Determinants of International Tourism

Contributor Notes

Author’s E-Mail

The paper estimates the impact of macroeconomic supply- and demand-side determinants of tourism, one of the largest components of services exports globally, and the backbone of many smaller economies. It applies the gravity model to a large dataset comprising the full universe of bilateral tourism flows spanning over a decade. The results show that the gravity model explains tourism flows better than goods trade for equivalent specifications. The elasticity of tourism with respect to GDP of the origin (importing) country is lower than for goods trade. Tourism flows respond strongly to changes in the destination country’s real exchange rate, along both extensive (tourist arrivals) and intensive (duration of stay) margins. OECD countries generally exhibit higher elasticties with respect to economic variables (GDPs of the two economies, real exchange rate, bilateral trade) due to the larger share of business travel. Tourism to small islands is less sensitive to changes in the country’s real exchange rate, but more susceptible to the introduction/removal of direct flights.


The paper estimates the impact of macroeconomic supply- and demand-side determinants of tourism, one of the largest components of services exports globally, and the backbone of many smaller economies. It applies the gravity model to a large dataset comprising the full universe of bilateral tourism flows spanning over a decade. The results show that the gravity model explains tourism flows better than goods trade for equivalent specifications. The elasticity of tourism with respect to GDP of the origin (importing) country is lower than for goods trade. Tourism flows respond strongly to changes in the destination country’s real exchange rate, along both extensive (tourist arrivals) and intensive (duration of stay) margins. OECD countries generally exhibit higher elasticties with respect to economic variables (GDPs of the two economies, real exchange rate, bilateral trade) due to the larger share of business travel. Tourism to small islands is less sensitive to changes in the country’s real exchange rate, but more susceptible to the introduction/removal of direct flights.

I. Introduction

A record one billion tourists crossed international borders in 2012, an event that was celebrated by the United Nations World Tourism Organization (UNWTO). The World Travel & Tourism Council (WTTC) estimates that travel and tourism accounts for 9 percent of global GDP from “direct and indirect activities combined”. Even when measured more conservatively, international tourism is a major source of income and cross-country linkages.

In 2006–10, international tourism receipts represented about 6 percent of international trade of goods and services, and nearly 2 percent of the world’s GDP.2 In comparison, international trade in fuels accounts for 10 percent of total trade, while international remittances stand at ¾ percent of the world’s GDP.3 Global averages conceal the importance of international tourism for many economies. In 45 of the 191 countries and territories for which 2009 data was available, international tourism accounted for over 20 percent of total exports. And while the term “tourism-dependent economy” may first evoke images of paradise islands, tourism plays an important role for countries spanning the full spectrum of economic size and development. A number of OECD countries (Australia, Greece, New Zeeland, Portugal, Spain, Turkey) derive 14 to 25 percent of export earnings from foreign tourists, and even the extremely diversified economies of US, France and Italy are in the 8 to 10 percent range.4

Given the magnitude of tourism flows and the importance of the industry for the global economy, it would only be natural to investigate the drivers of tourism flows and the cross-border spillovers generated by these flows. To date, most studies on the subject attempt to answer these questions using small datasets covering one or a small group of tourism-dependent countries.5 There are very few studies that analyze the universe of international tourism flows. This is in stark contrast to other components of the balance of payments. Goods trade, FDI and, more recently, remittances have been analyzed—usually in the framework of gravity models—using rich datasets spanning decades and encompassing most countries in the world. There are prolific streams of research focusing on important categories of goods trade (manufactured goods, fuels, other natural resources), with recent literature going as far as literally dissecting flows at the product level.

An important reason why tourism has not garnered the same level of attention is scarcity of cross-country data. This paper fills this gap by analyzing the determinants of international tourism flows and cross-border spillovers associated with tourism using an extensive panel dataset, constructed on the basis of UNWTO raw data, covering the universe of bilateral tourism flows over a decade. The paper makes several contributions to the literature on international tourism and the larger body of empirical research using the gravity model approach. The study presents results for equivalent specifications for tourism and merchandise trade, allowing for direct comparison between the two. It undertakes an in-depth analysis on the impact of the real exchange rate on tourism flows, and proposes a more robust way for introducing bilateral trade flows as an independent variable in the gravity estimations for non-merchandise flows. The paper estimates the impact on tourism of a number of non-traditional variables, such as presence of a direct flight, hotel rooms, climate and touristic attractions, and conflict magnitude. Finally, unlike previous studies which only focused on tourist arrivals, the paper also analyzes the determinants of the average length of stay. The remainder of this section summarizes the main results of the paper.

The gravity equation performs very well at explaining bilateral tourism arrivals. For some equivalent specifications, the fit of tourism gravity equations is in fact better than that for goods trade, despite the fact that the dataset is twice smaller. The fit is also considerably better than the one reported in the literature for FDI and remittances flows.

Macroeconomic variables and economic ties have a large impact on tourism arrivals. The elasticities of bilateral tourism with respect to the GDPs of the origin and destination countries are large, although smaller than the near-unitary elasticity estimated for goods trade. Strong trade ties are positively associated with higher tourism flows, a result that highlights the considerable share of business travelers in measured tourism flows; especially so in the case of intra-OECD travel. The real exchange rate has the expected effect: an appreciation of the origin’s currency increases bilateral tourism, while the appreciation of the destination reduces it. This estimated elasticity is large and robust to a variety of estimation techniques and measures of the real exchange rate.

Traditional gravity variables have mostly the expected impact on tourism flows. Distance between the two countries has a negative impact on tourism with an elasticity that is nearly identical to that for merchandise flows. Language ties are more important for tourism than merchandise trade, while historical colonial relationships are less important. The presence of regional trade agreements between the two countries is associated with a small positive effect on tourism. Common currency is in fact associated with a reduction in both tourism and merchandise flows, a result that is fully driven by Eurozone countries. Unlike in the case of merchandise trade, economic remoteness of the destination country is associated with higher tourism flows, suggesting a premium placed by tourists on destinations off the beaten path.

The effect of two supply-side variables is analyzed: presence of direct flights and number of hotel rooms in the destination country. The presence of a direct flight is positively associated with tourism flows, but reverse causality dominates—a direct flight is added the year after bilateral tourism flows see an increase. The same positive relationship is observed between number of hotel rooms in the destination country and tourism inflows, but in this case reverse causality does not seem to play a role. Both supply variables are less important for intra-OECD tourism flows, suggesting that supply factors are not binding constraints to tourism in developed countries.

Several non-traditional variables are examined on the demand side. Tourists generally prefer travelling to regions with similar climates, but there is also a strong preference for travelling to warmer countries. Even after controlling for distance, time difference has a negative impact on tourism flows, suggesting that jet lag is a concern entering into the decision process. “Cultural capital” (as proxied by the number of UNESCO World Heritage sites) also plays a role in explaining tourism flows. Tourists avoid countries with ongoing conflicts.

Using a smaller dataset on tourist-nights, the study shows that tourism flows respond to changes in tourism determinants both through changes in the number of tourism arrivals, as well with a change in the average duration of stay. In particular, the real effective exchange rate has a strong effect on duration of stay—the real appreciation in the destination country is associated with both fewer tourists and shorter stays.

Finally, the study revisits the main findings for the case of small islands, many of which are tourism-dependent. In particular, it finds that—unlike in the case of other countries—the small island’s own real exchange rate has little impact on tourism arrivals.

The paper is organized as follows. Section II discusses the data and presents some summary statistics, Section III discusses the empirical strategy, Section IV presents estimation results, and Section V concludes.

II. Data

The smaller volume of research on trade in tourism services is in part due to the lack of a true equivalent to the Comtrade database for tourism flows. In the case of Comtrade (IMF’s Direction of Trade Statistics (DOTS), CEPII’s BACI are derived from the same source), the quality of data on merchandise trade is ensured by the fact that variables of interest—quantity, value, origin or destination—are (i) collected in a centralized fashion at customs, (ii) codified using standardized nomenclatures (Harmonized System, SITC, etc.) and (iii) aggregated using standardized information systems (e.g., some flavor of ASYCUDA). Moreover, aggregators of trade in goods have the advantage of seeing the same transaction through recorder by both importing and exporting country, which allows for ex-post adjustments and corrections (as done in BACI).

Despite ongoing efforts by international organizations, such uniformity has not yet been achieved in the case of statistics of bilateral tourism flows. Countries don’t record the destination of outbound tourists, so bilateral tourism data only comes from the destination country (the equivalent of the exporter in goods trade). Most countries measure tourist arrivals at the border, but some measure arrivals to hotels. Country practices differ in terms of who is counted (tourists or visitors6) and how they determine their country of origin (by nationality or residence). Important variables of interest—spending over the duration of the stay—are rarely collected and reported. The number of tourist-nights is only collected by about 40 percent of all countries.7 Finally, many statistical offices of destination countries choose to group visitors from some countries into categories (e.g., “Other Africa”, “Benelux”, “other CIS”), which makes it sometimes impossible to positively identify the country of origin.

Despite heterogeneity in tourism data collected at the national level, UNWTO aggregates and publishes an annual compendium of all bilateral tourism flows. This data has seen relatively little use since the format in which data is published is not easily convertible into a standard panel dataset.8 The electronic data is distributed in five year blocks, two of which were available, one for 1999–2004 and another one for 2005–2009.9 The definition used in the collection of data is specified for each country: whether the data is collected at the border or at hotel establishments, whether it is a measure of tourists or visitors, and whether the origin country is determined based on nationality or residence.

Table 1.

Dataset Summary Statistics

article image

The resulting unbalanced panel dataset contains 128,304 observations, each observation containing the number of tourist arriving from the origin country to the destination country in a given year.10 Of these, about 22 percent o observations had to be dropped because the country of origin was not unambiguously identified (8 percent) or because at least one key explanatory variable (GDPs, bilateral distance) was not available (14 percent). However, the impact on the share of tourism flows captured was smaller, as these observations accounted for only 11 percent of all tourism arrivals over the 10 years of data. To reduce noise in fixed effects and first differences regressions, I also eliminated all country-pairs for which the number of tourists from one country to the other in at least one year fell below 100.11

For the purpose of this study, the UNWTO dataset was then supplemented with a series of “gravity variables”. Macroeconomic indicators come from WDI, the IFS and the Penn World Tables (PWT). In particular, PWT provides the widest country coverage and is therefore the preferred source of GDP data. Bilateral trade statistics comes from IMF’s Direction of Trade Statistics (DOTS). A number of CEPII datasets provide standard gravity model variables, including bilateral distances, cultural, linguistic and colonial relationships. CEPII was also the source of data on common trade agreements and common currencies.12

Several standard measures of real exchange rates are used: (i) a bilateral real exchange rate (RER) computed from the IFS, (ii) real effective exchange rates (REER) of origin and destination countries reported by the IFS, (iii) PWT’s PPP factors for origin and destination countries. The latter is a bilateral exchange rate in purchasing-power-parity (PPP) terms vis-à-vis the United States, i.e. it measures the amount of U.S. dollars that buys the same basket of goods that $1 bought in the U.S.13 Imposing the restriction that changes in PPP factors in origin and destination have the same impact on tourism flows, a fourth measure of real exchange rate is introduced—the PPP factor ratio of the PPP factor of the destination to the PPP factor of the origin (an increase in the PPP factor indicates an appreciation—in dollar terms—of the destination vis-à-vis the origin).14 Each of these measures has advantages. The bilateral real exchange rate is the most commonly used measure in the gravity literature but, by definition, it does not incorporate the effects of exchange rate movements vis-à-vis third countries. The REER appears to be the preferred option as it accounts for third-party trading partners. However, the weights used in the REER are heavily slanted towards merchandise trade, and therefore may be less relevant for the analysis of trade in tourism, especially in the case of tourism-dependent economies (e.g., small islands). The PWT-sourced measures benchmark country-pairs to a single reference country, which conceptually places it between bilateral RER and REER. The PPP factor ratio is the preferred measure through most of the paper with the exception of section IV.D, which specifically deals with the impact of real exchange rates on tourism.

I also introduce several non-standard gravity variables. First of all, I constructed a bilateral “climate similarity index”, using a dataset published by Portland State University and based on the world climate map using the Koppen-Geiger classification.15 The dataset shows the distribution of area and population (as of 1995) of each country across various climate zones. The climate similarity index varies between 0 (climates of the two country have no overlap) and 1 (countries have exactly the same climate zone composition), and is constructed in the same manner as the export similarity index, introduced by Finger and Kreinin (1979). Specifically, Simi,j = ∑c Min(Zoneci, Zonecj), where Zoneci and Zonecj are the shares of climate zone c in countries i and j respectively.

Data on global bilateral passenger air travel flows was supplied by Diio, an aviation business intelligence company. Data on the number of hotel beds came from the UNWTO dataset itself. The list of World Heritage sites was downloaded from the UNESCO website. Data on conflicts comes from the Political Instability Task Force (PITF) and from the UCDP/PRIO Armed Conflict Dataset.

III. Empirical Strategy

The gravity model framework has been well tested and scrutinized during five decades of research on bilateral merchandise trade and, more recently, on other cross-border economic flows (FDI and remittances, in particular). It has been empirically established that trade between two countries is proportional to the economic size of the two countries and inversely proportional to a number of “trade resistance factors”, chief among them transportation costs, which is usually proxied by the distance between the two countries. Other resistance factors often included in gravity model regressions are linguistic characteristics (countries speaking the same language trade more), historical ties (countries trade more if they have ever been in a colonial relationship or were ever part of a single country), common border (the “average” distance between the two countries becomes a misleading measure of transportation costs when border regions face effectively zero distance), membership in trade agreements and monetary unions. It is reasonable to assume that many of the same factors influence tourism flows.

The preferred specification of the gravity equation has gone through a number of changes over the years. The traditional regression specification used to estimate the gravity equation for tourism in the case of a panel dataset would take the following form:


where Todt is a measure of the tourism flow from country of origin o (importing country) to destination d (exporting country) in year t, Yot and Ydt are the gross domestic products (measured in constant US$) of the origin and destination country respectively, Dod is the distance between the two countries, Xodt is a 1 × k vector of other variables proxying other resistance factors; and ηt is a set of T year dummies capturing common time effects.

As pointed out by Anderson and van Wincoop (2003), this specification suffers from omitted variable bias. It only accounts for the individual characteristics of o and d, and doesn’t recognize the fact that the flows from o to d also depend on the attractiveness of going from o to d compared to going from o to any other destination. In short, bilateral flows depend on multilateral parameters. The standard (and simplest) econometric approach for dealing with “multilateral resistance” is to introduce dummies for origin and for destination countries.16 The estimated regression then becomes:


where ωo and δd are origin and destination dummy variables. Note that this specification makes it impossible to estimate the coefficient on time-invariant country characteristics, such as geographical characteristics (exit to sea, climate zone, etc.). There is still scope for omitted variable bias in this specification. For example, the dataset holds no data on bilateral visa regimes, which can bias the estimated coefficient if the visa regime that a country faces is correlated with other regressors (e.g., own GDP level). This can be addressed by using a fixed effects specification where the panel variable is the country-pair or, equivalently, introducing country-pair dummies φod17:


Note that the introduction of origin-destination dummies makes it impossible to estimate coefficients on time-invariant variables such as distance, common cultural and historic ties. This regression will produce unbiased results under the assumption that the country’s multilateral resistance is constant over time (and therefore fully taken care of by country fixed effects).

Bayoumi and Eichengreen (1999) propose a first-differences specification for estimating the gravity equation, which produces unbiased results if disturbances follow a random walk, and which is particularly useful for investigating the impact of real exchange rates on trade flows:


The inability of the fixed effects and first differences specifications to estimate the impact of time-invariant variables on tourism can be addressed in two ways. First, one could use a random effects specification (with the country-pair as the panel unit). The main problem with a random effects specification is the fact that it doesn’t account for multilateral resistance. A partial solution is to proxy it with a measure of economic remoteness, which is a GDP-weighted average of the distance to all other countries.18 The second approach is to use the Hausman and Taylor (1981) estimator, which allows estimating coefficients on time-invariant variables by imposing assumptions on the endo-/exogeneity of each variable. This approach has been used in the gravity literature, among others, by Serlenga and Shin (2004).

Estimation results below rely on most estimation techniques discussed above. Baseline results use country fixed effects (CFE, equation 2) and country-pair fixed effects (CPFE, equation 3). Random effects (RE) and the Hausman-Taylor (HT) estimators allow analyzing a wider set of determinants, while the first differences specification (FD, equation 4) is well suited to estimate key macro-determinants (the importance exchange rate in particular).

IV. Estimation Results

A. Tourism Versus Merchandise Trade19

Benchmarking the effect of “traditional” gravity variables on tourism against merchandise trade using fixed effects regressions is a natural starting point for the analysis. I start by comparing the determinants of international tourism with those of merchandise trade using origin and destination fixed effects for tourism and, respectively, importer and exporter fixed effects for trade (“country fixed effects” or CFE for short). This model, which corresponds to equation 2 in Section III., allows the estimation of coefficients for a number of time-invariant variables, while at the same time accounting for multilateral resistance forces. The regressions cover the same time period with the exception of 2004, which is included in the trade regressions, but missing from the tourism dataset.

Two specifications of the gravity equation are estimated: a barebones specification with GDPs and distance, and an extended one with the usual geographical, historical and linguistic controls. In the case of tourism arrivals, the two specifications are estimated for the entire universe of bilateral flows (Table 2 regressions 1 and 2), as well as for the intra-OECD flows only (regressions 3 and 4). Unless indicated otherwise, the discussion focuses on estimates for the extended specification on the full sample (regression 2 for tourism and 6 for trade).

Table 2.

Gravity Equation for Tourism and Trade with Country Fixed Effects

article image
Note: Robust standard errors in parentheses. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

The first notable result is that the gravity equation works equally well for tourism and merchandise trade, as the R2 is roughly the same for comparable specifications and samples. In fact, the fit for tourism regressions is slightly better (regression 1 vs. 5 and 2 vs. 6).20

Tourism and trade exhibit very similar responses to a number of factors. The sign on most “trade resistance” variables is the same. The magnitude of coefficients is also close in many instances. Most notable is the distance between the countries—the elasticities computed for comparable samples/specifications are 1.59 (regression 2) for tourism and 1.54 for trade (regression 6).

An important difference between tourism and goods trade is that the elasticity of tourism arrivals with respect to the origin’s income is much lower than that of goods trade with respect to the importer’s income. This result is robust to the sample used (worldwide vs. intra-OECD) and to the introduction of additional variables in the extended specification.21 Results in regressions 1 and 2 contradict the conventional view that tourism is a superior good, which would imply an income elasticity above 1 for the origin country. The superior good hypothesis can be in principle reconciled with a sub-unit income elasticity of arrivals once it is recognized that growth in the origin country does not accrue uniformly to all residents. The elasticity of tourism spending with respect to origin GDP could still be one or higher if the upper segments of the distribution spend a higher share of their income on international tourism. However, this does not appear to be the case, as Appendix Table 4 shows that for countries with reliable balance of payments statistics the coefficient of arrivals on receipts is in fact 1 or slightly below. Once the sample is restricted to tourism/trade among OECD countries only (representing 51 percent of global tourist flows and 52 percent of global trade), the elasticity climbs to just above one (regression 3 and 4). Still, it remains considerably lower than for goods trade within the same group of countries (regressions 7 and 8), suggesting that tourism is less income-elastic than the average good. Clearly, there is nothing supporting the superior good hypothesis, although a definite answer probably needs to wait until data on bilateral tourism receipts becomes available.

Table 3.

Summary of Regressions on Tourist Arrivals

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively. Additional geographical controls include: Log areas of the two coutnries, origin/destination landlocked dummies and destination small island dummy. Historic and liguistic ties include: Ever in a colonial relationship, comon colinizer post 1945, common official or primary language, same language spoken by at least 9%, were the same country. Full results are presented in Table A5.
Table 4.

Gravity Equation for Tourism and Trade with Country-pair Fixed Effects

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

Contiguity is much more beneficial for tourism than for trade—a common border leads to 150 percent larger number of tourists between two countries, compared to only 50 percent in the case of goods flows.22 Tourism also responds stronger to linguistic ties. On the flip side, historical ties (colonial relationship, common colonizer, whether the two countries were ever part of the same country) have a relatively stronger effect on trade than on tourism.

The one troubling result is the negative coefficient estimated for countries sharing a single currency for both tourism and trade flows. This contradicts most findings since Rose (2000) has ignited the debate on the magnitude (but not the sign) of the effect a common currency has on trade.23 Restricting the sample on intra-OECD trade/tourism reverses this finding, but the reported magnitude is still lower than in previous studies.24 Robustness checks show that the negative sign is driven by countries in the Eurozone; for non-Euro currency unions the estimated coefficient varies between 0.3 and 0.6 depending on the specification.

B. Non-traditional Determinants of International Tourism

Some variables which are natural candidates for a tourism gravity equation are time-invariant (geographical remoteness, climate, number of World Heritage sites, time zone difference) and therefore cannot be estimated using fixed-effect specifications. Two specifications discussed above are suitable to estimate the importance of time-invariant determinants of international tourism: country-pair random effects (RE) and the Hausman-Taylor (HT) estimator. It should be noted that both the standard Hausman test and the Arellano (1993) test, which is robust to heteroskedasticity and autocorrelation of arbitrary form, reject the random effects specification. Nevertheless, RE results are presented here, as they serve as a natural comparator for HT results and allow for at least tentative conclusions on the impact of time-invariable variables on tourism. In addition, I augment the CFE specification with a number of country-pair characteristics (time-invariant variables describing relationships between the two countries but not country attributes).

Results for these three estimation techniques are summarized in Table 3. All specifications are estimated using the reduced sample, which excludes all country-pairs for which the flow drops below 100 tourists in at least one year.25

The basic gravity equation with only GDPs of the two countries and the distance between them is estimated in regressions 1 (RE) and 4 (CFE). Origin’s income elasticity is estimated at well below one, a result discussed in previous section and that is reversed for the OECD subsample. Also, even with this extremely limited specification, the equation explains over half of the variance in the case of random effects and over three-quarters in the case of the fixed effects specification.

Regressions 2 (RE), 5 (CFE) and 7 (HT) in Table 3 estimate a fairly typical extended gravity equation specification, which include a number of geographical, historic and linguistic controls. Negative coefficients for the populations of the two countries indicate that residents of rich countries travel more, and that tourists prefer travelling to richer countries. The random effects and HT specifications also allow for measuring the impact of economic remoteness. While the remoteness of the origin country is negative (as expected), both RE and HT find a positive coefficient on the destination’s remoteness–the opposite of the standard finding in the goods trade literature. This suggests that tourists may in fact place a premium on destinations that are “off the beaten path”, i.e. relatively far from larger economic centers. The negative coefficient on the common currency union estimated in fixed effects and HT regression has been discussed in the previous section.

Regressions 3, 6 and 8 add a number of gravity variables not usually featured in the goods trade literature in an attempt to quantify demand and supply forces specifically driving tourism flows.

First, I introduce bilateral goods trade as a proxy for bilateral economic activity and therefore a control for business tourism. Bilateral trade has been previously used as explanatory variable in gravity equation studies pertaining to migration, FDI and tourism.26 However, since goods trade is driven by the very same gravity forces, the introduction of bilateral trade as an explanatory variable for tourism will inevitably bias down the coefficients on gravity variables (GDPs, distance, colonial and language relationships, etc.). To avoid this downward bias, a two-step approach is used. The first step consists in computing the residuals from the trade gravity equation estimated in Table 2 equation 6. These residuals are then used as an independent variable in the tourism gravity equation. As results in regressions 3, 6 and 8 show, trade enters with the expected sign and is highly significant. This confirms that economic relations between two countries represent an important determinant of tourism flows. The introduction of trade flows allows to some extent to control for business travel, ensuring that the estimated GDP elasticities for origin and destination countries apply, in fact, to leisure travel (i.e., “proper” tourism) as well.

Second, I introduce the PPP factor ratio of the destination and origin countries as a measure of the real exchange rate. A one percent real appreciation of the destination country vis-à-vis the origin reduces arrivals by around 0.18 percent as estimated by the country fixed effects regression, with other models providing estimates which are slightly lower (0.16 for the RE regression) or slightly higher (0.2 for the HT regression). The result is significant at the 1 percent confidence interval for all three estimation methods.

I also look at the effect of two supply-side variables. The presence of a direct flight between countries is associated with a large increase in tourism, although the estimates vary significantly depending on methodology: from 20 percent (HT) to 80 percent (CFE). I also find a strong positive correlation between the number of hotel beds and tourist arrivals. However, it is impossible at this stage to ascertain causality in the case of either variable—the airline and hotel industries may be merely responding to increased demand.

Even controlling for distance and other geographical variables, tourism decreases with the difference in time zones, suggesting that jet lag plays a role when choosing a destination. Tourists avoid armed conflicts, as measured by the conflict magnitude in the PRIO dataset.

The number of UNESCO World Heritage sites (WHSs) in both origin and destination countries is introduced as proxy for the stock of “cultural/historical” capital.27 One would expect that tourists are drawn to countries with a larger stock of “cultural capital”, The expected sign on the origin country WHS is less certain. Residents of countries with relatively abundant cultural capital may have a lower impetus to travel abroad. But since they have been already exposed to cultural capital at home, they may also be more interested in discovering it abroad. Using the full sample, residents of origin countries with more WHSs tend to travel more, a result that is robust across specifications. Results for the destination’s cultural capital are mixed: number of WHSs enters positively and is significant in the RE specification (each WHS is associated with 1 percent higher arrivals), but the HT estimator finds no statistically significant effect.

I test whether tourists prefer visiting countries with a different climate than that of their home country. Arguments could be made both ways. If tourists are guided primarily by “love of variety” preferences, then the coefficient on the climate similarity index should be negative. If, however, tourists prefer the familiar (at least in terms of climate), a positive coefficient is to be expected. Results across specifications indicate that the second effect dominates—there is a positive and significant correlation between bilateral tourism and climate similarity.

The remainder of this section discusses robustness checks, results of which are presented in annex tables. Table A5 features random effects results for the full sample (not restricted to country-pairs with over 100 tourists per year), Table A6 looks at intra-OECD tourism, and Table A7 looks at alternative specifications.

Results for the full sample (Table A5 regressions 9 through 11) are broadly in line with baseline results. The coefficient on GDPs is closer to one, but the magnitude and significance of other key variables are broadly unchanged.

Economic variables have much larger effects in the intra-OECD sample (Table A6): GDP elasticities are close to one, the impact of the real exchange rate jumps to nearly 0.3 (from less than 0.2 estimated for the world), the coefficient on bilateral trade is between 0.13 and 0.3 across specifications (up from the 0.05 to 0.2 range for the world). Results for intra-OECD tourism also solve the World Heritage Sites puzzle: each WHS is associated with 2–4 percent higher arrivals to the destination, while the coefficient on origin WHSs is negative.28

Table A7 tests a few additional specifications. Regressions 1 and 3 decompose the currency dummy into a Eurozone dummy and non-Eurozone currency dummy. I find that the negative coefficient on currency union in Table 3 is driven by the Eurozone, while the coefficient for non-Eurozone currency areas is positive and significant. The negative result for the Eurozone may be explained by the relatively higher prevalence of same-day visitors in Schengen countries, at the expense of a relatively smaller number of multi-day visitors. Same-day visitors are not classified as tourists according to UNWTO definition, but nonetheless play a similar economic role. For random effects, an alternative specification for climate variables is tested in regression 3. Although the overall finding remains the same—tourists prefer travelling to countries with a similar climate to their own—some additional relationships can be teased out. Tourists travel less to cold countries and more to warm countries, while tourists from cold countries travel more. As a preview to Section E below, regression 6 decomposes the PPP factor ratio into origin and destination PPP factors for the CFE estimator. The two enter with the expected sign: the origin’s PPP factor enters positively while that of the destination country enters negatively. Both are statistically significant at the 5 percent level. However, the coefficient on the origin PPP factor is about twice larger than on the destination’s PPP factor.

C. Country-pair Fixed Effects and First Differences Regressions

As discussed in section III, country-pair fixed effects and first differences regressions are better equipped to handle multilateral trade resistance at the expense of preventing the estimation of the impact of all time-invariant variables. I revisit results in the previous two sections using these two estimating techniques.

Table 4 compares gravity equation results for tourism and goods trade using country-pair fixed effects (CPFE, corresponding to equation 3 in section III). Results are broadly in line with those in Table 2. Origin’s GDP is less important for tourism than importer’s GDP for goods trade. The hypothesis that tourism is a superior good gains some support when looking at intra-OECD tourism only: the elasticity of tourism with respect to the origin’s GDP is significantly above 1 (regressions 3 and 4), but still remains lower than the importer GDP elasticity estimated for goods trade (regressions 7 and 8). Regional trade agreements (RTA) and currency unions become irrelevant for tourism, but have the expected sign and a high level of significance for goods trade (regression 6). Surprisingly, the results for intra-OECD (regression 8) suggest a negative impact of regional trade agreements on goods trade.

A comparison of reported within-group R2 suggests that the gravity equation may be in fact better at explaining tourism than trade flows. For the full sample, the R-squared is higher for tourism (around 20 percent) than for trade (around 8 percent). However, the ranking flips when comparing intra-OECD flows only.

The impact on bilateral tourism of non-traditional gravity variables is estimated in Table 5 using both CPFE and first-difference (FD, corresponding to equation 4 in section III) regressions. The sample is again restricted to include only country-pairs for which flows exceeded 100 tourists per year in all years.

Table 5.

Regressions on Tourist Arrivals, Country-pair Fixed Effects and First Differences

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

The coefficients obtained from the first differences model for key macroeconomic variables are virtually identical to other estimators. The elasticity with respect to the origin GDP remains in the 0.5–0.6 range; the elasticity with respect to the real exchange rate (as measured by the PPP factor ratio) is around 0.2. Despite the loss of efficiency associated with the first differences estimator, standard errors for these variables remain small.

While CPFE and FD cannot directly estimate the impact of geographical variables, it is possible to measure their impact on the sensitivity of other variables. Table 5 introduces an interaction effect between the origin GDP and the common border dummy. The hypothesis tested is that at least some local border traffic is driven by factors of non-economic nature (e.g., visiting relatives across borders), which should result in a lower sensitivity to GDP variations in the origin country. I find the expected negative sign on the interaction effect, but it is not statistically significant.

A number of variables lose magnitude and/or significance in the FD specification. Changes in bilateral trade are still associated with high (presumably business-related) tourism, but the coefficient is much smaller: 0.01, compared to 0.05 for the CPFE and HT (in Table 3). Common currency and RTA come out as completely irrelevant. The results for origin and destination population growth are somewhat ambiguous,—the signs and magnitude of the coefficients are sensitive to the introduction of additional controls, sample (full vs. intra-OECD) and model (CPFE vs. FD). However, there is relatively strong evidence that tourists prefer richer countries (a negative sign on the destination population).

A comparison of CPFE and FD results sheds light on the importance of the two supply side factors: presence of a direct flight and number of hotel beds. First, coefficients on both of these variables are large and highly significant for the full sample, but small and insignificant for the intra-OECD sample. This suggests that the presence travel routes and accommodation capacity do not represent a binding constraint for developed OECD countries, but both are relevant issues for other countries.

Second, CPFE and FD results can help sort out causality of these variables—does tourism respond to an increase in hotel rooms and addition of flights or do airlines and hotel expand when they observe larger tourism flows? In the case of direct flight, CPFE results for the full sample (regression 2) show the presence of a direct flight is associated with 22 percent higher bilateral tourism flows (computed as exp(0.2)–1), while FD (regression 6) reports that the addition of a flight is associated with a 4 percent increase in tourism. The difference in coefficients could be interpreted in two ways: (i) establishment of a direct route does not significantly boost tourism in the year the route is established, but may be associated with increased tourism over longer time horizons29 or (ii) the causality goes in the opposite direction, with increased tourism leading to the subsequent establishment of direct routes, which is then measured by other econometric models (CPFE or CFE in Table 3). The first hypothesis should produce a positive coefficient the addition of a direct flight in the past, whereas the second hypothesis suggests a positive coefficient for the addition of a flight in the future. Regressions using both forward and lag of direct flight are presented in Table A8, regressions 2 though 4 (regression 1 is identical to regression 6 in Table 5 above). Clearly, both the lag and forward are highly significant, but the coefficient on the forward is higher in magnitude; it is in fact equivalent in magnitude to the coefficient on direct flight at time t.30 In short, while causality goes both ways, the addition of a direct flight follows the increase in bilateral tourism.31

Results for hotel rooms are very different. The coefficient reported by the FD model (Table 5 regression 6) is twice higher than that reported by CPFE (regression 2), suggesting that the short-lived effect of adding new hotel rooms (as measured by same-year elasticity in the FD regression) is larger than the effect over the longer period of time (measured by CPFE). One potential explanation is that new hotels often run deeply discounted opening promotions. Once these promotions end, arrivals drop somewhat. There is little doubt that causality in this case goes from accommodation capacity to tourism, since the decision to add hotel rooms can rarely be implemented within the same year. These conclusions are also corroborated by results regressions 6 through 9 in Table A8, which suggest that the coefficients on hotel rooms in t − 1 and t + 1 are close to zero and not statistically significant.

D. Tourist Arrivals and the Real Exchange Rate

Every result in the preceding section suggests a strong correlation between the bilateral real exchange rate and tourism arrivals. However, most results relied on a single, somewhat unconventional, metric of the real exchange rate—the ratio of PPP factors from the Penn World Tables. Moreover, to achieve parsimony, this metric combines the real exchange rates of both origin and destination countries. From the point of view of tourist destinations, it is the impact of the destination’s exchange rate on tourism that is of most interest. This section test alternative some measures of real exchange rates. The econometric specification of the gravity equation uses the first difference estimator, following Bayoumi and Eichengreen (1999).

Regression 1 in Table 6 serves as the baseline—it utilizes the preferred measure of real exchange rate, the PPP factor ratio—and drops all gravity variables except the GDPs of the two countries.

Table 6.

First Differences Regressions with Alternative Measures of Real Exchange Rate

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

Regression 2 splits the PPP factor ratio into the origin and destination PPP factors. Results are in line with robustness checks discussed at the end of section III.C—the signs of the two exchange rates are in line with expectation and both coefficients are significant at the 1 percent level. However, the origin’s exchange rate is estimated to be three times larger, in absolute magnitude, than that of the destination country (0.266 vs. 0.84).

Regression 3 utilizes a measure of the origin and destination real exchange rate misalignment proposed by Rodrik (2008)—PPP factors adjusted for the Balassa-Samuelson effect.32 Both coefficients and standard errors on the exchange rate variables are identical to those in regression 2 all the way through the third or even fourth decimal. However, the coefficients on the GDPs change—the coefficient on the origin’s GDP goes up (from 0.504 to 0.545), while that on the destination’s GDP goes slightly down (from .82 to .807). This is in line with expectations. Origin GDP and origin’s PPP factor (not adjusted for the Balassa-Samualson effect) are positively correlated, and both are positively correlated with tourism. When both variables are included in the regression, the coefficient on origin GDP is biased downward. The PPP factor misalignment, on the other hand, is stripped of the effect of growth on the real exchange rate. When entered into the regression with together with the GDP, this exchange rate measure no longer exerts a negative bias on the estimated coefficient for the origin GDP. Similarly, the destination’s unadjusted PPP factor and GDP are positively correlated, but have opposite correlations with tourism (negative with the PPP factor, positive with GDP). Therefore, the coefficient on destination GDP estimated in regression 2 is biased upward by the unadjusted real exchange rate, which is then corrected in regression 3.

Regression 4 uses the bilateral real exchange rate between the two countries, computed by adjusting the nominal bilateral exchange rate by the inflation rates in the two countries. Regression 5 uses IFS REER calculations, which weight bilateral real exchange rates by the trade shares (primarily goods trade shares) of trading partners. The results are remarkably close to those obtained in equations 2 and 3 despite fairly different methodologies for computing the exchange rates—the coefficient on origin’s REER is around 0.27, which is nearly three times larger than that of the destination’s REER.

Regressions 6 through 10 reuse the same five measures of real exchange rate, but include a set of additional variables. The introduction of these controls unambiguously raises the magnitude of the impact of real exchange rates on tourism. The elasticity of tourism arrivals with respect to the origin’s real exchange rate is close to 0.3, while the elasticity with respect to the destination’s real exchange rate is around 0.14–0.15.

Table A9 presents the results for the same specifications for intra-OECD tourism flows. Across specifications, the elasticities with respect to the various measures of the real exchange rate are about twice higher than for the full sample. The elasticity with respect to the destination’s own real exchange rate is more than twice larger than for the full sample. As tourists’ consumption basket in these countries is more heavily weighted towards domestic products (partly as a result of the Balassa-Samuelson effect, which makes non-tradables more expensive), changes in the country’s real exchange rate will have a tangible impact on the costs faced by the potential tourist, and therefore on her decision to undertake the trip.33

While the specifications above test the robustness of various measures of real exchange rate, they do not address the possibility of reverse causality. As noted by Rodrik (2008), a conventional instrumental variables approach is ruled out—there are no exogenous regressors that would affect tourism only via the destination’s real exchange route. A strong case could be made for the use of difference or system GMM, introduced by Arellano and Bond 1991 and Blundell and Bond (1998) respectively. I attempted both approaches, but lagged differences/levels turn out to be poor instruments for the real exchange rate, as the Hansen test of overidentifying restrictions was consistently estimated at, or very close to, zero.

However, reverse causality should not be a major issue for the destination’s real exchange rate. If causality went primarily from tourism to exchange rate, one would expect to find a positive sign on the exchange rate. Because this reverse causality channel does in fact operate, the coefficient obtained on the destination’s real exchange rate is likely to underestimate the impact of the real exchange rate on tourism. One could plausibly hypothesize that the coefficient measured for the origin country (around 0.28) provides the upper bound for the true coefficient on the destination’s real exchange rate.

E. Tourist-nights Versus Tourist Arrivals

The analysis above was based on regressions with tourist arrivals as the dependent variable. The UNWTO dataset also provides data on tourist-nights (or visitor-nights), albeit for a smaller sample of countries. Conceptually, tourist-nights are a closer proxy for tourism revenues, and it is possible that tourists adjust their behavior across both the extensive margins (to travel to a particular country or not) and the intensive margin (for how long).

Destination countries for which tourist-night data is available differ along several dimensions from the larger sample of countries for which tourist arrivals data is available. These countries receive more tourists, are richer and, by consequence, more expensive (Table 7).

Table 7.

Destination Countries Reporting Tourist-nights vs. Reporting Tourist Arrivals, 2009

article image

Table 8 presents the results of a subset of first differences specifications from Table 6.34 To correct for t sample bias discussed above, I first estimated the regressions for arrivals, limiting the sample to those observations that also have tourist-nights data (regressions 1 through 3).35 Regressions 4 through 6 present corresponding results for tourist-nights. Tourist-nights are more sensitive to real exchange rate movements in the destination country, but exhibit lower correlations with the destination’s GDP and bilateral trade. The magnitude of the intensive margin can be computed by subtracting the corresponding coefficient on arrivals (the extensive margin) from that on tourist nights. For example, a comparison of coefficients on the destination’s PPP misalignment in regressions 2 and 5, suggest that the intensive margin on the real exchange rate is around 0.41 (0.744–0.333).

Table 8.

First Differences Regressions on Arrivals, Tourist-nights and Average Tourist Stay

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

A more straightforward way to measure the intensive margin is to use as dependent variable the average tourist stay (the ratio of tourist-nights to tourist arrivals); results are presented in regressions 7 through 9. Several findings stand out:

  • The intensive margin of bilateral tourism on account of the origin’s GDP and real exchange rate is nil. When a country becomes richer (whether because of real growth or appreciation), it imports more tourism services by having more residents travelling abroad and/or by having old tourists travel to new destination, not by having tourists travel for longer periods of time to the same destinations.

  • Tourists respond to an increase in the real exchange rate of the destination country by reducing the length of stay. When prices in the destination country go up, some may not travel at all, whereas others cushion the price increase with shorter stays.

  • As destination countries grow, the average duration of the stay goes down. This can be linked to the dominance of business travel over leisure travel for richer countries.

  • The stronger the trade connections, the shorter the duration of stay. This is likely due to the fact that travel for business is generally much shorter than leisure tourism.

Regression 5 suggests that the overall impact of a 1 percentage change in the destination’s real exchange rate on tourism flows exceed 0.7 (as measured by the combined effect on the number of tourists, but also on their length of stay). Regression 6, which focuses on the bilateral real exchange rate, reports an elasticity of 0.45. Both estimates are considerably higher than that the 0.15 estimated by Eichengreen and Gupta (2012) for the impact of real exchange rates on “traditional service exports” which, in their definition, include tourism.

F. Tourism to Small Island States

A cursory look at Table A1 reveals that most of the tourist-dependent economies are small island states. Therefore, policymakers in these countries are particularly interested in identifying and quantifying drivers of tourism. The two main issues is the sensitivity of tourism to changes in origin GDP and changes in exchange rates. Tourism is a primary channel through which these countries are exposed to external shocks and, indeed, many of them have seen sharp drops in tourism arrivals following the global financial crisis. The impact on tourism of the exchange rate (both the regime and the particular level) is a subject of perennial policy debates in these countries. Results presented above quantify the magnitude of these forces for the entire world (and the OECD sample). However, the very large exposure of small islands to tourism combined with a very specific economic structure of these countries warrant a separate look.

Table 9 presents results of first differences regressions similar to those in section E, but adding several interaction effects with the “small island destination” dummy.36

Table 9.

First Differences Regressions on Tourist Arrivals with Small Islands Interaction Effect

article image
Note: Standard errors in parentheses are robust and clustered by country-pair. *, **, *** indicate that coefficients are significant at 10, 5 and 1 percent levels respectively.

The elasticities of tourism arrivals with respect to origin GDP are slightly smaller than for non-islands (regression 1), but the difference falls within the confidence interval.

Measures of the real effective exchange rate that do not discriminate between the origin and the destination (the PPP factor ratio in regression 2 and the bilateral real exchange rate in regression 3) could be interpreted as suggesting that tourism to small islands is more price-sensitive—the coefficients on the interaction terms of the respective exchange rate measures with the small island dummy are negative (and significant in the case of the bilateral RER). A more in depth examination leads to a different conclusion.

Regressions 4 and 5 disentangle the effect of origin and destination real exchange rates, as measured respectively by the PPP factor misalignment and REER. Small islands are more sensitive to changes in the real exchange rate in the origin country—the coefficients reported for the interaction term with the origin PPP factor are large and highly significant. However, policymakers and the tourism industry in small island states are most interested in the effect on tourism of their own real exchange rate. The coefficient of the corresponding interaction term in regression 4 is positive and highly significant. Computing the marginal effect results in an elasticity of on tourism arrivals to small islands with respect to the destination’s exchange rate of 0.08 (−0.124+0.200), with a standard error of 0.056 (t-stat=1.35). In short, the elasticity is close to nil. Regression 5 reaches the same conclusions using the more traditional REER measure.37 There are at least two potential mechanisms at work. First, it is much harder to engineer a real exchange rate movement for the basket of goods consumed by tourists to small islands, since it is even more import-dominated than that of locals. Second, the heavy reliance of packaged vacations (prices for which are usually set in foreign currency and negotiated by tour operators on an annual basis) limit the extent to which tourists benefit (or lose) from real exchange movements. This does not mean that a real exchange movement would not affect the current account of a small island in the expected direction, but in the short run, most of the impact will likely come from the import side, not tourism exports. In the longer run, however, a depreciated currency should—at least in theory—help expand tourism-related services (i.e., increase the domestic component of the tourist’s consumption basket).

A final point is that small islands are more susceptible to the addition/removal of direct flights than other destinations. The coefficient on the interaction term with the direct flight in regression 7 is positive and significant, implying a marginal effect of 0.1 (with a standard error of 0.03, significant at the 1 percent level) i.e., the addition or removal of a direct flight is associated with 10 percent higher/lower tourism arrivals from the respective market. This is in line with expectations, as small islands are more dependent on air transportation. It also explains why foreign airlines often obtain advantageous terms with small islands—a threat to terminate direct service may have a stronger impact on tourism and the economy.38

V. Conclusions and Policy Implications

The paper uses the gravity model to analyze the impact of dozens of variables on tourism; this section focuses only on a few key findings.

The gravity model does an excellent job at explaining tourism flows, often explaining a higher share of variation than equivalent specifications for international goods flows.

Tourism mostly responds in expected ways to standard gravity variables, although their relative importance often differs compared to goods trade. Most importantly, tourism flows exhibit a lower elasticity with respect to origin country GDP—around 0.6 compared to the unit elasticity commonly found (and confirmed here) for goods trade. The elasticity climbs back to around one for intra-OECD tourism, but remains lower than the one measured for goods trade. In short, the paper finds little support for the view that tourism is a superior good. Results from regressions on tourist-nights also suggest that tourists do not adjust their duration of stay in response to changes in real income.

Tourism and trade also go together—bilateral tourism is high where bilateral trade is higher than the gravity equation predicts. The relationship is particularly strong for intra-OECD flows, suggesting that a larger share of travel within these countries is driven by business travel.

One of the main objectives of the paper was to quantify the relationship between tourism and the real exchange rate—a topic that is hotly disputed in a number of tourism-dependent countries. First, tourism does react strongly to change in the real exchange rate, regardless of the chosen measure (bilateral, multilateral or with respect to a third country). The elasticity of tourism arrivals with respect to bilateral exchange rate is around 0.2. When focusing on the effect of the destination’s real exchange, the elasticity drops somewhat to around 0.15, i.e. a ten percent real depreciation is associated with a 1.5 percent increase in tourism arrivals. Using the set of destinations that also report data on tourist-nights, the paper finds that tourists also respond to changes in real exchange by adjusting the duration of stay. This as much as doubles the overall elasticity of tourism (when measured in tourist-nights) with respect to the destination’s real exchange rate. The magnitude of the impact varies across countries: intra-OECD tourism is much more sensitive to the real exchange rate (the elasticity w.r.t. the destination’s real exchange rate is around 0.35–0.4 after controlling for bilateral trade flows), while small islands exhibit a negligible response of tourism flows to changes in their own real exchange rate.

The paper looked at two supply side variables. The presence of a direct flight is associated with a large increase in bilateral tourism. However, much of this effect is driven by reverse causality—new flights are established wherever bilateral tourism increases. Reverse causality does not seem to affect the similarly strong correlation between tourism and number of hotel rooms. Again, the effect of these variables differs by country group: tourism to OECD countries is not capacity-constrained (both in terms of air connections and accommodations), whereas small islands exhibit a larger response to the addition/removal of a direct flight.

The policy implications are particularly strong with respect to market diversification and the exchange rate. In general, the lower sensitivity of tourism arrivals to GDP swings in origin countries can be a blessing if origin countries are experiencing a downturn. At the same time, reorienting tourism to fast-growing origin countries (e.g., China) is likely to be the best response to a slump in traditional markets (e.g., Europe in the case of Seychelles). Exchange rates play an important role in driving tourism flows, but the effect is not uniform across countries. In particular, when the basket of goods consumed by tourists and, more broadly, the inputs of the tourism sector are import-heavy (as is largely the case in small island states, for example), a real depreciation is unlikely to meaningfully lower the prices faced by tourists, and therefore attract them in greater numbers. In the long run, the depreciation may help increase tourism receipts even in a small island, as it could spur investment in tourism-related services, but the short-term improvement of the external balance is likely to come from the import side.

Future research could focus on utilizing the complete version of the same dataset (1995–2011). In particular, it could help understand whether tourism respondent differently to the drop in global demand during global financial crisis than during normal times—many small tourism-dependent economies saw larger dips in tourist arrivals than the sub-unit elasticities estimated here would suggest. A longer dataset without a gap year (2004 is missing from the dataset used in this paper) may also help address issues encountered when attempting to estimate dynamic/system GMM models. Future research could also expand the analysis to cover additional factors affecting international tourism—such as visa requirements (relevant for tourism to advanced economies) and natural disasters (relevant for small tourism-dependent countries)—which have not been incorporated here due to data limitations.