The spatial scope of human capital externalities and firms’ location in Brazil Open Access

Large cities may attract new companies because of agglomeration economies, which can offset the higher costs of space and commuting (Marshall, 1890; Carlton, 1983; Glaeser & Maré, 2001). These urban agglomeration gains can influence the locational choice of new companies due to the presence of specific skills of workers (better matching), productive interactions (learning possibilities), and sharing a wider variety of inputs and services (Duranton & Puga, 2004; Arauzo-Carod, Liviano-Solis, & Manjón-Antolín, 2010).

While acting to explain urban agglomerations, such sources of benefits supposedly present different spatial scopes of action (Rosenthal & Strange, 2003; Lavoratori & Castellani, 2021). For instance, while sharing a diverse range of resources and having access to workers with specialized skills may benefit all areas within a metropolitan region, the advantages of human capital spillovers are expected to primarily benefit the locations closest to where these skills are concentrated within cities (Rosenthal & Strange, 2020; Fu, 2007; Andersson, Quigley, & Wilhelmsson, 2009; Balsmeier, Fleming, & Lück, 2023). Analyzing the spatial scope of urban gains from agglomeration, however, requires highly detailed spatial information within cities. This may explain the lack of empirical exploration in the literature, even in developed countries (Rosenthal & Strange, 2008; Di Addario & Patacchini, 2008; Eppelsheimer, Jahn, & Rust, 2022).

As for developing countries, the available evidence on the spatial scope of agglomeration economies is even much more recent, incipient, and generally does not consider specific sources of agglomeration benefits (see Li, Li, & Liu, 2022; Campos & Azzoni, 2021; Almeida, Silveira Neto, & Rocha, 2023a, b). These few studies have shown that the spatial attenuation of agglomeration gains for China and Brazil is quicker as compared to developed countries. This initial evidence indicates that localization effects are stronger up to 1 km and generally disappear after 5 km, probably due to the poorer infrastructure of cities in developing countries. However, none of the existing studies on Brazil addressed the impact of human capital agglomeration on firms’ location decisions. It is important to emphasize that this represents a critical and understudied aspect, as such effects may influence firms’ total factor productivity rather than solely the productivity of the labor force.

In this paper, we empirically investigate the spatial scope of agglomeration economies stemming from the external benefits of human capital in Brazilian metropolitan regions. Unlike previous studies that have focused on wages and overlooked the spatial variation of the external benefits of human capital within Brazilian urban areas (Chauvin, Glaeser, Ma, & Tobio, 2017; Quintero & Roberts, 2023), we estimate how the number of new manufacturing establishments in a particular urban area is affected by the presence of individuals with higher education at different distances from that location.

Besides contributing to the scarce evidence on developing countries, some Brazilian particularities make the study of the country's spatial attenuation of human capital externalities appealing. Compared to other major developing countries like China and Russia, Brazil imposes no restrictions on worker migration and this lack of restrictions have contributed to the country's remarkably high urbanization rate, which is around 85% (Chauvin et al., 2017). Rapid urbanization has led to mobility challenges in cities due to insufficient transportation infrastructure, which may impact the extent of human capital externalities as individuals could be reluctant to move between locations. Second, as recently highlighted by Almeida et al. (2022), manufacturing activity is more concentrated in Brazil than in China, Russia, and developed nations. This spatial concentration is stronger over shorter distances and correlates with the share of college-educated workers (Almeida et al., 2022) ^[1]. Thus, external gains related to human capital with a smaller spatial scope may help explain this evidence.

Given the need for spatially disaggregated information, unveiling evidence of the presence of agglomeration gains within cities is far from easy. This paper uses unique micro-geographic data for Brazilian manufacturing activities to provide the first set of evidence about the spatial scope of the human capital spillover influence on the location choice of new establishments. We estimate the local determinants of establishments’ births as a function of the concentration of workers with a college education in different distance bands and other economic and environmental characteristics. The task involves a significant empirical challenge, as confounding variables (observable or not) simultaneously may affect firms' location decisions and the concentration of qualified workers. The research considers different expedients to face it.

First, we work with a rich database including information at different spatial levels. We consider 1 km × 1 km cells within cities as unity of analysis, with some variables defined at cell, district ^[2], and municipal levels. This attenuates the potential influence of firms' sorting and allows controlling time-fixed observed and unobserved characteristics specific to districts. Second, we include a comprehensive set of control variables at cell, district, and municipal levels representing other sources of agglomeration gains, previously existing transport infrastructure, geographic characteristics, local labor supply, and local development policies that may influence the locational choices of new establishments. Third, we use instrumental variables for the local numbers of college-educated workers thought a control function approach. Similar to the one of Moretti (2004a), these variables are based on the general expansion of Brazilian schooling levels promoted by the federal government policies during the 1990s and 2000s and the different age cohorts of municipalities.

Our results indicate that the positive effects of human capital spillovers, measured by the variation in the number of workers with a college education degree at different spatial bands, are far from being homogeneous in the urban space. Such positive external effects are stronger up to 1 km from the location chosen by the firm and disappear completely beyond 5 km. Thus, our set of evidence is consistent with the idea that locations with more skilled workers are more attractive for new establishments due to human capital spillovers deriving from face-to-face worker interactions. More particularly, our evidence also indicates that such effects associated with local human capital vary according to the degree of technological intensity of the sectors; while for lower technology-intensive industries the agglomeration effects are less localized (occurring only up to 5 km), for more technology-intensive industries such effects are more localized (occurring up to 1 km).

The remainder of the paper is organized as follows. Section 2 presents our data sources and the empirical approach based on count models of births of establishments. Section 3 presents and discusses the results of the investigation and section 4 presents concluding remarks.

Our main source of information is the Annual Report of Social Information (Relação Anual de Informações Sociais, RAIS) produced by the Ministry of Labor. This is the more embracing source of information about formal establishments in Brazil and contains establishment-level information, including the number of workers employed and their characteristics, the sector of activity ^[3], opening and closing dates (if applicable), and establishments' address.

We use the information about establishments’ address to georeference the manufacturing establishments each year in the period 2006–2014 ^[4]. Furthermore, each establishment has a unique identifier, the number on the National Registry of Legal Entities (CNPJ), and this make possible dividing establishments into new entrants and existing ones in each year. The former group corresponds to nonexistent establishments in the previous years (identified by opening year) and the latter to existing establishments. Especially important for the study of agglomeration gains, more than 90% of new establishments each year represented a new company in the period 2007–2014 ^[5].

Our sample consists of plants engaged in manufacturing for two main reasons. First, in the case of the Brazilian economy, the manufacturing sector is the one with the least informality, which, since we cannot geocode informal plants, contributes to the sample representativeness ^[6]. Second, as confirmed in previous studies, manufacturing activities appear to benefit more from agglomeration economies (see, e.g. Barufi, Haddad, & Nijkamp, 2016) and human capital spillovers tend to be greater when the sectors are economically close (in terms of interactions), circumstance more common between manufacturing activities (Moretti, 2004b). We also are able to classify firms by technological intensity using the classification proposed by Hatzichronoglou (1997) for OECD countries and recent revisions (e.g. ISIC, 2011) and the compatibility with CNAE 2.0 made by Cavalcante (2014) (see Table A3 in Appendix A).

Using the geographic coordinates, we have thus allocated new establishments into a uniform grid of 1 km × 1 km cells. We use these cells as our spatial unity of analysis and consider as eligible for a new manufacturing establishment only those that are within the 30 metropolitan regions that existed in 2006 (see Figure A1 in Appendix A) and that have at least one establishment in their territory. Both procedures seek to mitigate potential non-observable influences associated with differences in urban structures and spatial heterogeneity.

To capture the human capital externalities at various distances, we follow a strategy similar to that used by Rosenthal and Strange (2008). Specifically, from the centroids of these exogenous cells, we define four concentric ring variables, each of which measures the number of workers with a college degree present at a given distance from the individual's workplace: between 0–1, 1–5, 5–10, and 10–20 km. The motivation for choosing the size of the concentric rings is related to the spatial extension of Brazilian cities. The smallest ring takes into account the potential effects happening in the immediate vicinity of the cell. The next two distance ranges, 1–5 and 5–10 km, cover the distances of most common commuting distances within core cities ^[7]. The further away ring, the 10–20 km distance one, also encompasses commuting from nearby cities to the core city and, thus, interactions at the metropolitan region level.

As discussed in the following subsection, in addition to information about firms from RAIS, we use a comprehensive set of information from other official sources to build our control variables in our regressions. At the level of cell, we use information about the location of transport infrastructures, from the Ministry of Transport, and river and lakes, from IBGE (Instituto Brasileiro de Geografia e Estatística), to measure geographic distances. We also use information at municipal level to build variables that control for the general productive environment. Specifically, this later set of information includes municipalities' population from IBGE; capital expenditures, housing and urban planning expenses and tax revenue from Ministry of Finance; export and import values from the Ministry of Development, Industry, Trade and Services; and homicide and traffic fatality rates from the Ministry of Health.

Our task is to estimate the spatial extent of human capital externalities within cities. We empirically implement this by analyzing the effects of the presence of college-educated individuals in different neighborhoods on the number of new establishments in those areas (the cells). Considering the nature of our dependent variable (a non-negative integer), we employ a Poisson model ^[8]. This model relates the number of new establishments to local factors that influence firm gains, including the number of workers with university education at different distances from the firms.

Specifically, we assume that new establishments choose a specific cell within the exogenously determined grid (cell z = 1, …, Z) and estimate the following Poisson model:

where Y_zct+2 is the sum of the number of new establishments in cell z in municipality c in the two periods ahead and the explanatory variable of interest is $Sr(z)t$⁠, the number of workers with a college degree or higher in each concentric ring r around the centroid of cell z. The specification includes control variables at the levels of cells (X_z and Γ_zt), districts (γ_d) ^[9], and cities (H_ct) discussed below. Similarly to Rosenthal and Strange (2003) and Jofre-Monseny (2009), we calculate the outcome variable for the periods 2007–2008, 2009–2010, 2011–2012, and 2013–2014 and use the explanatory variables in the years 2006, 2008, 2010, and 2012, respectively.

Confounders, whether observable or not, that affect both the choice of a cell and the spatial concentration of college workers present challenges to obtaining unbiased and consistent estimates of the parameter of interest β_r, r = 1, …, 4 (Rosenthal & Strange, 2008). For example, proximity to transport infrastructure can enhance market access, productivity, and the employment of more college-educated workers. In the same sense, specific advantages, such as proximity to natural resources, can increase the return of those with more education or even serve as a source of amenities, affecting the sorting of these workers in both cases (Moretti, 2004c). Note also that such influences can operate at various spatial scales. For instance, a higher concentration of educated workers in a specific cell may reflect the benefits of agglomeration in the city as a whole, or it could be a result of more localized urban planning within its district. We employ two complementary strategies to address these challenges.

First, leveraging the richness and granularity of our dataset, we incorporate a significant set of control variables. As indicated in Equation (1), these variables include information at the cell, district, and municipality levels.

Specifically, we include a set of lagged time-invariant control variables at cell level, X_z, that may influence the location of new establishments and determine the concentration of skilled workers around a specific cell. We interact these time-invariant variables with time effects, τ_t, to control possible differences in trends across specific cells. In this regard, local transport infrastructure can significantly impact access to markets and urban mobility, ultimately influencing the productivity of firms (Holl, 2004a, b, 2016; Mayer & Trevien, 2017; Gibbons, Lyytikäinen, Overman, & Sanchis-Guarner, 2019) and the local concentration of workers (Baum-Snow, 2010; Haas & Osland, 2014; Gao et al., 2019). Similarly, proximity to rivers or lakes can influence firm profits due to their impact on input transportation costs (Ellison & Glaeser, 1999; Rosenthal & Strange, 2001; Ellison, Glaeser, & Kerr, 2010) and the spatial distribution of workers, as these bodies of water may provide natural amenities (Combes, Duranton, & Gobillon, 2008; Fajgelbaum & Gaubert, 2020; Gaigné, Koster, Moizeau, & Thisse, 2022). Thus, we include in X_z variables representing the availability of local transport infrastructure and proximity to rivers: the Euclidean distances of each cell's centroid to the nearest airport, public port, railway, federal highway, state highway, and the nearest river.

Note also that the relationship between new establishments in a particular area and the local availability of university-educated workers may merely indicate other local benefits related to agglomeration. These benefits could include a better matching between firms and workers, as well as a wider range of inputs and services available in the vicinity (Arauzo-Carod et al., 2010; Almeida, Silveira Neto, & Rocha, 2023). To address this possibility, we also include the total number of workers within a radius of 20 kilometers from the cell's centroid, represented by Γ_zt in Equation (1). Although it may be endogenous, making it a “bad control” (Angrist & Pischke, 2009), its inclusion is valuable for understanding the mechanisms behind the estimated effects of local human capital ^[10].

Bias arising from omitted variables can also stem from the characteristics of a district, such as the level of urban planning restrictions or the district's industrial nature. These factors can drive up urban space prices and influence the types of economic variables present in the area (Dantas, Duarte, Neto, & Sampaio, 2018). To mitigate possible bias related to district time-invariant influences, we include the district fixed effect in Equation (1), γ_d.

The final set of control variables, H_ct, encompasses city characteristics that may influence firm earnings and the concentration of highly educated individuals. This includes factors such as population size, the level of engagement with international trade, costs related to urban infrastructure, and measures of urban environmental quality.

The level of human capital and the size of the municipality both contribute to a wage premium in Brazil, as demonstrated by Chauvin et al. (2017) and Quintero and Roberts (2023). This suggests that urban economies operating at the city level can enhance productivity by promoting a greater diversity of inputs and services, as well as improving occupational matching and preferences within the municipality (Duranton & Overman, 2005). To account for these influences, we include the municipal population as a control variable ^[11].

The presence of businesses involved in international trade can indicate competition capacity and efficiency, influencing a municipality's appeal for new businesses. This may also influence the local labor demand by hiring more college-educated workers from other countries to meet their specific demands (Araújo & Salerno, 2015). Thus, we included in H_ct the value of municipal exports and imports (per 100,000 inhabitants) as a proxy for access to the international market.

In H_ct, we also consider variables associated with the structure, planning and quality of the urban environment, as these factors can affect the attraction of new companies and college-educated workers (Glaeser, Rosenthal, & Strange, 2010; Chatterji, Glaeser, & Kerr, 2014; Fisher, 1991; Detotto & Otranto, 2010; Rosenthal & Ross, 2010; Matti & Ross, 2016). Specifically, we include as additional controls the municipalities' capital expenditure, housing and urban planning expenses, the number of homicides, and the number of traffic fatalities per 100,000 inhabitants.

Finally, we introduced a linear trend for the dependent variable to reflect the influence of the general decline in the number of new firms within the Brazilian manufacturing sector during the period. As the numbers in Table A2 of Appendix A confirm, there is a reduction in the number of new companies per cell between 2007–2008 and 2013–2014.

In Table A2 of Appendix A we present descriptive information on our outcome variable (new plants), explanatory variables, and the number of cells, districts, and municipalities. Our entire sample contains about 50.4 thousand cells (observations). Following the country's trend, notice that our higher education variable (“college-or-more”) grows over time in all rings. There has also been a decrease in the number of new firms over the two years, reflecting a nationwide trend.

To address the possible remaining sources of bias, our second strategy utilizes exogenous variation in higher education levels in Brazil to construct a shift-share instrumental variables (SSIV) for the higher education variables in the cells’ rings.

Here, we use shifts in the national education policy between 1991 and 2004 as an exogenous source of variation in the number of college-educated people across concentric rings. In this regard, notice that relevant changes in the national education policy occurred in the early 90s in the country (Corbucci, 2002; Ferreyra et al., 2017). The approval of new laws establishing rules for the operation of the higher education system facilitated the system's expansion, favoring its diversification and opening space to increased participation of private universities and colleges (Corbucci, Kubota, & Meira, 2016; Reis and de Assis Pires, 2023). As shown in Figure 1, the expansion of the higher quality public educational system and the implementation of incentives for the growth of private institutions (tax exemptions, student loans, and other instruments to facilitate student’s access to those institutions) led the share of the 18–24 age range population enrolled in higher education to growth from 9.3% in 1995 to 45.1% in 2020.

Given our Poisson specification, we use the instrumental variable in a control function approach and also provide evidence about the endogeneity of $Sr(z)t$ (r = 1, …, 4) based on the coefficient of the first-stage residual described below (Cameron & Trivedi, 2013). This policy-driven instrument combines the growth of the college-educated population attributed to the education policies implemented by the federal government (shift component) with the past demographic structure (share component). The natural tendency for younger people to be more educated than previous generations (Moretti, 2004a) has been shifting upward as a result of educational policies (see, e.g. Corbucci, 2002; Corbucci et al., 2016; Reis and de Assis Pires, 2023) so that concentric rings with a lagged age structure made up of more young people would have a larger number of college-educated people when assigned the national shift.

Formally, the instrument for the number of workers with a college degree or higher is given by:

where $Ar∩c$ is the intersection area between concentric ring r and municipality c; $Ac$ is the total area of the municipality; $Pcm1991$ is the population ^[12] with a college degree or higher in the municipality c and cohort m (we defined three age groups: young 16–25, middle-aged 26–50, and old 51–70) in the reference year; and $Pmt$ is the national population with a college degree or higher in cohort m and year t = (2000, ⋯ , 2004).

For each r = 1, …, 4, the first step equation is thus given by:

where College_r(z)t is the shift-share instrumental variable, and w_r(z)t are omitted factors that can influence the local concentration of college-educated workers. Let the omitted term in Equation (1) be ϵ_zt = ρw_r(z)t + ψ_zt, with ψ_zt independent of w_r(z)t. We can therefore obtain consistent estimates for the parameters of interest by using the estimates of w_r(z)t $(wˆr(z)t)$ obtained from Equation (3) as a control in the second-stage equation:

Since we use $wˆr(z)t$ instead of w_r(z)t, we implement bootstrap simulations to take this extra source of variation into account (Petrin & Train, 2010; Cameron & Trivedi, 2013) ^[13]. This control function approach interestingly allows for testing the endogeneity of $Sr(z)t$ (r = 1, …, 4) by examining the statistical significance of the coefficients of the first-stage residuals (ρ_r). This information can help determine the appropriateness of the initial strategy (Cameron & Trivedi, 2013).

We initially presented the results following the more traditional approach that ignores possible variations in the spatial scope of human capital externalities (Moretti, 2004a, 2004c). Chauvin et al. (2017) and Quintero and Roberts (2023) applied this strategy to the Brazilian case and obtained evidence that local human capital generates an important wage premium. Here, our strategy introduces two key innovations related to these works.

First, we use the number of new firms as the dependent variable instead of wages. Note that the spatial location tends to be more directly tied to the decisions made by firms. Second, rather than considering the entire area of municipalities to assess local human capital, we focus solely on the area within a 20 km radius of the cell centroid. More specifically, we use Equation (1) with an important modification: instead of four distinct regions, we solely consider the number of workers with a college degree in a region defined by an entire circle with a radius of 20 km from the cells' centroids. In other words, we assume spatial homogeneity of capital human externalities up to 20 km.

Our initial set of evidence, thus, enables a comparison with previous findings in the literature that focus on the spatial homogeneity effects of human capital externalities. When contrasted with the following results, it also underscores the importance of considering the spatial scope of human capital externalities in Brazil. These initial estimates are presented in Table 1.

Notice that, in addition to the estimates from the Poisson model without any control variables (shown in Column 1), we have estimated three other specifications. Column 2 of Table 1 presents the estimates after including cell and municipal controls along with a linear time trend. Column 3 displays the estimates after adding district fixed effects. Finally, Column 4 presents the estimates obtained by using an instrumental variable (IV) and a control function approach. The IV in this case corresponds to the sum of values from Equation (2) for the areas within 20 km of the centroids of the cells. Due to space restrictions, here and in the next tables of results, we only present the estimates for our variable of interest, i.e., the coefficients associated with the local number of college-educated workers. The estimates for other variables are available by the authors upon request.

We highlight two pieces of evidence about this set of evidence from Table 1. Firstly, it's important to note that despite the reduction in the coefficient value when different controls are added, all specifications consistently show a positive and similar effect of the number of workers holding a college degree or higher on the number of new firms established. The decrease in the coefficient value when accounting for the district fixed effect indicates the challenge of obtaining within-district estimates using only four periods of data. In turn, the instrumental variables (IV) estimate indicates that confounding variables have little effect on the overall estimates. Overall, the evidence is in line with that obtained by Chauvin et al. (2017) and Quintero and Roberts (2023) when considering the effects of human capital externalities on wages in Brazil and, thus, favors the presence of such agglomeration gains in the country.

Notice, however, that this positive effect associated with the presence of workers with higher schooling appears small. For example, the estimated coefficient of Column 4 of Table 1 indicates that an increase of 1,000 in the number of workers with a college degree or more located in the circles (approximately 10% of the mean number of this kind of workers in the circles in 2012) generate an expansion of, on average, about 1.4% in the number of new firms in the cells ^[14]. This reduced impact of local human capital on firms' location decisions is somewhat expected, given that there are alternative locations within the same area or municipality ^[15]. As shown below, the apparent reduced magnitude of this source of agglomeration conceals various effects related to the location of more educated workers concerning the cells.

We now present estimates of the coefficients of interest from Equation (1), specifically those associated with the external effects of human capital at varying distances from the centroid of the cell. These results are presented in Table 2, which columns of results assume the same correspondent sets of control variables and strategies of Table 1.

Notice, first, that the results of the estimates of the residuals’ coefficients in Equation (4) of the control function strategy are presented in Table A4 of the Appendix B. As indicated by the numbers in this table, with the exception of the 1–5 km and 5–10 km rings, the coefficients of these first-stage residuals were not statistically significant, suggesting that the human capital variables for the other two rings can be considered exogenous. We then estimated the effects of human capital concentration in the rings in Equation (4) including only the estimated residuals for the 1–5 km and 5–10 km rings ^[16].

Some general patterns emerge from the numbers in Table 2. Firstly, we observe similar, positive, and statistically significant effects of proximity to college-educated workers within a 1 km radius. As for other rings, positive and statistically significant effects of local human capital are limited up to 5 km for specifications with control variables (Columns 1–3) and non-existent for the IV estimates (Column 4). Actually, in our IV estimates, we obtained negative effects of local human capital on the number of new firms for the more distant rings (5 to 10 km and 10 to 20 km).

Overall, our results indicate that as distance increases, the concentration of human capital rapidly loses its power in attracting new establishments, suggesting that distance is important for companies that take advantage of this agglomeration factor. In turn, at longer distances, the increase in the number of more educated workers appears to make cells less attractive in relation to others, suggesting prevalence of competition effects (Håkansson & Isacsson, 2019).

The magnitude of the external positive effect of local human capital on the number of new firms in the proximity of the cells is now important. Specifically, according to the numbers in Column 4 of Table 2, adding 1,000 and a standard deviation of the number of college-educated workers up to 1 km from a specific location increases the expected number of firm births, on average, by 29.6% and 5.5%, respectively. On the other hand, adding 1,000 college-educated workers in the farthest ring (10 to 20 km) decreases the number of new plants, on average, by about 4.4%.

This pattern of spatial attenuation conforms to the spatial scale of human capital externalities for which mechanisms are expected to act mainly at short distances (Fu, 2007; Rosenthal & Strange, 2008, 2020; Andersson et al., 2009; Eppelsheimer et al., 2022; Balsmeier et al., 2023) and appears in consonance with the difficulties of urban mobility in Brazilian cities that may limit the spatial scope of agglomeration gains (see, e.g. Silveira Neto, Duarte, & Páez, 2015; Silveira Neto & Moura, 2019). Our set of evidence also offers a possible economic fundamental for the recent results about the strong location pattern of manufacturing activity at short distances in Brazil shown by Almeida et al. (2022).

Considering that industries differ in the complexity of tasks performed by its workers and location patterns, we also explored the possibility of differences in spatial patterns of gains with the proximity to more educated workers according to the technological intensity of the manufacturing sectors.

More specifically, we obtained additional evidence by four groups of industries according to the pattern of technological intensity. We use the technological classification proposed by Cavalcante (2014), which it is based on the compatibility of the Brazilian CNAE (Classificação Nacional das Atividades Econômicas) with the OECD technological classification, and consider four groups of industry: high-tech, medium-high-tech, medium-low-tech, and low-tech (see Table A3 in Appendix A for the distribution of activities among these groups).

The following Table 3 presents the new results both with control variables and district fixed effects (Column 1) and the IV and the Control Function strategy (Column 2). Panel A exhibits the results for high-tech and medium-high-tech industries and Panel B exhibits the results for medium-low-tech and low-tech industries. Note that in general, we obtain similar results using both strategies and, when statistically significant, the coefficients present the same signals. Therefore, we focus our discussion on the IV estimates ^[17].

These new results strengthen our earlier findings and provide insights into the mechanisms of external gains related to human capital.

In this regard, it is important to note that the coefficients for the 1 km ring are the only ones positive and significant across all types of industry. Although there is no statistical difference among them, the coefficients numerically increase with the technological level of the sector. Notice also that, while the geographical extent of attenuation for other industries is limited to the first ring in the other three groups, for firms in the low-tech group it extends to the 1–5 km ring. The two pieces of evidence indicate that while the benefits of having human capital in high-tech industries are limited to the immediate neighborhoods, low-tech companies experience these advantages over a larger geographical area.

We highlight that this evidence partially aligns with the results presented by Almeida et al. (2022), who have shown that high-tech industries are concentrated at short distances. Thus, it suggests that one of the possible mechanisms responsible for the pattern of location of these industries is the limited spatial scope of the gains associated with the local concentration of workers with a college degree. The findings also support the notion that benefits primarily come from face-to-face interactions in non-low-tech industries, which can create learning opportunities. In contrast, for low-tech industries, there are wider advantages to engaging with more qualified individuals, often located farther away from the operational site (Wallsten, 2001; Woodward, Figueiredo, & Guimaraes, 2006; Rosenthal & Strange, 2008; Thisse, 2018).

Depending on their nature, agglomeration economies may be effective at different spatial scopes. Partially depending on face-to-face interactions, the gains associated with human capital externalities, in particular, may present limited spatial scope. This paper empirically explores the spatial scope of human capital externalities in manufacturing activities in Brazilian cities. We test whether human capital spillovers can make a location more attractive for the attraction of new establishments and whether the effects are attenuated with distance. The subject is still barely investigated in developing countries’ scenarios, despite their lack of adequate transport infrastructure that may affect individuals’ spatial mobility and, thus, potentially the benefits of spatial interactions.

To measure these effects, we take advantage of a unique micro-geographic data set and use count models for the number of new establishments in 1 km² cells in Brazilian metropolitan regions, considering the number of workers with a college degree or more at different rings from cells' centroids as measures of local human capital. Our strategy use a comprehensive set of control variable at the cell and municipality levels and consider the influence of time-fixed observed and unobserved heterogeneities of cells’ neighborhoods.

Our results revealed two important patterns. First, after controlling for the district fixed effect, previous local transportation infrastructure, local geographic amenities, local development policies, and the size of the local labor supply, we show that the effects generated by the concentration of college-educated workers on the number of new establishments are observed only at short distances. These findings are clearly in line with the arguments in favor of face-to-face interactions as the mechanism generating human capital spillovers. The positive effect of the local concentration of college-educated workers in a specific cell works up to a maximum of 5 km from its centroid, being stronger in the first concentric ring around the cell (up to 1 km). This pattern conforms to the precarious infrastructure of most Brazilian cities which may limit the spatial scope of agglomeration economies, especially human capital spillovers.

Considering how the human capital spillovers differ across technological levels, we also show that the effects generated by the local concentration of human capital are less localized for low-tech firms (spatial scope up to 5 km) and more localized for more technologically intensive firms (spatial scope up to 1 km). In all cases, the positive effect is stronger at short distances and tend to decrease with distance, indicating an attenuation pattern.

Besides contributing to the scarce literature about the spatial scope of agglomeration economies in developing countries, our results present implications for public policy. Note that the recent Brazilian expansion of higher education, despite being important from an individual point of view (Oliveira & Silveira Neto, 2022), tends to generate very localized agglomeration gains for manufacturing activities. In terms of attractiveness for firms, given the current lack of adequate transport infrastructure in Brazilian cities, this means that only specific locations within urban centers (with a higher concentration of educated people) would have more chances of expanding their productive capacities. It is important to note that this situation often applies to cities in developing countries with similar urban structures and inadequate transportation infrastructure. For example, this is particularly evident in cities across Latin America, as described by Fernández-Maldonado, Romein, Verkoren, and Parente Paula Pessoa (2014). Our results suggest, thus, that improving transport infrastructure can help to spatially expand the gains from the expansion of higher education.

The study presents some limitations that deserve to be considered in future work. Notice particularly that, despite the gains in terms of sector representation associated with considering solely firms from the manufacturing sector (since the majority of firms of manufacturing activities are in the formal sector and, thus, with information able for georeferencing), the absence of firms from services activities means disregarding the influence of local human capital on the location of a significant number of firms. The challenge of including such firms in the analysis must be faced in future research.

Edilberto Almeida would like to acknowledge the financial support provided by Edital Propesqi nº 05/2024, Federal University of Pernambuco (UFPE).

The supplementary material for this article can be found online.