anJ Binine.ys Reieunli. Vol. 29. No. .1. pp. 217-227, 1999 217

Are large auditors more accurate thansmall auditors?Clive Lennox*

.•\bstract—Theoretical research suggests that large auditors have more incentive to issue accurate reports comparedto small auditors (DeAngelo. 1981; Dye, 1993). Controiling for the client characteristics of large and small auditors,this paper shows that large auditors issue reports that are more accurate and more informative signals of financialdistress. These findings are consistent with the theoretical prediction of a positive relationship between auditor sizeand auditor accuracy.

1. IntroductionExisting theories predict that large auditors havemore incentive to issue accurate reports.' Reputa-tion arguments suggest that large auditors suffer agreater loss of rents as a result of inaccurate re-porting (DeAngelo, 1981). Moreover, auditorshave more incentive to give accurate reports, thegreater is the litigation penalty that is suffered forinaccuracy (Dye, 1993). Since large auditors havedeeper pockets than small auditors, they shouldhave more incentive to issue accurate reports.'

Much empirical evidence indicates that large au-dit firms are associated with higher quality. Largeauditors tend to receive higher fees than small au-ditors (Simunic and Stein, 1987; Beatty. 1989); in

•This paper is based on the author's PhD, thesis whilst atNuffield College, Oxford University. He would like to thankAnindya Banerjee, Steve Bond and anonymous referees forhelpful comments. Thanks are also owed to seminar partici-pants al the Institute of Economics and Statistics at OxfordI 'niversity. Financial assistance from the ESRC is gratefullyjcknowledged. Correspondence should be addressed to DrLennox at the Department of Economics, University of Bristol.8 Woodlands Road, Bristol BS8 ITN. The final version ofihis paper was accepted in January 1999. Email:C, Lennoxto.bristol.ac.uk

' Between 1987-94. the large UK audit firms were known asthe 'Big Six', comprising Price Waterhouse, KPMG Peat Mar-wick. Arthur Andersen. Touche Ross, Coopers & Lybrand andIirnst & Young. All other audit firms are referred to as smallin this paper,

- Large audit firms have deeper pockets than small audittirms because of joint and several liability. The Econoniislwrites (7 October, 1995): "As partnerships, the large account-ancies operate under ihe legal principle of joint and severalliability. This means that when a company collapses, its audi-tors who not only have deep pockets, but cannot abscond, maybe hit for the entire bill if they were negligetit. even if otherparties were careless too. Moreover, if the claim amounts tomore than an auditing firm's capital, all of the firm's partnersare liable righl down to their bootstraps for the bill—even ifihey had nothing lo do with the error." Similarly, the FinamialTimes writes (4 July, 1996): 'The big audit firms can find them-selves targeted for lawsuits because of their "deep pockets"—including their statutory insurance cover,'

IPOs, high reputation investment banks prefertheir clients to hire large auditors,, and companiesthat do hire large auditors are charged a smallerbanking fee (Menon and Williams, 1991; Balverset al., 1988). These results suggest that large au-ditors offer higher quality services; however, qual-ity is not really the same thing as accuracy. Largeauditors may give a higher quahty service, but thisdoes not necessarily mean that large auditors aremore accurate.

Theories on auditor selection have argued thata company has more incentive to hire an accurateauditor when it has favourable information andwhen agency costs are high. This is because anaccurate auditor is more likely to attest to thecompany's private information and because hiringan accurate auditor can be a signal to investorsthat the company has favourable private informa-tion (Titman and Trueman, 1986; Datar et al.,1991). Empirical studies have tested these theoriesby examining the stock market reaction to differ-ent types of auditor switch and by examining therelationship between agency costs and auditorchoice. If companies with favourable privateinformation prefer to hire large auditors, the stockmarket should react more favourably when com-panies switch to large auditors than to smallauditors.

Some studies have found that the stock marketreacts in the hypothesised direction (Nichols andSmith. 1983; Eichenseher et al., 1989); otherstudies have found no significant difference inmarket reaction for different types of switch (Firthand Smith, 1992; Johnson and Lys, 1990; Schwartzand Soo, 1995). Unfortunately, the event study ap-proach is not a useful way of testing whether largeauditors are more accurate, because a switch to alarge (small) auditor may signal good (bad) newsfor reasons that have nothing to do with accuracy.In particular, a company may switch to a large(small) auditor if it is expecting to grow (decline).Thus, the stock market reaction may reflect a com-

218 ACCOUNTING AND BUSINESS RESEARCH

Figure 1An auditor's expectations abont its clients' bankruptcy probabilities

("(P)

CuI-t)IT

pany's product market opportunities rather thandifferential accuracy between auditors. Studieshave generally confirmed the predicted relation-ship between agency costs and auditor choice—companies with high agency costs are more likelyto hire large auditors (Francis and Wilson, 1988;Johnson and Lys. 1990; DeFond. 1992). Althoughthis is consistent with the view that large auditorsare more accurate, there is an alternative explana-tion—managers may simply value the prestige thatcomes from being associated with large auditors inmuch the same way as they might value non-profit-maximising profits.

In contrast to these studies, this paper directlyinvestigates the relationship between auditoraccuracy and auditor size. Using a bankruptcymodel to control for differences in client charac-teristics, it shows that large auditors' reports aresignificantly more accurate indicators of financialdistress compared to small auditors' reports. Thenext section describes the methodology used toevaluate the accuracy and informativeness of auditreports. The following sections describe the dataand present the results.

2. Research methodology

2J. Evaluating the accuracy of audit reportsIn evaluating the accuracy of audit reports, it is

helpful to consider Figure I. The horizontal axisshows an auditor's expectations about its clients'bankruptcy probabilities (P); the vertical axisshows the density function f(P). Figure 1 illustratesthe case where most companies have low expectedbankruptcy probabilities. An auditor is assumed tochoose a "cut-off probability" P*, which determinesthe number of companies that are given qualifiedreports; companies with low bankruptcy prob-

abilities (those lying to the left of P*) are givenunqualified reports while companies with highbankruptcy probabilities (those lying to the rightof P*) are given qualified reports.

The main difficulty in measuring the accuracyof audit reports is that one does not directly ob-serve whether companies deserve clean or qualifiedaudit opinions. Therefore, previous research hasused the bankruptcy outcome as an ex postmeasure of whether a company should have beengiven a qualified report. This approach is not aperfect measure of accuracy—for example, a com-pany could enter bankruptcy after being given anunqualified report if there were no obvious prob-lems when the audit was undertaken; similarly, acompany could survive following a qualified reporteven if there were clear financial problems. There-fore, previous research has used bankruptcy mod-els as a benchmark for evaluating the accuracy ofaudit reports {Altman and McGough. 1974; Koh,1991).

For financially distressed companies that sur-vive, the bankruptcy model is likely to predictbankruptcy and the auditor may qualify; similarly,for companies that are apparently healthy prior tofailure, the bankruptcy model may incorrectly pre-dict survival and the auditor is likely to give anunqualified report. If bankruptcy is difficult (easy)to predict, one would expect error rates to be high(low) for both the model and auditors. If auditorshave access to private information that is usefulfor identifying failing companies, one might expectaudit reports to be more accurate than the model.If audit reports are affected by factors other thanthe probability of bankruptcy, one might expectaudit reports to be less accurate than the model.

SUMMER 1999 219

This paper compares the accuracy of audit re-ports and a bankruptcy model by measuring twoerrors. Following Koh (1991), the type I error rateis equal to the proportion of failing companies thatare given unqualified reports and the type II errorrate is equal to the proportion of non-failing com-panies that are given qualified reports.' Practition-ers might argue that Koh's definition of a type IIerror is problematic, since a qualified report sig-nals a potential problem rather than predictsbankruptcy. The reason for calculating both typeI and type II errors is best explained by consider-ing an extreme case, where an auditor always givesqualified reports (i.e. the auditor chooses P*^0).If one ignored type II errors and simply measuredaccuracy in terms of type I errors, one would havethe misleading impression that this uninformativereporting strategy maximises accuracy.

Calculating both type I and type II errors helpsclarify an important distinction between theaccuracy and conservatism of auditors. For agiven distribution of bankruptcy probabilities, anincrease in auditor conservatism raises the fre-quency of qualified reports (a more conservativeauditor chooses a lower P*). Therefore, a moreconservative auditor has lower type I and highertype II error rates for a given level of accuracy.On the other hand, a more accurate auditor haslower type I and type II error rates, for a givenlevel of conservatism (i.e. for a given choice of P*).

2.2. Evaluating the incremental informationcontent of audit reports

In addition to measuring the accuracy of auditreports, it is useful to test whether reports signaluseful incremental information about the prob-ability of bankruptcy (Hopwood et al., 1989). Theaccuracy of audit reports is not necessarily relatedto their incremental information content. For ex-ample, auditors might give going-concern qualifi-cations (clean opinions) when their privateinformation is less (more) favourable than publicinformation (Grout et al., 1994). In this view, au-dit opinions are used to signal auditors" privateinformation and do not fully refiect all publicinformation. Thus, it is theoretically possible thatthe incremental information content of audit re-ports could be high (low) even if accuracy is low(high). The role of public and private informationis shown in equations (1) and (2):

where:

F A I L S / - (1)

•* This does not imply ihitt ii qualified report is necessarilyequivalent to a prediction of bankruptcy. For example, if anauditor chooses a P* which is less than 0.5. qualified reportswould be given to companies which the auditor believes aremore likely to survive than fail. Table 5 shows that, on average,large (small) auditors give going-concern qualifications to com-panies whose bankruptcy probabilities exceed \%% (15%).

(2)

1 if company i issues its final annualreport in year t prior to entering bank-ruptcy;0 otherwise. F A I L S / > 0 ifFAILS, = 1; FAILSi,*<0 ifFAILS,, = 0.1 if company i receives a going-concernqualification in year t; = 0 otherwise.GQ,,*>0 ]f GQ,= l: GQ,,*<0 if

GQ,,

X,| = Publicly observable variables capturingfinancial health.

Zn = Privately observable variables captur-ing financial health.

Equation (1) states that the likelihood of failuredepends on variables that capture the financialhealth of the company. Some of these variables arepublicly observable {X,,); others are observable tothe auditor but not publicly observable (Z,,). Equa-tion (2) allows audit reporting to depend on thefinancial health of companies (X,, and Z|,). It isassumed that Efu , u,|J = O. This makes sense be-cause there is usually a lag of several months be-tween the audit reporting decision and the bank-ruptcy event.

One would like to test whether audit reports sig-nal valuable private information about the prob-ability of bankruptcy. However, one cannotdirectly test the null hypothesis y, = 0 because thestatistician does not observe Z,j. This problem canbe circumvented by including the audit report vari-able (GQ,|) in the bankruptcy model as shown inequation (3):

FAILS,,* = (3)

Under the null hypothesis that audit reports donot signal valuable private information about theprobability of bankruptcy, y^^^^O. If y i ^^O itmust also be true that E[GQ,, Vi,] = O, which impliesthat ^ , ^ 0 . If reports do not signal valuable pri-vate information about the probability of bank-ruptcy, audit reports should not be useful for pre-dicting bankruptcy after controlling for allpublicly observable financial distress variables(Xjj). Under the alternative hypothesis, a going-concern qualification is a signal of financial dis-tress (y,^-,>0) and audit reports should be usefulfor identifying failing companies (^3>0). Underthe null hypothesis that large and small auditors'reports are equally informative, one should be un-able to reject the null hypothesis that ^, is the samefor large and small auditors. Under the alternativehypothesis, large auditors' reports have greaterincremental information content and fi^ is signifi-

220 ACCOUNTING AND BUSINESS RESEARCH

cantly more positive for large auditors than forsmall auditors.

3. Data3.]. The sample

An attempt was made to collect data on all pub-licly quoted UK companies between 1987-94/ Forfailing companies, data were collected from all ac-counts published from 1987 until the first publicannouncement of failure; for non-failing com-panies, data were collected in all available yearsbetween 1987-94.^

Firstly, information on companies' auditors andaudit reports was collected from microfiche copiesof companies' annual reports.'' From each report,it was noted whether the auditor had given a quali-fied or unqualified audit opinion. When there is asignificant possibility of the company ceasing totrade in the foreseeable future, the auditor is re-quired to state that the accounts give a true andfair view subject to the company remaining a go-ing concern. The 'Auditing Guideline on GoingConcern' (August 1985) states: The going concernconcept identified in Statement of Standard Ac-counting Practice No. 2 is "that the enterprise willcontinue in operational existence for the foresee-able future". This means in particular that theprofit and loss account and the balance sheet as-sume no intention or necessity to liquidate...Theforeseeable future should normally extend to aminimum of six months following the date of the

^ A UK company is so defined if ils headqiiiirters and auditorwere based in the UK. Many of the companies in the samplehad overseas operations and overseas shareholders.

^ There were no cases where failing companies Issued ac-counts alter the first public announcement of failure.

'' These are located in the Corporate Information Library alWarwick University. Data on companies that were taken overwere collected up lo the point of takeover, and thereafter thesecompanies were treated as missing observations.

audit report or one year after the balance sheetdate whichever periods ends on the later date.'

In the sample, qualifications were given for fun-damental uncertainties and non-compliance withStatements of Standard Accounting Practice(SSAPs) as well as for going-concern reasons.'Failing companies that received qualified reportsmore than one year prior to failure were includedin the analysis so as to calculate type I and typeII error rates for alternative bankruptcy horizons.^A list of companies that entered administration,receivership or liquidation was obtained fromStock Exchange Financial Yearbooks.'' All suchcompanies are defined in this paper as failing. Mi-crofiche data were available for 1,086 companiesof which 123 failed between 1987-94.

^ 'Fundamental uncertainties" arose where auditors were un-certain about provisions made for tax, slow-moving stocks, baddebts or litigation—72 companies received qualified reports forissues other than going-concern while 17 companies receivedreports that were qualified for both going-concern and non-going-concern issues. The former were coded with a zero, whilethe latter were coded with a one.

^ For esample, a qualification two periods prior to failure isclassified as a type 11 error when the bankruptcy horizon is oneperiod and is classified as an accurate report when the horizonis two or more periods. Longer bankruptcy horizons have thedisadvantage that there is a greater probability of mis-classi-fying a type 11 error as an accurate report. As shown in Table2. the conclusions of this study are not sensitive to differentbankruptcy horizons.

In the UK. there are three types of reorganisation pro-cedure when a company becomes insolvent—liquidation, re-ceivership and administration. In a liquidation, the assets ofthe company are sold so as to meet the claims of creditors. Ina receivership, the receiver can decide whether it is in the cred-itors' interests to sell the company's assets. Generally, it is inthe creditors" interests to liquidate if the liquidation value ofthe company exceeds its going concern value. The possibilityof administration was introduced by the 1986 Insolvency Actin an attempt to reduce the number of inefficient liquidations.However, few administrators have been appointed comparedto the numbers of companies entering receivership or liquida-tion. The results m this paper are not sensitive to different kindsof reorganisation procedures.

Table IFinancial health and audit reporting (1987-94)

1987 1988 1989 1990 199! 1992 1993 1994 Tola!

s,F.N FA ILS,NGQ,

79945i

8625133

9036313

92?363520

925462627

90841826

89715322

8837221

7,104160123123

S, = Number of observatiotis iti year t for which tnicrofiche data were available.F,^Number of UK publicly quoted companies in the population which entered bankruptcy in year t.NFAILS, = Number of companies in the sample which issued their final annual reports in year t, prior toentering bankruptcy.NGQ, = Number of companies which received going concern qualifications in year t.

SUMMER 1999 221

Next, a search was made of Datastream for dataon the number of employees, cashflow ratios, prof-itability ratios and leverage ratios.'" Data on eachcompany's Standard Industrial Classification(SIC) codes were collected from Extel, because in-dustry sector is likely to be an important deter-minant of bankruptcy. Data for some of the 1086companies in the initial sample were unavailablefrom Datastream and Extel. The final sample com-prised 976 companies of which 90 failed over theeight-year period. The frequency of failure in thesample is l.73yo which is approximately equal tothe population frequency (Morris, 1997).

3.2. Bankruptcy and audit reporting between1987-94

Table 1 shows the pattern of bankruptcy andaudit reporting over the period 1987-94. Row 1shows the total number of companies for whichaudit data were available in the initial sample of1,086 companies. The number of companies in thefinal sample (976) exceeds the number of obser-vations in each year due to missing observations."Row 2 shows the number of failing companies inthe population. Given the timing of the economiccycle, it is unsurprising that most failures occurredbetween 1990-92. A comparison of rows 2 and 3shows that there is an average lag of around a yearbetween the last annual report of a failing com-pany (FAILS,) and its entry into bankruptcy (F,).'-Companies which received going-concern qualifi-cations were coded with a 1. whilst those that didnot were coded with a zero. Row 4 shows that thenumber of going-concern qualifications jumped upin 1990 and remained high during the recovery of1993-94.

3.3. The accuracy of large and small auditors'reports

Panel A of Table 2 shows that both type I andtype II error rates are smaller for large audit firmsthan for small audit firms. The difference in typeI error rates was statistically significant at the 10%level, while the difference in type II error rates was

'" These variables are defined in Table 3." Conipiinies with observations that were missing for rea-

sons other than bankruptcy or take-over were not droppedfrom the sample because this could cause problems of sampleselection bias. There are two reasons why the presence of miss-ing observations is not likely to cause problems. First, there arenot very many missing observations—the majority of com-panies (653) have a complete panel of data covering all eightyears. Secondly, when the sample proportion of companies dif-fers from the population proportion, the logit model has con-sistent coefficient estimates for all variables except the constant(Anderson, 1972). For all models reported in this paper, theresults from probit and logit models were found to be verysimilar and so sample selection bias does not seem to be aproblem.

'• The average tag was in fact 14 months.

not significant. Panel B reports the accuracy ratewhich is equal to the number of type I and type IIerrors divided by the total number of reports. Forexample, large auditors committed 42 type I errorsand 58 type II errors in 4511 audit reports; there-fore the accuracy rate for large auditors is 97.78%[(4511-^2-58)/451]]. The results indicate that thereports of large auditors were significantly moreaccurate than those of small auditors (at the 1%level). The difference in large and small auditors'accuracy rates was highly significant because bothtype I and type II error rates were lower for largeauditors."

In panels A and B the bankruptcy horizon wasassumed to be a single reporting period—thisdefinition makes sense because of the AuditingGuideline's definition of the 'foreseeable future'.However, the result that large auditors' reports aremore accurate is robust to considering alternativebankruptcy horizons. Panel C measures theaccuracy of reports when the horizon is two re-porting periods (a type I error occurs when a fail-ing company does not receive a qualification in itsfinal or penultimate reports; a type II error occurswhen a qualified report is not followed by bank-ruptcy within two reporting periods). Once again,large auditors' reports were significantly moreaccurate than small auditors' reports—similar re-sults were found for bankruptcy horizons of morethan two periods.

The superior accuracy of large auditors' reportscould have occurred because large auditors weremore accurate or because it was easier to predictbankruptcy for large auditors' clients. For exam-ple, large auditors' clients tend to be large and areunlikely to fail. Therefore, a bankruptcy modelwas used as a benchmark for evaluating theaccuracy of audit reports. If it is easier to predictbankruptcy for large auditors' clients, one shouldfind that the bankruptcy model is more accuratefor large auditors' clients. If large auditors are

'-' When large and small auditors" clients were matched onthe basis of size, large auditors" reports were again found to bemore accurate than those of small auditors. The data were par-titioned Into three groups;Group I: Company has < 15,000 employees and hires a largeauditor (.1,971 observations)Group 2: Company hires a small auditor {1,905 observations)Group 3: Company has > 15.000 employees and hires a largeauditor (540 observations)The mean number of employees in Groups 1, 2 and 3 were2,076, 2,081 and 52,556 respectively. If auditor accuracy de-pends on company size rather than auditor size, one wouldexfrect to find no difference in accuracy rates for Groups I and2 which have similar size companies. In Group 1, the type Iand type II error rates (76.36% and 1.48%) were both lowerthan small auditors' error rates (91.43V,, and 1.76Vr,). The dif-ference between large and small auditors' type I error rates wasstatistically significant at the WA\ level (;f- = 3.32*), while thedifference between type II error rates was not significant. Theseresults are similar to those reported in Table 2, where com-panies were not sorted into size groups.

222 ACCOUNTING AND BUSINESS RESEARCH

Table 2The correlation between auditor accuracy and auditor size

Panel A Type I and Type If error rates of large and small audit firms(Bankruptcy horizon = 1 year)

FAILS,, = 1Type 1 error (%)

FAILS,, = 0Type II error (%)

Large audit(AUD,,=

GQ,,=O42

76.36

4.3981.30

firms1)

GQ,,= I13

58

Small audit firms(AUD,,-O)

GQ,, = O GQ,,= 132 3

91.43

1.837 331.76

forequal error ratestarge and smallaudit firms

X2 = 3.28*

X' = l 92

Panel B Overall accuracy of large and small audit firms(bankruptcy horizon — 1 year)

Large audit firms Small audit firms H,,: equal accuracy for(AUD,, = 1) (AUD,, = O) large and small audit

finnsAccurate Inaccurate Accurate Inaccurate

4,41 i 100 1,840 65Accuracy (%) 97.78 96.59 X= = 7-63***

Panel C Overall accuracy of large and small audit firms(bankruptcy horizon=2 years)

Large audit firms Small audit firms H,,: equal accuracy for(AUD,,= 1) (AUD,, = O) large and small audit

firmsAccurate Inaccurate Accurate Inaccurate

4,360 151 1,813 92Accuracy (%) 96.65 95.17 X'=13.00***

•Significant at the O.I level (x^oi (1 d.f.) = 2.71)••Significant at the 0.05 level (x\,os d d.f.)-3.84)***Significant at the 0.01 level (x',,,,, (1 d.f.) = 6.63)When the bankruptcy horizon is one year, the failure variable is FAILS,,FAILS,, = I if company i issued its final report in year t prior to entering bankruptcy (90 observations); = 0otherwise (6,326) observations).When the bankruptcy horizon is two years, the failure variable is FAILS,,,,^,,FAILS,,,,+1^^ I if company i issued its final or penultimate report in year t prior to entering bankruptcy (178observations); ^ 0 otherwise (6,238 observations)GQ,,^1 if company i received a going-concern qualification in year t (107 observations); = 0 otherwise (6,309observations)AUD,,= I if company was audited by a Big Six auditor in year t; = 0, otherwise

more accurate than small auditors, one should find Previous research also indicates that cashflow, lev-that large auditors" reports are more accurate, erage, profitability and company size are relatedeven taking into account differences between large to financial distress (Altman, 1986). Data on theseand small auditors" clients. The explanatory vari- variables were obtained from Datastream and areables included in the model are described in the defined in Table 3.'-*next section. The return on capital variable (ROC,,) measures

profitability whilst capital gearing (CAPG,,) is ameasure of leverage. The gross cashflow (GCFi,),

3.4. Financial distress variablesData on the main activities of each company u c . J J . . i , .k .i. u .•

* / ' Some studies used total assets rather than number ot em-were collected because the effects of the economic pioyees to measure company size. However, assets were foundcycle are likely to differ across industry sectors. lohavelower predictive content than the number of employees.

SUMMER 1999 223

Table 3Company

Variable

EMP,,

Dl

D5

D8

DBTNi,

CASH,,

GCF,

CAPG,,

ROC,

size, industry sector, cashflow, leverage and profitability variables

Definition

Total number of employees

= 1 if company i operates in energy or water; ^ 0 otherwise

= 1 if company i operates in construction; ^ 0 otherwise

— 1 if company i operates in banking, finance or insurance; =0otherwise

(Total sales) X 100Total debtors

(Total cash) X 100Current liabilities

(Profits earned for ordinarysharehoIders+Depreciation+Tax equalisation) x 100

Capital employed+Current liabilities-Intangibles

(Preference capital+Subordinateddebt+Loan capital+Short-term borrowings) x 100

Capital employed+Short-term borrowing-Intangibles

(Total interest charged+Pre-tax profit) x 100Capital employed+Short-term borrowing-Intangibles

DalastreamCode

219

726

743

735

731

707

Interpretation

Company size

Industry effects

Industry effects

Industry effects

Debtor turnoverratio

Cash ratio

Gross cashflowratio

Capital gearingratio

Return on capitalratio

debt-turnover (DBTN,,) atid cash (CASHj,) ratiosmeasure cashflow. GCF,, measures profit-gener-ated cashflow. DBTNj, captures the fact that com-panies are more likely to fail if they are havingproblems in receiving payments for past sales.CASHj, is a measure of short-term liquidity. OtherDatastream financial ratios tried were incomegearing (leverage measure), credit-turnover, stock-turnover, the working capital ratio and the quickratio (cashflow measures).''^ These variables werefound to have insignificant and/or non-constant ef-fects on bankruptcy and were therefore omittedfrom the model.

4. ResultsTable 4 reports the results for two bankruptcymodels. Model 1 is used to evaluate the accuracyof audit reports while model 2 tests the incre-mental information content of going-concern qual-ifications. Both models show that cashflow (GCF,,)and leverage (CAPG,,) have non-linear effects onthe probability of bankruptcy—these tion-lineari-

'^ All these nitios are defined by Datastream. and have thefollowing codes; Income gearing (Datastream code, 732); Cred-itors' luniovcr (728): Stock turnover (724); Working capital ra-tio (741); and the Quick ratio (742),

ties were captured by including polynomial varia-bles (CAPG,,^ CAPG,' and CAPG/ and GCF,,^).Lennox (1999) has shown that a failure to takeaccount of these non-linearities causes heterosce-dasticity problems."'

The negative coefficients on company size(EMP,,) indicate that small companies were morelikely to fail. The Industry dummies show bank-ruptcy was more likely for companies operating inthe utility (Dl,), construction (D5,) or finance (D8,)sectors. The negative coefficients on the debtor-turnover (DBTN,,) and cash (CASH^,) ratios showthat companies were more likely to fail if paymentsfrom past sales were low or if cash reserves fell.The negative coefficients on gross cashflow (GCF|,)

"• Langrange Multiplier tests for omitted variable bias andheteroscedasticity show that the bankruptcy model does notsuffer from mis-specification problems. Lennox (1999) alsotested the model using separate estimation (1987 90) and hold-out (1991 94) samples, and found that the coefficient estimatesare stable over the period 1987-94. Standard errors were cal-culated using the Huber bootstrapping technique which givesconsistent standard errors under very weak assumptions; forexample, the Ibrmula does not require the standard assump-tions of no serial correlation or homoscedasticity (Huber,1967). Huber's formula is a canned part of the STATA statis-tical package and can be used in conjunction with the probitcommand.

224 ACCOUNTING AND BUSINESS RESEARCH

Table 4Bankruptcy models(z-statistics in parentheses)

Control variables

EMP,,

AUD,,* EMP,,

Dl,

D5,

D8,

DBTN,,

CASH,,

GCF,,

CAPG,,

CAPG,,

CAPG,,'

CAPG,/

ROC,,

AUD.,

CONSTANT

Experimental Variables

GQ,,

AUD,,* GQi,

Number of ObservationsNumber of companies

***Significant at the 0.01 level. **Significant atAll variables defined in Tables 2 and 3.

Model 1

-0.379e-04(-2.581)***

0.432(1.810)*0.328

(2.506)**0.398

(3.754)***- 0.200e - 03

(-2.010)**-0.472e-02

(-2.325)**-0.215e-03

(-2.340)**-0.205e-03

(-2.072)**-0.245e-03(7.698)***

-0.845e-08(-5.344)***

O.587e-13(4.802)***

-O.I]2e-I8(-4.491)***-0.323e-04

(-1.771)

-2.822(-19.519)***

6416976

the 0.05 level. *Significant at the 0.

Model 2

-O.llOe-05(-0.064)-0.606e-04

(-2.634)***0.439

(1.869)*0.368

(2,740)***0.406

{3.790)***-O.186e-O3

(-1.837)*-0.435

(-2.087)**-0.902e-02

(-1.334)-O.191e-O3

(-1.764)*0.254e-03

(8.046)***-0.877e-08

(-5.435)***0.613e-]3

(4.914)***-0.118e-]8

(-4.630)***-0.4I4e-04

(-1.559)-0.150

(-1.320)-2.806

(-18.462)***

-0.398(-1.135)

0.777(1.948)*

6416976

level.

also show that a compatiy was more likely to gobankrupt when profit-generated cashflow was low.

The positive coefficient on capital gearing(CAPG,,) shows that highly-levered companieswere more likely to fail. While the polynomialterms CAPG,,- and CAPG,,"* have negative coeffi-cients, the effect of leverage on bankruptcy wasmonotonically positive for values of CAPG,, lyingwithin two standard deviations of its mean. Fi-nally, the negative coefficient on profitability

implies that companies with low (or neg-ative) profits were more likely to fail.

Table 5 compares model l's predictive accuracyfor large and small auditors' clients to investigatewhether the positive correlation between auditoraccuracy and auditor size is due to client-specificor auditor-specific factors. Panel A reports thetype I and type II error rates for auditors andmodel 1 (auditors' error rates are the same asPanel A of Table 2). The cut-off probabilities (P*)

SUMMER 1999 225

Table 5Predictive accuracy of audit reports and model 1 for large and small auditors^ clients

Panel A Type I and Type II error rates of large and small audit firms

Large audit firms(AUD.,-1)

GQ,, = 0 GQ,, = 1FAILS;,-1 42 13Type I error (%) 76.36

FAILS,,-0 4,398 58Type II error (%) 1.30

Small audit firms(AUD,,-O)

GQ,,-0 GQ,,= l32 3

91.43

1,837 331.76

Type I and Type II error rates of model

Large audit firms'clients (AUD,, = 1)

Cut-ofr(P^) 0.187SURVIVE FAIL

FAILS,, = 1 40 15Type I error (%) 72.73

FAILS, -0 4,400 56Type 11 error (%) 1.26

Small audit firms'clients (AUD,,-O)

0.151SURVIVE FAIL

27 877.14

1,842 281.50

Ht,: equal error ratesfor large and small

audit firms

X' = 3.28*

X^=1.92

H : equal error ratesfor large and smallaudit firms' clients

X'-0.22

X = 0.58

Panel B Overall accuracy of large and small audit firms

Large audit firms(AUD,,^1)

Accurate Inaccurate4411 100

Accuracy (%) 97,78

Small audit firms(AUD,, = O)

Accurate Inaccurate1840 65

96.59

Overall accuracy of model I

Large audit firms'clients (AUD,,= 1)

Cut-off (P*) 0.187Accurate Inaccurate

4,415 96Accuracy (%) 97.87

•Significant at the 0.1 level (x^, (1 d.f.) = 2.7]).••Significant at the 0.05 level (x^on. d d.f.) = 3.84)•••Significant at the 0.01 level (x\^,, (I d.f.) = 6.63)FAILS,, = 1 if company i issued its final report in year totherwise (6,326) observations).

Small audit firms'clients (AUD,, = O)

0.151Accurate Inaccurate

1.850 5597.11

prior to entering bankruptcy

H,; equal accuracy forlarge and small audit

firms

X = 7.63***

H,,: equal accuracy forlarge and small audit

firms' clients

X = 3.36*

(90 observations); = 0

GQ,, = I if company i received a going-concem qualification in year t (107 observations); =0 otherwise (6309observations)-

for model 1 were chosen such that the number ofpredicted bankruptcies was equal to the number ofgoing-concerti qualifications. Since large auditorsgave 71 (13+58) going-concern qualifications, thecut-off probability for large auditors' clients was0.187, as this resulted in 71 predictions of failure

by model 1. Similarly, small auditors gave 36 go-ing-concern qualifications and the cut-off prob-ability for small auditors' clients was 0.151. Thisrule for choosing P* ensured that model 1 reflectedthe degree of auditor conservatism. Choosing alower (higher) P* would increase (reduce) the

226 ACCOUNTING AND BUSINESS RESEARCH

number of predicted failures. Since most com-panies are non-failing, choosing a lower (higher)P* would artificially reduce (increase) model l'saccuracy compared to audit reports.

There are three things to note from Table 5.First, model 1 was more accurate than both largeand small auditors' reports. Model l's type 1 andtype II error rates for large auditors' clients werell.iyV'y and 1.26%, while the error rates of largeauditors were 76.36% and 1.30% respectively.Model l's type I and II error rates for small au-ditors' clients were 17.\4% and 1.50%, while theerror rates of small auditors were 91.43% and1.76% respectively. This is consistent with previousresearch showing that audit reports are poor in-dicators of financial distress compared to bank-ruptcy models (Koh, 1991).

Secondly, the bankruptcy model was moreaccurate for large auditors' clients than for smallauditors' clients. Model l's type I and type II errorrates were 72.73% and 1.26% for large auditors'clients, but 77.14%, and 1.50% for small auditors'clients. This indicates that it is easier to predictbankruptcy for large auditors' clients than smallauditors' clients. This demonstrates the impor-tance of controlling for differences in client char-acteristics when comparing the accuracy of largeand small auditors.

Finally. Panel B shows that large auditors weresignificantly more accurate than small auditors,even after controlling for client characteristics. Thedifference in reporting accuracy between large andsmall auditors was highly significant (at the 1%level), while the difference in model l's accuracybetween large and small auditors" clients was lesssignificant (at the 10% level). While model I classi-fied 8 (O.I8yo) companies more accurately thanlarge auditors, it classified 20 (1.05%) companiesmore accurately than small auditors. Had modell's superior accuracy been the same for both largeand small auditors' clients, one would have ex-pected these numbers to be 19 7 and 8.3 respec-tively (rather than 8 and 20). One can reject thenull hypothesis that model I is equally moreaccurate for large and small auditors' clients at the5% level (x' = 6.40**). Since model l's superioraccuracy is significantly greater for small auditors'clients, the evidence indicates that large auditorswere significantly more accurate.

Model 2 in Table 4 tests whether audit reportssignal useful incremental information about theprobability of bankruptcy by including going-con-cern qualifications (GQ ,) as a predictor. To inves-tigate whether the information content of audit re-ports is related to auditor size, GQ,, is interactedwith an auditor size dummy (AUD,,) which takesa value of one if the auditor belonged to the BigSix.

Prior to estimating model 2, model 1 was re-estimated for large and small auditors' clients to

investigate any differences in coefficient estimates.The relationship between client size (EMP|,) andbankruptcy was found to be greater for large au-ditors' clients than small auditors' clients. There-fore, a variable capturing the interaction betweenclient and auditor size (AUD/EMP,,) is includedin model 2. The insignificant negative coefficienton going-concern qualifications (GQ,,) shows thataudit opinions did not signal useful incrementalinformation about the probability of bankruptcy.The positive coefficient on the interaction variable(AUD*GQ,,) is significant at the 10% level (but notthe 5% level). Although large auditors' going-con-cern qualifications may have signalled usefulinformation about the probability of bankruptcy,the statistical significance was not high. The stan-dard errors for the coefficients on GQ,, andAUDi,*GQ;, were 0.351 and 0.399, respectively.Thus, the difference between the coefficients onGQ, (-0.398) and on AUD,,*GQ,, (0.777) werestatistically significant at the 5% level—one can re-ject the null hypothesis that large and small audi-tors' going-concern qualifications had equalinformation content.

5. Conclusion

Large auditors are significantly more likely to givegoing-concern qualifications to failing companiesand clean opinions to non-failing companies. Re-sults from a bankruptcy model suggest that this ispartly because it is easier to predict bankruptcy forlarge auditors' clients than for small auditors' cli-ents. However, even after controlling for differ-ences between large and small auditors' clients,large auditors give significantly more accurate re-ports compared to small auditors. The incrementalinformation content of audit reports was tested byincluding going-concern qualifications in the bank-ruptcy model. While the statistical significance oflarge auditors' reports was weak, they were signifi-cantly better predictors of bankruptcy comparedto small auditors' reports. A caveat to themethodology of this and previous papers is thatgoing-concern qualifications are not equivalent tobankruptcy predictions by auditors—thus, the cor-relation between audit opinions and bankruptcyoutcomes may not fully capture differences inaccuracy.

Overall, the results are consistent with the rep-utation and deep pockets theories, which predictthat large auditors have more incentive to exerteffort in order to avoid issuing inaccurate reports.An alternative explanation, unrelated to effort orreporting incentives, is that large auditors aremore competent at obtaining or interpreting auditevidence. For example, large audit firms may havemore staff with client-specific knowledge, they maybe more experienced in auditing large quoted com-panies, or they may engage in greater industry spe-

SUMMER 1999 227

cialisation. Future research therefore needs to in-vestigate whether the superior accuracy of largeauditors is due to their deeper pockets, reputationeffects or greater competence.

ReferencesAllman. E. U986|. "Financial ratios, discriminant analysis andthe prediction of corporate bankruptcy", in Johnson. R. (ed.)./.V.VWC.V tinci reaclin^.\ in /mimigciiiil finuncc. New York; DrydenPress.

AUman. E. and McGough, T. (1974). "Evaluation of a com-pany as a going concern'. Journal of Avcoumancy: 51-57.

Anderson. J. (1972), "Separate sample logistic discrimination".Biomctrika 59: 19 35.

Balvers, R.. McDonald. B. and Miller. R. (1988). "Underpricingof new issues and the choice of auditor as a signal of invest-ment banker reputation'. The Accoiiminj; Review. 63: 605-621,

Beatty. R. (1989), "Auditor reputation and the pricing of initialpublic offerings'. The Accoiiniing Review, 64: 693-709.

Datar. S.. Eeltham. G. and Hughes, J. (1991). "The role of au-dits and audit quality in valuing new issues". Journal of Ac-iotiniing and Economics. 14: 3-49.

DeAngelo. L, (1981). "Auditor size and audit quality". Journalof Aecouniing ami Economics. 3: 183-199.

DeEond. M. (1992). "The association between changes in clientfirm agency costs and auditor switching". Audiling: A Journalof Theory anti Praiiice. !l: 16-31.

Dye. R. (1993). "Auditing standards, legal liability and auditorwealth". Journal of Poliliial Economy. 101: 887-914.

Eichenseher, J-, Hagigi. M. and Shields. D. (1989). "Marketreaction lo auditor changes by OTC conipanies". Auiliiing: AJournal of Practice ami Theory. 9: 29-40,

Eirth. M. and Smith. A. (1992), "Selection of auditor finiis bycompanies in ihe new issue market". Applied Economics. 24:247-255,

Francis, J, and Wilson. E, (1988). "Auditor changes: A jointtest of theories relating to agency costs and auditor differen-tiation". The Accounting Review, 63: 663 682,

Grout, P.. Jewitt. I.. Pong. C. and Whittington. G, (1994). "Au-ditor professional judgement: Implications for regulation andthe law". Economic Policy. 19: 307 344,

Hopwood. W,. McKeown, J. and Mutchler, J. (1989). "A testof the incremental explanatory power of opinions qualified forconsistency and uncertainty". The Accounting Review. 64:28^8,

Huber, P. (1967), "The behaviour of maximum likelihood esti-mates under non-standard conditions'. Proceedings of theFifth Berkeley Symposium on Malhemalical Statistics andProbahilily. \: 221-233.

Johnson. W. and Lys, T. (1990). "The market for audit services:Evidence from voluntary auditor changes'. Journal of Ac-counting and Economics. 12: 281 308.

Koh. H, (1991), 'Model predictions and auditor assessments ofgoing concern status'. Accounting and Business Research. 21:331 338.

Lennox, C. (1999). "Identifying failing companies: A re-evalua-tion of the logit. probit and DA approaches'. Journal of Ec-onomics and Business (forthcoming).

Menon, K. and Williams, D. (1991), "Auditor credibility andinitial public ofl'erings". The Accounting Review. 66: 313-332.

Morris, R. (1997). Early warning indicators of corporate failure:A critical review of previous research and further empirical ev-idence. Ashgate Publishing in association with the ICAEW.

Nichols, D, and Smith, D. (1983). 'Auditor credibility and au-ditor changes'. Journal of Aecouniing Research. 21: 534-544.

Schwartz, K. and Soo. B. (1995). "An analysis of form 8-Kdisclosures of auditor changes by firms approaching bank-ruptcy". Auditing: A Journal of Practice and Theory. 14:125-136.

Simunic, D. and Stein. M. (1987). "Production differentiationin auditing: A study of auditor choice in the market for newissues'. Canadian Certified General Accountants' ResearchFoundation.

Titman, S. and Trueman. B. (1986). "Information quality andthe valuation of new issues'. Journal of Aceounting and Eco-nomics. 8: 159-172.