Discussion of the relation between auditors' fees for nonaudit services and earnings management.

I. INTRODUCTION

Frankel et al. (2002) (hereafter FJN) present a timely paper using a new data set to test several propositions, including one suggested by the Securities and Exchange Commission (SEC) about the relation between nonaudit fees paid by a registrant to its auditor and the registrant's

We organize our discussion around a "Predictive Validity Model" (displayed later in Figure 1) that shows conceptual relations between independent and dependent variables, as well as their operational measures and treatment of other causal factors. We use this model to describe and evaluate the authors' approach, which blends complex behavioral and contractual relations with aggregate observable fee, accounting, and market price data. We hope that the discussion will encourage others to conduct additional research to address the important questions raised by both FJN and the SEC.

[FIGURE 1 OMITTED]

II. FJN's APPROACH

The underlying proposition examined by FJN is that the strength of the audit firm's economic dependence on the client, or economic bond, consciously or unconsciously reduces the auditor's independence, or willingness to resist client-induced biases in the reporting process. The lack of auditor independence results in lower earnings quality. The underlying cause (the economic bond) and the effect (lowered earnings quality) are the two theoretical constructs examined. Both are defined conceptually, and are not directly observable. Their relationship forms Link 1 in the first (or concepts) row of the Predictive Validity Model shown in Figure 1. (1)

At the operational level, FJN first adopt the SEC's proposition that the relevant operational measure of the auditor's economic bond to the client is the ratio of nonaudit fees to total fees received from a client (denoted FEERATIO). They also measure economic bond as the rank of nonaudit fees across clients for an audit firm in their sample (denoted RANKNON), assuming that the rank of the magnitude of nonaudit fees (irrespective of the magnitude of audit fees) compromises auditor independence. The rank of audit fees (RANKAUD) is also tracked, but is not presumed to affect auditor independence. The relation between the concept economic bond and its operational measures is Link 2 in the model.

The authors focus on one element of earnings quality, "management's deliberate intervention in the earnings process in the form of earnings management," and use three types of surrogate measures of earnings management: (1) P(benchmark) defined as the proportion of clients just meeting or beating earnings benchmarks (consensus analyst forecasts and prior period earnings), (2) ABSDACC or absolute discretionary accruals, and (3) ARET, the market price reaction to disclosure of fees, which implies a price reduction for presumed earnings management. These variables comprise the operational dependent variables. The relation between earnings quality and its three measures is Link 3 in the Predictive Validity Model. Other auditor and company characteristics comprise control variables whose relation to the operational dependent variables is Link 4 in the model.

As in all empirical studies, FJN draw inferences about Link 1, the association between concepts, by assessing Link 5, the association between operational measures of those concepts. The validity of this inference depends on the validity of Links 2 and 3, the operationalization of the concepts, and whether the effects in Link 4 are adequately controlled. Thus, most concerns about FJN's inferences, and directions for future research, can be classified by their relation to:

* Link 1--inconsistency in theory.

* Links 2 and 3--faulty operationalization.

* Link 4--failure to control for other potential causes.

* Link 5--data or statistical problems.

In the discussion to follow, we categorize our comments and those of the Conference participants in terms of the five links in Figure 1.

Link 1: Economic Bond and Earnings Quality

FJN propose that a stronger economic bond between audit/consulting firm and client will reduce the quality of reported earnings, other things equal. Auditor independence impairment, or reduced willingness to resist client-induced biases in reported earnings, is the presumed intervening process through which economic bond affects earnings quality, but is not measured in this study. As a consequence, much of FJN's discussion of SEC independence rules is not directly relevant to the issues examined in the paper. (2)

Better conceptual definitions can improve measurement of the concepts in all empirical work. We identify three issues related to the definition of economic bond that deserve increased attention in future research. First, FJN do not appear to consider whether high audit fees might also increase the economic bond. However, because the authors present no reason to distinguish between audit and nonandit fees (or profits) as determinants of economic bond, others might reasonably predict that operational measures of economic bond such as nonaudit fees, audit fees, and their sum would all be correlated positively with operational measures of earnings quality. This is a serious conceptual omission.

Second, the concept of economic bond could be refined to distinguish between total fees (or profits) and fees in excess of those expected from observable firm circumstances. As the authors suggest, both acquisitions and complex tax environments increase expected demands for both audit and consulting services. More insidious effects on economic bond may result from unexpected nonaudit and audit fees that may more accurately be likened to attempted bribes. FJN's estimation of unexpected nonaudit fees in the Appendix suggests they have considered this issue, but the same approach was not applied to audit fees and total fees, and the associations between unexpected fees and the earnings management measures were not proposed or reported. Unexpected fees may also better capture the profitability of the services provided.

The benefits to building better models of expected audit fees can be seen in an analysis of Enron's fee disclosures (audit fees of $25 million and nonaudit fees of $27 million). Enron's resulting FEERATIO of .52 is close to the median (.51) reported by FJN in Table 2, but the audit fee is more than 250 percent of the expected fee computed using a 1997 survey of audit fees as a proportion of client revenue. (3) Thus, while Enron would be rated as about average in economic bond based on FEERATIO, the unexpected audit fee may imply a much stronger economic bond, one that would remain even in the absence of nonaudit fees. The result of an expected fees exercise would be estimates of the unexpected or "excess" profitability of an audit client, irrespective of whether the excess results from a particular service or client-billing peculiarities of the audit firm.

The third conceptual issue related to economic bond that should be reconsidered is the form of the expected relationship between audit and nonaudit fees as components of the economic bond. As the results presented in FJN suggest, this relationship could be complex. For example, a willingness to lower audit fees to maximize potential total client fees or profits could cause unusual conditional relationships between the two fee components and economic bond. Other conceptual interactions probably exist.

Concerning the conceptual definition of earnings quality, we agree with the authors' choice of the most often studied element of earnings quality, earnings management. This choice allows the authors to compare their results with existing studies. Future research can broaden the number of earnings quality dimensions analyzed.

Link 2: Economic Bond and Nonaudit Fee Measures

Because FJN do not distinguish conceptually between audit fee and nonaudit fee sources of economic bond, both their interpretation of FEERATIO and their failure to include RANKNON and RANKAUD as components of economic bond are problematic. Since FEERATIO increases when audit fees decrease and RANKAUD is negatively related to earnings management, the results for FEERATIO, and for RANKNON and RANKAUD, are internally consistent (and consistent with prior research). They are inconsistent, however, with FJN's theory if the economic loss (bond) resulting from a client departure would be the total fees paid to the auditor. The lack of association between total fees and earnings management proxies adds to questions about the interpretability of the results.

Conference participants suggested a number of possible avenues for investigating these theory-inconsistent results. Some comments focused on other concepts for which RANKAUD may be proxying, including size (4) and audit fee low-balling. If larger clients (which have larger audit fees) are less likely to manage earnings, then audit fees may be negatively associated with earnings management. Similarly, a negative association could also result if low-bailing audit fees induces more auditor acceptance of earnings management attempts (Beeler and Hunton 2001).

Participants also suggested that future studies examine the role of nonaudit fees in audit partner compensation and whether the economic bond operates at the office level, as opposed to the firm level implicit in the RANKAUD and RANKNON variables. For example, a client with a low value for RANKNON or RANKAUD may represent a huge proportion of fees for a small office of an audit firm.

A number of questions were raised about the SEC's categorization of fees into the audit and nonaudit categories. It was suggested that, because 2000 was the first year for fee disclosure, the SEC categories might not be well understood by all registrants, leading to differences in their application across registrants both within a particular audit firm's clients and perhaps across audit firms (although at least one audit firm provided detailed instructions on how to break down nonaudit fees). These issues argue for an analysis by audit firm, including models that allow slopes to differ by audit firm as well as intercepts (now included by FJN in Table 7).

Measurement concerns were also raised about items normally related to audit services (e.g., pension plan audits, consultation on accounting standards or transactions, and tax work) that are included in the SEC definition of nonaudit fees. In addition, comparisons of U.S. results with studies in other countries will require reconciliation of the differences in definitions of nonaudit fees across countries. Finally, interpretation of results may be impaired by the use of a single year (the first year the disclosures were required, and concentration of almost 60 percent of the proxies filed in April 2001) that may be idiosyncratic with respect to the measured relations. While each of these concerns may add error to the analyses, no Conference participant suggested specific directional biases that may have been induced.

Link 3: Earnings Quality and Empirical Measures

FJN measure earnings management using two measures popular in archival studies: the probability of meeting or beating benchmarks (P(benchmark)), and the absolute value of discretionary accruals (ABSDACC). A strength of the paper is the fact that these two different types of measures provide some convergent results, reducing the possibility that the results are due to method bias.

Most Conference participants' concerns about measurement of earnings quality pointed to inherent contradictions in the earnings management literature, as opposed to specific problems in FJN's study. For example, the benchmark measure categorizes all firms that meet or beat the consensus forecast as having low-quality earnings, regardless of whether the goal was achieved through earnings management, expectations management, reduction of uncertainty, or improvements in operations. Conversely, a firm that is consistently well below or well above the consensus forecast will be deemed to have high-quality earnings by this criterion, even though recorded results may be far from the consensus due to a weak accounting system or attempts to manipulate earnings by large amounts. Similarly, discretionary accruals categorize, as earnings managers, firms that engage in transactions that take a long time to complete, or that involve complex judgments and estimates.

Finally, short-window price reactions to the release of proxy statements confound the release of absolute and relative fee information with other conditions or events noted in the proxy statement as well as other contemporaneous events occurring on the relatively few proxy filing dates. Also, high nonaudit fees may cause a negative price reaction because of the inference that the company "needed" a lot of consulting, instead of an inference that the audit and the resulting earnings were of low quality.

Link 4: Control Variables

Rather than using a system of equations (one to model fees, and one to model surrogates for earnings quality), FJN use a single equation simultaneously to model earnings quality surrogates and to control for nonaudit and audit fee determinants. They include a variety of control variables in the earnings quality regressions. Some of the control variables are alleged to decrease the operational dependent variable, earnings quality (litigation risk, growth, institutional ownership, and a reported profit vs. loss). Other included control variables are alleged to increase the operational independent variable, purchases of nonaudit services (acquisition and financing activity and poor performance). Still others measure additional factors that increase the conceptual independent variable, economic bond (longer auditor tenure and smaller auditor size). Firms with higher cash flows from operations are seen as less likely to manage earnings and more able to afford nonaudit services. So this variable is thought to be a determinant of both the operational dependent and independent variables. And, of course, client size is included as a control because it is related to most things.

Controlling for determinants of the operational independent (fee) variables in a single equation modeling earnings management makes the interpretation of the coefficients on the fee variables problematic. Including these variables could hide, accentuate, or even reverse the sign of associations between the fee variables and the earnings management variables. At the least, their inclusion eliminates sources of variation in nonaudit and audit fees that could affect earnings management.

Link 5: Nonaudit Fees and Earnings Quality Measures

Participants requested additional analyses to allow some assessment of the stability of the basic results, first among these being the correlation matrix for all of the variables (now added in FJN's Table 4). Second, in addition to the regressions provided, Conference participants suggested including the regressions with only the controls for the dependent variable followed by a regression including these same controls and the fee variable(s), to judge the marginal predictive power of the fee variables (analyses are now added to FJN's Tables 5-8).

As to the single equation approach to modeling, participants suggested separate but related models of determinants of both fees and earnings management in a system of equations. One approach was to include exogenous variables suspected to affect the independent variable nonaudit fees in a two-stage regression, using the predicted value of the nonaudit fee variable as an instrumental variable in the second stage. These same types of analyses were suggested for the total fee measure, RANKTOT, to probe the apparent lack of association between this measure of economic bond and earnings management. Other participants suggested a simultaneous-equations approach to account for the independent and dependent variables whose values may be jointly determined.

Participants also raised completeness issues. Several suggested incorporating as control variables corporate governance factors previously found to be related to earnings management. In addition, omissions in the discussion of the implications of certain results were noted, the most important being the lack of association of earnings management with RANKTOT and the negative association with RANKAUD. Several participants wanted to know why the economic bond argument should apply only to RANKNON and not to RANKTOT, and perhaps apply negatively to RANKAUD. Also noted was the lack of explanation for the disappearance of the nonaudit fee relation for larger firms. Is this an anomaly, or does the economic bond theory predict a conditional size relation? Several participants also questioned FJN's emphasis on the significant findings using ABSDACC and the earning forecast benchmark, and the lack of emphasis on the nonsignificant findings using the earnings change benchmark, where, of the fee measures, only the RANKAUD variable remain significant.

FJN conduct three tests of the stock price effects of public announcement of fees via registrant proxy filings. The authors take a short-window events study approach at the announcement date and predict that stock prices will react negatively to high, and higher than expected, nonaudit fees. The coefficients for the FEERATIO, and the RANKNON and RANKAUD models are insignificantly different from zero, suggesting no measurable effects. [R.sup.2]s for the two models are .01. The third model, using a within-sample calculation of unexpected nonaudit fees (FEERESIDUAL), also has an [R.sup.2] of .01, but the coefficient is significantly negative.

FJN interpret their overall stock price reaction results as "modest evidence consistent with investors reducing valuations as compensation for a perceived reduction in audit quality." Taken at face value, some participants wondered why unexpected FEERESIDUAL should compromise audit quality and lead to price declines while the nonaudit fee and audit fee measured by FEERATIO and both RANKNON and RANKAUD should not.

III. CONCLUSIONS

Overall, FJN have presented an ambitious analysis of a new data set and have made an important first pass at addressing some issues of great interest to regulators, practitioners, and scholars. Their data gathering was as comprehensive as was possible at the time, and their analysis provides documentation of empirical relations that suggest a basis for further investigation. The use of multiple measures helps in understanding possible causal links related to the constructs of economic bond and earnings quality. At the same time, however, various research design and measurement issues discussed above make their results difficult to interpret.

Issues related to the conceptual definition of economic bond cause the greatest problems. First among these is the lack of association between RANKTOT (the rank of total fees within each audit firm) and the earnings management measures. Second is the statistically negative relation between RANKAUD and the earnings management measures. Higher fees of either kind would presumably increase the economic bond. The authors' explanation that the negative relation to audit fee has been found in other studies does not explain why the relation should be negative in the present circumstance, nor help with the interpretation of its implications for their economic bond theory.

Finally, we suggest that the conceptual relations between management and auditor incentives and earnings management are more complex than suggested in FJN. The model in Figure 2 recognizes that management incentives affect both the magnitude of earnings management attempts and auditor incentives to resist those attempts (Nelson et al. 2002).

[FIGURE 2 OMITTED]

The model points out that an empirical association between fees and earnings management could result from firms with higher fees attempting more earnings management and/ or from reduced independence, neither of which is observable in archival studies. This suggests that careful modeling of both management and auditor incentives, combined with careful estimation of expected audit and nonaudit fees, may provide useful directions for future research.

(1) The Predictive Validity Model is based on Runkel and McGrath (1972) and Libby (1981), who used the model to analyze accounting research questions.

(2) In the final section, we discuss, as a suggestion for future research, a model that makes explicit some components implicit in the FJN independence argument.

(3) Based on a survey of 340 members of Manufacturers Alliance (Arlington, VA) reported in Public Accounting Report, November 30, 1997, the audit fee for firms with $20 billion or more in revenues was $92 per million in revenues. Enron's $100 billion in revenues would imply an audit fee of $9,200,000.

(4) The proxy used for size may not be fully effective.

REFERENCES

Beeler, J. D., and J. E. Hunton. 2001. Contingent economic rents: Precursors to predecisional distortion of client information. Working Paper, University of South Florida.

Frankel, R. M., M. F. Johnson, and K. K. Nelson. 2002. The relation between auditors' fees for nonaudit services and earnings management. The Accounting Review (Supplement): 71-106.

Libby, R. 1981. Accounting and Human Information Processing: Theory and Applications. Englewood Cliffs, NJ: Prentice Hall, Inc.

Nelson, M., J. Elliott, and R. Tarpley. 2002. Evidence from auditors about managers' and auditors' earnings-management decisions. The Accounting Review (Supplement): 175-202.

Runkel, P., and J. McGrath. 1972. Research on Human Behavior-A Systematic Guide to Method. New York, NY: Holt, Rinehart, and Winston, Inc.

William R. Kinney, Jr.
The University of Texas at Austin

Robert Libby
Cornell University