An analysis of taxpayer support for tourism promotion spending requires an econometric strategy with three objectives. The empirical model should capture the impact on tourism-related industries (here hotels and accommodations, amusement and recreation and arts and entertainment). This estimation should allow for the potential for nonlinear effects. Second, the model should permit estimation of the cross-border effects of tourism spending. Finally, the model should account for trend dynamics and spatial interactions that, if omitted, might bias the estimates of the impact of state spending on tourism promotion.
The approach used here calls for impacts estimated in a simple model that account for the magnitude of state spending on tourism promotion, temporal correlation and spatial correlation. It also controls for effects within each state that are invariant during the observed period, so-called fixed effects. This type of model is often referred to as a spatial Durbin model in that the specification incorporates the spatial and time autocorrelation features.45 This specification takes the form:
Y is the value of output or income (aggregate or disaggregated by industry, in either levels per capita or growth rates) in state i, in year t across the lower 48 conterminous states. The coefficient a is a common or fixed-effect intercept; X is real state tourism spending in state i, in year t; and, χ2i,t its squared value. This and the dependent variable is weighted in state j by the spatial weights matrix W̃, which is the first order contiguity matrix of states. The coefficient on this variable is ρ. The mxz space-recursive vector is denoted by θ, and δ is the usual temporal autoregressive coefficient, the optimal lags of which were selected by minimization of the Akaike Information Criterion.46 The disturbance term e, is considered to be iid×N(Ο,σ2).
In practice then the spatial component W̃Xj,t will be the mean value of the dependent variable in the contiguous counties in each state in the current time period t. The space-recursive vector is this variable in z lags.[*] We also include a time trend θΤ specifically to account for changes in household consumption pattern over the sample period. For example, real personal consumption expenditures on amusement parks more than doubled in per capita terms during the sample period.
For some specifications the dependent variables are demeaned in a process outlined in a 2005 paper by Pesaran to remove spatial heterogeneity which might not be captured in the spatial Durbin model. This is a cautionary step which will be applied to some of the estimates of tourism in levels. This step is motivated by the common fact that we do not know the source of spatial influence across states.
State government support for tourism expressed in direct marketing and support for development projects also offers some endogeneity concerns. States with a larger tourism sector might face stronger lobby efforts to engage in tourism promotion spending. This is not likely a large concern given that every state in our sample has, at some time, engaged in tourism spending. Nevertheless, the model employs two methods to address this concern. The first substitutes the per capita measures of tourism-related economic activity to growth rates. The second identifies what accounts for the likelihood that existing natural features in each state may influence support for state-funded, tourism-related development. To do this, the model estimates the relationship between state tourism spending in each state and the elevation span (distance between the tallest and lowest geographic point in each state). This offers the following relationship:
Χi,t = α + ψi + ei,t (Equation 2)
Χi,t — ei,t is the adjusted value of state tourism is related spending in each year t, in each state i.
The selection of an appropriate pooled econometric model is subject to ongoing debate. The chief concerns are the choice of using fixed effects or random effects. There is also some concern that fixed-effect estimates may not be compatible with spatial analysis. A spatial Durbin model, however, offers a method for analyzing spatial impacts in a fixed-effect setting. This is the strategy employed in this model, though it might limit the ability to directly extend this analysis, such as interacting tourism spending with the presence of a statewide hotel tax. Since there are other estimation procedures for use in such a setting, this is not likely to seriously limit the extension of this research.
Incomes in each of the affected sectors represent a good choice for policy considerations related to state tourism promotion spending. The impact of state tourism spending on sectional, tourismrelated economic activity can be deduced from these. This will be followed by tests that evaluate the sectorial adjustment within the counties where data permits. These disaggregated tests, however, are limited due to the lack of available data, specifically in rural counties.
These data are collected from the Bureau of Economic Analysis Regional Economic Information System. Data from state tourism promotion spending are reported in Graphic 2.
The model was populated with these data and analyzed with Equation 1 for hotel income, amusement and recreation income, and arts and entertainment income separately, since each of these sectors represent the core of tourism-related economic activity. While other sectors benefit from tourism activity, such as agriculture, travel services, retail sale of petroleum products, these are likely to be far more diffusely impacted by tourism spending than sectors that are directly influenced by visitor spending. In that sense, this analysis cannot be a complete benefit-cost analysis of state tourism spending which might require a general equilibrium modeling approach.[†]
The availability of data offer some opportunities and limitations. By employing both state GSP and personal incomes by sector during the observation period, the aggregate effects of tourism GSP can be isolated and analyzed as to whether those effects are transmitted to other inputs, such as labor, rent and proprietors’ income in these sectors. Ideally, capital, rents and profitability in these sectors, by state, would be estimated, but those data are unavailable. Also, national income accounts from the Standard Industrial Classification to North American Industrial Classification System were modified during the period of study. This reclassification imposed little change on hotel and accommodations and amusement and recreation data, but did impact the availability of arts and entertainment data. A NAICS binary variable to control for this accounting change was included, but GSP-to-income comparison for the arts and entertainment sector was not made, because this seems to be the most problematic of nonfarm transitions to NAICS.
Graphic 3 below reports results from estimates from the hotel and accommodations sector. Standard errors have been treated with White’s heteroskedasticity invariant, variance-covariance matrix. The major effects in levels (with Pesaran adjustments), growth rates and growth rates adjusted for endogeneity are reported, as are interpretation of the results and estimates on the impact of state-funded tourism promotion spending on hotels and accommodations. The asterisks mark the level of statistical significance for these figures, using the standard designation.
In general, the overall results of these models are the most informative. State-funded tourism promotion spending in neighboring states is largely not a factor in hotel and accommodations income nor GSP, as evidenced by the weak statistical significance of the spatial tourism spending coefficients. This holds for the nonlinear term (squared value), though in incomes it does become modestly statistically significant. The spatial and temporal autocorrelation terms are both statistically meaningful and offer reasonable magnitude and signs to hotel and accommodations GSP and incomes in each state.
The direct assessment of the GSP and income impact of state-funded tourism promotion spending finds that it boosts GSP in hotels and accommodations. This suggests that this model is sufficiently sensitive to detect impacts of this spending. However, the magnitude of the effect is small, with an elasticity of hotel and accommodation GSP to state tourism spending of roughly 0.02. To place this in context, the estimates indicate that a $1 million increase in state-funded tourism promotion spending in the average state during this period would have resulted in about a $20,000 increase in GSP in the hotels and accommodations industry. The data do not clearly specify the type of spending in a way that allow us to evaluate the elasticity of advertising or programmatic analysis, but it is clear that in aggregate, state-funded tourism promotion has a very small impact upon hotel and accommodation GSP, which is an important component of the overall tourism industry.
The two models analyzing growth rates yield no evidence that state-funded tourism promotion is influencing GSP growth in hotels or accommodations. This result holds when corrected for endogeneity, as described above.
Of interest also is the divergent results from GSP and personal incomes. While there is a very small impact of state-funded tourism promotion on GSP, there is no impact on hotel and accommodations income. So, the GSP growth generated by tourism promotion does not transmit to labor earnings through higher wages and salaries for hotel and accommodations employees. As previously noted, data are not available to more precisely estimate which factors of production (profit, interest or rent) these GSP increases flow to, but it appears that they do not accrue to labor.
This examination of hotel and accommodation sector GSP and income suggests a very small impact of state-funded tourism promotion on average across the 48 contiguous states. However, it still could be the case that there is an impact on other tourism-related industries, such as amusement and recreation or arts and entertainment. Graphic 4 reports results from estimates from the amusement and recreation sector. Standard errors have been treated with White’s (1980) heteroskedasticity invariant, variance-covariance matrix. As with the hotel and accommodation models, there is no evidence of shared unit root in these data.
Further, the full models repeat the sensitivity of amusement and recreation to state tourism spending as does the first model. Unlike the first model, state-funded tourism promotion is not correlated with changes in either GSP or income in the recreation and amusements sector. Thus, the low magnitude of impacts detected in the first models on hotels and accommodations finds even less impact on the second tests for amusement and recreation.
Graphic 5 below reports results from estimates based on arts and entertainment data. Because of the NAICS reclassifications, however, only the impact of state-funded tourism promotion on arts and entertainment income could be assessed. Standard errors have been treated with White’s  heteroskedasticity invariant, variance-covariance matrix, and again these data do not share a common unit root.
In this test of arts and entertainment income, the estimate in levels shows no statistically significant impact of state-funded tourism spending. However, in the growth model without endogeneity adjustments, there is a statistically significant, but economically miniscule impact on arts income. Calculating the impact in year one, spending $1,000,000 in additional state tourism promotion would increase total art and entertainment income by less than $35,000 in the average state. These results are statistically significant, but are far below the threshold of economic consequence.
In this one model of income growth in arts and entertainment, 13 states showed impacts for tourism spending that were statistically different from zero. All were negative, suggesting much of the observed negative impact tourism spending had on arts and entertainment was limited to those states, though another dozen more showed impacts that were close to generally accepted levels (10 percent).
[*]Spatial autocorrelation may occur in noncontiguous counties. So, it is possible to correct for the influence of distance using such measures as county centroid distances in a GIS setting. In practice, political boundaries are most commonly used since this is where data is collected. Also, estimating the influence of space in more distant counties increases calculation costs significantly.
[†]For the sector list and a CGQ modeling approach to tourism, see: Adam Blake and Jonathan Gillham, “A Multi-Regional CGE Model of Tourism in Spain” (Brussels: European Trade Study Group, Sept. 2001).