Tourism foreign direct investment led tourism gross value added: a co-integration and causality analysis of Croatian tourism

Abstract The purpose of this study was to investigate the causal relationship between foreign direct investment in tourism and tourism gross value added in Croatia. The study employed econometric techniques, such as the unit root test, Johansen co-integration, and the Granger causality test, in a vector error correction model (V.E.C. model), and the Toda–Yamamoto causality test in a vector autoregressive model (V.A.R. model), using quarterly time-series data from 2000(1) to 2012(4). The results confirm the existence of a stable co-integrated relationship between variables in the long term. A short-term relationship was also proved between foreign direct investment in tourism and gross value added, using the Toda–Yamamoto causality test. By including control variables, the two-way causality between the subject variables was proven using the Granger causality test.


Introduction
Despite the volatility of the tourism sector, Croatia is among those countries the economies of which are extremely dependent on tourism. Given the World Travel and Tourism Council's (W.T.T.C., 2011) forecasts, the situation will not change significantly in the future 1 Due to the non-investment activity in the Croatian economy since the onset of the global financial crisis, which refers to the lack of domestic capital, but also the reluctance of foreign investors, it seems logical to assume that further development of Croatian tourism needs foreign direct investment (F.D.I.).
The above-mentioned situation is consistent with The Master Plan and Strategy for the Development of Croatia's Tourism by 2020 (Institut za turizam, 2012), which states that tourism will be recognised as a priority economic activity by 2020. In that study, several possible scenarios for the development of Croatian tourism by 2020 are elaborated. A scenario that it is important to address, bearing in mind the topic of this paper, is that of major development investments. This scenario puts emphasis on the construction of new in tourism, and expectations of the Croatian government for further entry of foreign capital into the Croatian tourism sector, research findings have significant, wide-ranging socio-economic implications. The fact that there are no uniform results from previous research on causality between F.D.I. and tourism growth clearly points out the need for analysis of such on a concrete example. In order to establish such effects, this study is based on the example of Croatia.
The rest of the study is organised as follows: Section 2 describes the literature review on the subject topic; data and the methodological framework are described in Section 3; section 4 concerns interpretation of results; and, finally, conclusions and policy implications are drawn in Section 5.

Review of literature
A large body of literature is related to the general effects of F.D.I. on economic growth. Despite that, research results are still unclear and, thus, open to further research. It is possible to substantiate the positive results of this research (Asteriou, Dassiou, & Glycopantis, 2005;Borensztein, De Gregorio, & Lee, 1998;Mlinarević, 2004), but there are also negative results (Mencinger, 2003). The situation is even more complicated if the research is brought to the level of the sector, because the effects can vary, depending on whether this is a primary, secondary, or tertiary sector (UNCTAD, 2001, p. 138). In 1958, Hirschman concluded that not all sectors have the same potential to absorb foreign technology and connectivity as the rest of the economy (Hirschman, 1958, p. 109). Numerous recent studies have proven the different impacts of F.D.I. on the productivity of each sector (Alfaro, 2003;Cipollina, Giovannetti, Pietrovito, & Pozzolo, 2012;Tondl & Fornero, 2010).
With respect to tourism as a part of the tertiary sector, and despite the continuing increase in the volume of tourist travel (increase in the number of tourists, growth rates, and growth in tourism revenues), and the growth of F.D.I. over the past 20 years, the area relating to F.D.I. in the tourism sector remains practically unstudied. The initial research on the effects of transnational corporations on tourism is the work of Dunning and McQueen (1982). They concluded that the effects of foreign-owned hotels varied, depending on the type of tourism, the transnational corporations' country of origin, the host country, and the advantages and objectives of individual transnational corporations. Sinclair and Stabler (1991) argued that F.D.I. in tourism is a neglected area in studies related to tourism. Zhang (1999), who is considered a pioneer in the field of research on F.D.I. in tourism, believed that his research would be significant enough to spur further research in this area. The World Tourism Organisation (W.T.O.) has stressed that F.D.I. and tourism have only recently been defined and explained, so F.D.I. in tourism is considered a new, and little explored, area (U.N.W. T.O., 1999).
Recent research regarding the impact of F.D.I. on tourism is based on testing the so-called F.D.I.-led tourism growth hypothesis (Salleh, Othman, & Sarmidi, 2011, p. 251). This may indicate that F.D.I. preceding tourism growth is in some way associated with F.D.I., and not only that the change in the tourism growth is a result of changes in the level of F.D.I. Study of the long-and short-term relationships between F.D.I. and tourism growth could be conducted using the co-integration and causality tests. Such research methodology is very common in papers dealing with the tourism-led growth hypothesis, with respect to the issue of causality tests between tourism development and economic growth, exemplified by particular destination (Arslanturk, Balcilar, & Ozdemir, 2011;Belloumi, 2010;Bouzahzah & El Menyari, 2013;Schubert, Brida, & Risso, 2011;Surugiu & Surugiu, 2013), or a panel of countries (Aslan, 2014;Caglayan, Sak, & Karymshakov, 2012;Fourie & Santana-Gallego, 2011;Mello-Sampayo & Sousa-Vale, 2012). As far as papers related to F.D.I.-led tourism growth hypothesis are concerned (Salleh et al., 2011), with respect to the issue of causality tests between F.D.I. and tourism development, the same research methodology is used. Nevertheless, this represents a significantly smaller number of studies. Existing studies indicate different results for the association of tourism development and F.D.I., those in total and those aimed at tourism. Most of the research has proved a one-way causal relationship between F.D.I. and tourism growth (Bezić et al., 2010;Samimi, Sadeghi, & Sadeghi, 2013;Selvanathan, Selvanathan, & Viswanathan, 2012;Tang, Selvanathan, & Selvanathan, 2007). There is also one-way causality running from tourism growth to F.D.I. (Katircioglu, 2011;Khoshnevis Yazdi, Homa Salehi, & Soheilzad, 2015), as well as evidence of a two-way link between the observed variables (Othman, Salleh, & Sarmidi, 2012;Salleh et al., 2011).
Even though the amount of scientific papers that explore the relationship between tourism growth and F.D.I. in tourism have increased in the last few years, that number is still relatively small. Given the diversity of research results, and the fact that, according to the authors' knowledge, testing causality between selected variables in the case of Croatia has not been carried out so far, stated justified the purpose of research.

Data and methodological framework
This research employs quarterly time-series data from 2000(1) to 2012(4) in order to investigate the causal relationship between F.D.I. in tourism and tourism gross value added in Croatia. A variable for F.D.I. stock in tourism (F.D.I.-T) was obtained from the Croatian National Bank. The variable is deflated by the implicit deflator of gross investment, and reduced to the base year 2005. A variable for tourism gross value added (T.G.V.A.) was also obtained from the Croatian National Bank. This variable is deflated by the implicit GDP deflator, and reduced the base year 2005, and is seasonally adjusted (Census X-12). The productivity of the tourism sector is seen through the level of gross value added in the tourism sector 2 Both variables are in logarithm form.
As a first step of empirical analysis, a unit root test is conducted to determine the stationarity of the time-series data. Most of the time-series variables are non-stationary, and the use of such data leads to spurious regression, which cannot be used to get objective and accurate results from. The order of integration of the variables included in the model is determined by using the Augmented Dickey-Fuller (A.D.F.) test.
If the first step indicates that both variables are non-stationary, it is possible that there are short-and long-term relationships between the variables, so the second step examines their co-integration, using the Johansen co-integration test. The important pre-condition that needs to be fulfilled in applying the Johansen co-integration test is that the data must be of the same stationarity order. The V.A.R. model should also be very well specified before it approaches testing the co-integration. This requires selection of the optimal lag length through the usual information criteria, and evaluation of the quality of the model itself. Selection of the optimal lag length is crucial for the reliability of V.A.R. models (Liu, 2005).
The quality of the V.A.R. model will be verified by testing the normality of the distribution (Jarque-Bera test), the serial correlation (Lagrange multiplier [L.M.] test), and the heteroscedasticity (White's test) of the residuals.
Thirdly, if variables are integrated of order one, I (1), and there is no stable co-integration relationship between them, the V.A.R. model is built. Studies using a V.A.R. model in a situation where there is co-integration between the variables in the model are actually incorrectly specified. This is also the reason why this study first tests whether there is co-integration between the variables. If the variables have the same order of integration, and if the co-integration is established between them, the V.E.C. model will be assessed. The V.E.C. model is also called a restricted V.A.R. model. The V.E.C. model shows the long-and short-term relationships between the variables. The main purpose of the V.E.C. model is to emphasise the speed of adjustment of a short-term balance to a long-term equilibrium. Economic theory assumes that there is a tendency towards equilibrium that prevents the individual time-series to stray too far. Such a co-integration provides a so-called error correction model (E.C.M.). A lagged error correction term (ECT t-1 ) should have a negative value, indicating the return to the equilibrium state, and have a value ranging from 0 to 1, where 0 indicates no adjustments, while 1 indicates the total adjustment after a period of time.
As part of the evaluation of the V.E.C. model, the following models are specified: where logfdi_t and logbdv_t are, respectively, the logarithmic form of the F.D.I. in tourism and tourism gross value added in Croatia, Δ is the sign of differentiation, p is the optimal lag length, and ECT t-1 is a lagged error correction term. A small number of observations, usually no more than 40 per country, are a common feature of empirical studies involving causality testing (Zachariadis, 2006), especially in the case of south-eastern Europe. This is also the case in this study. So, finally, in order to reach more reliable conclusions, the authors used two methods of determining causality, regarding a small number of observations. Respectively, the Granger and Toda-Yamamoto causality tests were applied.
Granger causality is possible to express within the set V.E.C. model. Causality in co-integrated systems can be established if, and only if, the ECT t-1 , which takes into account long-term dynamics, and the sum of the coefficients of the lagged variables, which takes into account short-term dynamics, are both significant (Asghar, Awan, & Ur Rehman, 2012, p. 139).
The majority of existing papers researching the causality between variables use the standard Granger causality test. The classic Granger causality test is a way to implement the Wald test for the first parameters (p) of other variables in the V.E.C. model, and, if the Wald test is significant, it rejects the null hypothesis of no causality.
This study, as well as the Granger causality test, also uses the modified Granger causality test, or the Toda-Yamamoto causality test (Toda & Yamamoto, 1995). The Toda-Yamamoto causality test enables a more reliable conclusion to be drawn (Magnus & Fosu, 2008, p. 106). The Toda-Yamamoto test ignores any possible non-stationarity or co-integration between series, when testing for causality and fitting a standard V.A.R. in the levels of the variables (rather than first differences, as is the case with the Granger causality test) (Mavrotas & Kelly, 2001, p. 100). In this way, the risks associated with possibly incorrectly identifying the orders of integration of the series, or the presence of co-integration, are minimised and it also minimises the distortion of the tests' sizes, as a result of pre-testing (Chowdhury & Mavrotas, 2006, p. 4). The Toda-Yamamoto causality test involves estimation of an augmented V.A.R. (p + m) model, where p is the optimal lag length in the original V.A.R. system, and m is the maximal order of integration of the variables in the V.A.R. system. When a V.A.R. model is used for purposes other than testing for Granger non-causality, and the series are found to be co-integrated, a V.E.C. model should be used (Giles, 2011).
The Toda-Yamamoto causality test was applied in assessing the following pairs of bi-V.A.R. system: where logfdi_t and logbdv_t are, respectively, logarithmic forms of the F.D.I. in tourism and tourism gross value added in Croatia, p is the optimal lag length, and m is maximal order of integration of the variables in the V.A.R. system.
In order to ensure the reliability of the results herein, Granger and Toda-Yamamoto causality were tested in the V.A.R. by including specific control variables: good governance is measured by the Control of Corruption (logcc) and the Political Stability indices (logpolstab), which were taken from the World Bank. The value of annual data in both variables is attributed to the respective quarters. Furthermore, an educational variable is measured by the number of highly educated workers in the activity I (logskill), and was obtained from the Croatian Bureau of Statistics. Finally, movement of the exchange rate is measured by the movement of kuna in relation to euro (logexrt), and the variable is obtained from Eurostat. For the last two variables, the quarterly time-series were used, and these variables are seasonally adjusted (CENSUS X-12). All the control variables are in logarithmic form.
The Granger causality test was applied in assessing the following pairs in the V.A.R. system: where logfdi_t and logbdv_t are, respectively, logarithmic forms of the F.D.I. in tourism and tourism gross value added in Croatia, CV is control variables, and p is the optimal lag length.
The Toda-Yamamoto causality test was applied in assessing the following pairs in the V.A.R. system: where logfdi_t and logbdv_t are, respectively, logarithmic forms of the F.D.I. in tourism and tourism gross value added in Croatia, CV is control variables, p is the optimal lag length, and m is that maximal order of integration of the variables in the V.A.R. system.

Research results
Stationarity of the variables was tested for both time-series, and the results of the unit root test (A.D.F. test) indicate that both series are stationary after the first difference, which is evident from Table 1.
Therefore, both of the time-series are integrated in the same order, and denoted as I (1). The V.A.R. model must be extremely well specified before testing for co-integration. This requires the selection of the optimal lag length for the variables in the V.A.R. model, using the usual information criteria, and the evaluation of the quality of the model itself. Standard information criteria indicate the selection of different optimal lag lengths of each variable, from 1 to 4 (see Appendix). The quality of the V.A.R. model -testing the normality of the distribution (Jarque-Bera test), the serial correlation (LM test) and the heteroscedasticity (White's test) of the residuals -is satisfied if the maximum lag length is p = 4. It is also necessary to check whether the model is 'dynamically stable' .
As can been seen from Figure 1, none of the roots is outside the circle, which proves that the V.A.R. model is stable (stationary). The obtained result is very favourable, because the stability of the V.A.R. model means that the results and conclusions following from further analysis are not questionable. Since both the time-series are of the same order of integration I (1), the Johansen co-integration was tested. It was found that co-integration is present; one co-integration equation is observed.
The lagged ECT (ECT t-1 ) of equation (1) is negative and significant, which is an indicator of a stable long-term relationship between F.D.I. and gross value added in tourism. The coefficient of -0.11 indicates that about 11% of the imbalance is corrected in the next quarter.
Proven co-integration is an indication of Granger causality. The results of the test are shown in Table 3.
The test results indicate that the causality among variables has not been proven in either direction. Given the proven co-integration, the presence of a causal link is expected, so the Toda-Yamamoto causality test will also be applied. Although the testing of co-integration does not affect the further procedure of Toda-Yamamoto's causality test, it is very useful for evaluating the validity of the final result. If the variables are co-integrated, but the causality has not been proven, it is possible that the sample size is too small. As already stated, this test allows for reliable conclusions.  Lastly, Toda-Yamamoto causality is tested by adding extra m = 1 lags; the results of the test are shown in Table 4.
The results of the research indicate that, in the case of the first dependent variable (logb-dv_t), there is the presence of short-term causality running from total F.D.I. in tourism to tourism gross value added in Croatia. Therefore, the first null hypothesis of Granger non-causality, logfdi_t non cause logbdv_t, could be rejected. In other words, the short-term causality running from F.D.I. in tourism to tourism gross value added in Croatia has been confirmed at the 10% level of significance.
Regarding the other dependent variables, the second null hypothesis, logbdv_t non-cause logfdi_t, could not be rejected; the results of this study indicate that, in the case of the second dependent variable (logfdi_t), there is the absence of short-term causality running from tourism gross value added to F.D.I. in tourism.
Finally, by including the control variables, the Granger and Toda-Yamamoto causality testing was conducted. Stationarity of the variables was tested for all control variables, and the results of the unit root test (A.D.F. test) indicate that all series are stationary after the first difference, which is evident from Table 5, below. Therefore, all the respective time-series are integrated in the same order and denoted as I (1). The quality of the V.A.R. model -testing normality of distribution (Jarque-Bera test), serial correlation (L.M. test), and heteroscedasticity (White's test) of their residuals -is satisfied if the maximum lag length is p = 2. It is also necessary to check whether the model is 'dynamically stable' .
As can been seen from Figure 2, none of the roots is outside the circle, which proves that the V.A.R. model is stable (stationary).
The results of the Granger causality test are shown in Ttable 6. The null hypothesis of Granger non-causality, logfdi_t non cause logbdv_t, could be rejected at the 5% level of significance. If we take into consideration the influence of the control variables, the null hypothesis of Granger non-causality, all variables non cause logbdv_t, could be rejected at the 1% level of significance.
Regarding the other dependent variables, the null hypothesis, logbdv_t non-cause logfdi_t, could not be rejected; however, taking into consideration the influence of control variables, the null hypothesis of Granger non-causality, all variables non cause logfdi_t, could also be rejected at the 1% level of significance.
Lastly, Toda-Yamamoto causality test results are shown in Table 7.  The short-term causality running from F.D.I. in tourism to tourism gross value added in Croatia has been confirmed at the 1% level of significance. If we take into consideration the influence of control variables, the null hypothesis of Granger non-causality, all variables non cause logbdv_t, could be rejected at the 1% level of significance.
Regarding the other dependent variables, the null hypothesis, logbdv_t non-cause logfdi_t, could not be rejected; the results indicate that, in the case of the second dependent variable (logfdi_t), there is the absence of short-term causality running from tourism gross value added to F.D.I. in tourism. Moreover, the results remain the same, if we take into consideration the influence of control variables.

Conclusion
The results of the existing body of literature that deals with the issue of F.D.I.-led tourism growth hypothesis are rather diversified. The effects of F.D.I. in any sector, including the tourism sector, depend on the degree of development of a country (for example, the level of training of tourism staff), the legislation in the country, the stage of the life cycle of a tourist destination, etc. Given the mentioned particularities, research should be conducted for each individual country.
In this study, empirical research on long-and short-term causal links between F.D.I. in tourism and tourism gross value added in Croatia, using quarterly time-series data from 2000(1) to 2012(4), has been conducted. The study employed econometric techniques, such as the unit root test, Johansen co-integration, and the Granger causality test in a V.E.C. model, and the Toda-Yamamoto causality test in a V.A.R. model. The results indicate that, in the case of Croatia, it is possible to discuss the F.D.I.-led tourism growth hypothesis. Both long-and short-term causal relationships between F.D.I. in tourism and tourism gross value added has been proven. Established long-and short-term causal relationships between variables are in line with the F.D.I.-led tourism growth hypothesis. Since the diagnostic tests are satisfied, and the stability of the V.A.R. model is confirmed, it is possible to conclude that the results obtained are reliable. The inclusion of certain control variables, such as index of corruption control, index of political stability, the exchange rate, and the educational variable, also contribute to the reliability of these results. By testing short-term causality using the Granger and Toda-Yamamoto tests, it is also confirmed that, with the inclusion of control variables in the case of Croatia, it is possible to discuss the F.D.I.-led tourism growth hypothesis.
These results confirm the importance of attracting F.D.I. in tourism, since it has the potential to stimulate sustainable tourism growth, and, thus, given the significance of tourism to the Croatian economy, it also has the potential to stimulate sustainable economic growth. The most important thing is to highlight that Croatian tourism is seriously lacking large hotel capacity that is necessary to ensure its viability, and the presence of established hotel chains could definitely help in providing that. These results are also in line with certain previous research related to the F.D.I.-led tourism growth hypothesis (Samini et al., 2013;Zhang, Ebbers, & Zhou, 2011).
Our results also confirm the necessity to create favourable conditions to attract F.D.I. The Granger causality test has proven that political stability, the level of corruption, the movement of the exchange rate, and education, along with the gross value added in tourism, are having a significant impact on F.D.I. attraction in tourism in the short term. As some research suggests, and according to our results, the political authorities should seriously approach the process of attracting foreign investors to establish incentivised macroeconomic business conditions, which require a good credit rating, quality infrastructure, a realistic exchange rate, competitive tax incentives, and favourable interest rates, to ensure political stability and eliminate the administrative and legislative barriers, and the image of the country as a 'corrupt destination' (Alfaro, 2003;Khoshnevis Yazdi et al., 2015;Perić & Nikšić Radić, 2015).
The significant impact of F.D.I. in tourism on tourism gross value added in Croatia validates the necessity of public intervention by implementing various policies, including the government providing incentives to foreign investors in tourism. On the other hand, given the importance of the control variables on the inflow of F.D.I. in tourism, the study results suggest an improved regulatory framework to make Croatia more attractive to F.D.I. inflow in tourism.
Among issues for future research, the authors could broaden this research to include testing the F.D.I.-led tourism growth hypothesis on whether it attracts greenfield or brownfield investment.

Notes
1. According to W.T.T.C.s' (2011) estimates from 2011, the total tourism contribution to gross domestic product (GDP), in the case of Croatia, was 27.6%, while, at the European level, that contribution was only 7.7%. Forecasts for 2021 are not very different -the total tourism contribution to GDP, in the case of Croatia, will amount to 30.7%, and, at the European level, 7.8%. 2. Several authors use the gross value added as a proxy for the level of productivity when measuring the effects of F.D.I. in particular sectors (Alfaro, 2003;Cipollina, Giovannetti, Pietrovito & Pozzolo, 2011;Tondl & Fornero, 2008