Motorways, local economic activity and commuting

ABSTRACT This paper investigates the causal effects of the development of the Portuguese motorway network between 1981 and 2011 on the size of the local economy and the commuting of workers. We use instrumental variables based on transport networks from the late 18th century and 1945 as sources of pseudo-random variation for the location of motorways. The analysis shows that motorways have a strong effect on cross-municipal worker mobility, in terms of both incoming and outgoing commuters. We also find that motorways have a strong effect on the growth of business turnover and gross value added at the local level. Our results are robust to an extensive battery of sensitivity analyses.


INTRODUCTION
Scholars across several fields and disciplines have been forever concerned about spatial disparities in the distribution of economic activities.In this context, transport plays a key role in many spatial economic theories, which differ in terms of their assumptions and predictions.Many models focus on the micro-level, as in classical and neoclassical theories of industrial location, while other theories look more at the meso-and macro-levels, as in the case of neoclassical endogenous growth models and the more recent New Economic Geography.The relationship of transport infrastructure and economic growth has also been the object of numerous empirical studies, especially since the late 1980s, with most of the literature focusing on estimating effects at the subnational level.The meta-analyses of Elburz et al. (2017) and Melo et al. (2013) show that investment in roads, in particular, is normally associated with positive effects on output.A crucial problem, however, is that the spatial allocation of transport infrastructure is likely to be endogenous to regional growth due, for example, to reverse causality.If, for example, we assume that motorways are systematically allocated to high-growth regions to alleviate congestion, then studies that do not address the endogeneity of motorways will produce upward biased estimates and overestimate the true effect of motorways on economic output, which, in reality, could be zero, or even negative.
The endogeneity problem is at the crux of the survey by Redding and Turner (2015), which explores in detail the more recent literature that uses exogenous sources of variation in transport infrastructureinstrumental variables (IV), in most casesto identify effects on population density, economic output and other outcome variables.Baum-Snow (2007), for example, uses the highways of a 1947 plan as an instrument to estimate the causal effect of highways built between 1950 and 1990 on suburbanization in the United States.Duranton and Turner (2012) use the same 1947 plan and, in addition, the railroad network at the end of the 19th century and the routes of major expeditions of exploration of the United States between 1518 and 1850 as their IVs.They show a positive effect of highways on employment on US metropolitan statistical areas.Other interesting examples include, for example, Garcia-López et al. (2015), who focus on Spain and employs Roman roads and 1760 Bourbon roads as instruments, and Baum-Snow et al. (2017), who use road and rail networks from 1962 in China as a source of quasi-random variation in road and rail networks in 2010.Redding and Turner (2015) conclude, therefore, that the (then) available literature provides credible causal estimates of the effects of transport infrastructure.The research in the following years continued to offer similar results, reinforcing the general conclusion that transport networks have important socio-economic effects at the subnational level.Using Roman roads as instruments, Percoco (2016) finds that highway exits increase the number of plants and jobs across Italian municipalities.Möller and Zierer (2018) focus on West Germany and use historical IVs from an 1890 plan for the railroad network and a 1937 plan for the autobahn and major roads network; they find a significant positive causal effect of changes in autobahn-km on employment and the wage bill at the regional level.Levkovich et al. (2020) study how land development restrictions interfere with the effect of highways on suburbanization in the Netherlands, and show that when such restrictions exist, highwaysinstrumented in their analysis with 1821 roadsdivert population growth to locations further away from central cities.This results in a large-scale urban sprawl that skips the suburbs and 'leapfrogs' to peripheral towns.
Our paper adds evidence to this growing literature.Building on previous work on the link between motorways, urban growth and suburbanization in Portugal between 1981 and 2011 (Rocha et al., 2023), we attempt to estimate the causal effect of motorways on the growth of the local (municipal) economy, and also on the commuting mobility of workers, an aspect that has been largely overlooked in previous analyses.We address the endogeneity of the geography of motorways to long-run variations in the local economy by using IVs based on maps of roads in the late 18th century and the main roads of the 1945 National Road Plan.
More precisely, this paper contributes to the literature in the following ways.The first contribution is that we add the study of a different case, Portugal, to this literature.The country moved from having almost no motorways to having a large motorway network in about 30 years, which makes it a particularly interesting case to study.Indeed, while in the early 1980s Portugal had a total of less than 200 km of motorways, the network expanded to around 1500 km in 2000 and to around 3000 km in 2013, its current total length.According to Eurostat, in 2019 Portugal had the fourth highest ratio of motorways to population in the EU, with around 3 km/10,000 inhabitants, very close to the maximum of 3.3 km/ 10,000 inhabitants of Spain.This is approximately double the ratio of Germany or the Netherlands and 2.5 times that of Italy, the other EU countries that were analysed in the abovementioned studies.
We regard the sheer magnitude of this investment in physical infrastructure, which received ample support from EU funds, as something akin to a national-scale 'quasi-experiment', potentially with large effects on economic activity at the local level that should be investigated.Although there are already a few studies that look at the effects of motorways in Portugal, their approach is different from our study.Pereira andPereira (2016, 2019), on the one hand, estimate the impact of (monetary) investment in different types of infrastructure, including motorways, on gross domestic product (GDP), employment and private investment, but their vector autoregressive (VAR) models are implemented at the level of the country or the five mainland NUTS-2 regions.A noteworthy result is that the decline over time of the long-term output multiplier of motorway investments on GDP is extremely steep, decreasing from about €25 for the period 1979-88 to less than €4 from the mid-1990s onwards.
On the other hand, there are some recent studies at the municipality level, but they focus specifically on estimating the effects associated with the introduction in 2010 and 2011 of tolls on previously toll-free motorways (which represent about one-third of the total motorway network).This can be regarded as a source of exogenous variation, since it occurred in reaction to the sovereign debt crisis and was, therefore, a national-level decision motivated by budgetary reasons.Audretsch et al. (2020) show that the introduction of tolls caused a decrease in employment and firm formation in the affected municipalities, while Branco et al. (2021) find that firm turnover and profits also decreased in the firms located in the affected municipalities vis-à-vis firms in the remaining areas.We see our paper, which focuses on the effects of the construction of motorways at the municipal level over a 20-or 30-year horizon, as adding to this small but developing body of work on the effects of transport infrastructure in Portugal.
The second contribution of this study is that, in addition to examining employment growth, as done, for example, in Duranton and Turner (2012) and Möller and Zierer (2018), we also analyse separately jobs held by residents, 'incoming' workers (corresponding to jobs in a given municipality held by workers that reside elsewhere) and 'outgoing' workers (residents in a given municipality who work elsewhere).This breakdown of employment according to cross-municipal mobility allows us to offer new insights on the nexus between the opening of motorways and changes in the commuting patterns of workers.To the best of our knowledge, Fretz et al. (2022) is the only study to implement a similar analysis.They find that in Swiss municipalities, the number of in-commuters increases as a result of highways, while there is no effect on out-commuters.To anticipate our findings, we show that the effect of motorways is clearly larger for both incoming and outgoing commuters than for jobs held by residents.To illustrate, according to our estimates an increase of 1 SD (standard deviation) in motorways (around 12.6 km) would lead to a growth in incoming (outgoing) commuters of 21.2% (18.0%) between 1991 and 2011, whereas the effect on jobs held by residents would be 10.4% over the same period.In addition, we use our coefficient estimates to calculate the absolute number of additional jobs caused by motorways at the local level.We obtain a figure around 819,000 jobs, which is 3.3 times larger than the increase in the number of jobs at the country level over the same period.This suggests that motorways induced a vast spatial reorganization of the labour force within the country.
The third contribution is that we analyse the effect of motorways on local business turnover and gross value added (GVA), two measures of the size of the local economy.This is interesting because the studies in this literature that focus on small geographical units typically focus on outcome variables such as population, employment or the number of plants, but not on local economic output (e.g., Percoco, 2016;Levkovich et al., 2020).We circumvent existing data limitations for local business turnover and GVA by using proxy variables for their initial values in 1981 and 1991 (we employ different proxies to ensure that our results are not driven by measurement error).The main finding is that an increase of 12.6 km in motorways leads to an increase in local GVA of at least 30.4% between 1991 and 2011.
A noteworthy result is that our estimates show ordinary least squares (OLS) bias in a systematic way, indicating that the spatial allocation of transport infrastructure is indeed not exogenous.More precisely, we have produced 95 two-stage least squares (TSLS) estimates in this research and they are always clearly larger than their OLS counterparts (often by more than a factor of 2), regardless of the outcome variable being examined.According to Redding and Turner (2015), there is no a priori reason to believe that the sign of the OLS bias should be either positive or negative, as the direction of the bias should reflect differences in the political economy of infrastructure funding across countries.If IV estimates are larger in magnitude than OLS estimates, as in our paper, this suggests that the equilibrium allocation process assigns roads to places growing more slowly than places selected through a random allocation process.This type of difference occurs, for example, for the United States (Duranton & Turner, 2012) and Italy (Percoco, 2016), while the opposite appears to occur for Germany (Möller & Zierer, 2018).
The rest of the paper is organized as follows.Section 2 describes our methodology and the data we use to implement it.Section 3 discusses the validity of our IVs.Section 4 presents the main results on the effect of motorways on total employment, specific employment groups (defined according to cross-municipal mobility categories), business turnover and GVA, as well as a series of sensitivity analyses.Section 5 concludes and discusses ideas for future research.

EMPIRICAL METHODOLOGY AND DATA
We estimate a set of equations for municipality i of the following general type: where y is one of the following outcome variables: resident population, employed resident population, total employment, jobs held by residents, jobs held by non-residents ('in-commuters'), residents working in a municipality other than i ('out-commuters'), business turnover at the establishment level and GVA. 1 This equation corresponds, in effect, to a growth model, where the initial year t 0 is 1981 or the earliest year for which a specific outcome variable is available if this is after 1981. 2 Our coefficient of interest, b, measures the effect of the increase in the length of motorways, DH , on the growth of the outcome variable.
Vector X includes a number of control variables: surface area; average altitude; a measure of terrain ruggedness; log of the straight-line distance from the 1981 populationweighted municipality centroid to the coast; age of the municipality since its official establishment; a binary variable that equals 1 for Lisbon and Porto, and 0 otherwise, as the literature shows that central cities in metropolitan areas tend to lose population with motorways due to suburbanization (Baum-Snow, 2007;Garcia-López et al., 2015;Baum-Snow et al., 2017;Levkovich et al., 2020); 3 a binary variable that equals 1 if a municipality is a suburb of Lisbon or Porto, and 0 otherwise (a 'suburb' is defined here as a municipality that has a travel time to either Lisbon or Porto using 1981 roads of no more than 60 min, with travel times being calculated using 1981 population-weighted centroids); and a binary variable that equals 1 if a municipality is a district capital (with the exception of Lisbon and Porto), and 0 otherwise.We also control for motorway length in 1981 (this variable is equal to 0 for 256 municipalities), electricity consumption per capita in 1981 (as a proxy for local economic development) and district-level fixed effects w d (the country has 18 districts).See Table A1 in the supplemental data online for descriptive statistics, definitions and source details.
There are two cases for which y is not available for the initial year: business turnover and GVA.We circumvent this limitation by using several proxies such as the log of employment and log of electricity consumption to approximate the cross-sectional variation of the log of y in the initial year.Note that for the purpose of producing a point estimate for b, we can be agnostic about the exact shape of the function linking y to its proxy variable.To illustrate this, let r denote the proxy and assume that y t 0 = lr c t 0 , with l, c .0. Substituting this term in equation (1), we obtain: with a ′ = a + ulnl and u ′ = uc.That is, in this case we would estimate a model in which the constant and the coefficient of the proxy variable in the initial year are not directly interpretable, but, importantly, we are able to estimate the coefficient of interest b (as well as those of the other variables), for any arbitrary unobserved l and c.The relationship between a variable and a proxy is not, of course, an exact one, but an approximation.In other words, we should consider that y t 0 lr c t 0 instead of the equality above.This is essentially tantamount to the existence of measurement error in ln y i,t 0 in equation (1).Note that in OLS regressions measurement error in one explanatory variable may also contaminate the coefficients of the other variables (e.g., Abel, 2018).A possible concern is, therefore, that the coefficient of motorway expansion, DH i,2011−1981 , could be affected.Yet, this should not represent a problem in the context of our empirical implementation.First, we use IV methods to estimate our coefficient of interest, and there are no obvious reasons to believe that the instruments we use could be correlated with the abovementioned measurement error.Second, in practice measurement errors are likely to have a small magnitude, as suggested by the very high correlations between the log of business turnover or the log of GVA and all the proxy variables in 2011 (see section 4.2 below).Third, we use a number of different proxiesresident employed population, jobs, electricity consumption and a measure of local GDPto further mitigate concerns that our results could be influenced by measurement error.
We show in Figure 1a the growth between 1991 and 2011 of total local employment, that is, the jobs located in a municipality's territory (this variable is not available for 1981).Total employment grew in 121 municipalities and decreased in 154.Most of the municipalities with employment losses are located in the interior regions, while employment growth is clearly concentrated in coastal areas.Figure 1b and c show the growth of the two components of local employment, that is, jobs held by residents and jobs held by non-residents ('in-commuters'), respectively.It is interesting to note that jobs held by residents decreased in 208 municipalities, while jobs held by non-residents increased in 266 municipalities.This is a clear indication that the cross-municipal commuting of workers increased in general. 4Our empirical analysis below suggests that the development of the motorway network contributed to this trend.
The motorways that were built between 1981 and 2011, our main explanatory variable, represent around 86% of the current total length of motorways. 5As Figure 2a shows, in 1981 existing motorways served essentially the two metropolitan areas of the country, Lisbon and Porto; indeed, only 19 municipalities (out of 275) had motorways within their boundaries.Three decades later, that number was 157. 6The A1 corridor connected those two central cities in 1991 and, in the following two decades, the network expanded to other regions, creating connections to Spain and serving low-density regions in the interior of the country.At the same time, however, more motorways were built in the metropolitan areas of Lisbon and Porto and in the coastal strip between Lisbon and the north, increasing the network density in this part of the country.
As noted above, a critical problem is that the spatial distribution of transport infrastructure is likely to be endogenous due, for example, to reverse causality, as motorways may be preferentially allocated to high-growth areas or, in contrast, to relatively depressed regions as a way to foster their development.To address the potential endogeneity of motorways, we follow the literature that uses historical transport networks as IVs for modern-day networks and instrument DH with the length of the (mainly dirt) road itineraries of c.1800 and the 1st class roads of the 1945 National Road Plan (Figure 2b, c).The correlation between the endogenous regressor and the two instruments is, respectively, 0.39 and 0.47 (for scatterplots, see the supplemental data online), that is, both variables appear to be relevant predictors of the configuration of the motorway networksomething confirmed by the first-stage regressions below.Also, the identification of a causal effect requires the orthogonality of the dependent variable and the instruments, conditional on control variables.In particular, vector X includes geography-and history-related controls, since factors of this type may have an effect on both the instruments and the dependent variable.Likewise, it is advisable to control for the initial value of the dependent variable (or a proxy), since historical transport networks could have an influence on the size of the local economy observed at the beginning of the period of analysis, and this, in turn, may be correlated with local economic growth between 1981 or 1991 and 2011.

INSTRUMENTAL VARIABLES: ITINERARIES FROM CIRCA 1800 AND MAIN ROADS FROM 1945
The identification of our coefficient of interest hinges on the assumption that, conditional on control variables, our instruments are valid, that is, not correlated with the error term in equations ( 1) or (1´).Our first IV comes from a map of road itineraries in the late 18th century. 7More specifically, we measured the total length of itineraries that existed then in the areas that correspond to the modern-day 275 municipalities (the administrative division of the country was very different from the current one, as there were more than 800 municipalities in mainland Portugal).Roads c.1800 in Portugal were clearly worse than in other richer European countries at the time, and were often in very poor condition, in particular due to rain during the winter.Indeed, many times they simply could not be used (Link, 1803;Matos, 1980;Justino, 1988;Alegria, 1990;Pacheco, 2004).These roads were mostly used to travel by foot or by horse or donkey.An express courier travelling from Lisbon to Portoa road distance of around 313 km todaywould take about three days in 1810 (Matos, 1980), which means that the transport of goods would take much longer.Hence, roads were used essentially for short distances, while inland waterways or coastal navigation were seen as the country's primary transport network and the best option for longdistance transport of goods.In general, few people would travel, since travelling was unpractical, dangerous and too expensive (Alegria, 1990).It does not seem very likely that, after controlling for historical and geographical factors, these precarious dirt roads from the late 18th century could be correlated with the growth of employment or economic activity between 1981 or 1991 and 2011, apart from the fact that they contain information about 'natural' routes between nearby points in space and should display, because of this, a correlation with the shape of transport networks in our era.
The second instrument is the length of the 1st class roads in the 1945 National Road Plan. 8This is similar in spirit to Baum-Snow ( 2007) and Duranton and Turner (2012), who use the 1947 plan of the US interstate highway system as an instrument for, respectively, the change in highway rays between 1950 and 1990 and the level of highway-km in 1983.The authors argue that it is plausible that this instrument is valid, as the plan was designed to facilitate trade and national defence, not to facilitate metropolitan area development (Baum-Snow, 2007); in particular, the plan did not require planners to anticipate employment growth (Duranton & Turner, 2012).By a similar token, Baum-Snow et al. ( 2017) use roads and railroads in 1962 to instrument for highways and railroads in 2010 in China.As noted by the authors, the validity of their identification strategy depends crucially on the fact that Chinese roads and railroads served different purposes in 1962 than they do today.In particular, in 1962 roads existed primarily to move agricultural goods to local markets, whereas the highway system built after 1990 is designed to serve a modern urbanized economy where places of work and residence are separated and commutes are common.In other words, the 'unique context of China's transition to a market economy' (p.448) provides a credible source of pseudo-randomization of transport infrastructure by the authors' 1962 network instruments.
In our case, using the 1945 road plan as an additional instrument is also convenient, as it enables us to implement over-identification tests.To advance results, our empirical application always shows non-rejections of the joint null of no correlation with the error term of the second-stage equation.More precisely, assuming that the 'old' instrument from the 18th century is exogenous, these results suggest that the more recent instrument is also exogenous.Indeed, it is not evident why the main 1945 roads should display, conditional on control variables, a significant correlation with the local growth of economic activity between 1981 and 2011that is, other than through their correlation with the geography of motorways.
The 1945 National Road Plan involved essentially a major reclassification and modernization of pre-existing roads; some new connections were also planned.While many important works were carried out in the 1950s, in particular on the main roads (Pacheco, 2004, ch. 3), in general the plan was executed at a relatively slow pace in the 20-25 years following its approval (Sousa, 2013).The plan was developed by the autocratic Estado Novo regime , in a context where motorization rates were growing but were very low compared with more developed countries.In 1950 (1960) there were only 8.6 (21.9) cars per 1000 inhabitants (INE, 2001).9 Nothing suggests that the design of the 1945 plan could have anticipated the massive motorization era that corresponds to the period of analysis in our paper.In reality, by the mid-1970s the road network was already considered to be obsolete, with roads that were too narrow and sinuous (Pacheco, 2004, ch. 3).
The country was poor and rural.In 1950 (1960), 47.6% (42.2%) of the active population worked in the agricultural sector (INE, 2001).It is important to note that this rurality was a key element of the conservative ideological identity of the Estado Novo regime, one that permeates the motivations at the basis of the 1945 road plan.According to the original Decree-Law, it was necessary to improve road accessibility in regions with 'fertile soils' which could have a 'higher population density', in tandem with measures such as the 'arborisation of their mountains', the use of non-cultivated land, and the construction of water dams to support hydro-agricultural developments (DL34: 593, 1945, p. 373).This suggests that a main objective of the plan was to help promote a type of economic development largely based on the agricultural sector.As we argue in Rocha et al. (2023), such an anachronistic strategy has little to do with the channels through which transport infrastructure usually contributes to economic growth in open and modern economies, for example, through the attraction of industrial investments and foreign direct investment, the facilitation of international trade and innovation, etc.Another interesting aspect is the residual importance given to tourism in the 1945 plan, revealed by the fact that touristic areas were served by 3rd class roads, the least important national-level roads.This stands in contrast with the great importance of tourism in the modern Portuguese economy. 10 The arguments above suggest that it is likely that the 1st class roads of the 1945 plan do not have a 'direct' effect on the variation of employment-and output-related variables between 1981 and 2011 and, therefore, should be excluded from equation ( 1).The fact that the Estado Novo 'insulated' Portugal from the socio-economic development and modernization trends that were observed in the democratic West provides, in the context of our paper, a plausible source of identification, in that the 1945 plan was conceived to serve a rural and illiterate country, with very low levels of motorization and a largely closed economy, that is, radically different from the society and economy that developed after the democratic 1974 Revolution and the accession to the European Communities in 1986.In any case, in order to fully rule out the possibility that our results could still be contaminated by endogeneity due to the use of the 1945 instrument, we also implement models that only use instruments from 1800.

Employment and workers' commuting
As a first approximation, we look at the effect of motorways on the growth of resident population. 11Column 1 of Table 1 presents OLS and TSLS estimates of equation (1).Since t 0 is 1981we have census data for this initial year coefficient b relates the expansion of the motorway network between 1981 and 2011 to the growth of population over the same period.There is a substantial difference between the OLS and TSLS estimates, which suggests that the former may be biased due to endogeneity.The OLS estimate is much smaller.In principle, part of the motorway network may have been built in response, on the one hand, to local population, employment or income growth and, on the other, to help develop regions that were lagging behind.The low OLS coefficient suggests that the latter source of bias dominates.To advance findings, this is a strong regularity in the rest of the paper, regardless of the outcome variable being examined.While we cannot but speculate about what could be the specific causes for this difference, we suspect that the low OLS coefficient reflects, in part, the construction of motorways in low-density regions 12 that experienced negative population and/or economic growth in this period.This is certainly consistent with the fact that, in the context of the integration of Portugal in the EU, public investment in transport infrastructure was often seen as a key policy instrument to promote regional cohesion and the development of lagging regions in the interior of the country.As described by Pacheco (2004, ch. 3), the connection between the modernization of the country's road network and regional development objectives is Motorways, local economic activity and commuting 169 REGIONAL STUDIES evident in a number of official documents from the 1990sthese included generic but explicit references to the need of counterweighting negative trends in the interior regions, which, we add, in general continued to lose population in the following two decades.
According to the TSLS estimate in column 1, an SD of DH (i.e., an increase of motorway length of 12.6 km) leads approximately to a population growth of 8.8%, an effect with a substantial magnitude.The Kleibergen-Paap F-statistic is 31.5, that is, well above the usual threshold of 10, which means that the instruments are jointly relevant to explain the variation of the endogenous variable (see the first-stage coefficients in panel B1).In addition, the Hansen J-statistic is statistically non-significant, suggesting that the instruments are valid.This is also a strong regularity in the remaining of the paper.Such consistency gives us additional confidence that our excluded instruments can be regarded as exogenous (after controlling for geographical and historical factors) and, therefore, that the TSLS results can be interpreted as showing a causal effect.
In column 2 of Table 1, the dependent variable is the log of the resident employed population, which represents a more direct connection to local economic activity.The estimated coefficient is 68% larger than that for the resident population.That is, the effect on population growth is disproportionally driven by residents who have a job.This suggests a strong impact of motorways on the size of local economic activity.Columns 3 and 4 are analogous to columns 1 and 2, respectively, but now we establish 1991 as the initial year (this allows us to compare results with those in Table 2).Again, the coefficient on resident employed population is clearly larger (by 69%) than the general effect on resident population.Some of the employed people who reside in a given municipality are, naturally, commuters who work in a different municipality.In column 1 of Table 2, we consider jobs located in municipality i.In this case, we have to impose 1991 as the initial year, t 0 , as this variable is not available for 1981.The effect is substantial: 12.6 new km of motorways lead approximately to an increase in jobs of 14.0% between 1991 and 2011.In the following columns, we examine in greater detail the links between motorways and cross-municipal worker mobility.In column 2, we consider residents in i who work in i; in column 3 we consider 'in-commuters', that is, jobs located in i held by non-residents, whilst in column 4 we consider 'outcommuters', that is, residents in i who work in a different municipality.The estimates reveal that motorways promoted cross-municipal worker mobility substantially more than they promoted local jobs for residents.The effects on in-and out-commuters are very similar.More precisely, 12.6 new km of motorways lead to an increase in in-commuters of 21.2% and to an increase of out-commuters of 18.0% between 1991 and 2011.This stands in contrast to the effect estimated for local jobs held by residents (column 2) of only 10.4% over the same period.Overall, the results in Tables 1 and 2 suggest that motorways have important effects on local economic activity.Indeed, using the TSLS result of column 1 in Table 2, we estimate that motorways generated, at the local level, a total of around 819,000 jobs in mainland Portugal between 1991 and 2011.Note, however, that the actual increase in jobs in mainland Portugal in this period was approximately 245,000 jobs, or 6.4%, although this was simultaneous with a large spatial reorganization of the labour market (indeed, the average growth rate per municipality in the number of jobs in this period is 5.5%, but with a very large SD of 24.6%; for example, in 25% of the municipalities there was a decrease in the number of jobs larger than 14.8%, whereas at the other extreme of the distribution the increase in the top 25% municipalities was larger than 12.7%).
This difference in magnitudes suggests that motorways played a role in terms of the observed long-run spatial reorganization of economic activity in Portugal.It is likely, thus, that the expansion of the motorway network throughout the country has had a negative effect in many municipalities, possibly in those in the interior regions that remained, in relative terms, more distant from the network.In this respect, Redding and Turner (2015) discuss how a more complete understanding of the effects of infrastructure entails determining the extent to which the observed (gross) positive effects of transportation infrastructure reflect mainly 'new' growth or the spatial reorganization of existing activity.Yet, the authors also note that the existing reduced-form literature does not provide a basis for the separate identification of the two effects, adding that progress on this issue appears to fundamentally require an econometric framework which is capable of dealing with general equilibrium effects.The issue is complex and, as it lies well beyond the scope of this study, we leave it for future research, noting nevertheless that all results in this paper should be interpreted as reflecting, a priori, a mixture of the two effects.

Business revenues and GVA
According to the estimates in the previous section, employed population and jobs grew significantly more in the municipalities that received motorways than in those without motorways.While this is a clear indication that motorways have induced the growth of the local economy, it should be noted that the usage of a production factor (labour) is, of course, an indirect and incomplete measure of the size of economic activity.In this section, we look at the effect of motorways on non-financial firms' overall turnover and GVA.These complementary analyses allow us, on the one hand, to confirm if the same general pattern holds, and, on the other, to obtain a better sense of the approximate magnitude of the effect of motorways on the size of local economic activity.1); the coefficient of the endogenous variable is multiplied by 100 to increase readability.The number of observations is 270.In parentheses: t-statistics based on robust standard errors; *p < 0.1, **p < 0.05, ***p < 0.01.R 2 (dif) is computed using the log-difference form for the dependent variable (see note 2).All estimations include a constant and control for surface area, average altitude, terrain ruggedness, log of distance to the coast, official municipality age, length of motorways in 1981, log of electricity consumption per capita in 1981, a dummy variable for Lisbon and Porto, a dummy variable for suburban municipalities (i.e., municipalities with a travel time to either Lisbon or Porto in 1981 of no more than 60 min), a dummy variable for district capitals, and district-level fixed effects (not reported).
Table 3 considers overall turnover (revenues) at the establishment level in 2011 as the y in equation (1´).Since this variable is only available for recent years, we have to use proxies for the initial values.We use different variables to ensure that our results do not hinge on the choice of a particular proxy. 13In columns 1 and 2, we focus on the case in which t 0 , the initial year, is 1981.Given the availability of data, the proxy for the initial value corresponds respectively to resident employed population and the consumption of electricity.For 2011, the correlation between the log of revenues and the log of employed population is as high as 0.947; similarly, the correlation between the log of revenues and the log of consumption of electricity is 0.961.That is, both variables appear to be appropriate proxies for the revenues of establishments (for scatterplots, see the supplemental data online).The estimated coefficients are very similar in the two models and show that an increase of 12.6 km of motorways would generate an increase of around 53-57% in business revenues between 1981 and 2011. 14 In columns 3 and 4 the initial year is 1991.The proxies for the initial value of the dependent variable are respectively jobs located in i (recall this is not available for 1981) and, again, consumption of electricity.The correlation between the log of revenues and the log of jobs in 2011 is 0.965, which is an indication that the latter approximates well the cross-sectional variation of the former.According to the two TSLS models, an SD (i.e., 12.6 km) of motorways increases business revenues by about 43-44% between 1991 and 2011.Finally, in the last column we use 'local GDP' as the initial valuewhich is calculated by multiplying the estimates of municipal GDP per capita in 1994 produced by Ramos (1998)  15 by the census population in 1991and obtain an effect of motorway expansion on business revenues of around 48%.
In sum, the results in Table 3 show an important effect of the expansion of the motorways on business revenues and, therefore, on the size of local economic activity.If, for comparison purposes, we convert the abovementioned effects to a theoretical annual average growth rate, we estimate effects that range between 1.43% and 1.98% on an annual basis.As noted above, we use a variety of proxies for the initial values, but, importantly, the results do not appear to change significantly when a different proxy is used (interestingly, there appears to be a smaller annual effect for the longer 1981-2011 period vis-à-vis 1991-2011, which suggests that the effect of motorways on economic activity may have been smaller in the beginning when the network was still relatively small).Note: Estimates are based on equation (1´); the coefficient of the endogenous variable is multiplied by 100 to increase readability.The number of observations is 270.In parentheses: t-statistics based on robust standard errors; *p < 0.1, **p < 0.05, ***p < 0.01.All estimations include a constant and control for surface area, average altitude, terrain ruggedness, log of distance to the coast, official municipality age, length of motorways in 1981, log of electricity consumption per capita in 1981, a dummy variable for Lisbon and Porto, a dummy variable for suburban municipalities (i.e., municipalities with a travel time to either Lisbon or Porto in 1981 of no more than 60 min), a dummy variable for district capitals, and district-level fixed effects (not reported).In column 5, '1991(a)' means that local gross domestic product (GDP) was obtained through multiplying the estimates of municipal GDP per capita for 1994 in Ramos (1998) by the census population in 1991.
Table 4 has the same structure as Table 3 in terms of the proxies for the initial values.The difference is that the dependent variable is GVA, which is a usual measure of economic activity.Indeed, the correlation between the log of this variable and the log of GDP in 2011 at the NUTS-3 level is 0.9914, that is, very close to unity (for a scatterplot, see the supplemental data online).A relevant aspect is that GVA is provided at the firm level and not at the establishment level, that is, GVA is allocated in its totality to the headquarters of the firm.In order to adjust the variable for this locational discrepancy, we make use of the fact that turnover is provided at both the firm and establishment levels and compute the ratio between the two to obtain a municipality-specific measure of this discrepancy.More formally, first we calculate: Turnover firms in i Turnover establishments in i ; this ratio has a mean (median) of 0.91 (0.93) and, unsurprisingly, the maximum value (1.47) occurs for Lisbon, the capital, where the headquarters of many firms are located.Second, we use this ratio to compute a location-adjusted version of GVA that is meant to approximate the GVA that is generated at the establishment level.That is: The log of this adjusted GVA is, therefore, the dependent variable in the regressions of Table 4; the correlation for 2011 between this variable and the proxies for the initial values are 0.951 for resident employed population, 0.963 for the consumption of electricity and 0.969 for jobs.
The results are similar to those in Table 3, albeit the estimated coefficients are somewhat smaller.The effect of 12.6 new km of motorways on GVA is of 42-45% between 1981 and 2011 (columns 1 and 2) and around 30-36% between 1991 and 2011 (columns 3-5).As a comparative exercise, in terms of annual average growth rates, this would translate to a minimum of 1.16% (column 1) and a maximum of 1.53% (column 5).Even if we select the most conservative estimate, motorways appear to generate a considerable contribution to the growth of the local economy.

Robustness tests
The analysis in sections 4.1 and 4.2 shows in a systematic way that motorways induce the growth of the local economy, regardless of the specific variable used to capture it.
In part due to limitations in data availability, the analysis uses different variables for different periods.Results, however, are very similar when fully comparable and overall internally consistent.For example, as expected, effects estimated over a 20-year period are always smaller than effects estimated over a 30-year period for the same dependent variable, regardless of the proxy for the initial value of the dependent variable that was used.Also, the coefficient estimated for jobs between 1991 and 2011 (1.038) is much smaller than the coefficient estimated for GVA for the same period (2.266), reflecting the fact that local output grows not only with the increase in labour but also with increases in capital and productivity.In this section, we submit our analysis to an extensive battery of additional tests (to avoid repetition, most of these supplementary results are reported in the supplemental data online).

Location-adjusted GVA
First, we implement a different way of adjusting GVA for the fact that, as noted above, this is fully allocated to firms' headquarters.In this alternative approach, we run an auxiliary regression in which GVA in 2011 is estimated using two predictors that do not suffer from that problem, business turnover at the establishment level and jobs.We obtain the following relationship (t-statistics shown in The coefficient of determination is close to 1, that is, almost all of the cross-sectional variance of the log of GVA is explained by these two predictorsit is plausible that the remaining part can be ascribed, to a significant extent, to the abovementioned locational distortion.The fitted value for the log of GVA is, thus, the dependent variable in the regressions of Table 5.The results are very similar to those in Table 4, although the estimated TSLS coefficients are slightly larger.We consider the more conservative estimates in Table 4 as our preferred ones and report these additional results as a supplementary analysis that confirms the main message that motorways have a strong impact on the growth of the local economic activity.

Motorway access nodes
Second, we take into account the fact that some municipalities are traversed by a motorway but do not have any access ramps.More precisely, there are 23 municipalities for which our dependent variable, the variation in the length of motorways between 1981 and 2011, is positive and the number of access ramps in 2011 is zero.If the effect of motorways on the size of economic activity is smaller in these municipalities, then the average effects that we report in Tables 1-4 could underestimate the effects of (connected) motorways on local economic activity.We adapted our explanatory variable by changing it to zero for those 23 municipalities and re-estimated all the models (see Tables A2 and A3 in the supplemental data online).While all the 17 TSLS coefficients that have a (more or less quantitatively direct) connection with the size of local economic activity 16 are, indeed, larger than the correspondent coefficients in Tables 1-4, the differences are small.This may mean that, in most cases, those municipalities are sufficiently close to an access ramp to benefit from the motorway network.Third, we consider as the endogenous explanatory variable the log of the travel time by road from each municipality centroid to the nearest motorway access node.This allows us to consider specific information for each of the municipalities that have no motorways, as our length-based measure (km of motorways) is equal to zero for these 122 municipalities.The results are shown in Tables A4 and A5 online.Although the first-stage identification is not as strong as beforethe Kleibergen-Paap rk Wald F-statistic is typically around 22-24, compared with 27-31 in Tables 1-4 we find patterns that lead to analogous conclusions.That is, economic activity grows less in those municipalities that are more distant from a motorway access node.The ordering in the size of the estimated coefficients (in absolute terms) is essentially the same as beforein increasing order: population, employed population, employment, (location-adjusted) GVA and establishments' turnover.4.3.3.Roads of the late 18th century in just-and over-identified estimates Lastly, it could be argued that the road plan of 1945 constitutes a relatively recent instrument and thus may not be completely exogenous.Note, however, that we control for a range of factors, including, by construction, the initial value of the outcome variables (arguably this may capture channels through which roads in 1945 could have an influence on the growth of the outcome variable between 1981 or 1991 and 2011).Indeed, so far we have applied the Sargan-Hansen test 59 times, and failed to reject the joint null hypothesis that the instruments are uncorrelated with the error term in all the cases.Yet, in order to fully Note: See Table 3.The first-stage coefficients of the excluded instruments are also the same as in Table 3. GVA is adjusted as described in section 4.3.1.

REGIONAL STUDIES
rule out the possibility that the use of this instrument could induce bias, we present TSLS estimates that use the itineraries of c.1800 as the sole IV in Tables A6 and A7 in the supplemental data online.In addition, we use our geographical information system (GIS) to create a second instrument from 1800the log of the distance of the 1981 population-weighted municipality centroid to the nearest road itinerary.As shown in Tables A8 and A9 online, we continue to reject the null in overidentification tests.These just-and over-identified estimates with 'old' instruments are similar to our main TSLS results, suggesting that these can be regarded as reflecting a causal effect of motorways on the size of local economic activity.

CONCLUSIONS AND FURTHER RESEARCH
This paper analyses the impact of the creation of the motorway network in mainland Portugal on the size of the local economy and the patterns of workers' commuting.We show that motorways have a strong effect on the growth of employed population and local employment, establishments' turnover and GVA.Whilst local economic development dynamics are certainly the result of a mix of interlinked determinantswhich may include other types of infrastructure, human capital, local public policies, among othersour results indicate that motorways are a relevant contributing factor.We find, in addition, that the effect of motorways on the growth of local employment is larger for the jobs of non-residents (i.e., in-commuters) than for the jobs of residents in a given municipality.The effect on the growth of jobs of residents who work outside their municipality is also particularly large.These findings suggest that motorways caused an important expansion of cross-municipal commuting by expanding the pool of potentially interesting jobs available to people within a certain travel time radius.The estimates in this paper reflect, in principle, a causal effect, as we use IV methods to address the issue that the spatial distribution of motorways is unlikely to be exogenous.The OLS estimates are also positive and statistically significant, but systematically smaller than the TSLS estimates.This is evidence suggesting that endogeneity is present and that the bias is large and negative.As the sources of bias remain unknown to the researcher, we cannot be sure about the extent to which reverse causality is, in this context, an important cause of bias.Assuming this is the case, this may mean that investments in motorways are seen by policymakers as a way to promote the development of lagging regions, a hypothesis that is certainly consistent with the broad aims of EU-funded infrastructure development programmes.Yet, the rhythm of investment and the specific location of motorways in low-density regions may be, in principle, influenced by political economy factors, in which motorways (and even motorway access nodes) can be regarded as important political assets in the interplay between local political actors and the national government.We intend to explore this issue as part of our future research programme.
Our results are robust to alternative ways of capturing the expansion of the motorway network (i.e., the use of different endogenous variables), changes in the specification of the first-stage equation, and, perhaps more importantly in the context of this paper, the use of different proxies for the initial values of turnover and GVA and the implementation of an alternative method to allocate GVA to the municipality where this is generated (this is necessary because in the original variable GVA is fully allocated to the firms' headquarters, although many firms have establishments in other municipalities).Indeed, as our attempt to circumvent existing data limitations is a relevant part of the analysis, we take particular care to ensure that the thrust of our results is not sensitive to the choice of, for example, a particular proxy variable for the initial value of turnover or GVA.The fact that results change little with using different proxy variables suggests that our approach is reasonable.
According to our estimates motorways created around 819,000 jobs at the local level between 1991 and 2011.However, total jobs in mainland Portugal only increased by about 245,000 over the same time span.We regard this huge disparity as indirect evidence that the motorway network must have induced a large-scale spatial redistribution of employment, and, therefore, of economic activity in the country.This is an issue that we expect to investigate in detail in the near future.As emphasized by Redding and Turner (2015), although the distinction between 'genuine' growth and the spatial reorganization of economic activity is central to the understanding of the role of infrastructure in an economy, unfortunately relatively little effort has been directed to the study of this topic.
It should be mentioned that our focus is on estimating average effects over a long period of time.We do not attempt to characterize the shorter-run dynamics that are concealed in these effects.As noted, again, by Redding and Turner (2015), the literature has devoted little attention to the dynamics of how transportation infrastructure affects economic development, a topic that the authors see as an important, though difficult area for further research.In our case, as probably for many other countries, the lack of data for several variables at the municipality level (in particular for earlier years) precludes the implementation of panel data analyses.Finally, in the future it would be interesting to investigate whether specific motorways in the network could offer other possibilities for the identification of causal effects.Ciani et al. (2022), for example, use the original plans for a motorway in Calabria, south of Italy, which envisioned three feasible routestheir identification is elegant in exploiting the fact, revealed by historical studies, that the selection of the final route was driven by the desire to make it pass in the hometown and constituency of two powerful politicians.As more archival research is conducted in this area in Portugal and more information on the planning and decision processes of motorways becomes available in the future, it may be possible to identify particular settings that allow for the plausible estimation of causal effects.
Motorways, local economic activity and commuting 175 REGIONAL STUDIES NOTES 1.These variables were sourced from INE (Statistics Portugal).Business turnover and GVA refer to non-financial firms (including independent workers) and are calculated by INE using administrative firm-level data.In its original form, GVA is allocated in its totality to the headquarters of the firm.We adjust GVA for this geographical distortion by considering its relationship with both business turnover at the establishment level and local employment.The procedures are detailed in sections 4.2 and 4.3.2. That is, equation ( 1) is equivalent to: ln y i,2011 − ln y i,t 0 = a + u ˘ln y i,t 0 + . . ., with u ˘= u − 1 .
3. The country has two metropolitan areas, which were officially established in 1991.They concentrate about half of the population and jobs in mainland Portugal.
For the suburbanization dynamics that were observed between 1981 and 2011, see Rocha et al. (2023, passim).4. At the (mainland) country level, in 1991 approximately 75% of the jobs were held by residents in the municipality where their jobs were located.That share decreased to around 66% in 2011. 5.As part of our robustness analysis in section 4.3, we consider other explanatory variables.First, we exclude motorways with no access nodes; and second, we consider the distance of the municipality centroid to the nearest motorway access node.6.These numbers refer to the pre-1998 administrative division of the country.Currently mainland Portugal has 278 municipalities.In our empirical implementation we use 270 municipalities, as, to ensure fully comparability for all variables across time, we had to exclude the five municipalities from which the three municipalities created in 1998 were formed.7. The arguments for the identification strategy in this section draw from section 4.2 in Rocha et al. (2023).
The first instrument comes from Matos (1980), which was digitalized with geographical information system (GIS) software by Sousa (2010).According to Martins (2014), the map by Matos (1980) reflects essentially information published in 1767.The same author notes that in 1748 the historical list of itineraries is almost the same as in 1767.8.The maximum speed limit on 1st class roads is 100 or 80 km/h.The maximum speed limit on current motorways is 120 km/h.9.For comparison, motorization rates in recent years are around 500 cars per 1000 inhabitants.10.In 2017, the travel and tourism sector represented a direct (total) contribution of 6.8% (17.3%) of GDP.For comparison, in the EU this share was 3.9% (11.7%) (World Travel & Tourism Council (WTTC), 2018).11.This is the main objective of Rocha et al. (2023).
Although our present specification is slightly different (e.g., we use a more complete set of control variables) and the sample is slightly smaller, since we consider 270 municipalities instead of 275 (see note 6 above), the estimates in both papers are similar.12. Between 1981 and 2011, motorways were built in 153 municipalities.Of these, 90 (63) had a population density lower than 150 (100) inhabitants/km 2 in 1981 (as a reference we consider the threshold of 150 inhabitants/km 2 used by the Organisation for Economic Cooperation and Development (OECD) to identify 'predominantly rural' regions; www.oecd.org/regional/regionalstatistics/geographical-definitions.htm).13.As noted in section 2, what is needed is a proxy whose logarithm approximates closely the cross-sectional variation of the log of (unobserved) turnover in 1981 (i.e., it is indifferent if we use as a proxy variable r or lr c with arbitrary l, c .0; the estimates of our coefficient of interest, b, would be identical).14.Our results on business turnover and GVA are in real terms.If we deflate the dependent variable for the 30-or 20-year inflation, we obtain ln y i,2011 + ln (1/(1+ inflation)).The second term is a constant and, therefore, has no influence on the estimation of the coefficient of interest.15.Ramos's approach to estimate local GDP per capita is detailed in Table A1 in the supplemental data online.16.That is, 18 minus the coefficient on 'outgoing' jobs, that is, workers who reside in municipality i, but work elsewhere.For completeness, we also report this coefficient in the supplemental data online.

Figure 1 .
Figure 1.Employment growth and the cross-municipal commuting of workers, 1991-2011: total jobs (a), jobs held by residents (b) and jobs held by non-residents (c).

Figure 2 .
Figure 2. Road networks: evolution of the motorway network (a), road itineraries c.1800 (b) and 1st class roads of the 1945 National Road Plan (c).

Table 1 .
Motorways and the employed population.
Note: Estimates are based on equation (1); the coefficient of the endogenous variable is multiplied by 100 to increase readability.The number of observations is 270.In parentheses: t-statistics based on robust standard errors; *p < 0.1, **p < 0.05, ***p < 0.01.R 2 (dif) is computed using the log-difference form for the dependent variable (see note 2).All estimations include a constant and control for surface area, average altitude, terrain ruggedness, log of distance to the coast, official municipality age, length of motorways in 1981, log of electricity consumption per capita in 1981, a dummy variable for Lisbon and Porto, a dummy variable for suburban municipalities (i.e., municipalities with a travel time to either Lisbon or Porto in 1981 of no more than 60 min), a dummy variable for district capitals, and district-level fixed effects (not reported).170BrunoT. Rocha et al.
Note: Estimates are based on equation (

Table 3 .
Motorways and business turnover.

Table 4 .
Motorways and gross value added (GVA).The first-stage coefficients of the excluded instruments are also the same as in Table3.GVA is adjusted as described in section 4.2.