Tax Planning and Multinational Behavior∗

Agency decision

Ask Donna

What actually matters in this document.

Text

Rosanne Altshuler

Lysle Boller

Rutgers University Penn Wharton Budget Model

Juan Carlos Suárez Serrato

Stanford GSB & NBER

February 2024

Abstract

We study the adoption and use of a specific form of tax planning by US multinational

corporations (MNCs). Using IRS data, we identify “hybrid” tax planning structures (HTPs)

which can be used to avoid corporate income tax by targeting mismatches between US and

Irish, Dutch, and Luxembourgish tax law. By 2016, more than 35% of the foreign profits

of US MNCs were linked to HTPs. Difference-in-differences models comparing adopting

and non-adopting MNCs reveal that after HTP adoption, MNCs intensify behaviors related

to profit shifting, significantly increasing related-party loans, foreign intangible assets, and

profits held abroad. These changes result in stark reductions in foreign effective tax rates.

Adopting MNCs also experience larger increases in foreign tangible assets and in global

R&D, payroll, and investment.

Keywords: international taxation, profit shifting, Double Irish, Reverse Hybrid Mismatch

JEL Codes: D22, H25, H26, H32

This research was conducted as part of a Joint Statistical Research Program with the Internal Revenue Service

(IRS). Any views expressed are those of the authors and not those of the IRS. The statistics reported in this

paper have been reviewed and cleared for disclosure by the IRS.

∗

We are very grateful for comments from David Agrawal, Hunt Allcott, Alan Auerbach, Michael Best, Jeff

Clemens, Julie Cullen, Dhammika Dharmapala, Rebecca Diamond, Tim Dowd, Clare Doyle, Jesse Drucker, Naomi

Feldman, Roger Gordon, Mindy Herzfeld, Jim Hines, Chad Jones, Wojciech Kopczuk, Wouter Leenders, Rebecca

Lester, Antoine Levy, Li Liu, Clément Malgouyres, Peter Merrill, Mathilde Muñoz, Karthik Muralidharan, Paul

Oosterhuis, Mathieu Parenti, Emmanuel Saez, Navodhya Samarakoon, Shoshanna Vasserman, Daniel Xu, Danny

Yagan, Gabriel Zucman, and from seminar participants at CESifo, Columbia University, the Federal Reserve

Bank of San Francisco, the National Tax Association Annual Meetings, the NBER Public Economics Meetings,

the Oﬀice of Tax Analysis of the US Treasury Department, the Oxford University Centre for Business Taxation,

the Paris School of Economics, the Penn Wharton Budget Model, Rutgers University, Stanford University, UC

Berkeley, UC San Diego, the University of Illinois at Urbana-Champaign, the University of South Carolina,

and the Statistics of Income unit of the Internal Revenue Service. We are grateful to Arnold Ventures, the

International Tax Policy Forum, and the NBER for their support of this project. Special thanks to Chloe Gagin,

Nuria McGrath, Melanie Patrick, Ralph Rector, Michael Weber, and Jason Wenrich for assistance with data

access. Clare Doyle, Jason Harrison, and Kevin Roberts provided excellent research assistance. All errors remain

our own.

Introduction

The last quarter century has seen a remarkable increase in the complexity of tax planning

by multinational corporations (MNCs). Document leaks and special government reports have

revealed the existence of tax planning strategies that are designed to avoid corporate income

taxes in multiple jurisdictions by leveraging mismatches in tax laws across countries. While

media attention following these revelations and dissatisfaction with the current system have

motivated important international tax policy changes and multilateral projects to address tax

avoidance, little is known about the importance of these tax planning strategies.1 How prevalent

are these strategies among MNCs? Do they facilitate profit shifting and lower foreign effective

tax rates (ETRs)? Does tax planning influence the real economic activity of MNCs?

This paper uses tax data from the Internal Revenue Service (IRS) to answer these questions by

analyzing the adoption and use of a set of complex tax planning strategies that target mismatches

between US and Irish, Dutch, and Luxembourgish tax law. We first use data from multiple IRS

tax forms to reconstruct the ownership networks of the foreign aﬀiliates of US MNCs. We then

identify when a US MNC creates an ownership structure that can leverage mismatches in tax

laws across these countries. Data from the tax returns of US MNCs is crucial for the purposes

of identifying the adoption of these tax planning structures and understanding how MNCs use

them to shift profits across countries. This paper is the first to systematically uncover these tax

planning structures and to study how their adoption is related to changes in tax avoidance and

real economic activity.

We use these administrative data to document the growth and prevalence of these tax planning strategies. Although these strategies were extremely rare in the early 1990s, they were

gradually adopted by MNCs following the 1997 regulations known as “Check the Box” (CTB),

which facilitated this form of tax planning. By 2016, 17.5% of US MNCs in our sample had

adopted at least one of the structures we identify, and these companies were responsible for

more than 60% of foreign profits. Data on the foreign operations of adopting MNCs show that

more than 50% of their foreign profits flow through one of these structures and that more than

35% of the foreign profits of all MNCs in our sample are linked to these structures. MNCs

that use these strategies are also responsible for significant shares of domestic economic activity,

These changes and projects include important aspects of the recent Tax Cuts and Jobs Act (TCJA) of 2017,

the European Anti-Tax Avoidance Directive, and the OECD’s Base Erosion and Profit Shifting project.

Figure 1: Comparison of Foreign Effective Tax Rates

20%

10%

Other MNCs

Hybrid Tax Planners

1995

2000

2005

2010

2015

Notes: This figure compares the aggregate annual foreign ETR for two groups of US MNCs. The light green line

shows the ETR for a group of MNCs that eventually adopt at least one of the hybrid tax structures described in

Section 2. The dark blue line shows the ETR for MNCs that did not adopt any of these structures during the

sample window. The combined sample includes most large US C corporations as described in Section 3.

including 25% of domestic corporate payroll and 15% of domestic capital investment.

Relative to MNCs that never adopt these structures, those that do engage more intensely

in the kinds of financial transactions that could be used to shift profits to low-tax countries:

they increase their foreign holdings of intangible capital, they have larger loan balances between

related foreign aﬀiliates, and they collect more royalty income and accumulate more cash abroad.

Most strikingly, as shown in Figure 1, over the period during which these structures are gradually

put in place, the foreign ETRs of adopting MNCs experience a dramatic decline. By 2016, the

foreign ETR of adopters was roughly half that of other MNCs.

We use difference-in-differences event study regressions to conduct more formal comparisons.

The models measure changes in firm outcomes surrounding the adoption of hybrid tax planning

structures, relative to non-adopting firms. These models confirm that the marked increase in

behaviors related to profit shifting coincides with the adoption of these structures and that these

mechanisms of tax avoidance lead to significant declines in foreign ETRs, as in Figure 1.

Concurrent with declining foreign ETRs, we estimate that adopting US MNCs have larger

increases in foreign investment and accumulate 40% (p < 0.01) more depreciable capital in

foreign aﬀiliates than MNCs that do not adopt one of these structures. We also estimate that

the adoption of these tax planning structures is followed by a 20% (p < 0.01) increase in domestic

payroll and a 50% (p < 0.01) increase in expenditures on research and development (R&D).

We develop our results in three steps. First, we reconstruct the foreign ownership structures of

US MNCs using information from three key tax forms contained in the IRS data files. Parent-level

data from Form 1120 and related tax forms provide information on domestic activity, including

assets, payroll, and domestic investment. Parents also file an information form (Form 5471) for

each of their controlled foreign corporations (CFCs), which includes data on foreign assets, taxes,

and earnings and profits (E&P), as well as related transactions between CFCs. With the advent

of CTB, US MNCs could elect to “disregard” their foreign aﬀiliates. These foreign disregarded

entities (FDEs) are hybrid structures that are considered corporations in the host country but

pass-through entities from the US perspective, making them transparent to the US Treasury.

The IRS collects information about these entities in a separate information return, Form 8858.

A novel aspect of this paper is the integration of FDE data with their CFC owners.

Second, with this information in hand, we flag MNCs that adopt tax planning structures

that have previously only been revealed in leaks and special government reports. These planning

structures use entities that facilitate tax planning by creating mismatches in their tax treatment

between the United States and foreign countries. The first structure we identify, known as

the Double Irish, uses FDEs to leverage aspects of Irish and US tax law and shift profits out

of high-tax foreign countries into tax havens. The second and third structures use a strategy

known as a Reverse Hybrid Mismatch to route foreign profits to low-tax foreign aﬀiliates. We

identify usage of Reverse Hybrid Mismatches for foreign aﬀiliates located in the Netherlands

and in Luxembourg. These three hybrid tax planning structures are well known and have been

targeted by European countries through legal investigations.2 They have also motivated policy

agendas such as Action 2 of the OECD Base Erosion and Profit Shifting (BEPS) project, which

aims to close down hybrid mismatch arrangements.

Using these unique indicators of tax planning, we measure the growth, prevalence, and importance of these hybrid tax planning (HTP) structures relative to the aggregate economic activity

of US MNCs. As mentioned above, the three structures we identify are connected to large shares

of aggregate foreign activity by US MNCs by 2016, the end of our sample period. Although there

have been numerous investigations and case studies that reveal how these arrangements work,

this paper is the first attempt to systematically measure their magnitude.

Use of hybrid tax planning structures by US MCNs has been previously reported by Drucker (2010), Duhigg

and Kocieniewski (2012), Guardian (2019), Guardian (2018), and Kleinbard (2013).

In our third and final step, we compare the foreign and domestic operations of US MNCs

that adopt one of these strategies to those of MNCs that do not. In the early 1990s, these two

groups of firms experience similar trajectories with regard to their foreign ETRs and measures of

domestic and foreign activity. After the 1997 CTB regulations are put in place, foreign aﬀiliates

of MNCs that engage in hybrid tax planning experience significantly larger declines in ETRs;

more rapid foreign and domestic growth; and sharp increases across several proxies for profit

shifting and tax deferral, including related-party loans, intangible assets, payments related to

cost sharing agreements, royalty income, and cash held abroad.

To ensure that these results are tied to the adoption of HTPs, we estimate staggered differencein-differences models around the first year that a given MNC adopts an HTP. Estimates using

a “stacked” difference-in-differences estimator (e.g., as in Cengiz, Dube, Lindner and Zipperer,

2019) provide evidence that these structures were put in place for tax avoidance purposes.3

Specifically, we show that MNCs engage in more financial transactions that can be used to

shift profits and reduce foreign ETRs following the adoption of an HTP. We also find that

adopting MNCs experience larger increases in foreign capital, domestic payroll, and R&D than

non-adopting MNCs in the years surrounding the adoption of an HTP. One potential concern is

that these complex tax structures are adopted by specific types of firms that were also subject to

other macroeconomic shocks over this time period. We address this concern by showing that we

obtain similar results when we include industry-by-year fixed effects, when we flexibly control

for firm size bins interacted with year fixed effects, and when we flexibly control for differences

in intangible assets across firms. We also obtain similar results when we additionally use inverse

propensity score weights to attain more similar firm-to-firm comparisons, when we use a two-way

fixed effects (TWFE) specification, and when we use an alternative staggered estimator from Sun

and Abraham (2021).

Estimates of our staggered difference-in-differences models have a causal interpretation under

the assumption that the outcomes of MNCs that adopted an HTP would have otherwise trended

similarly to those that did not adopt such a structure. Event study results generally show that

important outcomes for adopting MNCs, such as growth in payroll, foreign assets, R&D, and

foreign ETRs, follow similar patterns to those of non-adopting MNCs prior to HTP adoption.

These results suggest that MNCs do not select into HTPs based on prior trends in economic

Throughout the paper we use tax avoidance to refer to legal strategies used to minimize tax obligations.

outcomes. However, it is possible that MNCs select into HTPs based on the gains from adopting

an HTP, such that adopting MNCs would have different post-period trends absent an HTP.

Interpreting these results through the lens of an economic model could help decompose firmlevel changes into two components, one that captures selection on gains from tax planning and

another that captures the effects of the HTP tax advantage.

Overall, our results provide the first systematic documentation of the prevalence of hybrid

tax planning structures among US MNCs. We show that, by 2016, these structures accounted for

about a third of their foreign profits and that these MNCs represent a large fraction of domestic

corporate activity. We estimate that adoption of these strategies is followed by large increases

in financial transactions that can be used to shift profits to low-tax countries and by declines in

foreign ETRs. We also estimate significant relative changes in domestic and foreign economic

activity following the adoption of an HTP.

This paper contributes to studies that quantify the importance of profit shifting. In a seminal

contribution, Hines and Rice (1994) describe and measure the importance of tax havens to

the operations of US MNCs. Clausing (2016) uses the sensitivity of reported profits to tax

rate differentials to estimate the magnitude of profit shifting of US MNCs. Tørsløv, Wier and

Zucman (2018) use macroeconomic data and differences in the profitability of different aﬀiliates

to estimate the magnitude of profits shifted to tax havens. Bilicka (2019) uses tax data from

the UK to argue that the large differences in the profitability between domestic UK firms and

the aﬀiliates of foreign MNCs (in the UK) are driven by profit shifting. In a recent survey,

Dyreng and Hanlon (2021) highlight the cross-sectional variation in tax avoidance and conclude

that a large portion of the variation in tax avoidance remains unexplained. Our focus on HTPs

contributes to understanding the importance of specific tax planning strategies.

While several papers document the existence of profit shifting, the magnitude of this problem has been hard to pin down. Using tax data from US firms, Dowd, Landefeld and Moore

(2017) argue that the sensitivity of reported profits to tax differentials can be non-linear and

that accounting for non-linearities increases estimates of profits shifted to low-tax countries.

In contrast, Blouin and Robinson (2020) argue that prior estimates using tax and survey data

can be plagued by double-counting of profits and that accounting for direct investment income

between aﬀiliates can significantly lower estimates of profit shifting. Following the suggestions

in Blouin and Robinson (2020), we subtract dividend income from related foreign corporations

when computing aggregate foreign earnings for US MNCs. Rather than provide estimates of

profit shifting, we document the prevalence of widely used tax planning structures, show that

MNCs use them in transactions that are likely related to profit shifting, and find that close to

one third of the foreign profits of US MNCs in our sample flow through these structures by 2016.

The characterization of the structure of the foreign activities of US MNCs prior to the passage

of the Tax Cuts and Jobs Act of 2017 (TCJA) is an important contribution of this paper, as the

response of US MNCs to the many changes and new incentives in the TCJA likely depends on

these preexisting structures.4

We also contribute to our understanding of CTB regulations by directly examining the adoption and consequences of complex tax structures facilitated by the policy. Using tax return data

from US MNCs, Altshuler and Grubert (2006) find reductions in foreign effective tax rates after

the enactment of CTB in 1997 that are consistent with the use of the tax planning strategies we

examine. Mutti and Grubert (2009) use multiple data sources to show that, after the implementation of CTB, MNCs increased profit shares in low-tax jurisdictions and transferred intangible

assets abroad. Blouin and Krull (2014) show that MNCs had more tax haven aﬀiliates and

longer ownership chains after the enactment of CTB. Faulkender, Hankins and Petersen (2019)

study the 440% increase in cash held abroad by US MNCs between 1998 and 2008 and argue

that this rise is driven by tax factors, including CTB. While prior research assumed that MNCs

disregarded foreign aﬀiliates following the CTB regulations, ours is the first paper to use tax

information to confirm when an aﬀiliate is disregarded and to systematically identify MNCs that

use a specific set of tax planning structures. In a contemporaneous paper, Samarakoon (2022)

uses tax data to identify firms that use a Double Irish structure and examines how the closure

of this structure impacts repatriation of deferred earnings by MNCs.5 Our results show that the

bulk of the decrease in foreign ETRs and increase in cash held abroad over the sample period

The TCJA lowered the corporate tax rate to 21% and made significant changes to US taxation of international

income. It also introduced four new provisions. First, due to the transition to territorial taxation, US MNCs

now can deduct dividends received from foreign aﬀiliates from their US taxable income, thereby eliminating any

repatriation tax. Second, the TCJA introduced a new tax on Global Intangible Low-Taxed Income (GILTI)—

defined as income that includes low-tax foreign income exceeding 10% of an MNC’s tangible foreign capital

investment (adjusted for depreciation). Third, a new category of income—Foreign-Derived Intangible Income

(FDII)—is subject to a reduced tax rate. FDII encompasses income derived from intellectual property held in

the United States that generates foreign sales. Finally, the TCJA introduced the Base Erosion and Anti-Abuse

Tax (BEAT) to curb the erosion of the tax base by both US and foreign MNCs.

Hardeck and Wittenstein (2018) use data from the Luxembourg Leaks to identify firms with hybrid tax

structures and find that hybrid tax structures reduce MNC tax rates, as measured by financial statements data.

was driven by MNCs that adopted a particular set of tax planning structures.

Finally, our paper contributes to the literature on how profit shifting impacts real behavior.

Grubert and Slemrod (1998) study profit shifting opportunities through Puerto Rico and argue

that US MNCs changed their investment decisions in response to these opportunities. Suárez

Serrato (2018) studies the reduction in profit shifting opportunities through Puerto Rico and

shows that US MNCs decrease their domestic investment in response.6 Albertus (2019) uses

Bureau of Economic Analysis data to compare US MNCs with different average foreign tax

rates prior to the implementation of CTB. He finds that MNCs with higher initial tax rates

experienced a larger decline in average rates and increased their R&D intensity after 1996. By

using tax data to identify specific tax planning structures and to demonstrate how they are used

for profit shifting, we reveal substantial heterogeneity regarding the types of firms that benefited

from CTB and shed light on the mechanisms through which CTB lowered foreign ETRs and

affected real economic activity.

The remainder of the paper is organized as follows. Section 2 discusses how CTB facilitated

the creation of foreign disregarded entities and describes the three hybrid tax planning structures

we examine. Section 3 provides an overview of the data. Section 4 discusses how the creation

of foreign disregarded entities and hybrid tax planning structures has grown over time. Section

5 estimates firm-level changes in foreign tax rates and economic activity of US MNCs following

the adoption of tax planning strategies. Section 6 concludes. We conduct additional analyses

in the appendices. Appendix A provides robustness checks for our main analyses. Appendix B

discusses the measurement of foreign earnings and taxes. Appendix C studies how MNCs can

structure cost sharing agreements to shift profits from the US parent to foreign aﬀiliates.

Hybrid Tax Planning Structures and Check The Box

This section describes the three hybrid tax planning structures that we study. We first describe

the “Check the Box” regulations that facilitated their adoption and then describe the structures

in detail.

de Mooij and Liu (2018) study the impact of transfer pricing regulations and show that these policies can

reduce investment. Bilicka, Qi and Xing (2019) show that a worldwide debt cap that limited interest stripping

as a form of profit shifting also impacts the investment decisions of UK MNCs.

2.1 Check The Box Regulations

During the period we study, the United States imposed a corporate tax on the worldwide income

of US corporations, with a credit for foreign taxes paid to avoid double taxation. The credit was

limited to what US tax would have been on the foreign income. Taxes were not due on active

foreign business income until it was repatriated to the US parent corporation. This deferral

feature of the US tax code made it attractive to hold income generated abroad in tax havens.

To prevent profit shifting, deferral was not extended to certain types of “tainted income”

under what is generally referred to as Controlled Foreign Corporation rules. These rules are

contained in Subpart F of the tax code, and foreign income that is subject to current US tax

is referred to as “Subpart F” income. Tainted income includes passive portfolio income and

the payment of interest, dividends, and royalties from one CFC to a related CFC in another

jurisdiction.

In 1996, the US Treasury promulgated regulations effective on January 1, 1997, that made

it easier for US corporations to change the entity classification (e.g., pass-through or corporate)

of domestic and foreign aﬀiliates. This policy change became known as Check the Box, referring

to the ease with which US corporations could change entity classifications. CTB was originally

intended to simplify tax filing for domestic firms. However, it also facilitated certain types of

international tax planning strategies that leverage mismatches in tax laws across countries. These

strategies make use of foreign aﬀiliates referred to as “hybrid entities” that are treated differently

for tax purposes at home and in host countries.

Below we discuss how tax planning structures that use hybrid entities allow US companies

to avoid US tax levied on intercompany payments such as dividends, interest, and royalties, and

how these structures leverage mismatches in tax laws across countries to lower foreign tax bills.

2.2 Hybrid Tax Planning Strategies

CTB facilitates tax planning by allowing MNCs to easily create “hybrid” entities. A hybrid

entity is a business operation that is incorporated from the foreign country point of view and

a pass-through (unincorporated branch of another corporation) from the US point of view (or

vice versa, in which case it is referred to as a “reverse” hybrid). Since 1997, an MNC can simply

check a box on a tax form to disregard a foreign corporation, thereby creating an FDE, a type of

pass-through entity. If an entity is disregarded, the transactions with its entity parent and with

other FDEs owned by the same parent become transparent to the US Treasury, as they are all

viewed as part of one consolidated corporation.7

The simplest hybrid tax planning structure allows MNCs to take large deductions for interest

in high-tax jurisdictions through the use of tax haven finance aﬀiliates. Consider the following

planning structure to finance a subsidiary in a high-tax country. Instead of funding the high-tax

subsidiary directly, the parent injects equity into a tax haven aﬀiliate, which lends to the hightax subsidiary. The high-tax subsidiary then pays interest to the tax haven aﬀiliate. This profit

shifting strategy is commonly known as “interest stripping.” Though the interest is deductible

abroad against taxable income, it remains subject to immediate US tax under the CFC rules.

Prior to 1997, CFC rules made the use of a tax haven financing aﬀiliate unattractive for tax

purposes. Since 1997, however, the parent can check the box on the high-tax aﬀiliate, making it

a hybrid FDE. From the US point of view, the high-tax CFC is an unincorporated branch of the

tax haven FDE; the interest payment is thus transparent to the US Treasury, which regards the

combined tax haven/high-tax operation as one consolidated corporation. The interest payment

therefore avoids Subpart F taxes and the company can defer US income tax by holding profits

in the tax haven.

Panel A in Figure 2 depicts this simple hybrid tax planning structure using a tax haven

aﬀiliate. The green box around the two entities (the tax haven CFC and the high-tax FDE)

indicates that the structure is consolidated from the US point of view. This simple structure

allows the parent to capitalize a foreign aﬀiliate through a tax haven while making intercompany

payments transparent, thereby avoiding any current US tax on interest.

While this structure avoids Subpart F tax and defers US income tax, the MNC would still

be subject to corporate income tax in the tax haven (if it exists) and potentially to foreign

withholding tax on the interest payments between aﬀiliates.8 Moreover, to combat interest

stripping, many countries have adopted “thin-capitalization” rules that limit the tax deductibility

of interest payments, reducing the attractiveness of this option.

Although it was possible for MNCs to create foreign disregarded entities prior to 1997, there were strict rules

regarding what types of entities could be declared as such. In particular, such entities had to demonstrate that

they possessed at least three of four characteristics associated with partnerships. In practice, we observe very few

of these entities prior to the implementation of CTB in 1997.

After the adoption of “look-through” rules passed as part of the Tax Increase Prevention and Reconciliation

Act of 2005, MNCs could avoid Subpart F taxation on distributions of interest, rents and royalties across CFCs

without relying on FDEs.

2.2.1

CTB and Cost Sharing Agreements

Another form of income shifting is available to MNCs with intellectual property (IP). This

method uses cost sharing agreements (CSAs) to develop IP that can be licensed abroad. These

agreements are particularly tax advantageous when combined with CTB.

Under a cost sharing agreement, the tax haven aﬀiliate makes a “buy-in payment” that funds

a part of the parent’s R&D project. This gives the aﬀiliate the right to license resulting IP to

other foreign subsidiaries in exchange for royalty payments. Royalty payments are not subject to

current tax under Subpart F if the parent checks the box on the aﬀiliate making the payment.9

The key is that with CTB, any payments for the use of the IP abroad are contained within one

consolidated company from the view of the US Treasury. This structure has the same foundation

as in Panel A of Figure 2, but replaces the equity injection with a transfer of IP (via a cost sharing

agreement) and uses royalty payments instead of interest to shift profits.

It is important to note that determining the right arm’s length payment for the buy-in is

usually quite diﬀicult. Typically the IP is not fully developed at the time the buy-in payment is

made, so there is uncertainty regarding future profits. While the US has rules under which buy-in

payments must be adjusted if the profits associated with the IP are too high relative to payments,

it is still possible for MNCs to underprice the IP. This allows US MNCs to shift income to lowtax aﬀiliates. MNCs are then able to use hybrid tax planning structures, as discussed below, to

minimize tax on the foreign profits generated from their IP.

Even if IP is not underpriced, MNCs have historically attempted to strategically allocate

allowable costs to generate tax savings through cost sharing agreements. We discuss these cost

allocation strategies as well as their legal challenges in Section 5.3 and Appendix C.

As with interest stripping, royalty payments may still be subject to corporate income taxes in

a tax haven and to withholding taxes that are meant to prevent profit shifting between countries.

We now describe complex tax planning strategies that aim to reduce exposure to income and

withholding taxes across multiple jurisdictions, including the US.

Profits could be further accumulated in a tax haven if MNCs overprice the royalty. The absence of comparable

transactions makes it hard for tax authorities to value intellectual property and correctly price royalty payments.

Figure 2: Diagrams of Hybrid Tax Planning Structures

(A) Simple CTB Structure

(B) Double Irish CTB with Cost Sharing

Parent MNC

Transfer IP

Equity

Consolidated

Corporation

Irish

Holding Company

(M&C in Bermuda)

Tax Haven CFC

Loan

License

Interest

Royalty

Irish

FDE A

High-Tax

FDE

Sales outside of US

Parent MNC

Transfer IP

(D) Reverse Hybrid Mismatch:

Dutch CV-BV

US #1

US #2

Managing

Partner

Silent

Partner

Transfer IP

Royalty

Dutch

Conduit

FDE

Irish

Holding Company

(M&C in Bermuda)

Dutch CV

(Partnership

Ltd.

Co.)

License

Royalty

License

Irish

FDE

Royalty

Dutch BV

(Private Ltd.

Liability Co.)

Sales outside of US

Notes: Panel A of Figure 2 depicts a hypothetical financing CTB structure; Panel B describes a Double Irish cost

sharing structure with CTB; Panel C illustrates the Double Irish with a Dutch Sandwich; and Panel D describes

a Reverse Hybrid Mismatch structure, otherwise known as a CV-BV (or SCS-SARL in the case of Luxembourgish

entities). In each of these diagrams, the green rectangle depicts the combined structures as perceived by the IRS,

and squares denote corporations. Squares with circles inside denote hybrid entities, which are corporations in the

local country but disregarded for US purposes. In Panel D, the CV (or SCS if Luxembourg entities are used) is

shown as a triangle to denote that it is a reverse hybrid: is it a partnership for Dutch purposes but a corporation

for US purposes.

2.2.2

CTB, Ireland, and Intellectual Property

The first tax planning strategy we study is known as the Double Irish and involves setting up

a network of aﬀiliates in Ireland and a tax haven country such as Bermuda. To motivate this

structure, consider a parent MNC that develops IP in the US that it wants to sell around the

world. The parent can transfer the IP to a holding company in a tax haven using a cost sharing

arrangement. The tax haven holding company then licenses the IP to an Irish CFC (CFC A),

which pays royalties back to the holding company from the sales revenue it receives selling the

IP abroad.

This initial cost sharing structure creates three tax problems for the parent. First, the parent

will be subject to Subpart F taxes (current US tax) on the royalties paid from Irish CFC A to

the tax haven holding company. Second, taxes will be due in Ireland on any profits that remain

in CFC A after royalties are paid to the haven holding company. Finally, the parent will owe

Irish withholding taxes on the royalty transfers to the haven.

The first two of these tax problems can be solved using CTB and a Double Irish tax planning

structure, as summarized in Panel B of Figure 2. In this structure, the parent transfers the

IP to a holding company managed and controlled in a tax haven (e.g., Bermuda) but legally

incorporated in Ireland. Though the US considers this an Irish holding company, under Irish

tax law, the holding company is a Bermuda company and therefore not subject to Irish tax.

The parent also checks the box on Irish CFC A to avoid current US tax on the royalties: the

CFC becomes an FDE and is therefore fiscally transparent to the US Treasury. This eliminates

Irish tax on any profits remaining in Ireland (i.e., the holding company) after royalties are paid,

as well as current US taxes on the royalties. However, the transfer from the Irish FDE to the

IP-holding company still generates Irish withholding tax.

MNCs can eliminate this withholding tax, thus solving the final tax problem, by inserting a

Dutch conduit—a “Dutch Sandwich”—between the Irish aﬀiliates. With the Dutch conduit in

place, the parent owes no withholding tax on payments between the conduit and the Irish aﬀiliate

(FDE A), as no withholding taxes are due between European Union companies. Further, no

withholding taxes will accrue on the royalties between the Dutch conduit and the Irish holding

company because no withholding tax is imposed on these transfers under Dutch law. To avoid

Subpart F taxes on these royalty payments, the parent also checks the box on the Dutch conduit,

making it an FDE. This “Double Irish with a Dutch Sandwich” hybrid tax planning structure,

as shown in Panel C of Figure 2, solves all three tax problems we identified above.

2.2.3

Reverse Hybrid Mismatch

The final two tax planning structures we examine use a strategy known as a Reverse Hybrid

Mismatch. While this strategy can be employed using aﬀiliates in different countries, we describe

a common structure using Dutch companies. To set up this structure, a US MNC creates two

US-based aﬀiliates to act as managing/silent partners in a Dutch closed limited partnership called

a CV (commanditaire vennootschap in Dutch). The partnership is a “reverse hybrid” entity: it is

treated as a pass-through company by the Netherlands and as a corporation by the US. The CV

owns a Dutch private limited liability company, called a BV (besloten vennootschap in Dutch),

which acts as a holding company. The BV owns foreign (non-US) subsidiaries (e.g., in Europe).

The BV also holds the license for the US IP, sells the IP to foreign companies, and pays royalties

to the CV. Panel D of Figure 2 depicts this structure.

The Reverse Hybrid Mismatch allows the MNC to avoid tax on foreign income by solving

three tax problems. First, tax may be due in the Netherlands. By Dutch tax law, the CV is a

pass-through entity, so corporate tax is not levied in the Netherlands. Second, payments from

BV to CV can generate Subpart F tax. The (reverse hybrid) CV is a corporation from the US

perspective, and if the parent “checks the box” to disregard the (hybrid) BV, the US sees the

two entities as a consolidated operation. Thus, no Subpart F tax will be due on the royalties.

Finally, payments from the BV to the CV could trigger Dutch withholding tax. However, during

our period of analysis, a 2005 decree by the Dutch Finance Ministry exempted US-based CVBVs from withholding tax. With this Reverse Hybrid Mismatch structure in place, profits from

US-developed IP sold abroad were not subject to corporate tax in the Netherlands and enjoyed

indefinite deferral from US tax (under pre-TCJA law). The CV-BV structure is effectively a

“sink” for foreign profits.

A Reverse Hybrid Mismatch structure can be set up through other countries. In particular, a combination of two types of Luxembourgish companies, known respectively as SCS and

SARL, yields a structure similar to the Dutch CV-BV.10 We study both CV-BV and SCS-SARL

structures for the purpose of our analyses.

SCS and SARL are short for société en commandite simple and société à responsabilité limitée, respectively.

In this case, the SCS is the reverse hybrid company and the SARL is disregarded from the US perspective.

While the description of these structures emphasizes their potential to minimize tax obligations, it is important to note that tax planning is also costly. MNCs have to pay for accounting

and legal advice and to engage in transactions to form the structures. In addition, company

executives differ in their perceived cost of adopting tax-aggressive positions. To the extent that

MNCs incur these costs to avoid paying taxes, tax planning is distortionary from an economic

perspective.

Data and Sample Construction

3.1 IRS Business Tax Data

We rely primarily on several IRS datasets for our analysis. These administrative datasets provide

parent and (both CFC- and FDE-level) aﬀiliate-level information disclosed in tax returns that

allows us to measure the domestic and foreign activity of a large sample of US corporations, both

private and public.

The first dataset, commonly referred to as the Statistics of Income (SOI) Corporate Sample,

is an annual stratified sample of US corporations that SOI uses to produce publicly available

aggregated business income statistics.11 The SOI Corporate Sample contains information from

unaudited tax returns for approximately 100,000 US corporations annually, and has been used

in the business tax literature to study the behavior of domestic firms (e.g., as in Yagan, 2015;

Zwick and Mahon, 2017). Our data focus on C corporations that were sampled between 1992 and

2016. The data primarily contain information from Form 1120, the US Corporate Income Tax

Return, as well as some information from related forms. In our analysis, we also use information

from Form 6765, which is used to claim the R&D tax credit, and Form 4562, which is used to

calculate tax deductions for depreciation on capital assets.

The second dataset, which reports information related to foreign aﬀiliates of US corporations,

is used by SOI to publish aggregate statistics for international business taxes (IRS, 2022a). This

dataset contains a subset of C corporations from the SOI Corporate Sample that file Form 5471

or Form 8858: Form 5471 provides financial information and activity of CFCs, and Form 8858

provides similar information for FDEs, the entity type enabled by CTB. We refer to this sample

as the “SOI International Business Tax Sample.” Unlike the SOI Corporate Sample, which is

provided annually, CFC data is collected only in even years. FDE data is collected for four of

Statistics are available at IRS (2022b), and the sampling procedure is described in IRS (2011).

Table 1: Data Sources and Selected Outcomes

SOI Corporate Sample

Form Description

Selected Outcomes

1120

Corporate Income Tax Return

Domestic Assets

Domestic Wages

6765

R&D Tax Credit

Domestic R&D Expenses

Domestic R&D Wages

R&D Tax Credit

4562

Depreciation and Amortization

Capital Investment

SOI International Business Tax Sample

Form Description

Selected Outcomes

5471

CFC Information Return

Country of Incorporation

Foreign Assets

Foreign E&P

Foreign Taxes

Transactions Between CFCs

Transactions Between US Parent and CFCs

FDE Information Return

Country of Incorporation

Date Disregarded

Foreign Assets

Foreign E&P

Pass-through Owners

Tax Owner

8858

Compustat Data

Description

Selected Outcomes

Consolidated Public MNC Data

Deferred Foreign Taxes (txdfo)

Foreign Taxes (txfo)

Net Income (ni)

Pretax Foreign Income (pifo)

R&D Expense (xrd)

Revenue (sale)

Total Assets (at)

the years in our sample period (2006, 2008, 2012, and 2016). Prior to 2004, SOI statistics only

included information related to CFCs for large MNCs with more than $500 million in assets.

Starting in 2004, the sampling procedure became much broader.

In some instances, we also supplement the IRS datasets with financial statements data on

Table 2: Sample Sizes

(1)

(2)

Int’l. Business Sample SOI Corp. Sample

(3)

Stable Sample

20,029

322,538

57,685

3,635

43,941

53,141

MNC Count

CFC Count

FDE Count

23,222

333,438

58,690

Notes: This table provides the size of three different samples of US MNCs, along with their related controlled

foreign corporations (CFCs) and foreign disregarded entities (FDEs). Column (1) provides sample sizes using all

MNCs in the SOI International Business Tax Sample. Column (2) provides sample sizes after removing MNCs

that were not C corporations. Column (3) applies a size filter that removes smaller MNCs from the sample so

that the sample composition is similar in earlier and later years.

public companies from Compustat. Table 1 summarizes the tax forms described above along

with selected outcomes that we use in our analysis, both from SOI data and from Compustat.

3.2 Sample Construction

Table 2 shows the size of several different samples of MNCs, along with their foreign aﬀiliates

(CFCs and FDEs). We consider a firm to be an MNC if it files Form 5471 for at least one

CFC. Column (1) reports the number of MNCs that have coverage in our data from the SOI

International Business Tax Sample. Column (2) shows the number of MNCs from Column (1)

that are C corporations contained in the SOI Corporate Sample. As mentioned above, there was

a sampling change in the international business tax study starting in 2004 that resulted in a large

increase in the sample, especially for smaller MNCs. To stabilize the firm sampling distribution

between earlier and later years, we remove MNCs that did not have at least one CFC with $50

million in foreign assets as well as those with fewer than $500 million in domestic assets. Column

(3) shows sample sizes after applying this filter. Our analysis primarily uses the sample shown in

Column (3) to study firm-level outcomes. Summary statistics for this stable sample are provided

in Table A.1.

3.3 Measuring Foreign Effective Tax Rates

One possible concern when measuring the income of foreign aﬀiliates of corporations relates to

the measurement of foreign earnings. Blouin and Robinson (2020) suggest that aggregated IRS

statistics may inadvertently double count foreign earnings. This is due to accounting quirks of

MNCs. Consider a hypothetical US firm with two CFCs (A and B). Suppose that CFC A is a

holding company that holds a 100% stake in CFC B and has no economic purpose other than to

collect dividends from its subsidiaries, and further suppose that CFC B discloses E&P of $100

million, which is issued as a dividend to CFC A. CFC A will then also report E&P of $100 million.

A simple aggregation of the firm’s foreign profits will result in an estimate of $200 million in

foreign E&P even though the true figure is $100 million. If firms with hybrid structures tend to

issue more dividends between their foreign aﬀiliates, then this exercise would overestimate their

E&P and in turn underestimate the ETR. However, each CFC must also file an attachment to

Form 5471 that discloses transactions between the focal CFC and related CFCs, including any

dividends that the CFCs may transfer to each other.

To ensure that there is no double-counting of foreign profits, we subtract these related dividends from our calculations. Specifically, we compute the firm-level foreign ETR as

ET R =

Foreign Taxes

Foreign Taxes + Foreign Earnings and Profits

where total foreign tax payments are taken from Form 5471, Schedule E.12 To calculate pretax

foreign earnings and profits, we obtain pretax E&P for each aﬀiliated CFC (using Schedules H

and E from Form 5471). Following the suggestion of Blouin and Robinson (2020), we remove

dividends received from related CFCs from E&P (using Schedule M from Form 5471).

In Appendix B, we examine the performance of this correction. We construct a proxy for aggregation error by generating a link between IRS data, which provide disaggregated information

about foreign aﬀiliates, and Compustat, which provides data from public disclosures of consolidated MNCs. By comparing the disaggregated data to the consolidated figures, we can quantify

the extent to which commonly-used aggregation techniques may result in double-counting of

foreign earnings in tax data.

This book-tax comparison reveals that aggregation error has been increasing over time, likely

because the mechanism by which this error occurs suggests that it grows as MNCs create more

complicated aﬀiliate networks. The linked sample also reveals large inconsistencies in the reporting of corporate income tax across firms’ books and tax filings. These inconsistencies are

particularly noticeable in extractive and financial industries.13 Applying the correction proposed

There does not appear to be any such double-counting concern related to the payment of foreign taxes. We

exclude unprofitable firm-years from this calculation.

Extractive industries often operate under contracts with foreign governments that include forms of revenue

sharing, which can be misreported as a corporate income tax.

by Blouin and Robinson (2020) yields a 30% reduction in the magnitude of foreign earnings as

measured in tax data in 2016 and significantly reduces book-tax differences. Furthermore, as

shown in Figures B.7 and B.8, this correction breaks the systematic relationship between booktax differences and the size of multinationals’ foreign aﬀiliate networks. Unadjusted book-tax

differences are increasing over time. After applying the correction, this is no longer true. Both of

these exercises indicate that the correction appears to significantly reduce measurement error.14

3.4 Measuring Changes in International Corporate Structures

Both the SOI Corporate Sample and the International Business Tax Sample have been used to

study domestic and international business taxation. Relatively little work, however, has utilized

the wealth of information regarding FDEs.15 Although data from Form 8858 are collected less

frequently than other samples, they allow us to observe two important features of US MNC

structures. First, they reveal the date when an entity was first disregarded by an MNC, which

allows us to measure adoption of CTB among MNCs. Second, they allow us to observe the

tax ownership structure of each CFC along with its FDEs. These ownership structures reveal

important cross-national linkages within MNCs and, most importantly, allow us to identify CFCs

and FDEs that have particular structures associated with the tax planning strategies described

in Section 2.

3.4.1

Detecting the Double Irish

As described previously, the Double Irish involves two Irish entities—a top-level entity that is

incorporated in Ireland, but managed and controlled in another low-tax foreign country, and a

lower-level Irish entity that merchandises the IP and pays a royalty. Typically, the lower-level

entity is “checked” and is classified as an FDE for US tax purposes. Alternatively, the MNC

may “check” both types of entities which are then classified as FDEs under the “tax ownership”

of a separate CFC. As a result, we flag two types of CFCs that could be used in a Double Irish

arrangement. First, we flag any CFC that is incorporated in Ireland and that checks the box on

an Irish FDE. Second, we flag any CFC that checks the box on two separate Irish FDEs. Note

In Appendix B.3, we also provide corrected estimates of the elasticity of foreign earnings with respect to

foreign tax rates following the methodology of Dowd, Landefeld and Moore (2017). We show that their general

finding that earnings are more sensitive to rates in haven jurisdictions is robust to this correction.

A notable exception is a recent working paper, Samarakoon (2022), that examines the impact of the closure

of the Double Irish tax structure in Ireland.

that this classification method flags “simple” Double Irish arrangements that involve a direct link

between two Irish entities, but also more complex arrangements, such as the Double Irish with

a Dutch Sandwich, that might involve intermediary aﬀiliates through which profits are routed.

3.4.2

Detecting Reverse Hybrids

The other type of structures we consider are Reverse Hybrid Mismatch arrangements common

in the Netherlands and Luxembourg. This arrangement, also described in Section 2, involves a

top-level entity that is classified as a partnership and a bottom-level entity that is classified as

a private limited company (PLC) in the associated country of incorporation. SOI data typically

provide the acronym that is associated with the management form of foreign aﬀiliates on Form

5471 and 8858. In the Netherlands, for example, partnerships are associated with the acronym CV

and the equivalent form of a PLC is associated with the acronym BV. The equivalent acronyms

in Luxembourg are SCS (for a partnership) and SARL (for a PLC). To classify potential reverse

hybrid structures, we flag any CFC classified as a CV (incorporated in the Netherlands) or SCS

(incorporated in Luxembourg) that check the box on an FDE classified as a BV (Netherlands)

or SARL (Luxembourg). We also flag any CFC that checks the box on a CV-BV or SCS-SARL

pair of FDEs.

For both the Double Irish and Reverse Hybrid Mismatch arrangements, we use the first

date that all flagged FDEs were disregarded to measure the year that an MNC first adopted a

particular structure.

Adoption and Prevalence of CTB and Hybrid Tax Planning

Structures

Below, we describe how US MNCs used CTB starting in 1997. We show that after its implementation, usage of tax-transparent FDEs quickly became widespread among US MNCs. A large

share of these entities are connected to well known tax havens.

Next, we focus on the tax planning structures described in Section 2. We show that MNCs

gradually adopted these structures in the decade after the implementation of CTB. By the 2010s,

MNCs that adopted at least one of these structures generated a majority of foreign earnings

among firms in our sample and within these MNCs, a majority of foreign earnings were connected

to these structures. We show that these MNCs also represent a large share of domestic corporate

Figure 3: Adoption of Foreign Disregarded Entities

(A) Cumulative Number of FDEs

Cumulative FDEs

60,000

40,000

20,000

1993

1997

2001

2005

2009

2013

2017

(B) Prevalence of FDEs among US MNCs

Share of MNC Profits

100%

80%

60%

40%

20%

1995

2000

2005

Share of MNC Foreign Profits

2010

2015

Share of MNCs

Notes: Panel A of Figure 3 plots the cumulative number of foreign disregarded entities (FDEs). There were fewer

than 50 FDEs prior to 1997 and this number grew rapidly following the adoption of CTB regulations. Panel B

plots the fraction of US MNCs with an FDE, as well as the share of foreign profits that accrue to US MNCs with

FDEs. By 2008, close to 80% of US MNCs have a FDE, and these MNCs account for close to 100% of foreign

profits.

Table 3: Disregarded Entities by Country of Incorporation

Country Name

Unadj. E&P (billions)

Num. FDEs

Ireland

Netherlands

United Kingdom

Switzerland

Cayman Islands

Singapore

Luxembourg

Bermuda

Canada

Australia

224

136

130

4,844

12,236

26,982

2,138

5,277

3,458

4,592

2,126

7,427

6,780

Notes: This table shows the largest ten countries by total foreign earnings generated by foreign disregarded

entities. Column (2) provides aggregate unadjusted E&P, generated by FDEs in the country listed in Column

(1). This includes E&P for all years that we observe Form 8858 filings (2006, 2008, 2012, and 2016). Column (3)

shows the number of unique entities across all years of this sample.

activity, generating 20% of domestic payroll and holding 15% of domestic capital assets by 2016

among US C corporations.

4.1 Adoption of Check the Box

Panel A of Figure 3 shows the cumulative number of FDEs created between 1992 and 2016. Prior

to 1997, usage of these transparent entities was relatively rare—the IRS used a resource-intensive

system that required firms to show that their aﬀiliates possessed a set of characteristics that

were more consistent with either a partnership (transparent) or corporation (non-transparent)

classification. Starting in 1997, the Treasury relaxed these restrictions, as described in Section

2. As a result, usage of FDEs became widespread over the next two decades, with over 60,000

foreign aﬀiliates classified as FDEs by the end of 2016. Panel B shows that by 2008, about 80%

of MNCs used CTB to declare at least one FDE, and that these MNCs generated nearly all of

foreign E&P.

Table 3 shows the largest ten countries according to total earnings generated by FDEs. The

Netherlands and Ireland are some of the largest domiciles for these types of foreign aﬀiliates.

FDEs also generate large amounts of earnings in well known tax havens, such as the Cayman

Islands and Bermuda.

4.2 Adoption of Hybrid Tax Planning Structures

Panel A of Figure 4 shows the evolution of the share of MNCs in our sample that adopted

particular hybrid tax planning structures. After the implementation of CTB, there was steady

adoption of these structures, with more than 17.5% of MNCs adopting at least one by 2016.

Panel B demonstrates that by 2008, these MNCs generated a majority of foreign E&P. Panel C

shows that the CFCs linked to HTPs generated the majority of profits within MNCs that use

them by 2016. Panel D shows the share of foreign E&P linked to HTP structures relative to

aggregate foreign E&P for MNCs in our sample—by 2016, more than 35% of all foreign E&P of

US MNCs was routed through an HTP.

Figure 5 shows that MNCs with tax planning structures comprise a large share of domestic

economic activity. As a share of all C corporations in the SOI Corporate Sample, which includes

domestic corporations as well as MNCs, MNCs with tax planning structures paid more than 20%

of domestic wages and accounted for about 15% of domestic investment by 2010.

The results in this section demonstrate that several hybrid tax planning structures became

widely adopted by US MNCs after the implementation of CTB, with large shares of foreign

profits flowing through these structures in the decades after adoption.

Figure 4: Adoption of Hybrid Tax Planning (HTP) Structures

(B) Share of Foreign Profits

in MNCs with HTP Structure

Share of MNC Foreign Profits

(A) Share of MNCs with an

HTP Structure

Share of MNCs

15%

10%

2000

SCS−SARL

2005

CV−BV

2010

20%

Double Irish

2015

1995

Multiple

2000

SCS−SARL

2005

CV−BV

2010

Double Irish

2015

Multiple

(D) Share of Foreign E&P Connected

to an HTP Structure (All MNCs)

Share of Foreign E&P

to an HTP Structure (HTP MNCs)

Share of Foreign E&P

40%

1995

40%

20%

1995

60%

30%

20%

10%

2000

2005

2010

2015

SCS−SARL

CV−BV

Double Irish

Multiple

1995

2000

2005

2010

2015

SCS−SARL

CV−BV

Double Irish

Multiple

Notes: These figures show that a growing share of US MNCs have adopted hybrid tax structures over time (Panel

A) and that these MNCs are responsible for a large share of the overall foreign E&P of US MNCs (Panel B).

Panel C shows that a large share of foreign E&P within adopting MNCs is connected to HTPs. Panel D shows

the share of foreign E&P that is connected to an HTP relative to all the MNCs in our sample. In each panel, the

blue area comprises MNCs that have adopted more than one structure.

Figure 5: Hybrid Tax Planning Structures and Domestic Economic Activity

(A) Share of Corporate Domestic Payroll

Share of Domestic Wages

25%

20%

15%

10%

1995

2000

SCS−SARL

2005

CV−BV

2010

Double Irish

2015

Multiple

Share of Domestic Investment

(B) Share of Corporate Domestic Investment

15%

10%

1995

2000

SCS−SARL

2005

CV−BV

2010

Double Irish

2015

Multiple

Notes: These figures show the share of domestic wages (Panel A) and domestic capital investment (Panel B)

paid by US MNCs that adopt one of the hybrid tax structures described in Section 2. This share is computed

as a fraction of all domestic wages and capital investment among C corporations in the IRS Statistics of Income

Corporate Sample. See Section 3 and Table 2, Column 2 for a description of the sample.

Figure 6: Predicted Probability of HTP Adoption by Industry

Information

Manufacturing

Professional, Scientific, and Technical Services

Wholesale Trade

Retail Trade

Management of Companies (Holding Companies)

Mining

Finance and Insurance

0.0

0.1

0.2

Notes: This figure reports the predicted probability of HTP adoption by sector according to a simple logit model.

We remove industries with fewer than ten MNCs in either group.

4.3 Which Firms Adopt Hybrid Tax Planning Structures?

To examine characteristics of MNCs that adopt hybrid structures, we estimate a series of simple

logistic regressions that predict HTP adoption and estimate industry shares within adopting and

non-adopting groups. These analyses support anecdotes that aggressive tax planning MNCs tend

to be larger firms that operate in industries with large amounts of IP. We also examine whether

a set of additional characteristics are predictive of adoption and discuss patterns that emerge

from this analysis.

We start by examining industry variation in the adoption of hybrid tax planning by sector.16

Figure 6 reports the predicted probability of adoption by sector using estimates from a simple

logit model with industry dummies. There is considerable variation in adoption across different

sectors. For example, MNCs that are classified within the Information sector are more than twice

as likely to adopt HTP structures compared to MNCs classified under Finance and Insurance.

We use IRS industry classifications that are analogous to 2-digit SIC codes. For consistency, we use the most

recently observed industry classification for each MNC across all years.

Figure 7: Sector Shares, HTPs vs. Other MNCs

Manufacturing

Other

Information

Wholesale Trade

Professional, Scientific, and Technical Services

Finance and Insurance

Retail Trade

Management of Companies (Holding Companies)

Mining

10%

20%

Hybrid Tax Planners

30%

40%

50%

Other MNCs

Notes: This figure reports industry shares by sector. Sectors with fewer than 10 firms in either group are collected

into the “Other” category.

Figure 7 shows industry shares for MNCs that adopt HTPs and for those that do not. A plurality of MNCs are classifed within the Manufacturing sector. MNCs that adopt HTP structures

disproportionately come from the Information and Manufacturing sectors.

Figure 8 provides predicted probability estimates from a logit model that is analagous to

Figure 6 but for subindustries.17 Adoption is stronger within subindustries that contain tech and

pharmaceutical firms (e.g., data processing and chemical manufacturing). Other IP-intensive

industries, such as publishing, also adopt HTP structures at relatively high rates. Figure 9

displays subindustry shares for adopting and non-adopting MNC groups, providing a similar

takeaway.

These IRS subindustry classifications are roughly equivalent to 3-digit SIC codes.

Figure 8: Predicted Probability of HTP Adoption by Subindustry

Data Processing, Hosting, and Related Services

Other Information Services

Miscellaneous Manufacturing

Publishing Industries (except Internet)

Merchant Wholesalers, Nondurable Goods

Chemical Manufacturing

Computer and Electronic Product Manufacturing

Primary Metal Manufacturing

Food Manufacturing

Professional, Scientific, and Technical Services

Machinery Manufacturing

Electrical Equipment, Appliance, and Component Manufacturing

Fabricated Metal Manufacturing

Transportation Equipment Manufacturing

Management of Companies (Holding Companies)

Merchant Wholesalers, Durable Goods

Mining

0.0

0.2

0.4

0.6

Notes: This figure reports the predicted probability of adoption by major group according to a simple logit model.

We remove groupings with fewer than ten MNCs in either group.

Figure 9: Major Shares, HTPs vs. Other MNCs

Computer and Electronic Product Manufacturing

Chemical Manufacturing

Professional, Scientific, and Technical Services

Publishing Industries (except Internet)

Merchant Wholesalers, Durable Goods

Merchant Wholesalers, Nondurable Goods

Machinery Manufacturing

Miscellaneous Manufacturing

Transportation Equipment Manufacturing

Management of Companies (Holding Companies)

Mining

Food Manufacturing

Fabricated Metal Manufacturing

Electrical Equipment, Appliance, and Component Manufacturing

Primary Metal Manufacturing

Data Processing, Hosting, and Related Services

Hybrid Tax Planners

10%

Other MNCs

Notes: This figure reports industry shares by major group. We remove groupings with fewer than ten MNCs in

either group.

Figure 10: Predicted Probability of HTP Adoption by Firm Size

0.15

Domestic Assets

Domestic Sales

Foreign Assets

0.10

Foreign Sales

0.05

0−25th Percentile

25−50th Percentile

50−75th Percentile

75−100th Percentile

Notes: This figure reports the predicted probability of adoption by size according to four simple logit models that

examine four different measures of firm size. We remove industry groupings with fewer than ten MNCs in either

group.

Figure 10 provides estimates of the predicted probability of HTP adoption according to

several different measures of MNC size. Each color provides estimates computed from a separate

logit model corresponding to a different measure of size (domestic assets, domestic sales, foreign

assets, and foreign sales). We bin the MNCs into size quartiles. All of these measures indicate

that larger MNCs tend to adopt HTP structures at higher rates.

In addition to industry and firm size, we examine whether a set of other observable characteristics is predictive of HTP adoption. Table A.4 reports coeﬀicient estimates for a battery of logit

regressions that predict HTP adoption based on whether an MNC claims a tax credit for R&D;

its age (binned by quartile with the youngest firms set as the reference category); the average

statutory foreign ETR that it faces (and its share of foreign sales in jurisdictions with unobserved

statutory rates); a measure of its geographic exposure to Check the Box; whether it operated in

in Ireland, the Netherlands, or Luxembourg prior to adoption; whether it has negative domestic

earnings; and its advertising to sales ratio (a proxy for intangibles used in, e.g., Grubert and

Slemrod, 1998).18 These models use observations from adopting MNCs in the period prior to

To compute a firm-level measure of exposure to Check the Box, we first compute a country-level measure of

exposure γc = πcp /πc , where πcp are aggregate foreign earnings generated by pass-through foreign aﬀiliates (FDEs)

for country c, and πc are aggregate foreign earnings for all foreign aﬀiliates in country c for the years that we

observe FDE earnings (2006, 2008, 2012, and 2016). Next, for each MNC i we compute the share of foreign

sales by country sic in the period prior to adoption. Finally, for each

∑ firm, we compute the exposure measure

as a weighted average of the country-level exposure measures, γi = c∈Ci γc sic for the set of countries Ci where

MNC i has positive sales in the period prior to adoption, where the weights sic are the firm-level country shares

calculated in the second step. For firms that never adopt, we compute this exposure measure in every year.

adoption, and include all observations for never-adopters. Table A.5 combines the variables from

these regressions into a single logit regression. Both tables provide specifications with year fixed

effects only and with interactions between industry, sales quartile bins, asset quartile bins, and

year fixed effects.19 Finally, Table A.6 provides logit estimates from a set of MNCs that also

appear in Compustat to examine whether the identity of firms’ auditors plays a role in adoption

of hybrid structures.20

Of the characteristics listed above, only a few appear to be predictive of HTP adoption. Unsurprisingly, MNCs that previously operated in the jurisdictions where we detect HTP structures

(Ireland, the Netherlands, and Luxembourg) are more likely to adopt them. More generally, the

geographic distribution of MNC activity is predictive of adoption—MNCs that operate in countries where Check the Box is more heavily used may be able to shift income more easily, which

may explain why this measure of exposure is predictive of adoption. Finally, MNCs that are in a

domestic loss position are less likely to adopt HTP structures—this also has an intuitive explanation as these firms may have tax credits that offset income tax levied on repatriated foreign

earnings. Confidence intervals for these estimates are generally wider in the combined regression,

and when more granular fixed effects are included. These tables show that, even though some

characteristics are predictive of adopting an HTP, most variables are not statistically significant

after the inclusion of industry and size fixed effects. While some variables statistically correlate

with adoption, it is hard to predict which specific firms will adopt an HTP, even conditional on

a rich set of covariates.

Overall, the results in this section confirm the conventional wisdom that larger firms in

industries that rely on intellectual property are more likely to adopt HTPs. At the same time,

there remains considerable unexplained variation in HTP adoption even after controlling for a

large set of firm characteristics that have been used as proxies for tax planning, indicating the

existence of idiosyncratic costs and benefits of adoption that vary across firms.

Note that the number of observations goes down in the second Column (from 8,608 to 4,920). This is because

to be included, each bin must have both HTP adopters and non-adopters.

We divide the firms into three groups, multinationals with a Big 4 auditor (Ernst & Young, PWC, Deloitte,

and KPMG), medium size auditors, and small auditors. These results suggest that auditors do not play a

significant role in the adoption of HTP structures.

Hybrid Tax Planning Structures and Multinational Activity

Below, we examine changes in domestic and foreign economic activity of adopting MNCs. We

start with a set of descriptive facts that compare select aggregate outcomes for MNCs that

utilize HTPs to other MNCs that do not rely on these arrangements. We then estimate staggered

difference-in-differences models that control for firm characteristics in order to compare outcomes

for MNCs that do and do not adopt hybrid tax structures. These models allow us to tie changes

in firm activity to the timing of HTP adoption.

5.1 Comparison of Aggregate Trends

While the structures we study have been suspected of being used for profit shifting, lack of tax

data prevented prior researchers from confirming this role in a systematic manner. We therefore

start by examining whether hybrid tax planning firms engage in the kinds of transactions that are

associated with profit shifting. Figure 11 compares hybrid tax planners to other MNCs along a

number of these dimensions. For a given outcome, this and related figures plot selected outcomes

as a percentage of 1996 levels. The secondary y-axis to the right of each graph indicates the level

values for the group of hybrid tax planning firms. In Panel A, we first document that hybrid

tax planning firms generate a much larger aggregate loan balance between their related CFCs

when compared to other MNCs. These balances may be related to interest stripping strategies,

as discussed in Panel A of Figure 2.

Panel B of Figure 11 shows that hybrid tax planning MNCs experience a much faster rise in

the book value of foreign intangible assets when compared to non-HTP MNCs. This growth is

consistent with the use of HTP structures to shift income generated by intangible assets. Panel

C shows that these MNCs also increased compensation for services paid by CFCs to parent

companies—which includes cost sharing payments that are used to transfer intangible assets

from the US to foreign aﬀiliates. Starting in 2008, Schedule G of Form 5471 allows us to observe

whether a parent had any cost sharing agreements with one of its CFCs; Panel D shows that

hybrid tax planning firms are also more likely to engage in these agreements.

Figure 1 compares average foreign ETRs of hybrid tax planners with that of other US MNCs

in our sample. At the beginning of the sample period, both types of MNCs paid taxes on foreign

E&P at similar rates. Starting in 2002, however, there is a striking divergence in the evolution of

Figure 11: Mechanisms for Profit Shifting

(A) Loans Between Related CFCs

(B) Foreign Intangible Assets

400

1,750

2,000%

Other MNCs

1,500

Hybrid Tax Planners

1,500%

Hybrid Tax Planners

1,250

200

1,000%

500

500%

$ billions

750

300

1,500%

$ billions

1,000

1,000%

Other MNCs

100

500%

250

1995

2000

2005

2010

2015

800%

1995

2000

2005

2010

2015

(D) Share with Cost Sharing Agreements

30%

Other MNCs

Hybrid Tax Planners

600%

400%

$ billions

Other MNCs

20%

Hybrid Tax Planners

10%

200%

1995

2000

2005

2010

2015

2008

2010

2012

2014

2016

Notes: These figures show the evolution of aggregate loans between related CFCs (Panel A), foreign intangible

assets (Panel B), payments from CFCs to US parent companies for technical services (Panel C), and the share

of MNCs in each group with active cost sharing agreements with a CFC (Panel D). For comparability, aggregate

values for both groups are normalized to 100% as of 1996 for Panels A through C. The light green line shows

values for a group of MNCs that eventually adopt at least one of the hybrid tax structures described in Section

2. The dark blue line shows aggregate values for MNCs that did not adopt any of these structures during the

sample window. For most outcomes, the right-hand axis displays dollar value in billions, relative to the aggregate

1996 dollar value for HTP MNCs.

each group’s ETR. By 2016, hybrid tax planners faced a foreign ETR that was about half of that

incurred by other MNCs. Figure 11 provides important context for the decline in foreign ETRs.

While one may suppose that declining statutory rates around the world may be responsible for

this decline, Figure 11 shows that MNCs with the largest reduction in foreign ETRs were also

engaging in behavior that has been linked to aggressive tax planning. Indeed, as we see in Figure

1, firms that did not adopt HTP structures experienced a much smaller decline in their foreign

ETR during our sample period.

Having shown that hybrid tax planning structures are likely used for profit shifting, we now

examine whether hybrid tax planning firms also deferred more income abroad. Panel A of Figure

12 shows that, relative to foreign E&P, hybrid tax planning firms saw faster declines in royalty

Figure 12: Evidence of Deferral

(A) Royalties to Parent

(B) Cash Held Abroad

150%

1,500%

700

Other MNCs

600

Hybrid Tax Planners

500

1,000%

400

300

50%

500%

$ billions

100%

200

Other MNCs

100

Hybrid Tax Planners

1995

2000

2005

2010

2015

1995

2000

2005

2010

2015

Notes: These figures show the evolution of royalty payments from CFCs to domestic parent entities as a share

of foreign E&P (Panel A) and of aggregate foreign cash balances (Panel B) for two groups of US MNCs. For

comparability, aggregate values for both groups are normalized to 100% as of 1996 in Panel B. The light green

line shows values for a group of MNCs that eventually adopt at least one of the hybrid tax structures described

in Section 2. The dark blue line shows aggregate values for MNCs that do not adopt any of these structures

during the sample window. For most outcomes, the right-hand axis displays dollar value in billions, relative to

the aggregate 1996 dollar value for HTP MNCs.

payments from CFCs to parents. This result is consistent with a transition away from undeferred

royalty income. Consistent with this interpretation, Panel B also shows that hybrid tax planning

firms saw large increases in cash held abroad compared to MNCs that did not adopt HTPs.

While the fact that MNCs accumulated cash abroad during the last two decades is well known,

this figure shows that the bulk of this growth occurred among the 300 firms that we observe with

hybrid tax planning agreements through Ireland, Netherlands, and Luxembourg.

The results in Figures 11 and 12 provide prima facie evidence of the specific mechanisms

through which hybrid tax planning strategies operate. The ability to shift profits to lower-tax

countries allows MNCs to avoid foreign income tax and defer US income tax.

These figures also highlight the value of using tax data, because they allow us to (i) identify the

adoption of specific tax planning structures (Figure 4), (ii) link the adoption of HTP structures to

specific profit shifting mechanisms (Figures 11 and 12), and (iii) measure the associated impact

on ETRs (Figure 1).

We now examine whether firms that benefited from tax planning also experienced differential

evolution in their real operations. Figure 13 shows that, while hybrid tax planning firms and

non-HTP MNCs had similar patterns of economic activity prior to 1997, their economic activities

diverged over the same time period that hybrid tax planning strategies were adopted. Panel A

shows that hybrid tax planners had larger increases in domestic capital assets; Panel B shows

Figure 13: Hybrid Tax Planning and Real Economic Activity

(A) Domestic Capital Assets

(B) Foreign Capital Assets

250%

1,000

200%

500

100%

500

50%

200%

400

150%

100%

250

Other MNCs

300

200

Other MNCs

50%

Hybrid Tax Planners

2000

2005

2010

100

Hybrid Tax Planners

1995

$ billions

750

$ billions

150%

2015

1995

2000

2005

2010

2015

300%

250

200

150

100

100%

Other MNCs

$ billions

200%

Hybrid Tax Planners

1995

2000

2005

2010

2015

Notes: These figures show the evolution of aggregate domestic capital assets (Panel A), foreign capital assets

(Panel B), and domestic wages (Panel C) for two groups of US MNCs. For comparability, aggregate values for

both groups are normalized to 100% as of 1996. The light green line shows aggregate values for a group of MNCs

that eventually adopt at least one of the hybrid tax structures described in Section 2. The dark blue line shows

aggregate values for MNCs that do not adopt any of these structures during the sample window. The right-hand

axis displays dollar value in billions relative to the aggregate 1996 dollar value for HTP MNCs.

that these firms also accumulated more foreign capital assets; and Panel C shows larger increases

in domestic payroll. Across all of these measures, declines in foreign ETRs were accompanied by

increases in foreign and domestic economic activity.

5.2 Estimating Staggered Difference-in-Differences Models

We now show that the changes in firm outcomes described above are closely tied to the adoption

of HTPs. Recent literature has provided several alternative models that researchers may use

to produce difference-in-differences estimates in staggered contexts. We provide estimates for

three of these models. Our main specification relies on the “stacked” design from Cengiz, Dube,

Lindner and Zipperer (2019). We choose this estimator as our main specification because it lends

itself easily to the addition of propensity score weights, which we describe later in this section.

The stacked design creates a data set for each cohort. This dataset includes MNCs that adopt

a hybrid structure, as well as MNCs that do not adopt a hybrid structure in the six years before

and after the adoption year c. Non-adopting MNCs may therefore be repeated in the regression

dataset as comparison units for different cohorts. Formally, the stacked design estimates the

regression equation,

Yict = αic + λct +

∑

µℓ 1 {t − c = ℓ} + vict ,

(1)

ℓ

where i indexes MNCs, c is the year in which a particular cohort first adopts a hybrid structure,

and t indexes years. ℓ is an indicator for the relative number of periods after MNC i adopts a

foreign tax planning structure.21 αic and λct are MNC-by-cohort and year-by-cohort fixed effects.

We estimate this regression for various outcomes Yict . In all specifications, we cluster standard

errors at the MNC level.

We interpret results of Equation 1 as measuring dynamic changes in firm-level outcomes of

adopting MNCs relative to non-adopting MNCs. Relative to the results in the prior section,

these estimates help tie changes in firm outcomes to the timing of adoption. This approach also

addresses the concern that firm outcomes are driven by concomitant shocks to firms with characteristics that are related to tax planning (e.g. larger firms, more IP-intensive firms, or firms

in different industries). To do so, we estimate alternative specifications that interact year fixed

effects with a set of pre-adoption covariates for MNCs to allow for time-varying heterogeneity

across industries, across foreign and domestic firm sales bins, and across foreign and domestic

bins for intangible assets.22 As a robustness check, we also estimate effects using the stacked specification with inverse probability weights (IPW) as well as an alternative specification proposed

by Sun and Abraham (2021) and a standard TWFE estimator.

While these firm-level comparisons are informative of the role of HTPs in driving the aggregate

changes described in the previous section, a key question is whether the estimates of Equation 1

can be interpreted as causal effects of HTPs. The usual assumptions for a causal interpretation

For comparison, a standard two-way fixed effects specification does not use repeated comparison units and

instead estimates the regression equation

∑

µℓ 1 {t − c = ℓ} + vit .

(2)

Yit = αi + λt +

ℓ

To be precise, this implies an augmented version of Equation 1 where λct is replaced by

G is a set of groups for which we include group-by-cohort-by-year fixed effects.

∑

g∈G λgct , where

include parallel trends and no anticipatory behavior. The descriptive evidence provided in the

previous section shows that the evolution of outcomes for adopting and non-adopting MNCs was

strikingly similar prior to the bulk of adoptions in the mid-2000s. Additionally, pre-trend coeﬀicients in the event study plots provided below are generally insignificant. Regarding anticipation,

recall that our data only provides observations in even years. MNCs would therefore have to

adjust behavior two years in advance for this form of bias to be present. Because the parallel

trends assumption is inherently untestable and that HTP adoption is an endogenous choice of

the firm, it is important to consider that MNCs may select into HTPs because they have more

to gain from tax planning.

5.2.1

Estimates of Changes in Financial and Tax Outcomes

Figure 14 plots estimates of Equation 1 for a set of financial and tax outcomes. We report two

specifications for each outcome. Specification 1 (in black) does not include additional controls.

Specification 2 (in orange) includes year-by-cohort-by-industry and year-by-cohort-by-group fixed

effects, where groups include domestic and foreign sales quartiles and domestic and foreign intangible asset quartiles. Quartiles are computed using the period prior to adoption for each

cohort.

Panel A provides estimates for the log of the balance of loans between CFCs, Panel B for the

log of intangibles held abroad, and Panel C for the log of cash held abroad. Across these three

outcomes, we observe similar trends for HTP-adopting and non-adopting MNCs prior to the

adoption of a hybrid structure followed by relative increases for HTP-adopting MNCs after the

period of adoption. Consistent with these mechanisms and the result of Figure 1, Panel D shows

that the foreign ETRs of HTP-adopting MNCs gradually declined relative to non-adopting MNCs

following adoption. Six years after adoption, MNCs experience a reduction in their foreign ETR

of between three and four percentage points. For all of these outcomes, we find that inclusion

of size-bin-by-cohort-by-year fixed effects and industry-by-cohort-by-year fixed effects does not

significantly impact the estimates.

Table 4 provides aggregated estimates using a similar specification that replaces relative

time dummies with pre and post dummies. The estimates in Column (1) indicate that loans

between CFCs increased by 40%, foreign intangible assets increased by 68%, and cash held

abroad increased by 56% on average relative to the period prior to adoption. We also estimate

Figure 14: Profit Shifting Mechanisms, Deferral, and Foreign ETRs

(B) Foreign Intangible Assets

Relative Change in Logged Outcome

(A) Loans Between Related CFCs

1.0

0.5

0.0

−0.5

−6

−4

−2

1.0

0.5

0.0

−6

−4

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

(D) Foreign Effective Tax Rate

Relative Change in Outcome

Relative Change in Logged Outcome

−2

Time Relative to Adoption

1.0

0.5

0.0

0.05

0.00

−0.05

−0.5

−6

−4

−2

−6

Time Relative to Adoption

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding foreign outcome listed in

each panel, using 2-year pooled average data from Compustat and dropping odd years to match the IRS SOI

International Business Tax Sample. R&D Intensity (Panel A) is calculated as the ratio of annual R&D to the

MNC’s most recent sales value pre-adoption, and is restricted to be less than 1. Specification 1 (in black) does

not include additional controls. Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and

year-by-cohort-by-group fixed effects, where groups include domestic and foreign sales quartiles and domestic and

foreign intangible asset quartiles, and where quartiles are computed using pre-adoption values for each cohort.

an average decline in foreign ETRs of 3.9 percentage points in the period after adoption. Column

(2) shows that these estimates are stable across specifications that include interactions between

bins of firm size and year fixed effects, bins of intangible assets interacted with year fixed effects,

and industry-by-year fixed effects, suggesting that our results are not driven by comparisons

across firms in different industries, in different domestic and foreign size categories, or that are

more or less dependent on intellectual property.

Table 4: Profit Shifting Mechanisms and Foreign ETRs

(1)

(2)

(3)

(4)

(5)

0.681*** 0.735***

(0.172)

(0.177)

248

1524

1490

0.637***

(0.193)

208

1221

0.590***

(0.171)

250

1532

0.709***

(0.175)

229

1757

0.563*** 0.555***

(0.133)

(0.136)

257

2054

2000

0.469**

(0.149)

214

1518

0.539**

(0.173)

257

2023

0.472**

(0.144)

252

2037

0.401*

(0.157)

233

1263

0.524**

(0.161)

233

1240

0.505**

(0.174)

197

1041

0.344+

(0.176)

238

1294

0.491**

(0.163)

205

1597

-0.039**

(0.015)

257

2100

-0.032*

(0.016)

257

2043

-0.024

(0.018)

214

1529

-0.037*

(0.017)

257

2046

-0.042**

(0.015)

252

2042

Yes

Stacked

Yes

Stacked

Yes

Stacked

Yes

TWFE

Panel A

Foreign Intangibles

Num. Treated

Num. Control

Panel B

Foreign Cash

Num. Treated

Num. Control

Panel C

Rltd. CFC Loans

Num. Treated

Num. Control

Panel D

Foreign ETR

Num. Treated

Num. Control

MNC & Year x Cohort FEs

MNC & Ind., Size x Yr FEs

MNC & Ind., Size x Yr x Cohort FEs

Inverse Prob. Weights

Model

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table provides estimates of the difference-in-differences model discussed in Section 5.2 for the corresponding foreign outcome listed in each panel, where Columns (1) - (3) use Equation 1 and Columns (4) and

(5) use Equation 2. Column (1) does not include additional controls. Columns (2) - (5) include year-by-cohortby-industry and year-by-cohort-by-group fixed effects, where groups include domestic and foreign sales quartiles

and domestic and foreign intangible asset quartiles, and where quartiles are computed using pre-adoption values.

Column (3) uses inverse probability-weighted data. Column (4) estimates an alternative specification from Sun

and Abraham (2021). Column (5) estimates a standard TWFE specification.

We conduct an additional exercise to provide an alternative summary of the average change in

the foreign effective tax rate to reduce potential measurement error. In contrast to the estimates

above, which use year-by-year foreign ETRs, Table 5 shows pooled estimates where ETRs are

Table 5: Pooled Foreign ETRs

Unweighted Foreign ETR

Weighted Foreign ETR

Num. Treated

Num. Control

Sample

Outcomes

Years Included

(1)

(2)

−0.035+

(0.021)

−0.092**

(0.034)

−0.032*

(0.016)

−0.061*

(0.028)

155

1101

SOI

Even

308

Compustat

All

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table estimates a difference-in-differences model using a two-period specification. Foreign effective

tax rates are computed using aggregate taxes and earnings from each period to reduce measurement error

relative to an annual specification. Column (1) provides estimates using measures from SOI data. Column (2)

provides estimates using measures from Compustat. Results are provided for an unweighted specifciation as well

as a weighted specification where firms are weighted by aggregate pre-period earnings. Firms are only included

if they have positive earnings in both pre and post periods.

computed at the firm level for all pre and post periods. We include firms that have positive

aggregate earnings in both periods.23 This reduces year-to-year volatility in the firm-level ETR

that can be generated by losses and tax credits. We winsorize these rates to ensure they do not

exceed 100% so that outliers do not skew the average. Finally, we compute the average foreign

ETR using two different measures. The first measure, which is also used for Figure 14 above, uses

SOI tax data to compute the foreign ETR following the methodology described in Section 3.3.

The second measure uses foreign taxes and pretax income as reported in Compustat following the

method used to calculate foreign effective tax rates in Dyreng, Hanlon, Maydew and Thornock

(2017). There are advantages and disadvantages to both measures—SOI data provide a larger

sample of multinationals, but only even years are present in our data. Compustat has smaller

coverage, excluding private firms, but these firms are observed annually. Finally, we include two

specifications for each measure—the first is an unweighted regression and the second weights firms

by aggregate pre-period foreign income. Results are shown in Table 5. The unweighted pooled

results are broadly similar to the event study estimates, showing declines of between 3.2 and

We use the stacked panel in this analysis in order to create distinct pre and post periods for the comparison

group of firms that do not adopt hybrid structures.

3.5 percentage points among firms that adopt hybrid structures. Weighted results demonstrate

much stronger reductions in foreign ETRs after adoption of hybrid structures, between 6.1 and

9.2 percentage points depending on the measure used. These results suggest that although firms

generally experienced reductions in foreign tax rates after adopting hybrid structures, these

reductions were concentrated in the largest adopting firms.

5.2.2

Estimates of Changes in Real Economic Activity

To measure changes in real activity, we first estimate Equation 1 for a set of outcomes that

includes foreign assets, domestic wages, and domestic capital assets. Foreign assets are reported

on IRS Form 5471, domestic wages and capital assets are reported on Form 1120, and domestic

investment is measured as the sum of reported assets placed into service on Form 4562. Figure

15 reports the results of the event study analyses for these outcomes. As with the financial and

tax outcomes, estimates are not significantly different from zero in the pre-adoption period for

any of the real outcomes. Following the adoption of a hybrid tax planning structure, we observe

significant relative increases in foreign and domestic capital assets and domestic wages.

Table 6 summarizes these estimates using an aggregated specification. The specification with

more granular controls (Column (2)) reports an increase in domestic investment of 27% and a

corresponding increase in domestic capital assets of 16%. Panel C reports a 40% increase in

foreign capital assets, while Panel D reports a 20% increase in domestic payroll.

Figure 15: Event Studies: Hybrid Tax Planning and Real Economic Activity

(B) Domestic Capital Assets

Relative Change in Logged Outcome

(A) Domestic Investment

0.6

0.4

0.2

0.0

−6

−4

−2

0.3

0.2

0.1

0.0

−0.1

−6

−4

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

0.25

0.00

−0.25

−2

(D) Domestic Wages

Relative Change in Logged Outcome

MNC & Industry, Sales, Asset x Yr x Cohort FEs

0.50

−4

MNC & Year x Cohort FEs

−6

−2

Time Relative to Adoption

0.4

0.2

0.0

−6

Time Relative to Adoption

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding outcome listed in each panel.

Specification 1 (in black) does not include additional controls. Specification 2 (in orange) includes year-bycohort-by-industry fixed effects and year-by-cohort-by-group fixed effects, where groups include domestic and

foreign sales quartiles and domestic and foreign intangible asset quartiles, and where quartiles are computed

using pre-adoption values for each cohort.

Table 6: Hybrid Tax Planning and Real Economic Activity

(1)

(2)

(3)

(4)

(5)

0.170*

(0.079)

241

1802

0.267**

(0.088)

241

1756

0.307**

(0.109)

209

1472

0.285**

(0.104)

242

1798

0.206*

(0.086)

240

1919

0.098*

(0.046)

250

1882

0.156**

(0.051)

250

1828

0.147**

(0.056)

214

1528

0.076

(0.048)

252

1856

0.105*

(0.049)

247

1964

0.351*** 0.404***

(0.101)

(0.100)

257

1906

1852

0.377***

(0.106)

214

1492

0.340** 0.366***

(0.111) (0.105)

257

247

1866

1978

0.150**

(0.055)

250

1872

0.204***

(0.059)

250

1819

0.182**

(0.062)

214

1519

0.186*

(0.078)

252

1850

0.148*

(0.060)

247

1961

Yes

Stacked

Yes

Stacked

Yes

Stacked

Yes

TWFE

Panel A

Domestic Investment

Num. Treated

Num. Control

Panel B

Domestic Capital

Num. Treated

Num. Control

Panel C

Foreign Capital

Num. Treated

Num. Control

Panel D

Domestic Wages

Num. Treated

Num. Control

MNC & Year x Cohort FEs

MNC & Ind., Size x Yr FEs

MNC & Ind., Size x Yr x Cohort FEs

Inverse Prob. Weights

Model

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table provides estimates of the difference-in-differences model discussed in Section 5.2 for the corresponding outcome listed in each panel, where Columns (1) - (3) use Equation 1 and Columns (4) and (5)

use Equation 2. Column (1) does not include additional controls. Columns (2) - (5) include year-by-cohortby-industry and year-by-cohort-by-group fixed effects, where groups include domestic and foreign sales quartiles

and domestic and foreign intangible asset quartiles, and where quartiles are computed using pre-adoption values.

Column (3) uses inverse probability-weighted data. Column (4) estimates an alternative specification from Sun

and Abraham (2021). Column (5) estimates a standard TWFE specification.

Figure 16: Event Studies: Hybrid Tax Planning and R&D Data from Compustat

(B) Log R&D

Relative Change in Outcome

(A) R&D Intensity

0.06

0.04

0.02

0.00

−0.02

−6

−4

−2

0.25

0.00

−0.25

−6

Time Relative to Adoption

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding R&D outcome listed in each

panel. R&D Intensity (Panel A) is calculated as the ratio of annual R&D to the MNC’s most recent sales value

pre-adoption, and is restricted to be less than 1. Specification 1 (in black) does not include additional controls.

Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and year-by-cohort-by-group fixed

effects, where groups include domestic and foreign sales quartiles and domestic and foreign intangible asset

quartiles, and where quartiles are computed using pre-adoption values for each cohort.

5.2.3

R&D Activity

To examine R&D activity, we merge in data from Compustat, which gathers R&D expenditures

from MNCs’ consolidated financial statements. Because this procedure introduces sampling

attrition, as not all MNCs in the tax data are public, we separate this analysis from the estimates

contained above that rely on administrative tax data.

In order to match the biannual frequency of observations in the SOI sample, we create a

two-year pooled average for each Compustat variable of interest and focus the estimation on the

same years available in the tax data. We study two measures of R&D using the same event study

model discussed above. The first measure, R&D intensity, is computed as the ratio of R&D to

revenue. We fix the denominator to the period prior to HTP adoption. A valuable feature of

this measure is that it does not exclude firm-years in which zero R&D expense is reported. The

second measure is the log of R&D. This measure leads to further sample attrition because of the

large number of firm-years that report zero R&D expenditures.

The results of these estimations are shown in Figure 16. As with other measures of real

economic activity, this figure shows that adopting firms experience a significant increase in both

R&D intensity and log R&D following HTP adoption. These results are summarized in Table 7.

Table 7: Hybrid Tax Planning and R&D Data from Compustat

(1)

(2)

(3)

(4)

(5)

0.026**

(0.008)

136

452

0.051

0.054

0.026**

(0.010)

124

385

0.051

0.054

0.032**

(0.010)

113

335

0.047

0.054

0.033***

(0.009)

124

380

0.051

0.054

0.034***

(0.009)

124

380

0.051

0.054

0.319*** 0.285***

(0.068)

(0.078)

298

252

0.362***

(0.077)

235

0.241***

(0.069)

266

0.268**

(0.084)

266

Yes

Stacked

Yes

Stacked

Yes

TWFE

Panel A

R&D Intensity

Num. Treated

Num. Control

Avg. R&D Intensity (All Firms)

Avg. R&D Intensity (Treated Firms)

Panel B

Log R&D

Num. Treated

Num. Control

MNC & Year x Cohort FEs

MNC & Ind., Size x Yr FEs

MNC & Ind., Size x Yr x Cohort FEs

Inverse Prob. Weights

Model

Yes

Stacked

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table provides estimates of the difference-in-differences model discussed in Section 5.2 for the

corresponding R&D outcome listed in each panel, where Columns (1) - (3) use Equation 1 and Columns (4) and

(5) use Equation 2. Estimations use 2-year pooled average data from Compustat, where odd years are dropped

to match the IRS SOI International Business Tax Sample. R&D Intensity (Panel A) is calculated as the ratio of

annual R&D to the MNC’s most recent sales value pre-adoption, and is restricted to be less than 1. Column (1)

does not include additional controls. Columns (2) - (5) include year-by-cohort-by-industry and

year-by-cohort-by-group fixed effects, where groups include domestic and foreign sales quartiles and domestic

and foreign intangible asset quartiles, and where quartiles are computed using pre-adoption values. Column (3)

uses inverse probability-weighted data. Column (4) estimates an alternative specification from Sun and

Abraham (2021). Column (5) estimates a standard TWFE specification.

Column (2) reports that R&D intensity increased by 0.026, which corresponds to a 48% increase

relative to the average R&D intensity of 0.054 for HTP-adopting firms. Panel B shows that log

R&D increased by 0.285 following adoption of an HTP. The larger estimate for R&D intensity

suggests that incorporating extensive-margin responses is important in this setting and leads to

larger estimates of the change in R&D activity following the adoption of an HTP structure.

5.2.4

Further Robustness Checks

To ensure that we are comparing similar firms in our regressions, we extend the estimation

of Equation 1 by applying inverse probability weighting to the sample of MNCs. A standard

approach to this form of weighting in an unstaggered difference-in-differences setting estimates a

single propensity score for each unit. These scores are canonically estimated using data in periods

prior to HTP adoption. In staggered designs, however, control units may serve as comparisons for

multiple treated cohorts, and there is not a well defined pre-period. The advantage of the stacked

estimator is that it can be thought of as combining a set of unstaggered difference-in-differences

datasets, one for each treated cohort.

As shown in Section 4.3, we find evidence for considerable variation in rates of HTP adoption

by industry and by firm size, which motivates our usage of industry and firm-size fixed effects in

the preceding difference-in-differences estimates to control for time-varying heterogeneity along

these dimensions. Section 4.3 also shows that several other observable characteristics predict

MNC adoption of hybrid tax structures (see Tables A.4 and A.5). Propensity score weighting

offers a convenient way to control for potential selection bias that may be introduced by a lack

of balance along these observable dimensions between adopting and non-adopting MNCs.

We approach the computation of propensity scores with a similar logic to the design of the

stacked estimator by computing a set of propensity scores for each treated cohort and nevertreated observations that are used as cohort-level comparisons. We use the same predictor

variables examined in Section 4.3 and Tables A.4 and A.5.24 This gives us a set of propensity

scores for never-treated units that are used as comparison units for multiple cohorts. These

scores increase comparability within cohorts across adopting and non-adopting MNCs for other

observable characteristics. The score incorporates pre-adoption measures of R&D activity; firm

age; average foreign statutory rates; exposure to geographies where it may have been easier to

use Check the Box; prior activity in Ireland, the Netherlands, and Luxembourg; whether or not

the MNC had negative domestic income; and the firm’s advertising intensity relative to sales.

Because the stacked estimator also explicitly duplicates never-treated units when they are used

as comparisons for multiple cohorts, we can apply these sets of propensity scores directly in

In contrast to the logit estimates shown in Section 4.3, we use a random forest model to compute propensity

scores. In our context, random forest models outperform logit models from a predictive standpoint, and can

flexibly account for non-linearities. They are also theoretically invariant to transformations such as the natural

log, which would require dropping observations with non-positive values.

the stacked design in a similar way to an unstaggered propensity score weighted difference-indifferences model.

Columns (3) of Tables 4, 6, and 7 show that we obtain similar estimates of changes in firm

outcomes when we expand the model in Equation 1 by including inverse probability weights. See

Figure A.1 and A.2 for corresponding event studies that also use IPW.

We also explore the robustness of our results to using alternative estimators. Columns (4)

of Tables 4, 6, and 7 show that we generally obtain similar estimates when using the estimator

of Sun and Abraham (2021). Columns (5) of Tables 4, 6, and 7 show that we also find similar

effects when using a simpler two-way fixed effects estimator.25

Overall, the results in this section show that the adoption of HTP structures precedes large

increases in tax avoidance behavior as well as increases in real economic activity, both at home

and abroad.

5.3 Domestic to Foreign Profit Shifting and the Role of Cost Sharing

Agreements

In Section 2.2.1, we briefly discuss how MNCs can use a contract known as a cost sharing

agreement (CSA) to transfer IP to foreign aﬀiliates. In Section 5.1, we show that HTPs utilize

CSAs at a much higher rate than non-HTP MNCs. One underlying reason for this is that MNCs

with hybrid tax planning structures may be able to use CSAs to shift profits from the US to

their low-tax hybrid entities.

Note that this behavior may in some instances be permitted under US tax law. One such

legal approach to profit shifting, which is described in publicly available legal briefs, was used

by MNCs in the early 2000s. During this period, MNCs were able to exclude certain R&D

costs that the IRS argued were subject to a CSA (and therefore, would have been allocated to

an MNC’s foreign aﬀiliates). Specifically, 1995 regulations issued by the IRS did not explicitly

require stock-based compensation to be covered by CSAs.

After extensive litigation, a technology firm won a ruling in the US Tax Court that permitted

the exclusion of these costs. Although the US Treasury attempted to clarify its regulations to

disincentivize this type of tax planning, other MNCs continued to engage in similar strategies

during a long period of regulatory uncertainty. Appendix C provides additional institutional

See Figure A.3 and A.4 for event studies using the estimator of Sun and Abraham (2021) and Figure A.5

and A.6 for event studies using the TWFE estimator.

detail about this strategy, along with empirical analysis that demonstrates a substantial behavioral shift by MNCs that held CSAs during the mid-2000s. We find evidence that MNCs

exposed to this regulatory uncertainty expanded their use of stock-based compensation, both

in absolute terms and as a fraction of their overall wage bill. We also find evidence that these

MNCs increased their overall R&D activity and reported larger amounts of their wage expenses

as eligible R&D costs for a domestic tax credit.

Conclusion

Complex tax planning strategies have been a focus of media attention and have played an important role in motivating international tax reforms. Despite this focus, policy makers, practitioners,

and academics lack a comprehensive understanding of the prevalence of these strategies and their

role in explaining some key facts surrounding international taxation.

Using an unique integration of tax data covering the domestic and foreign operations of US

MNCs, we help to fill this gap by reconstructing the ownership networks of foreign aﬀiliates of

US MNCs. This allows us to identify the adoption of three important tax planning strategies:

the Double Irish and two forms of Reverse Hybrid Mismatch arrangement, one through the

Netherlands, the other through Luxembourg. We show that these structures account for a

significant fraction of the foreign profits of US MNCs. We also link the use of these structures

to financial transactions that companies could use to shift profits to low-tax (and tax haven)

countries. Although only 17.5% of US MNCs adopt these structures, those that do obtain

a significant tax advantage over other MNCs. Remarkably, this small fraction of companies

generates a majority of foreign earnings by US MNCs and is responsible for the bulk of the

increase in cash held abroad over the period we study.

Our analyses use this tax data to reveal that the adoption of hybrid tax planning structures

is accompanied by significant changes in real economic activity. We find that firms adopting

these structures also have experience significant foreign and domestic growth.

While our analysis sheds light on multinational tax planning behavior over the last several

decades, a large number of policy changes have reshaped the incentives and feasibility of tax

planning in recent years. As new data becomes available, researchers should be able to determine

how firms have reacted to the new landscape and the extent to which current policy attempts to

curb profit shifting and tax avoidance have been successful.

References

Albertus, James F. 2019. “Does Foreign Tax Arbitrage Promote Innovation?” SSRN.

Altera Corporation & Subsidiaries v. Commissioner. 2019. “Altera Corporation & Subsidiaries

v. Commissioner.” 926 F.3d. 1061.

Altshuler, Rosanne, and Harry Grubert. 2006. “The Three Parties in the Race to the Bottom: Host

Governments, Home Governments and Multinational Corporations.” Florida Tax Review, 7(3): 137–

209.

Avi-Yonah, Reuven S. 2009. “Xilinx and the Arm’s-Length Standard.” Tax Notes, 123(10): 1231–

1236.

Bilicka, Katarzyna Anna. 2019. “Comparing UK Tax Returns of Foreign Multinationals to Matched

Domestic Firms.” American Economic Review, 109(8): 2921–53.

Bilicka, Katarzyna Anna, Yaxuan Qi, and Jing Xing. 2019. “Real responses to anti-tax avoidance

policies: Evidence from UK’s worldwide debt cap.” Working Paper.

Blouin, Jennifer L., and Leslie Robinson. 2020. “Double Counting Accounting: How Much Profit

of Multinational Enterprises Is Really in Tax Havens?” SSRN September 1, 2020.

Blouin, Jennifer L., and Linda K Krull. 2014. “Does Organizational Form Affect Firms’ Foreign

Operations? The Role of “Check-the-Box” on Multinational Tax Planning.” Vol. 107, 1–31, JSTOR.

Bodie, Zvi, Robert S. Kaplan, and Robert C. Merton. 2003. “For the Last Time: Stock Options

Are an Expense.” Harvard Business Review Web Site https://hbr.org/2003/03/for-the-last-t

ime-stock-options-are-an-expense (Accessed: July 4, 2023).

Bound, John, and Alan B Krueger. 1991. “The extent of measurement error in longitudinal earnings

data: Do two wrongs make a right?” Journal of labor economics, 9(1): 1–24.

Bound, John, Charles Brown, Greg J Duncan, and Willard L Rodgers. 1994. “Evidence

on the validity of cross-sectional and longitudinal labor market data.” Journal of Labor Economics,

12(3): 345–368.

Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer. 2019. “The effect of

minimum wages on low-wage jobs.” The Quarterly Journal of Economics, 134(3): 1405–1454.

Clausing, Kimberly A. 2016. “The Effect of Profit Shifting on the Corporate Tax Base in the United

States and Beyond.” National Tax Journal, 69(4): 905–934.

de Mooij, Ruud A., and Li Liu. 2018. “At A Cost: the Real Effect of Transfer Pricing Regulations

on Multinational Investment.”

Dowd, Tim, Paul Landefeld, and Anne Moore. 2017. “Profit shifting of U.S. multinationals.”

Journal of Public Economics, 148: 1–13.

Drucker, Jesse. 2010. “Google 2.4% Rate Shows How $60 Billion Is Lost to Tax Loopholes.”

Dube, Arindrajit, Ethan Kaplan, and Suresh Naidu. 2011. “Coups, Corporations, and Classified

Information.” Quarterly Journal of Economics, 126(3): 1375–1409.

Duhigg, Charles, and David Kocieniewski. 2012. “How Apple sidesteps billions in taxes.”

Dyreng, Scott, and Michelle Hanlon. 2021. “Tax avoidance and multinational firm behavior.”

Dyreng, Scott, Michelle Hanlon, Edward L Maydew, and Jacob R Thornock. 2017. “Changes

in Corporate Effective Tax Rates Over the Past Twenty-Five Years.” Journal of Financial Economics.

Faulkender, Michael W, Kristine W Hankins, and Mitchell A Petersen. 2019. “Understanding

the rise in corporate cash: Precautionary savings or foreign taxes.” The Review of Financial Studies,

32(9): 3299–3334.

Foley, Sean, and Hardi Kuvadia. 2021. “Stock-based compensation cost-sharing regs. following

Altera.” The Tax Adviser Web Site https://www.thetaxadviser.com/issues/2021/jun/stock-b

ased-compensation-cost-sharing-regs-altera.html (Accessed: July 4, 2023).

Gravelle, Jane G. 2013. “Tax Havens: International Tax Avoidance and Evasion.” Congressional

Research Service Web Site https://crsreports.congress.gov/product/details?prodcode=R40

623 (Accessed: July 4, 2023).

Grubert, Harry, and Joel Slemrod. 1998. “The Effect of Taxes on Investment and Income Shifting

to Puerto Rico.” Review of Economics and Statistics, 80(3): 365–373.

Guardian. 2018. “From Seattle to Luxembourg: How tax schemes shaped amazon.”

Guardian. 2019. “EU investigates Nike’s tax status in the Netherlands.”

Hardeck, Inga, and Patrick U. Wittenstein. 2018. “Assessing the Tax Benefits of Hybrid Arrangements - Evidence from the Luxembourg Leaks.” National Tax Journal, 71(2): 295–334.

Hines, James R., and Eric M. Rice. 1994. “Fiscal Paradise: Foreign Tax Havens and American

Business.” The Quarterly Journal of Economics, 109(1): 149–182.

IRS. 2011. “SOI Sampling Methodology and Data Limitations.” U.S. Internal Revenue Service https:

//www.irs.gov/pub/irs-prior/p946--2002.pdf (Accessed: February 7, 2022).

IRS. 2022a. “SOI Tax Stats - Controlled Foreign Corporations .” U.S. Internal Revenue Service Web

Site https://www.irs.gov/statistics/soi-tax-stats-controlled-foreign-corporations

(Accessed: February 7, 2022).

IRS. 2022b. “SOI Tax Stats - Corporation Complete Report.” U.S. Internal Revenue Service Web Site

https://www.irs.gov/statistics/soi-tax-stats-corporation-complete-report Accessed:

February 7, 2022).

IRS Acquiesces in Result, But Not Reasoning, of Xilinx Transfer Pricing Case. 2010. “IRS

Acquiesces in Result, But Not Reasoning, of Xilinx Transfer Pricing Case.” Internal Revenue Service

https://www.taxnotes.com/research/federal/irs-guidance/actions-on-decisions/irs-acq

uiesces-in-result%2C-but-not-reasoning%2C-of-xilinx-transfer/1fchf (Accessed: July 4,

2023).

Kleinbard, Edward D. 2013. “Through a latte, darkly: Starbucks’s stateless income planning.” Tax

Notes, 1515–1535.

MacKinlay, A. Craig. 1997. “Event Studies in Economics and Finance.” Journal of Economic Literature, 35(1): 13–39.

McWaters, Eleanor Moran. 2020. “Supreme Court Declines Certiorari in Altera, Stock-Based Compensation Cost Sharing Regulations Stand.” Crowell & Moring LLP Web Site https://www.crowel

l.com/en/insights/client-alerts/supreme-court-declines-certiorari-in-altera-stock

-based-compensation-cost-sharing-regulations-stand (Accessed: July 4, 2023).

Mutti, John, and Harry Grubert. 2009. “The effect of taxes on royalties and the migration of

intangible assets abroad.” In International trade in services and intangibles in the era of globalization.

111–137. University of Chicago Press.

Samarakoon, Navodhya. 2022. “The Effect of the Closure of the Double Irish Loophole on the

Location of U.S. Multinational Companies’ Profits.” SSRN November 23, 2022.

Suárez Serrato, Juan Carlos. 2018. “Unintended Consequences of Eliminating Tax Havens.” Duke

University Working Paper.

Sun, Liyang, and Sarah Abraham. 2021. “Estimating dynamic treatment effects in event studies

with heterogeneous treatment effects.” Journal of Econometrics, 225(2): 175–199. Themed Issue:

Treatment Effect 1.

Tax Reform Act of 1986. 1986. “Tax Reform Act of 1986.” 99th United States Congress.

Tørsløv, Thomas R, Ludvig S Wier, and Gabriel Zucman. 2018. “The Missing Profits of Nations.” National Bureau of Economic Research Working Paper 24701.

Yagan, Danny. 2015. “Capital Tax Reform and the Real Economy: The Effects of the 2003 Dividend

Tax Cut.” The American Economic Review, 105(12): 3531–3563.

Zucman, Gabriel. 2014. “Taxing across Borders: Tracking Personal Wealth and Corporate Profits.”

Journal of Economic Perspectives, 28(4): 121–48.

Zucman, Gabriel. 2015. The Hidden Wealth of Nations. The University of Chicago Press.

Zwick, Eric, and James Mahon. 2017. “Tax Policy and Heterogeneous Investment Behavior.” American Economic Review, 107(1): 217–48.

Appendices

These appendices include supplemental information and additional analyses. Appendix A

provides additional tables and figures related to the main analyses. Appendix B provides an

overview of measurement issues in several different sources of multinational data. Appendix C

describes how multinationals with cost sharing agreements may have used a legal strategy to

reclassify foreign costs as domestic costs for tax planning purposes.

A Additional Tables and Figures

Table A.1: MNC Summary Statistics

Domestic Outcomes

Tangible Assets

Wages

Investment

R&D

All MNCs

Mean

3024

583

267

Foreign Outcomes

Tangible Assets

Intangible Assets

Pretax E&P

Income Taxes

Cash

Sample Sizes

P25 P75

226 1981 9977

58 447 1636

10 141 1619

38 343

All MNCs

Mean

886

355

279

496

Foreign Aﬀiliate Counts

FDE Count

CFC Count

Hybrid Tax Planners

P25

P75

409

136

105

125

P25

P75

3635

P25 P75

SD Mean P25 P75

341 2894 14023 2659 201 1762 8557

114 928 2440

440

49 366 1308

17 216 1242

243

9 125 1704

0 106

644

27 184

Hybrid Tax Planners

3874

1856

1607

284

4579

All MNCs

Mean

4407

1125

356

179

Mean

1706

898

771

117

1429

P25

P75

811

462

323

430

Mean

P25

P75

512

Other MNCs

SD Mean P25

6598

668

3467

211

3062

148

378

8368

250

Hybrid Tax Planners

103

Other MNCs

188

P75

313 2692

91 1039

74 840

16 250

90 2775

Other MNCs

Mean P25

P75

3123

Notes: This table contains summary statistics for our stable sample as described in Section 3.2. This sample was created by identifying MNCs with

coverage in both the SOI Corporate and SOI International Business Tax samples, not including MNCs that did not have at least $500 million in domestic

assets as well as at least one CFC with $50 million in foreign assets. To ensure we do not disclose information about individual MNCs, the values listed

for P25 and P75 are the means of ten observations surrounding a given percentile.

Figure A.1: Event Studies: Profit Shifting Mechanisms and Foreign ETRs (IPW)

(B) Foreign Intangible Assets

Relative Change in Logged Outcome

(A) Loans Between Related CFCs

1.0

0.5

0.0

−0.5

−6

−4

−2

1.2

0.8

0.4

0.0

−0.4

−6

−4

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

(D) Foreign Effective Tax Rate

1.0

Relative Change in Outcome

Relative Change in Logged Outcome

−2

Time Relative to Adoption

0.5

0.0

0.05

0.00

−0.05

−0.5

−6

−4

−2

−6

Time Relative to Adoption

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding foreign outcome listed in each

panel, using inverse probability-weighted data. Specification 1 (in black) does not include additional controls.

Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and year-by-cohort-by-group fixed

effects, where groups include domestic and foreign sales quartiles and domestic and foreign intangible asset

quartiles, and where quartiles are computed using pre-adoption values for each cohort.

Figure A.2: Event Studies: Hybrid Tax Planning and Real Economic Activity

(IPW)

(B) Domestic Capital Assets

Relative Change in Logged Outcome

(A) Domestic Investment

0.50

0.25

0.00

−6

−4

−2

0.3

0.2

0.1

0.0

−0.1

−6

−4

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

0.25

0.00

−0.25

−2

(D) Domestic Wages

Relative Change in Logged Outcome

MNC & Industry, Sales, Asset x Yr x Cohort FEs

0.50

−4

MNC & Year x Cohort FEs

−6

−2

Time Relative to Adoption

0.4

0.2

0.0

−6

Time Relative to Adoption

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding outcome listed in each panel,

using inverse probability-weighted data. Specification 1 (in black) does not include additional controls. Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and year-by-cohort-by-group fixed effects,

where groups include domestic and foreign sales quartiles and domestic and foreign intangible asset quartiles, and

where quartiles are computed using pre-adoption values for each cohort.

Figure A.3: Event Studies: Profit Shifting Mechanisms and Foreign ETRs

(Sun Abraham)

(B) Foreign Intangible Assets

Relative Change in Logged Outcome

(A) Loans Between Related CFCs

0.5

0.0

−0.5

−6

−4

−2

1.2

0.8

0.4

0.0

−0.4

−6

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

(D) Foreign Effective Tax Rate

0.05

1.0

Relative Change in Outcome

Relative Change in Logged Outcome

−4

Time Relative to Adoption

0.5

0.0

0.00

−0.05

−0.5

−6

−4

−2

−6

−4

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

Notes: This figure provides estimates of µℓ from Equation 2 for the corresponding foreign outcome listed in each

panel, using the staggered estimator from Sun and Abraham (2021). Specification 1 (in black) does not include

additional controls. Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and year-bycohort-by-group fixed effects, where groups include domestic and foreign sales quartiles and domestic and foreign

intangible asset quartiles, and where quartiles are computed using pre-adoption values for each cohort.

Figure A.4: Event Studies: Hybrid Tax Planning and Real Economic Activity

(Sun Abraham)

(B) Domestic Capital Assets

Relative Change in Logged Outcome

(A) Domestic Investment

0.6

0.4

0.2

0.0

−6

−4

−2

0.2

0.1

0.0

−0.1

−6

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

0.25

0.00

−0.25

−2

(D) Domestic Wages

Relative Change in Logged Outcome

MNC & Year FEs

0.50

−4

−2

Time Relative to Adoption

−6

−4

Time Relative to Adoption

0.4

0.2

0.0

−6

−4

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

Notes: This figure provides estimates of µℓ from Equation 2 for the corresponding outcome listed in each panel,

using the staggered estimator from Sun and Abraham (2021). Specification 1 (in black) does not include additional

controls. Specification 2 (in orange) includes year-by-cohort-by-industry fixed effects and year-by-cohort-by-group

fixed effects, where groups include domestic and foreign sales quartiles and domestic and foreign intangible asset

quartiles, and where quartiles are computed using pre-adoption values for each cohort.

Figure A.5: Event Studies: Profit Shifting Mechanisms and Foreign ETRs

(TWFE)

(B) Foreign Intangible Assets

Relative Change in Logged Outcome

(A) Loans Between Related CFCs

0.5

0.0

−0.5

−6

−4

−2

1.2

0.8

0.4

0.0

−0.4

−6

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

(D) Foreign Effective Tax Rate

0.05

1.0

Relative Change in Outcome

Relative Change in Logged Outcome

−4

Time Relative to Adoption

0.5

0.0

0.00

−0.05

−0.5

−6

−4

−2

−6

−4

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

Notes: This figure provides estimates of µℓ from Equation 2 for the corresponding foreign outcome listed in

each panel. Specification 1 (in black) does not include additional controls. Specification 2 (in orange) includes

year-by-cohort-by-industry fixed effects and year-by-cohort-by-group fixed effects, where groups include domestic

and foreign sales quartiles and domestic and foreign intangible asset quartiles, and where quartiles are computed

using pre-adoption values for each cohort.

Figure A.6: Event Studies: Hybrid Tax Planning and Real Economic Activity

(TWFE)

(B) Domestic Capital Assets

Relative Change in Logged Outcome

(A) Domestic Investment

0.4

0.2

0.0

−0.2

−6

−4

−2

0.2

0.1

0.0

−0.1

−6

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

0.4

0.2

0.0

−0.2

−2

(D) Domestic Wages

Relative Change in Logged Outcome

MNC & Year FEs

0.6

−4

−2

Time Relative to Adoption

−6

−4

Time Relative to Adoption

0.4

0.2

0.0

−6

−4

−2

Time Relative to Adoption

MNC & Year FEs

MNC & Industry, Sales, Asset x Yr FEs

Notes: This figure provides estimates of µℓ from Equation 2 for the corresponding outcome listed in each panel.

foreign sales quartiles and domestic and foreign intangible asset quartiles, and where quartiles are computed

using pre-adoption values for each cohort.

Figure A.7: Event Studies: Compustat R&D (Robustness)

Relative Change in Outcome

(A) R&D Intensity

0.050

0.025

0.000

−6

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs (Unrestricted)

Relative Change in Outcome

(B) Log R&D

0.50

0.25

0.00

−0.25

−6

−4

−2

Time Relative to Adoption

MNC & Year x Cohort FEs

MNC & Industry, Sales, Asset x Yr x Cohort FEs

MNC & Industry, Sales, Asset, R&D Intensity x Yr x Cohort FEs

Notes: This figure provides estimates of µℓ from Equation 1 for the corresponding R&D outcome listed in

each panel, using 2-year pooled average data from Compustat and dropping odd years to match the IRS SOI

International Business Tax Sample. Specifications 1 and 2 (in black and orange, respectively) are identical to

the results shown in Figure 16. For R&D Intensity (Panel A), Specification 3 (in purple) does not restrict R&D

intensity to be less than 1. For Log R&D (Panel B), Specification 3 includes an additional R&D intensity quartile

fixed effect in its year-by-cohort-by-group fixed effects, where R&D intensity is not restricted to be less than 1.

Table A.2: Hybrid Tax Planning and R&D Data from Compustat (Robustness)

(1)

(2)

(3)

(4)

0.028**

(0.010)

125

388

0.060

0.056

0.034***

(0.010)

114

337

0.050

0.055

0.033**

(0.010)

125

387

0.059

0.056

0.034**

(0.011)

125

387

0.059

0.056

0.285***

(0.078)

252

Yes

0.362***

(0.077)

235

Yes

0.241***

(0.069)

266

Yes

0.268**

(0.084)

266

Yes

Stacked

Yes

Stacked

Yes

TWFE

Panel A

R&D Intensity (Unrestricted)

Num. Treated

Num. Control

Avg. R&D Intensity (All Firms)

Avg. R&D Intensity (Treated Firms)

R&D Intensity x Yr FEs

R&D Intensity x Yr x Cohort FEs

Panel B

Log R&D

Num. Treated

Num. Control

R&D Intensity x Yr FEs

R&D Intensity x Yr x Cohort FEs

MNC & Ind., Sales, Asset x Yr FEs

MNC & Ind., Sales, Asset x Yr x Cohort FEs

Inverse Prob. Weights

Model

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table provides estimates of the difference-in-differences model discussed in Section 5.2 for the

corresponding R&D outcome listed in each panel, where Columns (1) - (3) use Equation 1 and Columns (4) and

(5) use Equation 2. Estimations use 2-year pooled average data from Compustat, where odd years are dropped

to match the IRS SOI International Business Tax Sample. R&D Intensity (Panel A) is calculated as the ratio of

annual R&D to the MNC’s most recent sales value pre-adoption. All specifications include

year-by-cohort-by-industry and year-by-cohort-by-group fixed effects, where groups include domestic and foreign

sales quartiles and domestic and foreign intangible asset quartiles, and where quartiles are computed using

pre-adoption values. Log R&D (Panel B) also includes an additional R&D intensity quartile fixed effect in its

year-by-cohort-by-group fixed effects. Column (2) uses inverse probability-weighted data. Column (3) estimates

an alternative specification from Sun and Abraham (2021). Column (4) estimates a standard TWFE

specification.

Table A.3: HTP Uptake by Size

Quartile 2

Quartile 3

Quartile 4

Num. Obs.

R2 Pseudo

Year FEs

Dom. Sales

For. Sales

Dom. Assets

For. Assets

0.350+

(0.198)

0.350+

(0.199)

0.949***

(0.184)

0.483*

(0.207)

0.504*

(0.207)

1.147***

(0.193)

0.193

(0.193)

0.482**

(0.187)

0.623***

(0.184)

0.442*

(0.210)

0.803***

(0.198)

1.018***

(0.195)

8809

0.019

Yes

8809

0.024

Yes

8809

0.012

Yes

8809

0.020

Yes

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table displays coeﬀicients from logit regressions, as discussed in Section 4.3, that predict HTP

adoption based on 4 determinations of firm size: domestic sales, foreign sales, domestic assets, and foreign

assets. The model uses observations from adopting MNCs in the period prior to adoption and includes all

observations from never-adopters. *, **, and *** denote statistical significance at the 10, 5, and 1% level.

Table A.4: Selection into Hybrid Tax Planning

(1)

(2)

0.568***

(0.128)

8809

0.215

(0.170)

5030

0.499**

(0.178)

0.205

(0.191)

0.452*

(0.185)

8809

0.446*

(0.200)

0.071

(0.221)

0.128

(0.222)

5030

0.045

(1.047)

-0.212

(0.344)

7832

-1.653

(1.148)

-0.529

(0.364)

4528

2.411***

(0.596)

8280

1.930*

(0.873)

4796

1.460***

(0.183)

0.708***

(0.139)

0.023

(0.220)

8809

1.356***

(0.223)

0.440**

(0.168)

-0.105

(0.249)

5030

-0.861***

(0.172)

8809

-0.824***

(0.200)

5030

-0.007

(0.010)

8608

0.561

(1.714)

4920

Yes

Panel A

Claim Research Credit

Num.Obs.

Panel B

Age Quartile 2

Age Quartile 3

Age Quartile 4

Num.Obs.

Panel C

Avg. Statutory Foreign ETR

Share of Foreign Sales with Unobserved Statutory Rate

Num.Obs.

Panel D

Exposure to CTB

Num.Obs.

Panel E

Presence in Ireland

Presence in Netherlands

Presence in Luxembourg

Num.Obs.

Panel F

Domestic Loss

Num.Obs.

Panel G

Advertising to Sales Ratio

Num.Obs.

Year FEs

Industry, Sales, Asset Quartiles x Year FEs

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table displays coeﬀicients from a series of logit regressions, as discussed in Section 4.3, that predict

HTP adoption based on various MNC characteristics (Panels A-G). The model uses observations from adopting

MNCs in the period prior to adoption and includes all observations from never-adopters. Column (1) includes

year fixed effects; Column (2) includes year-by-industry, sales, and asset quartile fixed effects. *, **, and ***

denote statistical significance at the 10, 5, and 1% level.

Table A.5: Selection into Hybrid Tax Planning (Combined)

Claim Research Credit

Firm Age Quartile 2

Firm Age Quartile 3

Firm Age Quartile 4

Avg. Statutory Foreign ETR

Share of Foreign Sales with Unobserved Statutory Rate

Exposure to CTB

Advertising to Sales Ratio

Domestic Loss

Presence in Ireland

Presence in Netherlands

Presence in Luxembourg

Num.Obs.

Year FEs

Industry, Sales, Asset Quartiles x Yr FEs

(1)

(2)

0.105

(0.144)

0.417*

(0.188)

0.000

(0.209)

0.010

(0.206)

0.588

(0.993)

0.117

(0.320)

1.608*

(0.795)

−0.147

(0.500)

−0.793***

(0.184)

1.371***

(0.199)

0.578***

(0.150)

0.027

(0.229)

−0.045

(0.178)

0.399+

(0.212)

−0.015

(0.241)

−0.041

(0.238)

0.041

(1.146)

0.035

(0.352)

1.605

(1.007)

−1.919

(2.111)

−0.910***

(0.216)

1.331***

(0.244)

0.446*

(0.182)

−0.077

(0.265)

7624

Yes

4483

Yes

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table displays coeﬀicients from logit regressions, as discussed in Section 4.3, that predict HTP

adoption based on various MNC characteristics. This specification combines all characteristics tested in Table

A.4 into a single logit regression. Column (1) includes year fixed effects; Column (2) includes year-by-industry,

sales, and asset quartile fixed effects. *, **, and *** denote statistical significance at the 10, 5, and 1% level.

Table A.6: HTP Uptake by Auditor

(1)

Big 4

(2)

(3)

(4)

0.342

0.044

0.378

0.805

(1.019) (1.082) (0.293) (0.348)

0.370

0.511

(1.054) (1.119)

Medium

Num.Obs.

Year FEs

Industry, Sales, Asset Quartiles x Year FEs

4610

Yes

2491

Yes

4610

Yes

2491

Yes

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table displays coeﬀicients from logit regressions, as discussed in Section 4.3, that predict HTP

adoption based on the identity of an MNC’s auditors, determined using Compustat data. Columns (1) and (2)

include dummy variables for a “Big 4” auditor as well as a medium-sized auditor; Columns (3) and (4) include

only the Big 4 dummy. Odd columns include year fixed effects and even columns include year-by-industry, sales,

and asset quartile fixed effects. *, **, and *** denote statistical significance at the 10, 5, and 1% level.

B Measurement of Foreign Earnings and Taxes

In this appendix, we examine the potential for aggregation error in the measurement of foreign

earnings and taxes when using IRS data. The potential for measurement error is clearly summarized in Blouin and Robinson (2020, henceforth B&R). Appendix B.1 provides an overview

of several different commonly used data sources that are used to measure the activity of US

multinationals, as well as a discussion of the relative advantages and disadvantages of tax data

compared with other sources of accounting data for US MNCs. Appendix B.2 describes how we

construct a matched sample of MNCs which appear in both tax data and public filings compiled

by Compustat. We use this sample to examine differences in how foreign earnings and income

taxes are reported, to measure the extent of aggregation error that may be generated by naive

aggregation of foreign aﬀiliates’ tax filings, and to assess the performance of a simple correction

that aims to remove aggregation error from tax data. In Appendix B.3, we reproduce estimates

from earlier studies that attempt to measure tax semi-elasticities, i.e. how the allocation of foreign earnings by US multinationals correlates with average and statutory foreign tax rates. We

also provide corrected estimates of average foreign effective tax rates by jurisdiction.

B.1 Overview of Data Sources for Measuring Tax Outcomes of US

Multinationals

Below, we provide an overview of several different datasets that researchers have used to study

US MNC activity. Much of the literature discusses measurement issues, with a focus on data

compiled from surveys run by the Bureau of Economic Analysis (BEA). We review the main

issues encountered in this data. We then discuss the extent to which these measurement issues

may translate to the sources used in this study, Compustat and SOI.

B.1.1

Compustat

Compustat is a database that, among other things, compiles information about firms based on

their public filings. While Compustat is a useful source of information for a large set of firms, it

is not comprehensive—private firms are not covered by the database. Furthermore, information

about MNCs’ foreign operations is relatively limited in Compustat. It does, however, provide

some information about the aggregate consolidated foreign income and foreign tax expense for

a set of MNCs.

B.1.2

BEA Multinational Data

The BEA provides two different sources of data on US multinationals that are frequently used

in the literature. The first is typically referred to as “Direct Investment Income.” The second

source, which provides two different foreign income measures, is published under “Activities of

US Multinational Enterprises.”

Direct Investment Income

The BEA conducts quarterly surveys that collect information on investment positions and transactions with directly-owned foreign aﬀiliates of US MNCs. Research that examines foreign earnings by US MNCs have used the “Direct Investment Income” (DII) measure contained in this

data (e.g., Zucman, 2014, 2015). One advantage of DII is that it does not appear to suffer from

issues related to double-counting of equity income that are present in other data sources, as

discussed below. On the other hand, on a country-by-country basis, DII may not reflect the true

geographic distribution of where earnings are first realized. This is due to the fact that income

generated by indirectly-held foreign aﬀiliates will only be observed at the level of a directly-owned

aﬀiliate. In addition, DII does not provide measures of foreign corporate income tax. As discussed in B&R, although researchers may attempt to combine DII with other data sources that

provide information about tax payments, this can result in underestimation of average effective

rates for tax havens to the extent that taxes on equity income do not match the jurisdiction in

which they appear in DII data.

Activities of US Multinational Enterprises

The BEA also conducts annual surveys that collect detailed information on the activity of foreign

aﬀiliates directly or indirectly owned by US MNCs. These data are compiled into annual reports

titled “Activities of US Multinationals” (AMNE). Much of the international tax literature focuses

on statistics provided on the Income Statement (Tables starting with “D”). AMNE data provide

disaggregated information that separates income from indirectly and directly owned foreign aﬀil-

iates, thereby avoiding the misclassification error present in the DII data. As discussed in B&R,

however, simple aggregation of foreign net income may double count so-called “equity income”

that is included on the income statement of aﬀiliates that are parents of other aﬀiliates. B&R

provide a simple adjustment to prevent such double-counting by subtracting related dividend

income, which is also reported by the BEA.

Another advantage of AMNE data is that, unlike DII data, measures of corporate income tax

are also provided. This allows for more reliable estimates of taxes paid and average tax rates by

jurisdiction.

B.1.3

Administrative Tax Data

A full description of the SOI data is provided in Section 3. The discussion below focuses on the

potential for aggregation error in this data.

Like the BEA’s AMNE data, SOI international tax data include information about foreign

aﬀiliates whether or not they are directly or indirectly owned. One key difference from AMNE

data is that, as alluded to above, foreign aﬀiliates may be classified for tax purposes as opaque

entities (CFCs) or pass-through entities (FDEs). US parents can control their foreign aﬀiliates

directly or indirectly through another CFC or FDE; CFCs that are structured as parents of

FDEs report consolidated financial information on Form 5471. This method of reporting can

cause misclassification error similar to that observed in the BEA’s DII series.

To see how misclassification error can occur in the tax data, consider a US MNC with two

foreign aﬀiliates—aﬀiliate A is a holding company located in a tax haven, e.g. Bermuda, and

aﬀiliate B is an aﬀiliate that sells merchandise to customers, e.g. in Germany. If the US parent

classifies aﬀiliate B as a pass-through (FDE) and aﬀiliate A as an opaque entity (CFC), then

the information contained on aﬀiliate A’s Form 5471 will provide foreign income and tax data

that include its German operations. As noted above, because information covering pass-through

entities is available only for a handful of years, in general it is not possible to “undo” this

misclassification error. Unlike DII data, however, the SOI also reports information about foreign

tax payments.

The fact that the SOI contains misclassification error is not necessarily a “disadvantage” of

the tax data. This instead depends on the researcher’s goal. If researchers are attempting to

accurately measure effective tax rates imposed by countries on businesses that operate in their

jurisdiction, SOI data will generally be unreliable when compared with BEA data. If researchers

are instead interested in the effective tax rate faced by combined foreign structures, i.e. in the

consolidated operation of Bermudan-German aﬀiliates, then the tax data may be preferable as

the BEA does not provide information about whether foreign aﬀiliates are classified as passthroughs or opaque entities for tax purposes. Thus, while misclassification error may be present

in SOI data, this may not be problematic as long as researchers have the correct interpretation

of the data.

SOI data may also suffer from aggregation error in a manner that is similar to the BEA’s

AMNE dataset. This may occur if there is indirect ownership of one opaque entity by another.

In this case, foreign earnings that are distributed as dividends to the parent aﬀiliate may appear

as foreign earnings on each aﬀiliate’s Form 5471. However, the SOI also provides information

that can be used to correct for this type of error. Schedule M of Form 5471 reports transactions

between related foreign aﬀiliates, including dividends. B&R suggest removing these dividends

to accurately aggregate foreign income. Based on a comparison with BEA data, however, they

question whether this adjustment is suﬀicient to fully correct for aggregation error.

Aggregation error potentially poses a much more serious problem for researchers. If SOI data

overstate foreign earnings, estimates from research that examine foreign earnings in tax data may

not be reliable. Worse, to the extent this error is growing over time, which one might suspect if

MNCs’ ownership structures are also becoming more complex, standard economic approaches to

control for such error (e.g. unit fixed effects) might only exacerbate the problem (Bound, Brown,

Duncan and Rodgers, 1994; Bound and Krueger, 1991).

B.2 Assessing Aggregation Error in Foreign Earnings and Taxes

In this section, we use data from Compustat and SOI to measure the extent to which SOI data

may suffer from aggregation error, and examine whether the correction proposed by B&R appears

to suﬀiciently correct for this error. First, we describe how we create a matched sample of US

MNCs that appear both in SOI and Compustat. Compustat reports foreign income and foreign

tax figures on a consolidated basis for a set of MNCs.26 The consolidated nature of Compustat’s

The Compustat sample typically includes large, publicly-traded firms.

reporting implies that it should not be contaminated by aggregation error. This provides a

baseline comparison that can be used to measure aggregation error for a common set of firms,

avoiding the possibility that aggregate differences may simply be due to differences in sample

composition as in the analysis contained in B&R.27

B.2.1

Sample Construction

Figure B.1 shows attrition associated with the sample construction process. The top line in

the left panel shows the annual sample size for US MNCs contained in the SOI sample, which

ranges between 7,597 and 9,157. US MNCs are defined as US companies that file a corporate

tax return and that own at least one opaque foreign aﬀiliate (CFC) that files Form 5471. SOI

includes a firm identifier (EIN) that can be used to link the corresponding firm from Compustat.

The size of this linked sample, shown in the middle dashed line in the left panel of Figure B.1,

ranges between 1,845 and 2,334. The considerable attrition rate relative to the full SOI sample

is not surprising given that Compustat does not include information about privately-held firms

or smaller multinationals.

To ensure that the match is of high quality, we use information from Form 1120, Schedule M3, which must be filed by large US MNCs to reconcile book and tax financial information. Firms

must report book earnings on their M-3 that in theory should be the same as those disclosed

in their public filings. These earnings should be directly comparable to the Net Income figure

reported in Compustat; therefore, we drop firms from the matched sample that report M-3 Net

Income with a different sign from Compustat or firms that report Net Income that is not within

1% of the Compustat Net Income figure. The size of the final linked sample, which ranges

from 1,680 to 2,023, is shown on the bottom line of the left panel of Figure B.1. The figure

demonstrates that only a small share of firms are dropped when using Schedule M-3 to match

on net income, which is encouraging and indicative that the comparison is of high quality.

The right panel measures attrition by examining aggregate pretax income for each step of

the sample construction. A much smaller proportion of foreign earnings are dropped relative to

the attrition in sample size, confirming that Compustat does not provide coverage of relatively

This is not a criticism of that analysis. B&R only have access to highly aggregated data made public by

SOI, and therefore the concern about differences in sample composition is unavoidable.

small US MNCs.

Figure B.1: SOI MNCs vs. Matched Sample

Count

Pretax Income ($ billions)

600

7,500

400

5,000

2,500

200

2004

2007

2010

2013

SOI Sample

2016

2004

Matched to Compustat

2007

2010

2013

2016

Similar Book Income

Notes: This figure demonstrates the attrition associated with the sample construction process. The “SOI Sample”

contains all US MNCs from the SOI data sample. The “Matched to Compustat” sample contains SOI Sample

MNCs matched to Compustat via unique firm identifier (EIN). The final linked sample, “Similar Book Income,”

was created by dropping Compustat-matched MNCs whose M-3 Net Income, as reported on IRS Form 1120,

Schedule M-3, has a different sign than or is not within 1% of the firm’s Compustat Net Income. The left panel

displays sample sizes in terms of number of MNCs, while the right panel shows sample sizes in terms of pretax

income. These panels showcase Compustat’s lack of coverage for small and/or private MNCs, but underscore that

the match quality of MNCs that are able to be linked is high.

Figure B.2 provides an overview of attrition starting with the full Compustat sample. An

additional filter is applied to remove Compustat observations that report missing pretax foreign

income. As shown in the left panel, a much smaller portion of Compustat MNCs are dropped

in the match, which is unsurprising given that SOI is the more comprehensive database. In

2016, for example, the matched sample includes 68.8% of Compustat MNCs. In the right panel,

attrition is presented in terms of aggregate foreign earnings as reported in Compustat, with the

matched sample including 84% of the aggregate foreign earnings reported in Compustat in 2016.

Figure B.2: Compustat MNCs vs. Matched Sample

Count

Pretax Income ($ billions)

500

400

1,500

300

1,000

200

500

100

2004

2007

2010

2013

Compustat Sample

2016

2004

Matched to SOI

2007

2010

2013

2016

Similar Book Income

Notes: This figure demonstrates the attrition associated with the sample construction process, following the same

methodology as described in Figure B.1 but starting with the Compustat sample rather than the SOI sample.

Observations reporting missing pretax foreign income are removed from the initial Compustat sample. The left

panel shows sample sizes in terms of number of MNCs, while the right panel shows sample sizes in terms of pretax

income. These panels show that the majority of Compustat MNCs were able to be linked to an SOI MNC, and

that the match quality of linked MNCs is high.

B.2.2

Comparison of Aggregate Foreign Income and Tax

Having constructed a sample of MNCs for which both public filings and tax filings are observed,

we now examine aggregate foreign earnings and tax outcomes over time within the sample.

Figure B.3 shows three aggregate measures of foreign income tax. The solid line plots aggregate foreign tax from Compustat. The dotted lines provide two different estimates from tax

data: the series represented by short dashes are computed from Form 5471, Schedule C (Income Statement), and the series represented by the long dashes are computed from Form 5471,

Schedule E (Income, War Profits, and Excess Profits Taxes Paid or Accrued). We use Schedule

E as the preferred measure. The left panel contains aggregates computed from MNCs in all

industries. The right panel excludes MNCs classified as operating in Financial, Utilities, Mining,

Agriculture, or Oil & Gas.

Examining the full sample, it appears that Compustat tends to report much larger estimates

of foreign income tax than what is contained in tax filings. This difference is markedly lower

when excluding the industries selected above. One key reason for this discrepancy appears to

be related to how extractive industries operate foreign concessions. These projects are often

Figure B.3: Comparison of Aggregate Foreign Tax Measures

Foreign Income Taxes ($ billions)

All Industries

Finance and Extractive Industries Removed

125

100

2000

2010

Foreign Income Tax (Compustat)

2000

Foreign Income Tax (5471, Sch. E)

2010

Foreign Income Tax (5471, Sch. C)

Notes: This figure shows three measures of aggregate foreign income tax as respectively reported by Compustat;

IRS Form 5471, Schedule C; and IRS Form 5471, Schedule E. The left panel contains aggregates computed from

MNCs in all industries, while the right panel excludes MNCs operating in Financial, Utilities, Mining, Agriculture,

or Oil & Gas. The panels show that Compustat tends to report larger estimates of foreign income tax than is

contained in tax filings; however, this difference is lessened when excluding the industries mentioned above. We

use Schedule E as the preferred measure.

structured so that the foreign state receives a share of revenue or profits. Firms have discretion

over whether to report these profit-sharing arrangements as income tax in their books. It does

not appear that they have the same amount of discretion when they disclose information about

foreign taxes in their returns.

Even after removing this set of industries, Compustat tends to report larger estimates of

corporate income tax than what firms disclose in their tax returns. This suggests that aggregation

error is likely not a significant problem when computing a US MNC’s foreign tax bill from its

5471 filings.

Figure B.4 shows three aggregate measures of foreign pretax earnings for the matched sample.

As before, the solid lines plot aggregate foreign pretax income from Compustat and the dotted

lines provide two different estimates from tax data: the series represented by short dashes represents unadjusted E&P computed from Form 5471, Schedule H, and the series represented by

the long dashes is adjusted to remove dividends received from related aﬀiliates as reported on

Schedule M. For the SOI data, we add back foreign taxes as reported on Schedule E so that

both series represent pretax earnings. The left and right panels are inclusive and exclusive of the

Figure B.4: Comparison of Aggregate Foreign Income (Pretax)

Pretax Foreign Income ($ billions)

All Industries

Finance and Extractive Industries Removed

500

600

400

300

200

100

2000

2010

Unadjusted E&P (SOI)

2000

Adjusted E&P (SOI)

2010

Pretax Foreign Income (Compustat)

Notes: This figure shows three measures of aggregate foreign pretax income: aggregate foreign pretax income

as reported by Compustat; unadjusted E&P computed from IRS Form 5471, Schedule H; and Schedule H E&P,

adjusted to remove dividends received from related aﬀiliates as reported on Schedule M. Both SOI measures

include foreign taxes, as reported in Schedule E, added back to represent pretax earnings. As in Figure B.3, the

left panel contains aggregates computed from MNCs in all industries and the right panel excludes MNCs operating

in Financial, Utilities, Mining, Agriculture, or Oil & Gas. The panels indicate that the E&P adjustment proposed

by B&R is at least partially effective.

finance and extractive industries described above.

Notably, unadjusted E&P appear to diverge from the other measures over time. Compustat

and adjusted SOI foreign income measures are more closely aligned, although SOI seems to report

a slightly higher figure on average. This provides suggestive evidence that B&R’s suggested

correction for aggregation error is at least partially effective, removing the bulk of disagreement

between data sources. Figure B.5 provides a similar figure, but with aggregate post-tax earnings

(instead of pretax), demonstrating a similar pattern.

Posttax Foreign Income ($ billions)

Figure B.5: Comparison of Aggregate Foreign Income (Post-Tax)

All Industries

Finance and Extractive Industries Removed

400

300

200

100

2000

2010

Unadjusted E&P (SOI)

2000

Adjusted E&P (SOI)

2010

Posttax Foreign Income (Compustat)

Notes: This figure displays the same three aggregate foreign income measures as Figure B.4 on a post-tax, rather

than pretax, basis. It similarly underscores the effectiveness of B&R’s adjustment for aggregation error in the

SOI E&P data.

Finally, Figure B.6 presents a comparison of average annual foreign effective tax rates from

Compustat and SOI (one adjusted for aggregation error, one unadjusted). Compared to the

adjusted series, the average ETR computed from the unadjusted time series is downward-biased,

as would be expected given that it overestimates the denominator. There is still considerable

residual disagreement, however, when comparing adjusted ETRs from SOI with those computed

from Compustat data. Compustat estimates are much higher throughout the sample period.

This is a result of its higher tax estimates (Figure B.3) and lower foreign earnings estimates

(Figure B.4).

Figure B.6: Comparison of Aggregate Foreign Effective Tax Rates

All Industries

Finance and Extractive Industries Removed

Foreign Effective Tax Rates

50%

40%

30%

20%

10%

2000

2010

ETR (Compustat)

2000

Adjusted ETR (SOI)

2010

Unadjusted ETR (SOI)

Notes: This figure presents a comparison of average foreign effective tax rates from Compustat, SOI (adjusted

for aggregation error), and SOI (unadjusted). These tax rates are based on pretax income as reported in Figure

B.4 and foreign tax as reported in Figure B.3. The left panel contains aggregates computed from MNCs in all

industries and the right panel excludes MNCs operating in Financial, Utilities, Mining, Agriculture, or Oil & Gas.

The panels demonstrate that lack of adjustment to the SOI data creates downward bias in the tax rate estimates.

Alternately, they also show that comparatively higher taxes and lower foreign income in the Compustat data lead

to larger Compustat ETR estimates than adjusted or unadjusted SOI.

B.2.3

Measurement Error and MNC Corporate Complexity

In this section, we show that aggregation error is closely related to the size of MNCs’ foreign

networks. More complex foreign networks create more potential for double-counting, as dividends

may be reported on the books of a potentially large number of foreign aﬀiliates. Once foreign

earnings are adjusted to remove related dividends, however, there is no evidence that residual

measurement error is systematically related to the complexity of an MNC’s aﬀiliate structure.

Furthermore, we show that these adjustments result in residual measurement error that appears

to be roughly constant over time. These results are encouraging for researchers, as they suggest

that standard techniques to control for time-invariant heterogeneity and measurement error may

be valid.

The aggregate trends shown above suggest that aggregation error primarily distorts estimates

of foreign earnings, not foreign taxes. Given that the primary mechanism behind aggregation

error relates to the double-counting of foreign earnings that are distributed up through an MNC’s

foreign ownership network, we would expect aggregation error to increase with the size of an

Figure B.7: SOI/Compustat Foreign Earnings Differential

Difference in Log Foreign Earnings (SOI − Compustat)

60%

40%

20%

100

# CFCs

Adjusted Difference

Unadjusted Difference

Notes: This figure presents the results of two cubic spline regressions where the dependent variable is the difference

between log SOI foreign E&P and log Compustat foreign income. The orange line represents the estimates of

a regression using unadjusted SOI E&P, while the black line uses adjusted SOI E&P. The orange triangles and

black circles represent respective binscatter estimates produced for MNCs in 10 quantiles of foreign aﬀiliate

network size, as measured by the number of CFCs in the SOI data. The upward trend in the orange line and

binscatter estimates indicates that aggregation error increases with the size of an MNC’s foreign aﬀiliate network.

Alternately, the flat spline of the adjusted E&P regression indicates that residual differences between Compustat

estimates are unrelated to a firm’s foreign aﬀiliate network size.

MNC’s foreign aﬀiliate network. Figure B.7 confirms that this is the case. The orange line and

surrounding ribbon plot the estimates of a cubic spline regression. The dependent variable is

the difference between log unadjusted foreign E&P as reported in SOI and log foreign income as

reported in Compustat. The orange triangles are binscatter estimates produced for MNCs in 10

quantiles of foreign aﬀiliate network size, as measured by the number of opaque foreign entities

(CFCs) in the tax data. Note that there is a clear upward trend, indicating that aggregation

error is closely related to MNCs’ foreign network size.

The black line and circles plot the same statistics computed for a dependent variable that uses

adjusted SOI E&P. The spline is flat, suggesting that the residual differences between Compustat

1996−2000

(SOI − Compustat)

Difference in Log Foreign Earnings

Figure B.8: SOI/Compustat Foreign Earnings Differential by Period

2002−2008

2010−2016

60%

40%

20%

100

# CFCs

Adjusted Difference

Unadjusted Difference

Notes: This figure plots the same spline estimates as Figure B.7 above, split into three time periods. The increase

in the slope of the orange line over the three periods further confirms that aggregation error in the unadjusted SOI

data has been increasing over time. However, the roughly constant black line indicates that residual measurement

error has not been increasing.

estimates are unrelated to the complexity of a firm’s foreign network. If this adjustment were

imperfect, and missed a portion of double-counted profits, we would expect this slope to perhaps

be attenuated but not flat.

The black line does appear to be significantly positive, indicating that SOI foreign earnings

are on average about 8% larger than the figures reported in Compustat. One reason for this may

be that tax data include earnings by foreign aﬀiliates that are associated with sales to US-based

customers. These sales would be accounted for on the books of foreign aﬀiliates, but might not

be reported as foreign income on a firm’s consolidated public filings.

Figure B.8 plots spline estimates for the same sample, but split into three distinct periods.

This confirms that aggregation error has been increasing over time, as reflected by the increasing

slope of the orange splines. Adjusted E&P, however, remains flat and roughly constant at the

same level. This again suggests that residual measurement error has not been increasing over

time, and instead might be due to other reporting differences between tax and public filings.

Table B.1 provides a regression formulation of the graphical analysis shown above that estimates the following specification,

SOI

) − log(Πitf,Compustat ) = β log(NitCFC ) + µt + ui .

log(Πf,

(B.1)

Table B.1: Measurement Error and MNC Foreign Aﬀiliate Network Size

(Intercept)

Log # CFCs

Num.Obs.

FE: year

Remove Outliers:

Adjusted Earnings

(1)

(2)

0.095***

(0.027)

0.043***

(0.009)

0.082**

(0.027)

−0.001

(0.009)

9993

Yes

(3)

(4)

(5)

(6)

(7)

(8)

0.043***

(0.009)

−0.012

(0.009)

0.081***

(0.016)

0.047***

(0.005)

0.079***

(0.016)

0.005

(0.005)

0.050***

(0.006)

0.000

(0.005)

9993

Yes

9040

Yes

9040

Yes

9040

Yes

+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Notes: This table displays coeﬀicients from regressions of the form B.1. The dependent variable is the difference

between log SOI foreign E&P and log Compustat foreign income. Odd columns use unadjusted SOI E&P and

even columns use adjusted SOI E&P. Columns (3), (4), (7), and (8) add year fixed effects, and Columns (5)

through (8) remove MNCs operating in Financial, Utilities, Mining, Agriculture, or Oil & Gas. *, **, and ***

denote statistical significance at the 10, 5, and 1% level.

Again, the dependent variable is the difference between log foreign E&P as reported in SOI

and log foreign income as reported in Compustat. Odd columns use unadjusted E&P and even

columns introduce the double-counting adjustment. The independent variable is the log of the

number of CFCs in a firm-year. For the unadjusted difference, the coeﬀicients on this variable

are positive, with large t-statistics. The adjusted coeﬀicients, however, are close to zero with

narrow confidence intervals. Columns (3), (4), (7), and (8) include the year fixed effects µt , and

Columns (5) through (8) remove the industries excluded in the aggregate plots shown in the

previous section.

B.3 Measurement Error and Tax Elasticities of Foreign Income

In the previous section, we show that introducing a simple correction to remove related dividends appears to remove systematic aggregation error in the measurement of foreign earnings

in tax data. In this section, we recompute tax semi-elasticities of foreign earnings following

the methodology in Dowd, Landefeld a

This text is long and has been trimmed here. Open the source document for the complete record.

This is a copy of a public record, reproduced as it was published. It is not legal advice, and it may not be the version a court would rely on. Check the official source before you cite it.