080)
EXIT RLTDN (RLTDNX)		−0.0118*** (0.0042)	−0.0090** (0.0038)		−0.0180*** (0.0060)	−0.0168** (0.0075)
ENTRY CMPLX (CMPLXTE)		−0.0012 (0.0061)	−0.0034 (0.0050)		−0.0132 (0.0109)	−0.0180* (0.0102)
EXIT CMPLX (CMPLXTX)		−0.0067 (0.0052)	−0.0078 (0.0074)		−0.0035 (0.0076)	−0.0052 (0.0113)
RECESSION (RCSSN)	−0.0895*** (0.0319)	−0.0899*** (0.0322)	−0.0971*** (0.0034)	−0.0872*** (0.0326)	−0.0879*** (0.0328)	−0.0955*** (0.0037)
RCSSN*RLTDN				−0.0054 (0.0045)
RCSSN*CMPLXT				0.0121* (0.0073)
RCSSN*RLTDNE					0.0032 (0.0095)	0.0091 (0.0099)
RCSSN*RLTDNX					0.0115** (0.0056)	0.0151* (0.0085)
RCSSN*CMPLXTE					0.0218* (0.0110)	0.0259** (0.0125)
RCSSN*CMPLXTX					−0.0053 (0.0067)	−0.0036 (0.0084)
EMPLOYMENT	−3.98E–06 (3.18E–06)	−4.01E–06 (3.18E–06)	−0.0000** (5.75E–06)	−4.33E–06 (3.23E–06)	−4.35E–06 (3.24E–06)	−0.0000** (5.75E–06)
MANUF_SHARE	−1.4896** (0.7258)	−1.4856** (0.7282)	−1.5402* (0.8001)	−1.3991* (0.7325)	−1.3910* (0.7270)	−1.4339* (0.7744)
PUBLIC_SHARE	−0.3973 (0.3615)	−0.4476 (0.3661)	−0.4576 (0.3592)	−0.4330 (0.3613)	−0.4651 (0.3697)	−0.4750 (0.3645)
N_RCA	0.0018 (0.0023)	0.0019 (0.0023)	0.0014 (0.0023)	0.0019 (0.0023)	0.0019 (0.0023)	0.0014 (0.0023)
HIGHED_SHARE	2.0186** (0.9243)	1.9972** (0.9227)	2.8447*** (0.9724)	1.9861** (0.9367)	1.9629** (0.9342)	2.8294*** (0.9896)
SME_SHARE	0.2019 (0.1722)	0.2075 (0.1745)	0.1792 (0.1607)	0.2363 (0.1763)	0.2297 (0.1739)	0.1988 (0.1596)
CONSTANT	0.9676*** (0.2395)	0.9720*** (0.2369)	0.2007 (0.3759)	0.9620*** (0.2363)	0.9714*** (0.2317)	0.2007 (0.3759)
Reg and year FE	Yes	Yes	Yes	Yes	Yes	Yes
R^2 (within)	0.1583	0.1617	0.3256	0.1644	0.1699	0.3360
Observations	648	648	621	648	648	621

Note: Coefficients (standard errors) are shown. Significance levels are posted as *p < 0.10, **p < 0.05 and ***p < 0.01. All models incorporate robust standard errors. Models 1–2 and 4–5 are run with all 72 functional metropolitan areas; models 3 and 6 exclude the metropolitan regions of Stockholm, Gothenburg and Malmo.

Models 1–3 of Table 2 present results from estimating different models of local employment growth using increasingly disaggregated measures of occupational relatedness and complexity. Models 4–6 follow the same form as models 1–3, adding interactions of the recession dummy with the relatedness and complexity terms. These interactions reveal whether relatedness and complexity measures operate differently between periods of growth and crisis. Models 1–2 and 4–5 are estimated across all 72 Swedish labour market regions, while models 3 and 6 drop the three metropolitan areas of Stockholm, Gothenburg and Malmo. Separating the performance of smaller labour-market areas in Sweden from activities in the three large metropolitan regions is important. First, there is more inertia to the occupational structure in the larger metropolitan areas. Second, we do not wish our overall results to be driven by processes found only in metropolitan regions that have been characterized as leading to increases in the geographical polarization of employment (Eriksson & Hane-Weijman, 2017) and incomes (Enflo & Henning, 2016). There are few differences in the results between models 2 and 3 and between models 5 and 6, and so we are confident that the results are not driven by the larger metros. Before turning to the main results, note that the control variables perform consistently across all models.¹

Turning to the core variables, model 1 shows that the partial regression coefficient on occupational relatedness is positive and significant. Occupational relatedness here is the sum of its entry and exit components. This result is broadly consistent with the findings of Frenken et al. (2007). While positive, the coefficient on the aggregate complexity term is insignificant. In model 2, we separate the relatedness and complexity terms into their occupational entry and exit components. While relatedness overall has a positive influence on employment growth within Swedish labour markets, the results of model 2 show that this finding is driven more by occupational exit rather than entry. Combined, these results imply that the positive influence of relatedness on employment growth in Swedish labour markets occurs largely via regions stripping away less related occupations and focusing on their core labour market capabilities. The results from model 3 are consistent. Neither the entry nor exit terms for occupational complexity are significant in models 2 and 3, when all time periods are considered together.

A more nuanced set of findings appears when we separate the influence of occupational entry and exit on employment growth across periods of growth and recession. These results are generated by interacting entry and exit variables with the recession dummy (0/1). Model 4 reveals that the influence of relatedness on the rate of growth of employment was positive prior to the recession and that this effect did not change significantly after the onset of the financial crisis. Model 4 also reports that while complexity had no significant influence on employment growth before 2008, there was a significant positive change in the impact of aggregate complexity on employment growth after 2007. That said, this shift in the value of the complexity coefficient from before the crisis to its onset, was not large enough to turn the complexity coefficient significant for the period after 2007. Thus, while we cannot claim that the development of more complex occupations during the crisis played a significant role in limiting the decline in employment growth, we can report that on average across Swedish labour markets there appears to be a significant change over time in employment related opportunity costs of moving into more complex occupations.

Disaggregating the effects of relatedness and complexity in terms of entry (building new specializations) and exit (departing existing specializations) before and after the financial crisis offers further insights regarding employment growth. According to model 5, the coefficient on the interaction between the relatedness exit variable and the recession dummy is positive and significant. Thus, there is a significant difference pre- and post-crisis in terms of the effect of exiting unrelated occupations on employment growth. Examining the data for the post-crisis years further shows that after 2007 there is no significant impact of exiting occupational specializations on employment growth.

While changes in entry/exit complexity have no influence on the dependent variable in the pre-crisis period, through the crisis the coefficient on entry complexity changes significantly in a positive direction. This indicates that compared with periods of normal economic growth, developing specializations in more complex occupations has a positive influence on employment growth. Again, however, while positive in sign, this interaction effect is not strong enough to turn the influence of entering complex occupations significant for the period after 2007. Dropping the metro regions does not alter these findings in a major way, though we note that the coefficient on entry complexity is negative and significant in the pre-recession period in model 8. Thus, moving into more complex occupations appears to have been riskier in terms of employment growth outside the major metropolitan labour markets in Sweden, as we would expect.

Robustness checks involving the use of additional covariates (population density, income levels, etc.) yielded no broad changes to our findings, but did raise problems of collinearity. Concerns with annual data led to experimentation with multi-year periods of change and use of three-year running means of all right-hand-side variables. These experiments caused some variation in the relative magnitudes of core variables, but no appreciable change to the general story outlined above.

CONCLUSIONS

The primary objective of this paper was to explore whether the concepts of occupational relatedness and complexity were useful in terms of understanding regional employment growth. In addition, we examined whether those concepts operated differently in periods of normal growth and in periods of crisis, following the initial ideas of Frenken et al. (2007) that variations in the structure of regional economies produce different effects. A secondary aim was to contribute to the growing literature on relatedness and complexity and to extend that literature into the domain of occupations where there has been relatively little applied work to date. In this way we respond to the calls of Boschma (2017) and Whittle and Kogler (2020) to broaden work on relatedness. We add to that work in another way, following Neffke and Henning (2013) and Neffke et al. (2017), using worker flow data between occupations to build our relatedness measure, but examining different questions. This methodological choice is important for co-location measures of relatedness remain somewhat opaque.

Our main findings report how occupational relatedness and complexity changed in Swedish labour markets between 2002 and 2012, and how these changes impacted subsequent local employment growth rates. The results indicate that over the study period as a whole regions developing more related occupations experiencing faster employment growth, while the complexity of occupational specializations has no discernible impact. Disaggregating these effects, the relationship between relatedness and employment growth was driven in large part by labour market exits in relatively unrelated occupations and thus by consolidating employment around core sets of related occupational capabilities. Greater nuance was added to the findings by splitting the analysis into a period of normal economic growth before the 2008 financial crisis, and a subsequent period of recession. The positive effect of exit from unrelated occupations on employment growth was limited to the pre-crisis years. Across the growth and recession periods, the results show a significant positive change in the relationship between entry into complex occupations and employment growth. However, this effect is not large enough to produce an overall significant positive relationship between occupational complexity and employment growth during crisis years. There is also evidence, at least outside the major metro regions in Sweden, that entry into more complex occupations lowers employment gains in periods of economic expansion. This effect is mitigated after 2007. In terms of complexity, it appears that the opportunity cost of building more complex occupational specializations in local labour markets are lowered in periods of crisis. More work is clearly required to understand how the relationship between the relatedness and complexity of occupations and employment growth change over time and under what conditions they might exert a significant influence on growth in regional employment.

These results provide a much more fine-grained analysis of the processes through which relatedness and complexity impact regional economic performance than is usually provided in the literature, unpacking the importance of processes of entry and exit over time. In line with Moro et al. (2021), we argue that policy makers should focus investments on occupations that are well embedded in the local labour market in order to maintain growth. This may be especially important in periods of economic expansion. While earlier research has focused on targeting new growth paths as a core policy tool, our results suggest that investment in existing occupational specializations within regions maybe is more effective. Indeed, outside the three main metropolitan areas in Sweden, building new specializations in complex occupational classes negatively impacted employment growth prior to the financial crisis. Resilience in this case appears to be less about creating new possibilities than maintaining core strengths. As labour market conditions shift from a period of relative fast growth to a period of slower growth, policy must adapt as the relationships between complexity and relatedness and regional employment change. This result raises questions for those building Smart Specialisation and related policy tools around the concepts of relatedness and complexity.

A series of more specific policy issues also emerge from our investigation. New research in Sweden and elsewhere shows that economic turbulence triggers periods of intense job switching (Hane-Weijman, 2021) with redundant workers often pushed into unrelated employment (Eriksson et al., 2018), experiencing skill mismatch (Nedelkoska et al., 2015), underemployment (Hane-Weijman, 2021) and poor-wage development (Holm et al., 2017). To mitigate polarization and a rising discontent arising from such processes (cf. Rodríguez-Pose, 2018), it is vital that policies target place-sensitive retraining and educational investments in smaller and/or peripheral regions. Whether such policies should foster the ability of local labour markets to branch into complex and related activities remains an open question.

ACKNOWLEDGEMENTS

The authors are grateful for the comments made by Zoltán Elekes on a previous version of the paper. They also thank the editor and two anonymous reviewers whose comments substantially improved this paper. Any remaining errors are the responsibility of the authors.

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

Additional information

Funding

This paper was financed by the Marianne and Marcus Wallenberg (MMW) Foundation [grant number 2017.0042].

Notes

1. The recession dummy is always negative and significant. On average across all labour markets, the average annual compound rate of growth of employment was 0.089% lower after 2007 than before 2008. The lagged value of employment is negative and not significant, except in the case where the metro areas are removed from the sample, when the variable turns significant at the 0.1 level. This indicates that there is some convergence in employment volumes only outside the metro regions of Sweden. In all models, the manufacturing share and public sector share of employment are negatively related to employment growth, though significantly only in the case of manufacturing. The number of specializations found in the labour market areas is not a significant determinant of employment growth and neither is the SME share of jobs. The share of the workforce with a high education is significant and positively related to employment growth. The value of the coefficient on the high education share rises as we drop metro areas from the data.