A. Empirical Strategy

This paper examines the impact of tax treaties on U.S. outbound FDI destined to 78 countries over the period of 2007–2018. In our empirical analysis, we consider the gravity model as a benchmark and augment it with other key predictors. Our conceptual framework can be summarized by the following equation:

where fdi_it represents the volume of U.S. FDI destined to a country i in year t, GRAV_it is the vector of gravity variables, such as GDP (GDP_it) and the physical distance between the U.S. and country i ( dist_i ). KNOW_it represents the extent of knowledge capital of recipient country i , T_COST_it is the bilateral trade cost between the U.S. and country i. Tax_treaty_it is a dummy variable indicating whether a bilateral tax treaty is in effect between i and the U.S. in year t , and Z_it is the vector of other bilateral and multilateral economic agreements in effect at time t . TAX_it represents the vector of tax environment variables for recipient country i .

Among the gravity variables, GDP is expected to be a robust determinant of FDI, as horizontal FDI is often destined to countries that boast large markets and great purchasing power. 5 While physical distance is inarguably a crucial factor that determines trade flows as it is indicative of trade costs, its impact on FDI flows is ambiguous. On one hand, physical distance could be an indicator of costs related to FDI activities, such as transport, communication, market search, among others, implying that distance could affect FDI flows. On the other hand, firms often locate production in direct proximity to a foreign market to avoid distance-related costs. Therefore, all other factors being equal, it is plausible that the greater the distance, the greater the horizontal FDI incentives for firms.

As shown in Equation (1), we consider the degree of knowledge capital of a recipient country as a determinant of FDI flows. Markusen (2007) presents a theoretical model showing that FDI provides knowledge-intensive services to recipient countries for whom developing their own knowledge-intensive inputs would be cost prohibitive. At the same time, the absorptive capacity of recipient countries matters when multinationals decide on a location for FDI. In this paper, we use the relative level of human capital ( HC_it ) and total factor productivity (TFP_it) for country i compared to the U.S. as a proxy for its absorptive capacity.

Our trade cost variable, T_COST_it , is the geometric average of bilateral tariff rates between the U.S. and country i . High tariff rates induce foreign firms to avoid tariff barriers by locating their production within the destination market. This implies that, other things being equal, we can expect a positive relationship between tariff rates and FDI. However, the likelihood of such tariff-jumping FDI is influenced by other factors, including differential production costs, relocation costs and local demand conditions. Furthermore, high tariff rates may also be related to the inclination to protect domestic producers. Therefore, the estimated coefficient of this trade cost variable could be ambiguous.

The vector of Z_it includes dummy variables for tax information exchange agreements (TIEA hereafter), free trade agreements (FTA hereafter) and WTO membership (WTO hereafter). TIEA allows for the exchange of tax information to address harmful tax practices. It can complement tax treaties or can be used by countries for whom taxes on income or profits are low or even zero, making tax treaties inappropriate.6 FTAs often contain investment provisions to foster FDI flows between member countries. More importantly, FTAs and WTO membership can assure foreign investors that recipient countries adhere to global norms in trade and investment practices.

Finally, TAX_it consists of several variables to capture country-specific tax environments. These include tax haven status ( tax_ hvn_i ), transfer pricing rules ( trn_ prc_it ), controlled foreign corporation ( cfc_it ) rules, anti-avoidance regulations ( anti_ avd_it ) and corporate income tax rates ( tax_cp_it ). Tax havens tend to attract a large amount of FDI relative to their market size, especially by enabling multinationals to attain tax rates that are effectively close to zero. These impacts are not confined to tax havens but indeed apply to all countries with which an investing country has a tax treaty. Therefore, tax haven status should definitely be included in the estimation of a tax treaty–FDI nexus. Transfer pricing rules require firms to establish prices based on similar transactions between unrelated parties, and CFC rules prevent the artificial diversion of profits to a related company to minimize tax liabilities. Anti-avoidance regulations are designed to discourage or prevent tax avoidance in advance rather than addressing it after the fact.

Taking the abovementioned discussions into account, our estimation specification is as follows:

where T_t represents the vector of year dummies; ψ, ξ and ω are the vectors of the coefficients; and ε_it is the error term. While we employ several different estimators, including ordinary least squares (OLS hereafter), a fixed effects estimator, and a random effects estimator, we consider the potential impact of time-invariant unobserved heterogeneity across countries using a fixed effects estimator. In addition, as a robustness check, we also employ Arellano and Bond (1991)’s GMM estimation method to control for potential endogeneity bias.

B. Data Description

Our country panel data come from various data sources, including the U.S. Bureau of Economic Analysis (BEA hereafter), the Tax Treaty database, Penn World Tables, the ESCAP-WB trade cost database and CEPII, as presented in Table 1. The set of tax haven countries is created based on Garcia-Bernardo et al. (2017), and the tax environment variables come from Schanz et al. (2017). Our dependent variable is the U.S. outbound stock for each recipient country obtained from the U.S. BEA.7 We use the CEPII dataset as the source of the gravity variables. TFP_it is measured as the PPP-adjusted TFP level relative to that of the U.S. This measure, along with the human capital index ( HC_it ), comes from Penn World Tables 10.0. Moreover, the trade cost variable, proxied by the geometric average of bilateral tariff rates between the U.S. and country i , comes from the ESCAP-WB trade cost database.8

TABLE 1

DATA SOURCES AND DESCRIPTION

For tax-related regulation variables such as anti-avoidance regulations, transfer pricing rules and CFC rules, the higher the value of these variables, the weaker the regulations of recipient country i . For example, if transfer pricing rules are not well established or are not applied appropriately for country i , then the related dummy variable has a value of 1, implying that the likelihood of tax evasion or avoidance increases. Therefore, if the estimated coefficients for these variables are positive for country i , it means that the weaker the relevant regulations are to prevent tax evasion, the greater the amount of U.S. direct investment is into the country. In addition, as countries with low tax rates are more likely to attract FDI, we also include the corporate tax rate ( tax_cp_it ) of recipient country i in the estimation. Schanz et al. (2017) construct this variable in the following way: first, they note the maximum observed tax rate among all the countries in their sample data, after which they subtract each country’s tax rate from this value and divide that by the maximum rate. Thus, the corporate tax rate is normalized to a range between zero and one, with a higher value indicating a more attractive statutory tax rate. We use this variable in our regression and expect its estimated coefficient to be negative.

Defining which countries are tax havens is a complicated challenge. Traditional methods for identifying tax havens are based on differences in the tax and legal structures for base erosion and profit-shifting (BEPS) practices.9 Using this method, the EU includes 12 countries in its 2021 tax haven blacklist, mostly Caribbean and Channel Islands economies.10 Likewise, the OECD defined a total of 35 locations as tax havens in 2000, but by 2017, only Trinidad & Tobago was still listed as a tax haven. Recently, Garcia-Bernardo et al. (2017) identified a larger set of tax havens using an analysis of big data on the ownership networks of 98 million global companies across countries. They identify a total of 55 tax haven countries, including some advanced countries that function as offshore financial centers (OFCs hereafter), such as the Netherlands, the U.K., Switzerland, Ireland and Singapore.11 We adopt the approach of Garcia-Bernardo et al. (2017) to define tax havens.

As our panel data are collected from multiple sources produced by various institutions, there is underlying variation in the data coverage across data sources. As a result, our final panel data contain a total of 78 countries over the period of 2007–2018. The country list is presented in Table 2.12

As of 2018, the U.S. had a total of 60 tax treaties and 32 TIEA in effect. As shown in Table 2, our dataset comprises the majority of the countries that have a tax treaty with the U.S. Meanwhile, many TIEA treaty signatories are excluded from our sample due to the unavailability of information on taxation-related regulations and rules. Another observation is that the lion’s share of these countries consists of tax havens that have neither a tax treaty nor a TIEA with the U.S.

TABLE 2

COUNTRY LIST BY STATUS OF TAX TREATIES AND TIEA WITH U.S.

Note: * indicates a tax haven country as identified by Garcia-Bernardo et al. (2017).

Tables 3 and 4 contain summary statistics and present the correlations among the variables, respectively. Tax treaties are positively correlated with FDI flows, as are TIEA but with a much lower correlation coefficient. The correlations of FDI flows with tax environment variables except for tax haven status are negative, meaning that the better and stronger the tax systems and rules, the larger the amount of FDI inflows. However, because these are simple correlation coefficients, if the regression analysis controls for other determinants of FDI, the relationship between these variables could change. Tax haven status is positively correlated with FDI, while physical distance appears to have less of an impact on investment than it does on international trade. Tax treaties are negatively correlated with TIEA and FTA.

TABLE 3

SUMMARY STATISTICS

TABLE 4

CORRELATION MATRIX

A. Main Results

In this section, we report our estimation results from several different estimators. Table 5 contains the empirical results of the OLS, fixed effects, and random effects models. As presented in Column (1), a negative coefficient for tax treaties emerges when we run the OLS regression only using the gravity and absorptive capacity variables, which is often the case in existing studies (Blonigen and Davis, 2004; 2005). The size of the estimated coefficient becomes smaller if additional economic agreements are included in the estimation, but the negative sign remains, with high statistical significance (Column 2). In this regression, TIEA seemingly increases U.S. FDI into the partner countries. The effect of distance also becomes statistically insignificant. As we expect, the absorptive capacity variables all increase FDI.

TABLE 5

ESTIMATION RESULTS I: OLS, FE AND RE ESTIMATIONS

Note: 1) All regressions include year dummies. 2) Figures in parentheses are heteroscedasticity-robust standard errors. 3) *, ** and *** indicate significance at the 10%, 5% and 1% levels, respectively.

On the other hand, as shown in Column (3), once we control for country-specific tax environments, the negative impact of tax treaties on FDI disappears. This implies that the omission of these variables in the estimation could lead to biased estimated coefficients and create some spurious inference regarding the impact of tax treaties on FDI. Despite the fact that the tax environments of recipient countries, such as their local tax systems and regulations, are inarguably decisive factors in investment decisions, previous studies have failed to consider country-specific tax environments as a determinant of FDI, leading to omitted variable bias in their estimations. Among the tax environment variables, tax haven status appears to be the most decisive factor affecting FDI flows. Other factors being equal, tax havens’ FDI stock from the U.S. tends to be approximately 2.3% higher than that of non–tax havens. The regression results also indicate that the weaker a country’s CFC rules, the higher their FDI inflow from the U.S. One puzzling finding in this regression is the negative coefficient of the corporate tax rate. As mentioned above, a higher value of this variable indicates a more attractive statutory tax rate. Therefore, we can expect a positive coefficient for this variable if a lower tax burden increases FDI inflows from the U.S. However, our estimate finds the exact opposite, with strong statistical significance. We suspect that the OLS regression does not sufficiently control for cross-country characteristics to produce a sensible marginal effect of the independent variables.

One way to control for country-specific heterogeneity in the estimation is to use a fixed effects or random effects estimator. Our panel data allow for the use of these estimators in order to control for time-invariant unobserved characteristics across countries. We report the estimation results using these estimators in Columns (4) and (5). As depicted in Table 5, the impact of tax treaties is statistically significant, with a positive sign for both the fixed effects and random effects estimations. Both models suggest that tax treaties increase FDI stock into a recipient country by about 0.6%. The estimated coefficients of the absorptive capacity variables are now insignificant. Similarly, the effect of FTA is statistically insignificant. Transfer pricing rules, instead of CFC rules, emerge as one of the key variables that determine the magnitude of FDI stock.

While our Hausman test suggests that the fixed effects model is the more appropriate model, the sizes of the estimated coefficients for tax treaties are similar across the two models. A pitfall of the fixed effects estimator, however, is that one cannot examine time-invariant causes of the dependent variables separately, as time-invariant predictors, such as tax haven status and physical distance, are perfectly collinear with the individual fixed effects.

As discussed above, Blonigen and Davis (2004) and Egger et al. (2006) found that recently signed tax treaties tend to decrease FDI flows. They claim that new tax treaties or the revision of old ones may reduce the incentive for FDI, as they contain a more sophisticated incentive scheme for limiting FDI for tax avoidance purposes. Given that our sample contains tax agreements concluded in more recent years, we may find even more negative impacts if this claim is correct.

In Table 6, we report the correlation coefficients between tax treaties and other explanatory variables after classifying tax treaties into three groups based on the time of entry into force of the agreement. As shown in the table, for tax agreements that took effect before the 1990s, the target countries are mainly countries with a large GDP and high levels of productivity and human capital during our sample period. The strong negative correlation with tax environment variables suggests that countries with well-organized domestic regulations to prevent tax avoidance are more likely to have signed tax agreements with the United States.

TABLE 6

CORRELATIONS BETWEEN TAX TREATIES AND OTHER VARIABLES

Note: 1) ^† The period classification is based on the year when the first tax agreement came into force, 2) * indicates the significance at the 1% level after Bonferroni adjustment.

On the other hand, tax treaties that came into force in the 1990s have a slightly lower correlation coefficient with GDP compared to the previous period, and the correlation with productivity and human capital level is not statistically significant. In addition, the correlation with tax haven status appears to have a negative relationship, indicating that there had been a strong tendency to enter into agreements with countries other than tax havens. For tax treaties that took effect after the 2000s, we find a negative correlation with GDP but a positive correlation with human capital. In addition, these treaties have no statistically significant relationships with tax haven status, CFC or transfer pricing rules. Hence, it appears likely that the United States had signed tax treaties with countries with large economies, high productivity and human capital levels, and well-organized tax regulations until the 1980s but that the target countries of tax treaties became more diversified after that decade.

Table 6 also includes correlation coefficients among these variables when the sample is divided into OECD countries and non-OECD countries. The results suggest that the size of GDP shows a positive correlation only in cases of tax treaties with the non-OECD countries. The negative correlation between tax treaties and TIEA in the non-OECD sample implies that the TIEA is used complementarily in countries that do not enter into tax treaties. We also find that the quality of domestic regulations to prevent tax avoidance matters for non-OECD countries. Countries with well-organized tax regulations are more likely to have signed tax treaties with the United States.

Table 7 contains the analysis results estimated by comparing the impacts of recently signed tax treaties with those of older ones. In Columns (1) and (2), we compare the effects of tax treaties that took effect before the 1980s and those that took effect thereafter. In the fixed effects model, for the tax agreements that took effect before the 1980s, the coefficient cannot be estimated because the agreements are perfectly collinear with the fixed effects. On the other hand, according to the random effects estimation results, both new and old tax treaties appear to have a positive effect on FDI in our sample. However, the size of the effect is smaller for the new treaties compared to older ones. Columns (3) and (4) show the results of a re-estimation when the tax agreements are divided into those that went into effect before the 1990s and those that went into effect afterward. These results are qualitatively similar to those shown in columns (1) and (2). Therefore, our results suggest that tax treaties have a positive effect on direct investment regardless of the time they went into effect, but the investment promotion effect is somewhat reduced in the case of newer agreements.

TABLE 7

ESTIMATION RESULTS II: OLD VERSUS NEW TAX TREATIES

Note: 1) All regressions include year dummies. 2) Figures in parentheses are heteroscedasticity-robust standard errors. 3) *, ** and *** indicate significance at the 10%, 5% and 1% levels, respectively.

Table 8 includes the estimates of the effects of tax treaties when the sample is divided into OECD countries and non-OECD countries. Columns (1) and (2) include the results estimated for the entire sample after creating interaction terms between the tax treaty dummy and the OECD country dummy. In this case, the estimated coefficient represents the effect of tax agreements in other groups compared to non-OECD countries without tax treaties.13 According to the random effect analysis, the group of non-OECD countries with tax treaties has, on average, 0.96% more FDI stock invested from the U.S. than the group of non-OECD countries that do not have tax treaties with the U.S. While OECD countries have positive estimated coefficients regardless of whether they have a tax treaty, it should be noted that the size of the estimated coefficient for the OECD group with tax treaties is smaller than that of the OECD group without tax treaties. This implies that among OECD countries, tax agreements may not have a positive effect on FDI inflows from the U.S.

To verify this, after dividing the entire sample into OECD countries and non-OECD countries, we separately estimate each subsample.14 These results are presented in Columns (3) through (6). As shown in Table 8, we find that tax treaties appear to increase FDI among non-OECD countries. The estimated coefficient of tax treaties for the non-OECD sample is statistically significant, and tax treaties appear to increase the FDI stock invested from the U.S. by about 1%. On the other hand, there is no statistically significant effect of tax treaties on FDI in the OECD sample. Our results can be interpreted as evidence, as Baker (2014) suggests, that developed countries have various institutional mechanisms to prevent double taxation other than tax treaties, implying that the net effect of tax agreements may not appear. Brooks and Krever (2015) also argue that tax treaties could be redundant in developed countries, taking into account that most advanced economies have domestic tax laws stipulating either an exemption for tax income derived from other countries or a tax credit for taxes paid in the source country.

TABLE 8

ESTIMATION RESULTS (III): OECD VERSUS THE NON-OECD SAMPLE

Note: 1) All regressions include year dummies. 2) Figures in parentheses are heteroscedasticity-robust standard errors. 3) *, ** and *** indicate significance at the 10%, 5% and 1% levels, respectively, 4) F-test statistic based on the correlated random effects approach.

B. Robustness Check

We perform robustness tests of the regression results in the following way.15 First, it is possible that our analytic results may be affected by endogeneity problems, especially reverse causality bias. While our testing hypothesis is the causal impact of tax treaties on FDI, it may also be likely that tax treaties tend to be signed with countries with large FDI flows. To mitigate this potential bias, we employ Arellano and Bond (1991)’s GMM estimation method. In the meantime, we should recall that our tax treaty variable is a dummy variable and thus the first-differencing of this variable implies that the tax treaty effect can be estimated only for the exact year following the year the tax treaty entered into force. Following Barthel et al. (2010), we overcome this problem by applying an alternative tax treaty variable that measures the total number of years that pass after a tax treaty enters into effect. In this case, the first-differenced variable has a value of one for all years after a tax treaty becomes effective and zero otherwise, which is exactly what we want to test for.

Our GMM results are presented in Table A1 in the Appendix. We confirm that the estimated effect of tax treaties on FDI remains statistically significant and positive, although its statistical significance becomes weaker than in the fixed effect estimation. The Hansen statistics for over-identifying restrictions indicate that the validity of our instruments is not rejected.

Second, we use the approach of Garcia-Bernardo et al. (2017) to define tax havens, which includes several OECD countries. Given that some OECD countries, notably the U.K. and the Netherlands, have more FDI stock invested from the U.S. than others, it is possible that they are outliers in the analysis, affecting the estimation results. Thus, we re-run the regression excluding the Netherlands, the U.K., Switzerland, Ireland and Singapore from the list of tax havens. We confirm that such a change does not greatly affect our results quantitatively or qualitatively.

Third, considering the possibility that the tax prevention regulations have been strengthened in recently revised tax treaties, we conduct a regression analysis based on the time of entry into force of the most recently revised tax treaties. We find that despite these changes, the robustness of our analytic results is preserved. This result is consistent with findings by Davies (2003) showing that treaty renegotiations have no robust impact on FDI.

Fourth, we add more tax-related variables, such as country-specific tax withholding rates for dividends, interest, and royalties, as explanatory variables in the estimation. The estimated coefficients of these variables are largely insignificant. At the same time, our results suggest that a lower tax withholding rate levied on dividends attracts more FDI among the non-OECD sample, while a lower tax withholding rate levied on royalties increases FDI in the OECD sample.

Finally, Blonigen and Davies (2004) employ a slightly different specification compared to ours. In particular, basing on the empirical specification of Carr, Markusen and Maskus (2001), they include the following GDP-related variables in their estimations to distinguish between horizontal and vertical motivations for FDI: the sum of the two countries’ GDPs and the squared difference between the GDPs. Again we confirm that our results remain robust when these variables are taken into account.