Modelling of nonlinear asymmetric effects of changes in tourism on economic growth in an autonomous small-island economy

One of the most widely-researched, yet persistently elusive, subjects of tourism research over the last two decades has been the tourism-led growth hypothesis (TLGH). Beginning with a seminal study by Balaguer and Cantavella-Jordá (2002), a substantial literature has developed around establishing a theoretical basis for the TLGH, which envisages a causal relationship between the growth of tourism and economic growth in the destination’s economy (Brida et al., 2016; Marrocu & Paci, 2014, Nunkoo et al., 2020). This effect is considered to be especially important in the case of small-island economies, which often rely upon tourism as a source of foreign currency to fund the purchase of raw materials and technology from overseas (Apergis & Payne, 2012; Archer, 1995; Narayan, 2004; Roudi et al., 2019). Indeed, empirical studies attempting to verify the TLGH have concentrated disproportionately on island economies, with studies focusing on, inter alia, Antigua (Schubert et al., 2011); Bahamas (Singh et al., 2010); Barbados (Schubert et al., 2011; Singh et al., 2010); Cook Islands (Kumar et al, 2020); Fiji (Kumar & Patel, 2022; Narayan, 2004); Hong Kong (Jin, 2011); Jamaica (Singh et al., 2010); Malta (Katircioglu, 2009b); Mauritius (Solarin, 2018), Papua New Guinea (Kumar et al., 2022b); São Tomé and Principe (Da Costa Ribeiro & Wang, 2019); Taiwan (Kim & Chen, 2006); Tonga (Kumar & Patel, 2022; Kumar et al., 2021, 2022a); and Vanuatu (Kumar & Patel, 2022). Akadiri and Akadiri (2021), Apergis and Payne (2012), Fahimi et al., (2018) and Seetanah (2011), meanwhile, all have conducted studies using panel data from multiple island destinations.

The empirical literature on the TLGH has been reviewed by Pablo-Romero and Molina (2013). They found that the studies initially tended to employ the autoregressive distributed lag (ARDL) model in an attempt to establish and measure a relationship between the growth of tourism receipts and economic growth in destination economies (e.g., Adnan Hye et al., 2013; Kyophilavong et al., 2018; Lawal et al., 2018; Othman et al., 2012; Suryandaru, 2020). The traditional ARDL model suffers, however, from two major limitations. First, it does not distinguish between the stimulating effect of tourism growth and the retarding effect of tourism recession, which are assumed to be symmetrical. Second, such models assume a linear relationship between tourism growth and economic growth. This means they are unable to capture short-run volatility in the variables of concern that may be associated with, for example, an external shock (or a structural break).

The asymmetric nonlinear autoregressive distributed lag model (NARDL) model, as proposed by Shin et al. (2014) can, however, investigate asymmetries both in the short term and in the long run. The NARDL approach allows the researcher to highlight differences both in terms of coefficient size and causality effects. It has therefore begun to be adopted by researchers interested in the TLGH (e.g., Bastidas-Manzano et al., 2020; Charfeddine & Dawd, 2022; Husein & Kara, 2020; Kumar & Patel, 2022; Kumar et al., 2020, 2022a, 2022b). Such studies have, however, tended to produce inconsistent results, with some suggesting strong asymmetry in the relationship between tourism receipts and growth in the destination economy and others not. The nature of the asymmetry also varies substantially in the findings of such studies. Some of these studies have been applied in contexts where only annual data, often with a relatively short range, have been available. The effect of this has been to overlook short-term effects, which can be particularly important when tourism exhibits a substantial degree of seasonality.

This study aims, therefore, to evaluate the TLGH using a NARDL model populated with quarterly data from 1976 to 2019. The data are from Madeira, an island economy located in the Atlantic Ocean, where tourism accounts for a major proportion of GDP and has a substantial degree of seasonality. Madeira’s major specialisation in the tourism sector, and its exposure to an increasing array of internal and external risks, provides a strong rationale for examining the TLGH using data from Madeira. It is also relevant to note that while Madeira is an island economy, it is also one of two autonomous regions of Portugal. NARDL has yet to be applied to test the TLGH in such a context, with previous studies tending to have been focused on independent island states.

The context of this study is the non-independent but autonomous island region of Madeira. As is the case with many other small-island economies (Bardolet & Sheldon, 2008), the tourism sector occupies a large share of the Madeiran economy (16.2% of GVA in 2019). Based on its direct and indirect contributions to the labour market, the sector accounted for 17% of total employment in Madeira’s economy (Direção Regional de Estatística da Madeira, 2022). Based on the tourism satellite account, meanwhile, the ratio of tourism expenditure to the GDP was 28.8% in 2019 (Direção Regional de Estatística da Madeira, 2022).

The sector is also believed to play a significant role in stimulating public investment in new infrastructure in the peripheral and rural hinterland areas by supporting (indirectly through taxation) business investment through strategic spatial development and promoting competition in locations outside the main capital city that have been characterised by low entrepreneurship (ISMERI Europa 2011; Majdak & de Almeida, 2022). As is true elsewhere, the region is trying to develop niche markets based on high-spending visitors who are attracted by the region’s natural and cultural resources. To this end, it uses a diversified cultural agenda of high-quality events spread over the tourist season (Almeida & Garrod, 2021).

In this study, the analysis of the relationship between tourism and economic growth is based on two annual series – tourism arrivals and GDP – which is in line with the practices of Kyara et al. (2021), Husein and Kara (2020), and Song et al. (2019). Annual data for the real GDP (measured in constant 2019 euros) was available for 1976 to 2019. Arrival data was also available from 1976 to 2019. The Madeira Statistics Office provided the datasets. Figure 1 illustrates these time series. As can be seen, tourism arrivals alone cannot provide an adequate explanation of trends in GDP, especially since 1986.

Figure 1:

Several tests were therefore carried out to identify the best proxy for tourism activity. On an annual basis, the variable “tourism receipts” presented itself as a potential proxy. In line with expectations, preliminary analysis of the data in logarithmic form produced an OLS estimate with a positive and significant coefficient. Moreover, the Johansen cointegration test, widely used to determine the number of independent linear combinations, indicated that both variables were integrated of order 1. However, because the range is relatively small for the annual data (1976–2019), it comes as no surprise that the NARDL approach fails to detect asymmetry in the short term. For this reason, frequency conversion techniques were applied to define a sub-annual periodicity more suitable to examining short-term impacts. Quarterly data were also compatible with an increase in the number of degrees of freedom. Trimestral data were already available for the arrivals variable; this was not the case for either GDP or tourism receipts.

Several methods were adopted to convert the annual data to a quarterly frequency, including the quadratic match sum (QMS) technique, known for its seasonality-adjustment proprieties, and the Denton method (Shahzad et al., 2017, 2018; Nadeem et al., 2020; Kumar et al., 2021a; Kumar et al., 2022). Unlike Kumar et al. (2022), this study opted for the Denton method in order to avoid potential problems related to either stable or moving seasonality (Shahzad et al., 2017; Shahbaz et al. 2018; Kumar et al., 2021; Kumar et al, 2022). Of course, other cointegration methods, not applied in this study, such as those developed by Granger (1988), Hylleberg et al. (1990) and Engle et al. (1993), are also of interest in this regard. The quarterly frequency conversion resulted in a 176-period database. When quarterly data are used, “arrivals” becomes a better proxy for studying the impact of the tourism dynamics on GDP, rather than “tourism receipts.”

This study employs both ARDL bounds testing and causality testing based on an error correction model (ECM) to confirm the hypothesis of unidirectional causality from tourism arrivals to economic growth (Liu & Song, 2018). In line with past studies, the empirical calculation of the impact of tourism uses the following linear equation: (1) GDPt=μ+β*Gt+εt {GDP_t} = \mu + \beta *{G_t} + {\varepsilon _t} where G stands for the explanatory variable (arrivals), GDP stands for the dependent variable, β is the elasticity coefficient of arrivals, μ stands for the intercept, and ɛt stands for the error term.

Equation (1), based on Hu et al. (2022), identifies the long-run coefficients, whereas the error correction approach, defined below in general terms, includes both long-run and short-run effects. The standard representation of a symmetric linear ECM, with both the long- and short-run dynamic relationship, is as follows: (2) Δyt=u+ρy*yt−1+ρx*xt−1+∑i=0p−1αiΔyt.i+∑i=1q−1βi*Δxt−i+εt \Delta {y_t} = {\rm{u}} + {\rho _y}*{y_{t - 1}} + {\rho _x}*{x_{t - 1}} + \sum\nolimits_{i = 0}^{p - 1} {{\alpha _i}\Delta {y_{t.i}}} + \sum\nolimits_{i = 1}^{q - 1} {{\beta _i}*\Delta {x_{t - i}}} + {\varepsilon _t} Equation (2), in the ARDL format (error correction approach), considers the following: (3) ΔGDPt=v+∑i=0p−1nviΔGDPt−j+∑i=0p−1nviΔxt−j+γ1GDPt−1+γ2xt−1εt \Delta {GDP_t} = {\rm{v}} + \mathop \sum \nolimits_{i = 0}^{p - 1n} {v_i}\Delta {GDP_{t - j}} + \mathop \sum \nolimits_{i = 0}^{p - 1n} {v_i}\Delta {x_{t - j}} + {\gamma _1}{GDP_{t - 1}} + {\gamma _2}{x_{t - 1}}{\varepsilon _t} The short-run parameters (identified by the symbol Δ) in both variables result from changes, while the symbols γ₁ and γ₂ identify long-term parameters. In the NARDL approach, the explanatory variables must disaggregate into positive and negative sub-variables to investigate both the occurrence of a nonlinear relationship between variables and the existence of asymmetric effects in the long and short run. It is noteworthy that the volume of arrivals generated by global tourism demand has been increasingly volatile and prone to unpredictable abrupt falls in external demand, as the COVID experience fully demonstrated. Since this study aims to identify occurrences of nonlinear relationships among variables, we define the equation that splits the explanatory variable into positive and negative accordingly: (4) Gt=G0+Gt++Gt− {G_t} = G{\it 0} + G_t^ + + G_t^ - Equation (1) is thus re-defined as follows: (5) GDPt=β0+β1*Gt++β2*Gt−+εt {GDP_t} = {\beta _0} + {\beta _1}*G_t^ + + {\beta _2}*G_t^ - + {\varepsilon _t} The β coefficients represent the long term. The following equations define the partial sum of the positive changes in the T_r variable: (6) Gt+=∑j=1tΔGj+=∑j=1tmaxΔGj,0 G_t^ + = \mathop \sum \nolimits_{j = 1}^t \Delta G_j^ + = \mathop \sum \nolimits_{j = 1}^t {\rm{max}}\left( {\Delta {G_j},0} \right) (7) Gt−=∑j=1tΔGj−=∑j=1tminΔG,0 G_t^ - = \mathop \sum \nolimits_{j = 1}^t \Delta G_j^ - = \mathop \sum \nolimits_{j = 1}^t {\rm{min}}\left( {\Delta G,0} \right) The nonlinear formula of the NARDL approach is as follows: (8) ΔGDPt=v+∑i=0p−1viΔGDPt−j+∑i=0q−1viΔGt−j+γ1GDPt−1+γ2Gt−1+εt \Delta {GDP_t} = v + \mathop \sum \nolimits_{i = 0}^{p - 1} {v_i}\Delta {GDP_{t - j}} + \mathop \sum \nolimits_{i = 0}^{q - 1} {v_i}\Delta {G_{t - j}} + {\gamma _1}{GDP_{t - 1}} + {\gamma _2}{G_{t - 1}} + {\varepsilon _t} The terms np and no represent the lag order. The terms G_t⁺ and G_t⁻ signify the positive and negative shocks in terms of the number of arrivals. The symbol αi represents the long-run coefficients, γi the short-run coefficients (Long & Zhang, 2022). This dynamic error correction approach is commonly used to evaluate the relationship among integrated variables of order I(1) (Kumar et al., 2022).

Previous researchers have tended to employ control variables, such as the real exchange rate, because the rate of real exchange proxies the evolution of tourism prices and the impact on the volume of tourism demand of countries with cheaper currencies (Katircioglu, 2009a; Kumar et al., 2021; Oh, 2005; Song et al., 2019; Tang, 2018; Kumar et al, 2021). Nowadays, as more than 80% of tourists arriving in Madeira come from the Eurozone (DREM, 2022), the topic of exchange rates is not a major issue except for the British market. Portugal joined the eurozone in 2001. Nevertheless, the impact of changes in the exchange rate were explored in this study, given that the British market corresponds to one-fifth of the total number of arrivals. The impact of real exchange rate and inflation were investigated, and both were found to be determinants of tourism demand. In some instances of high inflation rates, the decision of not travelling may be the most rational one, as Madeira may be not economically very attractive due to high air-transport costs. For that reason, this study also considers the impact of gains in purchasing power proxied by the evolution of GDP per capita across the European countries as a whole.

In an EU funds-dependent economy such as Madeira, it is also important to study the impact of the volume of financial support from the EU. This study therefore tested the impact of several variables and dummies reflecting the impact of the accession to the EC in 1986, including total (current and capital) government expenditures, a dummy variable representing the pre- and post-accession period, and dummies for the multiannual financial frameworks. The most statistically significant variables were found to be the dummies representing the various multiannual financial frameworks.

Annual data for the variable “Europe GDP” measured in constant 2015 US dollars for the period spanning 1976–2019 was retrieved from the World Bank database. Annual data on the inflation rate for Portugal was sourced from the PORDATA database, and quarterly data on tourist arrivals, tourism receipts and overnights were retrieved from the Madeira Statistical Office. Quarterly data for the inflation rate and exchange rate were retrieved from the World Bank database. Substantial preparatory work was carried out to identify the relationship between the economic dynamics and the ups and downs of the tourism market. As mentioned above, while the analysis based on annual data led to positive and significant results, it failed to identify short-term asymmetries.

Before checking for a long-run relationship between tourism-related variables and a proxy for economic growth, several traditional unit root tests were conducted to examine the degree of stationarity of the variables under the null hypothesis of a series having a unit root. It is advisable to check for structural breaks due to events such as economic shocks and methodological changes in the compilation of the datasets. Tests developed by Bai and Perron (1998, 2002), Clemente et al. (1998) and Zivot and Andrews (1992) were used to achieve this. Such tests identify the number and location of more than one breakpoint. It is for the researcher, then, to determine whether such breaks can be readily connected to real-world events. Break dummies were not retained in the present study, as they were found not to be significant in the final estimation.

Once the diagnosis of unit roots and structural breaks was completed, the study continued to estimate the long-run and short-run parameters and the cointegration equation based on the ARDL bounds testing technique proposed by Pesaran et al. (2001). The asymmetrical NARDL approach was then employed to analyse the occurrence and form of any asymmetric effects. Granger causality tests were then used to confirm the existence and direction of causality, if any, between the variables under analysis.

Several models were tested with different proxies and formats of variables to detect the most statistically suitable model. To avoid spurious results when using time-based data, unit root tests needed to be performed before employing any cointegration techniques (e.g., NARDL) to estimate the long-run and short-run parameters. Another reason for pre-testing is to establish the stationarity or otherwise of the variables. The augmented Dickey-Fuller (ADF) unit root test and the Phillips and Perron (PP) test (Phillips & Perron, 1988) were used for this purpose. One of the limitations of the NARDL approach is its ineffectiveness in handling the I(2) series because, in such cases, F-statistics are invalid (Hu et al., 2021; Ibrahim, 2015; Meo et al., 2018). Table 2 shows the results of the unit root tests. The variables are stationary at the first difference and non-stationary in levels, and none are stationary at I(2). Therefore, there is no cause for concern.

Owing to lack of space, the discussion here is limited to presenting the results of the unit roots test. Further details on the augmented Dickey-Fuller (ADF) unit root test, the Phillips and Perron (PP) test and unit root tests with structural breaks can be found in Zivot and Andrew (1992), Tang and Tan (2015), Mérida and Golpe (2016) and STATA (2023).

Three tests were used to consider the likely presence of structural breaks (Tables 3a, 3b and 3c). The results confirm that all the variables integrate in order one. The different tests pointed to a wide range of potential structural breaks with either positive or negative effects. A number of events (which could be potentially introduced into model using dummies) had an effect on the economy as a whole and on tourist arrivals: increases or decreases following the entry into force of several multiannual financing programmes adopted by the EU; the rise and decline of the construction sector; the 2010 natural disaster followed by recovery in 2012, and then by the worst forest fires in a century in 2016; the 2011 financial crisis, and the IMF interventions in Portugal and indirectly in Madeira, with their long-lasting effects; the Arab Spring, a favourable event for drawing a significant number of tourist arrivals to the region; and other minor events. A change in the calculation methodology of the GDP could also potentially account for some of the breaks identified below. Such a long period of time is bound to be affected by a number of such events, so what matters is to focus attention on the long-term relationship between tourism and economic growth, while acknowledging the occurrence of short-term upward or downward deviations.

Table 4 shows the results of the ARDL model. Given that the break dummies are essential for a cointegration relationship, they are included in the initial estimation. However, based on multiple preliminary tests, the conclusion was reached that only the dummies covering the multiannual financing programmes starting in 2007 and 2014 were statistically significant. It is also worth mentioning that the model was tested using preliminary tests both on levels and logs. The data interpreted in the following paragraph, uses the model with logs.

The bounds test was then performed using the null hypothesis (H₀) of no cointegration, which proposes that the coefficients of the lagged level variables in Equation 5 are jointly set to zero. The result of the bounds test leads to the rejection of H₀ and confirmation of cointegration at the 1% level (in both models). Moreover, the adjustment coefficient is statistically significant, together with the appropriate (negative) signal. The coefficient value suggests that about 5.9% of disequilibrium errors are corrected each year, from which we can infer that the convergence process to equilibrium is achievable in approximately 17 years. The short-term effect of arrivals on GDP amounts to 0.2237 (a 1% increase in receipts, leading to a 0.22% increase in GDP). Moreover, the control variables have a significant short-term effect on the level of GDP. The long-run coefficient associated with the number of arrivals is 0.809, which is statistically significant at the 1% level.

The cointegration bond test indicates that the value of the F-statistic for the ARDL model was 12.208 (8.221 in the “levels” model), which is statistically significant based on the critical values of the upper bond level of significance of 5%. This value of the F-statistic permits the rejection of H₀, indicating the existence of a long-run relationship between tourism arrivals and GDP growth.

The dummy variables used for structural breaks identified above did not lead any satisfactory conclusions in terms of statistical significance. For the reasons explained above, there are too many events to account for. In line with expectations, the impact of the economic growth dynamics in Europe is positive, as well as the impact of devaluations of the escudo and euro against a basket of European countries’ currencies (including, since 2011, against the pound sterling). An inflation rate above that recorded in the main countries of origin negatively impacts tourism dynamics. Again, in line with expectations, the sizeable reduction in the volume of funds allocated by the EU to Madeira after 2007 had a negative impact on GDP.

Given that Madeira has experienced a series of economic events with different degrees of impact, it is not surprising that the impacts of these events are not symmetric or of equal magnitude in terms of their impact on GDP. Table 6 shows the long-run coefficients for the asymmetric relationship between the independent and dependent variables based on the STATA output. This suggests that tourism growth measured by the number of arrivals positively influences economic growth at a 1% level of significance. Therefore, high tourist arrivals cause economic growth based on higher local production and sales. According to the NARDL estimates, the economic dynamics are influenced by G_NEG and G_POS, which implies that a reduction in arrivals leads to a decrease in GDP, while positive changes in the number of arrivals significantly increase GDP. Note that Table 5 confirms the presence of short-term impacts. It is worth noting that the “levels” model produces better statistical results than the “logs” version using the NARDL approach. While other specifications based on a different set of control variables produced a better fit, it was decided to keep the list of control variables employed in the ARDL model for consistency purposes. For that reason, the information provided in Table 5 is based on the model in levels.

The respective FPSS and tBDM test statistics (in Table 5) of 13.645 and 6.189 suggest rejecting the null hypotheses of no-cointegration at the 1% level. All in all, the NARDL estimates are statistically significant, given the high adjusted R-squares. The NARDL model, this was assessed using four diagnostic tests (see the end of Table 5), namely the standard diagnostic tests (serial correlation, normality, heteroscedasticity, and functional form). The tests examined in this study, with the notable exception of the normality test, can be deemed appropriate. Various solutions were attempted to address the violation of the normality assumption, such as the inclusion of dummy variables and the inclusion of other control variables. While it was possible to validate the “normality assumption” under certain circumstances, such efforts resulted in undesirable consequences in terms of the other tests. The explanation may be that there were too many outlier events, leading to residuals lying in the extremities of the normal distribution. In line with Knief and Forstmeier (2021), it was considered that “violating the normality assumption” was a lesser evil, especially given the huge number of shocks. Non-normal errors are a major issue when predicting from a linear model, but this is not the present case.

The presence or otherwise of asymmetric effects of arrivals on economic growth was investigated using WALD tests on the coefficients of the positive and negative shocks of tourism. The results of the WALD tests are provided in Table 6. Both F-tests suggested that there was asymmetry in the short-term or long-run relationships between arrivals and economic growth.

The results, in terms of the long-run effect, as shown in Table 6, suggest that when arrivals increase by 1 (perhaps due to a positive shock following a successful promotion campaign abroad, or increased security concerns in Madeira’s competitor destinations), the level of GDP increases by 0.001*10⁶€ (1000€), which is very much in line with the information provided by the Tourism Satellite Account from the Madeira Statistical Office. However, when the independent variable decreases by 1 (perhaps due a negative shock, such as the 2010 natural disaster in Madeira), the level of GDP decreases by 7000€ in the long run.

The long- and short-run Wald tests, termed WLR and WSR, reported in Table 5, lead to the acceptance of the null hypothesis of long- and short-run symmetry at the 1% level. The associated plot of the cumulative sum (CUSUM, not shown), suggests that regression coefficients are generally stable over the sample period.

When the results of the NARDL model confirm the existence of an asymmetric cointegration relationship, cumulative dynamic multipliers can be used to illustrate the asymmetric responses of economic growth during a change in the volume of arrivals. As shown in Figure 2, the asymmetric impact of arrivals on economic growth is significant in both the short term and the long run because the line marked “asymmetry” (which shows the effect regardless of whether the change is positive or negative) is significantly different to 0 over the entire period considered. It is also evident that a decrease in the number of tourist arrivals (indicated by lower dotted line) has a stronger impact on economic growth than the rise in number of arrivals, represented by the green dotted line. This confirms the main conclusions obtained from Table 6. In addition, Figure 2 depicts a relatively rapid decrease of GDP from 0 in the case of a downturn; the rate of change slows continuously until the 10th period, and stabilises by the 20th period. The impact of a reduction in the number of arrivals is constant and permanent. Increased levels of tourist arrivals do have a positive effect on GDP, but this is small in magnitude compared to the drop in demand. A rise in the number of arrivals therefore has a smaller effect on economic growth than a decrease.

Figure 2:

According to Kumar et al., in the absence of an “obvious or proven theoretical consensus [or] a priori empirical evidence on which direction the causality between the variables should hold, VAR analysis is useful.” (2022b, p.5). That is not the case in Madeira, however, as past evidence has corroborated the importance of tourism in the island context (Madjak & de Almeida, 2022). In line with the practice, therefore, Granger causality tests were carried out. To test for cointegration and to carry out Granger causality tests, it is necessary to specify the maximum lag length for analysis. The results stipulate a maximum lag length of 2, which we define as the maximum lag length for testing cointegration and estimating the long-run and short-run parameters. The causality results are shown in Table 7 with a maximum lag of 2 in the VAR models. As expected, unidirectional causality from tourism growth was found, which confirms the TLGH rather than the linear models.

This paper examines the effects of changes in the level of tourism receipts on economic growth in Madeira. It uses annual data for the period spanning 1976–2019 and employs a NARDL model to examine the short-term and long-run symmetry of both negative and positive changes in tourism arrivals. Some previous studies have used tourism arrivals (e.g., Brida et al., 2015; Kumar et al., 2020) or tourism nights (Mérida & Golpe, 2016) to proxy tourism receipts. This, however, requires the assumption of relatively constant prices, and it is possible for either of these to rise while tourism receipts fall, and vice versa. Other studies again have attempted to address this limitation by transforming annual data into quarterly data using some form of data interpolation (e.g., Kumar et al., 2022b). Such techniques may adversely affect the robustness of the findings, calling into question their conclusions.

The present study started with annual data from 1976 to 2019, which is arguably able to produce more reliable results. However, when it was not possible to identify short-run asymmetry, the annual data were transformed into quarterly figures, which were reassured and encouraged by the statistically significant results obtained with annual data.

With regard to causality, significant unidirectional causality was found to run from arrivals to GDP growth, thus corroborating the TLGH. This is in keeping with most empirical studies of the TLGH (Bilen et al., 2017; Brida et al., 2016; Nunkoo, 2020; Shahzad et al., 2017), though not all. As such, this study adds further evidence in support of the TLGH, characterising the growth of tourism as having a positive, substantial and significant effect on the size of the destination’s economy.

The results of this study indicate that both positive and negative shocks to tourism flows can significantly affect the rate of GDP growth experienced in the economy as a whole. This is in keeping with most previous studies, although there have been exceptions to this (e.g., Kumar et al., 2022b). Further analysis suggests, however, that the negative effect is significantly greater than the positive one. This result corroborates those of studies such as Kumar et al. (2022a), but is in direct contradiction with Kumar et al. (2020), which suggested that the long-run effect of positive changes in tourism was as much as three times that of the equivalent negative effect. Kumar et al. (2022b), meanwhile, found there to be an effect associated with positive changes, but no significant effect from negative changes. As such, this study adds to the body of evidence suggesting that a decline in tourism may have significantly greater implications for the destination economy than an increase. This implies that while the economy may benefit from the growth of tourism, it may also be particularly vulnerable to a reduction in tourism receipts.

Evidence of asymmetry was found with regard to the dynamics in both the short term and the long run, for positive or negative shocks. This corroborates the findings of previous studies, which have tended to suggest a nonlinear relationship (e.g., Brida et al., 2015) and hence prefer the use of NARDL (rather than simpler ARDL) models. Nevertheless, this result has important managerial implications, because the results obtained are dependent on the additions of a number of control variables.

Many empirical studies investigating the relationship between tourism and economic growth exist. As noted above, a disproportionate number have focused on small-island destinations. Many of these have been developing countries. To the authors’ knowledge, this is the first empirical study that has focused on an autonomous island region, one that is heavily dependent of EU funds and has an oversized public sector. This is important with respect to drawing out the managerial implications of this study, in that Madeira has substantial – but not unlimited – scope to responses to its findings. In short, Madeira’s scope for action is, at least to some extent, constrained within the broad policy-making context set by Portugal and, of course, the European Union, of which Portugal is a full member. An example of the effects of this dynamic is the IMF intervention in Portugal following the 2008–2011 financial crisis, which required Madeira to make a sharp reduction in its public spending. This point is made by Nunkoo et al. (2020), who recommend that more studies be carried out at the regional (sub-national) level.

This is important because, in the case of Madeira, there is evidence that negative changes in inbound tourism are likely to have a stronger effect of hindering economic growth than positive changes have in promoting it. Developing the tourism sector to attract tourists with higher incomes and willingness to spend thus represents something of a two-edged sword as a policy. The growth of tourism receipts will promote economic growth – so long as it continues – but the policy carries with it an inherent risk. Any decline in tourism, for example due to exogenous factors or loss of dynamics, can be expected to lead to a substantially larger recessive effect on the destination economy. This suggest that destination marketers need to develop ways to anticipate and mitigate the consequences of external shocks (Hernández-Martín, 2008). A programme of measures to decouple the tourism sector from the broader economy in times when tourism demand is in decline is therefore recommended.

In the meantime, Madeira also needs to harness tourism as an engine of economic growth and to cash in where demand is growing. The identification of unidirectional causality running from tourism to economic growth suggests that the region can effectively boost economic growth performance by implementing economic policies that actively promote the expansion of the tourism sector. However, as this sector has reached a mature stage in its life cycle, a different set of strategies is required, one that is not geared towards new openings of hotel establishments or infrastructural investment but rather focused on the quality of the tourism products. The sector needs to ensure that it meets international standards in terms of sustainable, smart and eco-friendly tourism that is based on circular economy principles. It also needs to attract new private investment in market niches and business models in line with the need to protect and enhance Madeira’s natural, cultural and historical resources.

This study contributes to knowledge in several important ways. First, it confirms the TLGH at the regional level. This is important because there are likely to be forces at work at the regional level that are not fully captured in national-level data. There are also likely to be policy measures available to independent states that are not available to autonomous regions, and perhaps vice versa. The findings of this study support the TLGH, which suggests that the growth of the tourism sector can serve as a long-run driver of growth in the destination economy.

Second, the use of the NARDL model allows the effects of positive and negative changes in tourism receipts to be assessed. In the case of the present study, significant effects are found in both cases, with the former being significantly larger in magnitude than the latter. This implies that a fall in arrivals will have a greater long-run effect in retarding economic growth than an equivalent increase in arrivals will have in promoting it. This recommends an ambitious-yet-cautious approach to developing the tourism economy in Madeira: one that harnesses the potential for tourism to boost the economy when tourism demand is solid or increasing, but also insulates the economy from a stronger recessive effect when tourism demand is weakening or in decline. Investing in and promoting high-end, sustainable tourism may be part of such a strategy, in that it enables the growth potential of tourism to be harnessed without necessarily creating an economic over-dependence on the sector.

This study has some shortcomings that must be duly acknowledged. First, it was not possible to incorporate control variables such as capital investment due to data limitations; these data are not available in the case of Madeira for the full span of such a long period (1978–2019). Second, it is possible that this study was not able to fully consider the short-run period. This is because other variables not recorded in the tables provided above may have not been acknowledged.

Future research should investigate the viability of complementary sectors, such as market niches in agriculture and in the tertiary sector, in order to identify which balance of growth across sectors would be most appropriate from a risk-minimisation viewpoint in order to promote overall economic growth. There could also be much to gain in terms of understanding the short-term dynamics of the adjustment process by making use of monthly data. In the present case, quarterly data series of sufficient quality were not available, which forced us to apply data conversion techniques.