Retail shareholders and the efficacy of proxy voting: evidence from auditor ratification

In contrast, many of the other issues on the proxy statement are confounded by company performance. For example, determining an informed vote on a director or audit committee member’s election is difficult because directors and audit committee members have direct influence over both financial statement quality and company performance. Similarly, determining the reasonableness of compensation paid to executives is a very difficult task that depends on a variety of complex factors, including (but not limited to) the personal preferences of the individual executive, the supply of competent alternative executives in the marketplace, and the intangible value of the individual executive due to synergies with employees or technical expertise.

Our objective is to provide evidence relevant to the SEC’s recent campaign to increase retail investor participation in the proxy voting process. While our results indicate that vote outcomes vary with the composition of a company’s shareholder base, we emphasize that we are documenting on-average effects. That is, we are not suggesting that all institutional shareholders are informed or that all retail shareholders are uninformed.

For example, in its DEF 14A filed on April 5, 2016, Amazon.​com states, “The Audit Committee will evaluate the shareholder vote [on auditor ratification] when considering the selection of a registered public accounting firm for the audit engagement for the 2017 fiscal year.”

As discussed above, we focus on a setting where developing an expectation for an informed vote is arguably much easier relative to other proxy vote settings. However, it is possible that retail shareholders could be engaging in so-called “vote no” campaigns in which they vote against all management proposals in an attempt to signal their dissatisfaction with company performance. While “vote no” campaigns likely limit the usefulness of vote outcomes on specific proxy items, we acknowledge that this type of voting behavior may be justifiable when considering individual shareholder incentives and when viewed more holistically (i.e., beyond a specific proxy item).

Increasing disclosure and improving the proxy process is currently at the forefront of SEC attention. According to SEC Commissioner Kara Stein, “Our current proxy system is arcane at best, … it has not necessarily helped to provide transparency to either the companies or their investors” (SEC 2018b). Consistent with this, Dasgupta and Noe (2019) suggest that shareholders often do not have access to information relevant to proxy voting decisions. Moreover, both the PCOAB and the U.S. Department of the Treasury’s Advisory Committee on the Auditing Profession (ACAP) have argued that increased disclosure related to audit quality could enhance the auditor ratification process (ACAP 2008; PCAOB 2013).

ISS and other proxy advisors (e.g., Glass Lewis) provide recommendations on the auditor ratification vote. We do not consider these in our analyses because we do not have access to proxy advisor recommendations data. However, we do not expect the exclusion to affect our inferences because 1) it is unclear the degree to which proxy advisors influence voting outcomes (Copland et al. 2018), 2) proxy advisor recommendations against auditor ratification are rare (e.g., Cunningham (2017) reports only 12 such recommendations in her sample of over 9000 company-years from 2009 to 2012), and 3) the small number of recommendations against auditor ratification would have to vary systematically across the distribution of retail ownership in a way that seems unlikely.

For example, Mayhew and Pike (2004) provide experimental evidence that transferring auditor retention/dismissal authority from managers to investors leads to higher audit quality. Dao et al. (2012) provide archival evidence to support the findings of Mayhew and Pike (2004) by documenting that subsequent restatements are less likely and abnormal accruals are smaller among companies that submit auditor ratification to a vote.

Further demonstrating that the auditor ratification vote has meaningful implications, Tanyi and Roland (2017) document that lower shareholder approval of the auditor is associated with a negative market reaction to the announcement of the auditor ratification vote results.

Liu et al. (2009) and Cunningham (2017) consider these characteristics when examining factors that influence auditor ratification voting outcomes, finding mixed results. Specifically, Liu et al. (2009) find a positive relation between a restatement announcement and votes against ratification, while Cunningham (2017) finds no such association. Moreover, Liu et al. (2009) find a negative relation between stock market returns and votes against ratification, while Cunningham (2017) finds no such association. While Liu et al. (2009) and Cunningham (2017) provide evidence on the relation between shareholder dissent on auditor ratification and restatements/negative returns (along with a number of additional explanatory variables), the on-average relation between votes against auditor ratification and restatements/negative returns is not the focus of our study. Rather, we rely on the idea that an informed voter is more likely to vote against the auditor following an audit failure (resulting in a relation between shareholder dissent and restatements) and examine whether retail ownership moderates this relation. Stated differently, we seek to assess whether retail ownership influences the extent to which the auditor ratification vote is informed.

If we instead have the proxy vote window begin either 30 days prior to the official meeting date or on the official meeting date itself, our inferences are unchanged throughout.

In untabulated analyses, we consider two alternative specifications to capture stock market performance. First, we re-specify our negative annual return indicator variable after measuring returns relative to the value-weighted market return (i.e., abnormal returns). Second, we specify the decile rank of annual stock returns. Our results are unchanged using either of these alternative measures.

Consistent with Hermanson et al. (2009) and Liu et al. (2009), we use the natural logarithm to reduce the skewness of the proportion of votes cast against auditor ratification. Our results are unchanged if we follow Cunningham (2017) and Tanyi and Roland (2017) by including abstaining votes in the numerator of our measure of shareholder dissent, or if we scale votes against ratification by common shares outstanding as of the fiscal year-end when the vote occurs instead of by total votes cast.

We use a simple interaction specification in our primary tests because it simplifies the interpretation of the results in some of our additional analyses. In untabulated analyses, we also consider a specification that allows the coefficients on all control variables to vary across the distribution of retail ownership (i.e., a fully interacted model) and find that our inferences are unchanged.

We obtain institutional investor holding data from investment managers’ Form 13F filings. One of the limitations of the institutional holdings data is that it is only reported on a quarterly basis. Because we cannot always determine a company’s precise level of institutional ownership—and thus retail ownership—leading up to the proxy vote date, we create RETAILOWN_t using the average proportion of institutional ownership over the two quarters that encompass the likely voting period. This reasonably approximates a company’s level of institutional ownership throughout the proxy voting window. For our sample, the median institutional ownership is approximately 71%, and the mean is approximately 61%. These levels are consistent with recent work that studies institutional investors (e.g., Cheng et al. 2013; Reid and Carcello 2017).

By construction, RETAILOWN_t classifies company insiders as retail investors. We use this approach in constructing our primary measure of retail ownership because incorporating data on insider holdings results in significant sample attrition. Nevertheless, as discussed in the Online Appendix, the correlation between our primary measure and a refined measure of retail ownership that excludes shares held by insiders exceeds 95% in the reduced sample. In addition, we find that our inferences regarding the influence of retail shareholders on the auditor ratification vote are unchanged using the reduced sample and the refined measure.

We control for the prior year proportion of dissenting votes since extant research documents that the audit committee and other stakeholders use the prior year vote outcome as a benchmark for evaluating the current year vote results (Dao et al. 2012). However, our inferences are similar when we exclude the prior year proportion of dissenting votes from the model.

We use OLS (i.e., a linear probability model [LPM]) to estimate average treatment effects (ATEs) for dismissals because it allows for a simple comparison of ATEs. Our inferences are not sensitive to this design choice. Specifically, the effects estimated using a logit model are nearly identical to the estimates obtained from LPM, and our statistical inferences are unchanged if we use a logit model.

To ensure that we appropriately link voting behavior to the auditor in place over the past year, we exclude observations where the auditor that is submitted for ratification is not the auditor that completed the most recent financial statement audit. We exclude observations with zero votes cast against ratification because our measure of dissenting votes is log-transformed.

The coefficient on NEGATIVE_RET_t in Column (2) represents the on-average effect for companies in the lowest decile of retail ownership (i.e., where RETAILOWN_t = 0). This effect is negative and significant, which seems inconsistent with intuition. However, this result is likely a product of the linear nature of the decile-ranked retail ownership variable. Consistent with this explanation, when we separately run Eq. (1) (without interactions) on the subsample of observations in the lowest decile of retail ownership, the results reveal an insignificant association between negative returns and votes against ratification and a significantly positive association between restatements and votes against ratification (untabulated).

The results in Table 3 indicate that the variable of interest in Table 4 (VOTES_AGAINST_t+1) is correlated with several of the explanatory variables included in the model. As such, we perform several (untabulated) sensitivity tests to alleviate any concerns that our inferences in Table 4 are impacted by multicollinearity. First, we calculate variance inflation factors (VIFs) for each of the explanatory variables in the baseline model in Table 4 (Column (1)). None of the VIFs are above 3 (the majority are below 2) suggesting that multicollinearity is unlikely to be a significant concern. We then consider several alternative specifications of the model in Table 4 and find that the results hold in each case. These include 1) re-estimating the tests after dropping VOTES_AGAINST_t from the model, 2) re-estimating the tests after replacing VOTES_AGAINST_t+1 with the residual of VOTES_AGAINST_t+1 from Eq. (1), and 3) re-estimating the tests using the change in votes against (VOTES_AGAINST_t+1 – VOTES_AGAINST_t) as the main explanatory variable.

In many of these cases, the auditor that is submitted for ratification is the party that identified and reported the prior auditor’s failure, suggesting they may be of particularly high quality (and, thus, deserving of support in the ratification vote).

There were 38 observations in our primary sample where the auditor in the year of the restatement announcement was not the auditor responsible for the audit of the misstated financial statements. For these observations, RESTATE_t is set equal to zero in our primary tests. In addition, we also add back observations that had a restatement announcement but 1) the auditor up for ratification was not the auditor during the prior fiscal year, 2) there was not an audit completed in the subsequent period, or 3) the company’s auditor subsequently resigned (an additional 46 observations). As described in the sample selection section, we initially remove these observations to ensure that we appropriately link voting behavior to the auditor engaged during the year leading up to the vote.

Given the magnitude of the coefficient on RESTATE_SEVERE_t*RETAILOWN_t, we perform joint tests of RESTATE_SEVERE_t and RESTATE_SEVERE_t*RETAILOWN_t at different points in the distribution of RETAILOWN_t. The results indicate that the positive relation between severe restatements and votes against ratification persists through the 25th percentile of retail ownership but becomes insignificant at the 50th percentile of retail ownership. Notably, the effect of RESTATE_SEVERE_t is never significantly negative (even at 100% retail ownership).

The results from tests that differentiate between severe and non-severe restatements are likely at least partially attributable to differential salience of the restatement announcement (i.e., while investors are more likely to respond to severe restatements, they are also more likely to hear about severe restatements via company disclosures, media attention, etc.). In an untabulated analysis, we further examine the potential effect of restatement salience using company media coverage as an alternative proxy. Specifically, we re-estimate our main model after replacing RESTATE_t with separate indicators for “high news” and “low news” restatements, which are defined based on whether the restating company is above or below the sample median number of news articles (compiled by Ravenpack) for each company-year. Inferences from these tests are consistent with those from tests that differentiate between severe and non-severe restatements.

Consistent with the measurement of our primary dependent variable, we use the natural logarithm to reduce skewness in our measure of voter turnout. We find similar inferences if we exclude abstentions in our measurement of voter turnout.

We follow Bethel et al. (2009) and scale by common shares outstanding to account for the size of the voting base, losing 29 observations that are missing this information. We note that this is not necessarily the exact number of votable shares in existence (i.e., our variable is not intended to be an exact measure of the percentage of votable shares that are cast); however, we can think of no reason that differences between total shares outstanding and total votable shares would systematically vary across our sample companies in a way that would bias our results. We choose the fiscal year end of the year in which the vote is cast instead of the prior fiscal year end to ensure there is not a mechanical association between the negative return variable and changes in shares outstanding that occur during the return measurement window (Pontiff and Woodgate 2008).

We also considered changes in S&P 500 index constituents as a plausible exogenous shock to retail investor ownership. As discussed in Aghion et al. (2013) and Agarwal et al. (2018), institutional investors are often benchmarked against the S&P 500 index and thus have strong incentives to own S&P 500 companies. Accordingly, a company’s addition to (removal from) the index should induce an increase (decrease) in passive institutional ownership and thus a decrease (increase) in retail ownership. This change should also be theoretically unrelated to votes against the auditor, suggesting that it satisfies both conditions discussed above. In untabulated analyses, we find evidence that addition to/removal from the S&P 500 index relates to changes in retail ownership. In addition, we find that our inferences using changes in S&P 500 index constituents are similar to those using brokerage closures and mergers.

Dao et al. (2012) estimate their selection model by including the fitted values from the first-stage regression as an independent variable in the second-stage regression. While our approach is conceptually similar to theirs, we follow the recommendation of Lennox et al. (2012) and perform a two-stage Heckman procedure using an exclusion restriction. Nonetheless, we note that our inferences are unchanged if we follow the same approach as Dao et al. (2012) and include the fitted values from the first-stage regression in the second-stage regression.

Similar to analyses in the prior section, we again considered changes in S&P 500 index constituents as an instrument. However, this variable is insignificant in a first-stage model of holding a vote, suggesting that it does not appear to be a valid instrument for this test.