The impact of taxes on the extensive and intensive margins of FDI

Given the size of the literature, the surveys of Gresik (2001), Fuest et al. (2005), or Voget (2015) provide useful starting points.

Examples include Hebous et al. (2011), Barrios et al. (2012), Merlo et al. (2016), Behrendt and Wamser (2018), Davies and Killeen (2018), and Lawless et al. (2018). Although these studies vary according to the data used (some use data from a single home country, others for several), measurement of taxes (which includes effective taxes, statutory taxes, and the tax difference between the home and host), and methodology (with conditional logit, nested logit, and mixed logit being employed across and within analyses), the consensus is that taxes tend to lower the probability of a firm choosing a potential host.

Yeaple (2009) examines both margins using US firm-level data. Davies and Kristjánsdóttir (2010) do so for industry-level Icelandic FDI. Neither controls for taxes.

Indeed, in the literature using aggregate FDI data, the home tax is often a “push” factor, i.e., investment often flees a high-tax home.

Using a single host country, Görg and Strobl (2015) examine the effect of Irish taxation on its inbound FDI at the industry level. On average, they find no effect of taxes at the extensive margin excepting that from Germany. Employment growth, their measure of the intensive margin, is meanwhile hampered by higher Irish taxes, but only for pre-existing firms.

A similar approach was used by Behrendt and Wamser (2018) when looking at how firm characteristics influence the effect of bilateral tax treaties on the extensive margin of FDI. Note that unlike Behrendt and Wamser (2018), Egger and Merlo (2011), and Davies et al. (2009), we do not include the existence of a bilateral tax treaty as a control variable. This is because, in our sample of mostly European home and hosts, there is very little variation in the use of treaties.

When estimating the extensive margin, Lawless et al. (2018), Merlo et al. (2016), Behrendt and Wamser (2018) and Davies and Killeen (2018) find firm-driven heterogeneous tax responses. They do so via sample splits and/or the use of mixed logit.

Note that this avoids the well-known problem of zeros in the log–log gravity specification. See Santos Silva and Tenreyro (2006) for discussion.

The idea of a minimum efficient scale is now well recognized in the FDI literature as embodied by seminal work such as Helpman et al. (2004).

This is because in the latter, competition amounts to a second price auction so that the efficient host, which generates the most surplus from investment and therefore offers the highest bid for the firm, hosts in equilibrium.

See e.g., Davies and Eckel (2010) or Krautheim and Schmidt-Eisenlohr (2011).

See Tirole (1988) for discussion.

In the original random utility model of McFadden (1974) on which logit estimators were based, a representative agent was assumed so that there was no variation by, for example, owner or home country.

Examples estimating (only) the extensive location choice include Basile et al. (2008), Merlo et al. (2016), Merz et al. (2017) and Behrendt and Wamser (2018).

Note that this is one parameter per owner, not per owner-investment. Thus, for MNEs with multiple investments this is assumed the same across all of their investment projects.

Alternatively, \(\alpha _{o}\tau _{h,t}\), together with \(\varepsilon _{o,m,h,t}\) can be interpreted as error components so that \(\mu _{o,h,t}=\alpha _{o}\tau _{h,t}+\varepsilon _{o,m,h,t}\). This creates correlations among the unobserved profits for different location alternatives, relaxing the IIA assumption.

Recall that each owner has a single \(\alpha {o}\), hence this approach.

Note that we include subscripts both for the owner o, who may have many affiliates, and an individual affiliate i, with this latter identification necessary as the inverse Mills ratio is affiliate specific.

In unreported results, as an alternative to our combination of mixed logit and OLS, we also used a Heckman two-step approach where sample selection is estimated as part of a maximum likelihood estimation rather than in a two-step fashion. This is the approach used by Davies and Kristjánsdóttir (2010). As with the reported results, we found significant tax effects only in the first stage which estimated the probability of investment. Note that in our multi-host context, however, this probit-based estimation procedure treats each investment decision as an independent choice, in contrast to the explicit comparison of alternatives as in the logit class of estimators. This is why we use mixed logit in the current results. The results of the Heckman estimator are found in the online appendix.

As he used a linear probability model, if Yeaple (2009) included the inverse mills ratio this would amount to including the exclusionary restrictions in the intensive stage.

We use two digit sector dummies. Note that we were unable to use sector-year dummies since 51% of our observations have a single sector-year observation, with a further 20% having only two.

These can be found at http://​www.​bvdinfo.​com.

As can be seen, the distribution of home countries reported in Amadeus has some peculiarities. First, the availability of investments from outside of Europe is limited. Second, two small European countries—Belgium and the Netherlands – appear to be potentially overrepresented. One approach for dealing with this is provided by Kalemli-Ozcan et al. (2019). This, however, requires access to vintages of the Amadeus data which we unfortunately lack. Therefore, we instead resort to sub-samples of the data. As reported in the appendix, when restricting our sample to European homes only, we find results quite similar to those reported. The same goes for when we exclude the Netherlands as a home country. When excluding Belgium, the EATR becomes insignificant although the point estimate remains negative. These additional results are in the online appendix.

These alternative results are available on request.

Thus, we are not considering the wave of corporate inversions via mergers which occurred during the sample period.

Note that there is a decline in investments toward the end of the sample, an issue arising from lagged reporting to the Amadeus data. Excluding 2015 or 2014–2015 has little impact on our results. These alternatives are available in the online appendix.

Note that in unreported results we examine the growth of the affiliates by using their size in year \(t+1\), \(t+2\), and \(t+3\). This does not affect our results and suggests that, if as according to Hartman (1985) firms underinvest to grow via retained earnings, that growth takes some time to manifest itself.

This excludes 524 investments, the bulk of which report zero affiliate assets.

NACE dummies are automatically controlled for in the extensive margin estimates and are included in the intensive estimates. In the online appendix, we report results for three sectors: manufacturing, services, and financial services. The latter two sub-samples confirm the overall results whereas those for manufacturing are quite different. While this is reminiscent of the results of Lawless et al. (2018) given the small number of manufacturing investment projects, we hesitate to make too much of these results. In addition, using repeated subsamples of the services investments suggests that the number of observations in manufacturing may play a part in the differences. See the online appendix for further discussion.

Note that as there is a great deal of missing information on total assets, particularly for non-European affiliates, we use the number rather than the size of these prior affiliates.

This was calculated as the number of investments prior to the first investment in the sample.

One firm had 198 prior investments and another had 918. When we exclude firms in the top percentile of prior investment experience, that is those with more than 52 prior investments, this had little impact on our results.

See Freund and Pierola (2015) for discussion.

In particular, we look to the work on discrete location decisions such as Head et al. (1995), Crozet et al. (2004), Brülhart et al. (2012), Basile et al. (2008), Chen and Moore (2010), Siedschlag et al. (2013a, 2013b).

Note that, as GDP and GDP per capita are in logs, this is equivalent to including GDP and population in logs. Also, the correlation coefficient between GDP and GDP per capita is small, 0.234 for the home country and 0.236 for the host.

The correlation between these two is 0.568. When only one is included, results are broadly similar, as demonstrated in the online appendix.

See http://​www.​cepii.​fr.

Note that this is not origin specific. Particularly in light of the EU, investment barriers may differ between EU members and country pairs involving a non-member. Potential issues are hopefully mitigated by the inclusion of dummies equal to one with both countries are EU15 or Eurozone. In any case, however, if this measure was a poor indicator of investment barriers one would expect insignificance rather than the consistently significantly negative coefficient we find.

This can be found at https://​ora.​ox.​ac.​uk/​objects/​uuid:​81f28d9a-fe6e-445b-8d34-a641b573d986.

Although we refer the reader to Devereux and Griffith (2003) for a detailed discussion, intuitively, this approach constructs a hypothetical firm with a given holding of industrial buildings, intangible assets, and so forth. For each of these, the country-specific tax offsets are calculated, assuming a depreciation schedule, and deducted from the statutory corporate tax to arrive at the effective rates. In unreported results, as an alternative to these constructed tax rates, we used those constructed by Spengel et al. (2014), who use a comparable methodology for European countries. When doing so, similar results were found.

As pointed out by Hines (1996) among many others, increases in taxes tend to have a larger impact when taxes are initially low. This motivates our use of logs which controls for this.

Confirmatory results of the equivalence of these two approaches are available on request.

Note that by including the inverse Mills ratio in the intensive estimation, we are controlling for the impact taxes may have on selection and mitigate potential bias from that source.

These numbers are given by \(100*\Phi (- \hat{\alpha }_{o}/\hat{s}_{o}),\) where \(\phi\) is cumulative standard normal distribution and \(\hat{\alpha }_{o}\) and \(\hat{s}_{o}\) are the estimated mean and standard deviation of the coefficient for the host tax.

This is reminiscent of the negative impacts on aggregate FDI found by Blonigen et al. (2007).

While one might be concerned that this arises from collinearity between these, when omitting them individually we found signs matching those here.

Note that we do this because, recalling that the logit estimators rely on variation across options, interacting a variable that does not vary across hosts (e.g., owner size) with one that does (e.g., the host EATR) would result in a coefficient primarily driven by variation in the EATR itself which is already being controlled for.

This would be reminiscent of Davies et al. (2018) who find that transfer pricing in tangibles is primarily in differentiated products where arms-length comparables are difficult to identify.

Note that although the coefficient in column (3) is positive, it is well outside the normal range of significance.