Audit fee residuals: costs or rents?

Rajib Doogar rdoogar@uwb.edu

University of Washington Bothell

Padmakumar Sivadasan psivadas@tulane.edu

Tulane University

Ira Solomon* isolomon@tulane.edu

Tulane University

October 14, 2013 *Corresponding author: A.B. Freeman School of Business, 7 McAlister Drive, New Orleans, LA 70118. Phone: (504) 865-5407

Audit fee residuals: costs or rents?

Abstract In contemporary research residual audit fees are used as proxies for researcher-unobserved audit costs as well as for auditor rents (profits in excess of audit costs). As a result, researchers draw conflicting policy implications from the same research finding. We show analytically how differences in residual fee persistence across continuing and new audit engagements can reveal the extent to which residual fees consist of unobserved audit costs, auditor rents, and transient, i.e., noise elements. Empirically, in a large sample of U.S. public company audit engagements, we find evidence indicating that fee residuals largely consist of researcher-unobserved audit production costs and are likely to be poor proxies for rents. This finding provides valuable guidance for how fee residuals should be used in future research, indicates promising avenues for future audit fee research, improves the ability to predict expected audit fees from past fee data and clarifies the policy implications that can reliably be drawn from extant and future fee-residuals-based research. Keywords Audit fees, audit fee residuals, auditor rents, audit costs, fee discounting. JEL Classification L11 L12 M41 M42 M48

1. Introduction

Although audit fee residuals (hereafter, fee residuals) are widely studied in contemporary

accounting research, their proper interpretation is far from settled.1 Some studies use fee

residuals as proxies for auditor rents, i.e., audit fees in excess of normal audit production costs,

while others use them as proxies for researcher-unobserved audit production costs (unobserved

costs).2 However, as Francis (2011:138) notes, “… we have no idea if fee residuals measure a

threat to independence. … (they) … might simply capture abnormally high audit effort or the

auditor’s pricing of (unobserved) client risk characteristics.” Our study presents large-sample

empirical evidence that helps resolve this conflict about the interpretation of fee residuals.

Uncertainty about what fee residuals measure complicates interpretation of the findings of

fee-residuals-based studies. Consider a finding that fee residuals are negatively associated with

auditee financial reporting quality. If one views fee residuals as a proxy for unobserved costs,

this finding would be read as (reassuring) evidence of audit costs (auditor effort) rising with

either engagement risk or auditee demand for external verification. By contrast, if one views

residual fees as a proxy for rents, this finding would be read as (troubling) evidence of impaired

auditor independence. And, if one views fee residuals as a mixture of unobserved costs and

rents, the evidence would be read as inconclusive. In sum, depending on the researcher’s choice,

the same research finding can be used to either support, oppose or be agnostic about the need for

regulatory interventions in the audit market! By elucidating what fee residuals measure, our

study permits reliable interpretation of the findings of fee residuals-based studies, placing that

substantial branch of the contemporary research literature on a much sounder footing.

1 Fee residuals are computed as the difference between actual audit fees and predicted audit fees from a regression of audit fees on engagement attributes known to affect audit production costs (see Simunic, 1980). 2 As in Simunic (1980), in this study, normal audit production costs include direct costs of audit production as well as the normal cost of auditor risk-bearing and reflect auditor effort while auditor rents represent fees in excess of all production costs, and therefore a potential threat to auditor independence (DeAngelo, 1981).

2

We show analytically that, in competitive audit markets, the extent to which audit fee

residuals consist of unobserved costs or rents can be discerned from their differential ability to

explain audit fees charged in subsequent years by continuing and new auditors (hereafter, fee

residual persistence). If fee residuals are composed mostly of unobserved costs, they will persist

(i.e. explain subsequent year audit fees) at close to a dollar for dollar rate in both continuing and

new engagements. If, on the other hand, fee residuals are mostly rents they will persist in

continuing engagements, but will be discounted during auditor transitions. The reason for the

discounting is that competition to become the new incumbent and earn future rents will lead new

auditors to rebate the capitalized value of all rents back to the auditee (DeAngelo 1981; Watts

and Zimmerman 1986: 314-315; Dye 1991; Kanodia and Mukherjee 1994).

We report three key empirical findings. First, in continuing Big N audit engagements, i.e.,

when there is no auditor change, after controlling for other fee determinants, a dollar of the

preceding year’s fee residual explains about seventy to eighty cents of the current fee. Second, in

lateral Big N auditor transitions a dollar of the preceding year’s fee residuals explains between

seventy to seventy-five cents of the fee charged by a successor Big N auditor. Third, in Big N to

non-Big-N auditor transitions a dollar of the preceding year’s fee residual explains about sixty to

seventy-five cents of the fee paid to a successor non-Big-N auditor. The first finding indicates

that fee residuals contain a large persistent component (albeit one that could be costs, rents or an

admixture of costs and rents). 3 The second finding shows that this persistent component largely

consist of costs common to all Big N auditors. The third finding further indicates that about 80%

of that common component consists of costs common to both Big N and non-Big-N auditors.

These findings are robust to alternative approaches to computing fee residuals.

3 The remaining twenty to thirty cents of the residual fee is transient, i.e., specific to the current period. It therefore is unlikely to reflect any persistent auditor rents (i.e., the kind of rents thought to impair auditor independence).

3

Our findings have four key implications. First, future research should use fee residuals as

proxies for researcher-unobserved audit costs common to all auditors. Second, given their

significant incremental explanatory power in the fee model, elucidating the factors that drive fee

residuals offers a promising avenue for future research. Third, given their high persistence,

lagged residual audit fees can be used to improve computations of expected audit fees in settings

where the researcher is interested only in predicting (rather than explaining) audit fees. Fourth,

when drawing policy implications, the association between fee residuals and measures of audit

quality should be interpreted as speaking to the effects of auditor effort rather than of auditor-

auditee economic bonding.

The rest of this paper is organized as follows. In section 2 we review prior literature, present

an analytical model of residual audit fee persistence in continuing and new audit engagements,

and outline our strategy for identifying rents in fee residuals. In section 3 we describe research

methods and the data used. In section 4 we report results and in section 5 we offer some

concluding remarks.

2. The analytics of fee residual persistence

2.1 Computing and interpreting fee residuals

Fee residuals are usually computed as the difference between actual audit fees and expected

audit fees predicted by a Simunic (1980) style audit fee model of the form

(1)

where yit is the audit fee for engagement i in period t, xit is a vector of engagement characteristics

(e.g., auditee size, complexity, and riskiness) and is an error term.

The fee residual computed from this model, , can be thought of as the sum of the effects

on yit of (i) any researcher-unobserved production costs, , i.e., of fee determinants omitted

4

from xit, (ii) of any auditor rents, ρit, and (iii) of a pure noise component ηit.4 Studies that adopt a

rent-centric interpretation of the fee residual effectively assert that the cost and noise components

of are both negligible. Studies that adopt a production cost–centric interpretation of the fee

residual, effectively assert that the rent and noise components of are both negligible. Lastly,

studies that explicitly or implicitly adopt an agnostic interpretation assert that all three

components of are non-negligible.

Our study is motivated by the observation that not all three readings of the fee residual can

simultaneously correctly describe the same data set and that depending on the meaning a

researcher ascribes to , an observed association between and any other measure of interest

(e.g. audit quality, financial reporting quality, auditee governance, auditee cost of capital) can be

given rather different policy interpretations. For instance, a negative association between and

auditee cost of capital (cf. Ball et al. 2012) can be interpreted as evidence that

(1) auditee cost of capital is lower when auditor rents are larger (rent-centric reading), or

(2) auditee cost of capital is lower when audits are more costly (cost-centric reading), or

(3) the inimical influence on auditor independence of the rent component of the fee

residual is more than offset by the beneficial impact of greater auditor due diligence

reflect in the cost component of the fee residual (agnostic reading).

Consequently, were such evidence to be invoked in a policy debate about limiting the rents

that auditors can earn (e.g., by capping auditor tenure), it would be possible for different interest

groups to point to the same finding (a negative association between and auditee cost of

capital), to argue that capping auditor rents would be undesirable (reading 1), to argue that the

evidence does not really speak to rents since the rent component of is likely to be negligible

4 Statistically, the noise term subsumes both transient impacts on fees as well as any estimation error.

5

(reading 2), or to argue that while may contain both rents and unobserved production costs

(reading 3), the evidence at hand does not support a regulatory intervention.

DeFond et al. (2002) is an early example of the use of fee residuals to measure auditor rents

(termed auditor-auditee economic bonding or excess profits).5 An influential commentary by

Kinney and Libby (2002: 109-110) also advocates using fee residuals to measure the economic

bond between auditor and auditee. Over the subsequent decade, the use of fee residuals to

surrogate for auditor rents has become quite popular. Notable studies include Srinidhi and Gul

(2007) who find no systematic association between fee residuals and earnings quality, Hope and

Langli (2010) who find no association between large (and positive) fee residuals and auditors’

propensity to issue going concern opinions for Norwegian companies, Kanagaretnam et al.

(2010) who find no association between positive fee residuals and under-provision of loan losses

of banks and Choi et al. (2010), who find a positive association between (positive) fee residuals

and larger abnormal accruals (lower financial reporting quality).

It is important to note at this point that many of these rent-centric studies have used multiple

measures of auditor rents, including non-audit fee residuals, total fee (the sum of audit and non-

audit fees) residuals as well as the magnitudes of the audit, non-audit, and total fees paid to

auditors. We would, therefore, be remiss if we did not caution the reader that our results should

be interpreted as questioning only those inferences drawn in prior research that are founded on

the use of audit fee residuals as a measure of auditor rents. Inferences based on other measures

used in those studies are beyond the scope of our present investigation.6

5 DeFond et al. (2002) use multiple measures, residual audit fees, residual non-audit fees and residual total fees as surrogates for auditor rents. They find no association between these measures and auditor propensity to issue going concern opinions, which they interpret as evidence that auditor rents do not impair auditor independence. 6 Notwithstanding the limited focus of our investigation relative to these prior studies, our findings have broad relevance for this stream of literature. First, if audit fee residuals are noisy proxies for auditor rents, potentially, so are any measures that are based on sums or ratios that involve either audit fee residuals, or more generally, audit fees. Second, to our knowledge there is no study that has investigated the extent to which non-audit fee residuals

6

In sharp contrast to the rent-centric interpretation of fee residuals as excess profits, Ettredge

and Greenberg (1990) view fee residuals as production efficiency measures: positive (negative)

residuals indicate that the auditor is an inefficient (efficient) producer relative to other auditors.

Hribar et al. (2010) view fee residuals as a measure of extra audit effort or risk premium charged

by the auditor when faced with poor auditee accounting quality. Similarly, Ball et al. (2012)

view higher fee residuals as a proxy for higher auditee demand for financial statement

verification and report that voluntary disclosures by auditees that have higher fee residuals are

both more accurate and more credible to investors. A cost-centric reading is also suggested by

the studies that use private data sets (e.g., O’Keefe et al. 1994; Bell et al. 2001, 2008) and

document many, not publicly available, drivers of audit labor usage and mix (e.g., number of

audit reports, number of auditee business locations, and perceived auditor business risk).7

A relatively small number of studies adopt an agnostic view of fee residuals. For instance,

Higgs and Skantz (2006) argue that fee residuals may be composed of both auditor rents and

unobserved audit costs. However, in their empirical analyses they do not attempt to separate the

fee residual into the rent and cost components.

themselves reflect non-audit costs or rents. In fact, our findings with respect to audit fee residuals suggest the need for future research along the lines of the present study to disambiguate the extent to which non-audit fee residuals contain researcher-unobserved non-audit services costs and rents. We do not investigate this issue because it would add to an already lengthy and complex analysis, and more importantly data on fees paid to auditors for non-audit services are available only for a very limited period. Third, the 2000 SEC Auditor Independence Rules having limited the provision of most non-audit services for auditee fiscal years ending after 31 December 2002, future research must necessarily confine attention to audit fees and audit fee residuals in order to measure auditor rents. 7 To the extent that all of these researcher-unobserved production-cost related factors affect audit fees, their overall impact on fees will be reflected in the error term of the fee model, i.e., in the residual fee. In the auditing context, many of the researcher-unobserved influences are likely to be correlated with observable determinants. The number of auditee business locations (a variable likely to have a positive impact on audit fees) is not publicly observable and is therefore excluded from most estimations of equation (1) (i.e., is a component of ). However, the number of auditee locations is also likely to (a) vary slowly over time and (b) be highly correlated with auditee size. Auditor litigation risk, another driver of the extent of audit procedures and of the risk premium charged by the auditor is also likely to be highly persistent and associated with observable auditee characteristics. As we explain in Section III, our statistical procedures account for such (relatively) time-invariant omitted correlated effects.

7

In sum, prior research has (a) made extensive use of fee residuals as a research construct, (b)

adopted conflicting interpretations of the fee residual, leading to (c) different interpretations of

similar empirical findings. In Appendix A we review some of the key studies that offer rent-

centric, cost-centric or agnostic interpretations of the fee residual, reporting in the authors’ words

the interpretations each study places on the fee residual and the resulting policy implications they

draw. In so doing, we establish, first, that the three streams of research appear to have evolved in

parallel with very little evidence of inter-play. Second, we elucidate how audit costs, auditor risk

premium, and auditor rents, key constructs for our study have been defined in prior research.

Our discomfort with the divergent readings of fee residuals being offered in parallel streams

of contemporary research studies is not isolated. For instance, Francis (2011: 138) observes that:

“I am skeptical of the use of abnormal fees to measure auditor independence because we have no idea if fee residuals measure a threat to independence. Alternatively abnormal audit fees might simply capture abnormally high audit effort or the auditor’s pricing of (unobserved) client risk characteristics.”

The purpose of our study is resolve this ambiguity regarding the interpretation of fee

residuals by investigating the following question: Do fee residuals reflect primarily audit costs or

primarily auditor rents or are they best viewed as an admixture consisting of significant

components of both costs and rents (and therefore, without further refinement, are not a good

proxy for either)?

2.2 Identifying the substantive information content of fee residuals

As noted earlier, in principle, fee residuals from equation (1) reflect the sum of researcher-

unobserved production costs (including normal risk premia), auditor rents, and noise. Further,

both the cost and rent component can contain elements of persistent factors (costs and rents) that

are common to all auditors, common to some auditors or are specific to the incumbent auditor.

As a result the residual from equation (1) can be recast as

8

(2)

where cc, ic, cr and ir represent, respectively, common costs, idiosyncratic costs, common rents

and idiosyncratic rents (and therefore are indexed only by i) and the last component is a term that

captures the sum of purely transient components and noise.

In continuing (steady state) audit engagements, ceteris paribus, the first four components of

the fee residual should continue to determine audit fees in every period and should persist at a

close to dollar for dollar rate while the last term being the sum of transient and noise elements

would not be expected to affect future audit fees. Consequently, in continuing engagements, the

lagged fee residual can be expected to persist at a positive rate close to one, with the difference

from unity being the likely proportion of transient and noise elements in the fee residual.

In auditor transitions, however, matters can be expected to be quite different. By definition,

the first and third terms in equation (2) can be expected to persist while the second and fourth

terms should, much like the last (transient/noise) term, not affect the successor’s fee. As with

continuing engagements, the first term can be expected to affect the successor’s audit costs at a

dollar-for-dollar rate. The third term, which reflects the rents the new auditor also expects to

earn in future periods, however, will be treated very differently by the new auditor. Since the

new auditor expects to earn an annual stream of rents equal to cr, in theory, competition among

auditors to become the new incumbent – and thereby earn the stream of annual rents – will force

all bidding auditors to rebate the present value of the entire expected rent stream back to the

auditee (DeAngelo 1981; Watts and Zimmerman 1986: 314-315; Dye 1991; Kanodia and

Mukherjee 1994). Consequently, in pricing a new engagement, the successor auditor’s weights

on the various components of the predecessor’s fee residual will be 1 for the first component, 0

for the second and fourth components and a present value factor, –K, for the third component.

9

By the foregoing logic, the expected fee equations for continuing and new engagements are:

(3C)

(3N)

where equation (3C) is for continuing auditees and (3N) for new auditees. In particular, note that

the coefficients of cc, ic, cr and ir are each close to unity in the first equation and unity, zero, -K

and zero respectively in the second equation.

In practice, the individual components of cannot be observed by the researcher. However

equations (3C) and (3N) can be recast in terms of to obtain the specification:

(4C)

(4N)

where, as we show next, the coefficients γC and γN will be functions of the extent to which the

lagged fee residual is comprised of the four components cc, ic, cr and ir. As a result, empirical

estimates of γC and γN, the respective coefficients of lagged fee residual in the two equations

should shed light on the extent to which fee residuals consist of auditor rents.

First consider the case where consists entirely of idiosyncratic noise, i.e., forces that do

not affect subsequent audit production by any auditor. In this polar case, the weight of the

lagged fee residual in both equations (4C) and (4N) should be zero, i.e., the lagged fee residual

should not explain subsequent audit fees for either continuing or new engagements. Thus, a test

of γC=0 can shed light on the extent to which residual audit fees on continuing engagements

contain persistent influences (costs or rents).

Next consider the case where the lagged fee residual consists purely of common costs, i.e.,

, with the shares of the other components including noise being zero. In this case, by

10

equations (4C) and (4N), we would expect to observe γC = γN = 1.8 By the same logic, should

consist partly of cc and a noise term , we would expect to observe γC = γN = 1-α where

α is the fraction of represented by , i.e., / . Generalizing the setting to one

where the lagged fee residual is a mixture of common and idiosyncratic costs and noise, i.e.,

, for continuing engagements, one would expect to observe γC = 1-α as

before. However, since the idiosyncratic cost should not affect the successor, on new

engagements one would expect to observe γN = 1-β where / .9 Much the

same would be true were were to include some idiosyncratic rent, iri, as well since that

component too would be expected to affect the continuing auditor’s costs but not those of the

new auditor. However in this case, while one would still expect to observe γC = 1-α as before, γN

would be expected to have the value 1-θ where θ = / .

By contrast, in a scenario where , i.e., the lagged residual consists purely of

auditor rents common to all auditors, we would expect to observe γC =1 and γN = -K and, in the

general case, where , one would expect to observe γC = 1-α

and γN = 1-δ, where δ = 1 / . Note that unlike all of the

previous cases, the coefficient on cri in the expression for δ is 1 , as a result of which,

even when the rent component of the residual fee is relatively small, δ can take values smaller

8 In this case, note that (3C) and (3N) both reduce to which implies γC = γN = 1 in equations (4C) and (4N). The coefficients in each of the special cases discussed below follow directly from analogous substitutions of the respective definitions of into equations (3C) and (3N) 9 Note that β will vary with the fraction of that is comprised of common costs, cc. As the relative contribution of cc to increases (equivalently, as the relative contribution of ic decreases), β will decrease (γN will increase). Consequently, when cost structures are materially indistinguishable across auditors, i.e., ic=0, the expression for β reduces to that for α and one would expect to observe γC=γN. As a result, an empirical finding that in lateral Big N transitions, γC = γN would provide strong evidence that the fee residual contains no appreciable amount of rents. By contrast, since non-Big-N auditors are expected to have lower costs than Big N auditors, even when contains no rents, one would expect to observe β≥α, i.e., γC≤γN in Big N to non-Big-N transitions. The persistence of fee residuals across lateral Big N auditor transitions, therefore, constitutes a stronger test of the extent to which fee residuals contain rents than does its persistence across Big N to non-Big-N auditor transitions.

11

than minus one and the coefficient γN can be expected to be negative.10 Our empirical analyses,

described next, build on this intuition.

3. Methods and data

3.1 Computing residual fees

We compute the residual fee charged by an incumbent auditor, RFeeit, as the unexplained fee

from the model:11

Ln(Afeeit) = b0 + b1 Ln(TAit) + b2 Lossit + b3 ROAit + b4Leverageit + b5 InvRecit + b6 ForOpsit + b7 √Employeesit + b8 Nsegmentsit + b9 NewFinit + b10 ExtDistit + b11GCOit + b12 ICWeakit + b13Busyit + b14 Delayit + b15 Afilerit + Year Indicators + (2)

where Ln(Afee) is the natural logarithm of the audit fee paid to the external auditor for the fiscal

year in question (expressed in constant 1999 US dollars using the US Bureau of Labor Statistics (BLS) Consumer Price Index (CPI) series as deflator).

Ln(TA) is the natural logarithm of auditee total assets (TA) in constant 1999 US dollars. Loss takes the value 1 if the company reports negative net income, 0 otherwise. ROA is auditee return on assets (operating income after depreciation divided by total

assets) winsorized at 1%. Leverage is the ratio of total liabilities to total assets winsorized at 1%. InvRec is the sum of auditee inventory and receivables divided by auditee total assets. ForOps takes the value 1 if the auditee reports a foreign currency translation adjustment, 0

otherwise. √Employees is the square root of the number of auditee employees (measured in thousands). Nsegments is the number of business segments

10 Specifically, in a case where the lagged residual fee consists only of common costs and rents, the successor’s fee is expected to be 1×cc-K×cr, so that when cr/cc ≥1/K, the expected coefficient of lagged fee residuals will be non-positive! For example, when K=5, if rents comprise more than 17% of the fee residual (i.e., cc ≤ 83% of ), the ratio cr/cc will exceed five and the expected value of γN will be 1×0.83-5×0.17=-0.02 (we have set ici=iri=η=0 and cri/ϵi,t-1=0.17, so that δ= {0+0-(5+1)cri+0}/ϵi,t-1=-6×cri/ϵi,t-1.=1.02 and 1-δ=-0.02). If fee residuals were to largely be comprised of rents, an priori necessary condition for fee residuals to be a credible proxy for rents, e.g., if 70%-80% of the fee residual were rents, the coefficient of lagged fee residuals will be in the vicinity of minus 3 to minus 4. 11 Our specification includes all variables consistently found to be significant in prior research (cf. Hay et al. 2006). We exclude from the specification certain (primarily corporate governance-related) measures that have been used in prior research for three reasons: first, data on these are available only for a limited set of firms and their inclusion would severely reduce the power of our auditor switching analysis. Second, these measures have been found to add relatively little to overall model explanatory power (Hay et al. 2006: 175). Third, most of these measures are likely to be time-invariant for a given auditee and their effects (and that of other time-invariant engagement characteristics, e.g., auditee industry) will be captured in the firm fixed effect component of residual fees. The effects of all such persistent unobserved audit production cost related characteristics included in that residual are expected to persist (affect the fees charged by the successor auditor) while any rent component of the predecessor auditor’s residual fees will not so persist. See Appendix B for formal variable definitions.

12

NewFin takes the value 1 if sum of new equity and debt issue exceeds $50,000, 0 otherwise.12

ExtDist takes the value 1 if the absolute value of extraordinary items or discontinued operations exceeds $10,000, 0 otherwise.

GCO takes the value 1 if the auditor opinion for the fiscal year includes a going concern qualification, 0 otherwise.

ICWeak takes the value 1 if the auditor reports an internal control weakness, 0 otherwise. Busy takes the value 1 if the auditee fiscal year ends in December, 0 otherwise. Delay is the number of calendar days elapsed between the auditee’s fiscal year end and

the date of the audit opinion. Afiler takes the value 1 if auditee market value of equity at the end of the fiscal year

exceeds $75 million, 0 otherwise.

Coefficients for all of the explanatory variables are expected to have a positive sign, except

for ROA which is expected to have a negative sign since more profitable auditees pose lower

audit risk. Three aspects of the RFeeit computation are worthy of note. First, fee residuals for

year t are obtained by estimating model (2) for all years up to and including year t. This ensures

that the fee residual for each year is based on all data that would have been available to auditors

in year t+1 (our test year) to prepare their bids, precluding introduction of ex-post information or

hindsight bias in the fee model coefficients (and, consequently, in fee residuals).13

Second, model (2) is estimated using only current and past data for continuing engagements:

any year during which an auditor change occurs is excluded from the estimation process. This

procedure ensures that the fee residual is computed from a model calibrated upon engagements

where the auditor is most likely to earn rents. Third, since Big N and Non-Big-N audit

production and fee functions differ significantly (Chaney et al. 2004; Sankaraguruswamy and

Whisenant 2009), we estimate model (2) separately, each year, for Big N and Non-Big-N

auditees. The second and third restrictions ensure that when an auditee switches from a Big N 12 New security issuances often result in an increase in the total number of reports rendered by the auditor, a key driver of audit labor usage (O’Keefe et al. 1994, Bell et al. 2008). In a recent study that uses publicly available data Sankaraguruswamy and Whisenant (2009, Table 3) find NewFin significantly explains audit fees. Excluding NewFin from the analysis does not materially alter any of our principal inferences. 13 While we believe that estimating fee model residuals using all available years of data is inappropriate (due to the potential hindsight bias introduced by use of future fee data), in unreported analyses we find that using fee residuals computed from such a pooled model does not materially alter our principal inferences.

13

(Non-Big-N) auditor to a Non-Big-N (Big N) auditor, fee residuals for all subsequent years are

computed using only data from periods audited by a comparable (Non-Big-N or Big N) auditor.

As noted earlier, the fee residual from this model will capture the combined influence of all

researcher-unobserved factors that influence audit fees, i.e., fee determinants excluded from

equation (2) as well as estimation error (noise).14 Following the approach used in GAO (2003),

we account for the (unknown) correlations between omitted and observed fee determinants and

obtain residual fees via fixed-effects panel estimation of model (2).15 Since many prior research

studies obtain fee residuals from pooled OLS estimation of model (2), i.e., without correcting for

omitted correlated variables, we also replicate and confirm the findings of the persistence tests

using pooled OLS residuals (see Section 4.3, Additional analyses).16

3.2 Preliminary: Testing first-year fee discounts

To investigate competition in the market for new auditees, we first estimate, separately for

Big N and Non-Big-N auditees, the model:

14 Potential fee determinants on which large-sample data are not presently publicly available to researchers (and, consequently, excluded from the model specification used in this study) include audit cost drivers such as number of audit reports issued and/or number of locations audited, measures of auditor assessments of business risk and auditor reliance on auditee internal controls (O’Keefe et al. 1994; Bell et al. 2008), intensity of auditee demand for credible auditing (Ball et al. 2012) , the nature and extent to which the auditor encounters unanticipated problems (Hribar et al. 2012), and measures of engagement efficiency (Ettredge and Greenberg 1980). Elucidating the large sample impact of each of these presently immeasurable audit cost drivers offers a promising area of future research. Note, however, that to the extent that these factors are likely to vary slowly over time, our panel data research design explicitly controls for the impact of such omitted variables. 15 See, Wooldridge (2002: 247-249). In the auditing context, Francis (2011: 132) also recommends using panel data methods to control for the influence of firm-specific omitted variables. An equivalent, albeit computationally more cumbersome, approach is to use pooled OLS and a full set of engagement-level indicators {ci}, i.e., to estimate the model and compute fee residuals as the sum of the estimated auditee fixed effects and the pure error term . 16 Note that both the panel and pooled OLS specification assume coefficient stationarity over time. Alternatively, we could estimate equation (2) year-by-year, thereby allowing coefficients to vary over time. We do not use this approach because if the fee model contains omitted correlated variables, using single year OLS would yield inconsistent coefficient estimates and, therefore, inconsistent estimates of the fee residual. Since our objective is to elucidate the composition of fee residuals, we prefer not to start our analysis with a potentially incorrect measure of the residual.

14

Ln(Afeeit) = b0 + b1 New+b2LResidual + b3 Ln(TAit) + b4 Lossit + b5 ROAit + b6 Leverageit + b7 InvRecit + b8 ForOpsit + b9 √Employeesit + b10 Nsegmentsit + b11 NewFinit + b12 ExtDistit + b13GCOit + b14 ICWeakit + b15Busyit + b16 Delayit + b17 Afilerit + Industry Indicators + Year Indicators + εit (3)

where New is equal to 1 for a first-year audit engagement, 0 otherwise. LResidual is the value of the lagged residual, , obtained from the first stage (i.e. the

preceding year’s fee residual obtained from model (2)) and other model variables are as defined earlier.

In this model, the coefficient on New (b1) captures the average initial fee discount and in a

competitive audit market, b1 can be expected to be negative and significant. Since LResidual is,

by construction, orthogonal to the levels of last year’s explanatory variables, it also is at best

loosely correlated with the current year’s levels of those variables and effectively serves as an

instrument for all time-invariant unobserved factors whose omission might render the model

inconsistent.17 Note that at this stage we include, but do not predict the sign of the coefficient

on, LResidual (b2).18 We defer that analysis to the next stage as part of our more systematic

investigation of residual fee persistence.

3.3 Testing fee persistence

To shed light on our key issue of interest, the extent to which the preceding year’s residual

fees explain audit fees on continuing and new engagements, we estimate the augmented model:

17 In fact, should the omitted time-invariant factors have a significant influence on audit fees, incorporating their effects via inclusion of LResidual in the fee model can also help the analyst develop more accurate predictions of expected audit fees conditioned on all publicly available data (as for instance a competitor interested in displacing an incumbent auditor might wish to do). However, note that improving model predictive power and fit in this way does not in any way enhance the explanatory power of the model (cf. Schmueli 2010, Section 1.6, Ebbes et al. 2011). 18 The reason for including LResidual and estimating model (3) using OLS is as follows. As noted earlier, if the omitted fee determinants are likely to be correlated with observed fee determinants, OLS estimation of model (3) without LResidual would yield inconsistent estimates of all model parameters including the coefficient of New. One way to render the estimation consistent is to include LResidual in model (3) as a (lagged) surrogate for the omitted factors, another is to omit LResidual from the specification and estimate the reduced model using a fixed-effects panel estimator (or after including a complete set of engagement-level indicators). We report results of estimating model (3) primarily to facilitate comparisons with the fee persistence analyses reported later.

15

Ln(Afeeit) = b0 + b1 LRFeePosit + b2 LRFeeNegit + b3 Ln(TAit) + b4 Lossit + b5 ROAit + b6 Leverageit + b7 InvRecit + b8 ForOpsit + b9 √Employeesit + b10 Nsegmentsit + b11 NewFinit + b12 ExtDistit + b13GCOit + b14 ICWeakit + b15Busyit + b16 Delayit + b17 Afilerit + Industry Indicators + Year Indicators + εit (4)

where LRFeePos is the value of the lagged residual, , obtained from the first stage if that

residual is ≥ 0 and 0 otherwise, and, LRFeeNeg is the value of the lagged residual, , obtained from the first stage if that

residual is < 0 and 0 otherwise. and other model variables are as defined earlier. We estimate this model, separately for the

continuing and new auditees using OLS (after including year and industry indicators).

In equation (4), the coefficients on LRFeePos and LRFeeNeg speak to the persistence of

positive and negative lagged fee residuals, after current levels of known fee determinants have

been accounted for. 19 As discussed earlier, we expect these coefficients to differ across

continuing and new auditees and, since prior research differs on the extent to which positive and

negative fee residuals likely reflect rents (DeFond et al. 2002; Hope and Langli 2010; Larcker

and Richardson 2004; Kanagaretnam et al. 2010; Choi et al. 2010; Asthana and Boone 2012), we

conduct all of our tests separately for positive and negative residual fees.

3.4 Data

Our sample consists of all public company audit engagements for which financial statement,

audit fee and opinion data are available in the Compustat and AuditAnalytics databases and that:

(1) are US-domiciled entities, and

(2) have a primary Standard Industrial Classification (SIC) code other than 4400-4999 or 6000-

6999 (both inclusive).

19 Although the dependent variable in equation (4) is total audit fees (not residual audit fees) in year t, b1 and b2 can be interpreted as speaking to the persistence of residual audit fees: the coefficient of ϵt-1 in a regression of yt on xt and ϵt-1 is identical to the coefficient obtained by (1) regressing yt on Xt and (2) regressing the residual from that regression, εt, on the component of ϵt-1 that is orthogonal to Xt (cf. Beaver, 1987: 139-140 , especially his Figure 1 and discussion thereof: as he notes, estimating equation 5 is equivalent to the 2 stage procedure of estimating his equations 6 and 9 after orthogonalizing Z2 using his equation 1).

16

Restriction (1) confines the sample to U.S. auditees since audits may be produced and priced

differently across national legal and institutional boundaries. Restriction (2) controls for inter-

industry differences in audit production and pricing (e.g. Stein et al. 1994; Fields et al. 2004). To

eliminate confounds due to differences in audit production functions across larger and smaller

auditors, we estimate equations (2) and (3) separately for Big N and non-Big-N auditees. Table

1 reports sample attrition due to the data restrictions imposed in this study.

Table 2, Panel A (Panel B) reports descriptive statistics for the sample used in the study. A

comparison of panel A with panel B reveals that Big N auditees are, on average, much larger

than their non-Big-N counterparts: mean audit fees and total assets of Big N auditees are about

$1.83 million (AFee) and $3,816 million (TA) respectively compared to fees and total assets of

about $274 thousand (AFee) and about $171 million (TA) respectively for non-Big-N auditees.

Big N auditee financial performance is also better than that of non-Big-N auditees: 31% of Big N

auditees report a loss, and mean ROA for Big N auditees is 0.02, compared to 52% and −0.14 for

non-Big-N auditees. Third, and related, Big N auditees also pose lower risk of imminent firm

financial distress than do Non-Big-N auditees (mean GCO 0.03 vs. 0.15). Fourth, Big N auditors

are more likely than Non-Big-N auditors to issue adverse internal control opinions (5% vs. 3%).

This difference likely reflects the higher proportion of Big-N auditees that are accelerated filers

(mean Afiler=0.92 vs. 0.76 for Non-Big-N auditees) and, therefore, subject to internal control

audits under the Sarbanes-Oxley Act.

Table 3 reports (year indicators omitted) the estimates of equation (2) used to compute fee

residuals for 2009 (which will be used as lagged fee residuals for 2010 when estimating model

(3)).20 All of the variables in the Big N fee model are significant in the expected direction at the

20 To estimate 2009 residual fees, equation (2) is estimated separately for Big N and Non-Big-N auditees, using data for all years during the period 2001-2009 when the auditee was a continuing Big N (or Non-Big-N) auditee. We

17

1% level of significance (or better). By contrast, for non-Big-N auditees, Busy as well as GCO,

Delay and AFiler are not significant, indicating lower capacity pressure and lower auditor

responsiveness to certain auditee risk characteristics relative to the Big N model. The intercept

as well as a number of other coefficients, most notably auditee size, also differ discernibly across

the two models. An untabulated F-test rejects the hypotheses of model equality (F(16, 5049) =

27.48, p<0.01), validating separate estimation of model (2) for Big N and Non-Big-N auditees.

4. Results

4.1 Preliminary: Competition in the market for new auditees

Table 4 reports the results of pooled OLS estimation of model (3) separately for Big N and

Non-Big-N auditees for the period 2003-2010. To account for repeated observations on the same

auditee, reported significance levels are based on standard errors clustered by auditee.

In the Big N auditee sample (results reported in the first two columns of Table 4), all of the

control variables are significant in the expected direction at the 5% level (or better) and model R2

is fairly high (about 0.92). In particular, the improvement in model R2 over that of the rolling-

window panel model reported in Table 3 (0.76) is interesting because it suggests that including

lagged values of fee residuals can help the analyst better predict – but not explain – expected

audit fees (much as incorporating information about serial correlation can help improve predicted

values of a time series without increasing explanatory power).

The key features of the results reported in these two columns is that relative to the fees paid

by their continuing auditees, Big N auditors offer significant discounts to new auditees. The

coefficient on New is negative and significant (-0.139, p<0.01), indicating that the new Big N

also performed a Hausman test to determine whether the fixed effects model or random effects model is more appropriate in our context. For both Big N and Non-Big-N models, the Hausman test rejects the null hypothesis that the random effects model is the preferred model (Big N: χ2(23)=2,939, p<0.01; Non-Big-N: χ2(22)=95.20, p<0.01 ). For reporting parsimony we do not tabulate the 2003-2008 residual fee models (all of which yield qualitatively similar coefficient estimates and model fit statistics).

18

auditor provides a first year fee discount of about 13% (100*(1- e -0.139)). The findings for Non-

Big-N auditees (results reported in columns three and four of Table 4) are qualitatively similar.

The most notable difference is that, as expected, Non-Big-N auditors offer their new auditees

significantly smaller fee discounts: the mean initial fee discount is only about 4.7 % (100*(1- e -

0.048)).21 Collectively, the findings in table 4 are consistent with prior findings in the literature

and indicate intra-Big-N auditor competition for new auditees (see, e.g., Francis 1984; Francis

and Simon 1987; Simon and Francis 1988; Ettredge and Greenberg 1990; Craswell and Francis

1999; Ghosh and Lustgarten 2006; Sankaraguruswamy and Whisenant 2009). In theory, the

magnitude of this initial fee discount reflects the sum of (a) any auditor optimism in bidding for a

new engagements and (b) the present value of all expected auditor rents.

4.2 Determining the information content of fee residuals

Table 5 reports the results of estimating equation (4) separately for continuing Big N

(columns 1 and 2), continuing non-Big-N (columns 3 and 4), laterally-switching Big N (columns

5 and 6), laterally-switching non-Big-N (columns 7 and 8) auditees and auditees switching from

Big-N-to-smaller-auditors (columns 9 and 10) for the period 2003-2010. Columns 1 and 2,

columns 5 and 6, and columns 9 and 10 correspond to the three principal tests described earlier.

The remaining four columns of table 5 provide additional context for interpreting these results.

As in table 4, reported significance levels are based on standard errors clustered by auditee.

Columns 1 and 2 of table 5 present the results of the first test. The coefficients on LRFeePos

(b1) and LRFeeNeg (b2), the main focus of interest, are both positive and significant (0.71 and

0.81 respectively, one-tail p value <0.01 in each case) indicating that fee residuals are highly

persistent from year-to-year: Big N auditees that pay above (below) average fees in one year are

21 Should they represent rents (rather than any auditor optimism), the smaller initial fee discounts documented in the non-Big-N auditee segment (about 4.7% vs. about 13%) likely reflect the lower lifetime rents that the Non-Big-N auditors can expect to earn over their expected tenure.

19

likely to pay above (below) average fees in the next year.22 Columns 3 and 4 provide additional

context for these findings by examining fee persistence in continuing non-Big-N engagements:

the coefficients on LRFeePos and LRFeeNeg for non-Big-N auditees are also both positive (0.82

and 0.78 respectively, one-tail p value <0.01). Overall, these four columns provide strong

evidence that, for both Big N and non-Big-N engagements, residual audit fees largely reflect

systemic engagement-specific influences rather than transient or noise factors.

Columns 5 and 6 of table 5 speak to the second test. The coefficients of LRFeePos and

LRFeeNeg are both positive and significant (0.74 and 0.72 respectively, one-tail p value < 0.01

in each case). The finding that the coefficient of LRFeePos is almost identical to that

documented in column 1 provides particularly strong evidence that residual audit fees from

continuing Big N audit engagements consist largely of researcher-unobserved audit production

costs common to all Big N auditors.23 Columns 7 and 8 provide additional context for the

second primary test and also set the stage for the third primary test by reporting the coefficient of

LRFeePos and LRFeeNeg for laterally-switching non-Big-N auditees. These coefficients, (0.79

and 0.76 respectively, one-tail p value <0.01 in each case) are not materially different from those

for laterally switching Big N auditees (values reported in column 5).24

22 In the terminology of Section 2, this finding indicates that α, the contribution of to is roughly of the order of about 30% and that only 70% of the lagged fee residual consists of potentially persistent components. 23 In the terminology of Section 2, this finding indicates that in lateral Big N auditor transition, for positive lagged residual fees, which are considered to be the most likely instances where auditors earn rents, γC (0.71 in column 2) = γN (0.74 in column 5). This finding provides strong evidence that positive lagged fee residuals consist largely of unobserved production costs common to the predecessor and successor auditor and contain very little by way of auditor rents. By contrast, for negative lagged fee residuals, the finding that γC (0.81 in column 2) ≥ γN (0.72 in column 5) indicates that below-average costs incurred by the predecessor do not persist at the same rate across lateral Big N auditor transitions. That is, when the predecessor auditor had charged below average audit fees, the successor auditor reverses part of the below-average fee paid to the predecessor auditor. Such a reversal could indicate either a loss of production efficiency (i.e. higher labor usage) or greater initial year effort and/or higher risk-premium recovery by the successor auditor (cf. Bell et al. 2008). 24 Two interesting implications of this finding are that, first, the extent of cost commonality between non-Big-N auditors is not materially different from that between Big N auditors. This finding is, to our knowledge, the first systematic empirical evidence on the extent of audit cost commonality across different classes of auditors. Second, since non-Big-N auditors are expected to earn considerably smaller rents than their Big N counterparts, the

20

Columns 9 and 10 of table 5 report the results of the third test, that of residual fee persistence

on Big N audit engagements that switch to smaller auditors. As noted earlier, this test is likely to

confound the effects of the successor auditor earning lower rents (driving the persistence to zero)

with the effect of rebating costs that the successor does not share with the Big N predecessor

(which also would drive residual fee persistence to zero).25 The results, however show that

coefficients on both LRFeePos and LRFeeNeg are positive and significant (0.57 and 0.76

respectively, one-tail p value <0.01). By the logic exposited earlier in section 2, the difference of

about 0.14 in the coefficients of LRFeePos between the values reported in columns 2 and 9

represents the sum of (a) ic, i.e., cost differences between Big N and non-Big-N auditors, (b) ir,

rents accruing to the Big N predecessor not expected to be earned by the successor, (c) less any

discounting by the successor of its expected rents. Should – as is often posited in prior research –

the non-Big-N successor be expected to earn no rents, the last component can be expected to be

zero and the difference of 0.14 can be interpreted as the sum of the Big N predecessor’s ic and ir

components. Were Big N and non-Big-N auditors’ costs to differ by as little as 10%, the extent

of rents in the lagged fee residual would not be more than 4% of the residual fee. Collectively,

therefore, we conclude that the results of the three tests described above provide consistent and

strong evidence that residual fees largely comprise of unobserved audit production costs and

contain little if any rents.

similarity of the coefficients in lagged fee residuals across columns 1, 3, 5 and 7 provides further evidence for a cost-centric interpretation of fee residuals: the fee residuals in columns 1 and 5, which are computed from Big N engagements, and therefore are more likely to contain rents, persist at about the same rate as residuals on non-Big-N engagements which are thought to contain little if any rents. 25 In some ways the analysis of Big N to non-Big N auditors is more powerful than the analysis of lateral Big N transitions (reported in columns 5 and 6), in that the successor non-Big-N auditor is less likely to earn rents than the successor auditor in the second test. As a result, fee persistence in this test can be expected to be lower than it was in that test. However, the successor auditor in this case also is expected to have lower audit production costs than did the successor auditor in the second test. Consequently, overall, the lateral Big N auditor transition test is, in our view, the more powerful test.

21

4.3 Additional analyses

4.3.1 Alternate model specification

Separately estimating model (4) on continuing and new auditees allows for the possibility

that continuing and new auditees may have different coefficients for the test and control

variables. However, such estimation does not allow us to test for the sample wide mean first-

year fee discounts offered by the auditors. Therefore, we also test the following interaction

model that includes a variable (New) that will capture the first-year fee discount, but will not

allow all the coefficients to vary across continuing and new auditees. The model we employ is

Ln(Afeeit) = b0 + b1 Newit + b2 LRFeePosit + b3 LRFeeNegit + b4 Newit*LRFeePosit + b5 Newit*LRFeeNegit + b6 Ln(TAit) + b7 Lossit + b8 ROAit + b9 Leverageit + b10 InvRecit + b11 ForOpsit + b12 √Employeesit + b13 Nsegmentsit + b14 NewFinit + b15 ExtDistit + b16GCOit + b17 ICWeakit + b18Busyit + b19 Delayit + b20 Afilerit + Industry Indicators + Year Indicators + εit (5)

where New is equal to 1 for a first-year audit engagement, 0 otherwise. LRFeePos is the value of the lagged residual, , obtained from the first stage if that

residual is ≥ 0 and 0 otherwise, and, LRFeeNeg is the value of the lagged residual, , obtained from the first stage if that

residual is < 0 and 0 otherwise.

and other model variables are as defined earlier. In this model, New captures the first year fee

discount while the coefficients on LRFeePos (b2) and LRFeeNeg (b3) speak to the persistence of

positive and negative residual fees in continuing engagements. When lagged residual fees reflect

systematic engagement-specific influences, both b2 and b3 can be expected to be positive and

significant. Finally, in this test, the sign and magnitude of the sum (b2 + b4) of the coefficients

on LRFeePos and New*LRFeePos and the sum (b3 + b5) of the coefficients on LRFeeNeg and

New*LRFeeNeg (b5) speak to the persistence of lagged residual in auditor switches. As a result,

our predictions for the sign and magnitude of these sums is the same as our predictions for the

persistence of the lagged residuals in auditor switches.

22

Table 6 reports the results of estimating model (5) for (a) continuing and laterally-

switching Big N auditees (column 1 and 2), (b) continuing and laterally-switching non-Big-N

auditees (columns 3 and 4) and (c) continuing auditees and auditees switching from Big N to

smaller auditors (columns 5 and 6). Columns 1 and 3 reveal that both Big N and Non-Big N

auditors offer significant initial fee discounts (the coefficients of New are each negative and

significant). Interestingly, the coefficient on New for the Big N to smaller auditor switches is

positive (0.32, p<0.01) indicating that auditees that switch from a Big N auditor to a non-

Big-N auditor pay about 37% (100*(e 0.317-1)) more than comparable non-Big-N auditees. 26

Second, the sum of the coefficients (i.e. b2+b4 and b3 +b5) is positive and significant for all

auditor transitions (last two rows of table 6). This finding is inconsistent with the lagged

residual largely consisting of rents: were that to be the case, one would expect both sums to

be negative and significant for lateral Big N switches (and zero for both lateral non-Big-N

switches and Big N to smaller Non-Big-N auditor switches). Therefore, as in table 5, the

results in table 6 strongly support the proposition that fee residuals are driven by unobserved

costs or risk factors and not by auditor rents.

4.3.2 Tests based on OLS estimation of residual fees

As noted earlier (cf. Section 2), the analysis thus far is predicated on the assumption that

researcher-unobserved influences on audit fees are correlated with the observed “explanatory”

variables used to predict audit fees. Most prior studies, however, estimate residual fees using

26 To the best of our knowledge, this result has not been reported in prior research. Interestingly, untabulated analyses show that fees paid to new Non-Big-N auditors are, on average, about 34% (100*(1-e-042)) lower than the “counterfactual” fees the auditee would have paid if they continued with a Big N auditor (cf. Chaney et al. 2004). In conducting this analysis, for each of the 310 auditees that switched, we computed the counterfactual fee as the expected fee from rolling widow estimates of model (3) for Big N auditees, i.e., after accounting for the impact of lagged fee residuals on Big N auditors’ fees for continuing engagements. In other words, we document that auditees switching from a Big N to a non-Big-N auditor pay lower fees than they would have paid the predecessor (obtain a relative “initial fee discount” from the non-Big-N successor), a result similar to Chaney et al. (2004), but that they pay more than a similar continuing non-Big-N auditee would pay its auditor (to our knowledge, a new finding).

23

OLS analysis that includes industry and/or year fixed effects (but not auditee-level fixed effects).

The results of such analyses can be interpreted as unbiased and consistent estimators of the fee

residual under the stronger assumption that researcher-unobserved influences that affect

observed audit fees are uncorrelated with the observed vector of fee regressors. In this section,

we report the results of replicating our analyses under this stronger assumption.

Table 7 reports results analogous to those reported in table 3 after adding to model (2) a full

set of industry fixed effects. The model used to estimate the results reported in table 7, therefore,

includes both industry and year fixed effects, ensuring consistency with the residual estimation

procedures used in the vast majority of prior audit fee research. Tables 8 and 9 report results

analogous to tables 4 and 5 respectively after replacing the auditee-fixed-effects panel residual

with the residuals computed from the model reported in table 7.

The most notable differences between the results reported in tables 3 and 7 is that the OLS

estimates reported in table 7 show (a) a coefficient of auditee size that is almost 1.5 times as

large as its panel estimation counterpart (the coefficient of Ln(TA) is 0.46 in table 6 vs. 0.33 in

table 3), (b) a substantially smaller intercept (3.86 in table 6 vs. 6.45 in table 3) and (c) larger

coefficient magnitudes in general than those obtained in the panel estimation. The overall

implication of these differences in parameter estimates of the fee model is that correcting for any

unobserved correlation of omitted fee determinants with the observed fee drivers attenuates the

impacts implied by the conventional OLS estimator which is estimated under the more restrictive

assumption that all such correlations are zero.

Comparing the results reported in table 8 to their analogs in table 4 reveals that substituting

the OLS residuals for panel residuals does not materially impact our findings. In both tables 8

and 4, the new auditee discount for Big N auditees is about 13% (the Big N coefficient of New is

24

about -0.14, p<0.01 in both tables) and about 4.7% for non-Big-N auditees (the non-Big-N

coefficient of New is about -0.05, p<0.01 in both tables). Likewise, lagged residual fees persist

at similar rates in both tables for the Big N (0.76 in both table 4 and table 8) as well as non-Big-

N auditees (0.81 in table 8 vs. 0.79 in table 4). Overall, the results in table 8 confirm the findings

of first-year fee discounts and residual fee persistence noted in table 4.

Comparing the results reported in tables 5 and 9 also reveals no significant differences. In

both tables continuing Big N and non-Big-N auditees exhibit similar degrees of residual fee

persistence. For instance, for continuing Big N auditees, the coefficients for LRFeePos and

LRFeeNeg are 0.71 and 0.81 in table 5 and about 0.67 and 0.84 in table 9 while for non-Big-N

auditees, they are 0.82 and 0.78 in table 5 and 0.84 and 0.78 in table 9. The coefficients of both

LRFeePos and LRFeeNeg for lateral Big N and Big N to smaller auditor switches (columns 5 and

6 and columns 9 and 10 respectively) are also qualitatively similar across tables 9 and 5. In sum,

our inference that residual fees comprise largely of unobserved production costs (and contain at

best negligible auditor rents) is not sensitive to the procedure used to estimate the fee residual.

We also replicated the analyses reported in table 9 after computing fee residuals using the

approach most frequently followed in prior research, namely pooled OLS estimation of model

(2) across all auditees (both Big N and non-Big-N). Our principal findings (untabulated) and

inferences are, in all material respects, unaltered by this methodological perturbation. For

instance, using pooled OLS model residuals, we find that the coefficients of positive lagged

residual fees are 0.66 (one-tail p value <0.01) for continuing Big N auditees and 0.70 (one-tail p

value <0.01) for laterally switching Big N auditees. The corresponding coefficients of negative

lagged residual fees are 0.84 (one-tail p value <0.01, continuing auditees) and are 0.78 (one-tail p

25

value <0.01, laterally switching auditees). The other coefficients, without exception, similarly

closely conform to those reported in Table 9.

5. Concluding remarks

We show analytically how differential persistence of residual audit fees across various types

of audit engagements can elucidate the extent to which residual audit fees consist of researcher-

unobserved audit costs, auditor rents and noise (purely transient influences and random error).

Empirically, in a comprehensive sample of U.S. public company audit engagements conducted

during 2003-2010, we document that residual fees are comprised almost entirely of unobserved

audit production costs and noise.

Four noteworthy implications follow. First, future research should use residual audit fees as a

surrogate for researcher-unobserved audit costs rather than for auditor rents (auditor-auditee

economic bonding). Second, since residual fees are persistent both over time and across auditors

and their inclusion materially increases model fit, future research that elucidates the factors that

drive residual fees can help improve the explanatory power of the Simunic fee model. Third,

since they are highly persistent, lagged residual fees can (and should) be used to predict expected

audit fees (i.e., in forward-looking applications of the Simunic fee model where the goal is to

predict rather than explain observed audit fees). Fourth, prior research findings of associations

between residual fees and measures of audit quality or auditee financial reporting quality should

be interpreted as association between researcher-unobserved audit production costs and the

quality measure of interest rather than as evidence of auditor-auditee economic bonding.

Our findings also suggest some intriguing conjectures. In our sample, the mean initial fee

discount in lateral Big-N transitions is between ten and fifteen percent. Since we find that fee

residuals almost entirely reflect costs, any rents paid to continuing Big N auditors cannot exceed

26

the amount that if capitalized would explain an such observed initial fee discount (i.e., of ten to

fifteen percent). This finding implies that annual rents to Big N auditors are quite small. In

principle, the magnitude of rents is indicative of both the extent of price competition in the audit

market and of the strength of auditors’ incentives to collude with auditees. Consequently a

finding that any rents accruing to public company auditors are quite low should be comforting to

financial statement users. Moreover, while our estimates are based on public data, our research

approach can be used by regulators such as the PCAOB to independently estimate the extent of

auditor rents using the much richer data sets that are likely available to them. Should this be

feasible, a regime of periodic reporting by regulator of the likely magnitude of rents earned by

auditors can be expected to improve investor confidence in the functioning of the audit market

and to reduce the cost of corporate capital. Even if regulators’ estimates of rents are much higher

than what we report, sunshine being the best disinfectant, the mere fact that they are estimable

and reportable can be expected to improve the functioning of the audit market over the long run.

27

References

Asthana, S.C., & Boone, J.P. (2012). Abnormal audit fee and audit quality. Auditing: A Journal of Practice & Theory 31(3), 1-22.

Ball, R., Jayaraman, S. & Shivakumar, L. (2012). Audited financial reporting and voluntary disclosure as complements: A test of the confirmation hypothesis. Journal of Accounting and Economics 53(1-2), 136-166.

Beaver, W.H. (1987) The properties of sequential regressions with multiple explanatory variables. The Accounting Review 62(1), 137-144.

Bell, T.B., Doogar, R., and Solomon, I. (2008) Audit labor usage and fees under business risk auditing. Journal of Accounting Research 46 (4), 729-760.

Bell, T.B., Landsman, W.R., & Shackelford, D.A. (2001) Auditors' perceived business risk and audit fees: Analysis and evidence. Journal of Accounting Research 39 (1), 35-43.

Chaney, P.K., Jeter D.C.,& Shivakumar, L. (2004). Self-selection of auditors and audit pricing in private firms. The Accounting Review 79 (1), 51-72.

Choi, J.H., Kim, J.B. & Zang, Y (2010). Do abnormally high audit fees impair audit quality? Auditing: A Journal of Practice and Theory 29 (2), 115-140.

Craswell, A.T., & Francis, J.R. (1999) Pricing initial audit engagements: A test of competing theories. The Accounting Review 74 (2), 201-216.

DeAngelo, L.E. (1981) Auditor independence, ‘low balling', and disclosure regulation. Journal of Accounting and Economics 3 (2), 113-127.

DeFond, M.L., Raghunandan, K. & Subramanyam, K (2002). Do non–audit service fees impair auditor independence? Evidence from going concern audit opinions. Journal of Accounting Research 40 (4), 1247-1274.

Dye, R.A. (1991). Informationally motivated auditor replacement. Journal of Accounting and Economics 14 (4), 347-374.

Ebbes, P., Papies, D. & van Heerde, H.J. (2011). The Sense and non-sense of holdout sample validation in the presence of endogeneity. Marketing Science 30 (6), 1115-1122

Ettredge, M., & Greenberg, R. (1990). Determinants of fee cutting on initial audit engagements. Journal of Accounting Research 28 (1), 198-210.

Fields, L.P., Fraser, D.R. & Wilkins, M.S. (2004). An investigation of the pricing of audit services for financial institutions. Journal of Accounting and Public Policy 23 (1), 53-77.

28

Francis, J.R. (1984). The effect of audit firm size on audit prices: A study of the Australian market. Journal of Accounting and Economics 6 (2), 133-151.

Francis, J.R. (2011). A framework for understanding and researching audit quality. Auditing: A Journal of Practice and Theory 30 (2), 125-152.

Francis, J.R., & D.T. Simon (1987). A test of audit pricing in the small-client segment of the US audit market. The Accounting Review 62 (1), 145-157.

Ghosh A., & Lustgarten, S. (2006). Pricing of initial audit engagements by large and small audit firms. Contemporary Accounting Research 23(2), 333-368.

Government Accountability Office (GAO), (2008). “Audits of public companies: Continued concentration in audit market for large public companies does not call for immediate action.” GAO-08-163, Washington, DC.

Higgs, J. L., & Skantz, T. R. (2006). Audit and nonaudit fees and the market’s reaction to earnings announcements. Auditing: A Journal of Practice & Theory 25(1): 1–26.

Hay, D.C., Knechel, W.R. & Wong, N. (2006). Audit Fees: A meta-analysis of the effect of supply and demand attributes. Contemporary Accounting Research 23 (1), 141-191.

Hope, O.K., & Langli, J. (2010). Auditor independence in a private firm and low litigation risk setting. The Accounting Review 85 (2), 573-605.

Hribar, P., Kravet, T. & Wilson, R. (2010). A new measure of accounting quality. University of Iowa Working paper

Kanagaretnam, K., Krishnan, G.V. & Lobo, G.J. (2010). An empirical analysis of auditor independence in the banking industry. The Accounting Review 85 (6), 2011-2046.

Kanodia, C., & Mukherji, A. (1994). Audit pricing, lowballing and auditor turnover: A dynamic analysis. The Accounting Review 69 (4), 593-615.

Kinney Jr., W.R., & Libby, R. (2002). Discussion of the relation between auditors' fees for nonaudit services and earnings management. The Accounting Review 77 (Supplement), 107-114.

Larcker, D.F., & Richardson, S.A. (2004). Fees paid to audit firms, accrual choices, and corporate governance. Journal of Accounting Research 42 (3), 625-658.

O'Keefe T.B., Simunic, D.A & Stein, M.T. (1994). The production of audit services: Evidence from a major public accounting firm. Journal of Accounting Research 32 (2), 241-261.

Sankaraguruswamy, S., & Whisenant, S. (2009). Pricing initial audit engagements: Empirical evidence following public disclosure of audit fees. Working paper available at http://papers.ssrn.com/sol3/papers.cfm?abstract_id=452680

29

Schmueli, G. (2010). To explain or to predict? Statistical Science 25 (3), 289-310.

Simon, D.T., & Francis. J.R. (1988). The effects of auditor change on audit fees: Tests of price cutting and price recovery. The Accounting Review 63 (2), 255-269.

Simunic, D.A (1980). The pricing of audit services: Theory and evidence. Journal of Accounting Research 18 (1), 161-190.

Srinidhi, B.N. & Gul, F.A (2007). The differential effects of auditors’ nonaudit and audit fees on accrual quality. Contemporary Accounting Research. 24(2), 595-629.

Stein, M.T., Simunic, D.A & O’Keefe, T.B. (1994). Industry differences in the production of audit services. Auditing: A Journal of Practice and Theory 13 (1), 128-42.

Watts, R.L., & Zimmerman, J.L. (1986). Positive accounting theory. Prentice-Hall. New Jersey.

Wooldridge, J.M. (2002). Econometric analysis of cross section and panel data. The MIT Press. Boston, MA.

30

Appendix A

Cost-centric interpretations of the fee residual

One of earliest explicitly cost-centric readings of residual audit fees is offered by Ettredge

and Greenberg (1990) who note that:

“… in a competitive setting cross-sectional differences in audit fees should be highly correlated with differences in audit costs. We employ the old auditor's fee as a surrogate for cost and compare it with the "expected" fee given the client's characteristics. The expected fee is derived from a regression model … Each client's fee residual is obtained from the regression … In similar fashion, the new auditor's fee is used as a dependent variable in a regression with the same independent variables. If a new auditor has a positive regression residual, that auditor should be a high-cost provider relative to other new auditors.”

More recently Ball et al. (2012) motivate a cost centric reading of the fee residual as follows:

“…We follow a substantial auditing literature starting with Simunic (1980) and Watts and Zimmerman (1983), and use the amount of excess audit fees paid by a firm as the proxy for the extent of its financial statement verification, based on the logic that incremental audit effort is priced by its auditors. Audit fees are affected by the choice of audit firm (notably, Big Four versus smaller firm), the seniority level of the audit engagement partner, the number of audit personnel on the job and their average hourly rate, the degree of verification of internal control systems and individual transactions required by the client, the frequency of communication with the audit committee, and other variables. Fees typically are negotiated by management and approved by audit committees in advance. They are directly linked to the quantity and price of audit activity, and hence to the extent of independent verification of financial reporting. Excess audit fees represent fees that are incremental to those associated with previously identified determinants”

while Hribar et al. (2010) argue that:

Economic theory suggests that in a competitive equilibrium, audit fees incorporate the expected cost of poor quality earnings. Auditors face significant reputation and litigation risk if their client’s financial reports are misstated. In response, auditors will increase audit hours and audit fees when they perceive accounting quality to be low. Isolating this part of the audit fee from the total audit fees can therefore provide a summary measure of the auditor’s assessment of the quality of the accounting system. We use a regression-based approach to remove the expected amount of audit fees based on the scope of the audit and other determinants, and use the unexplained fees as a measure of accounting quality. To conclude our review of the cost-centric view, we note that none of the cost-centric studies

makes any effort to isolate and exclude any auditor rents contained in the residual fee before

using it as a proxy for auditor rents. Consequently, from an econometric perspective, research

practice among cost-centric scholars can accurately be described as assuming that any rents or

noise components of residual fees are negligible and that the entire residual fee can be used as a

valid proxy for researcher-unobserved audit production costs.

31

Rent-centric interpretations of the fee residual

In a highly influential commentary, Kinney and Libby (2002: 109-110) argue that:

“ … 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. … More insidious effects on economic bond may result from unexpected nonaudit and audit fees that may more accurately be likened to attempted bribes. … 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 … 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.”

while DeFond et al. (2002) argue that:

“… auditor independence may be influenced by the amount of client fees relative to their expected amounts, rather than the nominal amounts … This notion is consistent with auditors’ being influenced by whether the client is a source of unusually high or low fees. Therefore, we draw on prior research that models audit and non-audit fees to develop a model that extracts the unexpected portion of fees in our sample firms … we use the error terms … to surrogate for the “unexpected” portion of each of our fee variables.”

Larcker and Richardson’s (2004) more nuanced rent-centric reading of fee residuals is that:

“We expect auditor behavior to vary depending on whether the auditor is being paid more or less than the economic benchmark for a specific client. When the abnormal fee is less than or equal to zero the auditor has little to lose if it imposes stringent accounting requirements on the client that result in lower levels of accruals. If this action causes the auditor to lose this client, we assume there are other more profitable uses for the staff previously assigned to this client. However, if the client remains with the auditor, there is considerable incentive for the auditor to be aggressive with this client to minimize any reputation loss due to an audit failure. Obviously, where the auditor would be most susceptible to client pressure is when the abnormal fee is greater than zero. If reputation concerns are not relevant to the auditor, we should observe lower earnings quality for firms that pay a positive premium to the auditor for audit and non-audit work.”

Subsequent research adopts one of two approaches to the use of fee residuals. Srinidhi and Gul

(2007) use all fee residuals, positive or negative as measures of rent and argue that:27

“we use a simultaneous expectations model to decompose audit and nonaudit fees and test the proposition in Kinney and Libby 2002 that economic bonding might result from unexpected rather than from expected fees.”

27 While Gul and Srinidhi (2007) motivate the use of residual audit fees using Kinney and Libby’s economic bonding argument, they find no association between residual audit fees and the magnitude of auditee abnormal accruals (their measure of audit quality), which they interpret as follows:

“This implies that when audit fees are unexpectedly high, it is difficult to distinguish whether it reflects an unexpectedly high audit effort (with consequent improvement in accrual quality) or excessive rents (that result in a deterioration of accrual quality).”

32

and following – and citing – them, Kanagaretnam et al., (2010) argue that:

… our main focus is on unexpected total fees and unexpected nonaudit fees because these fee measures are based on controlling for bank-level determinants of normal fees and prior research suggests that they are better measures for examining fee dependence (Kinney and Libby 2002; Srinidhi and Gul 2007).

Choi et al. (2010), however, like Larcker and Richardson, advocate focusing attention

primarily on positive fee residuals as the most likely repository of auditor rents:

“This study examines whether the association between audit fees and audit quality is asymmetric and thus nonlinear in the sense that the association is conditioned upon the sign of abnormal audit fees. We define abnormal audit fees as the difference between actual audit fees i.e., actual fees paid to auditors for their financial statement audits and the expected, normal level of audit fees. … As noted by Kinney and Libby (2002 109), abnormal fees “may more accurately be likened to attempted bribes” and can better capture economic rents associated with audit services or an auditor’s economic bond to a client than normal fees or actual fees. … We expect that the association between abnormal audit fees i.e., a proxy for economic rent and audit quality is negative when abnormal audit fees are positive i.e., when actual audit fees are higher than normal audit fees. This is because excessive audit fees can create incentives for auditors to acquiesce to client pressure for substandard reporting and thus erode audit quality. We expect, however, that the association between fees paid to auditors and audit quality fee-quality association hereafter is ambiguous or insignificant when abnormal audit fees are close to zero or negative. This is because auditors have few incentives to compromise audit quality in this case.”

In contrast to these studies, Asthana and Boone (2009) argue that positive fee residuals

represent auditor-auditee economic bonding (auditor rents) while negative fee residuals

reflect greater auditee bargaining power, both of can be expected to impair audit quality:

“If the audit fee model is well specified, the residual audit fee reflects abnormal profits from the audit engagement. To the extent that some factors are unobservable (and hence omitted from the audit fee model), the residual audit fee metric measures abnormal audit profitability with error. Abnormal audit profitability should be associated with both client bargaining power and economic bonding. With respect to the former, Casterella et al. (2004) show a negative association between proxies for client bargaining power and audit fees earned by industry specialists. Their research suggests that, ceteris paribus, below-normal audit fees may reflect billing concessions granted by the auditor due to client bargaining power. With respect to the later, Kinney and Libby (2002) note that ‘‘Unexpected fees may also better capture the profitability of the services provided . . . more insidious effects on economic bond may result from unexpected nonaudit and audit fees that may more accurately be likened to attempted bribes.’’

Hope and Langli (2010), on the other hand avoid taking any position on the interpretation of

positive or negative fee residuals and use both positive and negative fee residuals in their

main test and positive fee residuals in ancillary tests, a procedure they justify as follows:

“Our main tests examine the effects of (abnormal) audit and nonaudit fees on the probability of issuing going-concern opinions (either in an explanatory paragraph or as a disclaimer of

33

opinion)—generally the most serious audit modification provided. … According to Choi et al. (2009) and Hope et al. (2009a), economic bonding between the client and audit firm is less likely to occur for any level of negative abnormal fees. Consequently, in the seventh sensitivity test we run all tests using only the subsample with positive abnormal fees.”

To conclude our review of the rent-centric literature, it may be useful to note three stylized

facts about this literature that are relevant to the rest of our analysis and for future research on

audit fees and fee model residuals. First, none of the studies cited above alludes to cost-centric

readings of the residual fee proper in motivating the use of residual audit fees as a research

construct (as opposed to the raw audit fee, which most studies acknowledge does reflect auditor

effort), or attempts to isolate and exclude any cost-related components of the fee residual before

using the fee residual as a proxy for auditor rents.28 Consequently, from an econometric

perspective, research practice among rent-centric scholars can accurately be described as

assuming that any researcher-unobserved audit production cost or noise components of residual

fees are negligible and that the entire residual fee can be used as a valid proxy for auditor rents.

Second, note that while economic rents (excess profits) are, by definition, expected to be

non-negative, regression residuals are, by construction zero mean variables. This suggests the

need for a formal explanation of why a zero mean variable (residual fees) can be used as a valid

proxy for a variable that is expected to have a positive mean (auditor rents). The extant rent-

centric literature does not offer such a formal explanation. The passages cited from both Larcker

and Richardson (2004) and Choi et al. (2010) are representative of the usual treatment of the

issue in the prior literature: while the authors usually argue explicitly that positive fee residuals

are more-likely to contain rents, they generally are silent as to what negative fee residuals

contain. The implication seems to be that in a regression framework means rents will be swept

28 The one exception, as noted earlier, is Gul and Srinidhi (2007) which does allude, albeit in passing, to a possible cost-centric interpretation when offering an ex-post explanation of an observed lack of association between residual audit fees and their measure of audit quality. Asthana and Boone (2009) too allude to the possibility that fee residuals may include the effects of unobserved drivers of audit effort.

34

up in the intercept with fee residuals reflecting deviations from, i.e., above or below-average

rents.

Agnostic interpretations of residual fees

Higgs and Skantz (2006) is the only prior study we have found which explicitly argues that

fee residuals are likely to be an admixture of researcher-unobserved costs and rents:

“While it seems reasonable to expect fee residuals to be positively associated with engagement profitability, it is also likely that fee residuals reflect demand and risk effects. First, demand factors may be omitted from the fee prediction model. The auditor may recognize (and price) engagement complexity not captured by the model, or the client may have unique preferences for audit quality and thus acquire more or fewer services than the model predicts. … Second, risk factors apparent to the auditor but not specified in the model may be reflected in fee residuals. … Finally, residual fees will represent the abnormal profitability of the engagement. A negative relation between this profitability component and earnings quality would be expected if the market perceives extraordinarily profitable engagements as a threat to auditor independence.”

Higgs and Skantz (2006) make no attempt to decompose fee residual into cost and rent

components before testing their hypotheses. Consequently, from an econometric perspective,

research practice among agnostic scholars can accurately be described as assuming that residual

fees are an admixture of yet to be ascertained amounts of researcher-unobserved audit production

costs, auditor rents and noise.

35

Appendix B: Variable Definitions Variable Name Description AFee The sum of all audit and audit-related fees, in constant 1999 dollars, paid to the

external auditor (Audit Analytics data item matchfy_sum_afee). Ln(Afee) Natural logarithm of AFee, TA Auditee total assets in millions of constant 1999 dollars.Ln(TA) Natural logarithm of TA*106

Loss Takes the value 1 if net income (Data172, NI) is negative, 0 otherwise. ROA Auditee return on assets, computed as operating income after depreciation

(Data178, OIADP) ÷ Total assets (Data6, AT).Leverage The ratio of total liabilities (Data181, LT) to total assets (Data6, AT). InvRec Computed as (Total receivables (Data2, RECT) + Total inventories (Data3, INVT))

÷ Total assets (Data6, TA).ForOps Takes the value 1 if the auditee reports a foreign currency translation (Data150,

FCA) value other than zero, 0 otherwise.√Employees Square root of number of employees (Data29, EMP).Nsegments Number of business segments. Data is obtained from the Compustat segment data. NewFin Takes the value 1 if sum of new equity (Data108, SSTK) and debt issue (Data111,

DLTIS) exceeds $50,000, 0 otherwise.ExtDis Takes the value 1 if the absolute value of extraordinary items or discontinued

operations (DATA48, XIDO) exceeds $10,000, 0 otherwise. GCO Takes the value 1 if the auditor opinion for the fiscal year includes a going concern

qualification, 0 otherwise (Audit opinion from Audit Analytics Audit Opinion Database).

Busy Takes the value 1 if the auditee fiscal year ends in the month of December, 0 otherwise.

ICWeak Takes the value 1 if the auditor reports an internal control weakness (IC weak data is from Audit Analytics Internal Control database), 0 otherwise

DELAY The number of calendar days elapsed between the auditee’s fiscal year end and the date of the audit opinion (Audit Analytics audit opinion data set item date aud op minus Audit Analytics fee data set item fiscal year ended).

Afiler Takes the value 1 if auditee market value of equity at the end of the fiscal year exceeds $75 million, 0 otherwise.

New Takes the value 1 in the first year of an engagement, 0 otherwise. LRFeePos Takes the value of the lagged residual fee when that value is positive, 0 otherwiseLRFeeNeg Takes the value of the lagged residual fee when that value is negative, 0 otherwiseBigtoNBig Takes the value 1 for switches from Big N to Non-Big-N auditors, 0 otherwiseNBigtoNBig Takes the value 1 for intra-Non-Big-N auditor switches, 0 otherwise

36

Table1: Sample Selection Criteria (Test Sample) Merged Compustat-AuditAnalytics sample with all relevant data available for fiscal years 2003 to 2010 (both inclusive) 37,390Less observations: 1) Pertaining to companies headquartered outside the United States 4,6732) Pertaining to companies operating in the regulated and financial sectorsa 7,4273) Pertaining to companies with missing industry codes 5914) With delay less than zero 355) Missing lagged residuals 4,0616) Pertaining to years in which companies switch between big four and non-big four auditors 363Total number of qualifying observations used in main tests 20,240Number of Big N observations available for analyses 15,609Number of non-Big-N observations available for analyses 4,631 a Firms for which Compustat reports a primary SIC code (sic) between 4400 to 4999 (both inclusive) or between 6000 to 6999 (both inclusive).

37

Table 2: Summary Statistics (Test Sample) Variable Mean S.D. p(0) p(25) p(50) p(75) p(100) Panel A. Big N Auditees AFee ($103) 1,833 3,485 22 426 866 1,828 89,800 Ln(AFee) 13.69 1.12 9.99 12.94 13.64 14.39 18.15 TA ($106) 3,816 20,984 0 173 608 2,043 797,769 Ln(TA) 20.04 1.85 12.01 18.78 20.04 21.24 27.22 Loss 0.31 0.46 0.00 0.00 0.00 1.00 1.00 ROA 0.02 0.25 -2.34 0.01 0.07 0.12 0.38 Leverage 0.51 0.31 0.03 0.30 0.48 0.65 3.16 InvRec 0.26 0.18 0.00 0.12 0.23 0.36 1.00 ForOps 0.30 0.46 0.00 0.00 0.00 1.00 1.00 √Employees 2.38 2.81 0.00 0.71 1.53 2.92 45.83 Nsegments 2.01 1.59 0.00 1.00 1.00 3.00 10.00 NewFin 0.42 0.49 0.00 0.00 0.00 1.00 1.00 ExtDis 0.07 0.26 0.00 0.00 0.00 0.00 1.00 GCO 0.03 0.17 0.00 0.00 0.00 0.00 1.00 ICWeak 0.05 0.22 0.00 0.00 0.00 0.00 1.00 Busy 0.67 0.47 0.00 0.00 1.00 1.00 1.00 Delay 65.45 31.50 8.00 55.00 61.00 73.00 986.00 Afiler 0.92 0.27 0.00 1.00 1.00 1.00 1.00 New 0.01 0.11 0.00 0.00 0.00 0.00 1.00 Number of Observations 15,609 Panel B. Non-Big-N Auditees AFee ($103) 274 411 5 70 141 295 5,103 Ln(AFee) 11.88 1.07 8.44 11.13 11.83 12.55 15.45 TA ($106) 171 706 0.00 10 32 107 21,338 Ln(TA) 17.08 1.84 7.34 15.94 17.07 18.28 23.52 Loss 0.52 0.50 0.00 0.00 1.00 1.00 1.00 ROA -0.14 0.47 -2.34 -0.17 0.01 0.09 0.38 Leverage 0.56 0.56 0.03 0.23 0.41 0.66 3.16 InvRec 0.30 0.23 0.00 0.11 0.27 0.46 0.97 ForOps 0.15 0.36 0.00 0.00 0.00 0.00 1.00 √Employees 0.63 0.78 0.00 0.20 0.37 0.72 8.25 Nsegments 1.65 1.21 0.00 1.00 1.00 2.00 9.00 NewFin 0.11 0.31 0.00 0.00 0.00 0.00 1.00 ExtDis 0.02 0.13 0.00 0.00 0.00 0.00 1.00 GCO 0.15 0.36 0.00 0.00 0.00 0.00 1.00 ICWeak 0.03 0.27 0.00 0.00 0.00 0.00 9.00 Busy 0.60 0.49 0.00 0.00 1.00 1.00 1.00 Delay 76.41 31.53 14.00 64.00 75.00 88.00 642.00 Afiler 0.76 0.42 0.00 1.00 1.00 1.00 1.00 New 0.09 0.29 0.00 0.00 0.00 0.00 1.00 Number of Observations 4,631 Variable definitions: See Appendix B.

38

Table 3: Benchmark models for computing residual fees charged by Big N and non-Big-N auditors in continuing engagements performed during 2001 to 2009 (panel estimation)

Variable & Predicted Sign

Big N Non-Big-N (1) (2) (3) (4)

Coefficient t-value Coefficient t-value Ln(TA) + 0.328 21.01 *** 0.311 13.09 ***

Loss + 0.040 4.40 *** 0.037 2.44 ***

ROA - -0.180 -5.57 *** -0.178 -4.90 ***

Leverage + 0.176 6.46 *** 0.108 3.66 ***

InvRec + 0.400 5.17 *** 0.209 2.10 **

ForOps + 0.056 3.72 *** 0.071 2.05 **

√Employees + 0.076 5.08 *** 0.120 2.18 **

Nsegments + 0.020 3.60 *** 0.022 1.55 *

NewFin + 0.020 2.54 *** 0.076 3.10 ***

ExtDis + 0.060 5.12 *** 0.089 1.96 **

GCO + 0.041 1.64 *** 0.019 0.55ICWeak + 0.239 12.70 *** 0.150 2.46 ***

Busy + 0.226 4.50 *** 0.071 0.41Delay + 0.002 7.56 *** 0.000 0.34Afiler + 0.021 1.33 *** 0.011 0.53Constant + 6.446 21.13 *** 6.249 15.22 ***

Model Fit Diagnostics

Number of observations 20,096 6,324 Number of distinct auditees 3,874 1,798 % Variation explained by auditee fixed effects (σu) 0.54 0.63 % Variation explained by pure error (σe) 0.30 0.30 Fraction explained by auditee fixed effects (σu

2/(σe2+σu

2)) 0.76 0.82 Ho: All Auditee Fixed Effects (ui)=0 569.56*** 39.44*** Adjusted R 2 (overall) 0.76 0.70 Adjusted R2 including auditee fixed effects 0.94 0.93

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using auditee-level fixed effects panel estimation and include (unreported) year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

39

Table 4: OLS regressions of current audit fees on an indicator for first year audits, lagged (panel) fee residuals and other fee determinants in continuing audits and first year audits for Big-N to Big-N, non-Big-N to non-Big-N and Big-N to non-Big-N switches.

Big-N Auditees Non-BigN Auditees Variable &

Predicted Sign (1) (2) (3) (4)

Coeff . t-value Coeff. t-value Test Variables New - -0.139 -4.45 *** -0.048 -2.46 *** Lresidual ? 0.761 105.70 *** 0.794 65.03 *** Control Variables Ln(TA) + 0.357 97.61 *** 0.328 56.92 *** Loss + 0.044 6.39 *** 0.028 2.55 *** ROA - -0.235 -14.73 *** -0.225 -12.73 *** Leverage + 0.138 13.34 *** 0.110 8.60 *** InvRec + 0.287 12.47 *** 0.125 4.40 *** ForOps + 0.080 13.17 *** 0.081 5.82 *** √Employees + 0.073 31.35 *** 0.094 9.20 *** Nsegments + 0.020 11.84 *** 0.021 5.35 *** NewFin + 0.013 2.07 ** 0.040 2.49 *** ExtDis + 0.072 7.14 *** 0.084 2.35 *** GCO + 0.039 2.03 ** 0.005 0.24 ICWeak + 0.281 17.92 *** 0.156 3.34 *** Busy + 0.183 31.51 *** 0.184 17.61 *** Delay + 0.002 12.59 *** 0.001 2.27 ** Afiler + 0.065 6.23 *** 0.052 4.16 *** Constant + 5.262 76.51 *** 5.511 53.74 *** N 15,609 4,631Number of Clusters 3,367 1,492Adjusted R2 0.92 0.92

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using auditee-level fixed effects panel estimation and include (unreported) year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

40

Table 5: Separate OLS regressions of current audit fees on lagged (panel) fee residuals and other fee determinants for continuing audits (Big-N and Non-Big N) and first year audits (Big-N to Big-N, non-Big-N to non-Big-N and Big-N to non-Big-N switches).

CONTINUING AUDITS FIRST YEAR AUDITS

Big N Auditees Non Big N Auditees Big N Switches Non Big N Switches Big to N Big Switches

Variable & Predicted Sign

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Coeff. t-value Coeff. t-value Pred.Sign Coeff. t-value Coeff. t-value Coeff. t-value

Test Variables LRFeePos + 0.713 59.44 *** 0.817 49.94 *** +/- 0.739 6.07 *** 0.793 14.39 *** 0.568 6.57 ***

LRFeeNeg + 0.810 64.14 *** 0.777 26.52 *** +/- 0.720 6.17 *** 0.764 10.56 *** 0.755 9.12 ***

Control Variables Ln(TA) + 0.357 103.20 *** 0.324 54.80 *** + 0.381 9.26 *** 0.346 16.40 *** 0.333 10.08 ***

Loss + 0.044 6.37 *** 0.028 2.50 *** + 0.119 1.23 -0.008 -0.17 0.094 1.44 *

ROA - -0.234 -14.83 *** -0.222 -12.08 *** - -0.212 -0.98 -0.265 -4.11 *** 0.005 0.04Leverage + 0.140 13.91 *** 0.110 8.39 *** + 0.063 0.53 0.098 2.32 ** 0.098 1.46 *

InvRec + 0.279 12.35 *** 0.134 4.54 *** + 0.484 1.78 ** 0.096 0.96 0.300 1.67 **

ForOps + 0.082 13.28 *** 0.081 5.74 *** + -0.034 -0.42 0.068 1.14 0.150 2.03 **

√Employees + 0.074 35.00 *** 0.095 9.70 *** + 0.107 3.54 *** 0.056 1.08 0.074 1.70 **

Nsegments + 0.020 11.84 *** 0.023 5.81 *** + -0.039 -1.39 * -0.005 -0.27 0.021 0.81NewFin + 0.012 1.95 ** 0.039 2.27 ** + 0.102 1.21 0.036 0.53 0.113 1.18ExtDis + 0.073 7.12 *** 0.081 2.16 ** + -0.018 -0.14 0.164 1.18 -0.034 -0.22GCO + 0.040 2.10 ** 0.000 -0.02 + -0.078 -0.46 0.046 0.71 0.150 1.54 *

ICWeak + 0.280 17.50 *** 0.153 3.29 *** + 0.431 3.35 *** 0.386 2.49 *** 0.611 5.27 ***

Busy + 0.179 30.31 *** 0.187 17.50 *** + 0.293 3.69 *** 0.134 3.23 *** 0.125 2.43 ***

Delay + 0.002 12.40 *** 0.001 1.77 ** + 0.004 5.61 *** 0.001 1.35 * 0.001 0.61Afiler + 0.063 6.14 *** 0.053 4.21 *** + -0.019 -0.14 0.017 0.28 0.113 1.76 **

Constant + 5.301 79.99 *** 5.551 51.66 *** + 4.367 5.65 *** 5.489 14.93 *** 5.489 9.46 ***

N 15,403 4,192 206 439 310Number of Clusters 3,353 1,431 204 405 309Adjusted R2 0.93 0.92 0.86 0.88 0.82

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using auditee-level fixed effects panel estimation and include (unreported) year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

41

Table 6: OLS regression of current audit fees on lagged (panel) fee residuals and other fee determinants for continuing audits (Big-N and non-Big-N) and first year audits (Big-N to Big-N, non-Big-N to non-Big-N and Big-N to non-Big-N switches).

Continuing and Laterally Switching Big N

Auditees

Continuing and Laterally Switching Non-Big-N

Auditees

Continuing Non-Big and Big to Non-Big Switching

Auditees

Variable & Predicted Sign (1) (2) (3) (4) (5) (6)

Coeff. t-value Coeff. t-value Coeff. t-value Test Variables New - -0.192 -3.72 *** -0.056 -1.78 ** 0.317 7.54 ***

LRFeePos + 0.713 59.36 *** 0.813 50.09 *** 0.810 50.30 ***

LRFeeNeg + 0.811 64.37 *** 0.777 26.78 *** 0.775 26.40 ***

New X LRFeePos ? 0.009 0.10 0.001 0.02 -0.056 -0.74New X LRFeeNeg ? -0.221 -1.94 * -0.027 -0.37 -0.136 -1.71 *

Control Variables Ln(TA) + 0.357 102.60 *** 0.327 57.14 *** 0.328 55.79 ***

Loss + 0.044 6.39 *** 0.028 2.53 *** 0.034 2.96 ***

ROA - -0.234 -14.72 *** -0.225 -12.75 *** -0.217 -11.77 ***

Leverage + 0.141 13.73 *** 0.110 8.68 *** 0.109 8.44 ***

InvRec + 0.282 12.48 *** 0.128 4.52 *** 0.150 4.96 ***

ForOps + 0.080 13.03 *** 0.080 5.69 *** 0.090 6.21 ***

√Employees + 0.075 34.62 *** 0.093 9.22 *** 0.093 9.31 ***

Nsegments + 0.019 11.62 *** 0.021 5.43 *** 0.023 5.63 ***

NewFin + 0.013 2.01 ** 0.040 2.46 *** 0.037 2.16 **

ExtDis + 0.072 7.12 *** 0.080 2.25 ** 0.071 1.93 **

GCO + 0.038 1.99 ** 0.004 0.21 0.001 0.06ICWeak + 0.282 17.87 *** 0.156 3.34 *** 0.179 3.40 ***

Busy + 0.180 30.35 *** 0.184 17.71 *** 0.182 16.95 ***

Delay + 0.002 12.67 *** 0.001 2.19 ** 0.001 2.15 **

Afiler + 0.063 6.16 *** 0.049 3.89 *** 0.062 4.82 ***

Constant + 5.287 79.35 *** 5.506 53.13 *** 5.439 51.00 ***

N 15,609 4,631 4,502 Number of Clusters 3,367 1,492 1,540 Adjusted R2 0.93 0.92 0.92 Test of sum of coefficients Coeff. F (1,3366) Coeff. F(1, 1491) Coeff. F (1, 1539) LRFeePos + New X LRFeePos 0.721 54.40 *** 0.811 270.61 *** 0.754 99.18 ***

LRFeeNeg + New X LRFeeNeg 0.588 27.06 *** 0.749 112.50 *** 0.639 73.27 ***

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using ordinary least squares (OLS) estimation and include (unreported) industry and year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

42

Table 7: Benchmark models for computing residual fees charged by Big N and non-Big-N auditors in continuing engagements performed during 2001 to 2009 (OLS estimation)

Variable & Predicted Sign

Big-N Auditees Non-Big-N Auditees (1) (2) (3) (4)

Coeff. t-value Coeff. t-value Ln(TA) + 0.457 43.04*** 0.492 39.55***

Loss + 0.064 4.96*** 0.113 5.06***

ROA - -0.368 -12.20*** -0.305 -10.51***

Leverage + 0.168 6.54*** 0.173 7.96***

InvRec + 0.510 8.31*** 0.132 2.05 **

ForOps + 0.214 12.36*** 0.242 6.46***

√Employees + 0.045 4.89*** 0.078 2.76***

Nsegments + 0.034 6.61*** 0.028 2.61***

NewFin + 0.043 3.40*** 0.076 2.28 **

ExtDis + 0.196 10.60*** 0.360 5.93***

GCO + 0.064 2.35*** -0.009 -0.26

ICWeak + 0.384 15.96*** 0.218 2.79***

Busy + 0.104 5.97*** 0.067 2.45***

Delay + 0.002 9.77*** 0.000 0.56

Afiler + 0.057 3.20*** 0.024 0.89

Constant + 3.864 19.59*** 2.514 11.88***

Number of observations 20,096 6,324 Adjusted R2 0.82 0.74

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using ordinary least squares (OLS) estimation and include (unreported) industry and year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

43

Table 8: OLS regressions of current audit fees on an indicator for first year audits, lagged (OLS) fee residuals and other fee determinants in continuing audits and first year audits for Big-N to Big-N, non-Big N to non-Big N and Big-N to non-Big N switches.

Big-N Auditees Non-Big-N Auditees

Variable & Predicted Sign

(1) (2) (3) (4) Coeff. t-value Coeff. t-value

Test Variables New - -0.138 -4.31 *** -0.047 -2.37 *** Lresidual ? 0.757 96.77 *** 0.805 59.71 *** Control Variables Ln(TA) + 0.464 129.10 *** 0.477 93.46 *** Loss + 0.060 8.17 *** 0.068 6.11 *** ROA - -0.374 -22.13 *** -0.310 -17.71 *** Leverage + 0.202 18.62 *** 0.189 14.54 *** InvRec + 0.496 20.58 *** 0.149 5.06 *** ForOps + 0.193 30.03 *** 0.191 13.76 *** √Employees + 0.049 20.04 *** 0.081 9.17 *** Nsegments + 0.035 20.40 *** 0.029 7.65 *** NewFin + 0.007 1.12 0.038 2.25 ** ExtDis + 0.120 10.96 *** 0.166 4.43 *** GCO + 0.065 3.17 *** 0.027 1.41 * ICWeak + 0.293 18.12 *** 0.142 3.04 *** Busy + 0.100 16.51 *** 0.070 6.74 *** Delay + 0.002 14.01 *** 0.001 3.83 *** Afiler + 0.070 6.67 *** 0.028 2.19 ** Constant + 3.121 46.30 *** 2.942 32.18 *** N 15,609 4,631Number of Clusters 3,367 1,492Adjusted R2 0.92 0.91

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using auditee-level fixed effects panel estimation and include (unreported) year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.

44

Table 9: Separate OLS regressions of current audit fees on lagged (OLS) fee residuals and other fee determinants for continuing audits (Big-N and non-Big N) and first year audits (Big-N to Big-N, non-Big N to non-Big-N and Big-N to non-Big-N switches).

CONTINUING AUDITS FIRST YEAR AUDITS Big N Auditees Non-Big-N Auditees Big N Switches Non-Big N Switches Big to non-Big Switches

Variable & Predicted Sign

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Coeff. t-value Coeff. t-valuePred. Sign Coeff. t-value Coeff. t-value Coeff. t-value

Test Variables LRFeePos + 0.673 44.53 *** 0.843 36.13 *** +/- 0.701 4.89 *** 0.741 10.55 *** 0.527 5.61 ***

LRFeeNeg + 0.837 56.19 *** 0.780 20.83 *** +/- 0.766 6.27 *** 0.766 9.93 *** 0.790 7.21 ***

Control Variables Ln(TA) + 0.463 137.30 *** 0.475 89.28 *** + 0.474 10.71 *** 0.476 21.96 *** 0.433 13.40 ***

Loss + 0.060 8.24 *** 0.071 6.14 *** + 0.131 1.34 0.023 0.47 0.097 1.44ROA - -0.371 -22.15 *** -0.307 -17.10 *** - -0.374 -1.76 -0.331 -4.64 *** -0.137 -1.14 ***

Leverage + 0.206 19.73 *** 0.192 14.45 *** + 0.157 1.27 0.156 3.31 *** 0.145 2.05 ***

InvRec + 0.487 20.88 *** 0.163 5.36 *** + 0.694 2.40 ** 0.070 0.66 0.499 2.69ForOps + 0.193 30.06 *** 0.191 13.36 *** + 0.036 0.44 0.171 2.82 *** 0.269 3.53 ***

√Employees + 0.050 23.27 *** 0.083 9.30 *** + 0.083 2.67 *** 0.050 0.94 0.036 0.80Nsegments + 0.035 19.81 *** 0.032 8.05 *** + -0.019 -0.71 * 0.003 0.17 0.036 1.40NewFin + 0.006 0.95 0.037 2.15 ** + 0.106 1.22 0.013 0.18 0.126 1.22ExtDis + 0.120 10.84 *** 0.153 3.96 *** + 0.009 0.07 0.341 2.48 *** -0.018 -0.12 ***

GCO + 0.065 3.20 *** 0.017 0.86 + -0.110 -0.65 0.093 1.34 * 0.204 2.07 *

ICWeak + 0.291 17.59 *** 0.138 3.02 *** + 0.482 3.80 *** 0.384 2.19 ** 0.609 5.11 **

Busy + 0.097 15.93 *** 0.071 6.67 *** + 0.207 2.60 *** 0.043 0.98 0.051 0.95Delay + 0.002 13.98 *** 0.001 3.36 *** + 0.004 6.07 *** 0.001 1.65 ** 0.001 0.63 **

Afiler + 0.072 6.82 *** 0.032 2.55 *** + 0.048 0.35 -0.041 -0.69 0.125 1.86Constant + 3.178 49.33 *** 2.946 31.38 *** + 2.454 2.94 *** 3.180 8.46 *** 3.544 6.42 ***

N 15,403 4,192 206 439 310Number of Clusters 3,353 1,431 204 405 309Adjusted R2 0.92 0.92 0.86 0.86 0.80

Variable definitions: See Appendix B. The dependent variable is natural logarithm of audit fees. The models reported are estimated using auditee-level fixed effects panel estimation and include (unreported) year indicators. ***,**,* indicates significance at the 0.01, 0.05, and 0.10 level or better, respectively. Significance levels are based on one-tailed tests for directional predictions, two-tailed tests otherwise. t-statistics are based on heteroskedasticity corrected and auditee-level clustered standard errors.