The effect of tax-motivated income shifting on information asymmetry

The jurisdiction where firms report income for tax purposes is not necessarily the jurisdiction where the income is earned. We view the true location of income as being based on where the economic earnings activities took place. Income shifting, then, is the extent to which income is reported elsewhere, due to lower tax rates.

Collins et al. (1998) provide evidence that investors, on average, at least partially recognize the effect of income shifting on firm value. Even if investors understand the amount of shifting that occurs on average, however, it is unclear whether they will understand the amount of shifting that occurs at any given firm. Insofar as investors can ascertain the extent of shifting at a sufficiently low cost, we will fail to find support for a relation between tax-motivated income shifting and a firm’s external information asymmetry.

Similar to Rego and Wilson (2012), we use two-stage least squares to estimate our simultaneous system of equations, where proxies for tax-motivated income shifting and external information asymmetry are the endogenous, dependent variables. See Section 3 for more details.

Assuming managers make decisions to maximize firm value, an alternative interpretation of this result is that our estimates provide a lower bound of the benefits of tax-motivated income shifting.

Since 2013, the United States has the highest corporate tax rate among OECD countries. (See Part C, Table II.1, of the OECD Tax Database, available at http://www.oecd.org/tax/tax-policy/tax-database.htm .)

Our study differs from the work of Gallemore and Labro (2015) and McGuire et al. (2017) in that they focus on internal information quality as a determinant of tax avoidance and tax-motivated income shifting, respectively, whereas we consider the consequences of tax-motivated income shifting on the firm’s external information asymmetry.

Our treatment of loss firms is comparable to research by Collins et al. (1998) and Klassen and Laplante (2012). Collins et al. (1998, p. 216) “exclude observations with negative pretax domestic or foreign income,” and Klassen and Laplante (2012, p. 1262) “exclude firm years with negative five-year summed pretax domestic or foreign income (loss firms) because their income shifting incentives are more difficult to reliably estimate.”

We use four-digit GICS (or GICS groups) as our industry definition. The classifications have the advantage of being revenue stream-based rather than product-based (e.g., SIC and NAICS schemes). Historical GICS are broadly available from 1985 through the present, which covers our entire sample period. Results are robust to using two-digit SIC, three-digit NAICS, and Fama-French 48 industries to estimate industry membership fixed effects (untabulated).

For each three-year period, there are 36 trading months and approximately 753 trading days. Hence 32 trading months and 670 trading days represent roughly equivalent fractions of the respective totals.

Our results are qualitatively unchanged throughout when we (i) limit the analyses to the final quarter of each firm-year, (ii) include fiscal quarter fixed effects, or (iii) include calendar quarter fixed effects.

Our sample extends from 1995 to 2012, and we thus use 35 percent as the U.S. statutory rate. Beginning in 1993, the U.S. statutory rate for corporations with taxable income of at least $18,333,333 became 35 percent (https://​www.​irs.​gov/​pub/​irs-soi/​02corate.​pdf). Although statutory rates other than 35 percent exist for corporations with taxable income below $18,333,333, the statutory rates for lower levels of income rarely apply to the MNCs in our sample. The MNCs in our sample are profitable (only 5.2 percent of firm-years report a loss) and have a mean market capitalization of $2.6 billion.

We exclude the FTR_AVE main effect, because it is a linear combination of the firm-specific FTR_AVE interactions.

As in the work of Collins et al. (1998), a negative (positive) coefficient suggests the effects of income shifting dominate (are dominated by) implicit taxes.

We conduct sensitivity analyses requiring that each α_3i is statistically significantly less than zero at the 5 percent and 10 percent levels to be treated as evidence of income shifting (i.e., no evidence of shifting if insignificant or significantly positive). Each of these sensitivity checks (untabulated) produces qualitatively similar results to our reported results.

As do Klassen and Laplante (2012), we originally include a fourth lagged value of FTR and the ratio of foreign sales divided by worldwide sales as instruments, but based on t-statistics and an over-identifying restrictions test, we eliminate these two instruments and retain the more parsimonious equation (4).

Bid-ask spread is affected by the stock’s market microstructure, such as exchange rules, trading activity, execution costs, and dealer borrowing costs needed to support inventory positions (Lee and Masulis 2009). Bid-ask spread is also used as a measure of market liquidity in the literature (e.g., Daske et al. 2008). Therefore we use the adverse-selection component of the bid-ask spread (LAMBDA) to better capture information asymmetry. However, we acknowledge that LAMBDA may be difficult to estimate accurately in the presence of algorithmic trading.

We also examine the absolute value of analysts’ forecast errors. The results (untabulated) remain both statistically and economically significant when we use this alternative measure.

As a robustness check, we replace TA_GAAP with TA_CASH (industry-size-adjusted cash ETR), and the results are qualitatively unaffected.

Consistent with the approach of Hope et al. (2013), we verify that the firms in our sample disclose geographic earnings in the pre-SFAS 131 period.

We also conduct our main information asymmetry tests using outbound and inbound shifting measures separately. Results (untabulated) are consistent with H1 that both forms of income shifting increasing firms’ information asymmetry. The results are relatively stronger both in terms of significance level and economic significance for outbound shifting, consistent with outbound shifting being more prevalent than inbound shifting over our sample. Additionally, within inbound shifting only, our evidence for income shifting affecting information asymmetry is stronger for the first half of our sample, consistent with inbound shifting being more common in the early years of our sample.

As a robustness check, we repeat our main analyses using the income shifting model of Dyreng and Markle (2016). The results (untabulated) are qualitatively consistent with our main findings.

We estimate the average cost to traders by taking α₁ as estimated from equation (6) times the standard deviation of predicted SHIFT_AVE and predicted SHIFT_IV, respectively. We divide by 100 to convert percentage to decimals. We then multiply that by the average monthly volume and average price per share for our sample. We multiply by 12 to annualize our estimate of the average cost to traders.

In contrast to Daske et al. (2008), we examine the adverse selection component of the bid-ask spreads, rather than percentage bid-ask spreads. When using percentage bid-ask spreads as our information asymmetry proxy and a similar set of control variables as Daske et al. (2008), we find statistically significant results. The economic magnitude of these results are slightly smaller than the results we find using the adverse selection component bid-ask spread, which may be attributable to percentage bid-ask spreads being a noisier measure of adverse selection.

We use different measurement windows in our income shifting and information asymmetry proxies. The income shifting measures require five years of data, while we use three years of data to calculate LAMBDA and IDVOL. As a robustness check, we measure income shifting, information asymmetry, and control variables all using four and five years of data. The results (untabulated) are qualitatively similar under these alternative measurement windows.

Results are qualitatively similar when we allow the sample period to span years t–5 through t+5.

To have a measure of income shifting for all firms, including those that discontinue disclosure, we take the average of SHIFT_AVE (SHIFT_IV) over the pre-adoption period. Results are qualitatively unaffected when we instead take the value at t–1, t–3, or the largest value over the period t–3 through t–1.

We check whether the firm continues to list at least one significant tax haven subsidiary in years t+1 through t+3. If the firm does not, we remove that observation from this analysis. Some firms remove tax havens outside of this window. A removal could represent a divestiture or a liquidation of the tax haven subsidiary. Alternatively, it could represent a strategic reporting decision. We include a control variable MULTI that is an indicator if a firm (i) reports a tax haven entry, (ii) continues to report the subsidiary throughout the five-year window from entry year t to year t+3, (iii) at some subsequent year no longer reports a tax haven subsidiary, and (iv) then reports a new tax haven subsidiary entry. MULTI takes a value of one for 17 percent of the firms in the haven entry sample. Our results are robust to restricting the tests to firms where MULTI is zero. In additional sensitivity analysis, we find our inferences are unchanged when considering various windows (i.e., one year, two years, four years, and five years) around the initiation year.

In addition to the tax haven initiation tests, we also use tax haven counts as an alternative proxy of income shifting in our levels analyses. The results (untabulated) remain both statistically and economically significant when we use this alternative income shifting measure.

Initiating a tax haven requires beginning operations in a new country, which may increase organizational complexity and consequently information asymmetry. To examine this alternative explanation, we conduct a counterfactual analysis. In this analysis, we set the POST variable in Table 7 equal to one for a firm’s initial entry into a nonhaven country. Inconsistent with the alternative explanation, the POST variable is not statistically different from zero in any of the four regressions.