BOOTSTRAP PANEL CAUSALITY TESTING OF TOURISM AND GDP NEXUS FOR UPPER-MIDDLE INCOME COUNTRIES

Ulkelerin gelirlerini artiran kaynaklardan biri de turizmdir. Bircok calisma cesitli ulkelerde ve ulke gruplarinda GSYIH ile turizm arasinda nedensel iliskiler oldugunu gostermektedir. Bu cercevede, mevcut calismanin amaci ust-orta gelir grubundaki 33 ulke icin turizm ve GSYIH arasindaki nedensellik iliskisini incelemektir. Bu amaca yonelik olarak 1995-2018 donemini kapsayan yillik verilere Emirmahmutoglu ve Kose (2011) tarafindan onerilen bootstrap nedensellik testi uygulanmistir. Bu test hem yatay kesit bagimliligi ve bagimsizligi durumlarinda, hem de gozlem sayisi dusukken uygulanabilen guclu bir testtir. Sonuclar Grenada ve Guney Afrika icin turizmden GSYIH’ye tek yonlu bir nedensellik; Arnavutluk, Bulgaristan, Dominik Cumhuriyeti, Ekvador, Jamaika, Peru ve Sri Lanka icin GSYIH’den turizme tek yonlu bir nedensellik; Ermenistan, Mauritius ve Kuzey Makedonya icin turizm ve GSYIH arasinda cift yonlu nedensellik oldugunu; ancak geriye kalan 21 ulke icin nedenselligin olmadigini gostermistir.


INTRODUCTION
The development of the tourism sector has benefited directly and indirectly from many channels on economic performance. It is stated that the development in the tourism sector contributes to GDP, employment, investment, foreign exchange earnings, as well as causing a significant socio-cultural and environmental development. Therefore, many governments are striving to stimulate economic growth with the tourism sector (Archer, 1995;Balaguer and Cantavella-Jorda, 2002;Dritsakis, 2004;Durbarry, 2002;Nowak, Sahli and Cortés-Jiménez, 2007;Lee and Chang, 2008;Payne and Mervar, 2010).
Among great numbers of studies on tourism economics, investigation of tourism and economic growth relationship is paid special importance. After the initial study of Lanza and Pigliaru (2000), studies examining the association between tourism and growth have increased. One of the popular topics is to examine the causality between tourism and economic growth. This causal association is classified by the way of causality into four hypotheses: (1) Tourism-led growth hypothesis: Unidirectional (one-way) causality from tourism to economic growth, (2) Conservation (growth-led tourism) hypothesis: Unidirectional (one-way) causality from economic growth to tourism, (3) Feedback hypothesis: Bidirectional (two-way) causality between tourism and economic growth, and (4) Neutrality hypothesis: No causality.
The goal of this work is to explore causal relationship between tourism receipts and GDP for 33 upper-middle income countries using yearly data from 1995 to 2018. Our study differs from previous works in two aspects. First, there is no work, as far as we know, especially focuses on the causal relationship between tourism and GDP for upper-middle income countries. Therefore, there is no clear information on the way of causality. Tourism-led growth hypothesis is supported in many papers in the literature. We have tested whether this applies to the uppermiddle income countries as well. Second, most of other studies use conventional causality tests which take no account of cross section dependency. Here, we employ novel causality approach suggested by Emirmahmutoğlu and Köse (2011). Causality tests of Granger (1969) and Sims (1972) require stationary series. However, series can be integrated or cointegrated at different levels. Toda and Yamamoto (1995) present a modified-Wald statistic to overcome such problems. Since the series used in this paper are integrated at different levels and have cross-section dependency, the approach of Emirmahmutoğlu and Köse (2011) which is built upon Toda and Yamamoto's (1995) procedure is applied. The outcomes of causality tests show that tourism causes GDP in Grenada and South Africa; GDP causes tourism in Albania, Bulgaria, Dominic Republic, Ecuador, Jamaica, Peru, Sri Lanka and the whole panel; tourism and GDP mutually cause each other in Armenia, Mauritius and North Macedonia; and no causality in Azerbaijan, Botswana, Brazil, China, Colombia, Costa Rica, Dominica, Fiji, Guatemala, Guyana, Jordan, Malaysia, Mexico, Paraguay, Romania, Russian Federation, Samoa, St. Lucia, St. Vincent and The Grenadines, Thailand and Turkey. The remainder is organized as follows: Section 2 reviews the related literature, Section 3 presents data, model and methodology, Section 4 shows empirical findings, and the last section concludes.

RELATED LITERATURE
There are lots of studies focus on tourism and economic growth nexus. A brief summary of the works is given in Table 1. Among these studies, Balaguer and Cantavella-Jordá (2002) for Spain; Gündüz and Hatemi-J (2005) for Turkey; Lee and Chang (2008)  Oh (2005)  Studies that support feedback hypothesis are Dritsakis (2004) for Greece; Lee and Chang (2008) for non-OECD countries; Aslan (2013) (2014) for African countries (tourism variable changes) come across the findings that confirm the neutrality hypothesis.

Data
The annual data consist of GDP (constant 2010 US$), tourism receipts (constant 2010 US$) for 33 upper-middle income countries 3 spanning 1995-2018. Tourism receipts converted to constant 2010 US$ using consumer price index (2010=100). The series are sourced from World Bank's (2020) World Development Indicators. Each variable is employed in its natural log and abbreviated as ln for natural log of GDP and ln for natural log of tourism receipts. Descriptive statistics are given in Table 2. There are 792 observations for each variable. For 33 middle-income countries in 1995-2018, average real GDP is about 359 billion US$ (with 1 trillion US$ standard deviation, roughly) when average real tourism receipts are about 6.5 billion US$ (with 19.5 billion $ standard deviation, roughly).

Methodology
Before starting to test the causality between the variables, Im, Pesaran and Shin (IPS) (2003) unit root test for heterogeneous panels, which is built upon Dickey and Fuller's (1979) augmented unit root test (ADF), is applied. Emirmahmutoğlu and Köse (2011) have improved the lag augmented vector autoregression (LA-VAR) method recommended by Toda and Yamamoto (1995) using meta-analysis to test Granger causality in mixed heterogeneous panels and examined finite sample properties of the test by considering both cross-section independency and dependency via Monte Carlo simulations. Results uncover that the power of LA-VAR procedure is very high under considering both cross-section independency and dependency even when N (number of cross sections) and T (time periods) are small.
Following Emirmahmutoğlu and Köse (2011), level VAR model with + max lags in heterogeneous mixed panels below is considered: Here, max represents maximum integration order for each . Although we show only to test causality from ln to ln in Equation (2), same steps are also valid for Equation (1) to test causality from ln to ln .
The first stage is determination of maximum integration orders of variables for each using unit root tests and estimate Equation (2) using ordinary least squares for each and specify lag lengths ( 's) using information criteria.
The second stage is the estimation of Equation (2)  (3) In the third step, as proposed by Stine (1987), residuals have to be centered with herein below where ̂= (̂1 ,̂2 , … ,̂) ′ , = max( ) and = max( max ): Next, [̃, ] × is developed from these residuals. To conserve the cross-covariance structure of the errors, a full column with replacement from the matrix at a time is randomly selected. Bootstrap residuals are indicated as ̃ * ( = 1,2, … , ).
In the fourth stage, bootstrap sample of under the null hypothesis is generated as follows: ln , * =̂l n + ∑̂2 1, ln , − In the fifth step, ln , * is replaced for ln , with no parameter restrictions. Then the individual Wald statistics are calculated to test the null hypothesis of no causality on an individual basis for each . Individual p-values are estimated using corresponding individual Wald statistics. Afterward, Fisher test statistic is calculated as follows: where is the -value corresponding to the Wald statistic of the -th individual.
The bootstrap empirical distribution of Fisher test statistics is made by replicating third and fifth steps and specifying bootstrap critical values by choosing the proper percentiles of these sampling distributions. This paper also employs Pesaran's (2004) cross-sectional dependence test, which is proper method when time dimension ( ) is lesser than the cross-section dimension ( ). Pesaran's (2004) test statistic can be computed both for homogeneous/heterogeneous and nonstationary series.

CONCLUSIONS
Scarce sources that countries already have are turning these countries into diverse their income items. Tourism is one of the outstanding factors to stimulate the economy. Thus, lots of countries improve their balance of payments and provide economic growth by increasing export revenues from tourism. This paper investigates this relationship for 33 middle-income countries using causality test suggested by Emirmahmutoğlu and Köse (2011).
Tourism led-growth hypothesis is accepted Grenada and South Africa. This finding for South Africa is in accordance with Akinboade and Braimoh (2009). Unidirectional causality from tourism expenditures to GDP points that tourism expenditures play key role on economic growth as a complementary to factors of production. Therefore, expansionary policies on tourism expenditures could be supported in these countries.
Conservation hypothesis is valid for Albania, Bulgaria, Dominic Republic, Ecuador, Jamaica, Peru and Sri Lanka, as well as the whole panel. This finding for Bulgaria is in accordance with Aslan (2013) when incompatible with Chou (2013). Also, this finding on the whole 33 upper-middle income countries is in harmony with Ekanayake and Long (2011), who find that GDP causes tourism in the short-run in developing countries. Unidirectional causality from GDP to tourism receipts means that economic growth contributes tourism receipts. Therefore, policies on tourism receipts will not affect the economic growth.
Feedback hypothesis is accepted for Armenia, Mauritius and North Macedonia. Bidirectional causal connection between tourism expenditures and GDP entails these countries to take into consideration both variables when designing economic policies.
Finally, it is seen that neutrality hypothesis is acceptable for the remaining countries. Lack of causality between tourism and GDP for Malaysia is incompatible with Tang and Tan's (2015) results which indicate tourism causes GDP. No causality between tourism receipts and GDP shows that expansionary or contractionary policies on tourism will have no impacts on economic growth. Therefore, economic policies can be designed regardless of tourism related issues in these countries.
Future research may use time-varying parameter VAR models to test the causality and may find different results for different periods.