Spillovers in the Urban Wage Premium

Spillovers in the Urban Wage Premium Andrea R. Lamorgese, Elisabetta Olivieri and Marco Paccagnella April 2018 Abstract In many countries urban workers enjoy higher wages than non urban ones, and the wage premium increases with the size of the city. In this paper we show that this wage premium can go beyond the city boundaries and affect also workers located elsewhere through spillovers in productivity between plants/firms. In particular, we analyze spillovers in the urban wage premium along: i) the internal structure of the firm (i.e. between different plants belonging to the same firm), ii) the value chain (i.e. different firms linked through client-supplier relationships). We rely on a matched employers-employee dataset in order to control for firms and workers characteristics. Our results confirm that workers enjoy a wage premium when they work in an urban area; this premium is larger for workers in headquarters than in branches. Despite that, workers in branches benefit from the localization of their headquarter: when firm s headquarter is in an urban area, workers in non-urban branches receive a wage premium of about 3%. Furthermore, working for a firm supplying goods or services to urban firms provides a wage premium that is almost as large as the one a worker would get in a urban non-supplier firm. According to our results, these spillovers are not driven by workforce composition; they reflect the fact that plants/firms connected to cities have larger rents to share with workers. We also conclude that the presence of such within- and between-firms spillovers, by equalizing wages across urban and non-urban locations, results in a downward bias in standard estimates of the urban wage premium and in an under-appreciation of the economic importance of cities. Keywords: Urban wage premium, headquarters vs. branches, value chains, AKM. We wish to thank Riccardo Crescenzi, Stephen Machin, Paolo Naticchioni, Andrea Petrella, Paolo Sestito for their valuable advice, as well as seminar participants at Banca d Italia and OECD. The usual disclaimer applies. The views expressed herein are those of the authors and not necessarily those of Bank of Italy or OECD. Research Department, Bank of Italy, Via Nazionale 91, 00184 Roma, Italy. Email address: andrea.lamorgese@bancaditalia.it. Research Department, Bank of Italy, Via Nazionale 91, 00184 Roma, Italy. Email address: elisabetta.olivieri@bancaditalia.it. Directorate for Education and Skills, OECD. Email address: marco.paccagnella@oecd.org.

1 Introduction In many countries, spatial disparities in wages are large and are a source of policy concern. In particular, the urban economic literature provides ample evidence for the existence of an urban wage premium, estimating that in urban areas wages are between 1% and 11% higher than in non-urban areas. 1 This literature typically compare wages of individuals working in areas with different degrees of urbanization net of observable and non observable worker s and firm s characteristics. To the best of our knowledge, no previous papers in this literature has considered the possibility that workers in non-urban areas may enjoy a wage premium thanks to economic interactions with firms or plants located in urban areas. In this paper we look for evidence of spillovers in wage premia operating beyond city boundaries. We look at two different channels through which wage premia may propagate. The first operates between plants linked by corporate relationships, i.e. controlled by the same firm. Since plants in the same firm may share the same wage policy, it is likely that the localization of one plant impacts workers located in other plants. The second operates along the value chain. Firms that supply goods and services to firms located in an urban areas, may indirectly benefit of the localization advantage of their clients (e.g. because of productivity spillovers) and pay a higher wage to their workers. Indeed, there is evidence that participating in value chains raises the productivity of firms, providing them with know-how acquired from within the value chain, with best-practises both on technical and administrative grounds (Accetturo et al., 2013; Accetturo et al. 2016). Both channels would tend to equalize wages across urban and non urban locations, thus biasing downward standard estimates of the urban wage premium. To verify our hypothesis we rely on data from Italian National Institute of Social Security (INPS) and the Survey of industrial and service firm (Invind). Matching these two sources of information we are able to work with a matched employeremployees dataset which contains information on a representative sample of Italian big firms and on the entire working history of workers who have ever transited in one of these firms between 2005 and 2014. Importantly, our data allow us to identify different plants belonging to a same firm, as well as to infer client-supplier relationships between different firms. We can therefore the wage premium that a worker enjoys depending on the degree of urbanization of: i) his place of work (standard urban wage premium), ii) the place where the headquarter of the firm is located, iii) the place where customers of his firm are located. We focus on Italy, which is a country with wide spatial disparities in wages of the private sector. Beyond the large gap between wages across Northern and Southern regions of the country, another important source of heterogeneity is the urban localization of the worker. On average, in the period 2005-2014 average hourly wages in the private sector are 9% larger in urban areas than in non urban ones. 2 Such wage premium is increasing in population density and in city size. Figure 1 See Rosenthal and Strange (2004) and Puga (2010) for some reviews. 2 See section 3.2 for a definition of urban and non urban areas. 2

1 describes the relation between log of population (on the X-axis) and log of hourly wages (on the y-axis). It shows that urban areas are generally larger and pay a larger wage: when population doubles, hourly wages grow by 2.2%. When density doubles, they grow by 0.8%. The wage elasticity to an increase in population is qualitatively comparable, even if quantitatively smaller than the ones reported in comparable analyses for the United States, Spain, and France. Figure 1 Hourly wage, urban and non urban areas and population 2,3 2,3 2,1 2,1 1,9 1,9 1,7 1,7 1,5 1,5 8 10 12 14 16 Source: Istat, Rilevazione SLL urbani sulle forze di SLL lavoro. non urbani Note: Log of population on the X-axis, log of hourly wages on the y-axis. Data are averages of 2009-2014 values, net of year and quarter fixed effects. The solid line is the predicted value of a linear regression of log of wages on a constant, year and quarter fixed effects, and log of population. Green dots are non urban areas, red dots are urban areas. Our results show that the urban wage premium is not uniform across plants of the same firm. Indeed, workers employed in the firm s headquarter enjoy a wage premium which is 4 percentage point higher than the urban wage premium enjoyed by workers employed in urban branches. However, workers in non-urban branches share a part of the urban wage premium of their own headquarter: their wages are estimated to be 3% higher in case the headquarter of the firm is located in an urban area. Results hold using both the size of the population as a proxy of agglomeration and using a TSLS approach. Furthermore, we show that wage spillovers occur also between firms linked by client-supplier relationships. While suppliers in general pay lower wages than nonsupplier firms, they do pay 5% higher wages if they sell intermediates to sectors that are mainly located in urban areas. The literature has highlighted many possible reasons behind the observed fact that wages tend to be higher in cities than in non-urban areas. Among them, indi- 3

vidual sorting of more skilled workers into urban areas may account for a significant portion of the spatial wage variation (Combes et al., 2008). In other words, the wage gap across cities is partially due to a different composition of the labor force, so that workers with better (observable or non observable) characteristics, which the labor market rewards to a larger extent, might have a preference to locate in cities. Likewise, firms with better (observable or non observable) characteristics, which are more productive, might prefer to operate in cities and therefore pay a larger wage to their workers simply for rent-sharing reasons. In both cases the urban wage premium would be the consequence of the sorting of better firms or better workers in urban areas. In order to detect whether our results are driven either by workers or by firms selection into cities, we perform a standard AKM decomposition of wages (Abowd et al., 1999, see) and look for spillovers in both the firms and workers components of wages. As in Dauth et al. (2016), we consider that the employer specific component of wages captures the part of the actual wage paid to the employee that would be paid to any employee, which depends solely on the employer s intrinsic characteristics. The employee specific component of wage captures instead the part of the actual wage that would be paid to that employee by whichever employer, since it is linked to the employee intrinsic characteristics. Our results show that spillovers concern significantly only the employer component of wages. In other words, they are almost entirely related to the fact that plants/firms connected to the city have larger rents to share with their workers. This evidence imply that a standard measure of urban wage premium may underestimate spatial differences in wages driven by employer s intrinsic characteristics. The rest of the paper is structured as follows: in sections 2 and 3 we describe the related literature and our data. In section 4 we present our empirical strategies and in 5 we provide the results. Section 6 discusses robustness checks and extensions of our analysis. Section 7 concludes. 2 Literature The number of papers on wage disparities across cities is vast. Starting from the seminal contribution of Glaeser and Maré (2001) for the United States, the economic literature on urban wage premium has extent to various possible explanations, issues and countries of analysis. Long-run equilibrium wage differentials among similar workers can arise to the extent that there are differences in worker skills and/or productivity between urban and non-urban areas. Many papers looked at the extent to which sorting of highskilled workers into urban areas can explain spatial disparities in wages. If cities pay more because they attract the most skilled workers, a significant portion of the premium is then likely to be a return to observed and unobserved skills (Combes et al., 2008; Mion and Naticchioni, 2009; Matano and Naticchioni, 2012). The alternative hypothesis holds that workers are more productive in urban settings due to agglomeration economies (Ciccone and Hall, 1996; Glaeser, 2011). Other papers focused on the timing of the wage premium and investigated whether urban workers receive this premium immediately, or through faster wage growth over time. Among them, Glaeser and Maré (2001) distinguish a wage growth 4

effect (interpreted as human capital accumulation) from the wage level effect (agglomeration economies) and show that the wage growth effect plays an important role. De La Roca and Puga (2016) show that in bigger cities workers obtain an immediate static premium and accumulate more valuable experience, which persists after workers move elsewhere. Finally, some contributions deal with assortative matching of workers and firms in urban areas. Andersson et al. (2014) find stronger assortative matching in denser counties of the US. Mion and Naticchioni (2009) find a positive correlation of individual ability and firm size in Italian regions, but a negative correlation of assortativeness and density. Dauth et al. (2016) find a stronger assortative matching in denser German regions, which has become considerably more important over time, thereby contributing substantially to the rise in spatial wage inequality in Germany. The empirical literature on wage disparities across cities dealt with many developed countries. Yankow (2006) estimates a urban wage premium for US big cities workers of around 22%, which decreases to around 5% when unobserved worker heterogeneity is taken into account. For France, Combes et al. (2008) find that unconditional correlation between log of individual wages and log of lagged city s density is 4.9%, 3 which goes down to 3% when observed and unobserved characteristics of the city and the workers are taken into account. For the UK, D Costa and Overman (2014) find an unconditional urban premium of 14.1%, which decreases to 8.4 when considering controls for workers observable and to 2.3% when also conditioning out sorting on workers unobservable characteristics. Finally, for Spain De La Roca and Puga (2016) estimate an urban wage premium of 4.8% conditional on workers observables and of 2.5% when controlling for sorting on unobservables. As far as Italy is concerned, Di Addario and Patacchini (2008) find that adding 100,000 inhabitants in the local labor market raises earnings by 0.1 percent. This effect decays very rapidly with distance, losing significance beyond approximately 12 kilometers. Matano and Naticchioni (2012) provide evidence of a urban wage premium of around 1.8% for both industrial and service firms, conditional on firms and workers observables. The authors stress that the urban wage premium is particularly concentrated in the right tail of the wage distribution. Workers unobservables are shown to bias upward the estimate of the effect of density on wages: conditioning out workers fixed effects halves the magnitude of the mean estimates, and the bias is shown to be particularly large in the right tail of the wage distribution, meaning that better-paid workers are the ones more likely to be self-selected into denser cities. According to this literature, the urban wage premium in Italy is smaller than many other developed countries, even when observed and unobserved workers heterogeneity is taken into account. There might be several reasons for this to be the case, among which centralized wage setting (Mion and Naticchioni, 2009), the rich supply of non monetary compensation (i.e. amenities) in Italian big cities (Dalmazzo and de Blasio, 2011), the limited mobility of Italian workers (Di Addario and Patacchini, 2008). In the following sections, we focus on a different explanation of the reduced 3 The coefficient does not change much when the authors consider the correlation of the log of individual wage with log of lagged population. 5

wage premium accruing to Italian urban workers, i.e. network externalities in wage setting across plants within a firm and across customers and suppliers along the value chain. These externalities may go beyond city boundaries and lead to an underestimation of the urban wage premium particularly relevant in Italy, where firms are smaller and more connected. 3 Data 3.1 Matched employers-employee data The data used in this paper were constructed from the Bank of Italy s Survey of manufacturing and service firms (INVIND). INVIND is an open panel of around 4,000 firms per year, representative of manufacturing and service firms with at least 20 employees. It contains detailed information on firms characteristics, including the type of output produced and the municipality of the headquarter. The Italian Social Security Institute (Istituto Nazionale Previdenza Sociale, INPS) provided the complete work histories of all workers who were ever employed by an INVIND firm in the period 2005-14, including spells of employment in firms not listed in the INVIND survey. Overall, we have information on about 5 millions workers per year, 25% of whom are employed in firms surveyed by INVIND in any given year. The remaining ones are employed in about 1,800,000 firms not surveyed in INVIND, of which we only know the industrial sector and the average number of employees during the year. The information on workers includes age, gender, nationality, municipality of work, worker s qualification, annual gross earnings (including irregular payments such as overtime, shift work, and bonuses), number of days worked, and the firm identifier. We have deleted records with missing entries on either the firm or the worker identifier, those corresponding to workers younger than 15 and older than 65, and those in the first and last percentiles of the weekly earnings distribution. In table 1, we report summary statistics on workers characteristics (average gross weekly earnings, worker s qualification, nationality, average age of workers, gender, number of observations, both for the entire sample and for the sample of workers employed in an INVIND firm). Since the INVIND sample does not include small firms, which generally pay lower wages and hire less qualified workers, gross earnings in the INVIND sample are on average higher than in the full sample. In spite of that, other descriptive statistics for workers in the INVIND sample are quite similar to those of the total sample; this was expected, because the two samples are largely formed by the same workers at different times of their working life. Table 2 reports statistics on the variables used in our main regression analyses. A total of 8,400 firms are included in the INVIND sample in the period considered. The sample is unbalanced and the median INVIND firm employs about 60 workers. Small firms (with less than 50 employees) form 47% of the sample; medium firms, whose employment ranges between 50 and 249 employees represent 38%; large firms (with 250 or more employees) 15%. Since we are interested in spillovers in the firm wage policy across branches and along the value chain, table 2 also describes the structure of firms in our sample. 6

Table 1 Summary statistics: workers All sample INVIND Sample average gross daily earnings 85.3 95.4 % managers 0.5 0.5 average age of workers 39.1 41.1 % women 34.8 31.5 % foreigners 9.2 6.5 # of observations per year 5,081,122 1,290,973 Source: INPS and INVIND. Note: annual averages from 2005 to 2014. Table 2 Summary statistics: firms All sample INVIND Sample Number of firms per year 720,401 4,031 with 1 plant 692,605 2,614 with 2 plants 18,036 588 with 3 plants 3,899 225 with 4 plants 1,716 126 with 5 plants or more 4,145 478 firms with less than 50 workers 692,950 1,871 firms with 50-249 workers 23,699 1,547 firms with 250 workers or more 3,751 613 % suppliers - 56.8 % suppliers to urban firms - 7.3 Source: INPS and INVIND. Note: annual averages from 2005 to 2014. 7

In INPS data we observe for each worker the municipality of work, that we use to identify the local labor market of the establishment in which the worker is employed (LLMs; see section 3.2). 4 On the other hand, from INVIND we are able to identify the municipality (an therefore the LLM) in which the firm has its headquarter. The table shows that 35% of INVIND firms have plants in more than one LLM (4% of all INPS firms) and almost 12% have at least 5 plants. Only for INVIND firms we can identify the municipality of the headquarter. Table 3 shows that multi-plants firms are typically bigger, employ a higher share of females, and pay higher wages. The two latter features are even more noticeable in headquarters. Table 3 Summary statistics: single-plant and multi-plant firms Single-plant Multi-plants Headquarter average gross daily earnings 87.5 94.1 99.3 average age of workers 41.2 40.9 41.1 % women 23.5 27.4 29.8 % foreigners 7.4 6.4 6.2 employees 111 573 - % managers 0.8 0.9 1.3 Source: INVIND and INPS. Note: INVIND sample of firms. Annual averages from 2005 to 2014. We are also interested in the supplier-customer relations along the value chain. Unfortunately, we do not observe any direct measure of customer-supplier relation between firms in our data. However, INVIND allows us to distinguish among firms that supply goods and services to other firms (or to the public sector) and firms that directly sell goods and services to private consumers. Through this information, we identify supplier firms in the pool of INVIND firms. According to table 2, 57% of firms are suppliers, meaning that they mainly sell intermediate goods and services to other firms or to the public sector. They are generally smaller than the average firm in the sample and employ a lower share of women (table 4). We than try to identify among suppliers those who are most likely to sell goods and services to firms located in urban areas. Our working hypothesis is that these suppliers benefit from the localization of their customers. We identify suppliers of urban firms on the basis of their sector of activity and of the input-output matrix provided by the Italian national institute for statistics (Istat). More precisely, we first of all we defined urban sectors by dividing sectors of activities (two digits Italian ATECO 2007) in two groups, depending on the concentration of the firms belonging to the sector in urban LLMs (for our definition of urban LLM, see section 3.2). Urban sectors are defined as the ones with a share of 4 Since a firm could have more than one plant in the same LLM, when we count the number of plants of a firm we may underestimate the real number. If a firm has got both the headquarter and a branch in the same LLM, we treat all workers in that LLM as if they were employed in the headquarter. 8

firms in urban LLMs above the median (61.7%). Then, on the basis of the input-output matrix provided by Istat in 2010, we select sectors that sell more than 50% of their output to urban sectors. Supplier firms that belong to these sectors are defined as suppliers to urban firms. While the categorization of firms in the various categories is based on somewhat arbitrary thresholds, the results of our analysis are robust to choices of different thresholds. According to our definition, 13% of supplier firms supply to urban firms. These firms pay higher wages than average (table 4). Since our definition of supplier to urban firms depends on the sector of activity, this difference may also reflect differences in sector composition. In the following, our estimation results will always control for the sector of activity of the firm in order net out the effect of sector composition and to isolate the effect of being a supplier to an urban firm. Table 4 Summary statistics: suppliers and non-suppliers Non-Suppliers Suppliers to urban firms average gross daily earnings 89.6 92.5 102.5 average age of workers 41.2 41.1 41.7 % women 32.4 19.5 14.0 % foreigners 6.5 7.4 7.9 employees 389.3 249.4 346.7 % managers 0.8 0.9 1.2 Source: INVIND and INPS. Note: INVIND sample of firms. Annual averages from 2005 to 2014. A supplier is a firm that supplies goods and services to other firms or to the public sector. A Non-Supplier is a firm that sells goods and services to families. A Supplier to urban firms is a supplier that belongs to sectors that sell more than 50% of their outputs to sectors with a share a firms located in urban LLMs above the median. 3.2 Italian urban areas What is meant by city in the empirical urban literature is often vague. Conceptually, urban areas do not necessarily overlap with the administrative borders of single municipalities and in general not even with those of the administrative units at a lower level of breakdown (NUTS3 regions, provinces in Italy). In practice, in the Italian context, the relevant level of aggregation to define urban areas is somewhere in between the municipal and the provincial level. A functional agglomeration which stays in between municipalities and provinces is the Local Labor Market (LLM), which is conventionally deemed as a good representation of a spatial agglomeration. LLMs are the result of a partition of the Italian territory made of subset of municipalities chosen in such a way that they contain both the place of residence and the workplace of (a majority of) residents. 5 5 At the end of 2014 Istat issued the fourth classification of LLMs based on commuting flows of the 2011 Census (see http://www.istat.it/it/strumenti/territorio-e-cartografia/ sistemi-locali-del-lavoro). The definition is consistent with the European definition of LLM. 9

The analyses in this paper look at spacial differences in wages depending on the level of agglomeration in LLMs. We will proxy agglomeration with: 1) the log of population in the LLM, 2) a dummy variable which identifies urban areas. In order to identify urban LLMs we rely on a recent methodology proposed by OECD Eurostat definition which we adapt to take into account the Italian Bureau of Census (Istat) definition of commuting areas. The OECD Eurostat methodology defines a urban area as a homogeneous set of areas whose density of population exceeds a certain threshold. 6 This definition is consistent with the traditional view that urban agglomerations are the places where production and knowledge spillovers take place, for the simple reason that density creates thick markets and favors the matching between demand and supply. It is therefore consistent with the traditional sources of agglomeration (labor pooling, cost sharing and knowledge spillovers), which are at the core of the birth of industrial cities in XIX century. To define a urban area Eurostat performs a three-step procedure. First, it considers a partition of the territory of the European Union in a grid of 1 square km cells, select all those with a population density of at least 1,500 inhabitants per square km and clusters together in what it calls an urban center all the neighboring dense cells reaching a population of 50,000 inhabitants or more. Second, it considers the administrative borders and aggregate to the urban center all municipalities whose at least half of the population is resident in the urban center; the so formed agglomeration is called the metropolitan core. Third, OECD-Eurostat defines an urban area as the union of the metropolitan core and its commuting area. Such urban area is called a Larger Urban Zone (LUZ). 7 In this work we adopt the first two steps of the OECD-Eurostat definition, while we consider Italian LLM, rather than OECD-Eurostat LUZ in the third step, so that we define a urban area as an LLM containing a urban center as defined by OECD-Eurostat. This methodology identifies 73 urban areas (or urban LLMs) over a total of 611 LLMs in 2011. Non urban areas are the remaining LLMs. Table 5 shows that almost 56% of plants are located in urban LLM s (both in the full and in the INVIND samples). Headquarters and supplier firms have more or less the same distribution in urban and non-urban areas. Conversely, suppliers to urban firms are slightly more concentrated in cities; this may reflect the fact that geographic proximity increases the chance of a supplier-customer relation between firms, or that firms have an incentive to locate close to their clients or to their suppliers. Firms in urban LLMs are on average bigger than in non-urban areas: therefore the share of workers in these areas (63%) is greater than the one of firms. Workers in urban areas tend to be more qualified (80% of managers work in urban LLMs) 6 See http://ec.europa.eu/eurostat/statistics-explained/index.php/european_cities_ %E2%80%93_the_EU-OECD_functional_urban_area_definition. 7 The commuting area is similar to the LLM, with slightly different thresholds. Namely, OECD Eurostat s commuting zones is constituted by all municipalities with at least 15% of residents who work in a neighboring municipality, such that the LUZ is continuous and self contained. A LLM is characterized by the fact that: i) people commute for work reasons; ii) LLMs are self contained (at least 75% of resident work within the LLM; 25% at most outside it); iii) municipalities within a LLM are contiguous (commuting takes place between contiguous municipalities, non contiguous ones are excluded); iv) the core of the LLM is the municipality toward which commuting flows are maximum. 10

and to earn higher wages (89 euros per day on average, compared to 78 euros in non-urban LLMs). Table 5 Summary statistics: workers and firms in urban and non urban LLMs Urban Non-Urban # shares # shares plants 446,787 55.7 355,582 44.3 plants of firms >49 emp. 48,950 54.6 40,646 45.4 INVIND plants 342,178 57.1 256852 42.9 INVIND headquarter 2,005 53.1 1,771 46.9 INVIND suppliers 1,086 55.9 857 44.1 INVIND supp. to cities 148.4 60.0 98.8 40.0 employees 3,250,116 63.3 1,885,002 36.7 managers 21,686 79.8 5,473 20.2 women 1,162,096 65.1 623,209 34.9 foreigners 281,725 59.5 191,997 40.5 average age of workers 38.7-39.2 - gross daily earnings 89.0-78.0 - Source: INVIND and INPS. 4 Empirical design 4.1 The urban wage premium First of all, our empirical analysis aims at measuring the urban wage premium in our data. We do that using a standard specification in the literature. We estimate a log-linear Mincerian function augmented with urbanization: w ijlt = α + β 1 X 1it + β 2 X 2jt + γurban lt + ε ijlt, (1) where i is the individual, j the firm, l the LLM and t the year. X 1it are controls for observable workers characteristics (namely age, age squared, gender, worker qualification, nationality, part-time jobs) and X 2jt are controls for observable firms characteristics (number of employees and sector of activity); ε ijlt is the error term. We also control for regional and time dummies. Finally, we capture the urban effect with the variable urban, which is represented by either the LLM log population mass or a time-invariant dummy which identifies urban LLMs. In order to test our assumption that working in cities may benefit more workers in headquarters than workers in branches: i) a dummy which identifies workers in headquarters (HQ jlt ), ii) the interaction term urbanhq jlt between HQ jlt and urban lt, which identifies workers in urban headquarters. Thus, the specification is: w ijlt = α + β 1 X 1it + β 2 X 2jt + γ 1 urban lt + γ 2 urbanhq jlt + γ 3 HQ jlt + ε ijlt. (2) 11

In a further step, we look for the presence of spillovers in the urban wage premium among plants of the same firm. We do that by focusing on workers in branches and evaluating whether they benefit from a wage premium due to the fact that the headquarter of their firm is located in an urban area. We estimate equation 3 on workers in branches and we add the variable HQurbanLLM jlt, which is a dummy for branches with a headquarter in a urban LLM (or the log population in headquarter s LLM): w ijlt = α + β 1 X 1it + β 2 X 2jt + γ 1 urban lt + γ 2 HQurbanLLM jlt + ε ijlt. (3) Thus, γ 1 is the wage premium that workers in branches get when they are in a urban LLM; γ 2 is the premium that workers in branches get when their headquarter is in a urban LLM. Finally, we measure spillovers along the value chain looking at the effect of working for a supplier firm which sells goods and services to firms located in cities. Equation 4 includes both the dummy supplier jt for supplier firms and the dummy Usupplier jt for firms that supply goods and services to urban firms. w ijlt = α + β 1 X 1it + β 2 X 2jt + γ 1 urban lt + γ 2 supplier jt + γ 3 Usupplier jt + ε ijlt. (4) 4.2 Worker s and firm s wage components After establishing the presence of spillovers among the various dimensions discussed above, further insights can be gained by decomposing wages into two components: one related to the worker and the other to the firm. This allows us to look at urban wage premia (and related spillovers) separately by each component of wages. Following Abowd et al. (1999), Macis and Schivardi (2016), and Dauth et al. (2016), we decompose wages into two components, one specific to the firm and one specific to the worker, by applying the well-know methodology proposed by (Abowd et al., 1999) (henceforth AKM). We estimate the following Mincerian equation: w ijlt = α + βx it + γ 1 I i + γ 2 I jl I t + ε ijlt, (5) where the vector of covariates X it includes age, age squared, a set of dummies for workers qualification and part-time jobs. I i are individual dummies; I jl I t plant-year dummies. Using the predicted values of these regressors, we can decompose wages as: w ijlt = θ 1i + θ 2jlt + residual (6) where the worker and the plant-year specific components are respectively the predicted value of the worker s and the plant s dummies (namely θ 1i = ˆγ 1 I i and θ 2jlt = ˆγ 2 I jl I t ) and the residual includes also workers observable characteristics. Then we collapse the individual component at plant level, obtaining the average individual component of employees in every plant and year θ 1jlt. The AKM wage decomposition rests on an assumption of exogenous worker mobility conditional on observables. In section 6 we discuss some possible violations 12

of this assumption and provide the results of a series of tests proposed by the previous literature to detect violations. We conclude that the AKM assumption is a reasonable representation of the mobility process in our context. This methodology allow to establish whether the urban wage premium is due solely to heterogeneous workforce composition across urban and non urban areas, that is to the sorting of better individuals (according to their observable and unobservable characteristics) in urban areas, or also to the fact that more productive firms sort into urban areas and there end up paying higher wages to workers. We will label this other channel as rent sharing, following Macis and Schivardi (2016). According to the workers sorting channel, better workers are paid a higher wage irrespective of the quality of the firm that employs them at each point in time and location; according to the rent sharing channel, some workers get a higher wage than their market value because they are employed in better firms. Having disentangled these two components of the wage, we then use them as dependent variables to see whether they are related to the degree of urbanization. We estimate at plant-year level: θ 1jlt = α + βx jlt + γurban lt + ε jlt, (7) θ 2jlt = α + βx jlt + γurban lt + ε jlt. (8) Finally, we check whether spillovers within plants of the same firm affect the worker or the plant component of wage. In other words, we want to see whether plants that benefit of this spillover pay a larger wage either because of their workforce composition or because they have larger rents to share with workers. We do that by restricting equations 7 and 8 to branches and adding them the term HQurbanLLM jlt as in equation 3. 5 Results 5.1 The urban wage premium Table 6 provides the simplest measures of urban wage premium in Italy (equation 1). When controlling only for heterogeneity over time, workers in urban LLMs earn a wage 9% larger than workers in other LLMs. In a LLM with a double population with respect to another, earnings are on average 4.5 percentage points larger. This is a substantial effect, given that population varies a lot among LLMs (the biggest LLM has a population equal to more than 1,000 times the one of the smallest one). These effects are significantly weaker when we add some standard controls to our regressions (gender, age, nationality, part-time jobs, worker s qualification, firm s sector and size, region of work) and are reduced to 3 and 2% (columns 2 and 5). The estimated coefficients on the workers observable characteristics are in line with our expectations. Wages appear to be lower for females and foreign born workers and exhibit a concave age profile. The urban wage premium we observe in the data is mostly explained by sorting of higher quality individuals: when controlling for unobserved workers heterogeneity through workers fixed effects (columns 3 and 6), the urban wage premium drops 13

to less than 1% and a double population is associated with a wage less than 1% higher. The wage elasticity to an increase in population is quantitatively smaller than the ones reported in comparable analyses for United States, Spain, and France (see Section 2). Table 6 Urban wage premium in Italy Dependent variable: log of individual daily wages (1) (2) (3) (4) (5) (6) urban LLM 0.090*** 0.029*** 0.007*** (0.023) (0.004) (0.000) log pop 0.045*** 0.016*** 0.005*** (0.004) (0.001) (0.000) female -0.129*** -0.128*** (0.003) (0.003) part-time -0.580*** -0.465*** -0.576*** -0.461*** (0.005) (0.000) (0.005) (0.000) age 0.023*** 0.057*** 0.023*** 0.058*** (0.001) (0.000) (0.001) (0.000) age 2-0.000*** -0.000*** -0.000*** -0.000*** (0.000) (0.000) (0.000) (0.000) foreign born -0.058*** 0.002*** (0.002) (0.002) employees (1) 0.000*** 0.000*** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) Fixed effects: time Yes Yes Yes Yes Yes Yes sector No Yes Yes No Yes Yes qualification No Yes Yes No Yes Yes region No Yes Yes No Yes Yes individual No No Yes No No Yes Obs. 61,612,166 61,612,166 61,612,166 56,425,322 56,425,322 56,425,322 R 2 0.021 0.669 0.880 0.031 0.671 0.885 Source: Inps-Invind 2005-2014. Notes: robust SE clustered by LLM in parenthesis. (1) average number of firm s employees during the calendar year. For firms in the INVIND sample we can identify whether the worker is employed in a branch or in the headquarter of the firm. In order to check whether the urban wage premium is uniform across plants, in table 7 we restrict our analysis to multiplants firms in the INVIND sample (one-sixth of total observations). The size of the urban wage premium in this subsample is in line with the one we obtained in the full sample (compare columns 1 and 3 of table 7 with columns 2 and 5 of table 6). 14

In columns 2 and 4 of table 7 we add both a dummy variable which identifies workers in headquarters and its interaction with our main variables of interest (equation 2). Results show that once controlling for standard variables, we do not observe any differences between the average earning in the headquarter and the one in branches. Despite that, the urban wage premium for workers in headquarters is significantly larger (4 percentage points more than in other plants). This suggests that headquarters benefit more than branches from being localized in an urban area. This may be driven either by the fact that skilled work (typically concentrated in cities) is more relevant in tasks generally performed in headquarters (e.g. R&D), and by other urban peculiarities that may increase headquarter productivity more than branches productivity (e.g. knowledge spillover among firms). 5.2 Spatial spillovers in wages In the previous subsection we observed that urban localization has a deeper effect on wages for workers in headquarters rather than for the ones in branches. Now we want to check whether workers in branches share a part of this headquarter urban wage premium regardless of their place of work. Indeed, since plants of the same firm may have similar wage policies, the wage premium enjoyed by headquarters localised in urban areas could go beyond the boundaries of the LLM and affect also workers in plants located elsewhere. If this is the case, the standard measure of the urban wage premium in table 6, which is based on workers place of work and does not take into account spatial spillovers, may underestimate the importance of cities for productivity and economic outcomes.. In order to detect the presence of spillovers in the urban wage premium among plants of the same firm, we focus on branches and look at the effect on workers earnings to have an headquarter in a urban LLM (equation 3). Table 8 shows our results. Workers in urban plants enjoy a urban wage premium of around 1% over non urban ones (direct effect), to which a further 3% is added for workers in plants whose headquarter is located in an urban area. Results hold if agglomeration is measured using the size of the population in the LLM. Thus, wages in branches seem to reflect headquarters localization more than their own localization. All in all, this mechanism equalizes wages across urban and non urban locations because of spillovers across headquarters and branches within the corporate boundaries. Spillovers in the wage policy do not need to occur only within the corporate boundaries though. We wonder whether they occur even across firms connected by some sort of network along the value chain (equation 4). Indeed, table 9 shows that wage spillovers occur between customers and suppliers as well. Even if suppliers in general pay lower wages than non-supplier firms, they do pay 5% higher wages if they sell intermediates to sectors that are mainly located in urban areas. Working for a supplier of urban firms provides a wage premium that is almost as large as the one in a urban non-supplier firm, once observed heterogeneity is accounted for (column 2). 15

Table 7 Headquarters and branches Dep. var.: log of individual daily wages for workers in multi-plants INVIND firms (1) (2) (3) (4) urban LLM 0.021*** 0.007 (0.006) (0.006) urban LLM X headquarter 0.039*** (0.011) log population 0.012*** 0.008*** (0.001) (0.002) log pop. X headquarter 0.008** (0.003) headquarter -0.009-0.084** (0.009) (0.040) female -0.114*** -0.114*** -0.115*** -0.115*** (0.003) (0.003) (0.003) (0.003) part-time -0.538*** -0.537*** -0.535*** -0.534*** (0.008) (0.008) (0.008) (0.008) age 0.019*** 0.018*** 0.018*** 0.018*** (0.001) (0.001) (0.001) (0.001) age 2-0.000*** -0.000*** -0.000*** -0.000*** (0.000) (0.000) (0.000) (0.000) foreign born -0.022*** -0.022*** -0.023*** -0.023*** (0.004) (0.004) (0.004) (0.005) employees (1) 0.000** 0.000** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) Fixed effects: time Yes Yes Yes Yes sector Yes Yes Yes Yes qualification Yes Yes Yes Yes region Yes Yes Yes Yes Obs. 10,721,509 10,721,509 9,663,108 9,663,108 R 2 0.738 0.738 0.738 0.738 Source: Inps-Invind 2005-2014, only workers in multi-plants INVIND firms. Notes: robust SE clustered at LLM level in parenthesis. (1) average number of employees during the calendar year. 16

Table 8 Spillovers in the urban wage premium from headquarters to branches Dependent variable: log of wages of workers in branches of multi-plants INVIND firms (1) (2) urban LLM 0.009* (0.005) headquarter in a urban LLM 0.032*** (0.005) log population 0.007*** (0.001) log pop in headquarter s LLM 0.005*** (0.002) female -0.102*** -0.102*** (0.002) (0.002) part-time -0.545*** -0.541*** (0.006) (0.006) age 0.015*** 0.014*** (0.001) (0.001) age 2-0.000*** -0.000*** (0.000) (0.000) foreign born -0.022*** -0.023*** (0.004) (0.004) employees (1) 0.000*** 0.000*** (0.000) (0.000) Fixed effects: time Yes Yes qualification Yes Yes sector Yes Yes region Yes Yes Obs. 6,798,248 6,126,196 R 2 0.738 0.739 Source: Inps-Invind 2005-2014, only workers in branches of multi-plants INVIND firms. Notes: robust SE clustered by LLM in parenthesis. (1) average number of employees during the calendar year. 17

Table 9 Spillovers in the urban wage premium from customers and suppliers Dependent variable: log of wages of workers in INVIND firms (1) (2) urban LLM 0.085*** 0.022*** (0.016) (0.006) supplier -0.021*** -0.020 (0.012) (0.007) urban supplier 0.038*** 0.048*** (0.019) (0.015) female -0.124*** (0.004) part-time -0.542*** (0.008) age 0.021*** (0.001) age 2-0.000*** (0.000) foreign born -0.043*** (0.004) employees (1) -0.000*** (0.000) Fixed effects: time Yes Yes qualification No Yes sector Yes Yes region No Yes Obs. 13,642,490 13,642,490 R 2 0.359 0.714 Source: Inps-Invind 2005-2014, only workers in INVIND firms. Notes: robust SE clustered by LLM in parenthesis. (1) average number of employees during the calendar year. 18

5.3 Abowd-Kramarz-Margolis wage decomposition In table 10 we present the results from estimation of equation 5. The dependent variable is the natural logarithm of daily earnings and the vector of covariates includes age and age squared (proxying for labor market experience), a set of dummies for worker s qualification and a dummy for part-time jobs, as well as individual and plant-year dummies. Table 10 Estimating worker effects and plant effects Dependent variable log of daily wages Number of observations 10,627,234 Number of worker FEs 2,072,271 Number of plant-year FEs 126,524 F (Prob>F) 20,741.44 (0.00) R 2 0.958 Root mean squared error 0.112 Source: Inps-Invind 2005-2014. Notes: robust SE in parenthesis. The estimation is implemented by the Stata code reghdfe. Table 11 presents the variance-covariance matrix between log wages and the different components of wages. Similar to Abowd et al. (1999), Iranzo et al. (2008) and Macis and Schivardi (2016), a substantial role in earnings variability is played by heterogeneity in worker effects (the correlation between wages and worker effects is 0.37). Indeed, plant effects also play an important role (correlation equal to 0.29). The correlation between the worker and the plant effects is negative. Table 11 Variance-covariance matrix of workers and plants effects log wages workers effect plants effect log wages 0.489 workers effect 0.369 1.595 plants effect 0.289-0.104 0.567 Source: Inps-Invind 2005-2014. Notes: correlations; the diagonal entry reports the standard deviation. 5.4 Urban premium and spatial spillovers in wage components We estimate equations 7 and 8 using either worker s and plant s wage components as dependent variables, and report results in table 12. Each observation in these 19

regressions is a plant in a certain year. Worker s wage components are collapsed at plant-year level and their variability over time is given by turnover in firm s employment. All regressions include plant-year level workforce characteristics (average age, percentages of females and foreign born in the workforce), and firm characteristics (employment, sector of activity), as well as region and year dummies. The urban wage premium in the worker s wage component (6%) is larger than in the plant s one (1%), in line with the rest of the literature. This suggests that workforce composition is different in urban and non-urban areas and more populated LLMs attract workers with higher skills. For plant-year wage premia, the estimate is smaller whereas statistically significant. This indicates that the higher wages related to the urban localization are also due to a plant effect, which is common to all workers in the plant. This is consistent with the idea that the firm and the workers share the surplus from a higher firm productivity in cities. In table 13 we try to determine which of the two components of wages is affected by spatial spillovers. Again, we consider only branches and we look at the effect of headquarter s localization on the two wage components. We find that headquarter s localization has no effect on the worker component (columns 1 and 3), meaning that workers are not on average more skilled when they work for a branch with an urban headquarter. Conversely, headquarter s localization is really relevant on the plant component. This means that plants with a urban headquarter have more rents to share with all their workers, regardless of their skills. Finally, instead of decomposing wages into a worker and a plant components, we split them into a worker and a firm components, thus considering workers in firms branches as if they were working in the LLM of their headquarter. In this case, the urban wage premium in the firm component is almost equal in size to the previous plant component, whereas the premium in the worker component is about a half (table 14). This confirms the importance of headquarter localization in wage determination at branch level and the fact that workers tend to sort themselves into urban areas more than into firms with an headquarter in a urban area. 6 Robustness 6.1 Endogeneity concerns and instrumental variable approach The relation between wages and city size might be plagued by endogeneity concerns (De La Roca and Puga, 2016). More precisely, an omitted variable bias could arise if some city characteristics simultaneously boost earnings and attract workers to the city. Furthermore, a reverse causality problem could affect our empirical design, since higher earnings may attract workers from other LLMs leading to an increase in city size. Our estimation models use two alternative proxies of city size: i) a dummy which identifies urban LLMs, ii) the natural logarithm of population in the LLM. As far as the urban dummy is concerned, its definition is time-invariant and mainly depends on historical reasons that are unrelated with current wages. Thus, we are not particularly concerned with endogeneity issues. 20

Table 12 Urban premium in workers and plants wage components Dependent variable: wage components at plant-year level workers plants workers plants (1) (2) (3) (4) urban LLM 0.060*** 0.012*** (0.002) (0.001) log population 0.032*** 0.006*** (0.001) (0.001) female -0.174*** -0.035*** -0.169*** -0.029*** (0.004) (0.003) (0.004) (0.003) age 0.154*** 0.002*** 0.153*** 0.003*** (0.000) (0.000) (0.000) (0.000) foreign born -0.077*** -0.039*** -0.081*** -0.046*** (0.009) (0.007) (0.009) (0.007) employees (1) 0.000*** 0.000*** -0.000 0.000*** (0.000) (0.000) (0.000) (0.000) Fixed effects: time Yes Yes Yes Yes sector Yes Yes Yes Yes region Yes Yes Yes Yes Obs. 135,641 135,641 120,598 120,598 R 2 0.933 0.852 0.935 0.838 Source: Inps-Invind 2005-2014. Only plants of INVIND firms. Notes: Wage components collapsed at plant-year level and obtained through an AKM wage decomposition. Robust SE in parenthesis. (1) average number of employees in the firm during the calendar year. Conversely, log population may vary over time depending on the level of wages and a simple OLS estimator may yield biased estimates. This problem is typically considered negligible by the literature (Ciccone and Hall, 1996; Combes et al., 2008), since the relative size of cities is pretty stable over time and almost entirely unrelated with current earnings. Nevertheless, many papers use a Two-Stage-Least- Squares (2SLS) model with lagged local population as an instrument for current population, arguing that lagged population is not affected by current endogenous (and exogenous) migration flows. In this section we also use a Two-Stage-Least-Squares (2SLS) strategy, exploiting both lagged population and differences in population growth among nationalities. We use the stock of population in 1995 by nationality at LLM level and, in order to provide time-variability to our instrument, we predict the actual LLM population by augmenting 1995 s stocks on the basis of population growth at national level for each nationality. Specifically, for each LLM l, we compute a fictional population: 21

Table 13 Spatial spillovers in wage components Dependent variable: wage components at plant-year level workers plants workers plants (1) (2) (3) (4) urban LLM 0.064*** 0.007*** (0.002) (0.002) HQ in a urban LLM 0.002 0.023*** (0.002) (0.002) log population 0.032*** 0.005*** (0.001) (0.001) log pop. in HQ s LLM 0.007*** 0.002*** (0.001) (0.001) female -0.180*** -0.030*** -0.17169*** -0.024*** (0.005) (0.004) (0.005) (0.004) age 0.152*** 0.002*** 0.151*** 0.003*** (0.000) (0.000) (0.000) (0.000) foreign born -0.028*** -0.039*** -0.056*** -0.037*** (0.011) (0.009) (0.011) (0.009) employees (1) 0.000*** 0.000*** -0.000 0.000*** (0.000) (0.000) (0.000) (0.000) Fixed effects: time Yes Yes Yes Yes sector Yes Yes Yes Yes region Yes Yes Yes Yes Obs. 104,500 104,500 92,530 92,530 R 2 0.933 0.833 0.934 0.819 Source: Inps-Invind 2005-2014. Branches of INVIND firms only. Notes: Wage components collapsed at plant-year level and obtained through an AKM wage decomposition. Robust SE in parenthesis. (1) average number of employees in the firm during the calendar year. 22

Table 14 Urban premium in workers and firms wage components Dependent variable: wage components at firm-year level workers firms workers firms (1) (2) (3) (4) urban LLM 0.029*** 0.013*** (0.002) (0.002) log population 0.019*** 0.004*** (0.001) (0.001) female -0.167*** -0.030*** -0.165*** -0.031*** (0.007) (0.005) (0.007) (0.005) age 0.132*** 0.001*** 0.131*** 0.001*** (0.000) (0.000) (0.000) (0.000) foreign born -0.122*** -0.067*** -0.114*** -0.068*** (0.011) (0.008) (0.011) (0.008) employees (1) 0.000* 0.000*** -0.000 0.000*** (0.000) (0.000) (0.000) (0.000) Fixed effects: time Yes Yes Yes Yes sector Yes Yes Yes Yes region Yes Yes Yes Yes Obs. 35,558 35,558 31,918 31,918 R 2 0.924 0.903 0.923 0.889 Source: Inps-Invind 2005-2014. Only INVIND firms. Notes: Wage components collapsed at firm-year level and obtained through an AKM wage decomposition. Robust SE in parenthesis. (1) average number of employees during the calendar year. S l,t = i P op i,l,t0 π i,t (9) where P op i,l,t0 measures the stock of population with nationality i that was settled in the LLM l in 1995; π i,t is the growth rate of population with nationality i in Italy at time t. Summing across all nationalities, we obtain a measure of the predicted population in LLM l in year t. The exclusion restriction of this instrument relies on two assumptions. First, population in 1995 must not be correlated with the current trend in local wages. This assumption is standard in this literature and even more likely to be fulfilled when we include LLM dummies in our specifications (see section 6.2). Moreover, the composition by ethnicity has changed markedly in the last two decades for reasons that do not depend on local economic conditions. Second, our instrument is not valid as long as the total population growth in Italy with nationality i is affected by unobserved factors at LLM level. This is not likely to be the case, since each LLM 23

represents a tiny share of total population. Tables 15 and 16 show that all the main results presented in past tables using OLS carry forward when we adopt an IV strategy. 6.2 The effects of an increase in population on local earnings Until now, our specifications looked at the effects of population on wages by exploiting spatial variability in our data. Nevertheless, the amount of population in the LLM may be correlated with many city characteristics related to city size (such as amenities or the quality of public services) which may affect local earnings. Since most of these characteristics are generally time-invariant, in order to isolate the role of population in wage determination, we now control for all LLM s time-invariant characteristics. In other words, we pose a slightly different research question: what would happen to wages in a certain LLM after an exogenous increase in population? To answer this question, we add LLM dummies to our specifications and instrument log population with the variable described above. Table 17 shows that a 1% increase in population in a certain LLM would lead to an increase in earnings by 0.2% and confirms both that the urban wage premium is larger for workers in headquarters rather that for workers in other plants and that for workers in plants having a urban headquarter leads to a wage premium. 6.3 Tests of the exogenous mobility assumption The AKM wage decomposition has been criticized for failing to account for endogenous mobility. The error term might be structurally related to the assignment of workers to employers either through search dynamics (Mortensen, 2003; Lentz, 2010), coordination frictions (Shimer, 2005), or learning (Gibbons et al., 2005). In this section we present a series of tests of the exogenous mobility assumption suggested by Card, Heining and Kline (2013) and Card, Cardoso and Klein (2013). First of all, we verify whether mobility is based on the value of the worker-plant match. If the exogenous mobility assumption is violated due to sorting based on the value of the worker-plant match, then the wage premium would include a match component that would be specific to each worker-plant pair. To test for such sorting, we look at wage changes for job movers. We considered all job changers with at least two consecutive years both in the old and in the new plant. We classified plants in every year of analysis (from 2005 to 2014) in quartiles on the basis of the average daily wages. We exclude plants with less than 10 employees in order to avoid that a single worker may affect the average wage in a substantial way. Under the exogenous mobility assumption, workers who transition between plants in the same quartile should not experience any wage change. Also, workers who move from a low quartile to a high quartile should experience a wage increase; conversely, workers who move in the opposite direction should have a roughly symmetrical wage reduction. If workers change plants on the basis of a match component, then job changes in the same quartile will be associated with wage increases, and the loss for movers from plants in a high quartile to plants in 24

Table 15 IV estimates of the urban wage premium Dependent variable: log of wages of workers in multi-plants INVIND firms OLS IV OLS IV OLS IV lpop 0.012*** 0.012*** 0.008*** 0.008*** 0.007*** 0.007*** (0.001) (0.000) (0.002) (0.000) (0.001) (0.000) lpop if headq. 0.008** 0.007*** (0.003) (0.000) headquarter -0.084** -0.081*** (0.040) (0.002) lpop in HQ s LLM 0.005*** 0.005*** (0.002) (0.000) female -0.115*** -0.115*** -0.115*** -0.115*** -0.102*** -0.102*** (0.003) (0.000) (0.003) (0.000) (0.002) (0.000) part-time -0.535*** -0.535*** -0.534*** -0.534*** -0.541*** -0.541*** (0.001) (0.000) (0.008) (0.000) (0.006) (0.000) age 0.018*** 0.018*** 0.018*** 0.018*** 0.014*** 0.014*** (0.001) (0.000) (0.001) (0.000) (0.001) (0.000) age 2-0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) foreign born -0.023*** -0.023*** -0.023*** -0.023*** -0.023*** -0.023*** (0.004) (0.000) (0.004) (0.000) (0.004) (0.000) employees (1) 0.000*** 0.000*** 0.000** 0.000*** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) sector Yes Yes Yes Yes Yes Yes worker s qual. Yes Yes Yes Yes Yes Yes year Yes Yes Yes Yes Yes Yes region Yes Yes Yes Yes Yes Yes obs. 9,663,108 9,663,108 9,663,108 9,663,108 6,126,196 6,126,196 R 2 0.738 0.739 0.738 0.739 0.739 0.739 lpop lpop lpop ls 0.987*** 0.988*** 0.988*** (0.000) (0.000) (0.000) ls if HQ -0.004*** (0.000) ls in HQ s LLM 0.001*** (0.000) lpop if HQ lpop HQ s LLM ls -0.000 0.001*** (0.000) (0.000) ls if HQ 0.986*** (0.000) ls in HQ s LLM 0.987*** (0.000) Source: Inps-Invind 2005-2014. Notes: multi-plants INVIND firms omly. Robust SE clustered by LLM in parenthesis. (1) average number of firm s employees during the calendar year. 25

Table 16 IV estimates of the urban premium in plant s and worker s wage components Dependent variables: wage components at plant-year level worker component plant component OLS IV OLS IV log population 0.032*** 0.032*** 0.006*** 0.007*** (0.001) (0.001) (0.001) (0.001) female -0.169*** -0.169*** -0.029*** -0.029*** (0.004) (0.004) (0.003) (0.003) age 0.153*** 0.153*** 0.003*** 0.003*** (0.000) (0.000) (0.000) (0.000) foreign born -0.081*** -0.082*** -0.046*** -0.047*** (0.009) (0.009) (0.007) (0.007) employees (1) 0.000*** 0.000*** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) sector Yes Yes Yes Yes year Yes Yes Yes Yes region Yes Yes Yes Yes obs. 120,598 120,598 120,598 120,598 R 2 0.935 0.935 0.838 0.839 First stage: lpop lpop ls 0.988*** 0.988*** (0.000) (0.000) Source: Inps-Invind 2005-2014. Only INVIND firms. Notes: wage components collapsed at firm-year level and obtained through an AKM wage decomposition. Robust SE in parenthesis. (1) average number of employees during the calendar year. 26

Table 17 The effects of an increase in population on earnings Dependent variable: log of wages of workers in multi-plants INVIND firms OLS IV OLS IV OLS IV lpop 0.150* 0.208*** 0.142* 0.175*** 0.052 0.444*** (0.080) (0.018) (0.078) (0.018) (0.062) (0.026) lpop if headquar. 0.011*** 0.011*** (0.003) (0.000) headquarter -0.129*** -0.128*** (0.042) (0.002) lpop in HQ s LLM 0.006*** 0.006*** (0.001) (0.000) female -0.114*** -0.114*** -0.114*** -0.114*** -0.101*** -0.101*** (0.003) (0.000) (0.003) (0.000) (0.002) (0.000) part-time -0.534*** -0.533*** -0.533*** -0.533*** -0.540*** -0.540*** (0.008) (0.000) (0.008) (0.000) (0.006) (0.000) age 0.018*** 0.018*** 0.018*** 0.018*** 0.014*** 0.014*** (0.001) (0.000) (0.001) (0.000) (0.001) (0.000) age 2-0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) foreign born -0.021*** -0.021*** -0.021*** -0.021*** -0.021*** -0.021*** (0.004) (0.000) (0.004) (0.000) (0.004) (0.000) employees (1) 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) sector Yes Yes Yes Yes Yes Yes worker s qual. Yes Yes Yes Yes Yes Yes year Yes Yes Yes Yes Yes Yes LLM Yes Yes Yes Yes Yes Yes obs. 9,663,108 9,663,108 9,663,108 9,663,108 6,126,196 6,126,196 R 2 0.743 0.743 0.743 0.743 0.744 0.744 lpop lpop lpop ls 0.832*** 0.832*** 0.936*** (0.001) (0.001) (0.001) ls if HQ 0.000*** (0.000) ls in HQ s LLM -0.000*** (0.000) lpop if HQ lpop HQ s LLM ls 0.017*** 0.151*** (0.002) ls if HQ 0.098*** (0.000) ls in HQ s LLM 0.987*** (0.000) Source: Inps-Invind 2005-2014. Notes: multi-plants INVIND firms only. Robust SE clustered by LLM in parenthesis. (1) average number of firm s employees during the calendar year. 27

a low quartile will experience a smaller wage change with respect to movers in the opposite direction. Figure 2 Movers from 1st and 4th quartiles Source: INPS and INVIND 2005-2014. Table 18 and figure 2 show that workers who move from a low-paying plant to a high-paying plant experience wage increases that are increasing with the gap between origin and destination quartiles; workers who move in the opposite direction experience similar wage declines. This symmetry is in line with predictions of the AKM model. Then, according to figure 3, wages of movers who stay within the same quartile are essentially unchanged. The lack of a mobility premium suggests that idiosyncratic worker-plant match effects are not crucial in order to explain job mobility. The exogenous mobility assumption would be violated even if the idiosyncratic component of wages is associated with transitions between high-wage and low-wage plants. Thus, if wages of movers show an upward trend in the years before the move. Table 18 reveals that wages of movers show no systematic trend prior to job change. In other words, wage fluctuations do not predict mobility patterns. Another way to test for the importance of idiosyncratic worker-plant match effects is the comparison of the AKM decomposition and a match fixed effects regression. If match effects are important, such a model should perform better than the AKM model in terms of statistical fit. We find that the match fixed effects model has an adjusted R 2 that is slightly lower (0.918), and a Root MSE slightly higher (0.139) than those from the AKM regression (0.946 and 0.112). Thus, the match component in wages seems to have no relevance in explaining wages. Finally, we examined residuals from the AKM regression. Following the preceding literature, we formed deciles based on the estimated worker and plant effects, and computed average residuals in each cell. The mean residuals by cell are generally very small. In 99 cases out of 100, the mean residual is smaller than 0.01 in magnitude (in line with Macis and Schivardi, 2016). The largest deviations appear among the highest-decile workers and the highest-decile plants. 7 Conclusions In many countries workers located in big cities enjoy higher wages. In this paper we show that this wage premium can go beyond the city boundaries and affect also 28

Table 18 Mean wages before and after job change Quartiles Mean log wage of movers Wage change 2 years before 1 year before 1 year after 2 years after 1 to 1 40.1 40.8 38.5 39.4-3.6 1 to 2 47.9 47.1 53.6 53.5 12.7 1 to 3 49.6 50.5 61.1 61.0 22.0 1 to 4 53.0 56.3 70.3 73.6 31.7 2 to 1 51.3 54.7 43.0 44.6-17,4 2 to 2 58.1 59.6 58.1 58.8-0,7 2 to 3 62.8 63.5 66.5 66.0 4.9 2 to 4 66.2 72.8 74.3 76.2 8.3 3 to 1 61.3 67.1 43.8 45.6-30.4 3 to 2 66.6 71.3 61.3 63.3-9.6 3 to 3 75.9 77.6 78.0 78.9 2.2 3 to 4 84.9 84.9 96.0 92.9 11.3 4 to 1 74.3 85.0 49.7 52.7-35.7 4 to 2 78.4 88.8 67.1 70.5-17.7 4 to 3 92.4 100.2 91.3 94.3-3.6 4 to 4 103.4 106.6 110.2 112.2 5.9 Source: Inps-Invind 2005-2014. Notes: Average daily earnings for job changers observed for at least 2 years prior and 2 years after a job change. Quartiles are based on average daily wage in plants with at least 10 employees. workers located elsewhere through spillovers in productivity between plants/firms. In particular, we explore two dimensions through which wage premia may propagate: i) along plants within the same corporate structure of the firm, ii) along the value chain. Firms that supply goods and services to firms with a localization edge, may indeed indirectly benefit of this edge and pay a higher wage to their workers. Our results show that workers in branches share a part of the localization edge of their headquarter: when firm s headquarter is in an urban area, workers in branches get a wage premium of about 3%. Furthermore, working for a firm supplying goods or services to urban firms provides a wage premium that is almost as large as the one a worker would get in a urban non-supplier firm. According to our results, firms/branches that enjoy these spillovers do not hire better workers. Thus, these spillovers may reflect the fact that plants/firms connected to the city have larger rents to share with workers. All in all, these spillovers equalize wages across urban and non urban locations. According to the previous literature, in Italy the urban wage premium is smaller than in many other developed countries. Nevertheless, network externalities in wage setting across plants within a firm and across customers and suppliers along the value chain may be particularly relevant in Italy, where firms are typically smaller 29

Figure 3 Movers within the same quartile Source: Inps-Invind 2005-2014. and more connected. 30