Overconfidence and Corporate Tax Policy

See Rego and Wilson (2012) and Gallemore et al. (2014).

Malmendier and Tate (2005) test the validity of their measure by comparing the returns from unexercised in-the-money stock options to hypothetical option exercises coupled with an investment in the S&P 500 index. They find that investment in the index produces higher returns more often than the unexercised option strategy. Specifically, investing the proceeds from exercised in the money options in the S&P 500 index would beat the strategy of holding exercisable in-the-money options 54.14% of the time.

In sensitivity tests, we relax this assumption and find similar results (see Section 4.2).

The tax avoidance measures in our study are consistent with prior literature and include the cash effective tax rate, estimated tax shelter probability (Wilson 2009; Rego and Wilson 2012), and residual book-tax differences not attributable to accruals management (Desai and Dharmpala 2006).

See for example Malmendier and Tate (2005, 2008); Hirshleifer et al. (2012); Schrand and Zechman (2012); Ahmed and Duellman (2013); and Ben-David et al. (2013).

For example, if firm A experiences an endogenous turnover event in 1996 and an exogenous turnover event in 2001, then we delete observations for firm A prior to 1997.

To reduce the impact of outliers, we winsorize all continuous variables at the 1st and 99th percentiles.

The limited coverage of Execucomp, relative to Compustat, has a pronounced effect in reducing our sample from the 824 firms identified by Fee et al. (2013) to the 135 firms that are usable for our study.

We obtain similar results for CASH ETR after subtracting special items (SPI) from pretax book income, dropping firms with negative pretax income, or both. We omit the year after turnover from our CASH ETR analysis to avoid commingling cash tax payments across CEO regimes. Turnover years are excluded from all analyses.

In estimating TAX SHELTER SCORE, BTD is pretax book income (#PI) less estimated taxable income scaled by total assets (#AT), where estimated taxable income is (current federal tax expense, #TXFED, plus current foreign tax expense, #TXFO)/0.35, less the change in tax loss carryforwards, #TLCF. Leverage is total debt (#DLTT+#DLC) scaled by total assets (#AT). ROA is pretax book income (#PI) scaled by total assets (#AT). ForeignIncome is foreign pretax income (#PIFO) scaled by total assets (#AT); and 0 if #PIFO is missing. Finally, R&D is research and development expenses (#XRD) scaled by total assets (#AT); and 0 if #XRD is missing.

See for example Malmendier and Tate (2005, 2008); Campbell et al. (2011); Hirshleifer et al. (2012); Schrand and Zechman (2012); Ahmed and Duellman (2013); Hribar and Yang (2016).

Research suggests that our overconfidence measure is distinguishable from risk seeking (Malmendier and Tate 2005; Ben-David et al. 2013). Executives’ risk seeking would predict overinvestment in high-risk, high-return assets, but it would not predict overinvestment in one’s own firm. This is because the better-than-average effect combined with miscalibration would lead overconfident executives to underestimate the risk-return profile of their own firm to a greater extent than external investment opportunities. Hence, if overconfident managers were risk seekers, they would be more likely to overinvest in external assets, rather than in their own firm.

Our results are robust to variations on our overconfidence measure as well as wholly different measures of overconfidence. See Sections 4.2 and 4.3.

As Dyreng et al. (2010) point out, the use of firm fixed effects constrains our tests to only consider variation within the firm. Thus, if a firm always experiences lower tax rates than another firm because it operates in lower-taxed jurisdictions, this effect will be captured in its fixed effect.

As evident in correlations reported in Table 4, consistent with prior research, LEVERAGE is negatively and significantly correlated with OVERCONFIDENCE, and R&D is positively and significantly correlated with OVERCONFIDENCE (Hirshleifer et al. 2012). However, none of the correlation coefficients is high enough to raise concerns regarding multicollinearity.

Both time trends in tax avoidance and the somewhat coarse nature of our overconfidence measure could at least partially explain the increase in corporate tax avoidance we document when another overconfident CEO replaces an overconfident CEO.

In an untabulated analysis, we examine the association between CEO overconfidence and the three-year standard deviation of cash ETR (Guenther et al. 2017). The results from these tests suggest that overconfident CEOs do not suffer from more volatile cash payments. We interpret these results with caution, because many of the turnovers in our sample do not have a long enough period afterward to draw meaningful conclusions, regarding whether tax avoidance initiated by an overconfident CEO will result in more volatile cash ETRs in the long run.

In an untabulated analysis, we find a low (i.e. -0.06) correlation coefficient between raw CEO narcissism (NarcScore per Olsen and Stekelberg 2016) and CEO overconfidence. Since we lack narcissism data for over 25% of the sample, we code missing observations to 0 and include the MISSING_NARCISSISM variable to pick up the average effect of narcissism when it is missing.

For example, consider executive Z who works for two different firms, FIRM1 and FIRM2. We code OVERCONFIDENCE as 1 for firms employing executive Z if she demonstrates overconfidence at any point in the sample. The key reason for this is that it allows OVERCONFIDENCE to only vary with executive Z coming in or out of the firm. If we coded OVERCONFIDENCE only after executive Z demonstrates overconfidence, we could have cases where OVERCONFIDENCE varies within executive Z’s tenure in the firm (i.e., no variation in OVERCONFIDENCE due to turnover). Consider the case where the executive exhibits overconfidence in the third year of her tenure with FIRM1. Waiting until the third year to code her as overconfident would create variation in OVERCONFIDENCE within FIRM1, despite no change in CEO (i.e., in the absence of an exogenous shock).

Factor_5 captures the common variation in OVERCONFIDENCE, Net Purchase, OC_Firm5, Over-Invest_1, and Over-Invest_2.

There are at least two reasons why forecast error is likely a noisier proxy for executive overconfidence than the options-based proxies. First, all executives, regardless of their confidence, face a substantial probability of missing a forecast. In contrast, there is not a strong reason to believe that less-confident executives would have a strong inclination to leave in-the-money options unexercised. Hence the forecast-based measure might be less discriminating than the options-based measure. Second, overconfident executives might be willing to engage in more aggressive accounting, operating decision, or both (Schrand and Zechman 2012) to avoid missing forecasts, which would manifest in a lower likelihood of missing a forecast. This propensity to take aggressive actions might counter the propensity to make more optimistic forecasts, also making the forecast-based measure less discriminating.

While our paper relates to the work of Dyreng et al. (2010), we have different research questions, research designs, samples, and overconfidence proxies.

Armstrong et al. (2012) and Rego and Wilson (2012) suggest that tax directors are more directly related to the tax function than CFOs. Nevertheless, Rego and Wilson (2012) assert that the role of the CFO in financial reporting and in overseeing the maximization of after-tax cash flows suggests they could be important in setting corporate tax policy. Chyz and Gaertner (2017) present evidence that both CEOs and CFOs appear to be held accountable for tax outcomes. Although at least partially driven by a lack of endogenous turnover data for CFOs, their findings are considerably weaker for CFOs.

OC_CEO is equivalent to OVERCONFIDENCE from our primary analyses. We merely change the variable name to make it easier for readers to interpret CEO and CFO effects separately. OC_CFO is measured consistent with the approached used to measure OC_CEO.