برآورد کشش تقاضا در پانل های غیر ثابت: مورد گردشگری هاوایی

It is natural to turn to the richness of panel data to improve the precision of estimated tourism demand elasticities. However, the likely presence of common shocks shared across the underlying macroeconomic variables and across regions in the panel has so far been neglected in the tourism literature. We deal with the effects of cross-sectional dependence by applying Pesaran’s (2006) common correlated effects estimator, which is consistent under a wide range of conditions and is relatively simple to implement. We study the extent to which tourist arrivals from the US Mainland to Hawaii are driven by fundamentals such as real personal income and travel costs, and we demonstrate that ignoring cross-sectional dependence leads to spurious results.

The past several decades have seen tremendous growth in the literature which seeks to explain and forecast tourist flows (Song et al., 2012 and Li et al., 2005). Following a wide variety of empirical methods applied across different countries and time periods, researchers have produced an even wider array of estimates for demand elasticities central to marketing, forecasting and policy work. While the income elasticity of tourism demand is generally expected to lie between one and two, Crouch, 1995 and Crouch, 1996 found that nearly 5% of estimates from 80 international studies were negative. Analyzing 30 years of international tourism demand studies, Witt and Witt (1995) found income elasticity estimates ranging from 0.4 to 6.6 with a median value of 2.4. Although an “inferior” tourist destination could explain a negative income elasticity, and an elasticity of less than 1 might be explained by some “necessary” short-haul international trips, such as those from the US to Canada, the large variation in estimates calls into question their validity and limits their usefulness to decision-makers. Estimates of price elasticities fare about the same. Witt and Witt (1995) found estimates ranging from −0.05 to −1.5, and Crouch, 1995 and Crouch, 1996 found that about 29% of the estimates were positive. Finally, these studies found transportation price elasticity estimates ranging from 0.11 to −4.3. Crouch (1996) investigated a number of potential causes of such disparate results, and noted that model specification played an important role. We suspect that the wide range of elasticity estimates is due to the limited information and short samples used in time series models, and the use of panel estimation techniques that do not adequately deal with important characteristics of panel data. There is a rich literature making use of a variety of approaches to explain and forecast tourist flows. Conventional approaches range from exponential smoothing to vector autoregressive and error correction models using time series data for a single origin-destination pair (Witt and Witt, 1995 and Li et al., 2005). Recently some alternative quantitative tools, such as artificial neural networks, fuzzy time series, and genetic algorithms, have been showing up in the literature. (For a comprehensive survey of recent developments in tourism demand modeling, see Song & Li (2008).) Unfortunately, the entire literature on tourism demand is at the mercy of short time series samples. And the limited data available for estimation has likely contributed to imprecise estimates of demand elasticities. For example, Bonham, Gangnes, and Zhou (2009, p. 541), report an income elasticity for Hawaii tourism demand from the US that is “implausibly large and estimated quite imprecisely”. Fortunately, it may be possible to obtain better estimates of the parameters of interest by taking advantage of the variation in both the temporal and cross-sectional dimensions of panel data sets. This point has not been lost on the tourism literature, and as econometric tools have advanced, a trend to exploit the richness of panel data has emerged (Song and Li, 2008 and Seetaram and Petit, 2012). In their review article, Song et al. (2012, p. 1657) suggest that “future studies should pay more attention to the dynamic version of panel data analysis and to more advanced estimation methods… ”. While early panel studies ignored problems arising from nonstationarity and potential cointegration, the tourism literature has now begun to address such issues. Among others, Seetanah, Durbarry, and Ragodoo (2010) estimated a static model of demand for South African tourism using Fully Modified Ordinary Least Squares (FMOLS) developed by Pedroni (2001). Using the same technique, Lee and Chang (2008) investigated the long-run co-movements and causal relationships between tourism development and economic growth. Falk (2010) applied the dynamic heterogeneous panel technique of Pesaran, Shin, and Smith (1999) to estimate the effects of snow fall on winter tourism in Austria. One common thread running through this nascent literature is reliance on the assumption of cross-sectional independence, or that each unit contributes entirely new information to the dataset. Yet, cross-sectional units are almost certainly influenced by national or global shocks such as business cycles, technological innovations, terrorism events, oil crises or national fiscal and monetary policies. In fact, a large empirical macro and macro-finance literature (see Stock and Watson, 1989 and Stock and Watson, 1998) and results presented here for Hawaii tourism show that cross-sectional dependence is very common. And, neglecting cross-sectional dependence can lead to substantial bias in conventional panel estimators (Kapetanios, Pesaran, & Yamagata, 2011). An increasingly common solution to the problem of cross-sectional dependence is to model such dependence using a factor structure. To the best of our knowledge, this approach has not been used in the tourism literature where cross-sectional dependence is usually ignored. But at least one study has included observed proxies for unobserved common factors. Nelson, Dickey, and Smith (2011) used oil prices, indicator variables for the effects of the September 11, 2001 terrorist attacks, and a nonlinear time trend capturing the overall slow-down of tourism demand during recessions. Such proxy variables may be effective in mitigating the effects of cross-sectional dependence, but their choice involves judgement on the part of the researcher, and it is unclear whether they are adequate to capture all sources of common shocks. Alternatively, unobserved dynamic common factors can be approximated using the methods proposed by Bai et al., 2009 and Pesaran, 2006, or Kapetanios et al. (2011). These approaches have the benefit that they do not require selection of a set of observed proxies. We estimate tourism demand elasticities from a panel of tourist arrivals to Hawaii from 48 US states over 19 years using the common correlated effects (CCE) estimator of Pesaran, 2006 and Kapetanios et al., 2011. This technique offers many advantages. First, the CCE estimator allows us to deal with the possibility of cross-sectional dependence caused by common factors. Second, it does not require ex ante information about the unobserved common factors, allows the factors to contain unit roots and to be correlated with the regressors. Finally, the CCE estimator offers good finite sample properties (Kapetanios et al., 2011 and Westerlund and Urbain, 2011), and is relatively simple to implement. The rest of this paper is organized as follows: in Section 2 we outline our tourism demand model and describe the CCE estimator we use to deal with cross-sectional dependence in panels; in Section 3 we present panel estimates of demand elasticities for Hawaii tourism; and Section 4 concludes.

The dramatic growth in tourism over the past several decades has led to an extensive literature which seeks to quantify the effect of income and prices on tourist flows. Although estimates of demand elasticities are central to marketing, forecasting and policy work, the literature on tourism demand has produced a range of elasticity estimates that are occasionally at odds with economic theory and reduce their usefulness in decision making. To improve the precision of estimates, it is natural to turn to the richness of panel data exhibiting variation in both the temporal and the cross-sectional dimension. This point has also been realized by scholars in the tourism literature, and as econometric tools have advanced, a trend to exploit the greater information content of panel data has emerged. However, panel estimation using non-stationary data requires careful attention to the likely presence of common shocks in the underlying macroeconomic variables. Early panel studies of tourism demand have relied on the assumption of cross-sectional independence, or that each region contributes entirely new information to the dataset. Yet, cross-sectional units are generally influenced by national or global shocks such as business cycles, technological innovations, terrorism events, oil crises or national fiscal and monetary policies. We demonstrate that neglecting cross-sectional dependence leads to spurious estimation results. Our contribution to the literature lies in estimating tourism demand elasticities while accounting for unobserved non-stationary common factors in the data. We use the CCE estimators of Pesaran, 2006 and Kapetanios et al., 2011 to deal with cross-sectional dependence in panel regressions. This technique offers several advantages over competing methods: it does not require ex ante information about the unobserved common factors, allows them to contain unit roots and to be correlated with the regressors, allows for heterogeneous factor loadings, exhibits good finite sample properties, and is relatively simple to implement. We apply the CCE estimators to US state level quarterly data spanning the period from the first quarter of 1993 to the first quarter of 2012. We obtain an income elasticity for Hawaii tourism demand that is slightly greater than one, elastic demand with respect to hotel room prices, and inelastic demand with respect to airfare. Our estimates are more plausible than those of Bonham et al. (2009) whose pure time series model did not benefit from cross-sectional variation, and are in line with the results of Nelson et al. (2011) who included in their model observed proxies for common factors, such as oil prices and a non-linear time trend. However, the selection of observed variables as proxies requires some subjective judgement on the part of the researcher, and in general such proxies may not successfully capture all sources of unobserved common factors in the panels. In contrast, our CCE based approach does not require the selection of observable variables to capture the sources of common shocks. Instead, the method uses cross-sectional averages to directly approximate the unobserved factors in the analyzed series, and it consistently filters out those factors that are actually present in the panels. It is important to note that our analysis is subject to some limitations. Driven primarily by data constraints we limit our analysis to a subset of relative price effects. For example, we do not consider price competition between Hawaii and other long-haul destinations such as Mexico and the Caribbean. Furthermore, our pooled and mean group estimators produce elasticities of overall tourism demand from the US Mainland, and do not differentiate among market segments. It is likely that demand elasticities vary across regional and other types of market segments, and we intend to explore these issues in future research. Finally, while tourism demand may fluctuate with the business cycle, we are not analyzing that relationship. Because the influence of business cycles is filtered out in our estimation approach, the reported estimates represent elasticities, or isolated marginal effects of the regressors.