Tourism and GDP: A Meta-analysis of Panel Data Studies

1 TOURISM AND GDP: A meta-analysis of panel data studies Authors Mercedes Castro-Nuñoª(*), José A. Molina-Toucedob, María P. Pablo-Romeroc a. Department of Economic Analysis and Political Economy, Avda. Ramon y Cajal, 1 University of Seville, E-41018 Seville, Spain Tel.: (+34) 954 55 44 77 Fax: (+34) 954 55 76 29 [email protected] b. Department of Economic Analysis and Political Economy, Avda. Ramon y Cajal, 1 University of Seville, E-41018 Seville, Spain Tel.: (+34) 954 55 75 29 Fax: (+34) 954 55 76 29 [email protected] c. Department of Economic Analysis and Political Economy, Avda. Ramon y Cajal, 1 University of Seville, E-41018 Seville, Spain Tel.: (+34) 954 55 76 11 Fax: (+34) 954 55 76 29 [email protected] (*) Corresponding author: Mercedes Castro-Nuño. [email protected] Acknowledgements The authors would like to express their gratitude to Dr. Richard Perdue and the anonymous reviewers for their very helpful comments. Abstract This paper provides a meta-analysis of a selected sample of 87 estimates from studies based on panel data techniques published through until 2012. The purpose is to obtain a summary measure of the effects of tourism on economic growth by applying models for both fixed and random effects. The results show a positive elasticity between GDP and tourism, although the magnitude of the effect varies according to the methodological procedure employed in the original studies for empirical estimates. In this sense, when estimates exclude other explanatory variables of economic growth, elasticities are overvalued. KEYWORDS: Meta-analysis, Panel data, Economic Growth, Elasticity. 2 1. INTRODUCTION. Tourism has attracted relatively little attention in the literature on economic development until the last decade. The main thrust of research on the economic impact of tourism in recent times has been to measure direct and secondary impacts on local and regional economies (Pearce & Butler, 2010). As pointed out in Sinclair (1998), most studies on tourism published during the second half of the 20th century, were not expressly directed towards the topic of tourism's role in economic development. In these studies, it was considered that foreign exchange earnings from tourism could be used to import capital goods to produce goods and services (McKinnon, 1964), that tourism can favor employment or additional tax revenues (Davis et al., 1988; Belisle & Hoy, 1980; Durbarry, 2002; Khan et al., 1990; Uysal & Gitelson, 1994), and that tourism can generate economic growth by enhancing efficiency through competition (Bhagwati & Srinivasan, 1979; Krueger, 1980). Tourism can also facilitate the exploitation of economies of scale at a local level (Helpman & Krugman, 1985). Theoretical models that consider a causal relationship between tourism and economic growth are a relatively recent phenomenon (Kim et al., 2006). Lanza & Pigliaru (2007) were the first to investigate this relation from an empirical point of view, while Balaguer & Cantavella- Jorda (2002) were the first to analyze the tourism-led growth hypothesis (TLG) – i.e. the hypothesis according to which tourism generates economic growth – from an econometric perspective. From the outset, an increasing number of articles with the same objective – although for different countries, using different methodologies and obtaining different results –have been published. Most of the studies, such as those by Dritsakis (2004), Durbarry (2004), Ongan & Demiroz (2005), Gunduz & Hatemi (2005), Oh (2005), Kim et al. (2006), Katircioglu (2007, 2009, 3 2010), Lee & Chien (2008), and Brida & Risso (2009), are based on time series and refer to single national economies. Other time series studies, such as those by Brida et al. (2010), Chen & Chiou-Wei (2009), Jin (2011), Lean & Tang (2010) and Arslanturk et al. (2011), for example, provide support for the TLG hypothesis. The TLG hypothesis has also been investigated using cross-sectional analyses. Among the published reports, those by Lanza & Pigliaru (2007), Singh (2008), Po & Huang (2008) and Figini & Vici (2010) are particularly relevant. Finally, a third group of studies on the TLG hypothesis is based on panel data, this being a method of studying a particular subject at multiple sites, periodically observed over a defined time frame (Hsiao, 2003). The number of studies based on panel data is smaller than that for the time-series approach, although a larger sample of countries is included, thus making these analyses more global (Lee & Chang, 2008). The TLG hypothesis has thus been analyzed by different quantitative approaches and multiple outcomes have been obtained, although no scientifically rigorous systematic review has yet been made to determine the reason for these differences. In this sense, the meta-analysis proposed in this paper can fill that knowledge gap, because it aims to bring together empirical findings to explain the variation in results and thereby yield some generalizations and directions for future research. Our objective is to apply quantitative research synthesis –metaanalysis– as a technique for comparative research, by reviewing published empirical studies that analyze the TLG hypothesis using panel data. These studies tend to be more recent and provide global estimates on the basis of large samples of countries as outlined in Lee & Chang (2008). 4 Therefore, this analysis may extract generic conclusions that are valid at an international level. Further to this, the panel data method examines the relation between GDP and tourism by considering other variables that are also essential for economic growth. The use of meta-analysis emerged in 1976 (Glass, 1976). In contrast to the traditional narrative review, the basic purpose of meta-analysis is to provide the same methodological rigor to a literature review that is required for experimental research (Rosenthal, 1995). Despite being a relatively recent technique, its use has spread to different scientific disciplines such as social and behavioral sciences (Castillo-Manzano & Castro-Nuño, 2012; Cooper & Hedges, 1994; Lipsey & Wilson, 2001; Petticrew & Roberts, 2006; Schulze, 2004) and particularly health sciences (Eddy et al., 1995; Egger et al., 2001; Sutton et al., 2000). In the case of economic growth and development studies, the meta-analysis technique has been used to integrate findings on the effects of fiscal policies (Nijkamp & Poot, 2004; Phillips & Goss, 1995), the influence of income inequality conditions or political structures (De Dominicis et al., 2008; Doucouliagos & Ulubaşoğlu, 2008), the contribution of social capital to economic growth (Westlund & Adam, 2010) and population growth (Headey & Hodge, 2009), and the effectiveness of development aid (Doucouliagos & Paldan, 2008). In the field of tourism, general applications of this technique can be found in tourism research (Dann et al., 1988), tourism forecasting (Calantone et al. 1987), and more specifically on tourist and economic impact studies (Wagner, 2002). Meta-analyses of particular importance in this field concern those performed on tourism income multipliers (Baaijens et al., 1998), regional tourism multipliers (Baaijens & Nijkamp, 2000), and tourism demand (Crouch, 1995; Lim, 1999). More recently, reports have been published concerning specific branches of 5 tourism, such as that by Carlsen & Boksberger (2011) on wine tourism, Weed (2009) on sports tourism, and Sariisik et al. (2011) on sailing tourism. This paper is divided into six sections: after this introduction, Section 2 explains the metaanalysis technique. Section 3 analyzes and classifies the studies based on panel data to corroborate the TLG hypothesis, explaining how the estimates sample has been selected and organized. Section 4 describes the outputs obtained by meta-analysis. A discussion of the results is presented in Section 5, and conclusions are provided in Section 6. 2. METHODOLOGY. Following Glass et al. (1981) and Lipsey & Wilson (2001), the meta-analysis method consists of deducing a summary effect based on the combination of different estimates (effect sizes) from a selected sample of studies by means of different statistical techniques (Chalmers et al., 2002); the most usual are the fixed-effects model (FEM) and the random-effects model (REM). Under a FEM, the selected studies are combined on the premise that there is homogeneity among them and that, according to Borenstein et al. (2009), there is one true effect size (θ) which underlies all the studies. The inference made is only conditioned by the considered studies, without taking into account the variability among them. The only determinant of the weight of each study in the meta-analysis would be its own variance (within-study variance). The technique to combine studies according to a FEM, known as the inverse variance weighted method, was described by Birge (1932) and Cochran (1937). Each effect size is involved in the summary effect in a manner that is inversely weighted by its precision (statistical weight). Assuming a sample of "m" estimates or effect sizes, (i = 1, 2… m), 6 representing a measure of the analyzed effect called Ti, a summary effect may be formulated as in (1) (Borenstein et al., 2009): [1] where wi is the statistical weight of the i-th estimate: [2], and the variance of the i-th estimate, so that: . The variance of the summary effect is formulated as: [3]. It is possible that the variability among studies is higher than that expected simply due to pure randomness, which would be detected in the first instance by testing the hypothesis of homogeneity. The most widely used test was originally developed by Cochran (1954); it calculates the parameter Q, where: [4] Because of the low power of this test, highlighted by Takkouche et al. (1999), it is recommended that a subgroup analysis of studies with similarities to each other be performed, or that additional procedures are used to quantify the possible heterogeneity. Among these alternatives, I2 is a parameter proposed by Higgins et al. (2003), which indicates the proportion of the variation between studies (between-studies variance) in the total variation; that is, the proportion of total variation due to heterogeneity: [5], where (Tau-Squared) is the between-studies variance and the within-study variance. 7 If heterogeneity is detected a REM should be used, which considers that the estimated effects of the included studies are only a random sample of all those possible. Following Borenstein et al. (2009), the true effect (θ) under a REM could vary from study to study. If it was possible to perform an infinite number of studies, the true effect sizes for them would be distributed about a mean effect which would explain two possible sources of variation: within the studies (random error) and between studies (true dispersion), so that: [6], where ei is the error when Ti estimates the true effect θi. The variance is given by expression [7]: [7], where vi is the variance due to sampling error in the i-th estimate, and the betweenstudies variance. Applying the variance weighted method, expression [2] under the REM is transformed and we have, for each i-th estimate, adjusted weights (wi* ) according to [8]: [8], where is the between-studies variance and wi the statistical weight for an i-th estimate under a FEM. With regards to the summary effect (i.e. a mean effect obtained from "a distribution of effect sizes") and its variance, from [8] we can calculate, respectively: ; [9] 8 The possibility of obtaining a biased summary effect must be assessed, which is derived from the presence of publication biases as a result of the fact that many completed studies are not actually published because they do not achieve significant effects, because they are unfavorable, or because they have negative outcomes (Sterne et al., 2000; Thornton & Lee, 2000). Analytically, the publication biases can be detected by the statistical methods of Begg and Egger, which enable testing for the null hypothesis of no bias. The method of Begg (Begg and Mazumdar Rank Correlation Test: Begg & Mazumba, 1994), suggests an inverse correlation between study size and effect size; the rank order correlation (Kendall's tau b) between the treatment effect and the standard error can be determined. Egger's Test of the Intercept (Egger, Smith, Schneider & Zinder, 1997) suggests that we assess this same bias by using precision (the inverse of the standard error) to predict the standardized effect (effect size divided by the standard error). The size of the treatment effect is captured by the slope of a regression line (B1), while bias is captured by the intercept (B0). These methods are often supplemented by so-called funnel plot diagrams, which are plots of a measure of study size on the vertical axis as a function of effect size on the horizontal axis. Ideally, the cloud of data points resembles an upside-down funnel, indicating no publication bias. A funnel plot that is asymmetric, meaning that one of the tails is missing or is markedly shorter or more thinly populated by data points, indicates the possible presence of publication bias (Borenstein et al. 2009). The main limitation of this graphical method is that the symmetry is defined subjectively by the investigator, as evidenced in the literature (Thornton & Lee, 2000; Macaskil et al., 2001). This limitation requires the application of Duval & Tweedie's Trim and Fill algorithm (Duval & Tweedie, 2000), which estimates the number of 9 missing studies and yields an effect size estimate that is adjusted for the funnel plot asymmetry. Finally, to assess the robustness or stability of the calculated summary effect, a sensitivity analysis is performed based on an iterative repetition of the meta-analysis, omitting in turn each of the studies and including others. 3. PANEL DATA STUDIES AND ESTIMATES. We have identified 13 studies published through until 2011 which use panel data to analyze the relationship between tourism and GDP (details are given in Table 1). The studies were identified by search techniques using Scopus, ScienceDirect, Google Scholar and the main journals in tourism research (see Ryan, 2005). The main search terms were: “tourism, economic growth and tourism led growth hypothesis”. References from other studies were also used, which include not only articles in scientific journals listed in Journal Citation Reports (JCR) or other databases, but also working papers (Wpaper) that have achieved a certain scientific recognition on account of their quality or number of citations. All the studies included are shown with an identification code in Table 1. The last column shows the estimates in each study which may differ depending on the estimation model: whether additional variables were used to explain economic growth, the type of variable used to measure tourism, whether the sample was classified into subsamples and the inclusion or not of instrumental or dummy variables for econometric estimates. [PLEASE INSERT TABLE 1] 16 although they reach lower random point estimates. Even though the non-dynamic scenarios involved lower estimates, they showed a remarkable dispersion or heterogeneity and higher summary random point estimates. Most of the dynamic scenarios in Table 3, exhibited a publication bias, while non-dynamic scenarios did not in general. The re-adjusted random point estimate in the dynamic models has a value of 0.00063 for the overall sample when all estimates are considered. In the other scenarios of the overall sample, the re-adjusted random point estimate is between 0.01238 (when travel income is taken as a proxy for tourism) and 0.00039 (when tourism arrivals are taken as the proxy). As shown in Table 3, the random point estimate for non-dynamic models has a value of 0.25800 for the overall sample when all estimates are considered. In the other scenarios of the overall sample, the estimate fluctuates around the same value, with a minimum value of 0.191 and a maximum value of 0.344. With the exception of the estimates that use arrivals as a tourism indicator, none of the estimated values are biased. In this case, the re-adjusted random point estimate is 0.20182. The random point estimates for non-dynamic models are therefore much higher that those obtained for dynamic models. Nevertheless, one must keep in mind that the small value obtained for dynamic models only reflects the effect of tourism on economic growth for the current period. However, the dynamic nature of the model really means that this effect may persist in time. In order to know what is the full impact of tourism on GDP in these dynamic models, or to determine a representative value of the total tourism-productivity elasticity over time, the value of the long-term cumulative dynamic multiplier (CDM) of tourism must be calculated. 17 For this, it is necessary to determine the estimated value of the parameter  that relates the current period productivity to the productivity of previous periods. However, to calculate this cumulative effect, the estimated function should be stable. This occurs when  in Equation (10) is less than unity. Otherwise, the trajectory is divergent and the effect tends to multiply with time. In the case where the dynamic functions are stable with a single time delay, the long-term CDM is equal to:    1 CDM [12] The value of this cumulative multiplier is similar to the concept of the tourism-productivity elasticity of non-dynamic studies, which can help to interpret the effect of tourism on economic growth. In Table 4, estimated values of  and  are given along with the value of the CDM calculated when the functions are stable. Parameters whose estimates proved to be not significant in the original studies have been omitted from Table 4, meaning that all estimates included in Table 4 are of type A. In the last column, it can be seen that the CDM value is substantially higher than the estimated value of  . The last row of Table 4 shows that the average value of the CDM, which summarizes this effect, is 0.0488. An interval of values for the long term CDM has been calculated, which depends on the diverse specifications of estimates. Extreme values of the re-adjusted random point estimate 18 obtained for dynamic models have been used. The minimum value of the CDM is calculated from the coefficient  of the estimates using tourism arrivals and  equal to 0.00039. The maximum value is calculated from the coefficient  of the estimates using travel income and  equal to 0.01238. This gives a range between 0.002 and 0.11. It can be seen that the maximum value of long-term CDM (0.11) is significantly lower than the re-adjusted random point estimate in non-dynamic models (0.25), even at the latter’s minimum value (0.191). It should be noted here that the choice of the type of panel model used may directly impact on the validity of the results. If the data generating process is bestdescribed as a dynamic panel, then the result from a non-dynamic panel would suffer from biases that come with misspecifications. Therefore, the results of the non-dynamic models should be considered with care. This difference in values may also be explained when the dynamic and non-dynamic scenarios are divided into two subgroups (A and B). In the dynamic model scenarios, 38 estimates are of type A and only 2 are of type B, while in the non-dynamic model scenarios, 23 estimates are of type B and only 2 are of type A. In both cases, it is noted that the random point estimates, or the re-adjusted values of the estimates type A, are lower than those obtained in each overall scenario (dynamic overall scenario or non-dynamic overall scenario). On the other hand, the random point estimates of the estimates type B are higher than those obtained in each overall scenario. That is, when additional variables that explain economic growth in the production function are included, in addition to the tourism variable, the impact of tourism on GDP decreases. In this sense, it can be argued that when estimates exclude other explanatory variables of economic growth, 19 elasticities are overvalued. As pointed out in Cortés-Jiménez and Pulina (2010), the variables which represent a decisive contribution to growth should not be excluded from the analysis if we aim to adequately show the way in which tourism also contributes to it. Nevertheless, this conclusion must be considered with care since there are only two estimates type B for dynamic models and two estimates type A for non-dynamic models. This sub-disaggregation of estimates into A and B does not seem to be sufficient to eliminate the heterogeneity among them, except for type A non-dynamic scenarios. This suggests that other circumstances exist that also affect the value of the random point estimates. Therefore, other classification criteria based on methodological aspects (discussed in Section 3) have been considered and applied to the overall sample for dynamic and non-dynamic scenarios, as well as to their respective A and B subgroups. It was found that the inclusion of temporal variables and the use of instrumental variables tended to decrease the random point estimates value in all of the type 2 scenarios, and that the random point estimate values obtained for these scenarios were higher when travel income was used to measure tourism than when the number of arrivals was used. Furthermore, the random point estimate tended to be slightly greater when the estimates that consider only specific samples of countries were used in dynamic models. This sample includes groups of countries considered to be poor, small, specialized in tourism, or from a specific geographic zone. The literature considers that tourism in such groups of countries has a considerably higher impact on economic growth. However, this small difference does not seem to reaffirm that the major effect is true in all groups. In this regard, some studies consider that country size does not affect the relationship between tourism and economic 20 growth (Lanza et al., 2003; Sequeira & Nunes, 2008), but the degree of specialisation in tourism is indeed quite relevant (Adamau & Clerides, 2010; Holzner, 2011; Narayan et al. 2010; Sequeira & Campos, 2005; Sequeira & Nunes, 2008). It also seems that there is evidence in favour of tourism fostering growth to a greater degree in countries with a lower income level (Eugenio-Martín et al. 2004; Lee & Chang, 2008; Seetenah, 2011). 6. CONCLUSIONS. Theoretical models that consider a causal relationship between economic growth and tourism (namely the TLG hypothesis) are a recent phenomenon. This paper performs a quantitative systematic review (meta-analysis) that synthesizes the findings from the previous panel data empirical evidence. According to the meta-analysis presented, we can conclude that tourism contributes to economic growth, thereby corroborating the initial hypothesis and suggesting that economic efforts to promote tourism will have a positive effect on the country's overall economic activity. Nevertheless, the magnitude of the effect was found to vary according to the methodological procedure employed in the original studies using empirical estimates. In general, we find that the value of these elasticities is affected by a range of features used in the estimates carried out. We deduce that as a model becomes more specific, the value of elasticity (productivity with respect to tourism) tends to decrease. Thus, the inclusion of explanatory variables for economic growth, in addition to that of tourism, tends to reduce the value of the elasticity. Also, when temporal variables and instrumental variables are considered, the value of the elasticity tends to diminish.The estimates also depend on the proxy used to measure tourism. When the proxy is travel income, the elasticity is higher than when the number of arrivals is used. 21 We consider essential that future research includes economic growth explanatory variables in estimates carried out and that functions to be estimated, including temporal and instrumental variables, are adequately specified. Also, it may be appropriate that estimations use both proxies of tourism – arrivals and travel income – to enable a range of values to be specified concerning tourism’s effect on GDP. If available data are insufficient to calculate these proxies, other variables of a similar magnitude may be used. For example, the sub-items of balance of payments related to tourism as an alternative to the magnitude of travel income. The sample estimates can be clearly divided into two subsamples. Those based on dynamic models and those based on non-dynamic models. The meta-analysis applied to estimates based on dynamic functions shows that elasticity in the short-term is small, yielding a readjusted random point estimate ranging between 0.0004 and 0.0123. The initial effect is extended in time, so that in the long-term the average value of the elasticity is in a range between 0.002 and 0.110. The value of this long-term CDM is obtained from estimates that include variables for economic growth in addition to that of tourism (type A). Hence, although the initial effect of the tourism on GDP is small, it is increased over time. Therefore, policy makers and the tourism industry should plan long-term actions, whether public, private or mixed, in order to optimize the resources invested. The meta-analysis applied to estimates based on non-dynamic functions shows that the elasticities had an average value of 0.266 for the overall sample. The results of the nondynamic models should be considered with care because the data generating process is bestdescribed as a dynamic panel. Also, this value may be overestimated because the majority of estimates were type B for non-dynamic scenarios, which do not include other explanatory 22 variables of economic growth. The random point estimate for type A non-dynamic scenarios was only 0.038, which is within the range of values for the long-term CDM. Finally, the meta-analysis results show that the tourism effect tends to be slightly greater for specific samples of countries. However, there were insufficient panel data studies relating to groups of similar countries for this result to be more precisely stated. Thus, future research focused on specific groups of countries (specialized in tourism, from specific geographic regions, with a similar income level…) would likely extend our findings. REFERENCES Adamou, A., Cleride, S. (2010). “Prospects and limits of tourism-led growth: The international evidence”. Review of Economic Analysis, 3, 287–303. Arslanturk, Y., Balcilar, M., Ozdemir, Z. A. (2011). “Time-varying linkages between tourism receipts and economic growth in a small open economy”. Economic Modelling, 28, 664– 671. Baaijens, S.R., Nijkamp, P., Van Montfort, K. (1998). “Explanatory meta-analysis for the comparison and transfer of regional tourist income multipliers”. Regional Studies, 32, 839-849. Baaijens, S.R., Nijkamp, P. (2000). “Meta-analytic methods for comparative and exploratory policy research: An application to the assessment of regional tourist multipliers”. Journal of Policy Modeling, 22, 821–858. Balaguer, J., Cantavella-Jordà, M. (2002). “Tourism as a long-run economic growth factor: the Spanish case”. Applied Economics, 34, 877-884. Begg, C.B., Mazumbar, M. (1994). “Operating characteristics of a rank correlation test for publication bias”. Biometrics, 50, 1088-1101. 23 Belisle, F., Hoy, D. (1980). “The perceived impact of tourism by residents”. Annals of Tourism Research, 8, 83-97. Bhagwati, J., Srinivasan, T. (1979). “Trade policy and development”. In International Economic Policy: Theory and Evidence, edited by Dornbunsch, R., & Frenkel, J., Baltimore: Johns Hopkins University Press. Pp. 1-35. Birge, R.T. (1932). “The calculation of errors by the method of least squares”. Physical Review, 40, 207–227. Borenstein, M., Hedges, L., Higgins, J.P.T., Rothstein, H.R. (2009). Introduction to metaanalysis. Chichester, UK: Wiley. Brida, J.G., Lanzilotta, B., Lionetti, S., Risso, W.A. (2010). “The tourism-led growth hypothesis for Uruguay”. Tourism Economics, 16 (3), 765–771. Brida, J.G., Risso, W. A. (2009). “Tourism as a factor of long-run economic growth: An empirical analysis for Chile”. European Journal of Tourism Research, 2(2), 178-185. Calantone, R.J., Di Benedetto, C.A., Bojanic, D. (1987). “A comprehensive review of the tourism forecasting literature”. Journal of Travel Research, 26, 28-39. Carlsen, J., Boksberger, P. (2011). “The blending of wine and tourism experiences: A metaanalysis”. In CAUTHE 2011 National conference tourism: creating a brilliant blend, edited by M.J. Gross. Adelaide, S.A.: University of South Australia. School of Management. Pp. 983-989. Castillo-Manzano, J.I., Castro-Nuño, M. (2012). “Driving licenses based on points system: an efficient road safety strategy or a global fashion transport policy? a worldwide metaanalysis”. Transport Policy, 21, 191-201. Chalmers, T.C., Hedges, L.V., Cooper, H. (2002). “A brief history of research synthesis”. Evaluation & the Health Professions, 25, 12-37. 24 Chen, C.-F., Chiou-Wei, S. Z. (2009). “Tourism expansion, tourism uncertainty and economic growth: New evidence from Taiwan and Korea”. Tourism Management, 30, 812–818. Cochran, W. G. (1937). “Problems arising in the analysis of a series of similar experiments”. Journal of the Royal Statistical Society, Supplementary, 4, 102-118. Cochran, W. G. (1954). “The combination of estimates from different experiments”. Biometrics, 10, 101-129. Cooper, H., Hedges, L.V. (1994). The handbook of research synthesis. New York: Russell, Sage. Cortés-Jiménez, I. (2010). “Which type of tourism matters to the regional economic growth ? The cases of Spain and Italy”. International Journal of Tourism Research, 10, 12-139. Cortés-Jiménez, I., Pulina, M. (2010). “Inbound tourism and long-run economic growth“. Current Issues in Tourism, 13, 61-74. Crouch, G. I. (1995). “A meta-analysis of tourism demand”. Annals of Tourism Research, 22, 103-118. Dann, G., Nash, D., Pearce, P. (1988). “Methodology in tourism research”. Annals of Tourism Research, 15, 1-28. Davis, D., Allen, J., Consenza, R. M. (1988). “Segmenting local residents by their attitudes, interests, and opinions toward tourism”. Journal of Travel Research, 27, 2-8. De Dominicis, L., Florax, R.J., De Groot, H.L. (2008). “A meta-analysis on the relationship between income inequality and economic growth”. Scottish Journal of Political Economy, 55, 654-682. Doucouliagos, H., Ulubaşoğlu, M.A. (2008). “Democracy and economic growth: A metaanalysis”. American Journal of Political Science, 52, 61–83. Doucouliagos, H., Paldam, M. (2008). “Aid effectiveness on growth: A meta study”. European Journal of Political Economy, 24, 1-24. 25 Dritsakis, N. (2004). “Tourism as a long-run economic growth factor: an empirical investigation for Greece using causality analysis”. Tourism Economics, 10, 305-316. Dritsakis, N. (2011). “Tourism Development and Economic Growth in Seven Mediterranean Countries: A Panel Data Approach”. Tourism Economics, forthcoming. Durbarry, R. (2002). “The economic contribution of tourism in Mauritius”. Annals of Tourism Research, 29, 862-65. Durbarry, R. (2004). “Tourism and economic growth: The case of Mauritius”. Tourism Economics, 10, 389-401. Duval, S., Tweedie, R. (2000). “Trim and fill: a simple funnel plot based method of testing and adjusting for publication bias in meta-analysis”. Biometrics, 56, 455-463. Eddy, D.M., Hasselblad, V., Shachter, R. (1995). Meta-analysis by the confidence profile methods: The statistical synthesis of evidence. Boston, MA: Academic Press. Egger, M., Smith, G.D., Schneider, M., Zinder, Ch. (1997). “Bias in meta-analysis detected by a simple, graphical test”. British Medical Journal, 31, 629-634. Egger, M., Smith, G.D., Altman, D.G. (2001). Systematic reviews in health care: Metaanalysis in context, (2nd ed.). London: BMJ Books. Eugenio-Martín, J. L., Morales, N. M., Scarpa, R. (2004). Tourism and economic growth in Latin American countries: a panel data approach. Fondazione Eni Enrico Mattei Working Paper Series, Nota di Lavoro 26. Fayissa B., Nsiah C., Tadasse B. (2008). “Impact of tourism on economic growth and development in Africa”. Tourism Economics, 14, 807–818. Fayissa B., Nsiah C., Tadasse B. (2009). Tourism and economic growth in Latin American countries (LAC): Further empirical evidence. Working Papers 200902, Middle Tennessee State University, Department of Economics and Finance. 32 Table 2. Tourism led growth: characteristics of the panel data estimates. Code Coefficient Student’s t-test or Std Error* p-value A B Dynamic Model Time dummies Instrumental variables Proxy of tourism expansion Sample of Countries yes no yes no yes no Tourism receipts Arrivals Overall Specific Adam 1 0.002 0.00056x 5% x x x x x x Adam 2 0.011 0.0007 x no x x x x x x Adam 3 0.00018 0.000045 x 5% x x x x x x Adam 4 0.00012 0.00006 x 10% x x x x x x Adam 5 0.00117 0.00071 x no x x x x x x Adam 6 0.0041 0.00096 x 5% x x x x x x Adam 7 0.0039 0.0012 x 5% x x x x x x Adam 8 0.00048 0.00012 x 5% x x x x x x Adam 9 0.00021 0.00015x no x x x x x x Adam 10 0.0027 0.0012 x 10% x x x x x x Cort 1 0.001 5% x x x x x 24 Italian & Spanish coastal regions Cort 2 0.006 1% x x x x x Cort 3 0.001 5% x x x x x Cort 4 0.006 1% x x x x x Cort 5 -0.001 no x x x x x 13 Italian & Spanish interior regions Cort 6 -0.015 10% x x x x x Cort 7 -0.001 no x x x x x Cort 8 -0.017 5% x x x x x Cort 9 0.001 10% x x x x x 14 Italian & Spanish Mediterranean coastal regions Cort 10 0.007 1% x x x x x Cort 11 0.001 5% x x x x x Cort 12 0.006 1% x x x x x Drit 1 1.235 1% x x x x x 7 Mediterranean countries Drit 2 0.077 10% x x x x x 33 Eug 1 0.00036 1.68 10% x x x x x 21 Latin American countries Eug 2 -0.00019 2.54 10% x x x x x 7 Latin American high-income countries Eug 3 0.00063 1.92 10% x x x x x 11 Latin American mediumincome countries Eug 4 0.00062 2.63 10% x x x x x 3 Latin American lowincome countries Fa y i 1 0.042 0.0082 x 1% x x x 17 Latin American countries Fa y i 2 0.021 0.0103 x 5% x x x Fa y i 3 0.0243 0.0071 x 1% x x x Fa y i 4 0.0266 0.0082 x 1% x x x Fa y i 1 0.0378 0.0085 x 1% x x x x x 42 sub-Saharan countries Fa y i 2 0.0388 0.01 x 1% x x x x x Fa y i 3 0.0249 0.0081 x 1% x x x x x 30 sub-Saharan countries Fa y i 4 0.0256 0.0081 x 1% x x x x x Holz 1 0.011 2 5% x x x x x x Holz 2 0.018 2.84 1% x x x x x x Holz 3 0.008 2.08 5% x x x x x x Holz 4 -0.041 -0.97 no x x x x x x Lee 1 0.36 11.84 5% x x x x x 23 OECD countries Lee 2 0.17 14.36 5% x x x x x Lee 3 0.24 26.41 5% x x x x x Lee 4 0.13 13.47 5% x x x x x Lee 5 0,5 52,69 5% x x x x x 32 OECD countries Lee 6 0.5 35.21 5% x x x x x Lee 7 0.61 5.38 5% x x x x x Lee 8 0.17 -0.3 no x x x x x Lee 9 0.13 -0.62 no x x x x x 5 Asian countries 34 Lee 10 0.17 8.98 5% x x x x x Lee 11 0.32 4.06 5% x x x x x Lee 12 0.24 6.89 5% x x x x x Lee 13 0.15 1.1 no x x x x x 11 Latin American countries Lee 14 0.09 8.9 5% x x x x x Lee 15 0.36 16.36 5% x x x x x Lee 16 0.23 12.08 5% x x x x x Lee 17 0.3 8.23 5% x x x x x 16 sub-Saharian countries Lee 18 0.18 6.75 5% x x x x x Lee 19 0.03 2.44 5% x x x x x Lee 20 0.08 0.83 no x x x x x N ara 1 0.24 3.84 1% x x x x x 4 islands N ara 2 0.72 15.29 1% x x x x x Seet 1 0.12 1.95 10% x x x x x 19 Pacific islands Seet 2 0.06 1.95 10% x x x x x 20 developing countries Seet 3 0.064 1.96 10% x x x x x 10 developing countries Seet 4 0.14 2.04 10% x x x x x 19 Pacific islands Seet 5 0.033 1.87 10% x x x x x 20 developing countries Seet 6 0.08 1.89 10% x x x x x 10 developing countries Sequ1 -14.21 -1.37 no x x x x Sequ2 -0.62 -1.78 10% x x x x Sequ3 -1.89 -1.29 no x x x x Sequ4 1.44 0.39 no x x x x Sequ5 0.329 2.22 5% x x x x Sequ6 1.104 2.26 5% x x x x Sequ 1 0.013 1.05 no x x x x x x Sequ 2 0.041 2.42 5% x x x x x x 35 Sequ 3 0.026 1.92 10% x x x x x x Sequ 4 0.025 1.85 10% x x x x x x Sequ 5 0.048 3.77 5% x x x x x x Sequ 6 0.041 2.69 5% x x x x x x Sequ 7 0.095 4.44 1% x x x x x x Sequ 8 0.049 3.89 1% x x x x x x Sequ 9 0.042 2.48 1% x x x x x x Sequ 10 0.106 4.14 1% x x x x x x Sequ 11 0.047 3.32 1% x x x x x x Sequ 12 0.041 2.98 1% x x x x x x Sequ 13 0.077 2.5 1% x x x x x 47 small countries Sequ 14 0.06 1.79 10% x x x x x Sequ 15 0.055 1.87 10% x x x x x 45 small countries Sequ 16 0.092 3.45 1% x x x x x 58 poor countries Sequ 17 0.063 3.28 1% x x x x x Sequ 18 0.055 2.73 1% x x x x x Sou 1 0.1006 1.55 no x x x x 7 NUT II Portuguese regions Sou 2 0.1082 1.7 5% x x x x Sou 3 0.1522 2.05 10% x x x x Sou 4 0.0401 1.54 no x x x x 30 NUT II Portuguese regions Sou 5 0.0056 2.31 10% x x x x Sou 6 0.009 2.55 10% x x x x A and B are referred to estimates which include or not adding explanatory variables of GDP. * All data for this column are t-student except for those marked with an x, which signifies std. error according to the original format from each study. Source: Own elaboration 36 Table 3: Meta-Analysis Results Type 1 SCENARIOS Type 2 SCENARIOS Type 3 SCENARIOS No. of Point Estimates FIXED POINT ESTIMATE / Z-value RANDOM POINT ESTIMATE / Z-value HETEROGENEITY MEASURES PUBLICATION BIAS Duval & Tweedie’s Trim and Fill Q test (ρ) I2 Begg (Kendall's tau b) Egger (B0) No. of missed studies Readjusted Random Point Estimate DYNAMIC OVERALL overall 44 0.000 (4.902)*** 0.00089 (2.818)*** 0.000*** 47.016 (-0.241) 0.833 16 0.00063 t 22 0.000 (4.706)*** 0.0059 (2.055)** 0.000*** 57.897 0.112 1.041 10 0.00044 no t 22 0.00742 (4.738)*** 0.00742 (4.738)*** 0.968 0.000 (-0.141) 0.279 0 - inst 34 0.000 (4.813)*** 0.00077 (2.449)** 0.000*** 52.562 (-0.327) 0.888 11 0.00073 no inst 10 0.00576 (2.816)*** 0.00576 (2.816)*** 0.874 0.000 0.667 0.3505 0 - travel income 15 0.01238 (5.139)*** 0.01238 (5.139)*** 0.639 0.000 0.190 0.301 1 0.01238 arrivals 23 0.000 (4.591)*** 0.00031 (4.591)*** 0.710 0.000 (-0.145) 0.387 5 0.00039 specific countries 30 0.000 (4.775)*** 0.0065 (2.339)*** 0.002*** 48.261 (-0.311) 0.971 12 0.00071 A overall 38 0.000 (4.732)*** 0.00055 (2.379)** 0.031** 32.272 (-0.263) 0.641 13 0.00044 t 22 0.000 (4.706)*** 0.00059 (2.055)** 0.000*** 57.897 0.112 1.048 10 0.00044 no t 16 0.01047 (2.036)*** 0.01047 (2.036)*** 1.000 0.000 (-0.284) (-0.136) 0 - inst 32 0.000 (4.731)*** 0.0059 (2.242)** 0.006*** 43.240 (-0.356) 0.765 10 0.00052 no ins t 6 0.001 0.00128 1.000 0.000 0.667 0.1448 2 0.00076 37 (0.099) (0.099) travel income 15 0.01238 (5.139)*** 0.01238 (5.139)*** 0.639 0.000 0.190 0.301 1 0.01238 arrivals 23 0.000 (4.591)*** 0.00031 (4.591)*** 0.710 0.000 (-0.144) 0.387 5 0.00039 specific countries 24 0.000 (4.604)*** 0.00043 (2.650)*** 0.165 21.987 (-0.371) 0.690 8 0.00044 B (INST) overall 2 0.009 (3.406)*** 0.063 (0.911) 0.042** 75.819 - - - - no t 2 0.009 (3.406)*** 0.063 (0.911) 0.042** 75.819 - - - - specific countries 2 0.009 (3.406)*** 0.063 (0.911) 0.042** 75.819 - - - - Table 3: Meta-Analysis Results (Cont…) Type 1 SCENARIOS Type 2 SCENARIOS Type 3 SCENARIOS No. of Point Estimates FIXED POINT ESTIMATE / Z-value RANDOM POINT ESTIMATE / Z-value HETEROGENEITY MEASURES PUBLICATION BIAS Duval & Tweedie´s Trim and Fill Q test (ρ) I2 Begg (Kendall's tau b) Egger (B0) No. of missed studies Re-adjusted Random Point Estimate NO DYNAMIC OVERALL overall 25 0.266 (126.649)*** 0.258 (6.754)*** 0.000*** 99.618 0.187 (-0.775) 0 - t 10 0.198 (68.873)*** 0.208 (4.036)*** 0.000*** 99.592 0.356 2.662 0 - no t 15 0.344 (111.753)*** 0.290 (5.375)*** 0.000*** 99.515 0.000 (-2.606) 0 - no inst 25 0.266 (126.649)*** 0.258 (6.754)*** 0.000*** 99.618 0.186 (-0.775) 0 - travel income 14 0.321 (116.239)*** 0.2783 (4.563)*** 0.000*** 99.718 0.121 (-3.446) 0 - arrivals 11 0.191 (58.894)*** 0.309 (6.773)*** 0.000*** 98.631 (-0.073) 1.918 1 0.202 38 specific countries 25 0.266 (126.649)*** 0.258 (6.754)*** 0.000*** 99.618 0.186 (-0.775) 0 - A overall 2 0.038 (1.362) 0.038 (1.362) 0.9860 0.000 - - - - no t 2 0.038 (1.362) 0.038 (1.362) 0.9860 0.000 - - - - no inst 2 0.038 (1.362) 0.038 (1.362) 0.9860 0.000 - - - - travel income 2 0.038 (1.362) 0.038 (1.362) 0.9860 0.000 - - - - specific countries 2 0.038 (1.362) 0.038 (1.362) 0.9860 0.000 - - - - B overall 23 0.267 (126.902)*** 0.279 (6.993)*** 0.000*** 99.646 0.221 0.082 0 - t 10 0.198 (68.873)*** 0.208 (4.036)*** 0.000*** 99.592 0.355 2.662 0 - no t 13 0.344 (112.277)*** 0.290 (5.762)*** 0.000*** 99.150 (-0.012) (-1.251) 0 - no inst 23 0.267 (126.902)*** 0.279 (6.993)*** 0.000*** 99.646 0.221 0.082 0 - travel income 12 0.324 (116.669)*** 0.313 (4.872)*** 0.000*** 99.756 0.167 (-1.848) 0 - arrivals 11 0.191 (58.894)*** 0.240 (6.773)*** 0.000*** 98.631 (-0.072) 1.917 1 0.202 specific countries 23 0.267 (126.902)*** 0.279 (6.993)*** 0.000*** 99.646 0.221 0.082 0 - Note: Significance at ***1%, **5%, *10%, respectively. Overall is related to all the estimates; each scenario A or B is respectively related to estimates with / without added exogenous variables of the economic growth; t and no t mean estimates with /without time dummies; Inst and no inst are respectively related to estimates with / without instrumental variables; Travel income and arrivals refer to estimates based on proxy variables of tourism obtained from data series of travel income or tourist arrivals; specific countries are related to estimates obtained for a concrete sample of countries. Source: Own elaboration. 39 Table 4. Long-Term cumulative dynamic multipliers of tourism on economic growth. CODE Model type   Cumulative Dynamic Multiplier (CDM) Adam 1 A 0.002** (0.00056) -0.1** (0.0077) 0.02ª Adam 3 A 0.00018** (0.00045) -0.01** (0.0073) 0.018ª Adam 4 A 0.00012* (0.00006) -0.09** (0.017) 0.001ª Adam 6 A 0.0041** (0.00096) -0.101** (0.0076) 0.041ª Adam 7 A 0.0039** (0.0012) -0.101** (0.018) 0.038ª Adam 8 A 0.00048** (0.00012) -0.1** (0.0074) 0.004ª Holz 1 A 0.011** (2.00) 0.941*** (35.98) 0.186 Holz 2 A 0.018*** (2.84) 0.95*** (35.49) 0.360 Holz 3 A 0.008** (2.08) 0.97*** (52.93) 0.267 Cort 1 A 0.001** (n.a) 0.895*** (n.a) 0.010 Cort 2 A 0.006* (n.a) 0.884 *** (n.a) 0.052 Cort 3 A 0.001** (n.a) 0.919*** (n.a) 0.012 Cort 4 A 0.006*** (n.a) 0.907*** (n.a) 0.065 Cort 6 A -0.015 (n.a) 0.891*** (n.a) -0.138 Cort 8 A -0.017** (n.a) 0.942*** (n.a) -0.293 Cort 9 A 0.001* (n.a) 0.831*** (n.a) 0.006 Cort 10 A 0.007*** (n.a) 0.830*** (n.a) 0.041 Cort 11 A 0.001** (n.a) 0.869*** (n.a) 0.008 Cort 12 A 0.006*** (n.a) 0.857*** (n.a) 0.042 Eug 1 A 0.00036* (1.68) 0.777* (19.30) 0.007 Eug 2 A -0.0002* (2.54) 0.765* (12.64) -0.001 Eu g 3 A 0.00063* 0.738* 0.002 40 (1.92) (10.16) Eug 4 A 0.00062* (2.63) 0.597* (4.14) 0.002 Fayi 3 A 0.0249*** (0.0081) 0.568*** (0.073) 0.058 Seet 1 A 0.12* (1,95) 0.24** (215) 0.158 Seet 2 A 0.06* (1.95) 0.23*** (2.52) 0.078 Seet 3 A 0.064* (1.96) 0.34*** (2.43) 0.097 Seet 4 A 0.14* (2.04) 0.17* (2.17) 0.169 Seet 5 A 0.033* (1.87) 0.25** (2.15) 0.044 Seet 6 A 0.08* (1.89) 0.37** (2.19) 0.127 Average value of DM - - - 0.0488 0.0527 Note: Significance at ***1%, **5%, *10%, respectively. n.a. Not available. a. The estimated function is ttt tyy    1 . So ttt tyy    1 )1( and   /MD Source: Own elaboration.