Do PCAOB Inspections Improve the Accuracy of Accounting Estimates? Sarah B. Stuber Michigan State University

Transcription

1 Do PCAOB Inspections Improve the Accuracy of Accounting Estimates? Sarah B. Stuber Michigan State University Chris E. Hogan* Michigan State University *Corresponding author We appreciate helpful comments and suggestions on earlier drafts from Lauren Cunningham, Sean Dennis, Joe Schroeder, Jonathan Stanley, and Mike Wilkins, and from workshop participants at Michigan State University, University of Nebraska, University of Pittsburgh, University of Connecticut and the 2017 Auditing Section Midyear Meeting. We gratefully acknowledge financial support from the Eli Broad College of Business.

2 Do PCAOB Inspections Improve the Accuracy of Accounting Estimates? Abstract: Accounting estimates are increasing in importance in financial reporting, yet are subjective and difficult to audit. Despite issuing guidance intended to improve the accuracy of audited estimates, the PCAOB continues to identify deficiencies related to the audit of estimates through their inspections process. In this study, we examine whether PCAOB inspections are effective in improving the accuracy of audited accounting estimates by examining a significant estimate within the banking industry. More specifically, we use the association between the allowance for loan losses (ALL) and subsequent charge-offs to test the effectiveness of PCAOB inspections. We find that in contrast with the PCAOB s goal of more accurate and unbiased estimates (as defined in our introduction), ALL estimates become less accurate and overly conservative after the receipt of a PCAOB inspection finding related to the ALL. We then separately examine changes in ALL accuracy for public audit clients and private audit clients and find that ALL estimates become more conservative for public audit clients, but not for private audit clients, suggesting the reduced accuracy of the ALL following a PCAOB inspection finding is related to auditor concerns about future inspections rather than firm-wide changes in audit methodology. Overall, our findings are consistent with accounting estimates being difficult to audit and cast doubt on the efficacy of PCAOB inspections in improving estimate accuracy. Keywords: estimates, auditor, regulation, PCAOB, allowance for loan losses Data availability: Data are publicly available from the sources identified in the paper.

3 Do PCAOB Inspections Improve the Accuracy of Accounting Estimates? I. INTRODUCTION Accounting estimates are an important component of financial reporting; however, there is often a lack of precision and extreme uncertainty in financial statement estimates, potentially resulting in material misstatements (Christensen, Glover, and Wood 2012). Due to the inherent subjectivity in accounting estimates, developing and auditing estimates presents a challenge to both management and auditors. The PCAOB notes the important role that auditors play given the potential for management bias in developing estimates, as well as the increasing prevalence and significance of estimates in the financial reporting process (PCAOB 2017a). 1 At the same time, the PCAOB expresses concern that inspectors continue to identify deficiencies related to the audit of estimates in their ongoing inspection process for both large and small audit firms (PCAOB 2017a). In this study, we examine whether the PCAOB inspections process, specifically the identification of deficiencies related to the audit of estimates, is an effective means to improve the accuracy of financial statement estimates. While prior research provides evidence that the PCAOB inspections process improves overall audit quality (Fung, Raman and Zhu 2017; Krishnan, Krishnan and Song 2016; Lamoreaux 2016), there is limited evidence on the efficacy of the process in improving the audit of, and accuracy of, estimates. Accounting estimates require significant judgment and are material to the financial statements, making them important to financial statement preparers, users, auditors and regulators. 1 Prior research provides evidence that estimates are used to manage earnings (Gaver and Paterson 2004; Petroni 1992; Dietrich, Harris and Muller 2000; Beatty, Ke and Petroni 2002). Both the PCAOB and SEC have disciplined audit firms and companies for failures related to accounting estimates (PCAOB 2008; Kilpatrick Townsend & Stockton LLP 2013; PCAOB 2012). 1

4 In the wake of the financial crisis, the importance of estimates in conveying potential risk was highlighted even further (Balla, Rose and Romero 2012). The audit of estimates has been a priority item for the PCAOB for several years, and there is now a proposed standard related to the audit of estimates, including fair value estimates, under consideration (PCAOB 2017a). Proposed changes to the existing audit standard are informed at least in part by the PCAOB s inspection findings in this area; thus, providing evidence of the effectiveness of the inspections process in improving estimate accuracy may shed light on whether the proposed standard will be effective. Auditors have incentives to address deficiencies identified in the PCAOB inspections process to avoid additional negative consequences within their firm, and to avoid future PCAOB actions such as the release of Part II of an inspection report. To the extent inaccuracies related to estimates are due to a lack of professional skepticism on the part of auditors or due to inappropriate or insufficient audit procedures, we would expect the inspections process to result in changes in audit firm procedures and ultimately improvements in the accuracy of estimates. On the other hand, if estimates are simply difficult to develop and audit, or management and auditors face competing incentives, the inspections process may not be effective. Auditors may even prefer more conservative estimates, as opposed to more accurate estimates, to the extent they perceive this will reduce the likelihood of future inspection comments and/or reduce litigation risk resulting from negative inspection reports (Christensen, Lundstrom, and Newton 2018). Examining potential spillover effects of the PCAOB inspections process on the audits of private clients has the potential to provide insight on the possible outcomes of inspections. To the extent audit firms are making changes to their audit process, we would expect to observe changes in estimate accuracy for private clients similar to that observed for public clients. On the other hand, if concerns related to future inspection comments are driving changes in the response to 2

5 inspection deficiencies, we would expect to observe changes in estimate accuracy for public clients but not private clients. We extend our analyses to consider these possibilities to shed light on the spillover of PCAOB inspections, an area which has been not been explored in prior research to our knowledge. Consistent with guidance in the auditing standards, we consider the accuracy of an estimate to be the degree to which the estimate (e.g., allowance for loss, fair value) is developed by competent personnel using relevant and reliable data, and without bias (PCAOB 2501). While estimates will never be completely accurate due to inherent uncertainty in estimate development, an estimate developed with a disciplined and unbiased approach in the current period should closely match the economic reality (e.g., write-off, realized value) experienced in a subsequent period. Thus, for the purposes of our study, we define estimates that more closely match future economic reality to be more accurate. We use the allowance for loan losses (ALL) in the banking industry as a setting for our study. This setting is ideal for the study of estimates because (1) the accuracy of the ALL can be evaluated ex post 2, (2) the ALL calculation is dependent on subjective inputs that can be difficult for management to develop and auditors to evaluate, and (3) more recently the ALL is subject to high levels of regulatory attention, increasing the likelihood that inaccuracy could result in regulatory action (Financial Stability Forum 2009; Balla et al. 2012). The availability of data for both public and private banks allows us to examine possible spillover effects of the PCAOB inspections process to audits of private clients. In addition, most prior studies on the effectiveness of the PCAOB inspections process remove firms operating in highly-regulated industries (i.e., banks and 2 The failure of auditors to conduct an ex post review of the accuracy of the ALL is specifically identified by the AIPCA as being an area of audit failure identified in peer reviews (AICPA 2015). This underscores the assumption that the appropriateness of the allowance calculation can and should be evaluated ex post. 3

6 insurance companies); thus, we know little of how the inspections process affects the audits of these institutions. Examining a key estimate within the banking industry allows us to provide insight into the effects of the inspections process within an important, regulated industry. We assess the change in the accuracy of the ALL in response to PCAOB-identified deficiencies specifically related to the audit of the ALL using two approaches: (1) a model modified from Cantrell, McInnis, and Yust (2014) estimating the association between the current period ALL and subsequent charge-offs, 3 and (2) a model examining estimate error that allows us to assess whether differences between the ALL and subsequent charge-offs are due to conservative estimates or under-reserving. We estimate these models across all banks, and then separately for public banks and private banks. Using a sample of 16,675 bank-year observations from , we find that PCAOBidentified deficiencies specifically related to the ALL are associated with a decrease in association between the ALL and subsequent charge-offs (i.e., a decrease in accuracy of the ALL) as well as an increase in estimate error. Inconsistent with the emphasis of the PCAOB on unbiased, reasonable estimates, we find that ALL estimates become more conservative subsequent to PCAOB inspection findings related to the ALL. The movement towards more conservative estimates is most pronounced in banks operating in more lenient regulatory environments, which is consistent with our expectation that the auditor, and thus PCAOB inspection findings, would have greater influence in environments with weaker regulatory oversight. These findings suggest that in order to respond to PCAOB findings, the auditor makes changes to their audit approach, but because of the inherent difficulties of auditing estimates, these changes push the client to make more conservative, but not more accurate estimates. 3 Consistent with prior literature (e.g., Cantrell et al. 2014) we define the accuracy of the ALL as being the extent to which the ALL accurately predicts future charge-offs. 4

7 We also find evidence suggesting more conservative estimates are associated with auditors avoiding additional subsequent PCAOB findings related to the ALL, consistent with the inspections process incentivizing conservative estimates. The findings do not vary based on litigation risk, providing evidence that our results are driven by the auditor s desire to avoid future inspection findings, rather than to mitigate litigation risk. We note that our inferences on the impact of inspection findings on estimate accuracy hold after controlling for economic variations over our sample period, including controlling for the financial crisis. When we separately estimate the models for public and private banks to examine potential spillover effects of PCAOB inspections on private company audits, we find the decrease in accuracy (due to more conservative estimates) is significant for public banks but not for private banks. For private banks, we find there is no significant change in the accuracy of the ALL following PCAOB-identified deficiencies. Finally, we separately estimate the models for annually inspected and triennially inspected audit firm clients and find that the decrease in estimate accuracy is significant for both groups, but greater for the annually inspected audit firms, those subject to greater PCAOB scrutiny. Overall, our findings are consistent with estimates being inherently difficult to develop and audit, and auditors favoring more conservative estimates for their public banks possibly to avoid future inspection scrutiny. The findings cast doubt on the efficacy of the inspections process at improving estimate quality, at least for banks and their auditors. In addition, we find no evidence in our setting of a spillover effect on audit quality for private clients. In addition to documenting the findings discussed above, our study contributes to the literature on the influence of auditors and regulators on bank behavior (e.g., Nicoletti 2017). While bank regulators (e.g., OCC) have voiced concerns that auditors are responsible for delayed loss 5

8 recognition and potential under-reserving for loan losses (Dugan 2009), our results suggest that auditors influence the estimates of banks in a more conservative direction. Finally, we contribute to the literature examining the accuracy of the ALL (e.g., Cantrell et al. 2014) by providing insight into the (in)effectiveness of regulatory actions on addressing inaccuracies in ALL estimates and demonstrating the auditor s impact on the ALL. With the change from the incurred loss model to the expected credit loss model effective in 2018 (FASB 2016), the efficacy of regulatory action and auditors in monitoring the accuracy of the ALL will continue to be an important issue. II. BACKGROUND AND HYPOTHESIS DEVELOPMENT Research examining the effect of PCAOB inspections on financial reporting suggests that PCAOB inspections improve audit quality. Studies taking advantage of both cross-country and within-country variation in exposure to inspections demonstrate that audit quality, as measured by accruals quality, is higher for clients of audit firms subject to PCAOB inspections (Fung et al. 2017; Krishnan et al. 2016; Lamoreaux 2016). Relatedly, Aobdia and Shroff (2017) find that the increased regulatory scrutiny of PCAOB inspections is associated with increases in market share for non-u.s. auditors subject to PCAOB inspections, consistent with higher perceived audit quality. In contrast to these findings, Lennox and Pittman (2010) conclude that investors do not perceive an increase in audit quality as a result of PCAOB inspections. In addition to studies of the general effects on financial reporting quality, there are a few studies examining more targeted scrutiny. For example, Defond and Lennox (2017) find that audit firms improve the quality of their internal control audits when there is increased scrutiny in PCAOB inspections. Acito, Hogan, and Mergenthaler (2017) provide evidence that auditors respond to 6

9 audit deficiencies identified in the PCAOB inspection process by increasing audit effort, as proxied by audit fees, on clients with greater exposure to areas identified as deficient. Despite research supporting the effectiveness of the PCAOB inspection process in improving overall financial reporting quality, there is limited evidence on its effectiveness in improving the accuracy of accounting estimates. The PCAOB has identified estimates as pervasive to the financial statements, often substantially affecting a company s financial position and results of operations (PCAOB 2017b). 4 Despite its emphasis on the importance of estimates, the PCAOB continues to identify audit deficiencies related to auditing estimates and has issued additional guidance summarizing auditor s responsibilities in this area (PCAOB 2017b; PCAOB 2008). While one might expect the inspections process and additional guidance to result in improvements, there are several reasons why the scrutiny and guidance may not be effective in improving the audit of estimates. First, estimates are inherently difficult to develop and audit. Second, management incentives, coupled with the subjectivity of estimates, may make it difficult for auditors to significantly influence reported estimates. Third, auditors may have other concerns, such as litigation risk, that compete with the incentive to address PCAOB inspection comments. Estimates, by their nature, are inherently difficult to develop and audit 5. While testing transaction-based accounts is a mechanical process, auditing accounting estimates requires a thorough understanding of the process of developing the estimate, and audit procedures may require substantial modification year over year (Martin, Rich, and Wilks 2006). Auditing standards 4 The PCAOB also notes that while estimates rely on subjective inputs, estimates are not purely discretionary as according to auditing standards, the development of estimates is bounded by the requirements of the applicable financial reporting framework and reasonableness, such that financial statements are presented fairly in all material respects in conformity with applicable financial reporting framework (PCAOB 2017b; AS 2815). 5 Auditing standards outline three methods for auditing financial statement estimates: (1) develop an understanding of management s process and test management s assumptions, (2) develop an independent estimate to compare to management s estimate, or (3) compare the estimate to subsequent events (AIPCA 1988; AS 2501). PCAOB reports and academic studies identify the first option as the primary method used by auditors to audit estimates (Griffith et al. 2015). 7

10 require that auditors both understand the process of estimate development and assess the reasonableness of accounting estimates. Inconsistent with this guidance, Griffith, Hammersley, Kadous, and Young (2015) find that auditors instead tend to focus on verifying the accuracy of management s calculation. Furthermore, auditors have been shown to react to ambiguity with less conservatism than typical financial statement users, suggesting that the level of auditors professional skepticism when approaching subjective estimates may be less than optimal (Nelson and Kinney 1997). Adding to the challenge is some auditors lack of valuation knowledge and potential lack of training on the most effective methods of evaluating accounting estimates (Bratten, Gaynor, McDaniel, Montague, and Sierra 2012; Griffith et al. 2015). Management incentives to utilize the subjectivity of estimates to maximize their own benefit at the expense of financial statement users, coupled with the subjective nature of estimates, raises concerns that management may bias estimates either upward or downward. Both types of bias result in a departure from GAAP (PCAOB 2017a). While the role of auditors is to provide assurance on the financial statements, and to reduce management s opportunity to prepare and disclose biased or inaccurate financial information (PCAOB 2017a), findings in prior research suggest that at least historically, auditors may have been aware of earnings management using estimates, and tolerated it to some extent to preserve client relationships. 6 Such a lack of professional skepticism by auditors is of particular concern to the PCAOB and could explain the continued issues related to estimate accuracy. 6 Evidence from prior studies suggest that auditors either did not detect or did not correct earnings management in a variety of estimates including internal appraisals of investment property (Dietrich et al. 2000), insurance loss reserves (Gaver and Paterson 2004; Petroni 1992), and banking allowance for loan losses (Beatty et al. 2002; McNichols and Wilson 1988). The magnitude of the errors in accounting estimates often exceeds standard materiality levels, and even small adjustments to the inputs used to calculate estimates can materially impact reported numbers (Christensen et al. 2012; Petroni and Beasley 1996). 8

11 Auditors also have competing incentives and may choose to compromise when determining how best to address PCAOB concerns. While prior research documents increases in audit effort following PCAOB inspections (e.g., Acito et al. 2017), there is also research suggesting that deficiencies identified in PCAOB inspection reports increase litigation risk (Christensen et al. 2018). If auditors are concerned about heightened litigation risk associated with inspection deficiencies, they may be focused on more conservative, rather than more accurate, estimates. Auditor perception that conservative estimates may draw less PCAOB scrutiny may also result in more conservative estimates as auditors attempt to avoid future PCAOB findings. This would be in contrast to the PCAOB s stance that bias in estimates, whether aggressive or conservative, is undesirable. 7 To summarize, deficiencies related to the audit of estimates continue to be identified and reported despite the PCAOB s focus on the importance of professional judgment in this area and the additional guidance issued. This may be the result of the inherent difficulty in developing and auditing estimates, management incentives, and/or auditor incentives. As a result of these competing factors, we state our primary hypothesis in the null form. H1: The accuracy of accounting estimates for clients of an audit firm does not change following disclosure of PCAOB-identified deficiencies related to the audit of estimates. To further examine the impact of the PCAOB inspection process on the accuracy of estimates, and consider the competing predictions discussed above (i.e., impacts that may be due to changes 7 In a proposed auditing standard, the PCAOB states, a favorable estimate can reflect either an upward or downward bias, for example in earnings, depending on management incentives (PCAOB 2017a), confirming that the PCAOB considers both upward and downward bias to be undesirable. 9

12 in audit approach from those due to auditor incentives), we separately examine changes in the accuracy of estimates following PCAOB scrutiny for public audit clients versus private audit clients. To our knowledge, spillover effects of PCAOB inspections on the audits of private companies has not been studied. To the extent auditors make firm-wide changes to their audit approach based on PCAOB inspection feedback, we would expect auditors to change their approach to the audit of estimates on both public and private banks. This would be consistent with the vision of the PCAOB to improve audit quality. 8 However, if instead auditors only change their behavior on audits of public banks, this suggests they are concerned more about inspection risk (i.e., continued identification of deficiencies related to estimates), than improvements in audit process. We state the following hypothesis related to possible spillover effects in the null form. H2: The impact of PCAOB-identified deficiencies on the accuracy of accounting estimates does not differ for public bank audit clients versus private audit bank clients. III. RESEARCH SETTING AND DESIGN Research Setting: The Allowance for Loan Losses Estimate Banks provide an ideal setting to examine our research question for a number of reasons. First, estimates are material to bank financial statements. Bank financial reporting relies on estimates of the allowance for loan losses (ALL), fair values of investment assets, and foreclosed properties, and banks are required to disclose the fair value of most balance sheet items, including all loans and deposits. In this study, we focus specifically on the ALL to gain insight into the overall impact 8 The PCAOB Vision Statement as listed at pcaobus.org is as follows: The PCAOB seeks to be a model regulatory organization. Using innovative and cost-effective tools, the PCAOB aims to improve audit quality, reduce the risks of auditing failures in the U.S. public securities market and promote public trust in both the financial reporting process and auditing profession. 10

13 of PCAOB inspections on bank estimates, because the ALL requires significant judgment and is considered to be one of the riskiest areas of a bank audit (Beatty et al. 2002; Sutton 1997). In addition, the availability of charge-off and other credit loss information makes it possible to evaluate the accuracy of the allowance for loan losses ex post. 9 Second, the PCAOB specifically cites the ALL as an example of a key estimate that is susceptible to management bias, and auditing failures related to the ALL are often the subject of inspection findings (PCAOB 2017a; PCAOB 2008). The flexibility within GAAP for estimating the ALL allows for sufficient variation in models, suggesting there is an ability for auditors to influence the accuracy of this particular estimate if PCAOB scrutiny and guidance is effective. 10 The bank setting provides for competing incentives for management and the auditors. On the one hand, management may have incentives to increase income through a lower provision for loan losses and therefore lower ALL. On the other hand, banks may have incentives to over-reserve for loan losses in good times, and prior research has shown that banks use the ALL to smooth earnings by accelerating provision in profitable times (Liu and Ryan 2006; Greenawalt and Sinkey 1988). Oversight by bank regulators may encourage more conservative ALL estimates as regulators primary purpose is to ensure that financial institutions maintain safe-and-sound business practices (Curry, Coburn and Montgomery 1999). Consistent with this perspective, Wall and Koch (2000) note that bank regulators view the ALL as a type of additional capital cushion in case of economic downturn and thus favor a more conservative ALL. In contrast to these incentives, 9 We note that several papers examining estimates in the banking industry use the provision as the variable of interest, and often use quarterly provision and charge-off amounts, as these studies primarily focus on earnings management and the provision captures the income statement impact of changes in the allowance (Hribar, Melessa, and Small 2017; Beatty et al. 2002). Our focus is on the accuracy of the year-end audited allowance rather than the amount of the provision, although we note the two are highly correlated (0.88) in our sample. The allowance is the estimate that is the focus of the audit as opposed to the provision. 10 Beatty and Liao (2014) identify nine different allowance for loan loss models that fall within GAAP and illustrate the myriad of choices for management and the extent of judgment involved. 11

14 auditors responsibilities under the PCAOB auditing guidance is to provide assurance that estimates, and ultimately financial statements, are free from bias that results in material misstatement. Finally, data is available for both public and private banks, allowing for a comparison of changes in accuracy across these two client groups (a test of H2). Model of the Predictive Ability of the ALL To evaluate the accuracy of the ALL, we first measure the ability of the ALL to predict credit losses in the subsequent year using a model modified from Cantrell et al. (2014). 11, 12 The purpose of the ALL is to reserve for loan losses that are inherent in the existing loan portfolio. 13 These losses are recognized by the bank in the form of charge-offs, thus the ability of the ALL in time t to explain charge-offs in fiscal year t+1 is a feasible proxy for the accuracy of the ALL. According to the SEC, a registrants loan loss allowance methodology is considered valid when it accurately estimates the amount of loss contained in the portfolio. Thus, the staff would expect the registrant s methodology to include procedures that adjust loan loss estimation methods to reduce differences between estimated losses and actual subsequent charge-offs (SEC 2001). Because the goal of the ALL is to estimate or predict future credit losses, for the purposes of our discussion and to use 11 We note that Cantrell et al. (2014) also estimate a model using leading non-performing assets (NPA) as the dependent variable. In an untabulated robustness test, we estimate all models using NPA t+1 as the dependent variable and note no qualitative change in our results. Given the regulatory focus on the charge-off coverage of the ALL, we discuss only the model using CO t+1 as the dependent variable for the remainder of the paper. 12 Cantrell et al. (2014) use LOANHC as their variable of interest, calculated as the gross principal balance of loans less the loan loss reserve, or one minus ALL, as a percent of gross loans. They interpret a lower LOANHC as being indicative of higher expected losses. In our model, we use the inverse of this relationship, calculating ALL as a percentage of total loans and interpret a higher ALL as being indicative of higher expected losses. We consider this to be a more straightforward interpretation for the purposes of our study. 13 Credit losses are estimated following the incurred loss model during our sample period. Under the incurred loss model, the ALL captures expected losses related to non-performing loans on a loan-by-loan basis and expected losses related to the remainder of the portfolio based on FAS 5 (ASC ). 12

15 consistent terminology, we will refer to this predictive ability as the accuracy of the ALL. Both H1 and H2 are tested by estimating variations of the following model: COit+1 = β1allit + β2pcaobt + β3pcaob*allit + β4 SIZEit + β5coit + β6npait + β7capit + β8commercialit + β9consumerit + ai + ωi + εit (1) Consistent with Cantrell et al. (2014) and regulatory guidance, we use next period net chargeoffs (CO), calculated as charge-offs less recoveries, to measure subsequent credit losses. 14 The variable is scaled by gross loans. The primary variables of interest are the allowance at the end of year t (ALL) and an interaction between ALL and an indicator variable capturing the receipt of a PCAOB inspection finding (PCAOB*ALL). ALL is measured as the allowance for loan losses as a percentage of total gross loans. Ceteris paribus, the higher the value of ALL, the higher the level of the allowance for loan losses, which suggests an expectation of higher credit losses. If the allowance is appropriately reserving for future credit losses it should be positively associated with the level of subsequent charge-offs, thus we expect a positive coefficient on ALL. PCAOB is an indicator variable set equal to 1 if the audit firm s PCAOB inspection report includes a finding citing an audit failure specifically related to an ALL for any inspected engagement in the current or any of the previous two fiscal years based on the issue date of the report. We considered this period to be appropriate for two reasons. First, firms that audit fewer than 100 publicly-traded clients per year are inspected every three years, and thus it is reasonable to expect that the impact of a PCAOB finding would have a duration of approximately three years. Second, even for audit 14 The Office of the Comptroller of the Currency (OCC) considers coverage of one year s charge-offs to be a reasonable level of ALL (OCC 2012). 13

16 firms inspected annually, we expect auditors to continue to be sensitive to increased scrutiny for at least three years (Boone, Khurana and Raman 2015). We control for bank size (SIZE), using the log of total assets as an independent variable. We include non-performing assets as of the end of the year (NPA) and current year charge-offs (CO) since the allowance is determined in part by these factors. 15 Non-performing assets typically will lead charge-offs, as assets move from performing to non-performing before being eventually charged-off. In addition, we include the current period capital ratio (CAP) as the regulatory scrutiny faced by a bank and a bank s underlying risk are closely tied to its capital ratio. Loan portfolio composition can affect the accuracy of the ALL estimate, thus we include controls for loan portfolio composition by including the percentage of commercial (COMMERCIAL) and consumer (CONSUMER) loans in bank i s loan portfolio. Finally, we include year fixed effects (ω) to control for time trends and bank fixed effects (a) to control for time-constant bank characteristics. Standard errors are clustered at the bank level. Consistent with Beck and Narayanamoorthy (2013), we do not include economic variables, because the ALL estimation should consider economic conditions. However, our results are robust to the inclusion of economic condition variables. 16 The coefficient on ALL in Model (1) (β1) measures the change in the expected value of future credit losses if the ALL is increased by one unit, holding any control variables constant, for observations where PCAOB = 0. The coefficient on ALL then captures the association between 15 We note that the CO model specification exhibits multicollinearity. Because of this multicollinearity, we add each independent variable to the regression separately, and note that the inclusion of both CO it and NPA it does not result in a change in the sign or significance of the coefficients on ALL it or PCAOB*ALL it. Furthermore, multicollinearity would bias against us finding significant results when estimating the model, as standard errors are inflated in the presence of multicollinearity. We consider the inclusion of both NPA it and CO it to be appropriate due to their mechanical relationship with the ALL it and subsequent charge-offs. Thus, we do not consider multicollinearity to be a concern, as it does not cause bias or inconsistency of the coefficient estimates or invalidate standard errors. 16 Following Beatty and Liao (2014) we include variables capturing inflation and GDP as they identify these variables as appropriate for use in testing the accuracy of the loan loss provision. 14

17 ALL and future credit losses after controlling for the more mechanical portion of the allowance calculation. We do not have a prediction for the correct value of these coefficients but note the values intuitively should be between 0 and 1, with higher values indicating a greater association between the ALL and future credit losses. The coefficient on the interaction PCAOB*ALL captures the change in the association between ALL and future credit losses for observations where PCAOB = 1 relative to the association for observations where PCAOB = 0, a test of H1. As an alternative approach to evaluating estimate accuracy and to help us determine whether the differences in the association the between the ALL and charge-offs documented are due to under-reserving or a more conservative estimation of the ALL, we estimate the effect of regulatory actions on estimate error. One of the primary benefits of using the ALL to examine the accuracy of estimates is the possibility of evaluating an estimate error ex post. This error can be estimated by subtracting the subsequent period COs from the current period ALL. While we acknowledge that there will always be some amount of difference due to the uncertainty inherent in the estimate, a smaller difference between the ALL and subsequent COs is indicative of a more accurate ALL estimate. To examine the ALL error, we estimate the following model: ALL ERRORit = λ1pcaobit + λ2sizeit + λ3coit + λ4npait + λ5capit + λ6commercialit + λ7consumerit + ai + ωi + εit (2) where ALL ERRORit = ALLit COt+1. We use the signed error, rather than absolute value of the error, to help determine whether the error is caused by banks becoming overly conservative or by banks under-reserving; however, we also estimate the model using the absolute value of the 15

18 difference as an untabulated sensitivity test. 17 Given the majority (90.5%) of the values for ALL ERROR are positive, a positive coefficient on PCAOB in Model (3) would be indicative of the error resulting from banks becoming overly conservative. A negative coefficient on PCAOB in Model (3) would be consistent with the inspection process resulting in a decrease in the magnitude of the error, while a positive coefficient suggests the estimate becomes more conservative. 18 To test H2, we first estimate Model (1) separately for a sample of only public banks and a sample of only private banks. This estimation allows us to examine how the PCAOB finding impacts ALL accuracy for public and private banks separately. As a second test of H2, we use a difference-in-differences specification to estimate the difference in ALL accuracy in response to a PCAOB finding for public banks compared to private banks using the following model: COit+1 = δ1allit + δ2pcaobt + δ3pcaob*allit + δ4public + δ5public*all + δ6public*pcaob + δ7public*pcaob*all + δ8sizeit + δ9coit + δ10npait + δ11capit + δ12commercialit + δ13consumerit + ωi + εit (3) where variables are defined as described for Model (1), with the addition of PUBLIC, an indicator variable equal to 1 if bank i is publicly traded, and zero otherwise. Our primary interest is in the coefficient on the triple interaction of PUBLIC*PCAOB*ALL (δ7), which captures the incremental change in accuracy of the ALL for public banks compared to private banks when PCAOB = 1. Due to the limited variation in a bank s issuer status over time, we do not use bank fixed effects when 17 There is no change in inferences when the absolute value of the error is used or if observations with a negative error are eliminated. 18 It is important to note that there was no time during our sample period when banks were on average under-reserved. In fact, in our entire sample period there are only 1,560 firm-year observations that had a negative error, which would indicate that the bank was under-reserved. 16

19 estimating Model (2). Consistent with our estimation of Model (1), we include year fixed effects (ω) and standard errors are clustered by bank. We also conduct an analysis of ALL error as a supplemental test of H2. For this test, we estimate Model (2) separately for public and private banks. We also use a difference-in-differences approach to compare public and private banks by estimating the following model: ALL ERRORit = π1pcaobit + π2publicit + π3pcaob*publicit + π4sizeit + π5coit + π6npait + π7capit + π8commercialit + π9consumerit + ωi + εit (4) where the coefficient on PCAOB*PUBLIC (π3) is the coefficient of interest and captures the difference in ALL error for public banks compared to private banks when PCAOB = 1. A significant coefficient, whether positive or negative, on PCAOB*PUBLIC would indicate that auditor s response to a PCAOB finding differs for its public clients compared to its private clients. A negative coefficient on PCAOB*PUBLIC would be consistent with the inspection process resulting in a more significant decrease in the magnitude of the error for the public clients of an auditor following an inspection finding, while a positive coefficient would suggest that the estimate becomes more conservative for an auditor s public clients, compared to its private clients. IV. SAMPLE AND DESCRIPTIVE STATISTICS Financial Institutions and Auditors We obtain financial data for depository financial institutions from the FDIC Statistics on Depository Institutions (SDI) database and we obtain auditor information for each bank from the FR Y-9C bank holding company regulatory reports. We match the two databases to identify 17

20 auditors. We obtain the PCAOB inspection reports from the PCAOB website. Observations were collected from 2005 to 2015; however, the empirical test requires one-year-ahead charge-offs, thus the period of analysis is limited to 2005 to The auditor sample is limited to firms that issued at least 90 audit opinions for financial institutions in the FR Y-9C database over the period of (the full period of analysis) and are inspected by the PCAOB. The cutoff was set at 90 audit opinions as this seemed to be an appropriate level to capture the audit firms that audit a large cross-section of publicly traded banks. It also appeared to be a natural cutoff, as the firms that fell below the threshold generally audited fewer than 10 financial institution clients in all years in our sample period. Financial institutions are limited to depository institutions that are audited and have data available in SDI for total loans, allowance for loan losses, net charge-offs, non-performing assets and other real estate owned, and non-negative values for total assets, deposits, gross loans, net loans, and allowance for loan losses. These criteria result in a sample of 44 audit firms and 16,675 financial institution firm-years (details of sample selection provided in Table 1, Panel A). Banks included in the sample were limited to only those audited by one of the 44 sample auditors. The average number of client firm-years audited by each firm during the period was 919, with a maximum of 2,178 and a minimum of The median number of client firm-years audited by sample auditors over the period was 510. Table 1, Panel B reports descriptive statistics for the allowance, net charge-offs, and nonperforming assets as a percentage of gross loans, as well as other control variables, for all bank firm-years included in the sample. All continuous variables are winsorized at the 1 st and 99 th percentiles. The mean (median) allowance to loan ratio is 1.6 percent (1.4 percent), while the mean 19 The minimum number is lower than the 90 bank year cutoff used to determine firms included in the sample, because the cutoff was established prior to eliminating firms from the regulatory database that are missing necessary variables for the analysis (see table 1) 18

21 (median) charge-offs to total loans for the sample is 0.60 percent (0.20 percent) and mean (median) NPA to total loans is 2.1 percent (1.2 percent). Table 1, Panel B also reports the descriptive statistics for the PCAOB inspection findings, described in detail below. PCAOB is coded as 1 in the year of issuance and the two years following, and 0 otherwise. Panel B displays the proportion of observations per year for which PCAOB equals 1 (0.401), as well as the proportion of public clients in the sample (0.336). Table 2 presents a correlation matrix showing the pairwise correlation of key variables. Not surprisingly, ALL, CO and NPA are highly correlated. PCAOB Inspection Reports We download the PCAOB inspection reports for each of the 44 audit firms from the PCAOB website. The eight largest audit firms in our sample have annual inspection reports while the remaining thirty-six have an inspection report every three years. We search the inspection reports for any mention of the ALL and verify the inspectors identified a deficiency related to the audit of the ALL. There are a total of 85 audit deficiencies related to the ALL identified across all audit firms, with 29 of the 44 firms having at least one deficiency identified during the period. The findings in the PCAOB inspection reports are directly related to the audit process, but clearly take into account the accuracy of the allowance estimate. An example of the focus on both the process and accuracy of the estimate is in the following finding from one of the inspection reports in our sample (PCAOB 2010): In this audit, the Firm failed to perform adequate audit procedures to test the allowance for loan losses. Specifically, the Firm failed to evaluate the reasonableness of the issuer applying the same loss factors in its loan grading system as in the prior two years despite significant increases in the average net charge-offs, the total net chargeoffs, and the level of non-performing loans, as well as significant adverse changes in the economic 19

22 environment. (10) In this finding, the PCAOB notes deficiencies in audit procedures, but specifically cites failure to make proper adjustments given changes in the banks underlying risk factors, which is directly related to estimate accuracy. V. RESULTS Impact of PCAOB-Identified Deficiency: H1 To test H1, we examine the impact of a PCAOB-identified deficiency related to inadequate procedures in the audit of the ALL. Table 3, column (1) presents the results of estimating Model (1) using next period charge-offs as the measure of future credit losses controlling only for size, while column (2) presents the results of estimating Model (2), controlling for capital ratio, and current COs and NPAs. The coefficients on ALL are positive and significant (β1 = and β1 = 0.338, p-value < 0.01) suggesting the ALL is positively associated with future charge-offs, as we would expect, for observations where PCAOB equals 0. For example, moving from the first quartile value of ALL to the third quartile value of ALL is associated with an increase in next period CO of [( )*0.338)] for observations where PCAOB equals zero, holding other variables constant. The coefficient on PCAOB*ALLit is negative and significant in both model specifications, (β3 = and β3 = , p-value < 0.01). The negative coefficient indicates that the receipt of a PCAOB inspection finding by an audit firm is associated with a decline in estimate accuracy rather than an improvement in the accuracy of the ALL estimate. In terms of control variables, current period charge-offs (CO), non-performing assets (NPA) and the ratio of commercial to total loans (COMMERCIAL) are also significantly, positively associated with next period charge-offs. 20

23 We examine estimate errors to determine if the decrease in accuracy is the result of banks becoming overly conservative (i.e., over-reserving) or under-reserving following their auditor receiving a PCAOB inspection finding. Table 4, column (1) presents the results of estimating Model (2) controlling only for size, while column (2) presents the results for the full set of controls in Model (2). The results are consistent with our inferences from testing H1 in Table 3. The positive and significant coefficients on PCAOB in both model specifications (λ1 = 0.001, p-value < 0.05) provide evidence of an increase in estimation error following a PCAOB-identified deficiency, consistent with client banks becoming more conservative in their allowance estimation. There is evidence in prior research that audit firms adjust procedures and increase audit effort at both the individual client level as well as the national level in response to publicly-disclosed inspection deficiencies (Aobdia 2016; Acito et al. 2017). 20 However, it is not clear whether these changes improve financial reporting quality or whether they are overreactions in an effort to avoid future inspection deficiencies. Thus, we conduct a supplemental analysis to examine whether a movement towards overly conservative estimates following a PCAOB inspection finding is an effective strategy to avoid future findings. In the analysis we compare the ALL accuracy of clients of auditors that receive multiple PCAOB findings, to the ALL accuracy of the clients of auditors that successfully avoid subsequent findings. We add an interaction between ALL, PCAOB, and PCAOB*ALL and FUTUREFIND, an indicator variable equal to one if a bank is audited by an auditor that receives a subsequent PCAOB finding within the next three years, and zero 20 Acito et al. (2017) find that audit firm responses to PCAOB-identified deficiencies are similar for the Big 4 and other annually-inspected audit firms. However, given that there is an extensive literature that provides evidence that Big N auditors provide higher quality than non-bign auditors (DeFond and Zhang 2014), we test whether our results are driven by the subset of Big N audit clients or non-bign clients by including a Big N control variable. We find no change in our inferences. Based on this finding, we did not consider it necessary to control for Big N in our main analyses. 21

24 otherwise. 21 The coefficient on PCAOB*ALL*FUTUREFIND estimates the difference in response to a PCAOB finding for firms that receive another finding in the next three years, compared to the response for firms that successfully avoid a subsequent finding. Table 5 presents the results of this supplemental analysis. We find a positive and significant coefficient on PCAOB*ALL*FUTUREFIND (β=0.152, p-value<0.05) indicating that the ALL of clients of auditors that receive a second finding from the PCAOB has a greater association with future charge-offs, compared to the clients of firms that avoid a future finding. This finding indicates that the ALL accuracy for the clients of auditors that receive a subsequent finding is higher compared to those that do not receive a second inspection finding, suggesting that movement in a more conservative direction is an effective strategy for audit firms to avoid PCAOB inspection findings. This is consistent with the suggestion by Lennox and Pittman (2010) that the inspection findings of the PCAOB do not lead to an increase in perceived audit quality. Our evidence related to the ALL suggests audit firms may overreact and encourage clients to record overly conservative allowances rather than using targeted procedures to improve accuracy. Comparison of Public vs Private Clients: H2 We test H2 by examining if the impact of a PCAOB finding related to inadequate procedures in the audit of the ALL differs for public and private clients. We first separately examine public and nonpublic banks. Table 6, column (1) presents the results of estimation of Model (1) for public banks. Consistent with our results from H1, the coefficient on PCAOB*ALL is negative and significant (β3=-0.185, p-value <0.01), indicating a decrease in ALL accuracy following an auditor s receipt of a PCAOB inspection finding. In contrast, the coefficient on PCAOB*ALL for 21 We are unable to estimate a bank-level model of the likelihood of a future inspection finding as the client firms are not identified in PCAOB inspection reports. 22

25 private clients, presented in column (2), is insignificant (p-value>0.10), suggesting that a PCAOB inspection finding does not significantly affect the accuracy of private bank clients ALL. We then directly compare the effect of a PCAOB inspection finding on an audit firms public and private clients using a difference-in-differences specification. Table 6, column (3) presents the results of estimating Model (3). The coefficient on PUBLIC*ALL is positive and significant (δ4=0.173, p-value <0.01), suggesting that the ALL better predicts future charge-offs at public banks compared to private banks, when PCAOB = 0. The coefficient on PCAOB*ALL is insignificant (p-value>0.10), which is consistent with the results in column (2) indicating that an auditor s receipt of a PCAOB finding does not affect the ALL accuracy of its private clients. In contrast, the coefficient on PUBLIC*PCAOB*ALL is negative and significant (δ7=-0.136, p-value <0.01), indicating that the accuracy of the ALL decreases more for public banks compared to private banks when PCAOB = 1. Table 7, columns (1) and (2) present the results of estimating Model (2) to examine the ALL error for public and private banks separately. Consistent with the results from testing H2 in Table 5, we find a positive and marginally significant coefficient on PCAOB for public clients (λ1=0.001, p-value <0.10), and no significance on the coefficient on PCAOB for private clients (p-value >0.10). Column (3) presents the results of estimating Model (4), and consistent with the findings in columns (1) and (2), we find a positive and significant coefficient on PUBLIC*PCAOB (π3 = 0.001, p-value<0.10). Together, these findings indicate that when an auditor receives a PCAOB inspection finding, the ALL error for the auditor s public clients increases, while a PCAOB inspection finding does not have a significant impact on private clients. Overall, the results of our test of H2 indicate that PCAOB inspection findings significantly affect the ALL accuracy of auditor s public clients, but do not have a significant impact on the 23