Regional tourism convergence: a disaggregated analysis of Croatia

ABSTRACT

This study examines tourism convergence across regions within a country as a means to understand the extent to which there is segmentation of regional tourism markets and the differential impact of domestic versus foreign tourism. Additionally, our empirical analysis introduces a newly developed weak σ-convergence test alongside the club convergence approach to explore regional tourism convergence. Using disaggregated monthly county-level data for Croatia from 1998:1 to 2021:9, the results clearly identify the divergence across counties to suggest the regional segmentation of tourism in the country between coastal and continental regions. The policy implications of the results are also discussed.

KEYWORDS:

• Tourist arrivals and overnight stays
• club convergence
• σ-convergence
• Croatia

JEL Codes:

• Z32
• C33

Disclosure statement

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

Notes

¹ www.unwto.org/tourism-statistics/economic-contribution-SDG, accessed September 25, 2022.

² As noted by Payne, Gil-Alana, and Mervar (2022), earthquakes occurred in 2020 that affected the City of Zagreb along with Sisak-Moslavina and Krapina-Zagorje counties.

³ “Strategy for the Development of Sustainable Tourism until 2030” (2022).

⁴ Together with the recovery and resilience plans of the other 26 EU countries, it is a part of a wide-ranging response of the European Union to the negative economic and social impacts of the COVID-19 pandemic.

⁵ Hence, unlike the previous studies, we assume that the absence of convergence does not necessarily imply divergence since there is the possibility for the prevalence of convergent clubs across counties.

⁶ This division corresponds to the European Union official division for regional statistics at NUTS-3 level (NUTS-nomenclature of territorial units for statistics). Figure 1 parallels the presentation of the Croatian county map by Gil-Alana, Mervar, and Payne (2015) and Apergis, Mervar, and Payne (2017).

⁷ Since the log-t function is rapidly increasing initially, Phillips and Sul (2007) suggest eliminating the first third of the sample in recommending r = 0.3. It is worth noting that since u_t is serially correlated, Phillips and Sul (2007) employ the HAC (heteroscedasticity and autocorrelation consistent) long-run variance estimator proposed by Newey and West (1987) to construct the robust t-statistic.

⁸ The pattern of $K_t^y$ depends on any possible common factors in $y_{it}$. A factor structure definition is $y_{it}=a_i+{\lambda }_i{F}_t+u_{it}$ where $F_t$ is a vector of unobserved common factors and ${\lambda }_i$ is a vector of factor loadings. One can estimate $F_t$ with the cross-sectional averages of $y_{it}$ if there is one factor, or with the method of principal components if there is more than one factor. Then, the common factor is eliminated by ${\tilde{y}}_{it}=y_{it}-{\hat{\lambda }}_i{\hat{F}}_t$ by keeping fixed effects $a_i$ in ${\tilde{y}}_{it}$; and the linear trend regression is estimated with using ${\tilde{y}}_{it}$ instead of $y_{it}$.

⁹ Before proceeding with inferences, the Hodrick and Prescott (1997) filter was applied to the natural log transformed data for removing (seasonal and business) cycles in the data (see Phillips and Sul 2007, 2009).

¹⁰ Note four different truncation parameter settings are used (L = 3, 6, 9, 12) for the construction of the Newey and West (1987) HAC long-run variance.

¹¹ The relative convergence condition requires the relative transition curves $h_{it}$ converge to unity, and the variance of $h_{it}$ converges to zero as $t\to \mathrm{\infty }$.

¹² We also examined the results for the pre-COVID-19 period. The results for relative (club) convergence and weak σ-convergence still support divergence similar to the results inclusive of the COVID-19 period presented. Results are available upon request.

¹³ As shown at the bottom of Table 3, the principal components explain 90, 97, 81, and 96% of total variance for DOMAR, FORAR, DOMON, and FORON, respectively.

¹⁴ We are grateful to the anonymous reviewer for pointing out the need for robustness analysis by conducting other convergence tests. We estimate convergence models for absolute β-convergence with a battery of panel data estimators with common factors; and for stochastic convergence with a battery of panel stationarity tests, including both conventional first-generation tests without cross-section dependence and second generation tests with cross-section dependence. The results presented in the Appendix indicate that although the absolute β-convergence is supported overall, stochastic convergence is not supported consistent with the evidence in favour of divergence from relative and weak $\sigma$-convergence approaches.

¹⁵ To conserve space and for details on the clustering algorithm see Phillips and Sul (2007, 1798–1801).

¹⁶ The condition requires the relative transition curves converge to unity and the variance of the transition curves converges to zero as $t\to \mathrm{\infty }.$.

¹⁷ The results are similar for the pre-COVID-19 period.

¹⁸ See “National Development Strategy” (2021).