You are on page 1of 12

Transportation Research Part A 65 (2014) 44–55

Contents lists available at ScienceDirect

Transportation Research Part A


journal homepage: www.elsevier.com/locate/tra

Ports and regional development: A spatial analysis on a panel


of European regions
Anna Bottasso, Maurizio Conti, Claudio Ferrari ⇑, Alessio Tei
DIEC – Department of Economics, University of Genoa, Via Vivaldi 5, 16126 Genoa, Italy

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyses the impact of port activities on local development for a sample of 621
Received 18 October 2013 TL-3 regions located in thirteen European countries and observed over the period 1998–
Received in revised form 18 February 2014 2009. Using a spatial panel econometric framework which controls for spatial fixed effects,
Accepted 8 April 2014
the paper provides an estimate of both the direct and indirect (i.e. spillover) effects asso-
Available online 13 May 2014
ciated to port activities. Results suggest that ports might have non-negligible effects on
local GDP: interestingly, an important share of the effects takes place outside the region
Keywords:
where the port is located.
Port impact
Regional development
Ó 2014 Elsevier Ltd. All rights reserved.
Port spillovers

1. Introduction and theoretical background

Ports are important gateways for domestic and international trade since about 80% of the world trade depends on mar-
itime transport (UNCTAD, 2009). Given their crucial role as nodal infrastructure, ports have traditionally been considered as
a strategic economic endowment able to connect global and local markets favouring the globalization process. For instance,
in the case of developing countries, some studies (Radelet and Sachs, 1998; Chowdhury and Erdenebileg, 2006) suggest that
landlocked regions might suffer of a GDP gap with respect to port regions that could be as large as about 40%, ceteris paribus.
More generally, policy makers have always viewed transport infrastructure, of which ports are an important element, as a
key factor capable to foster territorial cohesion, mitigate economic disparities, favour economic development and conver-
gence. For instance, about 30% of the EU Regional Development Funds (ERDF) and Cohesion Funds planned for the period
2007–2013 have been devoted to infrastructure investments and a significant share of such budget has been allocated to
transport infrastructure investments aimed at improving regional endowments and completing trans-European Transport
networks (TEN’s): ‘‘.. roads, high speed rail lines, freight shipping ports, and airports financed through TENs schemes are
expected to bring about major EU-wide transformations, not only by removing bottlenecks and breaks in the EU transport
system, but also in terms of improving regional GDP per capita, promoting employment, facilitating mobility, and enhancing
accessibility, as reflected in the assessment criteria for these policy measures’’ (EspoN, 2009). Economic theory traditionally
has supported such beliefs since it suggests that services provided by transport infrastructures positively impact industry
productivity in different ways, the main channel being a reduction of time and transport costs, which in turn can have dif-
ferent implications such as higher productivity of other inputs, lower production costs, greater specialization, growth of
trade, more intensive competition, enlargement of relevant markets, improvement in the division of labour, better access
to foreign intermediates and exploitation of scale economies (Aschauer, 1989); furthermore, greater accessibility and lower
transportation costs contribute to raise the market potential of different locations (Niebuhr, 2006; Condeco-Melhorado et al.,

⇑ Corresponding author. Tel.: +39 010 2095235; fax: +39 010 2095511.
E-mail address: ferrari@economia.unige.it (C. Ferrari).

http://dx.doi.org/10.1016/j.tra.2014.04.006
0965-8564/Ó 2014 Elsevier Ltd. All rights reserved.
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 45

2011). However, the prediction of high returns of public infrastructure investment has been recently questioned on both the-
oretical and empirical grounds (Afraz et al., 2006; Romp and de Haan, 2007). The adoption of different econometric tech-
niques, as well as of different proxies for infrastructure and, especially, levels of data aggregation has provided mixed
results (e.g. Crescenzi and Rodriguez-Pose, 2012); moreover, some empirical researches have shown that infrastructure
investments have hampered regional convergence while, at the same time, have promoted national convergence, both
within the EU and the USA (e.g. Puga, 2002; Sloboda and Yao, 2008).1 Such issue can be better discussed within the theoretical
framework of New Economic Geography (NEG) models (e.g. Krugman, 1991; Puga, 2002) which account for the possibility that
variations in attractiveness and accessibility of locations induced by infrastructure investments might have potentially opposite
effects on the geographical distribution of economic activity. According to NEG models ‘‘the spatial concentration of economic
activities is fostered by more pronounced market size asymmetries (‘market seeking’) and production costs asymmetries (‘cost
saving’). Concentration is also promoted by easier tradability of products and factor services between locations as well as by
larger differentiation of firms in terms of products (‘horizontal differentiation’) and smaller differentiation in terms of
quality/productivity (‘vertical differentiation’). Once agglomeration economies are at work, market size and production cost
asymmetries become endogenous as they not only determine but are also determined by firms’ location decisions’’
(Ottaviano, 2008). Such asymmetries are partly determined by the presence of different types of infrastructure: in particular,
transport infrastructure improvements may reduce transport and trade cost and affect locations’ market potential and localiza-
tion of economic activities. On the one side, an infrastructural improvement may induce a more even distribution of economic
activities when the prices of non-tradables are much lower in less developed locations (Puga, 1999), when it promotes long-
distance commuting (Borck et al., 2007) and when it favours the transmission of knowledge from developed to less developed
locations: if knowledge spillovers generate strong agglomeration economies, production cost asymmetries fall, thus supporting
a more even economic geography (Baldwin et al., 2001). On the other side, transport infrastructure investments may exacerbate
geographical disparities: if a less developed region becomes better connected to a more developed location, unless the prices of
non-tradables (e.g. land rent, housing costs) are much higher in the developed region, it may be that better transportation
improves its market potential more than it improves the market potential of its less developed trading partner so that regional
disparities might increase (‘‘straw effect’’).
Moreover, improved transport infrastructures are more likely to favour the attractiveness of a location in the presence of
and close to hubs and gates (Krugman, 1993) which may enjoy the expansive effect of positive demand shocks occurred to
other locations. Finally, ‘‘the development of transport infrastructure, by increasing the accessibility of weaker regions, not
only gives firms in less developed regions better access to inputs and markets of more developed regions [. . .] but it also
makes it easier for firms in richer regions to supply poorer regions at a distance, and can thus harm the industrialization
prospects of less developed areas’’ (Puga, 2002).
By modelling the interaction among agglomeration/dispersion forces, transport costs and imperfect competition, NEG
models predict that the relationship between geographical concentration and transport costs is non-linear: when transport
costs are high or low, economic activities are more dispersed, while they tend to cluster when transport costs are at an inter-
mediate level. In fact, when the price of non-tradables increases as a consequence of agglomeration, dispersion may take
place if transport costs become lower than commuting costs; moreover, if transport and communication costs are suffi-
ciently low, firms may find it profitable to relocate activities to the periphery in order to lower production costs (e.g.
Fujita and Thisse, 2006). Transport costs are directly associated to the existence of a well developed and efficient network
of transport infrastructures whose nodal components, like ports, play an important role. For a long time ports have been con-
sidered as an engine for growth and development at both local and aggregate level since they allow the globalization process
to fully generate its benefits through the development of trade; moreover, the expansion of value added logistics activities
taking place near ports, like customization, labelling, final assembly, configuration, blending and so on, has increased the
multiplicative impact of port activity on the port industry itself and on port related activities. The most important theoretical
contribution supporting the idea that ports provide a comparative advantage to the cities where they are located is the Fujita
and Mori (1996) model which shows how a port may create endogenous growth by providing a competitive advantage to the
economic activities located around it and explains the positive effects of transport nodes (such as ports) in the process of
spatial economic development; furthermore authors suggest that the continued prosperity of port cities can be better under-
stood if lock-in effects of some self-enforcing agglomeration forces are taken into account.
Port efficiency has also been considered as an important factor which enables growth and development. For instance,
Clark et al. (2004) suggest that the effective protection rate for local producers enabled by transport costs is, in most cases,
higher than those entailed by tariffs and show that efficient ports considerably reduce shipping costs, while inefficient ports
reduce trade and growth. Also Haddad et al. (2010) suggest that port costs can be considered as an important barrier to trade
liberalization and develop a spatial CGE model integrated to a transport network system in order to simulate the impacts of
increases in port efficiency in a context of trade liberalization. Moreover, some studies suggest that port regions grow faster
than landlocked ones (e.g. Radelet and Sachs, 1998; Wilmsmeier et al., 2006), while Ferrari et al. (2010) and Bottasso et al.
(2013) find that port throughput positively affects local employment for a sample of Italian and European regions, respec-
tively. Finally, Kawakami and Doi (2004) show, by means of a VAR analysis, that ‘‘private capital is indirectly caused by port

1
Some authors have analyzed the existence of spillover effects of infrastructure investments (Holtz-Eakin and Schwartz (1995), Cohen and Morrison (2004),
Cohen (2010) and Boarnet (1998) among others) providing mixed evidence.
46 A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55

capital’’ in the case of the Japanese port sector.2 However, during the last decades, a series of important changes have occurred
in the port and shipping industries which, together with the new trends characterizing the economic growth process, have led
several authors to argue that ports economic benefits tend to be increasingly concentrated outside the region where the port is
located (Musso et al., 2000; Ducruet, 2009; Monios and Wilmsmeier, 2012; Iannone, 2012 among others). First, the container-
ization of traffic and the associated growth of ships average size have significantly reduced time and transport costs: such trends
have generated a selection and a reorganization process among ports and have favoured the diffusion of new practices, like hub
and spoke strategies, traffic concentration and service rationalization; furthermore, port related industrial sectors have become
more capital intensive, thus reducing possible positive employment effects associated to port activities. In second place, high
value added manufacturing and service industries have reduced traditional heavy industries in the growth process: the increas-
ing tertiarization of most advanced economies has reduced the demand for port logistics services while high value added prod-
ucts are increasingly transported by road, air and train. Third, the globalization process has induced the relocation of some
manufacturing activities towards countries characterized by lower production costs, thus reducing the demand for international
transport for some western ports (e.g. Gripaios and Gripaios, 1995; Lee et al., 2008; Grobar, 2008). Finally, a negative role has
also been played by congestion brought about by traffic growth which has reduced the attractiveness of port neighbouring areas
and have therefore generated relocation processes of activities and port related services: different case studies have in fact
shown that the latter tend to move from port cities to distriparks, free zones and multimodal platforms.3 This process, also
known as de-maritimisation, tends to increase, within the port regions, the conflicts associated to the existence of social costs
(congestion, pollution, etc.) generated by port activities that are not perceived to be accompanied by countervailing economic
benefits (Ferrari et al., 2006). The above discussion suggests that the agglomeration power of ports depends on the balance
between agglomeration and dispersion forces, which is not time invariant and can be affected by several factors associated
to technology development, changes in composition of economic activities, policy interventions and ecological consciousness,
among other things (Ducruet et al., 2012). Moreover, the agglomeration power of transport nodes like ports has been questioned
arguing that agglomeration forces are commodity specific and depend on regional specialization (Tabuchi and Thisse, 2002).
However, within the debate on the economic benefits of ports activity, most authors argue that positive economic effects are
getting more dispersed so that the relevant area which potentially benefits from multiplicative effects stemming from ports
activity is not well defined (e.g. Ducruet et al., 2012). The possibility that port benefits sprawl to a larger area, often wider than
the port region, raises different research questions: the main one is connected to the quantification of possible spillover effects
in the neighbouring regions. The evaluation of such spillovers could in fact be used both as a basis to ‘‘compensate’’ the pop-
ulation in the port region for the social costs associated to negative externalities generated by the port and as an economic jus-
tification for possible co-financing requests for new port-related infrastructures, or in order to avoid that some regions reap the
benefits arising from the presence of ports without paying any costs. At the same time the existence of positive spillovers makes
the governance of seaports a matter of interests also for the neighbouring (non-port) regions.
This work aims at studying the economic effects of ports’ total production on a sample of 621 TL3 regions of thirteen wes-
tern EU countries, observed over the period 1998–2009; in particular, we consider the port throughput of the largest 150
ports localized in 120 port regions. The novelty of this study is that we provide estimates of both the direct and indirect
(spillover) effects of port throughput using an appropriate empirical framework based on recent developments in spatial
panel data econometrics. A similar approach has been adopted by Cohen and Monaco (2010) for the estimation of a cost
function for the US manufacturing sector, who found that an increase in the level of port infrastructure (proxied by the
depreciated cumulative investments in port infrastructure) tends to reduce production costs within a state, but they also find
evidence of negative spillovers in neighbouring states. In turn, Cohen and Monaco (2009) use a panel of California counties
and estimate production functions for the retail sector ‘‘augmented’’ with the stocks of port infrastructure within the county
and in neighbouring counties and find that port infrastructure investments tend to increase production in the retail sector
both in the county where they are undertaken and in adjacent ones.4
The present paper is an empirical contribution on the economic importance of seaports that could usefully inform the
policy formulation on the port industry. It is organized as follows: Section 2 presents the data while Section 3 discusses
the main methodological issues that arise in the empirical work. Finally, Section 4 presents the main empirical results while
Section 5 concludes.

2. Data

The sample considered in this study is based on 621 TL-3 regions belonging to thirteen European countries, namely
Belgium, Denmark, Finland, France, Germany, Ireland, Italy, Norway, Portugal, Spain, Sweden, The Netherlands and United
Kingdom, observed over the period 1998–2009. Out of the total, 120 regions (about 20% of the total) have been classified

2
See also Deng et al. (2013) who analyze the impact of a sample of Chinese ports on the regional economy using a Structural Equation Modelling framework.
Coppens et al., 2007 and Danielis and Gregori, 2013 provide case studies based on the input–output approach.
3
The above discussion does not fully apply to less developed countries and to emerging countries which are at their first stages of the industrialization
process.
4
In both papers Cohen and Monaco deal with estimation issues arising from the spatial nature of their data through a Cochrane–Orcutt transformation of the
data, which is valid insofar as spatial autocorrelation in the data do not arise from spatial omitted variables, as we believe it might be the case when one studies
the effect of ports on local development.
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 47

Fig. 1. Sampled regions. Source: Authors’ elaboration. Black stars represent the location of main included ports.

as ‘‘port regions’’ because they host one or more ports; therefore we consider the activity associated to the largest 150 ports
located in the TL3 regions analyzed in this paper. Fig. 1 shows both the regions as well as the location of the main ports
considered in this work. In Fig. 1, black stars represent the localization of the biggest port regions within the sample.
Northern Europe accounts for the main share of port traffic: in fact major European seaports are located between Belgium,
the Netherlands and Germany; moreover, while northern countries count fewer ports than Mediterranean ones, they are
responsible for a greater share of total port activity in Europe, which in turn suggests the existence of some concentration
in the port sector.
The data used in this study are organized into two main categories: regional characteristics and port traffic information.
Regional data are defined at the OECD TL-3 level, although some information (i.e. the accessibility index) is available only at
the Eurostat NUTS 3 level, so that a conversion of NUTS 3 into TL3 has been necessary.5 Regional data have been collected
from the OECD database, Eurostat and the European Spatial Planning Observation Network (EspoN). From the first source we
have retrieved information on variables such as regional Area, GDP, Population and Patents, all aggregated at the TL-3 level. Data
on motorways length have been taken from Eurostat (at NUTS 2 level and converted at TL3), while EspoN was the source for
information on the regional accessibility index that we have considered in some model specifications. In particular, this index
computes accessibility through a ratio that considers the land characteristics and the infrastructural endowment of each region,
so that it may be considered a measure of the ‘‘connectivity degree’’ among regions (EspoN, 2009).
Port data have been extracted from the Eurostat database on port traffic. Minor ports have been excluded when their mar-
ginal contribution to the overall traffic was close to zero; in particular, as a measure of port activity we have considered the
port total throughput, thereby neglecting passenger traffic data. Indeed previous empirical evidence has shown that the eco-
nomic impact of passenger traffic is not particularly relevant, probably because a significant share of that traffic is related to
transit passengers (Bottasso et al., 2013). Table 1 reports summary statistics for the main variables considered in the empir-
ical analysis. In general, port regions are on average larger than non-port ones, while the latter are more densely populated
(mainly because of the presence of important metropolitan area such as Berlin, Madrid, Milan, Munich and Paris). Although
not reported in Table 1, per capita GDP (as well as patents per capita) is higher in the case of port regions. Similarly, the data
reveal that the density of economic activity, labelled as ‘‘Econ_Den’’ and computed as GDP/area, tend to be slightly higher in
non-port regions, suggesting that agglomeration economies might be more important in those regions with an higher con-
centration of economic activity. As far as the accessibility variables are concerned (both the accessibility index and motor-
ways network behave similarly), summary statistics suggest that non-port regions are slightly more accessible: this fact
could in part be explained recalling that both indexes refer to land infrastructures; however, the lower connectivity of port

5
It is important to bear in mind that the level of aggregation TL-3 and NUTS-3 are identical for most countries in our sample with few exceptions, namely
Belgium, Germany and The Netherlands.
48 A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55

Table 1
Summary statistics. Source: Authors’ elaboration based on Eurostat EspoN (2009) and OECD (2013) databases.

Mean Median Maximum Minimum St. dev.


Port regions (120)
Area 6429.8 3312 98910.7 40 12897.7
Population 723401.8 423099 5345603 9903 822450.5
GDP 24498.6 1988.3 234104.1 144.6 32613.4
Econ_Den 18.9 3.9 173.8 0.04 36.3
Patents 81.0 21.2 2483.6 0.0 201.3
Accessibility 89.1 84.6 167.2 27.3 33.1
Motorways 367.2 254.0 2379 5.0 379.7
Throughput 19139.2 10165.5 384210.0 0.0 34300.6
Non-port regions (501)
Area 4898.5 3509 55401.2 13.4 5662.8
Population 531124 390272 6295011 21,137 528662.2
GDP 16538.2 10255.9 229161.6 359 22151.9
Econ_Den 20.5 3.5 1789.8 0.05 109.4
Patents 60.8 19 1.679.4 0.0 127.6
Accessibility 94 93.2 193.4 28.5 34.1
Motorways 397.1 300 2379 5 365.1
Throughput / / / / /

Notes: Area: squared km; GDP: millions of 2005$ at purchasing power parity; Population: number of habitants; Patens: number of applications; Motorways:
kilometres of network; Density: GDP/area.

regions might suggest the existence of possible infrastructure bottlenecks around port areas which in turn might not allow
the exploitation of the potential economic benefits of EU ports. However, it is also important to highlight the relatively large
standard deviation for both port and non-port regions, which in turn suggests that within both groups there are non-
negligible differences in terms of accessibility and connectivity. We account for regional heterogeneity by including in
our model specification regional fixed effects as well as control variables such as motorways length, degree of innovation
and density of economic activity.

3. Methodology and empirical strategy

Let’s consider a simple panel linear regression model with spatial specific effects:
yit ¼ li þ xit b þ eit ð1Þ

where i denotes the cross section with i = 1, . . ., N, t is the index for the time dimension with t = 1, . . ., T, yit is an observation on
the dependent variable at i and t, xit an (1, K) row vector of observations on explanatory variables, and b a (K, 1) vector of fixed
but unknown parameters, eit is an independently and identically distributed error term with zero mean and variance r2,
while li denotes a fixed regional effect.
The inclusion of unit specific effects (li) should help to account for omitted specific time-invariant factors which charac-
terize each spatial unit whose omission might bias coefficient estimates: in fact, the empirical model cannot control for all
possible time invariant variables characterizing the general setting of each spatial unit, since they are often hardly available
or measurable.
Model (1) can be enriched in order to account for possible interaction effects among different units. In fact an observation
associated with a specific location may depend on observations at other locations either because decision-makers in a given
area are influenced by decisions taken in other areas or because decisions in a spatial unit are significantly influenced by
explanatory variables in other nearby areas; finally, decisions in different areas may be spatially correlated because of unob-
served characteristics that have similar effects in neighbouring areas, thereby giving rise to error terms that are spatially
correlated.
When such interactions effects are included in Eq. (1), the latter becomes:
X
N X
N
yit ¼ d wij yjt þ xit b þ wij xijt c þ li þ uit ð2Þ
j¼1 j¼1

with
X
N
uit ¼ q wij uit þ eit ð3Þ
j¼1

where wij is the element of a (N  N) nonnegative matrix W of known constants (spatial weights) which describe the spatial
organization of observed units and has zero diagonal elements, since no unit can be viewed as its own neighbour.6 The

6
The stationarity conditions that need to be imposed on the N  N spatial weights matrix W in a panel data setting are set forth in Yu et al. (2008).
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 49

construction of the weighting matrix W should account for the possibility that interaction effects are stronger for closer units:
inverse distance matrix or a contiguity matrix are often adopted in applied works.
The first spatial interaction effect (i) is modelled through the inclusion of a spatially lagged dependent variable term
P
(d Nj¼1 wij yjt ) where the parameter d is called the spatial autoregressive coefficient. A model which includes just such inter-
action effect is called ‘‘spatial lag model’’. According to Anselin et al. (2006), ‘‘the spatial lag model is typically considered as
the formal specification for the equilibrium outcome of a spatial or social interaction process, in which the value of the
dependent variable for one agent is jointly determined with that of the neighbouring agents’’. The spatial interaction effect
P
(ii) is accounted for by the inclusion of spatially lagged exogenous variables ( Nj¼1 wij xijt c) where c represents a K  1 vector
of fixed but unknown parameters. Such terms control for possible correlation among the dependent variable of a spatial unit
with the level of explanatory variables in neighbouring units: for example the growth path of a region might depend on ini-
tial levels of specific variables (investments, savings, etc.) assumed by the region itself and by neighbouring regions. When
both interaction effects (i) and (ii) are included in the model, the spatial econometrics literature labels it as the ‘‘spatial
Durbin model’’ (SDM). The third spatial interaction effect (iii) is incorporated in the model by assuming that the error term
uit is spatially autocorrelated, with q representing the spatial autocorrelation coefficient. Such term can reflect the effect of
omitted spatially autocorrelated variables or the impact of unobserved shocks which follow a spatial pattern. When the
above three spatial interaction components are jointly included in a regression specification, the literature refers to it as a
Manski model (Elhorst, 2010). A Model with only the (i) effect is usually labelled as Spatial Autoregressive model (SAR),
while if only the error term includes a spatial interaction term (iii) the model is known as Spatial Error model (SEM); finally,
when both (i) and (iii) effects are considered, LeSage and Pace (2009) call it Spatial Autocorrelated Model (SAC).
Despite it is possible to estimate the Manski model, it is impossible to separately identify each interaction effect from
estimated parameters. According to LeSage and Pace (2009), ‘‘the cost of ignoring spatial dependence in the dependent
variable and/or in the independent variables is relatively high since the econometrics literature has pointed out that if
one or more relevant explanatory variable are omitted from a regression equation, the estimator of the coefficients for
the remaining variables is biased and inconsistent. In contrast, ignoring spatial dependence in the disturbances, if present,
will only cause a loss of efficiency; furthermore, the spatial Durbin model produces unbiased coefficient estimates, also if the
true data generation process is a spatial lag or a spatial error model.’’ Moreover, the authors highlight that the SDM model
nests the SEM one since the latter can be derived after imposing an appropriate non-linear restriction on parameters. As a
result, LeSage and Pace (2009) suggest that the SDM model ‘‘does not ignore spatial dependence in the disturbances, but
implies a different type of specification for error dependence from that in the true SAC’’.
Hence we decided to consider a spatial Durbin model which includes spatial lags of the dependent variable and of the
explanatory variables. The possible existence of residual spatial autocorrelation in the error term will be checked by means
of a series of LM tests (see Florax and de Graaff (2004) for details). The basic specification which we estimate is:

X
N X
N
GDPit ¼ d wij GDPjt þ TPTit b þ wij TPTijt c þ li þ t þ eit ð4Þ
j¼1 j¼1

where the dependent variable is the log of GDP of each TL-3 region and the explanatory variables are the log of ports total
throughput (TPT) and the spatial lags of GDP and TPT; eit is an independently and identically distributed error term with zero
mean and variance r2, li denotes spatial specific effects and t is a time trend which should account for cyclical effects of main
macroeconomic variables. The adopted weighting matrix W is the row normalized matrix of the inverse of the squared dis-
tances between any region i and j. The spatial lag of the dependent variable accounts for possible spillover effects deriving
from neighbouring regions under the assumption that their GDP might be jointly determined. Indeed the GDP of each region
is a useful proxy for those factors able to favour agglomeration economies, like the relative geographical concentration of
economic activity and the absolute size of the region. Moreover, it is likely that one of the main important drivers of the level
of port throughput in a particular port region is the level of economic activity in the neighbouring areas (e.g.
P 
N
TPTit ¼ f j¼1 wij GDPjt : the omission of the spatial lag of GDP might therefore lead to an important omitted variable bias

that would plague the coefficient of the port throughput variable. The inclusion of ports total throughput and of its spatial lag
in the model is motivated by the idea that transport infrastructures generates externalities that may diffuse quite far and
may affect economic performance through a variety of mechanisms, not only related to their direct contribution to the regio-
nal stock of public capital, but also associated to their influence on the spatial organization of economic activities, both at
interregional and intraregional level. The reason is the ‘network character’ of the space economy itself which implies that
infrastructural changes in a certain region might also affect the performance of connected regions. The spatially lagged
throughput variable accounts for the possible impact of neighbouring ports activity on each region gross domestic product
and allows for the possibility that spillovers effects might arise also as local as opposed to global spillovers, i.e. those stem-
ming from the general equilibrium linkages deriving from the spatial interaction of neighbouring regions (LeSage and Pace,
2009). Our empirical model includes also a set of spatial specific effects (li) which can be treated as fixed or random effects.
The random effects model assumes that li are random variables that are independently and identically distributed with zero
mean and constant variance, independent of each other and from the explanatory variables. We consider such assumption
too restrictive and we prefer to treat li as fixed and possibly correlated with our explanatory variables; fixed effects should
account for those time invariant (or slow moving) regional characteristics which affect both GDP and port throughput, like
50 A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55

those associated to geographical aspects, infrastructure endowment, socio-economic conditions (like educational achieve-
ment and human capital quality), demographical composition, urbanization degree, political and historical factors, sector
composition of the economic activity and so on. Furthermore, the estimation of a fixed effects model is important because
the presence of ports within a region is mainly associated either to geography or to historical reasons, which in turn could
also explain GDP levels: by eliminating all time invariant omitted variables, our estimation procedure should allow for this
non-randomness in the presence of ports within a region and avoid the biases and inconsistencies that would characterize
pooled ordinary least squares (OLS) or random effects estimators.7 In particular, we use the maximum likelihood spatial fixed
effects estimator recently proposed by Lee and Yu (2010) who have shown that previous Maximum Likelihood (ML) estimators
proposed in the literature might provide inconsistent estimate of the variance r2 if N is large and T is small and have suggested a
bias correction procedure which we apply to our estimates (see Lee and Yu, 2010).
It is well known (LeSage and Pace, 2009) that when the spatial lag of the dependent variable is included in the
equation, parameter estimates lose their conventional interpretation as marginal effects, because the spatial lag gives rise
to a series of feedback loops and spillover effects across regions. LeSage and Pace (2009)8 propose the computation of
three different marginal effects:(a) the ‘‘average direct effect’’, which in our case would measure the impact, on GDP of
region i, of a change in the port throughput of the same region (averaged over all regions i); such effect is akin to the
conventional interpretation of a regression coefficient, however it captures also feedback loops between regions; (b) the
‘‘average total effect’’, which can be equivalently defined (LeSage and Pace, 2009) as either the average impact on GDP of
region i of a simultaneous unit change in port throughput in all other regions, or as the total impact over all regions
GDP of a change in port throughput of a particular region, (c) the ‘‘average indirect (spillover) effect’’, which is simply
the difference between the total and the direct effect. Our empirical strategy consists in estimating different extensions
of the basic specification shown in Eq. (4) by including additional control variables, like a proxy for the innovative capacity
of each region (number of patents)9 and its spatial lag: in fact, it might be the case that knowledge produced in one region
spills over into another, thereby influencing its economic performance; furthermore, such effect might be stronger the closer
the regions are, since proximity might favour the transmission of knowledge. The density of economic activities (GDP/area)
is further added to the model in order to better control for the existence of agglomeration economies linked to the
geographical concentration of economic activity.
Furthermore we check the robustness of our results by constructing spatially lagged variables with different weighting
matrices. We first weight the inverse distance matrix with regional GDP in order to account for uneven distributions of eco-
nomic activity among regions; alternatively, we weight the inverse distance matrix with a measure of reciprocal regional
accessibility which should account for different degrees of interregional connectivity: if internal infrastructures are not com-
plemented by adequate neighbourhood conditions, bottlenecks and criticalities may arise, so that spillover effects might be
negatively influenced.

4. Empirical results

This section provides empirical estimates of the impact of port throughput on local development by estimating several
specifications of Eq. (4). In particular, on the basis of the discussion outlined in Section 3, we start with a very parsimonious
Spatial-Durbin Model which only includes the log of port throughput,10 its spatial lag, the spatial lag of GDP and a time
trend,11 the latter proxying for EU-level macroeconomic developments, productivity growth, etc. Moreover, we consider a fixed
effects specification since we believe that there can be many time-invariant omitted variables possibly driving both the level of
port activity (indeed, the very presence of a port) and regional GDP.12
Despite the SDM model does not ignore the presence of spatial correlation in the disturbances, it might be worth explor-
ing whether some residual spatial correlation does persist in the error term. To the best of our knowledge an appropriate test
for spatial error dependence in a SDM panel model does not exist; therefore, we perform, year-by-year, an LM test for spatial
error correlation in spatial lag cross sectional models considered, among others, by Florax and de Graaff (2004), on the
residuals of our SDM panel model. In all model specifications reported in Tables 2 and 3 the LM tests do not reject the null
hypothesis of no spatial error correlation in most years.13

7
For the ‘‘right censoring’’ problem arising from the presence of non-port regions we refer to Bottasso et al., 2013.
8
See also Cohen (2010).
9
There is a long tradition in the industrial organization literature that considers the number of patents as the output of a firms research efforts: although
other proxies exist (e.g. the expenditure in research and development), they are generally highly correlated with the number of patents. See, for an exhaustive
discussion on the relationship between patents and innovation, Jaffe and Trajtenberg (2002).
10
In order to be able to take the natural log even for non-port regions, we have imputed (as typical in the literature when dealing with variables with zero
values) for these observations a very low level of throughput, namely 0.1 tons. It is important to note that, because in our estimation strategy we do not use the
between variation of the data, the choice of 0.1 does not affect our results in any way. However, our main results are confirmed if we re-run our main
regressions using port throughput instead of its natural log. A similar procedure has been undertaken in the case of the very few regions with zero patents.
11
Main results are broadly confirmed when we alternatively include a full set of year dummies.
12
In some specification we also control for regressors such as economic density, the number of patents or the length of motorways. Unfortunately, we cannot
control for other potentially important variables such as human capital, foreign direct investments, and the investment rates, because such variables are not
available at the level of disaggregation we employ in this paper.
13
Results are available from the authors upon request.
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 51

Table 2
Empirical estimates.

Model 1 Model 2 Model 3 Model 4 Model 5


TPT 0.0023 0.0023 0.0023 0.0023
(0.0009) (0.0009) (0.0009) (0.0009)
TPTt1 0.0011
(0.0006)
Patent 0.0040 0.0042 0.0034
(0.0015) (0.0015) (0.0014)
Patentt1 0.0037
(0.0128)
Econ_Den 0.0006 0.0006 0.0003
(0.0002) (0.0002) (0.00015)
Trend 0.006 0.006 0.006 0.005
(0.0005) (0.0005) (0.0005) (0.0005)
GDP_TL2 (0.0029)
0.00003
W  GDP 0.652 0.651 0.644 0.641 0.478
(0.021) (0.021) (0.022) (0.021) (0.026)
W  TPT 0.0044 0.0045 0.0047 0.0046 0.0025
(0.0027) (0.0027) (0.0027) (0.0027) (0.0019)
W  Patent 0.0083 0.0086
(0.0043) (0.0028)
TPT Direct effects 0.0030 0.0030 0.0030 0.0030 0.0013
(0.0008) (0.0008) (0.0008) (0.0008) (0.0006)
TPT Indirect effects 0.0166 0.0167 0.0182 0.0178 0.0056
(0.0080) (0.0072) (0.0070) (0.0068) (0.0031)
TPT Total effects 0.0196 0.0197 0.0212 0.0208 0.0069
(0.0084) (0.0075) (0.0073) (0.0071) (0.0033)
Patent Direct effects 0.0045 0.0046 0.0047 0.0045
(0.0018) (0.0018) (0.0017) (0.0015)
Patent Indirect effects 0.0076 0.0076 0.0296 0.0194
(0.0031) (0.0029) (0.0111) (0.0055)
Patent Total effects 0.012 0.012 0.0343 0.024
(0.0048) (0.0047) (0.012) (0.062)
Wald test 4.87 5.13 5.41 13.49 18.59
(0.027) (0.024) (0.020) (0.001) (0.000)

The dependent variable is the natural log of TL3 GDP. Patents and Throughput are in natural logs. All regressions include a full set of region specific fixed
effects and a time trend. Model 5 includes a full set of country specific time trends. Standard errors robust to heterosckedasticity and arbitrary correlation
within TL3 regions (in Model 5 standard errors are robust to arbitrary correlation within TL2 regions). W is a symmetric row normalized weight matrix of
the inverse of squared distances between regions i and j. Direct indirect and total effects refer to Port throughput and Patents. Wald test is a test for the SDM
versus the SEM model: it is distributed as a v2(1) or v2(2), depending on the number of non-linear restrictions. Number of observations is 6831 (6210 in
Model 5). All regressions run using the xsmle routine for STATA.
  
, , = statistically significant at 1%, 5% and 10%, respectively.

Our choice of estimating a SDM model is supported by the rejection of the non-linear restrictions implied by the SEM
model as shown at the bottom of Tables 2 and 3; moreover, the coefficient d of the spatial lag of GDP is large and highly
significant, thus suggesting the existence of important spatial interaction effects.
As we can see from the empirical estimates reported in Table 2 as Model 1, the coefficient of port throughput is positive and
statistically significant; finally, the spatial lag of port throughput is positive, but marginally statistically insignificant at the 10%
level. The bottom half of Table 2 reports estimated direct, indirect and total effects of port throughput on regional GDP which
can be interpreted as elasticities since both variables are measured in logs. In particular, the direct impact of 0.003 (statistically
significant at 1%), suggest that an increase of 10% in the level of port throughput in a given region tends to increase GDP in that
particular region by about 0.03%, which seems a reasonable magnitude. As far as the total is concerned, our results suggest that
an increase of 10% in port throughput in a particular region tends to increase GDP in the whole sample by about 0.19%, which in
turn implies indirect spillover effects as large as about 0.16%. In the next columns of Table 2 we extend the empirical model
along different dimensions. First, in Models 2–3 we add additional control variables, namely the log of patents (Model 2)
and the density of economic activity, measured as GDP per sq-km: the first variable is a proxy for the innovation activity carried
out in the region, while the latter is a possible measure of agglomeration economies (Ciccone and Hall, 1996).14 Results
obtained from estimation of Model 1 are confirmed: spillover effects of port throughput do exist and seem to be even slightly
larger. Furthermore, our empirical model allows us to compute the direct, indirect and total elasticities of GDP with respect to
patents which are found to be positive and statistically significant: the spillover effects tend to dominate the direct impact, as they
account for almost two-thirds of the total effect, thus suggesting that social benefits of innovative activity might be larger than

14
We have also checked that our results are robust to population density, as a proxy for urbanization. Given that the economic density variable might be
endogenous, it is important to note that all empirical results discussed in this paper do not depend on its inclusion in the regressions.
52 A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55

Table 3
Additional results.

Model 6 Model 7 Model 8 Model 9 Model 10


TPT 0.0019 0.0023 0.0012 0.0024 0.0013
(0.0009) (0.0009) (0.0004) (0.0009) (0.0004)
GDPt1 – – 0.542 – 0.254
– – (0.0139) – (0.014)
Patent 0.0023 0.0035 0.0018 0.0034 0.0006
(0.0014) (0.0014) (0.0009) (0.0014) (0.0008)
Econ_Den 0.006 0.0006 0.0003 0.0006 0.00026
(0.0002) (0.0002) (0.0001) (0.00019) (0.00015)
Trend 0.001 0.0052 0.0037 0.0052 0.0027
(0.0005) (0.0005) (0.0003) (0.0005) (0.0003)
Motorways – – – 0.0054 –
– – – (0.0021) –
W  GDP 0.840 0.653 0.50 0.637
(0.025) (0.0227) (0.015) (0.022)
W  GDP_growth 0.602
(0.0162)
W  TPT 0.0002 0.0049 0.0017 0.0048 0.0045
(0.0018) (0.0027) (0.0014) (0.0027) (0.0013)
W  Patent 0.0114 0.0079 0.013 0.0080 0.0085
(0.0065) (0.0046) (0.0019) (0.0043) (0.0018)
TPT Direct effects 0.0020 0.0030 0.0015 0.0032 0.0018
(0.0008) (0.0008) (0.0005) (0.0008) (0.0004)
TPT Indirect effects 0.0128 0.0183 0.0049 0.0185 0.0133
(0.010) (0.0067) (0.0027) (0.007) (0.0033)
TPT Total effects 0.0149 0.0214 0.0064 0.0217 0.015
(0.0100) (0.0070) (0.0028) (0.0074) (0.0034)
Wald test 5.14 12.92 8.62 58.65
(0.076) (0.002) (0.013) (0.000)

The dependent variable is the natural log of TL3 GDP in Models 6–9 and the rate of growth of GDP in Model 10. Patents and Throughput are in natural logs.
All regressions include a full set of region specific fixed effects and a time trend. Standard errors robust to heterosckedasticity and arbitrary correlation
within TL3 regions. In Model 6 W is a symmetric row normalized weight matrix whose ij entry is GDPi  GDPj/dij2 where dij is the distance between regions i
and j. In Model 7 W is a symmetric row normalized weight matrix whose ij entry is ACCi  ACCj/dij2 where ACC is the accessibility index described in the data
section. In Models 8–10 W is a symmetric row normalized weight matrix of the inverse of squared distances between regions i and j. Direct indirect and
total effects refer to Port throughput and Patents. Wald test is a test for the SDM versus the SEM model: it is distributed as a v2(1) or v2(2), depending on the
number of non-linear restrictions. Number of observations is 6831 (6210 in Model 5). All regressions run using the xsmle routine for STATA.
  
, , = statistically significant at 1%, 5% and 10%, respectively.

private (i.e. region specific, in this context) ones. A further extension is introduced in Model 4 where we include the spatial lag of
patents in order to take into account the possibility that spillovers induced by the innovative activity might be local, as opposed to
global (i.e. spillovers that operate only through the spatial lag of the dependent variable). Empirical results show that, while the
direct effects are about the same, the indirect ones are much larger, thus suggesting that not only knowledge spillovers do operate
through general equilibrium linkages, but they also tend to directly affect neighbouring units. This result is consistent with
previous empirical evidence which suggests that geographical proximity favours the diffusion of knowledge flows (see among
others Crescenzi and Rodriguez-Pose, 2011; Basile et al., 2011; Ertur and Koch, 2011).
So far we have assumed that – after controlling for spatial linkages arising from GDP in the nearby regions as well as for
spatial fixed effects – the level of port activity might be considered exogenous. However, there are still reasons to believe
that this might not be warranted: for instance, an increase in local GDP in a particular set of regions in a given year might
have triggered an increase in the demand for port services in the same set of regions, potentially biasing the port throughput
coefficient. Instrumental variable techniques would allow us to account for this issue but they are still not well developed in
the panel data spatial econometrics literature; more importantly, it would be hard to find convincing external instruments.
In order to mitigate such problems, in Model 5 we have included a full set of country-specific time trends in order to
account for country specific business cycles; in fact, if GDP growth had been stronger in some countries, thereby stimulating
port activity, there might be a positive correlation between the error term and port throughput thus biasing upwards its
coefficient: the inclusion of country trends is a way to capture this possibility.15 Secondly, we have included the GDP level
of the TL-2 region each TL-3 region belongs to: the rationale of this control is that in many EU countries TL-2 regions constitute
an upper tier of government with non-negligible economic powers16 so that is possible that omitted variables at the TL-2 level
(e.g. quality of regional institutions, capacity of using EU funds, improvements in the transport network) might affect both GDP
and port throughput at the TL-3 level. It might be important to remember that what matters in this case is not the physical
distance between TL-3 regions per se, but the fact that some TL-3 regions belong to the same higher tier level of government.

15
The inclusion of a full set of country/year dummies would be even better, but this would amount to include about 160 additional variables, which would
lead to computational problems in the estimation.
16
For instance, in Italy TL-2 regions have important degrees of freedom in the implementation of industrial policies or in allocating EU funds.
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 53

Finally, as a further control for possible reverse causality between GDP and port throughput, we have lagged the latter one
period.17 Empirical results show that both direct and indirect effects are lower, as expected if some problems of endogeneity
where indeed present; however, all effects remain non-negligible and statistically significant.18
Possible problems of endogeneity have also been addressed in two other different ways. First, we have estimated Model 1
with the GMM-SYS approach of Blundell and Bond (1998)19 which allows us to use lags of port throughput as instruments for
its possibly endogenous current levels in a GMM procedure.20 Empirical results (available from the authors upon request) sug-
gest that port throughput has a positive and statistically significant effect on regional GDP suggesting that, perhaps, endogeneity
is not likely to be a major concern in this framework.
Secondly, we have also estimated a cross sectional Spatial Autoregressive Model with spatial autoregressive disturbances
where we let the average GDP, computed over the entire sample period (1998–2009), to depend on port throughput and other
regressors defined at the beginning of the sample period,21 plus a set of TL2 dummies that capture unobserved regional specific
heterogeneity at the TL2 level. By considering port throughput at the beginning of the sample period, endogeneity bias stemming
from the possibility that GDP growth in nearby regions might have fostered port throughput is likely to be a minor issue. Also in
this case (results available upon request) we find a positive and statistically significant effect of port throughput on local GDP. In
particular, we find a total effect of about 0.018, a direct effect of about 0.014 and an indirect effect of about 0.004: if we compare
these results with our previous estimates we note that, while the total effect is very similar, the contribution of the direct and
indirect effects is reversed. This difference might be due to the fact that the panel data models allow for a more effective control
of spatial heterogeneity; however, the difference might also suggest that some of the large spillover effects we found in previous
estimates might indeed reflect demand of port throughput arising from GDP growth in nearby regions: as a result, while this cross
section regression confirms the positive effects of port activity on GDP, it might be safe to consider the indirect spillover estimates
in our panel data regressions as a sort of upper bound of the ‘‘true’’ spillover effect of port throughput.
We now turn to a set of regressions that provide additional robustness checks for our main findings outlined above.
First, we verify whether our results might be explained by the choice of a particular weight matrix that, although one of
the most common in the literature, is by no means the only one. While in Models 1–5 we have adopted the row normalized
matrix of the inverse of the squared distances between any region i and j as weight matrix, in Models 6 and 7 we have esti-
mated the Spatial Durbin Model 4 after assuming different weight matrices.
In particular, in Model 6 we have considered a row normalized matrix of the inverse of the squared distances multiplied by
the levels of GDP between regions i and j (measured at the beginning of the sample period in order to alleviate possible end-
ogeneity concerns as suggested by Corrado and Fingleton (2012)), i.e. we have considered a matrix W2 whose ij entry is:
2
GDPi  GDPj =dij : such matrix allows us to better take into account the uneven distributions of economic activity among regions.
2
Alternatively, in Model 7 we have considered a matrix W3 with ij element ðACCi  ACCj Þ=dij , where ACC is the accessibility
index described in Section 2: the rationale of using this weight matrix is in this case to ‘‘refine’’ the geographical distance
between any two regions by considering how accessible these regions are. Empirical results displayed in Models 6 and 7
do not show dramatic changes in elasticity in the case of port throughput: in particular, the use of W2 leads to somewhat
lower effects, particularly as far as the total effect is concerned, which are however not precisely estimated (the p-value is
0.12); in turn, the use of W3 leads to somewhat larger indirect effects. Further robustness checks are conduced along differ-
ent lines. Model 8 includes the first lag of the log of GDP in order to allow the economy to be outside its long run equilibrium
path: empirical results show that the lagged GDP coefficient is positive and statistically significant; moreover, we note a fall
in the coefficient of the spatial lag of GDP. As far as the elasticity of port throughput is concerned, the magnitude of the direct
effect is barely altered, while the indirect elasticity substantially falls, for a total elasticity of about 0.0064, about one third of
the magnitude reported in previous columns (with the exception of Model 5), which in turn highlights the importance of
allowing for some dynamics into the model.22
An additional regressor has been included in Model 9, namely the length of the motorways network at the TL2 level.
Indeed, it has been argued (see, for instance Ducruet, 2009) that the availability of a well developed transport network
(of which motorways are a major component) might be a crucial factor for a port to display its positive effects on the region’s
economic development. However, it is important to recognize that in our sample there is, in the case of motorways, an
important problem of missing data, amounting to about 20% of the sample which we have tackled by imputing data using
a Bayesian statistical procedure automated in Stata.23 Bearing in mind the problems that might arise with such a large number

17
Regression results are broadly consistent if we lag one period the other regressors. Results available from the authors upon request
18
We also note a fall in the magnitude of the spatial lag of GDP, even if this is quite natural, given our controls for country trends and TL-2 GDP.
19
All possible endogenous variables, namely port throughput as well as the spatial lag of GDP have been instrumented with their appropriate lags. In practice,
we have used these variables and their spatial lags as internal instruments (lagged two periods).
20
The GMM approach delivers consistent estimates as long as current levels of port activity are not correlated with future shocks to GDP levels, which seems
to us a reasonable hypothesis to defend. See for a more detailed discussion Bottasso et al. (2013).
21
We have regressed average GDP over the period 1998–2009 on the log of port throughput, the log of patents, economic density, the accessibility ratio (see
below) plus the spatial lag of GDP as well as a full set of TL2 regional dummies. This regression has been run using the spreg routine in STATA.
22
The presence of the lagged dependent variable in a fixed effects model might induce endogeneity problem due to the correlation between the region fixed
effect and the lagged dependent variable. However, the bias falls with the numbers of years considered in the panel which, in our case, is quite large (T = 12 in
our sample).
23
In particular, we have used the ‘‘mi impute regress’’ command that imputes missing variables using the normal gaussian regression imputation method. In
particular, we have created 20 imputations.
54 A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55

of imputed observations, empirical results suggest that the length of motorways network is statistically significant (with a total
elasticity, not shown, of about 0.016, similar to that identified in the case of employment by Bottasso et al., 2013) and that the
port throughput and patent elasticities are barely altered.
Finally, in Model 10 the dependent variable is the rate of growth of GDP instead of its level. We have again considered a
Spatial Durbin Model as in Model 4 but augmented with the lagged level of GDP in order to allow for convergence effects.
Empirical results show that port throughput exerts a positive and statistically significant effect on the rate of growth of GDP,
with most of the impact stemming from the indirect, feedback effects.
As a final robustness check we have also estimated a Spatial Durbin Model where we have entered the main regressors
(GDP, port throughput and patents) in per capita terms in order to tackle possible concerns related to scale issues, given the
large size differentials among regions in our sample. Main results are confirmed, with a port throughput total and direct
effects of 0.019 and 0.0023, respectively.

5. Conclusion

This study analyses the impact of ports activity on regional GDP, focusing in particular on its spillover effects on neigh-
bouring regions.
It is widely recognized that the economic relevance of seaports spreads well beyond the seaport regions and while there
are several academic contributions focusing on the economic (direct and indirect) impact of ports, the studies approaching
this topic from a spatial econometric perspective are still few.
This paper contributes to fill this gap by measuring the spillover effects of port activity on GDP in 621 TL-3 European
regions by means of a spatial panel econometric approach which controls for spatial fixed effects.
Our empirical findings reveal that ports tend to increase GDP in the area where they are located (direct effect) and that
there are large and positive spillovers on the GDP of nearby regions. The results are statistically and economically significant:
estimates of the direct effect suggest that an increase of 10% in the level of port throughput in a given region tends to
increase GDP in that particular region by about 0.01–0.03%. As far as the total effect is concerned, our estimates imply that
an increase of 10% in the level of port throughput in a given region is associated to an average increase in GDP in all regions
ranging between 0.06% and 0.2%, which in turn implies indirect spillover effects in the range 0.05–0.18%. These results are
robust to the inclusion of alternative control variables and hold when we apply different empirical strategies aimed at
accounting for possible endogeneity issues associated to the fact that an economic expansion of non-port regions might gen-
erate higher demand for port services in port regions: our findings indicate that a non-negligible share of economic benefits
associated to port activity are spread across nearby areas.
Our empirical results give also some policy hints about the model of governance of seaports that should be further
investigated.
Overall results seem to be consistent with the predictions of the economic literature that underlines how ports are fun-
damental infrastructures for regional economic growth, but also with recent discussions on the de-maritimisation process
and the enlargement of port hinterlands (e.g. Notteboom and Rodrigue, 2005). More in general the existence of positive total
effects of port activity on regional GDP confirms the opportunity to further invest in infrastructure development policies. In
particular, the existence of significant spillover effects confirm the importance of internalising the positive externalities gen-
erated by port activity:24 on the one side it seems reasonable to expect that non-port regions share the costs of future ports
infrastructure investments, given that benefits tend to move further away from port regions while costs, in terms of congestion,
pollution, etc., have mainly a local nature; on the other side it might be argued that non-port regions should participate in port
governance and that the rules of representations of general interests in the managing bodies of public infrastructures might be
revised in order to account for a larger number of stakeholders, i.e. similarly to the concept of extended Port Authority presented
by Notteboom and Rodrigue (2005).

References

Afraz, N., Aquilina, M., Conti, M., Lilico, A., 2006. Impact of Transport Infrastructure on Economic Growth, Annex 6 to Final Report of COMPETE Analysis of
the Contribution of Transport Policies to the Competitiveness of the EU Economy and Comparison with the United States Funded by European
Commission – DG TREN Karlsruhe, Germany.
Anselin, L., Le Gallo, J., Jayet, H., 2006. Spatial panel econometrics. In: Matyas, L., Sevestre, P. (Eds.), The Econometrics of Panel Data, Fundamentals and
Recent Developments in Theory and Practice, third ed. Dordrecht, Kluwer.
Aschauer, D.A., 1989. Is public expenditure productive? J. Monetary Econ. 23, 177–200.
Baldwin, R., Martin, P., Ottaviano, G., 2001. Global income divergence, trade and industrialization: the geography of growth take-off. J. Econ. Growth 6, 5–37.
Basile, R., Capello, R., Caragliu, A., 2011. Technological interdependence and regional growth in Europe: proximity and synergy in knowledge spillovers. Pap.
Regional Sci. 91 (4), 697–723.
Blundell, R., Bond, S., 1998. Initial Conditions and Moment Restrictions in Dynamic Panel Data Models. J. Econometrics. 87, 115–143.
Boarnet, M.G., 1998. Spillovers and the locational effects of public infrastructure. J. Regional Sci. 38 (3), 381–400.
Borck, R., Pflüger, M., Wrede, M., 2007. A Simple Theory of Industry Location and Residence Choice IZA. Discussion Paper N. 2862.
Bottasso, A., Conti, M., Ferrari, M., Merk, O., Tei, A., 2013. The impact of port throughput on local employment: evidence from a panel of European regions.
Transp. Policy 27, 32–38.

24
Proost et al. (2011) discuss congestion externalities as well as pricing and infrastructure financing schemes in the case of the trans-European EU transport
network.
A. Bottasso et al. / Transportation Research Part A 65 (2014) 44–55 55

Chowdhury, A., Erdenebileg, S., 2006. Geography against Development: A Case for Landlocked Developing Countries. United Nations, New York.
Ciccone, A., Hall, R., 1996. Productivity and the density of economic activity. Am. Econ. Rev. 86, 54–70.
Clark, X., Dollar, D., Micco, A., 2004. Port efficiency, maritime transport costs, and bilateral trade. J. Dev. Econ. 75, 417–450.
Cohen, J.P., 2010. The broader effects of transportation infrastructure: spatial econometrics and productivity approaches. Transport. Res. Part E: Logist.
Transport. Rev. 46 (3), 317–326.
Cohen, J.P., Monaco, K., 2009. Inter-county spillovers in California’s ports and roads infrastructure: the impact on retail trade. Lett. Spat. Resour. Sci. 2, 77–
84.
Cohen, J.P., Monaco, K., 2010. Ports and highways infrastructure an analysis of intra and interstate spillovers. Int. Regional Sci. Rev. 33, 239–263.
Cohen, J.P., Morrison, P.C., 2004. Public infrastructure investment interstate spatial spillovers and manufacturing costs. Rev. Econ. Stat. 86 (2), 551–560.
Condeco-Melhorado, A., Gutierrez, J., Garcia-Palomares, J.C., 2011. Spatial impacts of road pricing: accessibility, regional spillovers and territorial cohesion.
Transport. Res. Part A 45, 185–203.
Coppens, F., Lagneaux, F., Meersman, H., Sellekaerts, N., Van de Voorde, E., Van Gastel, G., Vanelslander, T., Verhetsel, A., 2007. Economic Impact of Port
Activity: A Disaggregated Analysis The Case of Antwerp. National Bank of Belgium, WP n. 110.
Corrado, L., Fingleton, B., 2012. Where is the economics in spatial econometrics? J. Regional Sci. 52 (2), 210–239.
Crescenzi, R., Rodriguez-Pose, A., 2011. Innovation and Regional Growth in the European Union. Springer, Heidelberg.
Crescenzi, R., Rodriguez-Pose, A., 2012. Infrastructure and regional growth in the European Union. Pap. Regional Sci. 91, 487–513.
Danielis, R., Gregori, T., 2013. An input–output-based methodology to estimate the economic role of a port: the case of the port system of the Friuli Venezia
Giulia Region, Italy. Marit. Econ. Logist. 15, 222–255.
Deng, P., Lu, S., Xiao, H., 2013. Evaluation of the relevance measure between ports and regional economy using structural equation modeling. Transp. Policy
27, 123–133.
Ducruet, C., 2009. Port regions and globalization. In: Notteboom, T., Ducruet, C., De Langen, P. (Eds.), Ports in Proximity: Competition and Coordination
among Adjacent Seaports. Aldershot, Ashgate.
Ducruet, C., Itoh, H., Joly, O., 2012. Material flows and local economic structure: port-region linkages in Europe, Japan, and the United States. In: World
Conference of Transport Research – SIG2 Proceedings, 21–22 May, Antwerp.
Elhorst, P.J., 2010. Applied Spatial Econometrics: raising the bar. Spatial Econometrics Analysis 5, 9–28.
Ertur, C., Koch, W., 2011. A contribution to the theory and empirics of Schumpeterian growth with worldwide interactions. J. Econ. Growth 3 (16), 215–255.
European Spatial Planning Observation Network – EspoN, 2009. Accessibility Project Reports. <www.espon.eu>.
Ferrari, C., Parola, F., Morchio, E., 2006. Southern European ports and the spatial distribution of EDCs. Marit. Econ. Logist. 8, 60–81.
Ferrari, C., Percoco, M., Tedeschi, A., 2010. Ports and local development: evidence from Italy. Int. J. Transp. Econ. 37, 9–30.
Florax, R.J.G.M., de Graaff, T., 2004. The performance of diagnostic tests for spatial dependence in linear regression models: a meta-analysis of simulation
studies. In: Anselin, L., Florax, R.J.G.M., Rey, S.J. (Eds.), Advances in Spatial Econometrics. Springer, Berlin, pp. 29–65.
Fujita, M., Mori, T., 1996. The role of ports in the making of major cities: self-agglomeration and hub-effect. J. Dev. Econ. 49, 93–120.
Fujita, M., Thisse, J.F., 2006. Globalization and the evolution of the supply chain: who gains and who loses? Int. Econ. Rev. 47, 811–836.
Gripaios, P., Gripaios, R., 1995. The impact of a port on its local economy: the case of Plymouth. Marit. Pol. Manage. 22, 13–23.
Grobar, L., 2008. The economic status of areas surrounding major US container ports: evidence and policy issues. Growth Change 39, 497–516.
Haddad, E.A., Hewings, G.J.D., Perobelli, F.S., dos Santos, R.A.C., Raul, A.C., 2010. Regional Effects of Port Infrastructure: A Spatial CGE Application to Brazil.
Int. Regional Sci. Rev. 33, 23–263.
Holtz-Eakin, D., Schwartz, A.E., 1995. Spatial productivity spillovers from public infrastructure: evidence from state highways. Int. Tax Pub. Finan. 2, 459–
468.
Iannone, F., 2012. The private and social cost efficiency of port hinterland container distribution through a regional logistics system. Transport. Res. Part A,
1424–1448.
Jaffe, A., Trajtenberg, M., 2002. Patents, Citations and Innovations. A Window on the Knowledge Economy. MIT Press, Boston.
Kawakami, T., Doi, M., 2004. Port capital formation and economic development in Japan: a vector autoregression approach. Pap. Regional Sci. 83, 723–732.
Krugman, P., 1991. Increasing returns and economic geography. J. Polit. Econ. 99, 483–499.
Krugman, P., 1993. The hub effect: or, threeness in international trade. In: Ethier, W., Helpman, E., Neary, P. (Eds.), Theory, Policy and Dynamics in
International Trade. Cambridge University Press, Cambridge.
Lee, L., Yu, J., 2010. Estimation of spatial autoregressive panel data models with fixed effects. J. Econ. 154, 165–185.
Lee, S.W., Song, D.W., Ducruet, C., 2008. The spatial evolution in global hub port cities. Geoforum 39, 372–385.
LeSage, J., Pace, K.R., 2009. Introduction to Spatial Econometrics. Taylor & Francis, Boca Raton.
Monios, J., Wilmsmeier, G., 2012. Giving a direction to port regionalization. Transport. Res. Part A 46, 1551–1561.
Musso, E., Benacchio, M., Ferrari, C., 2000. Ports and employment in port cities. Int. J. Marit. Econ. 2, 283–311.
Niebuhr, A., 2006. Market access and regional disparities: new economic geography in Europe. Ann. Regional Sci. 40, 313–334.
Notteboom, T., Rodrigue, J.P., 2005. Port regionalization: towards a new phase in port development. Marit. Pol. Manage. 32, 297–313.
Organisation for the Economic Cooperation and Development – OECD, 2013. Regional Aggregation Documentation.
Ottaviano, G., 2008. Infrastructure and economic geography: an overview of theory and evidence. Eur. Bank Invest. Pap. 13, 8–35.
Proost, S., Dunkerley, F., De Borger, B., Guhneman, A., Koskenoja, P., Mackie, P., Van der Loo, S., 2011. When are subsidies to trans-European network project
justified? Transport. Res. Part A 45, 161–170.
Puga, D., 1999. The rise and fall of regional inequalities. Eur. Econ. Rev. 43, 303–334.
Puga, D., 2002. European regional policies in light of recent location theories. J. Econ. Geogr. 2, 373–406.
Radelet, S., Sachs, J., 1998. Shipping Costs, Manufactured Exports, and Economics Growth. Harvard Institute for International Development, Cambridge.
Romp, W., de Haan, J., 2007. Public capital and economic growth: a critical survey. Perspektiven der Wirtschaftspolitik, Verein für Socialpolitik 8, 6–52.
Sloboda, B.W., Yao, V.Y., 2008. Interstate spillovers of private capital and public spending. Ann. Regional Sci. 42, 505–518.
Tabuchi, T., Thisse, J.F., 2002. Regional Specialization and Transport Costs. CEPR Discussion Papers N. 3542.
United Nations Conference on Trade and Development – UNCTAD, 2009. Maritime transport and the climate change challenge. In: Proceedings of the Multi
Expert Meeting, February 2009, Geneva.
Wilmsmeier, G., Hoffmann, J., Sanchez, R., 2006. The impact of port characteristics on international maritime transport costs. Res. Transport. Econ. 16, 117–
140.
Yu, J., De Jong, R., Lee, L., 2008. Quasi-maximum likelihood estimators for spatial dynamic panel data with fixed effects when both n and T are large. J. Econ.
146, 118–134.

You might also like