Economic Modelling 128 (2023) 106487

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

A nonparametric panel data model for examining the contribution of tourism

to economic growth✩
Ergun Dogan a , Xibin Zhang b ,∗
a Department of Economics, Cankaya University, Turkey
b
Department of Econometrics and Business Statistics, Monash University, Australia

ARTICLE INFO ABSTRACT

Keywords: We apply a nonparametric panel data model with cross-sectional and time-varying coefficients to examine the
Eurozone crisis relationship between tourist arrivals and economic growth in the Schengen area from 1995 to 2019. In contrast
Global Financial Crisis to the parametric models employed in other studies, our nonparametric model makes no assumption about
Nonparametric estimation
functional form and, hence, allows us to model the relationship nonlinearly. We find that the tourism–economic
Time-varying panel data model
growth relationship in the Schengen area is nonlinear and time-varying. While the relationship between tourism
Total factor productivity
Tourism-led economic growth hypothesis
and economic growth was positive and significant during 1995–2003, it was negative and significant during
the Global Financial Crisis (2007–2008) and the European recession of 2012–2013. One additional contribution
of the study is the finding that total factor productivity (TFP) has been growing at 1.45% per year. The results
also show that country-level TFP growth was disrupted during the aforementioned negative economic shocks.

1. Introduction et al. (2016a), Castro-Nuño et al. (2013), Ahmad et al. (2020), and Liu
et al. (2022). With few exceptions, the literature suggests wide support
Tourism is a major economic sector worldwide. According to the for the TLGH. For instance, in a review of 95 peer-reviewed studies
World Travel and Tourism Council, the sector, with all effects (direct, by Brida et al. (2016a), there was support for the TLGH in all but a
indirect, and induced) counted, was responsible for 10.4% of global handful of countries.
GDP and 10.3% of global employment in 2019.1 These figures show These studies, however, employ parametric models with restrictive
some variation across regions. The contribution of tourism to GDP functional form assumptions. A limitation of parametric models is that
(employment) ranges from 7% (5.6%) in Africa to 13.7% (15.6%) in the estimates may be biased and inconsistent, due to misspecification
the Caribbean; in Europe, the number is 9.3% (9.9%). if there is no exact prior knowledge of the functional form (Li et al.,
The tourism-led economic growth hypothesis (TLGH) states that 2011). In particular, the relationship may not be linear. Nonlinearities
tourism contributes to economic growth through multiple channels. can be introduced into the relationship by economic downturns in
In particular, tourism is a source of foreign exchange earnings, in- destination or source countries or by exogenous shocks, such as health
frastructure investment, employment, and positive economies of scale, scares, natural disasters, or terrorism in destination countries.
thus reducing production costs for local businesses (see, for example, Several studies find that there are structural breaks in tourism and
Antonakakis et al., 2015a). GDP (Lee and Chien, 2008; Lean and Smyth, 2009; Narayan, 2005;
Several studies have tested the TLGH. Examples published in the Smyth et al., 2009, among others), which are a source of complexity
leading tourism journals in the last 5 years include Antonakakis et al. and nonlinearities. As a consequence, parametric models may not
(2016, 2019), Chiu and Yeh (2017), De Vita and Kyaw (2016, 2017), uncover the underlying relationship between economic growth and
Wu and Wu (2017), Liu and Song (2018), Dogru and Bulut (2018),
tourism, nor the manner in which the relationship has evolved over
Lin et al. (2019), Eyuboglu and Eyuboglu (2020) and Zuo and Huang
time. The idea that nonparametric methods might be useful in analyz-
(2018). Reviews of much of the literature, dating back to the seminal
ing the relationship between economic growth and tourism has also
study by Balaguer and Cantavella-Jorda (2002), are contained in Brida
been recognized by Brida et al. (2020).

✩ The authors extend their sincere thanks to the Co-Editor, Professor Angus Chu, and anonymous reviewers for their constructive comments, which helped
improve the manuscript. Thanks also go to Russell Smyth for an enlightening discussion during an early stage of the paper and Song Li for excellent research
assistance.
∗ Corresponding author.
E-mail address: [email protected] (X. Zhang).
1
These figures were obtained from the United Nations World Tourism Organization (UNWTO).

https://doi.org/10.1016/j.econmod.2023.106487
Received 14 December 2022; Received in revised form 23 July 2023; Accepted 18 August 2023
Available online 22 August 2023
0264-9993/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Some studies employing parametric models to test the TLGH have important contribution. This methodology has been successfully used
recognized the possibility of instability and nonlinearities in the re- in several papers in other fields. Second, our study differs from others
lationship between economic growth and tourism. One approach has in its conceptual framework. Most studies do not provide a theoretical
been to include a squared term for tourism to capture nonlinearities in discussion of the tourism–growth relationship or provide only a brief
tourism (Adamou and Clerides, 2009; De Vita and Kyaw, 2016, 2017). overview. Our discussion starts by drawing attention to the substantial
Other studies have employed nonlinear cointegration and threshold multiplier effects of tourism reported in other studies, some of which
models (Po and Huang, 2008; Chang et al., 2012; Deng et al., 2014; are probably due to productivity gains. Although there is no direct
Wang, 2012; Phiri, 2015; Brida et al., 2015; Brida et al., 2016b; Chiu evidence showing that tourism can yield productivity gains, we make
and Yeh, 2017; Zuo and Huang, 2018). Some studies have applied time- a theoretical case that it does by drawing on the literature on the
varying models (Antonakakis et al., 2015a,b; Arslanturk et al., 2011; knowledge spillovers generated by exporters, FDI, and agglomeration.
Balcilar et al., 2014; Enilov and Wang, 2022) or time-varying copula Third, our study also differs from others in its explanation of why
functions (Pérez-Rodríguez et al., 2015). Yet other studies have em- the tourism–growth relationship is expected to be time-varying, which
ployed smooth transition regression (Pan et al., 2014; Wu et al., 2016) is largely missing in the literature. Fourth, our methodology lets us
or quantile regression (Shahzad et al., 2017). All these approaches, obtain findings on important variables, such as trend GDP, which is
however, make assumptions about the functional form, and none have an important indicator of an economy’s long-run growth potential.
the flexibility to fully capture the complexities of the relationship The remainder of the paper is set out as follows. The next section
between economic growth and tourism. provides a conceptual framework for the study. Section 3 outlines our
We address this gap in the literature by applying a nonparametric data sources and goes through some important economic developments
panel data model with cross-sectional and time-varying coefficients in Europe during the study period. We present our baseline parametric
to examine the relationship between tourism and economic growth results, which serve as a benchmark for the nonparametric results, in
for 17 Schengen countries in Europe over the period 1995 to 2019. Section 4. The nonparametric results are presented in Section 5. We
We also contribute to the literature on the TLGH by estimating the present the implications of our findings in Section 6, followed by the
common trend and country-specific trends, which has not been done Conclusions in Section 7.
before. Our novel nonparametric cross-sectional and time-series vary-
ing coefficient model employs a local linear dummy variable estimation 2. Conceptual framework
(LLDVE) method to estimate the trend and coefficient functions in a
highly nonlinear fashion. The LLDVE method we employ builds on the 2.1. Theoretical discussion
ideas initially presented in Li et al. (2011) and Zhang et al. (2012)
and recently developed by Silvapulle et al. (2017), Hailemariam et al. As mentioned in the Introduction, tourism’s overall effect on the
(2019), and Awaworyi Churchill et al. (2019) in the energy economics economy is large. These multiplier effects, which arise when tourism
literature. creates or increases demand for other sectors, are documented exten-
As pointed out by Fan and Zhang (2008), ‘‘the varying coeffi- sively in the tourism literature. One study on several EU countries
cient models are not stimulated by the desire of purely mathematical finds that gross value added increases by an amount that ranges be-
extension, rather they come from the need in practice’’. Many relation- tween 0.58 Euro to 1.26 Euro when tourist spending increases by 1
ships in economics are dynamic and evolving, for which time-varying Euro (Figini and Patuelli, 2022). These multipliers are calculated using
coefficient models instead of linear models would be more suitable. the Tourism Satellite Accounts; hence, they only account for direct and
Time-varying coefficient model is especially suitable for modeling eco- indirect effects. It is unclear to what extent these multiplier effects
nomic growth. Economies grow when capital and labor grow, but the reflect productivity increases, but it is plausible to believe they are
contribution of capital and labor need not remain constant over time partially due to productivity increases. In this section, we discuss how
for reasons such as changes in technology and labor efficiency. and why tourism can affect productivity.
This nonparametric modeling approach, however, has never been Our conceptual framework is based on the standard Cobb–Douglas
employed in the tourism literature. We use a wild bootstrap method production function model that is often used in the growth literature.
to generate the confidence intervals for the underlying trend and coef- We start with two inputs, capital and labor, and bring tourism into the
ficient functions. In so doing, we relax the restrictive functional form model as a productivity-enhancing variable:
assumptions in the existing literature testing the TLGH. Our approach
𝑌 = 𝐴𝐾 𝛼 𝐿1−𝛼 ,
has the advantage that it allows the common trend functions to evolve
to capture common global shocks due, for instance, to economic reces- where 𝐴 = 𝐸𝑇 𝜖 . In this representation, 𝐸 is the autonomous component
sion or terrorism, while the cross-sectional and time-varying coefficient of total factor productivity (TFP)—e.g., technological progress—𝑇 is
functions capture nonlinearities and heterogeneity across countries and the number of tourist arrivals, and 𝜖 is the tourism elasticity of TFP (𝐴).
time. This equation makes it clear that in a given year, two countries with
Our baseline parametric point estimates suggest that the relation- equivalent resources can have different levels of real GDP, with the one
ship between tourist arrivals and economic growth in the Schengen area possessing the larger tourism sector achieving the highest level.
is positive and significant. Parametric point estimates average out over Our treatment of tourism, which is a form of service export, as one
time, however, and do not capture relational nonlinearities or switching of the determinants of TFP, is consistent with the findings of many
between periods of positive and negative associations between tourist studies in the literature on the export–growth relationship. Exporters
arrivals and economic growth. usually have higher productivity than nonexporters, which increases
Our estimated nonparametric coefficient function for tourism sug- overall TFP and contributes to economic growth. The higher productiv-
gests that the relationship between tourist arrivals and economic ity of exporters is partly due to the economies arising from operating at
growth is highly nonlinear, being significantly positive between 1995 a larger scale. The productivity of exporters is also higher because of the
and 2003 but significantly negative during the 2007–2008 and 2013– knowledge spillovers in the form of learning-by-exporting, for which
2014 periods. the literature provides strong support (Keller and Yeaple, 2009, p.9).
In summary, our study differs from similar studies in several ways. We could expect to see these spillovers arising in the tourism sector, as
First, the methodology is different from the ones used in most of the well. Businesses operating in the tourism sector have to compete with
literature, where several variants of the Granger causality methodology their international peers intensely, whose efficiency levels are likely to
have been used. The use of a novel nonparametric varying coeffi- be very high, leading to an improvement in the quality of the services
cient model for the first time in the tourism literature is our most they provide (Jin, 2011).

2
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Tourism sector also attracts abundant foreign direct investment that limited to those areas but could spill over to other industries no matter
is likely to be more efficient than the domestic service providers. Hence, their locations. Alternatively, if, as in Faber and Gaubert (2019), it
tourism sector productivity is likely to be higher than the productivity is assumed that spillovers operate at the local level, then aggregate
in other service sectors, if the share of FDI of the tourism sector in productivity would increase whenever positive changes in productivity
overall FDI is relatively high. FDI might indirectly affect the productiv- in the areas where tourism activities cluster outweigh the possible
ity of the sector by generating knowledge spillovers. Spillovers occur negative changes in productivity in other areas.5
as workers move from foreign to domestic firms, through demonstra- We see no reason why some spillovers, indeed most, should not
tion effects and horizontal linkages (between the firms that are in occur at the aggregate level. Hence, our approach differs from Faber
the same industry), and when domestic firms strive to compete with and Gaubert (2019) in assuming that productivity-boosting effects spill
foreign firms (see Blomström and Kokko, 1998, for further details on over to other industries no matter their location.
the spillover channels). Evidence from other sectors show that these We can use the above equation to derive an equation that can be
spillovers could be substantial. Keller and Yeaple (2009) find that the used in econometric estimations as follows:
contribution of spillovers from the manufacturing FDI to the productiv-
ln 𝑌 = ln 𝐸 + 𝜀 ln 𝑇 + 𝛼 ln 𝐾 + (1 − 𝛼) ln 𝐿.
ity growth in the US manufacturing industries during 1987–1996 was
between 8% and 19%. After adding control variables, an equation that can be used for final
Consequently, the productivity in the tourism sector would probably estimations is obtained:
be higher than that of the other (service) sectors, either because most
of the sector’s ‘‘output’’ is exported or because of the heavy presence of ln 𝑌 = 𝛽0 + 𝛽1 ln 𝐾 + 𝛽2 ln 𝐿 + 𝛽3 ln 𝑇 + Control Variables + 𝑢.
foreign firms. If this is the case, output in the sector will increase more We use interest rate, consumer price index (CPI), exchange rate, and
than the decreased output in other service sectors as more resources openness (trade as a percentage of GDP) as control variables. There
are shifted to the tourism sector, increasing aggregate output. If the is a large literature showing that trade and economic growth are
tourism sector is not more productive than the other service sectors, related (see Singh, 2010, for a survey of this literature).6 Interest
however, a rise in tourism output would be offset by reduced output in rate, CPI, and exchange rate are included to control for the effects
other sectors, creating ‘‘Dutch’’ disease problems.2 of monetary policy on economic growth. These also control for the
Tourism activities might improve productivity in other sectors macroeconomic instability, which at times heavily affected economic
through the input channel, which contributes to aggregate TFP. Many growth during the period under study.
firms use at least some services as inputs. This provides an avenue
for the services to increase the overall productivity level of the econ-
2.2. Time variation
omy. Francois and Hoekman (2010) find evidence on this, and conclude
that producer services are major contributors to the overall productivity
Time-varying methodology we use in this study implies that coun-
growth in OECD and developing countries. Productivity improvements
tries would have a higher real GDP in 𝑡 + 1 even if they used the same
by service firms can improve the productivity of firms using their
resources as they did in 𝑡, if the contribution of tourism (𝜀) were higher.
services and thus increase the competitiveness of those firms, leading
Time variation can arise in several ways. One reason is that changes
to increased exports. Spillovers that arise as a result of any increase in
in productivity in the tourism sector due to sector-specific technological
exports this way would also be fed into the productivity.3 If tourism
shocks could increase 𝜀. Another reason is that the share of activity
could increase the efficiency of some of these inputs, the output or
generated by businesses owned and operated by foreigners change
value added of the industries using these inputs would increase. For
(FDI) in tandem with demand shocks to the tourism sector, increasing
instance, carrying a lot of leisure travelers, who are mostly tourists,
or decreasing spillover effects from these investments.
might also make airliners more efficient in the business travel segment,
Nonlinearity in the tourism–GDP relationship, as found in some
which in turn decreases the costs of businesses in all the other sectors.
studies, can cause time variation. The basic idea is that if the relation-
Tourism could also boost aggregate TFP because the operation of
ship is nonlinear at any given point of time, when specific threshold of
highly productive domestic and foreign service providers might gen-
a certain variable, such as tourism specialization, is reached, the effect
erate knowledge spillovers to other sectors in the economy through
of tourism on economic growth will change, generating a time-varying
vertical linkages (the backward and forward linkages with the firms
relationship between the two variables. Similarly, structural breaks
in related industries).4 The magnitude of these spillovers from tourism
in the tourism–GDP relationship that some studies find might cause
to other industries would depend on to what extent technological
time variation, as well. (See the Introduction for a list of studies that
knowledge, management, labor, and organizational practices of the
find nonlinearity and structural breaks). For instance, the relationship
tourism sector could be transferred to other sectors.
may change after severe recessions, such as the Global Financial Crisis
Faber and Gaubert (2019) mention several other channels through
(GFC).
which spillover effects could arise. A higher level of tourism activity
Suppose that elasticities do not change over time. Even in this
at a certain location could stimulate local credit and financial services
case, output (GDP) shocks that differ in magnitude across countries
growth, as well as other services such as accounting and consultancy,
would generate time variation in the tourism–real GDP relationship.
needed by businesses in that location. In addition, because of a larger
As mentioned above, a large tourism sector benefits some industries,
local tourism sector, the training level of local workers could increase,
whose outputs or services are used as inputs elsewhere in the economy
business opportunities could increase and lead to new business ven-
tures, and more people traveling across locations could help foster new
networks (Faber and Gaubert, 2019, p.2274). These spillovers can have 5
The argument in Faber and Gaubert (2019) for the existence of a long-run
an impact on aggregate productivity if the productivity-boosting effects effect of tourism on productivity and growth is based on the concept of ag-
of tourism activities clustered in certain areas may not necessarily be glomeration effects. Agglomeration of tourism activities increases productivity
in locations where this happens due to spillover effects on manufacturing firms
in the same location. However, this attracts firms from other locations to the
2 places where tourism activities agglomerate, leading to reduced productivity
See Copeland (1991), Chao et al. (2006), and Inchausti-Sintes (2015).
3
Hoekman and Shepherd (2017) find some evidence of both the direct and in those areas and suggesting an ambiguous effect overall (economy-wide).
6
the indirect effect (through exports) of the services sector on productivity. More recent evidence can be found in Chang and Mendy (2012), Dufrenot
4
For evidence of vertical spillovers in manufacturing industries through the et al. (2010), Hye and Lau (2015), Kim (2011), Sarkar (2008), and Shahbaz
backward and forward linkages, see Javorcik (2004). (2012).

3
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

by making them more efficient. This could only benefit the industries estimated using a set of countries with different levels of development
where these inputs are used if they can maintain the same levels may not be applicable to the developed ones.9
of revenues, in which case their value added would increase since Another reason for the selection of these countries is that they
their input costs would be lower. To explain how this can cause have economies with a services sector larger than their manufacturing
time variation in the tourism coefficient, assume an unexpected and sector, which allows us to safely assume that the additional labor
negative output shock hits the economy, as in a recession. This would the tourism sector might need would be diverted from other services
reduce the revenues of many firms and eventually affect their value sectors. In economies with a large manufacturing sector but a small
added. If these shocks had higher magnitudes in countries with larger services sector, comprising mostly developing economies, additional
tourism sectors, their real GDPs would be much lower; consequently, resources would likely come from the more productive manufacturing
the tourism coefficient would be negative.7 sector.10 This could lead to tourism becoming a ‘‘Dutch’’ disease.
Within the empirical framework we use, another possibility is that We use annual data for 17 countries that are part of the Schengen
the time variation in the tourism coefficient may be a result of time area as of 2019.11 In addition to real GDP (constant 2015 US dollars)
variation in the composition of demand. For example, certain segments, and tourism arrivals, we collected data on real capital stock (gross fixed
such as coastal and maritime tourism, might have different spillover capital stock formation in constant 2015 US dollars), labor force, the
effects on productivity than, say, mountain tourism. In certain years, CPI, the nominal exchange rate, population, and trade as a percentage
the share of high-impact tourists might be overwhelmingly greater than of GDP. All data, except nominal exchange rates, come from the World
the share of low-impact ones, which would be reflected in 𝜀.8 Some evi- Bank’s World Development Indicators database. Exchange rates were
dence indicates that the composition of tourism demand might matter. obtained from the OECD website because pre-euro exchange rates,
For instance, Tang and Tan (2015) show, using a recursive Granger which were converted to euro/dollar exchange rates using a conversion
causality test, that only arrivals from 8 of 12 markets contribute to factor, are available in the OECD dataset.
growth because the number of arrivals from the other 4 markets may
be insufficiently high or comprise mostly illegal workers rather than 3.2. Economic background
genuine tourists.
The study period includes several years when economic activity
slows down or contracts considerably. We trace out the effects of these
3. Data and economic background booms and busts that occurred during the study period by analyzing
the annual changes in the mean growth rates of real GDP and in the
3.1. Data number of tourist arrivals obtained from the country-level growth rates
in Fig. 1. The first such downturn occurred during the global economic
Named after the Schengen Agreement, the Schengen area consists slowdown of 2001, when the mean growth rate of real GDP dropped
of 26 European countries (22 of the European Union states and four to 2.58% from 4.2% in 2000. As a result of the Global Slowdown, the
other European countries), which have abolished their internal borders mean growth rate of tourist arrivals also decreased to 0.83% in 2001.
with other member countries. For tourists who need a visa to enter The mean real GDP growth rate continued to drop until 2003, then
the European Union, the creation of the Schengen area has greatly started to increase until 2008, when the GFC hit. The most severe
simplified travel within Europe because it allows them to travel on a economic contraction occurred in the year following the GFC, with
single visa or with a visa exemption (UNWTO, 2018). the growth rate turning negative at 4.72% in 2009. The mean growth
We focus on the Schengen countries for several reasons. First, rate of tourist arrivals bounced back to positive territory in 2002 and
Europe generally, and the countries included in our sample particularly, continued to increase until it peaked in 2004. As with the real GDP
are important destinations for tourism. Of the UNWTO regions, Europe growth rate, the growth rate of tourist arrivals turned negative in 2009,
accounted for the largest share of tourist arrivals, with 619 million with the arrivals decreasing at a rate of 3.29% from the previous year.
tourists, or 50% of the world’s total, in 2016 (UNWTO, 2018). It is Although the economic growth started to accelerate in 2010, it
estimated that tourism directly generates 5% of the European Union’s slowed down again in 2012 because of the Euro Debt Crisis, drop-
GDP and that when indirect links are accounted for, tourism contributes ping to almost zero in that year. Mean real GDP growth started to
over 10% of the European Union’s GDP (UNWTO, 2018). For this accelerate in the following year and remained above 2% from 2015
reason, the European Union has placed much emphasis on tourism as onward. Tourist arrivals increased by 13.13% on average in 2011. One
an engine of economic growth, consistent with the TLGH (Antonakakis reason for this is the recovery of the US economy following the GFC
et al., 2015a). around this time. The United States is Europe’s single biggest source of
tourists (UNWTO, 2018). Another reason for the recovery in inbound
The countries included in the sample are also similar in many other
tourism was the sharp increase in Chinese outbound tourists since 2009
aspects, such as education levels and labor force participation rates.
In short, they have similar levels of economic development. This may
be important in obtaining more reliable estimates as the coefficients 9
Advantages of homogeneous samples have been noted by other authors,
as well. See Sianesi and Reenen (2003), Madsen et al. (2018), and Yao et al.
(2020).
7 10
To illustrate, imagine a scatter plot that shows real GDP and tourist For the countries that are in our sample, gross value added of services
arrivals of 17 countries in logarithms in period 𝑡. Suppose a regression model as a percentage of total gross value added in 2020 is as follows: Austria
fitted an upward-sloping line to these data, indicating a positive relationship (70), Belgium (77.4), Finland (69.4), Germany (69.8), Hungary (67.2), Iceland
between tourism and real GDP (countries with higher tourist arrivals have (72.6), Italy (73.9), Latvia (72.6), Luxembourg (87.3), the Netherlands (78.2),
higher real GDP). However, percentage changes in real GDP (output) might Norway (67.5), Poland (64.9), Portugal (75.3), Slovenia (64.6), Spain (74.8),
differ across countries from one year to the next. If a sufficiently large number Sweden (74.2), and Switzerland (72.6). These data come from Eurostat,
of countries that received large numbers of tourists in year 𝑡 experienced accessed on May 22, 2023.
11
larger decreases in their GDP than the countries that received smaller tourist The countries in our sample are (with the year of first implementa-
arrivals in year 𝑡, the regression line would tilt down in 𝑡 + 1, that is, the slope tion given in brackets) Austria (1997), Belgium (1995), Finland (2001),
(elasticity) would become negative. Germany (1995), Hungary (2007), Iceland (2001), Italy (1997), Latvia
8
Time variation caused by changes in the composition of demand would (2007), Luxembourg (1995), the Netherlands (1995), Norway (2001), Poland
disappear if the shares of various groups could be controlled for in the (2007), Portugal (1995), Slovenia (2007), Spain (1995), Sweden (2001), and
estimations using disaggregated arrivals data, which we do not have. Switzerland (2008).

4
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Table 1
Pesaran’s 2007 CIPS test for unit roots.
Number of lags 0 1 2 3 4
Specification without trend
ln(GDP) 2.188 −0.379 −0.170 −1.314 −0.951
𝛥 ln(GDP) −6.924∗∗∗ −4.414∗∗∗ −2.969∗∗∗ −1.977∗∗∗ −0.671
ln(Arrival) −2.192∗∗ −2.615∗∗ −2.001∗∗ −1.937∗∗ −0.699
𝛥 ln(Arrival) −8.587∗∗∗ −6.052∗∗∗ −2.004∗∗ −2.497∗∗∗ 1.084
ln(CapitalStock) 2.433 1.337 0.827 1.893 −0.149
𝛥 ln(CapitalStock) −9.091∗∗∗ −5.777∗∗∗ −4.575∗∗ −0.627 0.926
ln(LabForce) −1.310∗ −1.664∗∗∗ 0.028 0.179 −0.852
𝛥 ln(LabForce) −9.881∗∗∗ −5.623∗∗∗ −3.811∗∗∗ −0.584 1.077
ln(CPI) −2.519∗∗∗ −2.104∗∗ −0.641 −1.518∗ 1.390
𝛥 ln(CPI) −7.001∗∗∗ −2.922∗∗ −2.543∗∗∗ −2.500∗∗∗ −1.985
ln(ExchangeRate) 0.154 −0.094 0.911 1.877 −0.367
𝛥 ln(ExchangeRate) −12.404∗∗∗ −7.75∗∗∗ −4.022∗∗∗ −15.863∗∗∗ −1.580∗
ln(Open) 0.758 0.236 0.408 −0.587 0.110
𝛥 ln(Open) −8.947∗∗∗ −4.267∗∗∗ −0.841 −0.491 1.126
Specification with trend
ln(GDP) 1.359 −1.631∗ 0.451 2.005 2.797
𝛥 ln(GDP) −5.190∗∗∗ −2.787∗∗∗ −1.609∗ −0.684 0.968
Fig. 1. Percentage changes in mean growth rates of real GDP and tourist arrivals, ln(Arrival) −0.044 −0.629 1.425 1.661 2.090
where mean growth rates are calculated from country-level growth rates. 𝛥 ln(Arrival) −7.356∗∗∗ −4.109∗∗∗ 0.381 −0.214 1.740
ln(CapitalStock) −0.021 −0.791 0.944 3.021 2.647
𝛥 ln(CapitalStock) −7.218∗∗∗ −3.890∗∗∗ −3.026∗∗ 0.110 2.868
ln(LabForce) −0.528 −2.246∗∗ 0.734 2.796 2.315
(UNWTO, 2018). However, because of Europe’s economic slowdown 𝛥 ln(LabForce) −8.964∗∗∗ −3.714∗∗∗ −0.922 2.162 4.409
the following year, the mean growth rate dropped to 4.27% in 2012 ln(CPI) −1.590 −1.111 −0.646 −2.588 0.917
and 2.64% in 2013. The mean growth rate of tourist arrivals accelerated 𝛥 ln(CPI) −4.124∗∗∗ −0.259 −1.026 −0.650 −0.009
again in 2014, continuing until 2017, when it peaked at 9.12%. In the ln(ExchangeRate) −2.240∗∗ −2.782∗∗ −1.393∗ −1.566∗ −1.564∗
𝛥 ln(ExchangeRate) −10.238∗∗∗ −5.666∗∗∗ −1.353∗ −3.654∗∗∗ 0.254
final 2 years of the sample period, the mean growth rates were 3.4%
ln(Open) 2.519 1.984 2.032 1.756 2.253
and 2.81%, respectively. 𝛥 ln(Open) −8.336∗∗∗ −2.984∗∗∗ 0.369 1.146 3.319
This discussion and Fig. 1 clearly show that the mean growth rate
Note: Null hypothesis is that the underlying series is I(1). CIPS test assumes cross-
of tourist arrivals was more volatile than the mean growth rate of real sectional dependence is in the form of a single unobserved common factor. 𝛥 denotes
GDP in this period. Fig. 1 also shows that the mean growth rate of first difference.
tourist arrivals was higher than the real GDP growth rate most of the
time—16 of 25 years, to be exact.
that affect both the dependent and the independent variables included
4. Baseline parametric estimates in the estimating equations. The average of the estimated coefficients
(𝛽𝑖 ) from country regressions indicates the effect of the independent
4.1. Empirical model variables on the dependent variable. The CCEP estimation is performed
by pooling all data across countries; hence, the estimated slope coef-
To examine the relationship between real GDP (denoted by GDP) ficients are the same for each country. Country-specific fixed effects,
and tourism, we use the following common correlated effects (CCE) and cross-sectional averages of each variable, including the indepen-
model: dent variable, interacted with country-fixed effects are added to the
estimating equations.12
ln GDP𝑖𝑡 = 𝛼𝑖 + 𝛽𝑖⊤ 𝑋𝑖𝑡 + 𝑢𝑖𝑡 , (1)
𝑢𝑖𝑡 = 𝜆⊤
𝑖𝑡 𝑓𝑡 + 𝜀𝑖𝑡 , (2) 4.2. Empirical findings from parametric estimations
for 𝑖 = 1, 2, … , 𝑛, and 𝑡 = 1, 2, … , 𝑇 , where 𝑖 denotes individual coun-
We start with cross-sectional dependence and panel unit root tests.
tries, and 𝑡 denotes the 𝑡th time period. Eq. (2) includes time-varying
We use the LM test proposed by Breusch and Pagan (1980) to test for
common factors (𝑓𝑡 ), which affect the dependent variable through
cross-sectional dependence since the number of cross-sectional units
country-specific factor loadings (𝜆𝑖𝑡 ). 𝜀𝑖𝑡 is a country-specific indepen-
(17) is smaller than the time periods (25).13 LM test statistics are all
dent and identically distributed error term. 𝑋 includes the following
significant at the 1% level, indicating that there is strong cross-sectional
variables in natural logarithms: real capital stock, labor force, tourism
dependence in all the data series used in this study.
as measured by number of tourist arrivals, CPI, nominal exchange rate,
Given the evidence of cross-sectional dependence in all the data
and trade openness (trade as a percentage of GDP).
series, we proceed with the Pesaran’s (2007) CIPS test for a unit root,
One reason why we adopt a CCE estimation methodology in which
which is robust in the presence of cross-sectional dependence. Results
cross-sectional dependence is handled by including common factors in
of the CIPS test with and without a trend are given in Table 1 and
estimations is that tourist arrivals and associated effects on GDP in
indicate that all series can be taken as 𝐼(1).
any given country may affect other countries (particularly neighboring
The point estimates for the parametric model are given in Table 2.
countries in Europe, as tourists may travel to more than one country),
The tourism coefficient is positive and significant in both models. The
causing cross-sectional dependence among the countries. Another rea-
son is that the economic crises in 1998, 2008, and 2012 are common
factors. It is possible that these crises affected all countries in the 12
We use the xtcce command in Stata to obtain the results reported in
sample through business cycle synchronization (Song et al., 2018). Table 2.
We use the CCE mean group (CCEMG) and CCE pooled (CCEP) 13
According to Pesaran’s (2015) CD test, the rejection of the null hypothesis,
estimators developed in Pesaran (2015) to estimate the model. The which is ‘‘errors are weakly cross-sectional dependent’’, implies strong cross-
CCEMG estimation is done by running a separate regression for each sectional dependence within the data series. The null hypothesis is rejected by
country with cross-sectional averages, which account for the factors the CD test for all series.

5
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Table 2 Li et al. (2011) proposed to estimate time-varying trends and coef-

Estimates of parameters for the parametric models.
ficients using the LLDVE method, which is based on the following as-
Variables CCEP CCEMG sumptions: (i) The error term 𝑢𝑖𝑡 satisfies certain martingale difference
ln(GDP) ln(GDP)
conditions along the time dimension; (ii) 𝑒𝑖𝑡 may be cross-sectionally
ln(Arrival) 0.0218∗∗ 0.0376∗∗
dependent for each 𝑖 and independent of 𝑋𝑖𝑡 ; and (iii) 𝑋𝑖𝑡 and 𝛼𝑖 can be
(0.0087) (0.0181)
ln(CapitalStock) 0.154∗∗∗ 0.157∗∗∗
correlated (see Silvapulle et al. (2017) and Hailemariam et al. (2019)
(0.0124) (0.0165) for further explanation on the estimation and bandwidth selection).
ln(LabForce) 0.332∗∗∗ 0.138 Let 𝑓̂(𝜏𝑡 ), 𝛽̂𝑡 and 𝛼
̂𝑖 denote the corresponding estimates obtained
(0.0672) (0.0953) through LLDVE. Country-specific individual trends can be estimated in
ln(CPI) −0.358∗∗∗ −0.231∗∗∗
a similar way to what Zhang et al. (2012) do. We use 𝑚 ̂ 𝑖 (𝜏𝑡 ) to denote
(0.0548) (0.0841)
ln(ExchangeRate) 0.164∗∗∗ 0.149∗∗ the estimates of these individual trends, and ̂ 𝑢𝑖𝑡 to denote residuals.
(0.0285) (0.0595)
ln(Open) −0.0394 −0.0527 5.2. Bootstrapping confidence intervals
(0.0282) (0.0436)
Constant 7.708 5.292
A wild bootstrapping method is used to construct confidence in-
(6.462) (3.463)
Observations 425 425 tervals for the time-varying common trend and coefficient functions.
Number of countries 17 17 Details of the bootstrapping procedure are as follows (see, for example,
Note: Standard errors are in parentheses. Triple stars and a single star refer to Wu, 1986; Mammen, 1993):
significance at 1% and 10% levels, respectively.
Step 1: Compute de-trended residuals 𝑣̂𝑖𝑡 = ̂ ̂ (𝜏 ), where ̂
𝑢𝑖𝑡 − 𝑚 𝑢 =
( 𝑖 𝑡 ) 𝑖𝑡
𝑌𝑖𝑡 − 𝑓̂(𝜏𝑡 ) − 𝑋𝑖𝑡⊤ 𝛽̂𝑡 − 𝛼
̂𝑖 , for 𝑖 =, 1, 2, … , 𝑁. Let ̂
𝒗𝑡 = 𝑣̂1𝑡 , 𝑣̂2𝑡 , … , 𝑣̂𝑁𝑡 .
Step 2: Resample the de-trended residuals 𝑣̂∗𝑖𝑡 = 𝑣̂𝑖𝑡 𝜂𝑖𝑡 , where 𝜂𝑡 is chosen
coefficients for the CCEP and CCEMG models are 0.0218 and 0.0376, √ √ √ √
respectively. The significant coefficient for tourist arrivals implies that to be −( 5 − 1)∕2 with a probability of ( 5 + 1)∕(2 5), and ( 5 + 1)∕2
a 10% increase in arrivals is associated with 0.22 or 0.38% higher real otherwise. Generate a bootstrapping sample of 𝑌𝑖𝑡 through
GDP, depending on the model used. Other variables that are significant 𝑌𝑖𝑡∗ = 𝑓̂(𝜏𝑡 ) + 𝑋𝑖𝑡⊤ 𝛽̂𝑡 + 𝛼 ̂ 𝑖 (𝜏𝑡 ) + 𝑣̂∗𝑖𝑡 ,
̂𝑖 + 𝑚
are real capital stock, exchange rate, and CPI when either estimator is
used, and the labor force when the CCEP estimator is used. for 𝑖 = 1, 2, … , 𝑁 and 𝑡 = 1, 2, … , 𝑇 .
The residuals obtained using the CCEP estimator can be used to Step 3: Based on the bootstrapped sample of {𝑌𝑖𝑡∗ , 𝑋𝑖𝑡 }, carry out LLDVE
check for cointegration among model variables (Holly et al., 2010). to obtain the estimates of the time-varying common trend 𝑓̂∗ (𝜏𝑡 ) and
That is, the residuals are examined using the CIPS test to determine coefficients 𝛽̂𝑡∗ , as well as the individual trend 𝑚
̂ ∗𝑖 (𝜏𝑡 ), for 𝑖 = 1, 2, … , 𝑁
whether they are stationary. If they are found to be stationary, cointe- and 𝑡 = 1, 2, … , 𝑇 .
gration is indicated between the dependent variable and independent
variables in the model. The CIPS test results show that the residuals ob- Step 4: Repeat Step 2-3 for 𝐵 = 1000 times and obtain the 90%
tained from the estimation are stationary up to four lags, indicating that confidence intervals for 𝑓 (𝜏𝑡 ), 𝛽𝑡 and 𝑚𝑖 (𝜏𝑡 ), for 𝑖 = 1, 2, … , 𝑁 and
there is cointegration between ln(GDP) and the independent variables 𝑡 = 1, 2, … , 𝑇 .
included in the model.14 Thus, we obtain the 90% confidence bands for the common trend
The parametric estimation methodology used here can capture the function, coefficient functions, and country-specific individual trend
time-varying effect of common factors; however, the estimated coeffi- functions.
cients are still time-invariant. We deal with this issue using the LLDVE
method described in the next section. 5.3. Empirical findings from nonparametric estimation

5. Panel data model with time-varying trends and coefficients 5.3.1. Tourism coefficient
The tourism coefficient function is plotted in Fig. 2. The con-
5.1. Time-varying trend and coefficients tribution of tourism to economic growth is positive and significant
during 1995–2003. After increasing steadily from 1997 onward, it
( )⊤ peaks around 2001 and then decreases until 2008. Tourism has a
Let 𝑌𝑖𝑡 be the dependent variable and 𝑋𝑖𝑡 = 𝑋𝑖𝑡,1 , 𝑋𝑖𝑡,2 , … , 𝑋𝑖𝑡,𝑘 be
a vector 𝑘 explanatory variables, for 𝑖 = 1, 2, … , 𝑁 and 𝑡 = 1, 2, … , 𝑇 . negative and significant effect on real GDP between 2007 and 2008
Li et al. (2011) proposed a fixed-effect panel data model with a and again from 2013 to 2014. The tourism coefficient ranges between
common time trend and time-varying coefficients: −0.05 and 0.05, with a mean value of 0.01. This indicates that the
contribution of tourism to economic growth has been positive on
𝑌𝑖𝑡 = 𝑓 (𝑡) + 𝑋𝑖𝑡⊤ 𝛽𝑡 + 𝛼𝑖 + 𝑢𝑖𝑡 , (3) average, but only for limited periods.
( The average coefficient reported here, as well as the coefficient
where 𝑓 (𝑡) are unknown country-specific trend functions, 𝛽𝑡 = 𝛽𝑡,1 ,
)⊤ obtained using parametric methods, compares well with the coefficients
… , 𝛽𝑡,𝑘 is an unknown vector of time-varying coefficients, 𝛼𝑖 is an
reported in panel data studies using methodologies comparable to
unknown individual effect, and 𝑢𝑖𝑡 is stationary for each 𝑖. For the
∑ ours. Xia et al. (2021) finds a coefficient of 0.041 using the AMG
purpose of identification, it is assumed that 𝑁 𝑖=1 𝛼𝑖 = 0 and that the estimator, implying that a 1% increase in the number of international
time variable 𝑡 is scaled by 𝑇 , such that 𝑓 (𝑡) = 𝑓 (𝜏𝑡 ) and 𝛽𝑡 = 𝛽(𝜏𝑡 ),
tourist arrivals is associated with a 0.04% increase in real GDP per
where 𝜏𝑡 = 𝑡∕𝑇 ∈ (0, 1].
capita. The coefficient estimated by Fayissa et al. (2008), who study
42 economies in Africa in the period of 1995-2004, implies that 1%
14
increase in tourism receipts is associated with a 0.03% increase in real
We tested for cross-sectional dependence first using the LM test, which
GDP per capita). Proença and Soukiazis (2008) find that 1% increase in
indicated cross-sectional dependence (𝜒 2 = 192 and 𝑝-value = 0.0009), with
the CD test of Pesaran (2015) indicating otherwise. The CIPS panel unit root tourism revenues is associated with 0.026 percentage point increase in
test statistics, 𝑧𝑡 for lags 0–4 range between −9.851 and −5.244 and are all real GDP per capita in their study of Greece, Italy, Portugal and Spain
significant at the 1% level. Maddala and Wu’s (1999) panel unit root test for the years between 1990 and 2004.
statistics, 𝜒 2 , for lags 0–4 gradually decrease from 291.209 to 141.914 and Our findings can possibly be interpreted within the context of Faber
are all significant at the 1% level. and Gaubert (2019), who posit that tourism’s contribution to aggregate

6
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Fig. 2. Graphs of estimated common trend and coefficient functions.

economic growth may be ambiguous. Our finding of a positive average values, then a significant negative tourism–growth relationship occurs.
effect during 1995–2003 may suggest that, at least during this period, Economic crises contribute to nonlinearities and complexity. Hence, a
productivity-enhancing effects of tourism activities, where they cluster, feasible explanation for the negative relationship between tourism and
are not offset by reduced productivity at other locations. economic growth during 2007–2011, in terms of the approach in Chiu
The negative coefficient we obtained for the period of GFC and the and Yeh (2017), is that the GFC changed the optimal threshold values.
recession in 2013 is probably driven by cross-country heterogeneity Wu et al. (2016) examine Granger causality between economic
in exposure to output shocks (how this could happen was explained growth and tourism both in the short run and in the long run in
in Section 2.2). In the conceptual framework above, a country with a Australia and nine countries located in Asia during the period of
larger tourism sector, as indicated by higher tourist arrival numbers, 1995–2013. The methodology they use is the panel smooth transition
could end up with a lower real GDP than a country with a smaller regression model with the real interest rate used as the threshold
tourism sector if the former experiences a more severe recession than variable. They find bi-directional causality between the two variables
the latter. The plots of common and country-specific trends in Fig. 3 that is nonlinear and varies with time and across countries, depending
clearly indicate the timing and the magnitude of these shocks. on the value of the threshold variable, the real interest rate.
Our finding of a time-varying coefficient aligns with the findings Another study that looks into time variation in tourism–growth
reported in several other studies. For instance, Antonakakis et al. relationship is that of Liu and Song (2018), who use a rolling Granger
(2015a) apply a time-varying model to examine the relationship be- causality test methodology to investigate the relationship between
tween tourism and economic growth in Europe from 1995 to 2012 and monthly visitor arrivals and quarterly GDP in Hong Kong between 1974
find that the tourism–economic growth relationship is not stable and and 2016. Their analysis shows that the relationship between tourism
becomes highly unstable during periods of crises. The relevant crisis and economic growth is unstable.
here is clearly the GFC, which slowed economic growth, as the tourism Shahzad et al. (2017) find that relationship between tourism–
coefficient function bottoms out during the height of the crisis. growth and growth in real GDP per capita in ten countries that they
Arslanturk et al. (2011), using Granger causality methods, find study is mostly positive. Contrary to what we find here, the relationship
that the time-varying effect of tourism (as measured by real tourism is found to be positive and especially pronounced during economic
receipts) in Turkey was insignificant from 1963 to 1975, negative from downturns. However, they also find that in a small number of countries,
1976 to 1983, and positive thereafter. The same methodology has also notably in China, the relationship between the two variables turns
been used in Balcilar et al. (2014) to examine the tourism–growth negative for some quantiles of economic growth.
relationship in South Africa during the period of 1960–2011. The study
finds that the effect of real tourism receipts on real GDP was positive 5.3.2. Common and country-specific trends
except for the period of 1985–1990. The common trend function plotted in Fig. 2 shows that logarithmic
Chiu and Yeh (2017) apply a threshold regression model to examine real GDP in the Schengen area countries included in the study, which
the tourism–economic growth relationship in 84 countries. They find reflects the autonomous growth in the real GDP and/or the joint effects
that if the threshold values are lower than certain optimal threshold of possible omitted variables. From the figures in the plot, it can be

7
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Fig. 3. Graphs of estimated country-specific individual trends.

calculated that during the 25 years of our study, real GDP increased Fig. 3 plots each country’s specific trend, which was obtained by
autonomously or due to the effects of possibly omitted factors by adding the country’s individual trend to the common trend. A coun-
approximately 43%, with annualized growth of around 1.45%. try’s specific trend reflects the country’s GDP growth pathway after
The most salient aspect of the plot of the common trend function accounting for the contributions of the six independent variables.
is that it exhibits a gradually increasing trend from 1995 to 2019,
The graphs show that the specific trends of Sweden and Switzerland
although the upward trend declines in 2000–2001 and again in 2013–
2014. The slowdown in the first period was part of a global economic are nearly the same as the common trend. The specific trends of Austria,
slowdown, attributable to higher energy and food prices, a downturn in Belgium, Germany, Italy, the Netherlands, Norway, and Portugal are
the information technology sector, and declining consumer confidence. clearly above the common trend before the period of GFC and below
The slowdown in the second period occurred because of an economic the common trend thereafter. Specific trends of Iceland and Luxem-
slowdown in the Euro area. bourg are below the common trend before the GFC period and above

8
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Fig. 4. Graphs of estimated common trend and coefficient functions with the lagged tourism being included.

afterward. Finland’s and Spain’s specific trends are above the common 2013. The CPI coefficient is positive and significant most of the time,
trend during the GFC period. suggesting a positive association between the CPI and real GDP. The
Specific trends of Hungary, Latvia, Poland, and Slovenia are below impact of the CPI on real GDP is positive until 2006, negative between
the common trend before 2004 and above afterward (with Poland’s 2007 and 2012, and positive again afterward. The negative relationship
break in 2007). Hungary, Latvia, Poland, and Slovenia joined the over the 2007–2012 period suggests supply shocks during this period,
European Union in May 2004 and have benefited from being European whereas the positive relationship during the Eurozone crisis indicates
members since that date. Their specific trends are not adversely affected shocks on the demand side.
by the subsequent GFC and European debt crisis.
The analysis of these trends clearly shows that the 2008–2009 GFC 5.4. Lagged effects of tourism
had a major effect in most countries. It is also clear from the plots
that the Eurozone debt crisis took its toll in most countries, excluding As described above, tourism’s contribution to long-term growth
Iceland, Luxembourg, Sweden, Switzerland, and the four countries arises through several channels. Some of these channels might affect
noted above. The effects of that crisis and the recession in 2013 appear GDP with a lag. For instance, the effects of learning-by-exporting or
to have been more severe in Italy and Portugal, which is not surprising learning-through-FDI can sometimes materialize with a delay. To check
as these countries were at the center of the debt crisis. this possibility, we repeat the time-varying estimation with a lagged
These plots support our conjecture that the sign of the tourism tourism term in the equations.15
coefficient changed to negative during the two recessions that hit the The plot of the lagged tourism coefficient shown in Fig. 4 indicates
economies under study based on the output shocks illustrated in the that, from 2009 onward, the time-varying coefficient of lagged tourism
plots. is largely significant, whereas the time-varying coefficient of tourism
is also significant during 2010–2011 but is insignificant after 2011.
5.3.3. Other variables During the period before the 2008 crisis, lagged tourism was significant
The coefficient functions of the other variables are also plotted in during 2002–2006, while tourism was significant during 1998–2004.
Fig. 2. Real GDP is influenced negatively by real capital stock during The common trend, country-specific trends, and the time-varying co-
1995–1996 and positively between 2001 and 2007 and after 2016. efficients of the variables other than tourism are largely the same as
The impact of capital stock has been increasing since 2016, becoming those in the original time-varying model (see Figs. 4 and 5).
stronger after 2018. The labor force has a positive and significant effect Finally, we note that the lagged tourism coefficient is positive and
on real GDP between 1996 and 2012. The effect turns negative and significant for the recession years of 2009, 2012, and 2013, whereas
significant during the 2013–2017 period. the tourism coefficient is not. Compared with the coefficients obtained
Trade openness is positive and statistically significant only between from the original model, this may suggest that the original estimates are
1996 and 2003; since then, its effect has been negative and insignif-
icant. The impact of the exchange rate on real GDP is negative and
15
significant, except for the period between 1995 and 2001 and after We thank an anonymous reviewer for suggesting this method.

9
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Fig. 5. Graphs of estimated country-specific individual trends with the lagged tourism being included.

biased upward. However, we do not think that the lagged effect could aligns with many studies finding that tourism contributes positively to
possibly be large enough to yield a positive coefficient in periods when economic growth. Therefore, we agree with these authors that it would
the economy experiences severe output shocks, such as during the GFC be a good idea to make some effort to implement tourism development
and the Eurozone crisis. Hence, while it is obvious that tourism has a policies to sustain a large flow of tourism.
lagging effect, determining its size calls for more precise modeling of We mentioned above that it is possible that the aggregate productiv-
the issue. ity could increase as more resources are allocated to the tourism sector
if the productivity in this sector is higher than that of at least some
6. Policy implications of the other service sectors. Governments can adopt policies that could
help to increase tourism productivity. This could be achieved by putting
Our finding that the tourism coefficient may make a positive contri- policies in place to facilitate the adoption of new technologies in the
bution to economic growth and time-varying has several implications. tourism sector. These policies can be in the form of financial support
The result, indicating a positive tourism–economic growth relationship, schemes or programs that can increase the absorptive capacity of the

10
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

service providers. Value added can also be increased by better training tourism and economic growth is highly nonlinear, with the tourism co-
of employees to increase the service quality. efficient function peaking in 1995 and 2001 and bottoming out in 2008
The segmented nature of the tourism demand suggests another way during the GFC. Our results are consistent with the view that crises can
to promote tourism. Governments and businesses could try to identify generate instability and nonlinearities in the tourism–economic growth
those demand segments or markets (countries) that create larger effects relationship and even change the sign of the relationship for a limited
first (for instance, these could be young travelers from countries with period.
economies on an up cycle). Then they can concentrate their promotion Policymakers might want to adopt policies to diversify demand in
and marketing efforts on these segments. addition to conventional tourism promotion policies. To this extent,
A related issue is that certain segments are seen as recession-proof. and considering that the categories of tourists, not just their numbers,
For instance, older travelers (retirees from developed countries, for might matter, it would be a good idea to analyze the tourism–growth
instance) are usually considered recession-proof – that is, their numbers relationship using tourism data disaggregated into several categories
do not decline much during recession times – but going solely by num- whenever possible, such as business versus leisure travelers.
bers could be misleading since the spillover effects of certain groups Our results emphasize that policymakers should consider the time-
could be lower than those of others. For instance, if the spillover effects varying link between tourism and economic growth when designing
generated by younger tourists, who tend to reduce their traveling tourism policies. It has become almost accepted wisdom that the rela-
during recessionary times, are higher than that of the older tourists, tionship between tourism and economic growth is positive. Our results
the net effect on productivity could be negative even if the numbers of suggest that this is generally the case for the group of countries included
older tourists did not change. in this study, but there can be times when the relationship becomes
Some authors also suggest that tourism could act as a buffer during unstable and even negative, particularly around times of crisis. Clearly,
recessions and downturns. For instance, Shahzad et al. (2017) empha- more research is needed to identify the nature and channels of produc-
size that when the tourism–economic growth relationship is nonlinear, tivity spillovers and how they change over time. Once this has been
policymakers should consider the specific phase of the economic cycle done, the time-varying nature of the tourism–growth relationship can
when designing tourism policies and that tourism-enhancing policies be fully utilized.
can be beneficial in periods of economic downturn. Analysis of both the common and specific trend functions shows a
Tourism promotion policies during recessionary periods can be of common upward trend in real GDP, and country-specific trends diverge
some help in stimulating the economy, but since recessionary shocks hit from the common trend asynchronously. This has several implications
several sectors at the same time, expansion in tourism demand would for tourism policy. First, these trends could be an indication of the
not be enough to overturn the fall in output in other sectors. substantial contribution of autonomous (independent of tourism) TFP
The plots of country-specific trends show that some countries trend to economic growth, as much higher than the contribution of tourism.
above the common trend, whereas others trend below it. Thus, those Thus, directing policy efforts into enhancing TFP by other methods
countries with below-average GDP trends might want to adopt aggres- rather than through tourism development and promotion policies might
sive tourism policies to promote economic growth during down cycles. be a better idea in the long run. Second, the asynchronous nature of
Such an approach could create conflicts among EU member countries, country-specific trends might make it difficult to coordinate tourism
however, and complicate the adoption of a common policy. policies, and indeed economic policies, among EU countries.
The growth in the trend real GDP can be interpreted as the growth Our results illustrate the importance of looking at how a coefficient
in the autonomous TFP, which is not a function of tourism. Hence, it behaves in the short run to better understand the nature of the long-run
can be said that in the countries under study, TFP has been growing relationship between two variables. The nonparametric time-varying
more than 1% a year during the period of 1995-2019 for reasons other methodology that we use in this study is an ideal tool for this. It
than the growth in tourism. This is much higher than the efficiency can be used to check the stability of the coefficients obtained in an
enhancing effects of tourism that we mentioned above. Hence, to empirical analysis to ensure that they can be used to make policy
the extent this efficiency growth can be attributed to the growth in decisions. There are other methods of dealing with time variation, but
technology, it would pay off more to concentrate policy efforts on they are mostly Granger causality-type analyses or involve assumptions
sustaining technological innovation, rather than promoting tourism. about the structure of the relationship under study. The nonparametric
Since the trend real GDP continued to grow during GFC, this also could time-varying methodology we adopt here, which is somewhat easier
prove to be a more effective way of boosting the economy in times of to implement, can be used as an alternative to these other methods or
recession. could be used in combination with them.

7. Conclusions Declaration of competing interest

A large body of literature examines the relationship between tourism We would like to confirm that the re-submitted manuscript is not
and economic growth and specifically tests the TLGH. This literature subject to any conflict of interest.
has increasingly realized that the relationship between tourism and
economic growth is nonlinear and unstable. However, existing stud- Data availability
ies that have employed parametric models to reflect instability and
nonlinearities in the tourism and economic growth relationship are Code and data will be available upon request.
not flexible enough to truly capture the complexity of the underlying
relationship. Our novel contribution has been to model the tourism– References
economic growth relationship in a nonparametric framework that
Adamou, A., Clerides, S., 2009. Prospects and limits of tourism-led growth: The
provides a much better insight into the time-varying link between
international evidence. Rev. Econ. Anal. 287–303.
tourism and economic growth over time. Ahmad, N., Menegaki, A.N., Al-Muharrami, S., 2020. Systematic literature review of
We have applied our nonparametric framework to model the rela- tourism growth nexus: An overview of the literature and a content analysis of 100
tionship between tourism and economic growth in 17 Schengen area most influential papers. J. Econ. Surv. 34 (5), 1068–1110.
countries from 1995 to 2019. We find that while our estimated non- Antonakakis, N., Dragouni, M., Eeckels, B., Filis, G., 2016. Tourism and economic
growth: Does democracy matter? Ann. Tour. Res. 61, 258–264.
parametric coefficient function for tourism was positive and significant Antonakakis, N., Dragouni, M., Eeckels, B., Filis, G., 2019. The tourism and economic
during the period of 1995–2003, it was negative and significant in growth enigma: Examining an ambiguous relationship through multiple prisms. J.
2007, 2008, 2013, and 2014. We also find that the relationship between Travel Res. 58 (1), 3–24.

11
E. Dogan and X. Zhang Economic Modelling 128 (2023) 106487

Antonakakis, N., Dragouni, M., Filis, G., 2015a. How strong is the linkage between Keller, W., Yeaple, S.R., 2009. Multinational enterprises, international trade, and
tourism and economic growth in Europe? Econ. Model. 44, 142–155. productivity growth: Firm-level evidence from the United States. Rev. Econ. Stat.
Antonakakis, N., Dragouni, M., Filis, G., 2015b. Tourism and growth: The times they 91 (4), 821–831.
are a-changing. Ann. Tour. Res. 50, 165–169. Kim, D.-H., 2011. Trade, growth and income. J. Int. Trade Econ. Dev. 20 (5), 677–709.
Arslanturk, Y., Balcilar, M., Ozdemir, Z.A., 2011. Time-varying linkages between Lean, H.H., Smyth, R., 2009. Asian financial crisis, avian flu and terrorist threats: Are
tourism receipts and economic growth in a small open economy. Econ. Model. shocks to Malaysian tourist arrivals permanent or transitory? Asia Pac. J. Tour.
28 (1–2), 664–671. Res. 14 (3), 301–321.
Awaworyi Churchill, S., Inekwe, J., Smyth, R., Zhang, X., 2019. R&D intensity and Lee, C.-C., Chien, M.-S., 2008. Structural breaks, tourism development, and economic
carbon emissions in the G7: 1870–2014. Energy Econ. 80, 30–37. growth: Evidence from Taiwan. Math. Comput. Simulation 77 (4), 358–368.
Balaguer, J., Cantavella-Jorda, M., 2002. Tourism as a long-run economic growth factor: Li, D., Chen, J., Gao, J., 2011. Non-parametric time-varying coefficient panel data
The spanish case. Appl. Econ. 34 (7), 877–884. models with fixed effects. Econom. J. 14 (3), 387–408.
Lin, V.S., Yang, Y., Li, G., 2019. Where can tourism-led growth and economy-driven
Balcilar, M., Van Eyden, R., Inglesi-Lotz, R., Gupta, R., 2014. Time-varying linkages
tourism growth occur? J. Travel Res. 58 (5), 760–773.
between tourism receipts and economic growth in South Africa. Appl. Econ. 46
Liu, A., Kim, Y.R., Song, H., 2022. Toward an accurate assessment of tourism economic
(36), 4381–4398.
impact: A systematic literature review. Ann. Tour. Res. Empir. Insights 3 (2),
Blomström, M., Kokko, A., 1998. Multinational corporations and spillovers. J. Econ.
100054.
Surv. 12 (3), 247–277.
Liu, H., Song, H., 2018. New evidence of dynamic links between tourism and economic
Breusch, T.S., Pagan, A.R., 1980. The Lagrange multiplier test and its applications to
growth based on mixed-frequency granger causality tests. J. Travel Res. 57 (7),
model specification in econometrics. Rev. Econom. Stud. 47 (1), 239–253.
899–907.
Brida, J.G., Cortes-Jimenez, I., Pulina, M., 2016a. Has the tourism-led growth hypothesis
Maddala, G.S., Wu, S., 1999. A comparative study of unit root tests with panel data
been validated? A literature review. Curr. Issues Tour. 19 (5), 394–430.
and a new simple test. Oxf. Bull. Econ. Stat. 61 (S1), 631–652.
Brida, J.G., Gómez, D.M., Segarra, V., 2020. On the empirical relationship between Madsen, J.B., Islam, M.R., Doucouliagos, H., 2018. Inequality, financial development
tourism and economic growth. Tour. Manag. 81, 104131. and economic growth in the OECD, 1870–2011. Eur. Econ. Rev. 101, 605–624.
Brida, J.G., Lanzilotta, B., Pereyra, J.S., Pizzolon, F., 2015. A nonlinear approach to Mammen, E., 1993. Bootstrap and wild bootstrap for high dimensional linear models.
the tourism-led growth hypothesis: The case of the MERCOSUR. Curr. Issues Tour. Ann. Statist. 255–285.
18 (7), 647–666. Narayan, P.K., 2005. Did Rabuka’s military coups have a permanent effect or a
Brida, J.G., Lanzilotta, B., Pizzolon, F., 2016b. Dynamic relationship between tourism transitory effect on tourist expenditure in Fiji: Evidence from Vogelsang’s structural
and economic growth in MERCOSUR countries: A nonlinear approach based on break test. Tour. Manag. 26 (4), 509–515.
asymmetric time series models. Econ. Bull. 36 (2), 879–894. Pan, S.-C., Liu, S.-Y., Wu, P.-C., 2014. Re-testing the tourism-led growth hypothesis
Castro-Nuño, M., Molina-Toucedo, J.A., Pablo-Romero, M.P., 2013. Tourism and GDP: using panel smooth transition regression models. Tour. Econ. 20 (1), 39–50.
A meta-analysis of panel data studies. J. Travel Res. 52 (6), 745–758. Pérez-Rodríguez, J.V., Ledesma-Rodríguez, F., Santana-Gallego, M., 2015. Testing
Chang, C.-L., Khamkaew, T., McAleer, M., 2012. IV estimation of a panel threshold dependence between GDP and tourism’s growth rates. Tour. Manag. 48, 268–282.
model of tourism specialization and economic development. Tour. Econ. 18 (1), Pesaran, M.H., 2007. A simple panel unit root test in the presence of cross-section
5–41. dependence. J. Appl. Econometrics 22 (2), 265–312.
Chang, C.-C., Mendy, M., 2012. Economic growth and openness in Africa: What is the Pesaran, M.H., 2015. Testing weak cross-sectional dependence in large panels.
empirical relationship? Appl. Econ. Lett. 19 (18), 1903–1907. Econometric Rev. 34 (6–10), 1089–1117.
Chao, C.-C., Hazari, B.R., Laffargue, J.-P., Sgro, P.M., Yu, E.S., 2006. Tourism, Dutch Phiri, A., 2015. Tourism and Economic Growth in South Africa: Evidence from Linear
disease and welfare in an open dynamic economy. Jpn. Econ. Rev. 57, 501–515. and Nonlinear Cointegration Frameworks. MPRA working paper, No.65000.
Chiu, Y.-B., Yeh, L.-T., 2017. The threshold effects of the tourism-led growth hypothesis: Po, W.-C., Huang, B.-N., 2008. Tourism development and economic growth – a
Evidence from a cross-sectional model. J. Travel Res. 56 (5), 625–637. nonlinear approach. Physica A 387 (22), 5535–5542.
Proença, S., Soukiazis, E., 2008. Tourism as an economic growth factor: A case study
Copeland, B.R., 1991. Tourism, welfare and de-industrialization in a small open
for Southern European countries. Tourism Economics 14 (4), 791–806.
economy. Economica 515–529.
Sarkar, P., 2008. Trade openness and growth: Is there any link? J. Econ. Issues 42 (3),
De Vita, G., Kyaw, K.S., 2016. Tourism development and growth. Ann. Tour. Res. 23–26.
763–785.
De Vita, G., Kyaw, K.S., 2017. Tourism specialization, absorptive capacity, and
Shahbaz, M., 2012. Does trade openness affect long run growth? Cointegration,
economic growth. J. Travel Res. 56 (4), 423–435.
causality and forecast error variance decomposition tests for Pakistan. Econ. Model.
Deng, T., Ma, M., Shao, S., 2014. Has international tourism promoted economic growth
29 (6), 2325–2339.
in China? A panel threshold regression approach. Tour. Econ. 20 (4), 911–917. Shahzad, S.J.H., Shahbaz, M., Ferrer, R., Kumar, R.R., 2017. Tourism-led growth
Dogru, T., Bulut, U., 2018. Is tourism an engine for economic recovery? Theory and hypothesis in the top ten tourist destinations: New evidence using the
empirical evidence. Tour. Manag. 67, 425–434. quantile-on-quantile approach. Tour. Manag. 60, 223–232.
Dufrenot, G., Mignon, V., Tsangarides, C., 2010. The trade-growth nexus in the Sianesi, B., Reenen, J.V., 2003. The returns to education: Macroeconomics. J. Econ.
developing countries: A quantile regression approach. Rev. World Econ. 146, Surv. 17 (2), 157–200.
731–761. Silvapulle, P., Smyth, R., Zhang, X., Fenech, J.-P., 2017. Nonparametric panel data
Enilov, M., Wang, Y., 2022. Tourism and economic growth: Multi-country evidence model for crude oil and stock market prices in net oil importing countries. Energy
from mixed-frequency granger causality tests. Tour. Econ. 28 (5), 1216–1239. Econ. 67, 255–267.
Eyuboglu, S., Eyuboglu, K., 2020. Tourism development and economic growth: An Singh, T., 2010. Does international trade cause economic growth? A survey. World
asymmetric panel causality test. Curr. Issues Tour. 23 (6), 659–665. Econ. 33 (11), 1517–1564.
Faber, B., Gaubert, C., 2019. Tourism and economic development: Evidence from Smyth, R., Nielsen, I., Mishra, V., 2009. ‘I’ve been to Bali too’(and I will be going
Mexico’s coastline. Amer. Econ. Rev. 109 (6), 2245–2293. back): Are terrorist shocks to Bali’s tourist arrivals permanent or transitory? Appl.
Fan, J., Zhang, W., 2008. Statistical methods with varying coefficient models. Stat. Econ. 41 (11), 1367–1378.
Interface 1 (1), 179–195. Song, H., Li, G., Cao, Z., 2018. Tourism and economic globalization: An emerging
Fayissa, B., Nsiah, C., Tadasse, B., 2008. Impact of tourism on economic growth and research agenda. J. Travel Res. 57 (8), 999–1011.
development in Africa. Tour Economics 14 (4), 807–818. Tang, C.F., Tan, E.C., 2015. Does tourism effectively stimulate Malaysia’s economic
Figini, P., Patuelli, R., 2022. Estimating the economic impact of tourism in the European growth? Tour. Manag. 46, 158–163.
Union: Review and computation. J. Travel Res. 61 (6), 1409–1423. UNWTO, 2018. European Union Tourism Trends. UNWTO, Madrid.
Francois, J., Hoekman, B., 2010. Services trade and policy. J. Econ. Lit. 48 (3), 642–692. Wang, Y.-S., 2012. Research note: Threshold effects on development of tourism and
Hailemariam, A., Smyth, R., Zhang, X., 2019. Oil prices and economic policy un- economic growth. Tour. Econ. 18 (5), 1135–1141.
certainty: Evidence from a nonparametric panel data model. Energy Econ. 83, Wu, C.-F.J., 1986. Jackknife, bootstrap and other resampling methods in regression
40–51. analysis. Ann. Statist. 14 (4), 1261–1295.
Wu, P.-C., Liu, S.-Y., Hsiao, J.-M., Huang, T.-Y., 2016. Nonlinear and time-varying
Hoekman, B., Shepherd, B., 2017. Services productivity, trade policy and manufacturing
growth-tourism causality. Ann. Tour. Res. 59, 45–59.
exports. World Econ. 40 (3), 499–516.
Wu, T.-P., Wu, H.-C., 2017. The influence of international tourism receipts on economic
Holly, S., Pesaran, M.H., Yamagata, T., 2010. A spatio-temporal model of house prices
development: Evidence from China’s 31 major regions. J. Travel Res. 871–882.
in the USA. J. Econometrics 158 (1), 160–173.
Xia, W., Doğan, B., Shahzad, U., Adedoyin, F.F., Popoola, A., Bashir, M.A., 2021. An
Hye, Q.M.A., Lau, W.-Y., 2015. Trade openness and economic growth: Empirical
empirical investigation of tourism-led growth hypothesis in the European countries:
evidence from India. J. Bus. Econ. Manag. 16 (1), 188–205.
Evidence from augmented mean group estimator. Port. Econ. J. 21, 239–266.
Inchausti-Sintes, F., 2015. Tourism: Economic growth, employment and dutch disease. Yao, Y., Ivanovski, K., Inekwe, J., Smyth, R., 2020. Human capital and CO2 emissions
Ann. Tour. Res. 54, 172–189. in the long run. Energy Econ. 91, 104907.
Javorcik, B.S., 2004. Does foreign direct investment increase the productivity of Zhang, Y., Su, L., Phillips, P.C.B., 2012. Testing for common trends in semi-parametric
domestic firms? In search of spillovers through backward linkages. Amer. Econ. panel data models with fixed effects. Econom. J. 15 (1), 56–100.
Rev. 94 (3), 605–627. Zuo, B., Huang, S., 2018. Revisiting the tourism-led economic growth hypothesis: The
Jin, J.C., 2011. The effects of tourism on economic growth in Hong Kong. Cornell Hosp. case of China. J. Travel Res. 57 (2), 151–163.
Q. 52 (3), 333–340.

12