Abstract
This paper investigates the impact of tax havens on non-tax haven countries in terms of foreign direct investment (FDI). We analyze the importance of agglomeration effects by including FDI inflow levels in tax havens and capture geographic spillovers by measuring proximity to the nearest tax haven. Our analysis yields several interesting findings. First, using panel data for 142 countries, we find evidence of positive spillovers from tax havens to nearby developing countries, but not to nearby developed countries. Second, restricting our panel to developing countries, we find the positive effect of tax haven FDI on developing countries to be robust. Third, we find that geographic distance matters for financial flows: developing countries which are the closest to a nearby tax haven benefit the most in terms of FDI inflows. This result is robust to accounting for spatial interdependence of FDI.
Acknowledgements
We thank Rossitza Wooster, Tom Fullerton, Dhammika Dharmapala, participants at the Southern Economic Association meetings, University of Texas-El Paso seminar and an anonymous reviewer of this Journal for helpful comments. We also thank Michelle Isenhouer, Josephine Huang, and Graham Veenstra for excellent research assistance.
Notes
1. Due to missing observations, the full sample forms an unbalanced panel framework. Non-tax havens are those countries which are not designated as a tax haven by Dharmapala and Hines (Citation2009). Countries are classified as developed and developing countries according to the World Bank’s Country and Lending Groups classification as of September of 2010 (World Bank Citation2010). See Online Appendix for details.
2. We thank the referee for pointing out this important point.
3. Dharmapala and Hines include countries with low business tax rates that have been identified as tax havens by multiple authorities, including the IMF, OECD, and previous research. See Table A1 in the Online Appendix for details regarding closest tax havens.
4. Miscoding Ireland as a tax haven is not likely to have a big effect on our estimates given that it does not lie in close proximity to developing countries. We thank the editor for noting this important issue.
5. Time dummy coefficients are not included for brevity. Country dummies are omitted because several independent variables are time invariant. The explanatory control for specific country characteristics and the cross-sectional panel-corrected standard errors account for cross-country variation in the error term.
6. Table A3 in the Online Appendix describes the variables and their sources.
7. Our exchange rate variable is the local currency unit per US dollar. The positive sign of the exchange rate variable suggests that FDI increases as currency depreciates (you need more of local currency to buy US dollars).
8. In some cases, linear interpolation was used to ensure that there were no gaps in the series, allowing us to estimate our model in a balanced panel.
9. All the variations of the model in were estimated using a variable that included FDI to nearest tax haven weighted by the distance. Previous results are robust to including the distance weighted tax haven FDI indicator. These results are not included but are available upon request.
10. Results for the endogeneity test are given in the Online Appendix. The estimations for the endogeneity test are unbalanced due to missing observations for the instrumental variable (natural log of FDI ouflows in US dollars, truncated for negative values with a value close to zero).
11. See Table A3 in the Online Appendix for an explanation. The standard deviation of the distance to the nearest tax haven category is almost 71 per cent of its mean value. Thus, using half a standard deviation allows for a more equal distribution of countries among groups.
12. It is possible to add the lag of the dependent variable as a regressor. This, however, might generate endogeneity and biased estimates. In results which are not reported, the estimated coefficient for the lag of the dependent variable is positive and statistically significant but the estimated TH_FDI coefficient becomes insignificant.
13. Blonigen et al. (Citation2007) use a spatial lag model to analyse the determinants of FDI and find that there is spatial interdependence.
14. Our MLE estimates show robust standard errors since the estimator of variance uses the Huber/White estimator instead of the traditional calculation.
15. Refer to LeSage (Citation1999) for a discussion on how to estimate the spatial error and lag models and how to test for the presence of spatially correlated errors and spatial autocorrelation.