ABSTRACT
Tourist’s expenditure is a key variable for the tourism industry with a high presence in the literature. Recent micro-econometric approaches estimating the factors driving the expenditure of visitors do not confer much relevance to destination features, despite its relevance shown by the theoretical tourism literature. The present investigation proposes a novel framework in tourism demand studies, expanding the previous modelling approach, in order to take into account such contextual destination effects. Particularly, we define a two-level hierarchical spatial model that includes spatial effects, operating at both the level of the destination and the neighbouring clusters, accommodating in this way for a spatial local dependence structure in the spending choices of visitors. Empirical evidence on the performance of this modelling framework is obtained by building on a relevant dataset for more than 136,000 international visitors reaching 1872 destinations in Spain. This allows for a multi-destination analysis showing that the destination contextual and local neighbourhood effects could account up to 50% of the variance of the expenditure of tourists in the case of most popular destinations. However, such effects do not show this salient role in the case of less developed destinations. Policy recommendations emerging from the results of the investigation are discussed further.
DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.
Notes
1. For example, Pierewan and Tampubolon (Citation2014) examine the determinants of individuals’ well-being across the European regions; Dong and Harris (Citation2015) evaluate the price of urban parcels; and Osland et al. (Citation2016) analyse the effect of the spatial structure of the region on housing prices.
2. For the classification, see https://reports.weforum.org/travel-and-tourism-competitiveness-report-2019/rankings/.
3. The index is computed by the Spanish private organization Fundación La Caixa and is periodically disseminated through its Spanish Social Yearbook Report.
4. The Moran’s test is based on a predefined structure of spatial dependence which is captured by a spatial weight matrix (W). This matrix indicates how the strength of the relationship between territorial units distributes across the whole space. In this case, the inverse of squared distance specification is chosen. Alternative specifications for the spatial weights, based on k-nearest neighbours or inverse distance, were also tried with non-remarkable differences in the empirical results, so we decide to keep this usual definition of the W matrix.
5. The main reason for employing a random effects multilevel/hierarchical model instead of a fixed effects specification lies on three main arguments. First, we are basically interested in capturing the distribution of the contextual effects, so employing a model that directly produces estimates of these components appears more desirable. Second, this type of specifications allows us to model spatial dependencies emerging at the destination-level. And third, random effects models offer more efficient estimations than other employed modelling strategies such as the full-pooling and no-pooling approaches as noted, for example, by Lacombe and McIntyre (Citation2017, p. 154) and Bell et al. (Citation2019, p. 1061).
6. Gelman and Hill (Citation2007) offer a general introduction to multilevel models.
7. Corrado and Fingleton (Citation2012) propose accommodating spatial effects in the contextual effects in multilevel models through a conditional autoregressive specification (see their equation 38). An equivalent approach can be found in Osland et al. (Citation2016), and an empirical comparison of these two alternative approaches is presented by Bivand et al. (Citation2017), who concluded that they produce results that are to a large extent comparable.
8. After denoting the dependent variable by , the ICC equals , where i and í represent individuals who travel to the same destination j. It is straightforward to check that in the HEM specification:
and
for .
By contrast, in the HSEM specification:
and
.
9. The estimation employs the Bayesian method introduced by Dong and Harris (Citation2015), as implemented in the R package HSAR v.0.5.1 (Dong et al., Citation2017).
10. In both cases, the continuous covariates are grand mean centred (expressed as deviations from the grand mean), so that the contextual effects may be interpreted as the deviation of the (logged) tourist expenditure from a reference individual evaluated at the mean.
11. When interpreting the coefficients for the dummy variables, it is assumed that approximates the ratio between the expected tourist spending associated with the alternative values of the covariate (all else equal). More specifically, denoting the level of the dependent variable by Y, the dummy explanatory variable by X and the associated coefficient, say , when the model specifies the dependent variable in logs, the following approximation may be considered for interpreting the estimated coefficient:
,
so that:
.
12. Estimates of the coefficients for the regional dummies are not reported in Table 3 although, as usual, they are available from the authors upon request.