Arbeitskreis Quantitative Steuerlehre

Quantitative Research in Taxation – Discussion Papers

Jens Müller, Arndt Weinrich

Tax Knowledge Diffusion via Strategic Alliances

arqus Discussion Paper No. 253

February 2020 revised May 2020 & August 2020

www.arqus.info

ISSN 1861-8944

Tax Knowledge Diffusion via Strategic Alliances

08/18/2020

Jens Müller Arndt Weinrich

Paderborn University Paderborn University

jens.mueller@upb.de arndt.weinrich@upb.de*

Acknowledgments: We thank Harald Amberger, Alissa Brühne (discussant), Paul Demeré, Alex Edwards,

Beatriz Garcia-Osma, Jochen Hundsdoerfer, Martin Jacob, Alastair Lawrence, Harun Rashid (discussant),

Leslie Robinson, Christina Ruiz (discussant), Harm Schütt and Jake Thornock for their helpful comments.

We also thank the participants at the 2019 annual meeting of the foundation Stiftung

Prof. Dr. oec. Westerfelhaus, the 2019 arqus Conference, the 5th Berlin-Vallendar Conference on Tax

Research, the 6th Annual MaTax Conference, the 2020 Hawaii Accounting Research Conference, the 2020

ATA Midyear Meeting in Fort Worth and the 82nd VHB Annual Business Researcher Conference 2020.

We further acknowledge the helpful discussions with our colleagues at the Department Taxation,

Accounting and Finance at Paderborn University. This work was supported by the

Stiftung Prof. Dr. oec. Westerfelhaus (Project-ID P02-1) and the Deutsche Forschungsgemeinschaft (DFG,

German Research Foundation, Project-ID 403041268, TRR 266 Accounting for Transparency).

*Send correspondence to Arndt Weinrich, Paderborn University, Faculty of Business Administration and

Economics, Warburger Str. 100, DE 33098 Paderborn.

1

Tax Knowledge Diffusion via Strategic Alliances

Abstract

This study examines strategic alliances as channels for tax knowledge diffusion between firms. Although

strategic alliances are primarily expected to foster their main business purposes, we investigate whether tax

knowledge diffuses as a second-order effect of peer-to-peer cooperation. To tease out the diffusion of tax

knowledge, we analyze changes in the tax planning behavior of high-tax firms in strategic alliances with

low-tax firms in comparison to high-tax firms in strategic alliances with other high-tax firms. We gain

insights into the business purpose of the strategic alliances by applying textual analysis of the deal

descriptions. Our results suggest an economically meaningful decrease in cash effective tax rates of high-

tax firms in strategic alliances with low-tax firms relative to high-tax firms in high-tax strategic alliances.

We find that the adjustment occurs on average not before the second year after a strategic alliance’s

initiation. We conjecture that strategic alliances are not intended to establish tax planning investments. We

triangulate our findings with regard to effects on textual sentiment of annual reports and tax haven

operations. Finally, we show that partner characteristics serve as a substitute rather than as a complement

for strategic alliances to low-tax firms. Overall, our results provide robust evidence for tax knowledge

diffusion via strategic alliances.

Keywords: Corporate Tax Planning/Avoidance, Knowledge Diffusion, Network, Strategic Alliance

JEL Qualification: C31, G34, H26

Data Availability: Public and/or subscription-based sources identified in the paper

Online Supplement: https://github.com/taxknowledge/diffusionviastrategicalliances

Declarations of Interest: None

2

1. Introduction

Do strategic alliances serve as channels for tax knowledge diffusion? In the form of contractual-based

cooperation between firms, strategic alliances are a relevant choice for optimal corporate institutionalization

(PwC, 2018). Strategic alliances are expected to foster their main business purposes and to facilitate

respective knowledge diffusion between the partners as a first-order effect. For instance, Li et al. (2019)

identify significant increases in firms’ innovative capacity when investing in R&D strategic alliances.

Because strategic alliances are not established as tax planning investments, tax knowledge is generally not

related to the activities within a strategic alliance. However, accounting research documents the impact of

cross-firm connections on the tax planning behavior of linked firms. Foci are on intentional transfers of tax

knowledge (e.g., see Cen et al. (2020)) and on evidence for the effect of intermediaries (e.g., see Brown &

Drake (2014)) in close relationships. With our study, we ease the assumptions of existing intermediaries

and intentional transfers of tax knowledge and shed light on whether tax knowledge diffusion occurs as a

(potentially unintended) second-order effect of strategic alliances.

Conceptually, diffusion of tax knowledge comprises gaining access to and being willing and capable of

employing relevant information and know-how. An exemplary mechanism could be a shift in a firm’s

management preferences and confidence toward implementing a specific tax planning activity when

observing unchallenged or successfully defended strategies of other partners in a strategic alliance. For our

inferences, we utilize a key feature of strategic alliances which is that they are, in contrast to equity joint

ventures, not subject to corporate taxation. This is useful for us since we are interested in whether strategic

alliances serve as channels for tax knowledge diffusion between firms and not in whether they provide

vehicles in tax planning activities themselves. Observing tax consequences at the level of the investing

firms allows us to identify tax knowledge diffusion via strategic alliances. This is the key innovation of our

study.

We empirically exploit information on strategic alliances, which we also refer to as “networks”, that were

established between publicly traded US firms from 1994 to 2016. Given that accounting data are available

3

for a network’s partners, we reshape the data from the alliance to the participant level (network-firm

observations). We measure tax knowledge by observing the outcome of a firm’s nonconforming tax

planning behavior (𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3) and classify the partners and the network as low-tax and high-tax. To tease

out tax knowledge diffusion after network initiation, we analyze changes in the tax planning behavior of

high-tax firms in strategic alliances with low-tax firms in comparison to high-tax firms in strategic alliances

with high-tax firms. We gain insights into the business purpose of a network by applying textual analysis

of the networks’ deal descriptions.

We find a substantial decrease in cash effective tax rates of high-tax firms in strategic alliances with low-

tax firms relative to high-tax firms in high-tax strategic alliances. Our results are also economically

meaningful because our analyses suggest reasonable cash effective tax rate levels for high-tax firms in low-

tax networks of 25.57%. Furthermore, we find that the adjustment occurs on average not before the second

year after a strategic alliance’s initiation. Additionally, we test for the effects on the reporting of operations:

textual sentiment of annual reports and tax haven usage. We find a negative response of textual sentiment

of 10-K filings to low-tax networks in comparison to high-tax networks for high-tax firms. This finding is

consistent with prior research indicating increased tax planning behavior when observing decreasing textual

sentiment of 10-K filings (Law & Mills, 2015). Further, we show that the identified effects from our

analyses seem not to not stem from changes in tax haven operations. Finally, we investigate whether partner

characteristics, such as geographical proximity, shared industry affiliation and an identical audit firm,

intensify or mitigate the identified effects.

Our study refers to the emerging accounting literature that identifies cross-firm connections to determine

increases in the tax planning behavior of linked firms. By focusing on close relationships via intermediaries

and intentional transfers of tax knowledge, recent research has analyzed board ties (Brown, 2011; Brown

& Drake, 2014), banks (Gallemore et al., 2019), human capital turnover (Barrios & Gallemore, 2019), and

auditors (Frey, 2018; Lim et al., 2018). Strategic alliances are, however, established on a peer-to-peer basis

without an intermediary. In recent work on peer-to-peer contracting, Cen et al. (2017, 2020) investigate

4

transfers of tax knowledge along the supply chain and document that both customers and suppliers increase

their tax planning activities once their relationship is considered dependent. They suggest that customers

and suppliers share tax planning benefits through lower product prices. Thus, the identification of

intentional tax knowledge transfer aligns the “supply chain channel” through tight bonds with research on

intermediaries. However, “dependency” is conceptually not a characteristic of strategic alliances, though

partners show mutual commitment not typically found in arm’s-length market transactions (Lindsey, 2008;

Chan et al., 1997). Our study analyzes tax knowledge diffusion as a second-order and potentially unintended

effect of strategic alliances.

We also contribute to the management literature which documents that firms often benefit from what they

learn in strategic alliances (Porrini, 2004). Empirical inferences are usually based on the contracting parties’

stock price performance (Boone & Ivanov, 2012; Mohanram & Nanda, 1996; Chen et al., 2015; Anand &

Khanna, 2000), return on equity, (cash flow) return on assets (Chan et al., 1997; Porrini, 2004),

postreorganization performance (Cai & Sevilir, 2012; Higgins & Rodriguez, 2006; Ishii & Xuan, 2014;

Porrini, 2004) and patent citations (Gomes-Casseres et al., 2006; Li et al., 2019). Nevertheless, not all

corporate practices diffuse in the same way (Cai et al., 2014). Since tax knowledge characteristics are

considered to lie somewhere between being substantially complex (Hoppe et al., 2019) and serving as mass-

market ideas (Lisowsky, 2010), ambiguity arises when analyzing the diffusion of tax knowledge. To the

best of our knowledge, we are the first to measure knowledge diffusion via a contractual cooperative

organizational form (i.e., strategic alliances) based on a firm’s tax planning behavior. Our study combines

accounting and management research to detect behavioral aspects of the tax planning process.

2. Conceptual Framework & Prior Literature

2.1. Cross-Firm Connections and Tax Knowledge Transfers

The extensive research that considers within-firm determinants of tax planning by firms underlines the

perceived importance of corporate taxes in economic theory, politics and society (for comprehensive

reviews, see Hanlon & Heitzman (2010) and Wilde & Wilson (2018)). Given the substantial economic

5

impact of tax planning activities (Clausing, 2016; Tørsløv et al., 2018; Blouin & Robinson, 2020), tax

advisors are intuitively linked to observations of tax planning. However, by investigating “tax planning

ecosystem[s]” (Dyreng & Maydew, 2017) between firms, recent research suggests the existence of

additional channels. The findings indicate that cross-firm connections impact tax planning behavior of

linked firms. Brown (2011) pioneered this strand of literature by examining the spread of a specific tax

planning tool, the corporate-owned life insurance shelter. Although she finds that board interlocks increase

the probability that a firm adopts the shelter from a prior user, she does not find significant shelter adoption

via shared audit firms. This latter result, although it is theoretically convincing due to independence

regulations on audit services, takes its place alongside a range of mixed inferences concerning auditors’

impact on firms’ tax planning behavior. For instance, Klassen et al. (2016) show that less tax aggressiveness

in the past is associated with the auditor preparing a firm’s tax return. However, there is also literature on

auditors who transgress their limited scope of function (Aobdia, 2015; Cai et al., 2016; Dhaliwal et al.,

2016; McGuire et al., 2012). By calibrating from the audit firm level to the individual audit engagement

partner, Frey (2018) suggests that the engagement of a German tax certified auditor, who signals high

competency in taxes, is associated with higher effective tax rates at client firms. In contrast, Lim et al.

(2018) find that Chinese firms with stronger connections to low-tax firms through individual audit partners

show lower effective tax rates. Consistent with the mixed evidence from prior literature, Nesbitt et al. (2020)

suggest that there are limits to the relation between auditor-provided tax services and clients’ tax

aggressiveness. Further disentangling the role of intermediaries, Barrios & Gallemore (2019) document that

firms exhibit increasing tax planning when they hire tax staff from sophisticated tax planners. This finding

is consistent with the inferences from analyzing board ties to low-tax firms (Brown & Drake, 2014).

Gallemore et al. (2019) show that firms experience meaningful tax reductions when they start a relationship

with a bank whose existing clients engage in tax planning.

[Figure 1]

6

These studies are conceptually aligned by the presence of intermediaries who implement tax planning

expertise in their set of contracts and intentionally transfer the tax knowledge gained to other parties with

whom they are contracting. Consistently, intermediaries are found to play a key role in the acquisition and

dissemination of information in many research fields (e.g., see Di Maggio et al. (2019)). Figure 1 aligns

institutional settings with identified channels. The research focuses on the role of intermediaries, whereas

Cen et al. (2017, 2020) analyze peer-to-peer contracting between firms. Specifically, they focus on transfers

of tax knowledge via the supply chain and document that both customers and suppliers increase their tax

planning activities once their relationship is considered dependent. Cen et al. (2020) suggest that customers

and suppliers share tax planning benefits through lower product prices. Although evidence for the intended

sharing of tax benefits is scarce (for instance, see Erickson (1998) and Erickson & Wang (1999)), an

intentional transfer of tax knowledge aligns the “supply chain channel” with research on intermediaries. In

contrast, we ease the studies’ assumptions of existing intermediaries and intentional transfers of tax

knowledge and focus on tax knowledge diffusion between firms as a (potentially unintended) second-order

effect when peer-to-peer contracting.

2.2. Strategic Alliances and Tax Knowledge Diffusion

2.2.1. Tax Knowledge Characteristics

Explicit knowledge can easily be codified and is systematically transferable, whereas tacit knowledge is

difficult to formulate and communicate because it “is deeply rooted in action, commitment, and

involvement in a specific context” (Nonaka, 1994, p. 16). Consequently, when (tax) knowledge qualifies

as more explicit, it should be more easily transferable (Meier, 2011). At first glance, one might refer to tax

knowledge as tacit given the substantial complexity of corporate taxes (Hoppe et al., 2019), the increased

uncertainty (Dyreng et al., 2019; Guenther et al., 2017) and the costs (Hundsdoerfer & Jacob, 2019) of tax

planning. Tax knowledge may also comprise know-how that is more tacit than pure information (Kale et

al., 2000). However, anecdotal evidence suggests that certain tax shelter strategies can be applied to

numerous firms and are not limited to a particular industry. Some findings suggest tax knowledge to be of

an explicit nature, given the inferences on corporate-owned life insurance shelters (Brown, 2011) or lease-

7

in, lease-out transactions (Wilson, 2009). Lisowsky (2010) further argues that a significant portion of tax

shelters serve as mass-market tax-saving ideas. However, not all firms make use of the tax planning ideas

that are heavily pitched by tax advisors.

2.2.2. Tax Knowledge Diffusion

Knowledge diffusion requires communication through channels over time among members of a social

system (Rogers, 2003). Thus, diffusion conceptually comprises gaining access to and being willing and

capable of employing relevant tax knowledge. Managers, who are seconded to a strategic alliance, are not

only involved in the alliance’s business as such but also monitor the network and the partner. Budget

meetings, for instance, could establish opportunities to gain insights into the partnering firms’ tax positions.

Observing effectively implemented tax planning strategies by partners in a network may not only reveal

unknown tax knowledge but could increase a firm’s management willingness to also implement the

respective tax planning activity. Moreover, it may provide a better assessment of costs and benefits of

certain tax planning strategies. Mulligan & Oats (2016, p. 70) note that “sharing information, particularly

about tax plans and technical advice about dealing with ambiguities in tax laws serves to provide legitimacy

to preferred tax positions, yielding a form of power […] when taking tax positions in dealing with Revenue

Authorities.” This is also consistent with knowledge diffusion being a gradual process of dissemination

(Szulanski, 1996) and with increasing probability of uniformity of actions in networks over time (Gale &

Kariv, 2003). Thus, tax knowledge diffusion could overcome the fear of costs from engaging in tax planning

(Gallemore et al., 2014; Hanlon & Slemrod, 2009; Graham et al., 2014; Austin & Wilson, 2017).

2.2.3. Determinants of Tax Knowledge Diffusion in Strategic Alliances

Firms are often found to benefit from what they learn in strategic alliances in other contexts (Porrini, 2004).

Empirical inferences are usually based on the contracting parties’ stock price performance (Boone &

Ivanov, 2012; Mohanram & Nanda, 1996; Chen et al., 2015; Anand & Khanna, 2000), return on equity,

(cash flow) return on assets (Chan et al., 1997; Porrini, 2004), postreorganization performance (Cai &

Sevilir, 2012; Higgins & Rodriguez, 2006; Ishii & Xuan, 2014; Porrini, 2004) and patent citations (Gomes-

8

Casseres et al., 2006; Li et al., 2019). In accomplishing a common goal, the partners in a strategic alliance

are engaged in joint problem solving via a social system (Rogers, 2003). Consistently, cooperation is found

to mitigate cultural differences (Kogut & Singh, 1988). Additionally, spillovers are more likely to occur in

cases of a high frequency of interactions between firms (Isaksson et al., 2016) and with increasing partner

trustworthiness (Jiang et al., 2016).

However, not all corporate practices diffuse in the same way (Cai et al., 2014). Major barriers are

knowledge-related factors, such as limits to a recipient’s absorptive capacity (Szulanski, 1996; Dyer &

Hatch, 2006). Furthermore, earning private benefits is valuable for a firm outside the scope of an alliance

(Khanna et al., 1998). Consequently, joint operations could induce complexity in the contracting parties’

organizations that reduces the ability to fine-tune the tax sheltering of their affiliates (Desai et al., 2004).

Corporate culture and governance could further impact a firm’s behavior with regard to implementing tax

planning strategies (Klassen et al., 2017; Armstrong et al., 2015). Additionally, both cooperation (Chen et

al., 2015) and tax planning (Dyreng et al., 2019) are found to increase a firm’s uncertainty. Even prudent

managers could expect the marginal disutility of uncertainty to exceed the benefits of received tax

knowledge. Thus, it remains an empirical question whether strategic alliances actually serve as channels

for tax knowledge diffusion between firms.

3. Data

[Figure 2]

3.1. Sample Selection

We exploit data on strategic alliances from the Securities Data Company (SDC) Platinum database on

strategic alliances over the 1994-2016 period. SDC is widely used in relevant research on corporate

cooperation (Anand & Khanna, 2000; Boone & Ivanov, 2012; Cai & Sevilir, 2012; Chen et al., 2015; Ishii

& Xuan, 2014; Li et al., 2019) and tracks a very wide range of agreement types (Schilling, 2009). SDC

distinguishes equity joint ventures from contractual-based strategic alliances on a first level, and flags on a

second level whether the cooperation compiles a research and development agreement, a sales and

9

marketing agreement, a manufacturing agreement, a supply agreement, and/or a licensing and distribution

pact. Our study focuses on strategic alliances and does not consider equity joint ventures.

SDC issues data at the strategic alliance level. We reshape data from the alliance to the partner level because

strategic alliances are not subject to corporate taxation but the (publicly traded) contracting parties are. For

instance, a strategic alliance between two partners translates to one network-firm observation for each of

the two firms. Compustat provides firm-year-level accounting information and we merge SDC and

Compustat data by using a firm’s six-digit CUSIP number (at the level of the ultimate parent of the

participant) as an identifier. Although SDC provides reliable network observations from the beginning of

1990 onwards, we start our sample in 1994 due to changes in accounting for income taxes and the domestic

statutory corporate income tax rate (Cen et al., 2017). We end our sample in 2016 to exclude any influences

from the 2017 US tax reform. Furthermore, we respectively consider strategic alliances between publicly

traded firms incorporated and headquartered in the US and in which all contracting parties are identified in

Compustat data (Figure 2).

3.2. Identification Strategy

3.2.1. Measuring Tax Knowledge

We measure tax knowledge by observing a firm’s nonconforming tax planning behavior. The lingua franca

in determining the degree to which a firm succeeds in this attempt is the effective tax rate, which puts tax

expenses and pre-tax book income into perspective. We base our inferences on the cash effective tax rate

(𝑐𝑎𝑠ℎ 𝐸𝑇𝑅) because 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 also captures tax deferral strategies (Hanlon & Heitzman, 2010; Edwards

et al., 2016). Furthermore, we apply a multiperiod (3-year) form of 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 (Brown & Drake, 2014;

Barrios & Gallemore, 2019; Gallemore et al., 2019) because we expect the likelihood of tax knowledge

diffusion to increase over time.

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=1 =∑ (3

𝑡=1 𝑡𝑥𝑝𝑑𝑖,𝑡)

∑ (3𝑡=1 𝑝𝑖𝑖,𝑡 − 𝑠𝑝𝑖𝑖,𝑡)

(1)

10

The terms 𝑡𝑥𝑝𝑑, 𝑝𝑖 and 𝑠𝑝𝑖 correspond to their Compustat data item equivalents of cash taxes paid, pre-tax

income and special items. Missing 𝑠𝑝𝑖 are reset to 0, while any 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 with a negative denominator is

reset to missing. Nonmissing 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 are winsorized at 0 and 1. By nature of this approach, 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3

would always be missing for the final and penultimate firm-year of a firm in our panel. For these firm-

years, we substitute 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 with 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅.1 Applying a forward-looking 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 has the

advantage that potential tax knowledge diffusion via strategic alliances can be directly linked to the year of

network initiation.

3.2.2. Low-Tax and High-Tax

For tax knowledge diffusion to occur, at least one network participant must possess sophisticated tax

knowledge. Therefore, we classify the strategic alliances’ partners into low-tax and high-tax firms.

Applying a forward-looking 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 for our analysis has the advantage of aligning any influence on the

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 directly with the year of network initiation. However, identifying low-tax firms based on a

forward-looking measurement would entail the disadvantage of concluding the type of input based on the

output. For the identification of sophisticated tax planners, we therefore consider the 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3,

which is constructed over a three-year preceding period:

𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=0 =∑ (0

𝑡=−2 𝑡𝑥𝑝𝑑𝑖,𝑡)

∑ (0𝑡=−2 𝑝𝑖𝑖,𝑡 − 𝑠𝑝𝑖𝑖,𝑡)

(2)

For every 𝑡 = 1 in which a new network is initiated, we consider the partners’ initial 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3, which

spans from 𝑡 = −2 to 𝑡 = 0.2 We require to observe 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 and 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 of all partners for a

network to be considered in our analysis. Figure 2 provides additional information regarding how we

identify low-tax and high-tax observations. We classify firms based on their

1 Example: given our sample period, 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 would always be missing for fiscal-year 2016. In this case, we

construct the numerator and denominator over one year, respectively. In robustness checks we exclude firm-edge-

years from our sample to ensure that variation in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 does not stem from this substitution (Table 6 Panel A). 2 We also refer to 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 and 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 as 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] and 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 [𝑡−2; 𝑡0] to highlight the

respective timing around a network initiation.

11

𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦 𝑦𝑒𝑎𝑟 𝑚𝑒𝑎𝑛 𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] and allocate firm-year observations into four

bins according to the quartiles of the distribution. Industry adjustment (Brown & Drake, 2014) and a

multiperiod measure (Dyreng et al., 2008; Dyreng et al., 2017) help us to validate identification of

sophisticated tax planners.3 In 𝑡 = 1 (network initiation), a partner is treated as low-tax firm in a network

when its adjusted 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] is in the first bin (i.e., lowest quartile). Firms that do not qualify

as low-tax firms are classified as high-tax firms. Consequently, a strategic alliance may be composed of

low-tax firms only, high-tax firms only, or a combination of high-tax and low-tax firms. In our analyses,

we focus on high-tax firms and thereby distinguish between high-tax firms that invest in strategic alliances

with low-tax firms (ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1, 𝑙𝑜𝑤‑𝑡𝑎𝑥 𝑛𝑒𝑡𝑤𝑜𝑟𝑘) and high-tax firms that invest in strategic

alliances with other high-tax firms (ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0, ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑛𝑒𝑡𝑤𝑜𝑟𝑘). We do not expect low-tax firms

to be exposed to tax knowledge diffusion because they already establish the group of sophisticated tax

planners (transmitter). Consequently, tax knowledge diffusion should respectively occur for the group of

high-tax firms (receiver). While there is a transmitter in a network for high-tax firms when ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 =

1, there is no transmitter for the group of high-tax firms when ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0. Thus, high-tax firms are in

a very similar situation except for potentially experiencing tax knowledge diffusion.4 Consequently, high-

tax firms in low-tax networks and high-tax firms in high-tax networks establish treatment and control group

for our analyses. Our identification strategy leads us to 201 observations of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1 and 627

observations of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0 (Figure 2).

3 Figure OS 1 in the Online Supplement shows details on industry-year-mean-(adjusted) effective tax rates throughout

the sample period. 4 This approach may be applied in various fields of research. For instance, Tan & Netessine (2019) use the catching

title “When You Work with a Superman, Will You Also Fly?”.

12

3.3. Information on Networks and Firms

[Figure 3]

Figure 3 maps networks and the investing firms from our sample.5 Each vertex (square) displays a firm in

its classification as low-tax (black vertices) or high-tax (gray vertices). A link between two gray vertices

translates into an observation of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0 for both high-tax firms. Correspondingly, a link between

a gray vertex and a black vertex translates into an observation of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1 for the high-tax firm.

Furthermore, it can be observed that there are firms with only one network observation in our sample

(n = 324). In contrast, there are also firms with multiple investments in strategic alliances during our sample

period (n = 178). Interestingly, these prominent partners largely consist of high-tax firms. This indicates

that firms do not strategically choose low-tax firms as partners in expectation of tax knowledge diffusion.

[Table 1]

Table 1, Panel A contains descriptive statistics with regard to firm-level accounting information

(𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠). Following Dyreng et al. (2010), we calculate 𝐸𝐵𝐼𝑇𝐷𝐴3, 𝑅𝑛𝐷𝐸𝑥𝑝3, 𝐴𝑑𝐸𝑥𝑝3, 𝑆𝐺𝐴3,

𝐶𝑎𝑝𝐸𝑥3, 𝐶ℎ𝑎𝑛𝑔𝑒𝑆𝑎𝑙𝑒3, 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒3, 𝐶𝑎𝑠ℎ3, 𝑀𝑁𝐸3, 𝑁𝑂𝐿3, 𝐼𝑛𝑡𝑎𝑛𝑔𝑖𝑏𝑙𝑒𝑠3, 𝑃𝑃𝐸3, and 𝑆𝑖𝑧𝑒3 as

𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠.6 Consistently with 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3, the 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 are constructed over rolling three-year

periods. From Compustat data, we can infer whether network partners share an audit firm and/or industry

affiliation in the year of network initiation (𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 & 𝑆𝑎𝑚𝑒𝐼𝑛𝑑) and whether their headquarters

are located in the same region as defined by the Bureau of Economic Analysis (𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛).7 To

increase the accuracy of our measures, we manually collect the geographical distance (as the crow flies)

between the zip codes of the network partners’ headquarters (𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦) to control for the potential impact

of geographical proximity in tax knowledge diffusion (Table 1, Panel B). We normalize 𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦

5 Networks of two partners may be displayed, respectively. However, the vast majority of networks in our sample

combine two firms. 6 All variables are defined in detail in the Appendix. 7 The respective BEA regions are Far West, Great Lakes, Mideast, New England, Plains, Rocky Mountains, Southeast

and Southwest.

13

between 1 for the closest and 0 for the farthest distance, which allows us to interpret the sign of a coefficient

in agreement with the sign of the indicator 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 (as, for instance, in Brown (2011)).

Controlling for these 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 ensures identifying the incremental impact of a strategic alliance

on a firm’s tax planning behavior (e.g., see Figure 1). We also test whether partner characteristics intensify

or mitigate the effects (Section 5.2).

[Figure 4]

Business activities in a strategic alliance are generally not limited and could thus exert an influence on the

options of tax planning that are available for a network participant. SDC provides information on a

network’s activities (e.g., see Section 3.1) and a deal description with every strategic alliance. To increase

the accuracy of our measures, we apply textual analysis on these deal descriptions to derive the main

business purposes of the strategic alliances in our sample (𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠). The word cloud depicted

in Figure 4 shows the 50 most common words used in the deal descriptions of our sample. Based on this

illustration and the activities described therein, we systematically search through the deal descriptions and

identify 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑊ℎ𝑜𝑙𝑒𝑠𝑎𝑙𝑒, 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑅𝑛𝐷, 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝐿𝑖𝑐𝑒𝑛𝑠𝑖𝑛𝑔, 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑆𝑒𝑟𝑣𝑖𝑐𝑒,

𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑀𝑎𝑟𝑘𝑒𝑡𝑖𝑛𝑔, 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑆𝑢𝑝𝑝𝑙𝑦, and 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑀𝑎𝑛𝑢𝑓𝑎𝑐𝑡𝑢𝑟𝑖𝑛𝑔 as major network activities.

Panel B of Table 1 shows the distribution of the respective indicator variables among ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and

𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ observations. Panel C of Table 1 presents information regarding the industry affiliation of

networks and firms. Industry affiliation is determined on a two-digit SIC-code basis. The majority of

networks and investing firms operate in business services and manufacturing.

3.4. Regression Analysis

If engaging in low-tax strategic alliances were associated with tax knowledge diffusion, one would expect

to identify increasing levels of tax planning at high-tax firms. Nevertheless, if the influence of a network

on a firm’s tax planning behavior as such were omitted, no inferences could be drawn. Therefore, our main

variable of interest ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 is constructed as an indicator variable to distinguish between high-tax firms

that enter into high-tax networks (ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0) and high-tax firms that engage in low-tax networks

14

(ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1). Consequently, ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 isolates the incremental effect a low-tax network exerts on

the high-tax firm’s 𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒:

𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒𝑖,𝑡=1

= 𝛽0 + 𝜷𝟏𝒉𝒊𝒈𝒉𝒕𝒐𝒍𝒐𝒘𝒊,𝒕=𝟏 + ∑ 𝛽𝑛𝑛

𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡=1𝑛

+ ∑ 𝛽𝑙𝑙

𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡=1𝑙 + ∑ 𝛽𝑘

𝑘𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡=1

𝑘

+ 𝛿𝑖𝑛𝑑 + 𝜏𝑡 + 휀𝑖,𝑡 .

(3)

By using indicator notation (ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤), the coefficient describes the effect of moving from one to another

condition. As we use a three-year rolling specification of a firm’s cash effective tax rate

(𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0]) for identification purposes, we primarily measure 𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒 by

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3]. Consistently, Kim et al. (2019) suggest that firms are generally able to adjust their tax

planning behavior within three years. The coefficient for ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 would load negatively if strategic

alliances served as channels for tax knowledge diffusion. To alleviate concerns about interpreting a level-

based dependent variable, we also construct the change-indicating variable 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3, which is

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] scaled by 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0]. Because strategic alliances go beyond linking high-

tax and low-tax firms, we include vectors of variables on 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠

(𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 & 𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦) and 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 in equation (3). Furthermore, we control for within-

firm determinants of tax planning by including a vector of 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠. We follow Dyreng et al. (2010)

and consider 𝐸𝐵𝐼𝑇𝐷𝐴3, 𝑅𝑛𝐷𝐸𝑥𝑝3, 𝐴𝑑𝐸𝑥𝑝3, 𝑆𝐺𝐴3, 𝐶𝑎𝑝𝐸𝑥3, 𝐶ℎ𝑎𝑛𝑔𝑒𝑆𝑎𝑙𝑒3, 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒3, 𝐶𝑎𝑠ℎ3,

𝑀𝑁𝐸3, 𝑁𝑂𝐿3, 𝐼𝑛𝑡𝑎𝑛𝑔𝑖𝑏𝑙𝑒𝑠3, 𝑃𝑃𝐸3, and 𝑆𝑖𝑧𝑒3. We include year (𝜏𝑡) and industry (𝛿𝑖𝑛𝑑) fixed effects

and cluster standard errors at the firm level (Petersen, 2009).

3.5. Difference in Differences (DiD)

The multiperiod dependent variables in the specifications of equation (3) allow us to tie tax knowledge

diffusion to the year of network initiation and to account for 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 & 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠. An alternative

approach for measuring tax knowledge diffusion is to maintain the panel structure of our data and apply a

15

DiD methodology. In this model, 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 is generally in alignment with ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤. However, we

create an embargo period around a network observation during which a firm may not invest in another

network (exclusion of overlapping events). Given a suggested average lifespan for strategic alliances of

five years (Chan et al., 1997), this embargo period contains the three years preceding (𝑝𝑜𝑠𝑡 = 0) and five

years subsequent (𝑝𝑜𝑠𝑡 = 1) to a network initiation.8 We adjust dependent and control variables from

multiperiod measures to their single-year versions:

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅𝑖,𝑡 = 𝛽0 + 𝛽1𝑡𝑟𝑒𝑎𝑡𝑒𝑑𝑖 + 𝛽2𝑝𝑜𝑠𝑡𝑡 + 𝜷𝟑𝒕𝒓𝒆𝒂𝒕𝒆𝒅 ∗ 𝒑𝒐𝒔𝒕𝒊,𝒕

+ ∑ 𝛽𝑘𝑘

𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡𝑘 + 휀𝑖,𝑡.

(4)

In this model, 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 measures the baseline difference in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 that is not due to the presence of the

treatment. The parameter 𝑝𝑜𝑠𝑡 captures changes in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 from before to after treatment. The parameter

of interest is the interaction 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 which measures the effect on 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 due to the treatment

(i.e., low-tax network of high-tax firm). Generally, the high-tax firms are in very similar situations except

for potentially experiencing tax knowledge diffusion. However, if characteristics of high-tax firms investing

into low-tax networks differed from the characteristics of high-tax firms investing into high-tax networks,

a concern about equation (4) would be that these differences drive observed differences in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅. In

addition to excluding overlapping events, we also employ entropy-balancing weighting (Hainmueller,

2012; Hainmueller & Xu, 2013) and use the entropy weights to reestimate equation (4). Observations are

balanced using 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 so that the means and the variance in the reweighted control group match

the treatment group (balanced sample).

8 We use an expected lifespan of five years because SDC does not provide sufficient information on the termination

of networks.

16

4. Results & Discussion

4.1. Descriptive Insights

[Table 2]

We start our analyses by performing a descriptive analysis of the changes in the tax planning behavior of

high-tax firms in strategic alliances with low-tax firms in comparison to high-tax firms in strategic alliances

with high-tax firms. Our focus is on our primary measure of 𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒, 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3]. First,

we compare the changes from 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] to 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] within the ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤

groups. We observe reductions in cash effective tax rates for both groups (Within-Group Change in

Table 2). While these decreases could comprise reversion to the mean, they would not explain differences

in the development between the groups. Therefore, we test for the difference between groups and between

periods (Difference in Within-Group Change). The respective difference of 4.3 percentage points is highly

significant (p‑value 0.0143). Observations of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1 are accompanied by a mean

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] of 25.57% and networks solely among high-tax firms (ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0) are aligned to

an average 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] of 28.04%. The descriptive statistics in Panel A of Table 1 on

𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 indicate a skewed distribution. A conclusion from this could be that our findings are

induced by increases in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 for high-tax firms in high-tax networks. Therefore, we not only test for

the difference of the means of 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 (p-value 0.0422) but also for the difference of the medians.

The result indicates a negative and significant difference (p-value 0.0935). We conjecture that our

inferences are not biased from increases in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 in the control group. We interpret these findings as

a first indication toward the existence of tax knowledge diffusion via strategic alliances.

4.2. Regression Results

[Table 3]

4.2.1. Baseline Regression

The main variable of interest in our regression analysis is ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 because it isolates the incremental

effect a low-tax network exerts on a high-tax firm’s 𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒. In Panel A of Table 3, we show

17

multiperiod specifications of 𝑇𝑎𝑥 𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒 for equation (3) with 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] and

𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 as dependent variables. By observing multiperiod measures, we can link our inferences

to the period of network initiation. For brevity and in support of refraining from discussing marginal effects

of control variables (Hünermund & Louw, 2020), we respectively report the coefficient estimates for

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠.9 The estimates for ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 are negative and significant in both

specifications. In the specification with 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] as the dependent variable, the estimate for

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 has a magnitude of -0.0273 (p‑value 0.0470). This is consistent with our descriptive inferences

on 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] in terms of direction and magnitude. Because an overall network effect is absorbed

by including high-tax firms in our regression and because the covariates of

𝑝𝑎𝑟𝑡𝑛𝑒𝑟, 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 & 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 account for a broad range of alternative explanations, we find it

plausible to associate the (relative) increase in tax planning activities for high-tax firms in low-tax networks

to be induced by the presence of a low-tax firm in the network. Extending equation (3) to the change-

indicating variable 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 yields corresponding implications (p‑value 0.0121). Interestingly, the

estimates for the 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠, namely, 𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦 and 𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟, do not surpass the common

levels of significance. In several additional analyses (see Section 5.2), we focus on interactions of

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 to investigate whether 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 complement or substitute the

identified effects. Furthermore, we do not observe that the business purposes of the strategic alliances

(𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠) drive our findings. Solely the coefficient for 𝑃𝑢𝑟𝑝𝑜𝑠𝑒𝑅𝑛𝐷 loads negative and

significant with 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 as the dependent variable (p-value 0.0769). This finding is consistent

with recent research that shows that strategic alliances in R&D lead to higher patent output (Li et al., 2019)

and that patents have a positive causal effect on corporate tax planning that is incremental to the effect of

R&D expenses on tax planning (Cheng et al., 2020). Generally, this supports our notion that strategic

alliances are not intended to establish tax planning investments. Therefore, we argue to identify tax

9 The Online Supplement contains Stata log-files and Table OS 1 which display all variables.

18

knowledge diffusion as a second-order effect of peer-to-peer cooperation. Our findings convey that

decreases in cash taxes paid are driven by the partner firm’s tax planning behavior.

4.2.2. Difference in Differences

Parallel Trend Assumption

[Figure 5]

Any DiD specification relies on the parallel trend assumption. Otherwise, one could not empirically identify

the posttreatment outcome absent the treatment. Usually, the parallel trend assumption is graphically

examined by observing pretreatment trends of the dependent variable among the treatment and control

groups. Accordingly, Panel A of Figure 5 provides visual documentation that the trends of 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 run

parallel for treatment and control firms prior to the treatment. Although a direct empirical test for the parallel

trend assumption is not possible, Patel & Seegert (2015) developed an approach to alleviate concerns about

potential confounding factors. They suggest regressing the treatment indicator, time fixed effects (i.e.,

𝑒𝑚𝑏𝑎𝑟𝑔𝑜 𝑝𝑒𝑟𝑖𝑜𝑑 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠) and the interaction of the treatment indicator and fixed effects on the

dependent variable:

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅𝑖,𝑡 = 𝛼 + 𝑡𝑟𝑒𝑎𝑡𝑒𝑑𝑖 + 𝑒𝑚𝑏𝑎𝑟𝑔𝑜𝑡 + 𝑒𝑚𝑏𝑎𝑟𝑔𝑜 ∗ 𝑡𝑟𝑒𝑎𝑡𝑒𝑑𝑖,𝑡 + 휀𝑖,𝑡. (5)

Failure to reject that the coefficient estimates for the interaction terms for

𝑒𝑚𝑏𝑎𝑟𝑔𝑜 ∗ 𝑡𝑟𝑒𝑎𝑡𝑒𝑑𝑖,𝑡 are jointly zero in the pretreatment period supports the parallel trend assumption. In

Panel B of Figure 5, we therefore present the coefficient and the 95% confidence interval of the interaction

of the treatment indicator and time fixed effects from equation (5) for the pretreatment years. None of the

presented individual coefficients are significantly different from zero. Moreover, the p‑value of the parallel

trend test (the coefficients of 𝑒𝑚𝑏𝑎𝑟𝑔𝑜𝑡 ∗ 𝑡𝑟𝑒𝑎𝑡𝑒𝑑𝑖 are jointly zero during pretreatment) is beyond the

usual levels of significance (p-value 0.4215). Consequently, we are able to provide visual and statistical

support for the parallel trend assumption.

19

Results

Panel B of Table 3 depicts two specifications of equation (4) that both include

𝑒𝑚𝑏𝑎𝑟𝑔𝑜 𝑝𝑒𝑟𝑖𝑜𝑑 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠 around the initiation of a network. This subsumes the 𝑝𝑜𝑠𝑡 indicator

(Gallemore et al., 2019). We include 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 and 𝑦𝑒𝑎𝑟 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠, which capture influences

that affect the tax planning behavior across all sample firms within a given year, as well as 𝛿𝑖𝑛𝑑. In the

second specification, entropy balancing weights are applied (balanced sample). The estimate for the

interaction of 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 is negative and significant in both specifications. Consequently, we find a

negative 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 response to low-tax networks in comparison to high-tax networks for high-tax firms.10

Taken together, the results from our descriptive and regression analyses are consistent with strategic

alliances serving as channels through which tax knowledge diffuses between firms.

4.2.3. Adjustment Speed

We are further interested in how fast high-tax firms adjust their tax planning behavior when cooperating

with low-tax firms, indicating the speed of tax knowledge diffusion. Therefore, we estimate five

specifications of equation (4). We extend the posttreatment period by one year with each specification.

Thus, we notate 𝑝𝑜𝑠𝑡 = 1 only for the year of network initiation [𝑡1] first and finish with 𝑝𝑜𝑠𝑡 equaling

one for the entire posttreatment embargo period [𝑡1; 𝑡5]. The coefficients of 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 ∗ 𝑝𝑜𝑠𝑡 present the

cumulative adjustment of a high-tax firm’s tax planning behavior with progressing time (𝑡1 to 𝑡5) when

cooperating with low-tax firms (i.e., adjustment speed). The results of this analysis are presented in Panel C

of Table 3. In accordance with the theory that suggests that diffusion of (tax) knowledge increases in

probability over time, the coefficient of 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 ∗ 𝑝𝑜𝑠𝑡 is not significant when the posttreatment period

is limited to the year of network initiation. This finding is consistent with our notion that strategic alliances,

per se, do not aim at facilitating tax planning. The coefficient of the interaction, however, turns significant

10 Our results are robust to (i) excluding 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠, 𝛿𝑖𝑛𝑑, and 𝜏𝑡 from the model (Table OS 2) , (ii) using the cash

tax differential developed by Henry & Sansing (2018) which reflects the extent to which a firm is tax-favored

(Table OS 3 and Figure OS 2 Panel A), and (iii) estimating the effect on profitability (Table OS 4).

20

when 𝑝𝑜𝑠𝑡 spans from 𝑡1 to 𝑡2. Furthermore, the estimated effect continues to be significant when extending

the posttreatment period to 𝑡3, 𝑡4, and 𝑡5. Consequently, we assess that high-tax firms, on average, are able

to adjust their tax planning behavior within two years of network initiation. These findings are consistent

with recent research by Kim et al. (2019) who suggest that firms generally are able to adjust their tax

planning behavior within three years and that high-tax firms may increase their tax planning behavior even

faster. Additionally, our results suggest that a firm’s adjustment of its tax planning behavior, once

implemented, stays rather constant over subsequent years.

5. Additional Analyses

5.1. Effects on Reporting of Operations

[Table 4]

To triangulate our findings, we analyze the effects of low-tax networks on firms’ reporting of operations.

Thereby, we identify drivers of the changes in tax planning behavior and corroborate our evidence that

these changes are due to tax knowledge diffusion.

Textual Sentiment of 10-K Filings

Law & Mills (2015) show that linguistic cues in firms’ qualitative disclosures provide incremental

information beyond traditional accounting variables to predict tax planning activities. They provide

evidence that the use of negative words in firms’ 10-K filings suggests (future) tax planning. Their

identification strategy relies on textual sentiment signaling external financial constraints. The authors show

that constrained firms react by increasing their tax planning behavior as a substitute for a more expensive

source of external financing. Quantifying language to measure firms’ fundamentals has received massive

interest in accounting and finance literature since Tetlock et al. (2008) and Loughran & McDonald (2011)

pioneered in this field (for comprehensive reviews, see Loughran & McDonald (2016; 2019) and Teoh

(2018)). Commonly, accounting-specific dictionaries (bag of words) that share common sentiments (e.g.,

positive, negative) are used to measure a document’s textual sentiment. While word classifications can

21

largely differ according to the investigated setting (Loughran et al., 2019; Loughran & McDonald, 2020),

the focus is generally on the use of negative words.

We test for this mechanism by analyzing whether the textual sentiment in firms’ 10-K-filings changes

differently for high-tax firms in strategic alliances with low-tax firms relative to high-tax firms in high-tax

strategic alliances. We estimate specifications of equation (4) with 𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡 (𝑈𝑠𝑒 𝑜𝑓 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑊𝑜𝑟𝑑𝑠)

as dependent variables.11 The results are depicted in Table 4. The estimate for the interaction of 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗

𝑝𝑜𝑠𝑡 is negative (positive) and significant in both specifications. We find a negative response of textual

sentiment of 10-K filings to low-tax networks in comparison to high-tax networks for high-tax firms. This

finding is consistent with our inferences on changes in firms’ tax planning behavior and with theory and

research on sentiment as a predictor of tax planning behavior.

Tax Haven Operations

An operational mechanism with regard to international tax planning is profit shifting from high-tax

jurisdictions to tax havens. US firms with material operations in at least one tax haven country are shown

to report lower effective tax rates (Dyreng & Lindsey, 2009). Firms, though, are required to disclose a list

of their subsidiaries under “Exhibit 21” when filing their annual reports (form 10-K). Exhibit 21 lists all

subsidiaries of a registrant, the state, or other jurisdiction of incorporation or organization of each, and the

names under which such subsidiaries do business. Exceptions apply for subsidiaries, which are not

considered “significant” (see Dyreng et al. (2020) and Demeré et al. (2019) for detailed explanations). If

low-tax networks paved the way for increases in international tax planning activities, one could observe

11 Data on textual sentiment is shared publicly by Bill McDonald. Please refer to the variable definitions in the

Appendix for further details. A graphical investigation of the parallel trend test is presented in the Online Supplement

(Figure OS 2 Panel B).

22

increases in tax haven usage with the purpose of profit shifting. We estimate specifications of equation (4)

with 𝑈𝑠𝑒 𝑜𝑓 𝑇𝑎𝑥 𝐻𝑎𝑣𝑒𝑛 and 𝑁𝑢𝑚 𝑜𝑓 𝑇𝑎𝑥 𝐻𝑎𝑣𝑒𝑛 𝑆𝑢𝑏𝑠𝑖𝑑𝑖𝑎𝑟𝑖𝑒𝑠 as dependent variables.12

The estimates do not suggest an increase in tax haven operations for high-tax firms in low-tax networks.

Conversely, the estimates for the interaction of 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 are negative and significant in both

specifications (Table 4). This suggests a decrease in tax haven operations for high-tax firms in low-tax

networks relative to high-tax firms in high-tax networks. These results appear inconsistent with the

identified main effects. However, Dyreng & Markle (2016) posit that the positive association of tax

planning activities and financial constraints (i.e., use of negative words) does not stem from profit shifting

to tax haven subsidiaries. Consistently, Edwards et al. (2016) show that constrained firms especially achieve

tax savings via deferral-based tax planning strategies. Consequently, we do not expect the probability of

outbound profit shifting to generally increase with decreasing 10-K sentiment. Furthermore, our focus is

on strategic alliances between US firms and not on international alliances. Data from Exhibit 21 also cannot

quantify the magnitude, extent, or legal structure of all tax haven activities. In particular, one cannot identify

specific tax planning transactions (Law & Mills, 2019). There is also some evidence of nondisclosure when

subsidiaries are in tax havens (Dyreng et al., 2020). Consistently, the concealment of tax haven operations,

when increasing tax planning behavior, would point toward changes in reporting but not in operations.

5.2. Partner Characteristics

[Table 5]

Thus far, we have considered the overall implications of low-tax networks for high-tax firms. We are also

interested in partnering firms’ characteristics, which could intensify or mitigate the identified effects.

Thereby, we focus on geographical proximity (Panel A of Table 5), identical industry affiliation (Panel B

of Table 5) and shared audit firms (Panel C of Table 5). Shared membership in a geographical region

12 Exhibit 21 disclosure data is shared publicly by Scott Dyreng. Please refer to the variable definitions in the Appendix

for further details. A graphical investigation of the parallel trend test is presented in the Online Supplement

(Figure OS 2 Panel C).

23

(𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛) could allow higher levels of interfirm interaction, as hypothesized by Brown (2011).

Furthermore, Brown & Drake (2014) assume that firms with the same industry affiliation (𝑆𝑎𝑚𝑒𝐼𝑛𝑑) share

the same operating environment and suggest that this could enhance the identified tax knowledge transfers.

We run specifications of equation (3) with (𝑑𝑒𝑙𝑡𝑎) 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 as the dependent variable and include

interaction terms of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and the respective 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙. The coefficient estimates for the

interaction terms comprise the incremental influence of geographical proximity, shared industry affiliation,

and shared audit firms on tax knowledge diffusion via low-tax networks.

In Panel A of Table 5, we interact ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 with the indicator variable 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛. We observe

significant and negative coefficient estimates of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 in both specifications. However, the estimate

for 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 and the interaction term do not surpass common levels of statistical significance.

While this finding is generally consistent with the inferences by Brown (2011) on geographical proximity,

Cen et al. (2020) report that the correlation of effective tax rates is stronger for the members of a supply

chain that are located within the same region. This emphasizes the importance of various channels for the

diffusion and transfers of tax knowledge. Our inferences for interacting ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 with 𝑆𝑎𝑚𝑒𝐼𝑛𝑑 are

similar. In particular, we cannot reject that the interactions of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and 𝑆𝑎𝑚𝑒𝐼𝑛𝑑 are significantly

different from zero. This finding is consistent with Brown & Drake (2014) and in line with our inferences

on geographical proximity. It supports the notion that tax shelters that are repetitive among firms are not

limited by industry barriers and serve as mass-market tax-saving ideas (Lisowsky, 2010). Finally, we

examine the role of shared audit firms (𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟). In specifications of equation (3) with

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] and 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 as dependent variables, the interaction terms for ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 ∗

𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 do not surpass common levels of statistical significance. Our findings indicate that a shared

audit firm serves as a substitute rather than as a complement for low-tax networks of high-tax firms. These

inferences take their place alongside a range of mixed inferences concerning auditors’ impact on firms’ tax

planning behavior. Brown (2011) does not find significant tax shelter adoption via shared audit firms, and

Klassen et al. (2016) show that less tax aggressiveness in the past is associated with the auditor preparing

24

a firm’s tax return. In contrast, Lim et al. (2018) and Cen et al. (2020) suggest that common auditors

facilitate tax planning. Consistent with the mixed evidence from prior literature, Nesbitt et al. (2020)

suggest that there are limits to the relation between auditor-provided tax services and clients’ tax

aggressiveness.

6. Robustness Checks

We examine the robustness of our findings regarding alternative explanations and potential concerns about

our identification strategy.

6.1. Alternative Explanations

[Table 6]

Prior research indicates that strategic alliances can serve as preliminary ties between successive acquirers

and targets (Ishii & Xuan, 2014; Porrini, 2004). Survivorship bias could thus exert an influence on our

inferences. We present a specification of equation (3) in Panel A of Table 6 in which we exclude firm-edge-

years from the analysis (i.e., firms with network-firm observations within the last three years of their

presence in our panel). This approach also ensures that variation in 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 does not stem from the

substitution with 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 for the final and penultimate firm-year of a firm in our panel. In this model, we

still find the loading of ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 to be significant (p‑value 0.0162) and the economic magnitude to be

consistent with our primary findings.

We identify tax knowledge diffusion for high-tax firms via low-tax networks. We do not consider the tax

position of low-tax firms because there is little reason to expect incremental, “negative” tax knowledge

diffusion for low-tax firms bound to high-tax firms. To empirically control for this notion, we construct

𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ, which is an indicator that equals one for low-tax firms in networks with high-tax firms and

zero for low-tax firms in low-tax networks. The results for this specification of equation (3) are presented

in Panel B of Table 6. The coefficient estimate for 𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ is far beyond common levels of significance,

25

with a p‑value of 0.3298. This finding further supports our inferences of tax knowledge diffusion occurring

among high-tax firms.

6.2. Alternative Identification Strategy

[Table 7]

There are some judgment calls involved in classifying strategic alliances as high-tax- and low-tax networks.

Therefore, we apply a modified identification strategy to check our strategy’s robustness. We enter all firms

in a regression and include an indicator variable ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚, the 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] of

a firm’s network-partner and the interaction 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 ∗ ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚:

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=1 = 𝛽0 + 𝛽1ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚𝑖,𝑡=1 + 𝛽2 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=0

+ 𝜷𝟑 𝒑𝒂𝒓𝒕𝒏𝒆𝒓 𝒑𝒓𝒆 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 ∗ 𝒉𝒊𝒈𝒉‑𝒕𝒂𝒙 𝒇𝒊𝒓𝒎

+ ∑ 𝛽𝑛𝑛

𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡=1𝑛 + ∑ 𝛽𝑘

𝑘𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡=1

𝑘

+ 𝛿𝑖𝑛𝑑 + 𝜎𝑖𝑛𝑑_𝑛𝑒𝑡𝑤𝑜𝑟𝑘 + 휀𝑖,𝑡 .

(6)

The results for this model are depicted in Table 7. The coefficient of ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚 loads positively and

is highly significant. This implies that the separation of low-tax and high-tax firms based on adjusted

𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] is robust. Furthermore, the coefficient of 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 ∗

ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚 is positive and statistically significant. Consequently, we observe a positive association

between partner firms’ cash effective tax rates for high-tax firms. This is not only consistent with the

findings from our main analysis (high-tax firms show higher cash effective tax rates when cooperating with

high-tax firms) but also with our theory and identified implications for low-tax firms. The latter argument

is strengthened by observing that the coefficient of 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 is far from common levels of

statistical significance.

7. Conclusion

The purpose of this study is to shed light on channels through which tax knowledge diffuses between firms.

Despite solid insights on the within-firm determinants of corporate tax planning, research on cross-firm

26

connections in this field is still developing. We contribute to this emerging literature by being the first to

identify strategic alliances as channels of tax knowledge diffusion between firms. With our study, we ease

prior studies’ assumptions of existing intermediaries and intentional transfers of tax knowledge. By using

data on strategic alliances between publicly traded US firms, we can distinguish between networks that

bring together high-tax and low-tax firms. Our results suggest an economically meaningful decrease in cash

effective tax rates of high-tax firms in strategic alliances with low-tax firms relative to high-tax firms in

high-tax strategic alliances. Furthermore, we find that the adjustment occurs on average not before the

second year after a strategic alliance’s initiation. We identify a negative response of textual sentiment of

10-K filings to low-tax networks in comparison to high-tax networks for high-tax firms, which predicts

increases in tax planning behavior. Further, we show that the identified effects from our analyses seem not

to not stem from changes in tax haven operations.

This study comes with limitations. We stipulate that strategic alliances are not intended to establish tax

planning investments. While our results support this notion, a firm’s decision to engage in a network is

intentional and not random. Once a firm experiences knowledge diffusion via a low-tax network, engaging

in subsequent networks could also depend on an expected learning effect. Following the recommendation

by Lennox et al. (2012), we emphasize that endogenous treatment assignment would affect the inferences

from OLS regressions. Nevertheless, research indicates that statutory tax rates are not negatively associated

with investment decisions on strategic alliances in multicountry investigations (Owen & Yawson, 2013).

More importantly, we would expect firms to cover their tracks by not disclosing the partner if a network’s

primary aim were to facilitate tax planning. These cases, however, are excluded from our sample since we

require the identification of all partners in a network. Finally, one could simply acquire tax planning

strategies from tax advisors if this were the prior concern of a firm’s management. Consistently, the most

prominent partner firms in our sample are high-tax firms. This conceptually underlines the identification of

tax knowledge diffusion via strategic alliances as a second‑order effect.

27

We think that our study is interesting for regulators because our results highlight that the determinants of a

firm’s tax planning behavior are not limited to within-firm characteristics or intentional transfers of tax

knowledge. Although prior research has dedicated substantial attention to corporate cooperation and firms’

tax planning individually, there are a number of unanswered questions regarding the interplay of the two.

Primarily, we encourage research on the actual mechanisms of (tax) knowledge diffusion.

28

References

Anand, B. N., & Khanna, T. (2000). Do Firms Learn to Create Value? The Case of Alliances. Strategic Management Journal,

21(3), 295–315. http://www.jstor.org/stable/3094189.

Aobdia, D. (2015). Proprietary information spillovers and supplier choice: evidence from auditors. Review of Accounting Studies,

20(4), 1504–1539. https://doi.org/10.1007/s11142-015-9327-x.

Armstrong, C. S., Blouin, J. L., Jagolinzer, A. D., & Larcker, D. F. (2015). Corporate governance, incentives, and tax avoidance.

Journal of Accounting and Economics, 60(1), 1–17. https://doi.org/10.1016/j.jacceco.2015.02.003.

Austin, C. R., & Wilson, R. J. (2017). An Examination of Reputational Costs and Tax Avoidance: Evidence from Firms with

Valuable Consumer Brands. Journal of the American Taxation Association, 39(1), 67–93. https://doi.org/10.2308/atax-51634.

Barrios, J. M., & Gallemore, J. (2019). Tax-Related Human Capital: Evidence from Employee Movements. Working Paper, SSRN.

https://doi.org/10.2139/ssrn.3285310.

Blouin, J., & Robinson, L. (2020). Double Counting Accounting: How Much Profit of Multinational Enterprises Is Really in Tax

Havens? Working Paper, SSRN. https://doi.org/10.2139/ssrn.3491451.

Boone, A. L., & Ivanov, V. I. (2012). Bankruptcy spillover effects on strategic alliance partners. Journal of Financial Economics,

103(3), 551–569. https://doi.org/10.1016/j.jfineco.2011.10.003.

Brown, J. L. (2011). The Spread of Aggressive Corporate Tax Reporting: A Detailed Examination of the Corporate-Owned Life

Insurance Shelter. The Accounting Review, 86(1), 23–57. https://doi.org/10.2308/accr.00000008.

Brown, J. L., & Drake, K. D. (2014). Network Ties Among Low-Tax Firms. The Accounting Review, 89(2), 483–510.

https://doi.org/10.2308/accr-50648.

Cai, Y., Dhaliwal, D. S., Kim, Y., & Pan, C. (2014). Board interlocks and the diffusion of disclosure policy. Review of Accounting

Studies, 19(3), 1086–1119. https://doi.org/10.1007/s11142-014-9280-0.

Cai, Y., Kim, Y., Park, J. C., & White, H. D. (2016). Common auditors in M&A transactions. Journal of Accounting and Economics,

61(1), 77–99. https://doi.org/10.1016/j.jacceco.2015.01.004.

Cai, Y., & Sevilir, M. (2012). Board connections and M&A transactions. Journal of Financial Economics, 103(2), 327–349.

https://doi.org/10.1016/j.jfineco.2011.05.017.

Cen, L., Maydew, E. L., Zhang, L., & Zuo, L. (2017). Customer–supplier relationships and corporate tax avoidance. Journal of

Financial Economics, 123(2), 377–394. https://doi.org/10.1016/j.jfineco.2016.09.009.

Cen, L., Maydew, E. L., Zhang, L., & Zuo, L. (2020). Tax Planning Diffusion, Real Effects, and Sharing of Benefits. Working

Paper, SSRN. https://doi.org/10.2139/ssrn.3213967.

Chan, S. H., Kensinger, J. W., Keown, A. J., & Martin, J. D. (1997). Do strategic alliances create value? Journal of Financial

Economics, 46(2), 199–221. https://doi.org/10.1016/S0304-405X(97)00029-9.

Chen, J., King, T.-H. D., & Wen, M.-M. (2015). Do joint ventures and strategic alliances create value for bondholders? Journal of

Banking & Finance, 58, 247–267. https://doi.org/10.1016/j.jbankfin.2015.03.020.

Cheng, C. A., Guo, P., Weng, C.‐H., & Wu, Q. (2020). Innovation and Corporate Tax Planning: The Distinct Effects of Patents

and R&D. Contemporary Accounting Research, Forthcoming. https://doi.org/10.1111/1911-3846.12613.

Clausing, K. 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. https://doi.org/10.17310/ntj.2016.4.09.

Conroy, R. M. (2012). What Hypotheses do “Nonparametric” Two-Group Tests Actually Test? The Stata Journal, 12(2), 182–190.

https://doi.org/10.1177/1536867X1201200202.

Demeré, P., Donohoe, M. P., & Liswosky, P. (2019). The economic effects of special purpose entities on corporate tax avoidance.

Contemporary Accounting Research, Forthcoming. https://doi.org/10.1111/1911-3846.12580.

Desai, M. A., Foley, C. F., & Hines, J. R. (2004). The costs of shared ownership: Evidence from international joint ventures.

Journal of Financial Economics, 73(2), 323–374. https://doi.org/10.1016/j.jfineco.2003.07.001.

Dhaliwal, D. S., Lamoreaux, P. T., Litov, L. P., & Neyland, J. B. (2016). Shared auditors in mergers and acquisitions. Journal of

Accounting and Economics, 61(1), 49–76. https://doi.org/10.1016/j.jacceco.2015.01.005.

Di Maggio, M., Franzoni, F., Kermani, A., & Sommavilla, C. (2019). The relevance of broker networks for information diffusion

in the stock market. Journal of Financial Economics, 134(2), 419–446. https://doi.org/10.1016/j.jfineco.2019.04.002.

Dyer, J. H., & Hatch, N. W. (2006). Relation-specific capabilities and barriers to knowledge transfers: creating advantage through

network relationships. Strategic Management Journal, 27(8), 701–719. https://doi.org/10.1002/smj.543.

Dyreng, S., & Maydew, E. L. (2017). Virtual Issue on Tax Research Published in the Journal of Accounting Research. Working

Paper, SSRN. https://doi.org/10.2139/ssrn.3068168.

Dyreng, S. D., Hanlon, M., & Maydew, E. L. (2008). Long‐Run Corporate Tax Avoidance. The Accounting Review, 83(1), 61–82.

https://doi.org/10.2308/accr.2008.83.1.61.

29

Dyreng, S. D., Hanlon, M., & Maydew, E. L. (2010). The Effects of Executives on Corporate Tax Avoidance. The Accounting

Review, 85(4), 1163–1189. https://doi.org/10.2308/accr.2010.85.4.1163.

Dyreng, S. D., Hanlon, M., & Maydew, E. L. (2019). When Does Tax Avoidance Result in Tax Uncertainty? The Accounting

Review, 94(2), 179–203. https://doi.org/10.2308/accr-52198.

Dyreng, S. D., Hanlon, M., Maydew, E. L., & Thornock, J. R. (2017). Changes in corporate effective tax rates over the past 25

years. Journal of Financial Economics, 124(3), 441–463. https://doi.org/10.1016/j.jfineco.2017.04.001.

Dyreng, S. D., Hoopes, J. L., Langetieg, P., & Wilde, J. H. (2020). Strategic Subsidiary Disclosure. Journal of Accounting Research,

58(3), 643–692. https://doi.org/10.1111/1475-679X.12308.

Dyreng, S. D., & Lindsey, B. P. (2009). Using Financial Accounting Data to Examine the Effect of Foreign Operations Located in

Tax Havens and Other Countries on U.S. Multinational Firms' Tax Rates. Journal of Accounting Research, 47(5), 1283–1316.

https://doi.org/10.1111/j.1475-679X.2009.00346.x.

Dyreng, S. S., & Markle, K. S. (2016). The Effect of Financial Constraints on Income Shifting by U.S. Multinationals. The

Accounting Review(91 (6)), 147–186. https://doi.org/10.2308/accr-51420.

Edwards, A., Schwab, C., & Shevlin, T. (2016). Financial Constraints and Cash Tax Savings. The Accounting Review, 91(3), 859–

881. https://doi.org/10.2308/accr-51282.

Erickson, M. (1998). The Effect of Taxes on the Structure of Corporate Acquisitions. Journal of Accounting Research, 36(2), 279–

298. https://doi.org/10.2307/2491478.

Erickson, M. M., & Wang, S.‐W. (1999). Exploiting and Sharing Tax Benefits: Seagram and DuPont. Journal of the American

Taxation Association, 21(2), 35–54. https://doi.org/10.2308/jata.1999.21.2.35.

Frey, L. (2018). Tax certified individual auditors and effective tax rates. Business Research, 11(1), 77–114.

https://doi.org/10.1007/s40685-017-0057-8.

Gale, D., & Kariv, S. (2003). Bayesian learning in social networks. Games and Economic Behavior, 45(2), 329–346.

https://doi.org/10.1016/S0899-8256(03)00144-1.

Gallemore, J., Gipper, B., & Maydew, E. L. (2019). Banks as Tax Planning Intermediaries. Journal of Accounting Research, 57(1),

169–209. https://doi.org/10.1111/1475-679X.12246.

Gallemore, J., Maydew, E. L., & Thornock, J. R. (2014). The Reputational Costs of Tax Avoidance. Contemporary Accounting

Research, 31(4), 1103–1133. https://doi.org/10.1111/1911-3846.12055.

Gomes-Casseres, B., Hagedoorn, J., & Jaffe, A. B. (2006). Do alliances promote knowledge flows? Journal of Financial

Economics, 80(1), 5–33. https://doi.org/10.1016/j.jfineco.2004.08.011.

Graham, J. R., Hanlon, M., Shevlin, T., & Shroff, N. (2014). Incentives for Tax Planning and Avoidance: Evidence from the Field.

The Accounting Review, 89(3), 991–1023. https://doi.org/10.2308/accr-50678.

Guenther, D. A., Matsunaga, S. R., & Williams, B. M. (2017). Is Tax Avoidance Related to Firm Risk? The Accounting Review,

92(1), 115–136. https://doi.org/10.2308/accr-51408.

Hainmueller, J. (2012). Entropy Balancing for Causal Effects: A Multivariate Reweighting Method to Produce Balanced Samples

in Observational Studies. Political Analysis, 20(1), 25–46. www.jstor.org/stable/41403737.

Hainmueller, J., & Xu, Y. (2013). Ebalance: A Stata Package for Entropy Balancing. Journal of Statistical Software, 54(7).

https://doi.org/10.2139/ssrn.1943090.

Hanlon, M., & Heitzman, S. (2010). A review of tax research. Journal of Accounting and Economics, 50(2-3), 127–178.

https://doi.org/10.1016/j.jacceco.2010.09.002.

Hanlon, M., & Slemrod, J. (2009). What does tax aggressiveness signal? Evidence from stock price reactions to news about tax

shelter involvement. Journal of Public Economics, 93(1-2), 126–141. https://doi.org/10.1016/j.jpubeco.2008.09.004.

Henry, E., & Sansing, R. (2018). Corporate tax avoidance: data truncation and loss firms. Review of Accounting Studies, 23(3),

1042–1070. https://doi.org/10.1007/s11142-018-9448-0.

Higgins, M. J., & Rodriguez, D. (2006). The outsourcing of R&D through acquisitions in the pharmaceutical industry. Journal of

Financial Economics, 80(2), 351–383. https://doi.org/10.1016/j.jfineco.2005.04.004.

Hoppe, T., Schanz, D., Sturm, S., & Sureth-Sloane, C. (2019). Measuring Tax Complexity Across Countries: A Survey Study on

MNCs. Working Paper, SSRN. https://doi.org/10.2139/ssrn.3469663.

Hundsdoerfer, J., & Jacob, M. (2019). Conforming Tax Planning and Firms’ Cost Behavior. Working Paper, SSRN.

https://doi.org/10.2139/ssrn.3430726.

Hünermund, P., & Louw, B. (2020). On the Nuisance of Control Variables in Regression Analysis. Working Paper.

https://arxiv.org/abs/2005.10314v2.

Isaksson, O. H.D., Simeth, M., & Seifert, R. W. (2016). Knowledge spillovers in the supply chain: Evidence from the high tech

sectors. Research Policy, 45(3), 699–706. https://doi.org/10.1016/j.respol.2015.12.007.

30

Ishii, J., & Xuan, Y. (2014). Acquirer-target social ties and merger outcomes. Journal of Financial Economics, 112(3), 344–363.

https://doi.org/10.1016/j.jfineco.2014.02.007.

Jiang, X., Bao, Y., Xie, Y., & Gao, S. (2016). Partner trustworthiness, knowledge flow in strategic alliances, and firm

competitiveness: A contingency perspective. Journal of Business Research, 69(2), 804–814.

https://doi.org/10.1016/j.jbusres.2015.07.009.

Kale, P., Singh, H., & Perlmutter, H. (2000). Learning and Protection of Proprietary Assets in Strategic Alliances: Building

Relational Capital. Strategic Management Journal, 21(3), 217–237. http://www.jstor.org/stable/3094186.

Khanna, T., Gulati, R., & Nohria, N. (1998). The Dynamics of Learning Alliances: Competition, Cooperation, and Relative Scope.

Strategic Management Journal, 19(3), 193–210. http://www.jstor.org/stable/3094095.

Kim, J., McGuire, S. T., Savoy, S., Wilson, R., & Caskey, J. (2019). How Quickly Do Firms Adjust to Optimal Levels of Tax

Avoidance? Contemporary Accounting Research, 36(3), 1824–1860. https://doi.org/10.1111/1911-3846.12481.

Klassen, K. J., Lisowsky, P., & Mescall, D. (2016). The Role of Auditors, Non-Auditors, and Internal Tax Departments in Corporate

Tax Aggressiveness. The Accounting Review, 91(1), 179–205. https://doi.org/10.2308/accr-51137.

Klassen, K. J., Lisowsky, P., & Mescall, D. (2017). Transfer Pricing: Strategies, Practices, and Tax Minimization. Contemporary

Accounting Research, 34(1), 455–493. https://doi.org/10.1111/1911-3846.12239.

Kogut, B., & Singh, H. (1988). The Effect of National Culture on the Choice of Entry Mode. Journal of International Business

Studies, 19(3), 411–432. https://doi.org/10.1057/palgrave.jibs.8490394.

Law, K., & Mills, L. F. (2019). Taxes and Haven Activities: Evidence from Linguistic Cues. Working Paper.

https://doi.org/10.2139/ssrn.2768605.

Law, K. K. F., & Mills, L. F. (2015). Taxes and Financial Constraints: Evidence from Linguistic Cues. Journal of Accounting

Research, 53(4), 777–819. https://doi.org/10.1111/1475-679X.12081.

Lennox, C. S., Francis, J. R., & Wang, Z. (2012). Selection Models in Accounting Research. The Accounting Review, 87(2), 589–

616. https://doi.org/10.2308/accr-10195.

Li, K., Qiu, J., & Wang, J. (2019). Technology Conglomeration, Strategic Alliances, and Corporate Innovation. Management

Science, 65(11), 5065–5090. https://doi.org/10.1287/mnsc.2018.3085.

Lim, C. Y., Shevlin, T. J., Wang, K., & Xu, Y. (2018). Tax Knowledge Diffusion Through Individual Auditor Network Ties:

Evidence from China. Working Paper, SSRN. https://doi.org/10.2139/ssrn.3229564.

Lindsey, L. (2008). Blurring Firm Boundaries: The Role of Venture Capital in Strategic Alliances. Journal of Finance, 63(3),

1137–1168. https://doi.org/10.1111/j.1540-6261.2008.01354.x.

Lisowsky, P. (2010). Seeking Shelter: Empirically Modeling Tax Shelters Using Financial Statement Information. The Accounting

Review, 85(5), 1693–1720. https://doi.org/10.2308/accr.2010.85.5.1693.

Loughran, T., & McDonald, B. (2011). When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of

Finance, 66(1), 35–65. https://doi.org/10.1111/j.1540-6261.2010.01625.x.

Loughran, T., & McDonald, B. (2019). Textual Analysis in Finance. Working Paper, SSRN. https://doi.org/10.2139/ssrn.3470272.

Loughran, T., & McDonald, B. (2020). Measuring Firm Complexity. Working Paper, SSRN. https://doi.org/10.2139/ssrn.3645372.

Loughran, T., McDonald, B., & Pragidis, I. (2019). Assimilation of oil news into prices. International Review of Financial Analysis,

63, 105–118. https://doi.org/10.1016/j.irfa.2019.03.008.

Loughran, T. I.M., & McDonald, B. (2016). Textual Analysis in Accounting and Finance: A Survey. Journal of Accounting

Research, 54(4), 1187–1230. https://doi.org/10.1111/1475-679X.12123.

McGuire, S. T., Omer, T. C., & Wang, D. (2012). Tax Avoidance: Does Tax-Specific Industry Expertise Make a Difference? The

Accounting Review, 87(3), 975–1003. https://doi.org/10.2308/accr-10215.

Meier, M. (2011). Knowledge Management in Strategic Alliances: A Review of Empirical Evidence. International Journal of

Management Reviews, 13(1), 1–23. https://doi.org/10.1111/j.1468-2370.2010.00287.x.

Mohanram, P., & Nanda, A. (1996). When do Joint Ventures create value? Academy of Management Proceedings, 1996(1), 36–

40. https://doi.org/10.5465/ambpp.1996.4978170.

Mulligan, E., & Oats, L. (2016). Tax professionals at work in Silicon Valley. Accounting, Organizations and Society, 52, 63–76.

https://doi.org/10.1016/j.aos.2015.09.005.

Nesbitt, W. L., Persson, A., & Shaw, J. (2020). Auditor-Provided Tax Services and Clients’ Tax Avoidance: Do Auditors Draw a

Line in the Sand for Tax Advisory Services? Working Paper, SSRN. https://doi.org/10.2139/ssrn.3556702.

Nonaka, I. (1994). A Dynamic Theory of Organizational Knowledge Creation. Organization Science, 5(1), 14–37.

https://doi.org/10.1287/orsc.5.1.14.

Owen, S., & Yawson, A. (2013). Information asymmetry and international strategic alliances. Journal of Banking & Finance,

37(10), 3890–3903. https://doi.org/10.1016/j.jbankfin.2013.06.008.

31

Patel, E., & Seegert, N. (2015). Using Strategic Interactions in Investment Decisions to Increase the Efficiency of Tax Policy.

Working Paper, SSRN. https://doi.org/10.2139/ssrn.2682827.

Petersen, M. A. (2009). Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches. Review of Financial

Studies, 22(1), 435–480. https://doi.org/10.1093/rfs/hhn053.

Porrini, P. (2004). Can a Previous Alliance Between an Acquirer and a Target Affect Acquisition Performance? Journal of

Management, 30(4), 545–562. https://doi.org/10.1016/j.jm.2004.02.003.

PwC (2018). New Entrants - New Rivals: How Germany's top companies are creating a new industry world.

Rogers, E. M. (2003). Diffusion of innovations. (Fifth edition, Free Press trade paperback edition). Social science. New York,

London, Toronto, Sydney: Free Press.

Schilling, M. A. (2009). Understanding the alliance data. Strategic Management Journal, 30(3), 233–260.

https://doi.org/10.1002/smj.731.

Szulanski, G. (1996). Exploring Internal Stickiness: Impediments to the Transfer of Best Practice Within the Firm. Strategic

Management Journal, 17(1), 27–43. http://www.jstor.org/stable/2486989.

Tan, T. F., & Netessine, S. (2019). When You Work with a Superman, Will You Also Fly? An Empirical Study of the Impact of

Coworkers on Performance. Management Science, 65(8), 3495–3517. https://doi.org/10.1287/mnsc.2018.3135.

Teoh, S. H. (2018). The promise and challenges of new datasets for accounting research. Accounting, Organizations and Society,

68-69, 109–117. https://doi.org/10.1016/j.aos.2018.03.008.

Tetlock, P., Saar-Tsechansky, M., & Macskassy, S. (2008). More Than Words: Quantifying Language to Measure Firms'

Fundamentals. Journal of Finance, 63(3), 1437–1467. https://doi.org/10.1111/j.1540-6261.2008.01362.x.

Tørsløv, T., Wier, L., & Zucman, G. (2018). The Missing Profits of Nations. Working Paper, National Bureau of Economic

Research. https://doi.org/10.3386/w24701.

Wilde, J. H., & Wilson, R. J. (2018). Perspectives on Corporate Tax Planning: Observations from the Past Decade. Journal of the

American Taxation Association, 40(2), 63–81. https://doi.org/10.2308/atax-51993.

Wilson, R. J. (2009). An Examination of Corporate Tax Shelter Participants. The Accounting Review, 84(3), 969–999.

https://doi.org/10.2308/accr.2009.84.3.969.

32

Appendix: Variable Definitions

Variable Definition [Compustat (low)/SDC (CAPITAL)/other (ITALICS) data items]

𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 Multiperiod cash effective tax rate:

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=1 =∑ (3

𝑡=1 𝑡𝑥𝑝𝑑𝑖,𝑡)

∑ (3𝑡=1 𝑝𝑖𝑖,𝑡 − 𝑠𝑝𝑖𝑖,𝑡)

Defined as cash taxes paid (txpd) divided by pre-tax income (pi) before special

items (spi); special items are reset to 0 when missing; numerator and

denominator are constructed as the sum of the current and two subsequent years;

observations with a negative denominator are reset to missing; for the final and

penultimate firm-year of a firm substituted by annual cash ETR; winsorized at 0

and 1.

𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 Cash ETR3 [t1; t3] scaled by pre cash ETR3 [t-2; t0]:

𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 =𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3[𝑡1;𝑡3]

𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅[𝑡−2;𝑡0]

Winsorized at p1 and p99.

𝒆𝒎𝒃𝒂𝒓𝒈𝒐 𝒑𝒆𝒓𝒊𝒐𝒅 Period of three years before [t-2;t0] and five years subsequent [t1;t5] to a network

initiation during which a firm may not invest in another strategic alliance

(exclusion of overlapping events).

Exhibit 21 Data on Exhibit 21 disclosure is shared publicly by Scott Dyreng:

https://sites.google.com/site/scottdyreng/Home/data-and-code/EX21-Dataset.

We consider disclosed tax haven subsidiaries (TAXHAVEN) and rely on the

classification of the jurisdictions as tax havens. We count TAXHAVEN

occurrences ourselves and do not use COUNT because manual inspections of

10-K filings show that country names are counted multiple times when occurring

in the subsidiaries’ names (NAME) as well (as indicated by the data description).

𝒉𝒊𝒈𝒉‑𝒕𝒂𝒙 𝒇𝒊𝒓𝒎 Inverse to low-tax firm; indicator variable in t1; equals 1 if the firm’s industry

mean adjusted pre cash ETR3 in t0 does not belong to the lowest quartile; 0 for

low-tax firms.

𝒉𝒊𝒈𝒉𝒕𝒐𝒍𝒐𝒘 Indicator variable; equals 1 for high-tax firms in low-tax networks; equals 0 for

high-tax firms in high-tax networks.

𝒍𝒐𝒘‑𝒕𝒂𝒙 𝒇𝒊𝒓𝒎 Indicator variable in t1; equals 1 if the firm’s industry adjusted pre cash ETR3 in

t0 belongs to the lowest quartile (“first bin”); 0 for high-tax firms.

𝒍𝒐𝒘𝒕𝒐𝒉𝒊𝒈𝒉 Indicator variable; equals 1 for low-tax firms in high-tax networks; equals 0 for

low-tax firms in low-tax networks.

∑ 𝒏𝒆𝒕𝒘𝒐𝒓𝒌 𝒄𝒐𝒏𝒕𝒓𝒐𝒍𝒔 Indicator variables for the main business purpose of a network, which is derived

from a network’s deal description (DEALTEXT) in SDC; comprises

PurposeWholesale, PurposeRnD, PurposeLicense, PurposeService,

PurposeMarketing, PurposeSupply and PurposeManufacture.

𝑵𝒖𝒎 𝒐𝒇 𝑻𝒂𝒙 𝑯𝒂𝒗𝒆𝒏 𝑺𝒖𝒃𝒔𝒊𝒅𝒊𝒂𝒓𝒊𝒆𝒔 Measure of material tax haven operations, calculated as

𝑁𝑢𝑚 𝑜𝑓 𝑇𝑎𝑥 𝐻𝑎𝑣𝑒𝑛 𝑆𝑢𝑏𝑠𝑖𝑑𝑖𝑎𝑟𝑖𝑒𝑠 = 𝑙𝑛 (1 + ∑ 𝑇𝐴𝑋𝐻𝐴𝑉𝐸𝑁𝑖,𝑡)

See Exhibit 21 for data origin.

∑ 𝒑𝒂𝒓𝒕𝒏𝒆𝒓 𝒄𝒐𝒏𝒕𝒓𝒐𝒍𝒔 See Proximity and SameAuditor for main analysis (equation (3)).

𝒑𝒐𝒔𝒕 Indicator variable; equals 0 [t-2;t0] and 1 for [t1;t5].

𝒑𝒓𝒆 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 Constructed as cash ETR3 but with numerator and denominator constructed as

the sum of the current and two preceding periods:

𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3𝑖,𝑡=0 =∑ (0

𝑡=−2 𝑡𝑥𝑝𝑑𝑖,𝑡)

∑ (0𝑡=−2 𝑝𝑖𝑖,𝑡 − 𝑠𝑝𝑖𝑖,𝑡)

Industry adjusted pre cash ETR [t-2;t0] is used to identify low- and high-tax firms

in t1.

33

Variable Definition [Compustat (low)/SDC (CAPITAL)/other (ITALICS) data items]

𝑷𝒓𝒐𝒙𝒊𝒎𝒊𝒕𝒚 Distance (as the crow flies) between the partners of a network according to the

zip code of the partners’ headquarters (addzip); collected from

freemaptools.com; normalized between 1 and 0 for closest and farthest distance.

𝑺𝒂𝒎𝒆𝑨𝒖𝒅𝒊𝒕𝒐𝒓 Indicator variable; equals 1 when all partners of a network share the same auditor

firm (au) in t1; 0 otherwise.

𝑺𝒂𝒎𝒆𝑰𝒏𝒅 Constructed as SameAuditor but for industry affiliation; industry is classified

using two-digit SIC codes (sic); see also Table 1, Panel C.

𝑺𝒂𝒎𝒆𝑩𝑬𝑨𝑹𝒆𝒈𝒊𝒐𝒏 Constructed as SameAuditor; equals 1 when all network partners are located in

the same BEA region in t1; 0 otherwise; the respective regions, as defined by the

Bureau of Economic Analysis, are Far West, Great Lakes, Mideast, New

England, Plains, Rocky Mountains, Southeast and Southwest.

𝑺𝒆𝒏𝒕𝒊𝒎𝒆𝒏𝒕 Measure of textual sentiment of the underlying 10-K filing, calculated as

𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡𝑖,𝑡 =(𝑛𝑝𝑜𝑠 − 𝑛𝑛𝑒𝑔𝑎𝑡𝑖𝑜𝑛) − 𝑛𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒

𝑛𝑡𝑜𝑡𝑎𝑙

See Textual Sentiment for data origin.

𝑺𝒕𝒓𝒂𝒕𝒆𝒈𝒊𝒄 𝑨𝒍𝒍𝒊𝒂𝒏𝒄𝒆 (𝑵𝒆𝒕𝒘𝒐𝒓𝒌) Contractual-based cooperation between US firms in the sample period 1994 to

2016, extracted from SDC Platinum (STRATEGICALLIANCE/SAF). All

requested data items are available in the Online Supplement.

𝑻𝒆𝒙𝒕𝒖𝒂𝒍 𝑺𝒆𝒏𝒕𝒊𝒎𝒆𝒏𝒕 Data on textual sentiment is shared publicly by Bill McDonald:

https://sraf.nd.edu/textual-analysis/resources/. We consider 10-K filings

(FORM_TYPE), positive, negative, total words (N_POSITIVE, N_NEGATIVE,

N_TOTAL), and negations (N_NEGATION).

𝒕𝒓𝒆𝒂𝒕𝒆𝒅 Treatment is set according to hightolow if strategic alliance falls into embargo

period (exclusion of overlapping events). Control observations (hightolow = 0)

are weighted to treatment observations (hightolow = 1) by entropy balancing

(weighting) on continuous firm controls (mean and variance) in the balanced

sample.

𝒕𝒓𝒆𝒂𝒕𝒆𝒅 ∗ 𝒑𝒐𝒔𝒕 Interaction of treated and post; main variable of interest in the difference-in-

differences model.

𝑼𝒔𝒆 𝒐𝒇 𝑵𝒆𝒈𝒂𝒕𝒊𝒗𝒆 𝑾𝒐𝒓𝒅𝒔 Measure of textual sentiment of the underlying 10-K filing, calculated as

𝑈𝑠𝑒 𝑜𝑓 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑊𝑜𝑟𝑑𝑠𝑖,𝑡 =𝑛𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒

𝑛𝑡𝑜𝑡𝑎𝑙

See Textual Sentiment for data origin.

𝑼𝒔𝒆 𝒐𝒇 𝑻𝒂𝒙 𝑯𝒂𝒗𝒆𝒏 Measure of material tax haven operations; Indicator variable, equals 1 if

∑ 𝑇𝐴𝑋𝐻𝐴𝑉𝐸𝑁𝑖,𝑡 > 0 (at least one tax haven subsidiary in Exhibit 21 in firm-

year); 0 otherwise; see Exhibit 21 for data origin.

Firm Controls*

𝑨𝒅𝑬𝒙𝒑𝟑 Advertising expense (xad) divided by net sales (sale); numerator and

denominator are constructed as the sum of the current and two subsequent years;

when missing reset to annual measure, thereafter reset to 0.

𝑪𝒂𝒑𝑬𝒙𝟑 Reported capital expenditures (capx) divided by gross property, plant, and

equipment (ppegt); numerator and denominator are constructed as the sum of the

current and two subsequent years; when missing reset to annual measure,

thereafter reset to 0.

𝑪𝒂𝒔𝒉𝟑 Cash and cash equivalents (che) divided by total assets (at); numerator and

denominator are constructed as the sum of the current and two subsequent years;

when missing reset to annual measure, thereafter reset to 0.

𝑪𝒉𝒂𝒏𝒈𝒆𝑺𝒂𝒍𝒆𝟑 Three-year annual average growth rate (geometric mean) of net sales (sale)

(√𝑠𝑎𝑙𝑒𝑡3/𝑠𝑎𝑙𝑒𝑡13 − 1); when missing reset to annual change, thereafter reset to

0.

34

Variable Definition [Compustat (low)/SDC (CAPITAL)/other (ITALICS) data items]

𝑬𝑩𝑰𝑻𝑫𝑨𝟑 Earnings before interest, taxes, depreciation and amortization (ebitda) scaled by

total assets (at); numerator and denominator are constructed as the sum of the

current and two subsequent years; when missing reset to annual measure,

thereafter reset to 0.

𝑰𝒏𝒕𝒂𝒏𝒈𝒊𝒃𝒍𝒆𝒔𝟑

The ratio of intangible assets (intan) to total assets (at); numerator and

denominator are constructed as the sum of the current and two subsequent years;

when missing reset to annual measure, thereafter reset to 0.

𝑳𝒆𝒗𝒆𝒓𝒂𝒈𝒆𝟑 The sum of long-term debt (dltt) and long-term debt in current liabilities (dlc)

divided by total assets (at); numerator and denominator are constructed as the

sum of the current and two subsequent years; when missing reset to annual

measure, thereafter reset to 0.

𝑴𝑵𝑬𝟑 Indicator variable; equals 1 if ∑ (3𝑡=1 𝑝𝑖𝑓𝑜𝑖,𝑡) > 0 (nonmissing, nonzero value for

pre-tax income from foreign operations); 0 otherwise.

𝑵𝑶𝑳𝟑 Indicator variable equals 1 if ∑ (3𝑡=1 𝑡𝑙𝑐𝑓𝑖,𝑡) > 0 (nonmissing, nonzero value of

tax loss carry forward); 0 otherwise; measured as the sum over three years.

𝑷𝑷𝑬𝟑 Gross property, plant, and equipment (ppegt) divided by total assets (at);

numerator and denominator are constructed as the sum of the current and two

subsequent years; when missing reset to annual measure, thereafter reset to 0.

𝑹𝒏𝑫𝑬𝒙𝒑𝟑 Research and development expenses (xrd) scaled by net sales (sale); numerator

and denominator are constructed as the sum of the current and two subsequent

years; when missing reset to annual measure, thereafter reset to 0.

𝑺𝑮𝑨𝟑 Selling, general, and administrative expense (xsga); divided by net sales (sale);

numerator and denominator are constructed as the sum of the current and two

subsequent years; when missing reset to annual measure, thereafter reset to 0.

𝑺𝒊𝒛𝒆𝟑 The natural log of total assets (at) for the respective and two subsequent periods;

when missing reset to annual measure, thereafter reset to 0.

*Continuous 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 are winsorized at p1 and p99 and not mean-centered.

Figures

Figure 1

This figure categorizes prior research regarding cross-firm connections and transfers of tax knowledge. It

aligns the framework (i.e., institutional setting) with identified channels.

Cross-Firm Connections and Tax Knowledge

Intermediary Peer-to-Peer

Intentional Transfer Diffusion

• Auditors [Frey (2018); Lim et al. (2018)]

• Banks [Gallemore, Gipper, Maydew (2019)]

• Board Ties [Brown (2011); Brown, Drake (2014)]

• Human Capital Turnover [Barrios, Gallemore (2019)]

Framework

Major

Channel

1 2

A B

1 A

2 A • Supply Chains [Cen et al. (2017); Cen et al. (2020)]

2 B • Strategic Alliances

Figure 2

This figure summarizes our sample selection identification strategy. The heartbeat pictogram at 𝑡1 indicates the year of initiation of a strategic

alliance.

t-2

industry adjusted

ranked according to quartiles

1 2 3 4

No Tax Knowledge Diffusion without a Sophisticated Tax Planner as Partner

main variable of interest:

Sample Selection & Identification Strategy

consider for analyses

Sample SelectionSDC Platinum (UPPER) and Compustat (lower) data items in parentheses

Network-

Firm Obs.Networks

Firm-

Year Obs.Firms

Compustat and SDC Platinum data matched (ULTPARENTCUSIP / cusip)

according to year of network initiation (DATEEFFECTIVE) // US firms only (loc,

fic, curcd, cik) // period 1994 - 2016 (fyear)

17,126 14,838 51,549 3,445

Identify all participants in a network 3,967 1,969 18,752 1,545

Identify and of all participants 1,092 543 7,114 502

t-1

t0

t1

t2

t3

201 627 198 66

1

Step #1

Step #2 Step #3

Step #4

Step #5

also onlyalso only

Figure 3

This figure maps networks and the investing firms from our sample (for networks of two partners). Each

vertex (square) displays a firm in its classification as low-tax (black vertices) or high-tax (gray vertices). A

link between two gray vertices translates into an ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 0 observation for both high-tax firms.

Correspondingly, a link between a gray vertex and a black vertex translates into an ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 = 1

observation for the high-tax firm. There are 324 firms each of which has one network observation in our

sample. In contrast, there are 178 firms with > 1 investments.

Figure 4

The word cloud depicted in Figure 3 shows the 50 most common words used in the SDC’s deal description

of the networks in our sample. By systematically searching through the deal descriptions, we identify

𝑤ℎ𝑜𝑙𝑒𝑠𝑎𝑙𝑒, 𝑙𝑖𝑐𝑒𝑛𝑠𝑖𝑛𝑔, 𝑚𝑎𝑟𝑘𝑒𝑡𝑖𝑛𝑔 and 𝑚𝑎𝑛𝑢𝑓𝑎𝑐𝑡𝑢𝑟𝑖𝑛𝑔 activities as well as

𝑟𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝑎𝑛𝑑 𝑑𝑒𝑣𝑒𝑙𝑜𝑝𝑚𝑒𝑛𝑡, providing 𝑠𝑒𝑟𝑣𝑖𝑐𝑒𝑠 and engaging in 𝑠𝑢𝑝𝑝𝑙𝑦 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛𝑠 as major business

purposes of the networks in our sample. The respective indicator variables are included in equation (3). All

variables are defined in detail in the Appendix.

Figure 5

Panel A

Panel A of this figure provides visual evidence that the trend of 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅 is similar for the treatment and

control firms prior to the treatment.

Panel B

In Panel B, we apply the approach of Patel & Seegert (2015) to provide statistical evidence for the parallel

trend assumption. The figure reports the coefficient and 95% confidence interval of the interaction of the

treatment indicator and embargo period fixed effects for pretreatment years. The p‑value for the parallel

trend test is reported at the bottom of Panel B.

Tables

Table 1 Information on Networks and Firms

Panel A Descriptive Statistics of Firm Controls [𝒕𝟏; 𝒕𝟑]

N 201 627 198 66

𝒉𝒊𝒈𝒉𝒕𝒐𝒍𝒐𝒘 == 1 == 0

𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ == 1 == 0

mean p50 mean p50 mean p50 mean p50

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 0.2557 0.2341 0.2804 0.2449 0.1557 0.1359 0.1631 0.1475

𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 0.9589 0.8995 1.0871 0.9554 2.1243 1.3715 2.8437 1.5392

𝐸𝐵𝐼𝑇𝐷𝐴3 0.1551 0.1504 0.1606 0.1517 0.1321 0.1255 0.1273 0.1279

𝑅𝑛𝐷𝐸𝑥𝑝3 0.0741 0.0517 0.0677 0.0475 0.0965 0.0834 0.1193 0.1208

𝐴𝑑𝐸𝑥𝑝3 0.0172 0.0036 0.0145 0.0000 0.0112 0.0000 0.0132 0.0018

𝑆𝐺𝐴3 0.2671 0.2720 0.2719 0.2613 0.3099 0.3052 0.3847 0.3704

𝐶𝑎𝑝𝐸𝑥3 0.1300 0.1159 0.1387 0.1170 0.1498 0.1382 0.1706 0.1412

𝐶ℎ𝑎𝑛𝑔𝑒𝑆𝑎𝑙𝑒3 0.0386 0.0275 0.0484 0.0267 0.0572 0.0484 0.0587 0.0447

𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒3 0.1911 0.1977 0.1932 0.1914 0.1781 0.1318 0.1674 0.1531

𝐶𝑎𝑠ℎ3 0.1797 0.1254 0.1725 0.1154 0.2448 0.2179 0.2901 0.2460

𝑀𝑁𝐸3 0.6169 1.0000 0.6411 1.0000 0.6667 1.0000 0.6818 1.0000

𝑁𝑂𝐿3 0.2786 0.0000 0.2823 0.0000 0.3889 0.0000 0.2879 0.0000

𝐼𝑛𝑡𝑎𝑛𝑔𝑖𝑏𝑙𝑒𝑠3 0.1909 0.1577 0.1765 0.1067 0.1509 0.0864 0.1251 0.0698

𝑃𝑃𝐸3 0.3958 0.3112 0.4141 0.3244 0.3165 0.2194 0.2947 0.2326

𝑆𝑖𝑧𝑒3 9.8628 10.6128 9.6368 10.0497 9.3460 9.5647 9.2779 9.0031

Panel A of Table 1 shows descriptive statistics for (𝑑𝑒𝑙𝑡𝑎) 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 and 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 in 𝑡1 (year of

network initiation). The 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 are constructed as multiperiod measures [𝑡1; 𝑡3]. All variables are

defined in detail in the Appendix.

Table 1 Information on Networks and Firms (continued)

Panel B Descriptive Statistics of Partner and Network Controls

𝑷𝒂𝒓𝒕𝒏𝒆𝒓 𝑪𝒐𝒏𝒕𝒓𝒐𝒍𝒔 𝑵𝒆𝒕𝒘𝒐𝒓𝒌 𝑪𝒐𝒏𝒕𝒓𝒐𝒍𝒔 (Purpose_)

n (mean) 𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 𝑆𝑎𝑚𝑒𝐼𝑛𝑑 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦 𝑊ℎ𝑜𝑙𝑒𝑠𝑎𝑙𝑒 𝑅𝑛𝐷 𝐿𝑖𝑐𝑒𝑛𝑠𝑖𝑛𝑔 𝑆𝑒𝑟𝑣𝑖𝑐𝑒 𝑀𝑎𝑟𝑘𝑒𝑡𝑖𝑛𝑔 𝑆𝑢𝑝𝑝𝑙𝑦 𝐶ℎ𝑎𝑖𝑛 𝑀𝑎𝑛𝑢𝑓𝑎𝑐𝑡𝑢𝑟𝑖𝑛𝑔

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤

== 1

44

(0.2189)

81

(0.4030)

45

(0.2239) (0.0253)

24

(0.1194)

52

(0.2587)

30

(0.1493)

86

(0.4279)

15

(0.0746)

15

(0.0746)

31

(0.1542)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤

== 0

122

(0.1946)

260

(0.4147)

130

(0.2073) (0.0255)

30

(0.0478)

183

(0.2919)

114

(0.1818)

270

(0.4306)

62

(0.0989)

42

(0.0670)

46

(0.0734)

𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ

== 1

44

(0.2222)

81

(0.4091)

46

(0.2323) (0.0262)

24

(0.1212)

51

(0.2576)

30

(0.1515)

84

(0.4242)

15

(0.0758)

14

(0.0707)

30

(0.1515)

𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ

== 0

18

(0.2727)

38

(0.5758)

28

(0.4242) (0.0496)

4

(0.0606)

16

(0.2424)

6

(0.0909)

40

(0.6061)

10

(0.1515)

0

(0)

6

(0.0909)

Panel C Industry Affiliation of Networks and Firms [two-digit SIC-code]

Industry of Networks (Network-Firm Observations) ∑ I II III IV V VI VII VIII IX X XI XII

Ind

ust

ry

of

Fir

ms

Agriculture, Forestry, & Fishing [01-09] I 0 0 0 1 0 1 0 0 0 0 1 0 3

Mining [10-14] II 0 6 0 1 2 0 2 0 0 0 1 0 12

Construction [15-17] III 0 0 0 0 0 3 0 0 0 0 1 0 4

Manufacturing: Chemical & Allied Products [28] IV 0 0 0 23 10 0 19 0 13 5 32 0 102

Manufacturing [20-39, except 28] V 0 4 2 12 110 16 30 3 42 109 43 7 378

Transportation & Public Utilities [40-49] VI 0 3 2 0 3 28 3 0 2 29 9 2 81

Wholesale Trade [50-51] VII 0 1 0 0 2 2 15 1 1 7 5 0 34

Retail Trade [52-59] VIII 0 0 0 0 4 3 9 1 5 10 3 2 37

Finance, Insurance, & Real Estate [60-67] IX 0 0 0 0 1 0 2 0 19 18 5 3 48

Services: Business Services [73] X 0 0 0 0 25 8 9 3 25 231 24 2 327

Services [70-89, except 73] XI 0 0 0 0 1 2 2 0 2 7 14 0 28

Nonclassifiable Establishments/Other XII 0 0 0 3 4 5 1 0 3 15 7 0 38

0 14 4 40 162 68 92 8 112 431 145 16 1092

Panel B of Table 1 shows the distribution of the 𝑝𝑎𝑟𝑡𝑛𝑒𝑟 & 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 variables among ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 and 𝑙𝑜𝑤𝑡𝑜ℎ𝑖𝑔ℎ observations. Panel C of Table 1 presents information regarding the industry

affiliation of networks and firms. Industry affiliation is determined on a two-digit SIC-code basis. All variables are defined in detail in the Appendix.

Table 2 Descriptive Analysis

Panel A Change from 𝒑𝒓𝒆 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕−𝟐; 𝒕𝟎] to 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑]

N 201 627

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 == 1 == 0

mean (SD) mean (SD)

𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] I 0.3039 (0.1789) 0.2856 (0.1341)

𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡1; 𝑡3] II 0.2557 (0.1874) 0.2804 (0.1936)

Within-Group Change I to II - 0.0481 *** - 0.0051

(p-value) (0.0039) (0.5418)

Difference in Within-Group Change - 0.0430 **

(p-value) (0.0143)

Panel B Differences in 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 ==1 ==0 difference (p-value)

- mean 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 0.9589 1.0871 - 0.1282 ** (0.0422)

- p50 𝑑𝑒𝑙𝑡𝑎 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 0.8995 0.9554 - 0.0560 * (0.0935)

All variables are defined in detail in the Appendix. Superscripts ***, ** and * indicate significance at the

1%, 5% and 10% levels, respectively, for two-tailed tests. We follow Conroy (2012) and apply quantile

regression to test for the difference in medians in Panel B. The results of this test are robust to using

(i) Mann-Whitney/Wilcoxon rank-sum and (ii) K-sample equality-of-medians tests (untabulated).

Table 3 Main Analysis

Panel A Regression Analysis

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑] 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

Coefficient (p-value) Coefficient (p-value)

𝒉𝒊𝒈𝒉𝒕𝒐𝒍𝒐𝒘 - 0.0273 ** (0.0470) - 0.1334 ** (0.0121)

𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦 - 0.0538 (0.2479) - 0.2317 (0.1663)

𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 - 0.0049 (0.7472) 0.0389 (0.5129)

Network Controls Yes Yes

Firm Controls Yes Yes

Fixed Effects Industry & Year Industry & Year

SE Cluster @ Firm Cluster @ Firm

N 828 828

Adjusted 𝑅2 0.1230 0.1210

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed

tests. The results for equation (3) are presented in Panel A. Our main variable of interest is ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤,

which is an indicator variable set equal to 1 for observations of high-tax firms cooperating with low-tax

firms and set equal to 0 for high-tax firms cooperating with high-tax firms. All variables are defined in

detail in the Appendix.

Table 3 Main Analysis (continued)

Panel B Difference in Differences

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹

Embargo Period Yes [𝑡−2; 𝑡5] Yes [𝑡−2; 𝑡5]

Entropy Balancing - Balanced Sample

Coefficient (p-value) Coefficient (p-value)

𝑡𝑟𝑒𝑎𝑡𝑒𝑑 0.0189 (0.2957) 0.0228 (0.2788)

𝒕𝒓𝒆𝒂𝒕𝒆𝒅 ∗ 𝒑𝒐𝒔𝒕 - 0.0411 * (0.0855) - 0.0555 ** (0.0430)

Firm Controls Yes (Annual Measures) Yes (Annual Measures)

Fixed Effects Industry & Year & Embargo

Period

Industry & Year & Embargo

Period

SE Cluster @ Firm Cluster @ Firm

N 1529 1529

Adjusted 𝑅2 0.0412 0.0596

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed

tests. Panel B depicts the results for equation (4). We create an embargo period around a network

observation during which a firm may not invest in another network (exclusion of overlapping events). The

embargo period contains the three years preceding (𝑝𝑜𝑠𝑡 = 0) and five years subsequent (𝑝𝑜𝑠𝑡 = 1) to a

network initiation. Including 𝑒𝑚𝑏𝑎𝑟𝑔𝑜 𝑝𝑒𝑟𝑖𝑜𝑑 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠 subsumes the 𝑝𝑜𝑠𝑡 indicator. We also

employ entropy-balancing weighting and use the entropy weights to reestimate equation (4). Observations

are balanced using continuous 𝑓𝑖𝑟𝑚 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 so that the means and the variance in the reweighted control

group match the treatment group (balanced sample). All variables are defined in detail in the Appendix.

Table 3 Main Analysis (continued)

Panel C Adjustment Speed

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹

Embargo Period Yes [𝑡−2; 𝑡5]

Entropy Balancing Balanced Sample

(#) of Specification Coefficient (p-value)

(1) 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 [𝑡1] - 0.0459 (0.1954)

(2) 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 [𝑡1; 𝑡2] - 0.0588 * (0.0615)

(3) 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 [𝑡1; 𝑡3] - 0.0522 * (0.0668)

(4) 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 [𝑡1; 𝑡4] - 0.0571 ** (0.0443)

(5) 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑝𝑜𝑠𝑡 [𝑡1; 𝑡5] - 0.0555 ** (0.0430)

Controls Firm Controls & Treated

Fixed Effects Industry & Year & Embargo Period

SE Cluster @ Firm

N 899; 1076; 1241; 1396; 1529

Adjusted 𝑅2 0.0624; 0.0660; 0.0707; 0.0624; 0.0596

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed

tests. Panel C depicts the results for five specifications of equation (4). The posttreatment period is extended

by one year for each specification (from 𝑝𝑜𝑠𝑡 = 1 𝑓𝑜𝑟 [𝑡1] to 𝑝𝑜𝑠𝑡 = 1 𝑓𝑜𝑟 [𝑡1; 𝑡5]). 𝑝𝑜𝑠𝑡 equals 0 for 𝑡−2

to 𝑡0 throughout all specifications. The coefficient estimates of 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 ∗ 𝑝𝑜𝑠𝑡 comprise the cumulative

adjustment (i.e., adjustment speed) of a high-tax firm’s tax planning behavior with progressing time when

cooperating with low-tax firms. All variables are defined in detail in the Appendix.

Table 4 Additional Analyses: Effects on Reporting of Operations

Textual Sentiment of 10-K Filings & Tax Haven Operations

Dependent Variable 𝑺𝒆𝒏𝒕𝒊𝒎𝒆𝒏𝒕 𝑼𝒔𝒆 𝒐𝒇

𝑵𝒆𝒈𝒂𝒕𝒊𝒗𝒆 𝑾𝒐𝒓𝒅𝒔 𝑼𝒔𝒆 𝒐𝒇 𝑻𝒂𝒙 𝑯𝒂𝒗𝒆𝒏

𝑵𝒖𝒎 𝒐𝒇 𝑻𝒂𝒙 𝑯𝒂𝒗𝒆𝒏 𝑺𝒖𝒃𝒔𝒊𝒅𝒊𝒂𝒓𝒊𝒆𝒔

Embargo Period Yes [𝑡−2; 𝑡5] Yes [𝑡−2; 𝑡5] Yes [𝑡−2; 𝑡5] Yes [𝑡−2; 𝑡5]

Entropy Balancing - - - -

Coefficient (p-value) Coefficient (p-value) Coefficient (p-value) Coefficient (p-value)

𝑡𝑟𝑒𝑎𝑡𝑒𝑑 0.0013 * (0.0923) - 0.0008 (0.1746) 0.0288 (0.6723) 0.0283 (0.7994)

𝒕𝒓𝒆𝒂𝒕𝒆𝒅 ∗ 𝒑𝒐𝒔𝒕 - 0.0015 ** (0.0174) 0.0012 ** (0.0126) - 0.1245 * (0.0500) - 0.1558 * (0.0884)

Firm Controls Yes (Annual Measures) Yes (Annual Measures) Yes (Annual Measures) Yes (Annual Measures)

Fixed Effects Industry & Year &

Embargo Period

Industry & Year &

Embargo Period

Industry & Year &

Embargo Period

Industry & Year &

Embargo Period

SE Cluster @ Firm Cluster @ Firm Cluster @ Firm Cluster @ Firm

N 1302 1302 1333 1333

Adjusted 𝑅2 0.1801 0.2598 0.2927 0.3646

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed tests. Table 4 depicts the results for

equation (4) with measures of textual sentiment of 10-K filings (𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡, 𝑈𝑠𝑒 𝑜𝑓 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑊𝑜𝑟𝑑𝑠) and Tax Haven Operations

(𝑈𝑠𝑒 𝑜𝑓 𝑇𝑎𝑥 𝐻𝑎𝑣𝑒𝑛, 𝑁𝑢𝑚 𝑜𝑓 𝑇𝑎𝑥 𝐻𝑎𝑣𝑒𝑛 𝑆𝑢𝑏𝑠𝑖𝑑𝑖𝑎𝑟𝑖𝑒𝑠) from Exhibit 21 disclosures as dependent variables. Data on textual sentiment

(Exhibit 21 disclosures) are shared publicly by Bill McDonald (Scott Dyreng). All variables are defined in detail in the Appendix.

Table 5 Additional Analyses: Partner Characteristics

Panel A Distance

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑] 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

Coefficient (p-value) Coefficient (p-value)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 - 0.0276 * (0.0868)

- 0.1156 * (0.0569)

𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 - 0.0145 (0.4786)

- 0.0411 (0.6504)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 ∗ 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 0.0008 (0.9825)

- 0.0856 (0.5526)

Controls Partner & Network & Firm Partner & Network & Firm

Fixed Effects Industry & Year Industry & Year

SE Cluster @ Firm Cluster @ Firm

N 828 828

Adjusted 𝑅2 0.1215 0.1199

Panel B Industry

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑] 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

Coefficient (p-value) Coefficient (p-value)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 - 0.0425 ** (0.0233)

- 0.1331 * (0.0900)

𝑆𝑎𝑚𝑒𝐼𝑛𝑑 - 0.0226 (0.1679)

- 0.0602 (0.3897)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 ∗ 𝑆𝑎𝑚𝑒𝐼𝑛𝑑 0.0358 (0.2697)

0.0030 (0.9791)

Controls Partner & Network & Firm Partner & Network & Firm

Fixed Effects Year Year

SE Cluster @ Firm Cluster @ Firm

N 828 828

Adjusted 𝑅2 0.1286 0.1212

Panel C Auditor

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑] 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

Coefficient (p-value) Coefficient (p-value)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 - 0.0353 ** (0.0330)

- 0.1710 *** (0.0074)

𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 - 0.0151 (0.3791)

- 0.0094 (0.8930)

ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 ∗ 𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 0.0377 (0.2683)

0.1789 (0.1811)

Controls Partner & Network & Firm Partner & Network & Firm

Fixed Effects Industry & Year Industry & Year

SE Cluster @ Firm Cluster @ Firm

N 828 828

Adjusted 𝑅2 0.1227 0.1212

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed tests. In Panel A, we test for the impact

of geographical distance between the headquarters of cooperating firms. Distance is measured by an indicator variable that is set equal to one if network partners are headquartered in the same region, as defined by the Bureau of Economic Analysis, at network initiation. Panel B follows the

approach of Panel A for the industry affiliation of network partners. We interact ℎ𝑖𝑔ℎ𝑡𝑜𝑙𝑜𝑤 with 𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 in Panel C. All variables are

defined in detail in the Appendix.

Table 6 Robustness Checks: Alternative Explanations

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed

tests. All variables are defined in detail in the Appendix.

Table 7 Robustness Checks: Alternative Identification Strategy

Interact Indicator with Continuous Variable

Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕𝟏; 𝒕𝟑] Coefficient (p-value)

ℎ𝑖𝑔ℎ‑𝑡𝑎𝑥 𝑓𝑖𝑟𝑚 0.0883 *** (0.0001)

𝑝𝑎𝑟𝑡𝑛𝑒𝑟 𝑝𝑟𝑒 𝑐𝑎𝑠ℎ 𝐸𝑇𝑅3 [𝑡−2; 𝑡0] - 0.0298 (0.5212)

𝒉𝒊𝒈𝒉‑𝒕𝒂𝒙 𝒇𝒊𝒓𝒎 ∗ 𝒑𝒂𝒓𝒕𝒏𝒆𝒓 𝒑𝒓𝒆 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 [𝒕−𝟐; 𝒕𝟎] 0.1301 * (0.0626)

𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦 - 0.0277 (0.4596)

𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 - 0.0075 (0.5402)

Network Controls No

Firm Controls Yes

Fixed Effects Industry & Network-Industry

SE Cluster @ Firm

N 1092

Adjusted 𝑅2 0.1646

Superscripts ***, ** and * indicate significance at the 1%, 5% and 10% levels, respectively, for two-tailed

tests. All variables are defined in detail in the Appendix.

Panel A Exclude Firm-Edge-Years Panel B Effect on low-tax Firms

Dependent Variable 𝒅𝒆𝒍𝒕𝒂 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑 Dependent Variable 𝒄𝒂𝒔𝒉 𝑬𝑻𝑹𝟑

Specification exclude nonsurvivors Specification only low-tax firms

Coefficient (p-value) Coefficient (p-value)

𝒉𝒊𝒈𝒉𝒕𝒐𝒍𝒐𝒘 - 0.1369 ** (0.0162)

𝒍𝒐𝒘𝒕𝒐𝒉𝒊𝒈𝒉 - 0.0207 (0.3298)

𝑃𝑟𝑜𝑥𝑖𝑚𝑖𝑡𝑦 - 0.2464 (0.1667) 𝑆𝑎𝑚𝑒𝐵𝐸𝐴𝑅𝑒𝑔𝑖𝑜𝑛 - 0.0003 (0.9909)

𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 0.0416 (0.5283) 𝑆𝑎𝑚𝑒𝐴𝑢𝑑𝑖𝑡𝑜𝑟 - 0.0067 (0.7529)

Network Controls Yes Network Controls Yes

Firm Controls Yes Firm Controls Yes

Fixed Effects Industry & Year Fixed Effects Industry & Year

SE Cluster @ Firm SE Cluster @ Firm

N 726 N 261

Adjusted 𝑅2 0.1680 Adjusted 𝑅2 0.0824

Online Supplement

R and Stata are used for the analyses. The do-files are available online at a GitHub repository.

• Version: 08/18/2020

• User: taxknowledge

• Repository: diffusionviastrategicalliances

• URL: https://github.com/taxknowledge/diffusionviastrategicalliances

The tables in this manuscript are reduced to the variables of interest. Therefore, the repository contains a

log-file of the analyses.

Impressum: Arbeitskreis Quantitative Steuerlehre, arqus, e.V. Vorstand: Prof. Dr. Ralf Maiterth (Vorsitzender), Prof. Dr. Kay Blaufus, Prof. Dr. Dr. Andreas Löffler Sitz des Vereins: Berlin Herausgeber: Kay Blaufus, Jochen Hundsdoerfer, Martin Jacob, Dirk Kiesewetter, Rolf J. König, Lutz Kruschwitz, Andreas Löffler, Ralf Maiterth, Heiko Müller, Jens Müller, Rainer Niemann, Deborah Schanz, Sebastian Schanz, Caren Sureth-Sloane, Corinna Treisch Kontaktadresse: Prof. Dr. Caren Sureth-Sloane, Universität Paderborn, Fakultät für Wirtschaftswissenschaften, Warburger Str. 100, 33098 Paderborn, www.arqus.info, Email: info@arqus.info

ISSN 1861-8944