Income shifting is a significant source of tax planning for U.S. corporations. We use confidential Internal Revenue Service (IRS) data to develop a firm-year measure of income shifting. Our measure captures the relative extent of U.S. multinational entity (MNE) net intercompany payments out of the United States to CFCs. Our data show that the majority of sample firms report net inbound intercompany payments on average. Sample firms report nearly $830 billion of outbound payments and over $1 trillion of inbound payments in total. Companies reporting net outbound payments are smaller and operate in high-tech industries. Supplemental analyses show that firms with outbound intercompany payments have a lower rate of IRS audit and are no more likely to be assessed additional taxes upon audit. Our study provides a measure based on publicly available data that researchers, investors, and policymakers can use to infer outbound income shifting.
Studies using intrafirm trade or price data are notable exceptions. In particular, Clausing (2003) observes prices charged between related and unrelated companies to more directly estimate tax-motivated transfer pricing.
Prior literature examines nontax motivations for foreign investment, production locations, and reported income, which can give rise to intercompany payments that shift income (e.g., Bartik 1985; Allred and Park 2007; Busse and Hefeker 2007; Lewin, Massini, and Peeters 2009; Alcácer and Zhao 2012; and Huang, Krull, and Ziedonis 2015).
Using Bureau of Economic Analysis data from 2000, Clausing (2006) estimates intrafirm trade represented 41% of U.S. international trade. This estimate likely understates the current portion given recent rising globalization and increases in cross-jurisdictional tax-motivated income shifting.
Because we construct our measure using a regression of net outbound payment intensity on firm-year determinants of income shifting, we caution researchers about using the measure as an outcome variable. As with similarly constructed measures (e.g., tax shelter propensity, financial constraint scores, etc.), using our measure as a dependent variable could lead to incorrect inferences if a true determinant of the outcome variable is uncorrelated with the variables used in the prediction model or vice versa.
E.g., Klassen et al. (1993), Hines and Rice (1994), Huizinga and Laeven (2008), Huizinga, Laeven, and Nicodeme (2008), Dharmapala and Riedel (2013), Dyreng and Markle (2016) and De Simone, Klassen, and Seidman (2017).
Chen et al. (2017) build on the work of Klassen and Laplante (2012b) to develop a continuous firm-year measure of outbound tax-motivated income shifting. In a concurrent working paper, De Simone et al. (2018b) use separate entity financials of global multinationals to estimate a firm-specific measure.
Collins and Shackelford (1998) also use confidential tax return information on intercompany payments to examine which types of transactions are most sensitive to explicit tax rate differentials across jurisdictions.
In general, any U.S. person that owns more than 50% of the total combined voting power of all classes of stock of a foreign corporation entitled to vote, or more than 50% of the total value of shares of all classes of stock of a foreign corporation, is required to file Schedule M.
Generating income directly in foreign locations can play a significant role in a firm’s ability to shift income to low-tax foreign operations, making the firm less willing to expend resources to shift further income via intercompany payments with CFCs. Other firms aggressively pursue both activities.
Although some research uses foreign assets to capture the extent of firms’ foreign operations, foreign assets are not widely reported in our sample period (Oler et al. 2007).
We rank by deciles because linear hypothesis tests of alternative ordered logistic regression models (e.g., using quartile or quintile ranks) reveal significant violations of the proportional odds assumption of ordered logistic regressions across several of our determinants. When ranking Net Payments into deciles, linear hypothesis tests suggest that the proportional odds assumption may be violated only for Tobin’s Q. However, because we aim to provide a measure that is parsimonious for other researchers to estimate, we require equal slopes for Tobin’s Q to eliminate the need for researchers to estimate 10 possible scores per firm-year observation.
The version of SAS software on IRS computers (which must be used to access and analyze confidential data) does not allow two-way clustering of standard errors. Because clustering standard errors by firm or by firm and time yields unbiased standard errors when a sample has many firm observations over a limited number of years (Peterson 2009), we do not believe inferences are sensitive to clustering.
Data availability for Schedule M increased significantly in 2005 after the IRS implemented the modernized e-File system. Prior to 2005, we can obtain data for only five Schedule M’s from the IRS.
Only those U.S. persons who had control of a foreign corporation for at least 30 days during the accounting period of the foreign corporation are required to file Schedule M. Therefore taxpayers filing Form 5471 but not filing Schedule M include MNEs that do not meet these ownership requirements.
Net Payments is calculated per return as total outbound payments less total inbound payments, scaled by worldwide sales. Thus it is possible to observe a positive value of average Net Payments for high-tech firms (i.e., net outbound intercompany payments for the average high-tech firm) while simultaneously observing average unscaled Total In exceeding average unscaled Total Out for high-tech firms.
To preserve the anonymity of firms in the IRS sample, we cannot provide explicit asset cutoffs. We therefore create four subsamples, each containing roughly 25% of sample observations, using rounded asset thresholds.
In untabulated analyses, we rank Net Payments by year. The decile rank by year is significantly correlated with Outbound Rank at 0.987. We confirm results are robust to ranking by year.
In additional tests (untabulated), we confirm this negative relation holds in univariate correlations as well as in subsample regression analyses, after partitioning the sample into high-tech and nonhigh-tech firms. We also estimate the regression by asset quartile and document a negative (insignificant) coefficient for firms in the bottom three quartiles (top quartile). Thus this unexpected finding is robust.
To facilitate computation, a database of scores for a sample of U.S. multinationals on Compustat and more detailed instructions on how to calculate the score are available on the authors’ websites.
For parsimony and consistent with the work of Hadlock and Pierce (2010), we calculate a score. The alternative is to identify the most likely Outbound Rank for each firm-year. Given that we rank our dependent variable into deciles, identifying the most likely rank requires calculating the probability associated with each of the 10 ranks per firm-year, then identifying the rank associated with the maximum of the ten probability calculations. Calculating a single probability following Wilson (2009) or Lisowsky (2010) is not possible in our setting because we use an ordered logit.
We consider the Collins et al. (1998) model to classify an observation as shifting income if the resulting residual is negative and the firm’s average foreign tax rate is less than the U.S. statutory tax rate. We repeat these comparisons described above replacing Outbound Score with the IRS payments data (Outbound Rank). Inferences remain unchanged, which further validates that our score successfully captures the information contained in the IRS payments data.
We define Audit based on codes in the AIMS database that indicate how a tax return case was closed. We set Audit equal to one when the code indicates any examination was conducted (e.g., “No Change,” “Agreed,” “Appealed,” and “Petitioned”). We set Audit equal to zero when the code indicates the return was accepted as filed or without an examination (e.g., “Accepted as Filed” and “Survey,” where the latter indicates the IRS initially selected a return for examination but closed it without contacting the taxpayer).
We estimate consistently signed coefficients across all three measures of Outbound. The two-tailed p value on the estimate for Outbound Rank is 0.1069, reflecting marginal significance at conventional levels.
We also identify potentially aggressive shifters as those reporting lower-than-expected inbound payments (or higher than expected outbound payments), based on the residual from an OLS regression of Net Payments on the determinants included in our equation (1). We find no evidence that the IRS is more likely to audit these potentially aggressive shifters (untabulated).
We acknowledge this analysis likely suffers from selection bias. However, our IRS software does not allow for tests addressing selection bias when both the selection model and the model of interest have binary dependent variables.