Testing the validity of the tourism-led growth hypothesis under long-range dependence

ABSTRACT

Traditionally, the relationship between tourism and economic growth has been studied using methods based on stationary or nonstationary models. In this paper, we explore the validity of the tourism-led growth hypothesis, applying fractional cointegration techniques and comparing the results with those obtained in previous studies. To do this, we study the long-run relationship between GDP and tourism series, using a two-step strategy. First, we apply fractional cointegration methods and then we test the null hypothesis of no cointegration against alternatives which are fractionally cointegrated. For the empirical analysis, we use seasonally adjusted quarterly data for GDP and tourist arrivals for seven European countries from 1990 to 2018. In the empirical exercise, evidence of cointegration is found only in very restricted cases. Therefore, the validity of the TLGH is less clear-cut when this methodology is applied.

KEYWORDS:

• Fractional integration
• narrow-band least squares (NBLS) and fully modified NBLS (FMNBLS) cointegration estimators
• fractionally cointegrated vector autoregressive (FCVAR)

JEL CODES:

• C01
• C52
• Z32
• Z38

Acknowledgement

María Santana-Gallego acknowledges support received from the CICYT Program (Spanish Government) through grant ECO201679124-C2-1-R (AEI/FEDER, UE). The views expressed here are those of the authors and not necessarily those of the institutions with which they are affiliated.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.

Notes

1 Mandelbrot (1977) characterized long memory processes as having ‘fractal dimensions’.

2 For instance, it is sometimes tedious to test for cointegration in the regression-based approach when integration orders are not confined to a particular region of the parameter space or when the residuals are not well defined, as is the case with semiparametric methods. In such cases, there is little consensus on how to test for fractional cointegration. On the other hand, cointegration rank analysis often requires tuning parameters that increase the sensitivity of the results (see Robinson & Yajima, 2002) but impair information on the strength of the relation.

3 According to Granger (1986) and Engle and Granger (1987), cointegration arises when $d_x>d_u\ge 0$. This case implies a dimensionality reduction, revealing a near-zero-frequency correlation between the two series. Let $b=d_x-d_u$ be the reduction parameter and $C\left(d_x,b\right)\ge 0$ the cointegration relationship. Following Aloy and De Truchis (2012), we can identify the following five cases: (1) Spurious regression: $0.5\le d_x\le 1,0.5\le d_u\le d_x,b=0$. (2) Standard cointegration$C\left(1,1\right)$: $d_x=1,d_u=0,b=1$. (3) Strong fractional cointegration $C\left(d_x,b\right)$: $0.5<d_x\le 1,0\le d_u<0.5,b>0.5$. (4) Weak fractional cointegration $C\left(d_x,b\right)$:$0<d_x\le 1,0\le d_u<d_x,b<0.5$. (5) Stationary fractional cointegration $C\left(d_x,b\right)$: $0<d_x\le 0.5,0\le d_u<d_x,b<0.5$.

4 These authors suggest an adaptation of the maximum likelihood estimator of Hualde and Robinson (2007) to jointly estimate β and d, and possibly other nuisance parameters, for a wide range of integration orders when the regressors are I(1).

5 The stationary fractional cointegration estimator, NBLS, is described in Appendix 2.

6 Tourism receipts could also be used as a proxy of tourism demand, but due to the limited availability of homogeneous quarterly data series for tourism receipts, we chose to use arrivals.

7 Important tourist destinations such as France or Turkey are not considered since their tourism data reporting does not cover the whole sample period.

8 Appendix 2 shows graphs for the time path and 100 sample autocorrelations for the logarithms of the GDP and tourist arrival series. As can be seen, both variables have increasing trends and autocorrelations decreasing slowly towards zero and with cyclical patterns indicating that they may be nonstationary.

9 Table A3 in Appendix 3 presents a unit root analysis with structural breaks. We present the Zivot and Andrews (1992) test for a unit root allowing for a single break in the intercept and the trend. In general, we cannot reject the null of a unit root. Similarly, the Lumsdaine and Papell (1997) and Lee and Strazicich (2003) tests do not allow us to reject the null of a unit root, in either case. These tests indicate, therefore, that the data series are not trend stationary with breaks in the intercept and trend at two time points. It is also noteworthy that the estimated breaks differ among the various alternatives. Finally, we reject the null of a unit root with breaks (one or two) in several cases (shown in bold and italic). Therefore, the series could be considered trend-stationary with breaks in both the intercept and the trend.

10 An adaptation of the Robinson and Yajima (2002) statistic to log-periodogram estimation could be used. The functional form of the test statistic can be found in Gil-Alana and Hualde (2009).

11 It is noteworthy that our sample is similar in length to some simulated examples used in Nielsen and Frederiksen (2011) to evaluate the finite sample properties of the GLS version of the FMNBLS and NBLS estimators. These simulations clearly demonstrated the superiority (in terms of both bias and RMSE) of the FMNBLS relative to NBLS in the presence of non-zero long-run coherence between the regressor and the error, even in short samples (e.g. 128 observations).

Additional information

Funding

María Santana-Gallego acknowledges support received from Ministerio de Ciencia, Innovación y Universidades through grant ECO201679124-C2-1-R (AEI/FEDER, UE).]