Estimating the Skill Bias in Agglomeration Externalities and Social Returns to Education: Evidence from Dutch Matched Worker-Firm Micro-Data

Abstract

This paper employs a unique set of micro-data covering almost one-third of the Dutch labor force, to estimate the heterogeneity of agglomeration externalities across education levels. This paper shows that there is substantial heterogeneity in the relationship between agglomeration and productivity of workers (proxied by their hourly wage) with different educational background. Apart from estimating the impact of the aggregate density of regional labor markets, we also estimate whether the composition of the local labor market in terms of education is related to the productivity of different types of workers. Using the presence of universities as an instrument, we estimate the effect of the supply of university graduates on wages, i.e. the social return to education. We find that agglomeration externalities are substantially higher for high- and medium skilled than for low-skilled employees. We find no positive effects from the presence of high-skilled on the productivity of low-skilled.

Appendix 1: Estimating Distance Decay Functions to Define Local Labor Markets

The size and shape of spatial units that is used to estimate regional patterns in economic outcomes matters a great deal for the research findings (Briant et al. 2010; Combes and Gobillon 2015). The work of Groot et al. (2014), who estimate agglomeration externalities on both the level of Dutch municipalities and on the level of NUTS-3 regions, provides a good illustration of this phenomenon. On the NUTS-3 level, they find that doubling the employment density is associated with 4.8% higher regional wages (controlling for worker heterogeneity), while applying the exact same methodology on the level of municipalities yields an estimated agglomeration elasticity of only 2.1%. Briant et al. (2010) find that the optimal choice for a regional classification depends on the spatial scope of the phenomenon under investigation, whereby the level of spatial disaggregation should match the level at which the forces under examination are expected to operate. In the case of agglomeration forces, this is the level of the local labor market.

A further problem is that data availability often limits the options that researchers have when choosing the appropriate regional classification. Even when the level of detail is sufficient, availabilities are often restricted to administrative units which may deviate substantially from what can be considered a regional labor market. Besides the fact that the average size of such administrative areas may not be appropriate, there is also a substantial heterogeneity in the (spatial) size of regional units, particularly on the level of municipalities. Another problem arises from taking a discrete approach to defining a regional classification: if two individuals are located just a few meters apart but on different sides of the regional border they are considered to be in different regions (or in our case local labor markets), while in reality there is no real difference in location.

In this paper, we have therefore chosen to consider the relevant local labor market for an economic actor at a certain location as a continuum. The farther away from the core of each individual actors local labor market, the less an area is considered to be part of the relevant local labor market. Following Thompson (1965) and Horan and Tolbert (1984), we conceptually define local labor markets as the area around an economic core where labor market transactions generally take place, which is bounded by the radius within which most of the commuting towards the core takes place. A straightforward operational definition that follows from this theoretical definition is to consider the extent to which an area at a given radius from a location where economic activities take place is part of the relevant local labor market (in other words, the spatial weight of the area at that radius) to be equal to the cumulative distribution function of the fraction of commutes that take place within that radius or further. Thus, as 100% (10%) of commutes takes place within a radius of 0 km (50 km) or more, we apply a spatial weight of ‘1’ (‘0.1’).

To formalize this relationship we have estimated a distance decay function. After experimenting with different functional specifications with up to three parameters, we found the following functional form to match the cumulative distribution function of observed commuting patterns almost exactly,

$$w(r) = \alpha + \beta \cdot \ln (r)\quad {\text{for}}\quad r > 0\quad \& \quad w > 0,\quad w(r) = 0\quad {\text{for}}\quad r = 0,\quad w(r) = 1\quad {\text{for}}\quad w > 1,$$

(3)

whereby w is the spatial weight of an area at a radius of distance r from the economic core. In the micro data that is available for this paper, we have both the residence and work municipality available for almost all Dutch employees (see Sect. 3 for a discussion of our data), as well as the x and y coordinates of the center of each municipality. Using OLS to estimate the above equation resulted in parameter estimates of α = 0.9385 and β = – 0.2219. The relationship between the spatial weight and distance to the core municipality of the average Dutch local labor market is presented in Fig. 4. Even though 50% commutes less than 7 km, around 10% of all commuters live more than 40 km from their jobs. At a distance of 68 km the estimated distance decay function crosses the horizontal axis. Even though there is a small percentage of commutes within our data that takes place at distances up to 200 km, the fact that these commutes account for less than 1% of total commutes supports the view that the estimated cutoff point is appropriate.

Fig. 4

Estimated distance decay function

Even though we could theoretically use actual the cumulative distribution function for each individual municipality as distance decay function, such a measure would be highly endogenous given our purpose of estimating agglomeration externalities. More productive regions characterized by high wages attract commuters from a very wide area compared to less productive and rural areas. Not in the least because of the increased demand for infrastructure that follows from these large commuting flows, there is generally more infrastructure connecting the large economic centers which results in better accessibility, attracting even more commuters (for this reason, estimating distance decay functions based on commuting time rather than distance is also problematic). If the distance decay function would be based on actual commuting towards a given municipality, the size of the spatial units would affect the size of agglomeration externalities, which is—given the findings of Briant et al. (2010) very likely to result in biased estimates. Therefore, we use the same distance decay parameters for all regions in our sample.

Figure 5 shows spatial weights for the local labor markets around three—for the purpose of illustration arbitrary chosen—municipalities: Amsterdam, Groningen and Maastricht. The weights quickly decline to 30–40% and decline more gradually from there onwards.

Fig. 5

Weights (in %) of municipalities around three local labor markets