Human capital spillovers in Dutch cities: consumption or productivity?

It is widely recognised in the literature that, in addition to private benefits from human capital accruing to individual workers, there are social benefits to having a sizeable pool of individuals with high human capitals within a region (Lucas 1988; Rauch 1991). It is argued that the beneficial effects of human capital are higher in dense economic environments such as cities because face-to-face interactions between individuals are less costly and happen more frequently. Such interactions speed the transmission of new ideas and knowledge in a regional economy. An overview of the theoretical work on the mechanism involved can be found in Duranton and Puga (2004). Moretti (2004a) gives an overview of empirical research on these human capital spillovers. Essentially, as workers become more productive in these urban environments, higher net wages will be paid by firms, thus attracting more workers. City land rents then increase to the point where an individual migrant faces the same real wage level across the economy, thereby bringing migration towards the city to a halt (Moretti 2004b). Through this process, the externality of human capital is internalised within the economy.

Human capital spillover is the mechanism through which the human capital stock exerts its effect on the economy. Assuming the process of internalisation described above, spillovers are often studied through the development of individual wages and land rent premiums in relation to regional aggregate human capital levels, using Mincerian frameworks (Rauch 1991; Acemoglu and Angrist 2000; Moretti 2004a, c; Glaeser and Maré 2001; Shapiro 2006). The wage effect of a relatively large supply of highly skilled workers consists of two components: a supply effect and a spillover effect. If complementarities between skill levels in the labour market are sufficiently high, and substitution between the skill levels is low, an oversupply of highly skilled workers will lead to lower wages (through lower marginal productivity) for this group. Higher wages for the other skills groups within the local economy can be expected as a result of the high complementarities. The spillover also serves to increase general productivity, for example through learning in dense environments. Any empirically measured effect of a region’s endowment of highly skilled inhabitants on skill-specific wage rates is therefore the sum of these two effects (Moretti 2004b). The positive effects on wages from the presence of a large stock of highly educated workers, as are generally found in the literature, are an indication that the spillover effect is sufficient to overcome any negative supply effect. However, greater insight is needed into the underlying mechanism. Broersma et al. (2015) show, for example, that, for the Netherlands, an important component of the wage spillover effect results from interactions between individuals within firms rather than interactions between individuals co-located within a region.

Wage and land rent signals may, however, be distorted by institutional issues related to the functioning of labour and housing markets such as central wage bargaining and planning restrictions on real estate development. Central wage bargaining negates regional differences in productivity levels. Constraints on local real estate development make the responses of land rents to productivity shocks hard to identify. A number of studies have therefore focused specifically on the effect of human capital endowments on either skill-specific or overall employment growth.¹

Human capital spillovers do not necessarily translate into employment growth. Combes et al. (2004) point out that technological development may be labour saving, or that labour supply may be inelastic. Suedekum (2009) notes that the employment effects of human capital spillovers may be driven by complementarities between skills groups. Also, a relatively skill-abundant local economy may, as a result of the negative supply effect on wage rates, attract fewer additional highly skilled workers (Suedekum 2009). A variety of studies have, however, demonstrated that high human capital endowments do have positive effects on subsequent city employment growth (Glaeser and Saiz 2004; Suedekum 2009; Glaeser et al. 1995; Shapiro 2006; for the Netherlands specifically see Marlet and Woerkens 2007).

Other studies have focussed on the effects on different skills groups. For Germany, it was found that the presence of highly skilled workers has positive effects on the employment opportunities for medium and lower skilled workers (Suedekum 2009; Schlitte 2010). Berry and Glaeser (2005), however, note an increasing accumulation of skilled workers in US metropolitan areas that are already well endowed with highly skilled inhabitants. Accumulation, or specialisation, in this context is studied by looking at the coefficient value of the lagged dependent variable in a model explaining levels or shares of human capital in cities. A value that is greater than one (the result in Berry and Glaeser 2005) implies an increasing concentration over time, but is also indicative of an explosive time series. Poelhekke (2006) noted that these results therefore need to be interpreted with some care: the share of skilled individuals cannot increase indefinitely, something the analysis by Berry and Glaeser (2005) suggests. Nevertheless, it was demonstrated that, in the period studied (1970–2000), skilled US cities became more skilled. This, however, was not due to the presence of skilled individuals as such. Rather, as Poelhekke (2006) concludes based on an analysis of the same Berry and Glaeser dataset, other city-specific factors seem to be in play.

Shapiro (2006) also notes that a positive relationship between human capital and city growth may be driven by other factors that affect both, such as amenities or the city’s industry structure. Glaeser (1999) has pointed out those cities with high endowments of amenities have grown faster than low-amenity cities. He notes that urban rents have increased faster than wages, suggesting that the desire to live in urban areas has increased for reasons beyond productivity increases. Poelhekke (2006) highlights the importance, as a pull factor in the localisation of skills, of the low-skilled service sector—i.e. the presence of local consumption opportunities. Bils and Klenow (2000) find a positive relationship between educational level and economic growth at the country level, but suggest that causality might in fact run the other way: that economic growth causes an increase in the average regional education level.

Spatial mobility may be a crucial factor in this respect. It has been demonstrated that regional human capital endowments can stimulate growth, as was discussed above. However, equally, human capital has been shown to flow to regions that are doing well (Duranton and Puga 2004; Faggian and McCann 2006, 2008, 2009; Carlino and Mills 1987). This spatial mobility of human capital has consequences for its relationship with regional economic development since structural outflows of human capital can negatively affect innovative capacity (Nijkamp and Poot 1998; Faggian and McCann 2009). Rodríguez-Pose and Vilalta-Bufí (2005) found that job satisfaction and migration measures contributed substantially to explaining regional GDP, over and above the more traditional measures of human capital endowments. Therefore, it seems that attention should be paid not just to existing endowments, but also to what attracts human capital to regions and whether these highly skilled workers are then able to utilise their skills in the regional economy.

Population and regional employment growth are thus found to be strongly related: directly through supply effects and spillovers, but also indirectly as a result of the aforementioned city-specific attractions and amenities. One line of literature has sought to more explicitly take these interrelationships into account, starting with Carlino and Mills (1987).² Following this approach, Boarnet (1994), Boarnet et al. (2005) and Gottlieb (1995) among others have pointed to the role that neighbouring areas play in employment and population growth. Boarnet (1994) highlighted spatial specialisation (working against living) in multi-centric metropolitan areas and finds that employment growth depends on the population growth in nearby residential areas. Gottlieb (1995) demonstrates that firms take the residential amenities in housing locations likely to be of interest to their staff into account in making location decisions, and that this extends to cities within commuting distance. Focussing on population growth, Partridge et al. (2007a, b) show that, for Canada, substantial, both positive and negative, effects may result from being close to larger metropolitan areas. In some cases, suburban areas are able to profit from providing residential opportunities for those working in the larger cities.

From this, we conclude that several factors are relevant in our study into the relationship between a city’s skills structure and the inflow of graduate human capital. Firstly, not only existing stocks but also the flows of graduate human capital are important in explaining regional development. Reverse causality may play a role, with human capital being attracted to successful regions and cities. These two variables are therefore best treated as endogenous. Secondly, it is important to control for factors that may drive both the skills structure and the inflow of graduate human capital. An important insight is that aspects that can be thought of as purely residential amenities may also affect the placement of employment. Thirdly, on the spatial scale of the Dutch city, we can expect strong interrelationships to exist between a city and other places, both close by and within commuting distances. In the next section we discuss how these matters were translated into our econometric approach.

In this study, we focus on the recursive relationship between a city’s employment skills structure and the inflow of graduate human capital onto the city’s labour and housing markets. Our analysis will be based on a sample of 287 Dutch cities, for which we have data spanning the period 1998–2008. In this section we outline our econometric approach. A discussion of our operationalisation and the dataset can be found in Sect. 3.

The main focus of our analysis is identifying the structural determinants of graduate inflow and a city’s skills structure. Therefore, in this paper, we analyse differences between cities, in terms of the spread of skills in total city employment and the inflows of human capital, rather than by considering annual growth rates.

In our modelling strategy, we have to be mindful of a number of potential econometric pitfalls. Firstly, the sought-after relationship is potentially obscured by endogeneity. In a number of similar studies, simultaneity was dealt with by lagging the exogenous variables by a considerable length of time. We model outcomes for the endogenous city-specific skills structure and graduate inflow using exogenous variables with ten-year lags.³ We also enter the lagged dependent variable into our models.

We apply a seemingly unrelated regression (SUR)⁴ model, taking into account the correlation that might exists between these equations as a result of endogeneity. We have applied the following specifications. Firstly we have an equation that defines \(\hbox {JobHigher/Lower}_{{i},{t}+10}\)—which is total city employment at higher and scientific skill levels, relative to the total number of jobs at medium, lower or elementary levels, in city i at time t:with i denoting the city index, \(i = 1,{\ldots },287\); and t the year of observation. \(\hbox {GrWork/Liv}_{{i,t}}\) denotes the lagged key explanatory variable. It is computed as the ratio of the inflow of recent graduate human capital into the city’s labour market to the inflow of recent graduate human capital into the city’s residential area. All other explanatory variables in the model are lagged to the base year and taken to be exogenous. We include a spatial lag of the key explanatory variable \((\hbox {WCGrWork/Liv}_{{i,t}})\) following Boarnet (1994) and Boarnet et al. (2005).⁵ \(\hbox {JobHigher/Lower}_{{i,t}}\) denotes the lagged dependent variable, which is intended to control for persistence in the city skill structure. Matrix \(X_{{i,t}}\) contains the other exogenous variables⁶ for city i, such as the education breakdown of the resident population—with the share of medium educated omitted, city size, labour and housing market characteristics and the number of graduates from a local higher education institution, plus spatially lagged values for a number of these variables. We also include year dummies.

The second equation describes the endogenous variable \(\hbox {GrWork/Liv}_{{i,t}+10}\), in which \(\hbox {JobHigher/Lower}_{{i,t}}\) is entered as main explanatory variable. The equation otherwise has a similar structure as the employment skills structure equation, including a lagged dependent variable, and the matrix \(X_{{i,t}}\) containing a set of additional exogenous variables:

In this section, we discuss the two main dependent variables in this analysis: city skills structure and the typology of the city in terms of inflow of graduates (leaning towards the labour market or towards the residential areas).

Traditionally, measures of human capital are based on the highest education level attained, or the years of schooling completed. However, as Schlitte (2010), among others, has pointed out, this measure does not take into account the effects of on-the-job learning or increased responsibilities as a result of tenure. Alongside this, as Di Pietro and Urwin (2006) point out, employers may increase the skills profile of their labour demands if skills are relatively abundant in their region, assuming that higher skilled employees will be easier to attract and can be trained at lower costs. The resulting patterns of over- and under-education mean that worker education level is not necessarily a good proxy for skill. As an alternative, the ‘International Standard Classification of Occupations’ (ISCO) job skills levels can be used as an indication of the skills structure of a local labour market. For the reasons outlined above, in this paper, we base our analysis of the city skills structure (endogenous variable \(\hbox {JobHigher/Lower}_{{i,t}}\)) on the ISCO classification, expecting it to better reflect the skills structure of the labour market.

We now turn to our second main variable, \(\hbox {GrWork/Liv}_{{ i,t}}\). In this study we measure a city’s ability to attract human capital by looking both at recent graduates, i.e. at around 18 months after graduation, starting a job on the city’s labour market as well as graduates moving into the city’s residential areas, as we are interested in disentangling production and consumption effects associated with these different flows. These graduates may have received their degree from a local college or university (i.e. they are retained) or elsewhere. Simply using the size of these flows or scaling them, for example using city population size, will not identify consumption or productivity spillovers over and above the baseline effect of city size since both residential and labour market oriented flows will be larger for larger cities (Venhorst et al. 2011a). Instead we have computed our second endogenous variable \(\hbox {GrWork/Liv}_{{i,t}}\) to generate the city-specific ratio of labour market inflow to residential inflow of graduates. This variable indicates whether a city’s inflow is relatively labour market dominated (values larger than one) or relatively strongly oriented towards residential areas (values below unity). Graduates who are working and living in the same city enter in the nominator as well as the denominator of this variable. A positive \(\beta _{1}\) coefficient for this variable in the \(\hbox {JobHigher/Lower}_{{i,t}}\) equation will then signal that cities with a labour market attraction profile will on the longer term exhibit a higher skilled job structure, much like the productivity spillover-driven developments in US cities discussed earlier. Likewise, cities with a more residential oriented inflow will exhibit a labour market skill structure oriented towards elementary to medium skilled workers, presumably in consumer services. A negative \(\beta _{1}\) coefficient signals the reverse. In this case, graduate inflows on city labour markets lead to opportunities for elementary to medium skilled workers, whereas cities with a graduate inflow dominated by residential motives feature more opportunities for higher and scientific-level jobs, likely in higher level services.

City-specific employment in terms of the skill levels demanded by the employers is derived from the Working Conditions Survey (WCS)⁷ published by the Dutch Ministry of Social Affairs and Employment. We use the School-leaver Information System (ROA-SIS) dataset⁸ from the Research Centre for Education and the Labourmarket (Maastricht University) and Statistics Netherlands data to compute the flows of graduates towards a city’s labour and housing markets (variable “GrWork/Liv”). Comparing the base values of our endogenous variables with the values ten years ahead, it becomes clear that, in our sample of cities, the ratio higher level jobs—lower level jobs has increased slightly. The mean value for GrWork/Liv has increased more substantially, from 0.68 to 0.78. This indicates that, at least in our sample of cities, the labour market inflows have on average become more important relative to the inflows into the residential market. This might imply that other smaller cities, not included in this study, over time have become more successful at attracting or retaining graduates to their residential areas.

In order to establish whether our results are driven a denominator effect of sorts, we have also estimated an alternative version of our SUR model, using JobShHigherSk rather than JobHigher/Lower. The variable JobShHigherSk simply measures the share of jobs in the city of higher or scientific level.

Based on the literature findings discussed earlier, we introduce a number of exogenous control variables alongside the time lags of the dependent variables. The exogenous variables in both equations include controls for city size, size of neighbouring cities, a set of controls for the city’s economic situation and amenity characteristics plus control variables describing the presence of universities. Below, we briefly introduce these variables.

A novel element in this study is separately controlling for the effects on the city’s skills structure of the skill levels required for the city’s existing stock of jobs (relevant to the productivity effects discussed earlier) and the education level of the city’s resident population (in relation to consumption spillovers). In the JobHigher/Lower equation, the lagged dependent variable controls for the existing stocks of jobs by the skill levels required. The education levels of the city’s inhabitants (variables “PopShLowEduc” and “PopShHighEduc”) are based on Statistics Netherlands data, using the three way (low, the omitted medium and high) ISCED classification. In this study we use the education level of the potential labour force (aged 19–65) by residential area, data which can be derived from the Statistics Netherlands Labour Force Survey. In addition to the effects of the consumption by higher educated residents on low-skilled service industries (Sassen 2001; Poelhekke 2006), Moretti (2004b) and Shapiro (2006) note that a higher educated resident population may uphold certain facilities in cities, alongside exhibiting higher rates of social activity. As such, it will be interesting to see whether the consumption effect of the skills structure of a city’s resident population is significantly associated with lower or higher skilled jobs, over and above the effect of the employment skills structure.

In both equations, we control for city size and the size of neighbouring cities. The question is whether the effect of the skills structure in general is significant, given the general demand effect of having access to a large market. In order to separate the effect of tastes and participation from the more general effective demand effect of higher wage levels for the higher educated population, we also enter the proportion of low income households (variable “ShLowIncHH”, again based on Statistics Netherlands data). Through this additional control, the variables measuring the resident population’s skills structure are ‘cleaned’ of any income effect and better reflect their different effects on the city’s skills structure, through tastes, participation or the relationship with the city’s facility structure. The income effect is itself indicative of the role a city’s purchasing power plays in determining its skills structure.

Earlier, from the literature, it was concluded that it is important to control for factors that simultaneously influence the presence of human capital, the city’s skills structure and employment growth (Shapiro 2006). In many studies, variables are added that reflect the economic structure of a city, alongside some assessment of amenity richness of the city or surrounding region. Often natural amenities, such as the vicinity of beaches, are used in this respect. The effect of such amenities on economic conditions has also been criticised (Storper and Scott 2009). Including amenities such as theatres and restaurants assumes that these are indeed valued by the group under study. To avoid this issue, we have decided to use somewhat more general measures reflecting housing costs and crime rates. In line with Gottlieb’s (1995) findings, we do not assume that residential amenities only affect the population since firms may well take residential amenities, in potential housing areas for their staff, into account when choosing a production location.

We use the LISA register of firms’ data to compute the average firm size (variable “FrmSize”) in a city. Larger firms are often assumed to be less economically dynamic (Blien et al. 2006) but may offer a wider range of employment opportunities for different skills groups. We also use the LISA dataset to compute the variable “Shock” that functions as a control for nationwide production shocks. Following Katz and Murphy (1992), we weigh national single-digit NACE sector-specific yearly employment growth by the city’s sectoral structure. In other words, we reweight the national growth rate by the city’s employment structure. This variable first controls for effects related to having a favourable, in terms of the business cycle, sectoral structure on the city’s skills structure. Given the growth in the Dutch service sector in the period studied, we would expect that this to largely translate into employment opportunities for the higher skilled. Second, it acts as a general control for a city’s industry structure. Data on housing prices (“LnHouseP”), crime rates (“Crime”) and unemployment rates (“Unem”) are all computed from Statistics Netherlands data.

We include the number of graduates that left university or colleges of higher education by city (the variable “Grads”). Naturally, this variable is 0 for all cities that do not have such institutions within their borders. This variable controls for the effect on the skills structure and graduate inflow of the presence of such institutions in cities.

Given the findings in Boarnet (1994) and Boarnet et al. (2005), we include spatial lags of some of the key endogenous and exogenous variables in our analysis. Florax and Folmer (1992) also recommended such an approach when faced with substantive spatial spillovers. The spatial scale of Dutch cities is relatively small and, as a result, strong interrelationships can be expected to exist between them. Boarnet et al. (2005) compare various spatially weighted matrix specifications in the context of modelling population and employment interactions. They conclude that there is no single best fit, and that the substantive interpretation of results is best served by using a spatially weighted matrix that leaves room for interpretation. In this study we therefore apply a row-normalised commuting matrix to compute the spatial lags of the relevant exogenous variables. We construct our row-normalised spatial interaction matrix using data on the 2009⁹ commuting flows, but including only those flows of a substantial size. The off-diagonal elements of the non-normalised base matrix were designated 0 if the flow was smaller than 100 commuting individuals and 1 if the flow was larger. Using a time-varying matrix with values reweighted according to flows would, in our view, introduces substantial levels of endogeneity in the specification, and therefore, after row-normalising, we do not reweight the spatially lagged variable according to flow size. Rather, the spatially lagged variable simply reflects the average value for municipalities with which a commuting relationship exists. All variables with the prefix ‘WC’ are computed using this matrix.

Our model found a positive \(\beta _{1}\) coefficient for the variable GrWork/Liv in the JobHigher/Lower equation. This implies that cities characterised by a relatively strong inflow of recent graduates onto their labour markets have higher shares of higher and scientific level jobs. Likewise, cities with a relatively strong inflow on their housing markets exhibit higher shares of elementary to medium share workers. This can be interpreted as production spillovers amongst the higher skilled and consumption spillovers between the higher and lower skilled. This in itself is in line with the findings for the USA (accumulation of the skilled in certain cities), but it is predominantly true for cities which are attractive work destinations for graduates. Consumption spillovers, likely pertaining to lower skilled service jobs, are thus found to be relevant as well, but in residential-type cities.

Looking at the reverse relationship, i.e. the effect of JobHigher/Lower on GrWork/Liv, we find a negative effect, which is only significant at the 10 % level. The result is confirmed in the Model 2 specification. Cities with wider job opportunities at the higher and scientific levels were found to be relatively well positioned to draw in graduate human capital for their labour markets (Venhorst et al. 2011a). In that light, this finding is surprising. The effect is (marginally) significant over and above the effect of having a large labour market and other city-level economic control variables. If these controls indeed capture the job attractiveness of the city, our result here serves to indicate that, ceteris paribus, smarter cities are indeed slightly more attractive as residential destinations.

We control for the inflow of graduates in neighbouring cities. In Model 1 we find a marginally significant negative effect on the city skill structure, which underlines the finding of accumulation of skills in a limited number of cities. However, this result is not confirmed in Model 2. We find a negative effect of this variable in the GrWork/Liv equation. This reflects the commuting system around cities, in that a high residential inflow in surrounding cities is associated with a stronger labour market inflow in the city under study, and vice versa.

Starting with the JobHigher/Lower equation, we have entered each city’s skills structure for the base year in order to control for the effect of existing human capital stocks. We find highly significant and positive effects for this lagged dependent variable. The interpretation here is that jobs of higher or lower skill levels tend to be concentrated in the same cities over time, even after controlling for the presence of universities and colleges and the sector structure.

The equation also features the breakdown of education levels in the resident population. Relative to the share of residents with a medium education level, we find positive effects for having a population with a higher education level (PopShHighEduc) on the city job skill ratio. This finding contrasts with the result for GrWork/Liv—here consumption spillovers do predominantly remain within the skill group (higher skilled services).

Turning to the graduate human capital inflow equation, we find that the lagged dependent variable is also positive and highly significant. It implies a degree of dynamic stability: a city’s ability to attract human capital to either its labour or its housing market depends strongly on having that ability in the past. We find no effects for the education structure of the resident population: having a relatively highly educated resident population as such does not lead to stronger residential or labour market inflows of graduates.

In this section, we discuss the results for the other control variables that were included in our models. We start with the effect of city size, before moving on to the variables pertaining to a city’s economic and residential amenity qualities and the presence of higher education institutions.