UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN BOOKSTACKS H or V THE HECKMAN BINDERY, INC North Manchester, Indiana JUST FONT SLOT TITLE H CC 1W ? H CC 1W H CC 1W H CC 7W R JULTY *o /RKING 19 PAPER 8 1984 7 NO. 1050-1070 330 B385<"CV"> no. 1050-1070 cop. 2 U. of ILL. LIBRARY URBANA )M] BINDING COPY PERICECAL D CUSTOM D STANDARD D ECONOMY D THESIS ACCOUNT LIBRARY NEW 66572 001 ACCOUNT NAME RUB OR TITLE I D SAMPLE BOOK: D CUSTOM D MUSIC D ECONOM Y I AUTH. 1ST D ACCOUNT mrrarv mcim I nun no t-,t, ,- . ~ *- 1 NO. VOLS. THIS TITLE FOIL COLO UNIV OP ILLINOIS ACCOUNT INTERNAL I.D. ISSN. 912400 I.D #2 NOTES STX4 COLLATING 35 ADDITIONAL INSTRUCTIONS BINDING WHEEL FREQUENCY SYS. I.D. 3 LIC E^Pt^ [BD W/O NO. 1054 BIND T+-2B;. SEP. SHEETS PTS. BD. PAPER I TAPE STUBS I CLOTH EXT. INSERT MAT. GUM SPECIAL PREP. PRODUCT TYPE 11 HEIGHT 11 ACCOUNT LOT NO. ACCOUNT PIECE NO. GROUP CARD I V 118 r« L , COVER SIZE X PIECE NO. HV365 001079] s~rx FACULTY WORKING PAPER NO. 1053 Subjective Prior Probability Distributions and Audit Risk Paul J. Beck Ira Solomon Lawrence A. Tomasslni JUL 6M4 UNIVERSITY OF ILLINOIS URBANA champaign College of Commerce and Business Administration Bureau of Economic and Business Research University of Illinois, Uroana-Champaign BEBR FACULTY WORKING PAPER NO. 1053 College of Commerce and Business Administration University of Illinois at Urbana-Champaign June 1984 Subjective Prior Probability Distributions and Audit Risk Paul J. Beck, Associate Professor Department of Accountancy Ira Solomon, Associate Professor Department of Accountancy Lawrence A. Tomassini University of Texas at Austin Acknowledgment: We wish to acknowledge the suggestions of the referee on earlier drafts and the comments of accounting workshop participants at the University of Illinois, Indiana University, Ohio State University, and the University of Minnesota. Other useful comments were provided by Ed Blocher, Dave Burgstahler, Dan Dhaliwal, Bill Felix, Jim Jiabalvo, D. Paul Newman, Jack Robertson, Mike Shields, and Bill Waller. Digitized by the Internet Archive in 2011 with funding from University of Illinois Urbana-Champaign http://www.archive.org/details/subjectivepriorp1053beck Subjective Prior Probability Distributions and Audit Risk Abstract This paper presents an analysis of the audit risk consequences of PPD ex- tremeness deficiencies and miscalibration. While there is empirical evidence that auditors, like many other decision makers, assess miscalibrated PPDs , the attendant inferential risk consequences of such deficiencies have not been addressed in the extant literature. The comparative statics analysis performed in this study indicates that the risk effects of miscalibration and extremeness deficiencies on the auditor's (substantive testing) evaluation decision are complex and cannot be predicted from an examination of the planning (sampling size) decision. 1.0 Introduction Bayesian models of the auditor's decision process (e.g., Bailey and Jensen [1977], Kinney [1975], and Scott [1973, 1975]) incorporate subjective beliefs formalized as prior probability distributions (PPDs). Recent studies by Solomon [1982] and Tomassini et al. [1982] investigated two conformance properties of auditors' account balance PPDs: extremeness and calibration. PPD extremeness is measured over a sequence of elicitation trials by computing the average subjective probability assigned to inter- vals containing the auditees' actual account values and, as such, can be viewed as a measure of predictive ability (see Seaver et al. [1978]). Calibration, in turn, is a measure of the ability to express an appropriate degree of confidence in such subjective estimates. Both the Solomon [1982] and Tomassini et al. [1982] studies reported that auditors' PPDs were raiscalibrated and that there was limited evidence of under confidence (see fn. 1). Other probability elicitation studies in psychology also have found miscalibration, but with the exception of experienced weather forecasters, these other studies have reported almost universal evidence of overconf idence (see Lichtenstein et al. [1982]). Although several empirical studies have investigated PPD cali- bration and extremeness, their specific decision-making consequences have not been analyzed formally in either the psychology or accounting literatures. Libby [1981] discussed the effect of mis- calibration on the audit sample size decision in compliance -9- testing and concluded that overconf idence impairs audit effec- tiveness, and underconf idence impairs audit efficiency. Tomassini et al. [1982] investigated the implications of underconf idence for the sampling decision in a substantive testing context, but did not consider the subsequent effect on the auditor's evaluation of account book values. Simulation studies (e.g., Blocher [1981] and Cushing [1980]) have been used to investigate the sensitivity of audit planning (sample size) decisions to misspecif ications of PPD parameters in the compliance and substantive testing contexts, respectively. While these studies indicated that planning decisions were sen- sitive to misspecif ications of both PPD means and/or variances, the findings permit only very limited conclusions about the adverse consequences of extremeness deficiencies and miscalibration. In our study, we identified specific audit risk consequences of PPD extremeness deficiencies and miscalibration for Bayesian hypothesis tests of account balances (which could be applicable to quality control and cost-variance decisions). Following Kinney [1975], we assumed that the auditor assesses account balance PPDs which are combined with mean-per-unit (MPU) sample estimates to form posterior probability distributions. The effects of extreme- ness deficiencies were then analyzed for a sequence of such posterior distributions by varying the PPD mean displacement rela- tive to the corresponding outcome sequence. Similarly, the effects -3- of mis calibration were determined by varying the correspondence between the assessor's perceptions of PPD mean displacement and the actual displacement. We subsequently identified the attendant audit risk consequences by interpreting our comparative statics results in the context of the Kinney [1975] audit decision model. Our study extends previous research first by formally analyzing the audit risk effects of miscalibration and extremeness deficien- cies on both planning and evaluation of substantive tests. Second, by allowing subjects' PPD variances to adjust in response to the perceived PPD mean displacement, we were able to avoid experimen- tally confounding the effects of calibration and extremeness defi- ciencies. Our analysis showed that the risk consequences of miscalibra- tion (overconf idence and underconf idence) were complex and depended upon the actual population mean and the nature and severity of any extremeness deficiencies. Furthermore, unlike Libby [1981], we found that overconf idence actually can predispose auditors to com- mit either effectiveness or efficiency errors. The remaining sec- tions of the paper are organized as follows: Section 2 presents our modeling assumptions which serve as a foundation for the com- parative statics analysis in Section 3. This is followed in Section 4 by a discussion of the audit risk implications of our comparative statics findings. Concluding remarks, with implica- tions for the design of audit training programs are provided in Section 5. -4- 2.0 Audit Decision Context Assumptions Audit verification of account balances is based upon both objective and subjective information. Objective information for asset accounts (such as inventory and receivables) is generally obtained by sampling from populations of subsidiary book values. Previous Bayesian studies in auditing have modelled such popula- tions using a probability density with an uncertain mean, u. In these models, a PPD is assessed for each D and then combined with sample information to form a posterior probability distribution. In practice, such PPDs would incorporate all of the available forms of evidence typically collected prior to substantive testing, including compliance tests. 2.1 Hvnothesis Testing Assumptions We model the audit evaluation decision as a Bayesian hypothe- sis test (i.e., we assume a two-point auditor loss function). This simplifying assumption implies that the losses borne by the auditor do not depend upon the magnitude of misstatement (see Moriarity [1975]) and ensures that the optimal decision rule is consistent both with the expected linear utility framework, adopted herein, and with traditional definitions of audit risk in the professional auditing literature. Specifically, for asset accounts, we assume that the null and alternative hypotheses are: -5- H~: U > X - M (i.e., account book value is not materially overstated) H : ji j^_ X - M (i.e., account book value is materially overstated) , where X is the auditee's reported account book value (expressed as an average), and M is the materiality limit defined with respect to the mean account book value. Note that the null hypothesis does not consider the case of material understatements, but other- wise is consistent with the hypothesis formulation initially pre- sented in Elliott and Rodgers [1972] and subsequently adopted in Section 4 of the recently issued Audit and Accounting Guide: Audit Sampling (AICPA [1983]). A discussion of the modified risk implications for cases in which the auditor employs a two-sided hypothesis test is presented below (see fns. 8 and 10). Figure 1 presents the four possible outcomes associated with the audit evaluation (acceptance-rejection) decision and the two hypothesized states of nature, H~ and H . When y _<_ X - M, the null hypothesis (H„) should be rejected. Thus, if H» were accepted, audit effectiveness would be compromised and the loss is represented by L • Alternatively, when y > X - M, audit effi- ciency would be compromised if H were to be rejected, and L represents the associated loss. Insert Figure 1 Here Assuming minimization of expected losses, the optimal audit decision rule (see Berger [1980]) is to reject H_ if and only if: -6- Pr(H ) L (1) 1 - Pr(H Q ) Lj 2 where Pr(H~) represents the posterior probability that S > X - M. The above formulation simplifies our analysis because, for a given loss ratio, L T /L , the optimal decision depends solely upon the posterior likelihood ratio, Pr(H Q )/(l-P (H Q )). The rela- tionship between L and L affects the audit risk implications of miscalibration and extremeness deficiencies and will be discussed further in Section 4. 2.2 Distributional AssumDtions Consistent with previous research (e.g., Kinney [1975] and Scott [1973, 1975]), we assume that the distribution of subsidiary book values and auditor's PPD are normally distributed and that the ordinary MPU estimator is employed. We also assume that the variance of each distribution is known. Given these assumptions, the posterior distribution for each audited account value is nor- mally distributed and parameterized (see Winkler [1972, p. 169]) by (m",v"), where: jm m = : (2) Ar = -V+-2- (3) V V V and: -7- m = sample estimate of the population mean, p; m 1 = PPD mean; m" = posterior estimate of the mean; n = quantity of sample evidence; v/n = variance of m; v 1 = PPD variance for the assessed m'; v" = posterior variance; The above assumptions are not as restrictive as they might appear since, as discussed below, our analysis of the simple MPU estimator can be generalized directly to the stratified MPU esti- mator. Also, the assumption of population normality is defensible when stratified sampling is employed since' the strata, which are more internally homogeneous, are less skewed than the population taken as a whole. More importantly, sample MPU estimates of the mean should be normally distributed, given the central limit theorem. Previous empirical results (e.g., Cushing and Romney [1981], and Solomon, et al. [1982]) suggest also that auditors' PPDs assessed for highly skewed populations, such as those described in Neter and Loebbecke [1975], are typically symmetric and can be approxi- mated by normal PPDs. Finally, relaxing the assumption of known sample variance by allowing an assessment of a second PPD for the variance parameter (actually a joint PPD for the mean and variance parameters) as in Felix and Griralund [1977] would not change our analysis in any significant manner. Calibration and extremeness, however, would have to be redefined with respect to each of the unknown parameters. Our assumptions establish a simplified environment in which there is only one uncertain parameter and there are no sample esti- mation problems due to skewness and/or an unknown sample variance. If miscalibration can be shown to be a significant problem in the simplified environment modeled herein, we can reasonably expect it to be as important or more so when the probability elicitation task is more complex and there are statistical estimation problems. 2 .3 Extremeness and Calibration Assumptions Since calibration and extremeness are defined with respect to a sequence of PPDs and corresponding outcomes (see Lichtenstein et al. [1982]), we assumed that the probability assessment-hypothesis testing procedures were to be repeated for a sequence of trials 3 with outcomes, (u^.; t=l, T}. This assumption is consistent with traditional definitions of inferential risks (e.g., alpha and beta risks) which are defined with respect to a sequence of estimates. PPD extremeness, as noted above, is measured ex post by imputing the probability assigned to a suitably chosen interval containing each y and then averaging these probabilities over the sequence (see Solomon [1982]). An achievement of high levels of extremeness would then require that the auditor-assessor specify distributions which are centered appropriately, relative to the u sequence, and are tight (see Peters [1978]). Otherwise, the -9- highest probabilities would be assigned to population values which do not occur most frequently. We modeled extremeness deficiencies in our study by PPD mean displacement (i.e., m = u + A, where A denotes the PPD mean displacement). By varying A in our comparative statics analysis, we manipulated the severity of the assessor-auditor's extremeness deficiency. The auditor's calibration, (see above and also fn. 1); reflects the ability to express an appropriate degree of confidence in sub- jective estimates of population means so it is defined with respect to a given level of extremeness (i.e., magnitude of A). Both overconf idence and underconf idence calibration defi- ciencies were modeled. An overconfident assessor's PPDs would be too tight and, therefore, over a sequence of elicitation trials, too many realizations would be captured in the outer fractile ranges, and too few captured in the inner fractile ranges (see Lichtenstein et al. [1982]). In contrast, underconf ident assessors would specify PPDs which are too diffuse, in that the inner fractile ranges capture too many outcomes, while the outer fractile ranges capture too few outcomes. The variance of a nor- mal PPD represents the assessor's uncertainty about the population mean and, thus, is related directly to the auditor's calibration (see Moskowitz and Bullers [1979]). Calibration was modeled in our comparative statics analysis by assuming that the auditor would implicitly estimate the average -10- mean displacement that would exist if the elicitation process were repeated for the entire u sequence, and then would adjust the assessed PPD variance on each trial to reflect the estimated (aver- age) displacement. Since each trial was assumed to be exchange- able, the assessor's perceptions about displacement would be the same for each trial in the sequence (see fn. 3). Furthermore, provided that an assessor were not completely miscalibrated, there should be some correspondence between the assessor's perceptions about the PPD mean displacement (A) and the actual A. Accordingly, we formalized this correspondence between A and A by expressing the PPD variance on each trial as a function of A (i.e., v' = v'(A), where A = A(A)). Note that, within the context of our model, perfect calibration would exist if A(A) = A for each pos- sible level of extremeness since the auditor would be able to spread the PPD variance appropriately (see fn. 2). Alternatively, A(A) > A would imply underconf idence, while A(A) < A would indi- cate overconf idence. The correspondence between A and A will be varied in our comparative statics analysis to identify the con- sequences of the two types of miscalibration. 3.0 Comparative Statics Analysis In this section we first investigate the effects of extreme- ness deficiencies and miscalibration on a sequence of posterior -11- distributions , after which (in Section 4) the audit risk implica- tions of these effects will be discussed. 3.1 Extremeness The effect of PPD mean displacement on the posterior mean for trial t can be determined by substituting u + A for m' into (2). Assuming that m = u , the following simplified expression for the posterior mean is obtained: m t = \ + (1 + n v'/v) ' (4) where n = n(m ,v'), m = u + A, and v' = v'(A(A)). The difference between m and \i represents the magnitude of t t F 6 posterior mean displacement on trial t. Since (1+n v'/v) > 0, equation (4) predictably indicates that the posterior mean on each trial will be displaced in the same direction as the PPD mean due to the presence of A. However, the magnitude of the posterior mean displacement can vary from trial to trial. Accordingly, we 4 take an expectation (average) with respect to {y ). E[nT] = E[u t J + E[A/(l+n t v'/v)]. (5) The consequences of altering PPD extremeness deficiencies for the sequence are determined by differentiating (5) totally with respect to A using the quotient and chain rules, -12- dn ._, „, (1 + n v'/v) [A[(n /v>^-+ ( v '/v>-f] aE [m J _ r t L t dA dA J 1 "dA~" = E[ T~. 7772" T~ 7772 ; (1 + n t v'/v) (1 + n t v'/v)' Simplifying (6) algebraically, dE[m"] f 1 V dv' v'A dn t i m dA " Wl + n v'/v) n . ,, ,2 ,. n . ,. ,2 dA '* U) t v(l + n v'/v) dA v(l + n v'/v) Equation (7) shows that the total effect of PPD mean displace- ment on the expected value of the posterior mean sequence can be decomposed into three terms. The first term on the right side of (7) represents the expected (upward) posterior mean displacement which would occur if ^v / 1.0 (i.e., auditor is underconf ident) , the assessed PPD variances would be too large given the actual level of PPD mean displacement. While (8) indi- cates that severe underconf idence can effectively drive the expected posterior mean displacement to zero, unnecessary sampling would still be performed. Further insight regarding the sample size effect is obtained by decomposing the dn /dA term in (7) which will show the depen- dence of n on A through the A and v' terms: dn 3n , , 3n , , * t_ = t dm' t dv' dA , g \ dA 8m' dA 3v* d £ dA * 3n The -r— T-J7— term in (8) can be either positive or negative, dm a/i depending on the auditor's loss function. Hence, general conclu- -14- sions about the sample size effect and the magnitude of the posterior mean displacement are not possible. Nevertheless, the decomposition in (8) indicates that there is a "calibration com- ponent" embedded in the ^ n t/dA term in (7). Since the sign of the terms multiplying c^dA i- n (7) is negative, the expected magni- tude of the posterior mean displacement will be reduced by the , • v • L dv' d'A x . . . dv' dA . _ calibration component when — z > 0, but increased when — s < 0. dA dA dA dA 3.2 Calibration This subsection extends the previous analysis by investigating the effects of PPD miscalibration on the posterior variance sequence. This is done by holding A constant, and varying the inaccuracy of the assessor's perception of the actual displacement. More specifically, we model the effects of miscalibration by assuming that A(A) ■ (1+)A, where represents the inaccuracy of the assessor's perception of A. In this framework, $ = would imply that the assessor's perception of the actual level of PPD mean displacement across the sequence of trials is accurate, so that the PPD variance can be adjusted appropriately. Alterna- tively, $ < and $ > would indicate overconf idence and under- confidence, respectively. As a first step here, we solve (3) for v". Upon simplifica- tion, the posterior variance for the t trial can be expressed as : v t = v/(n t + v/v'), (9) -15- where: n = n(m t ,v f ), m = u + A, v' = v'(A), and A = (l+)A. Taking an expectation with respect to the u sequence, we obtain the average variance of the posterior sequence: E[v t ] = vE[l/(n t + v/v')]. (10) Differentiating (10) with respect to , dE[v"] " dn #T7 f r , fl -2 r t dv' dA dv' dA, f 2i ,,,. = -vE [n + v/v f ] i-T-r—^--rr- v— s— -rr/v f } . (11) d t d v dA d $ dA d< J> Given that -rr = A and — s— > 0, (11) indicates that d * dA dE[v"]/ d 4> >(<) o when dn t 2 -7-r<(>) v/v' Z . (12) dv Since the direction of the inequality in (12) depends upon the specific parameter values, general conclusions about the effects either of underconf idence or overconf idence are not possible. An inspection of (12) indicates, however, that misspecif ication of the PPD variance will have the greatest impact on the posterior variance when the PPD variance is small in relation to the sample variance — i.e., the auditor has substantial subjective knowledge — or sampling costs are high so that t^dv' ^ s small. Hence, in situations in which the formal incorporation of subjective audit knowledge potentially would be most beneficial, underconf idence will cause the posterior variance to be biased upward, while over- confidence will cause the posterior variance to be biased downward. -16- 3. 3 Stratification So far our analysis of extremeness deficiencies and miscali- bration has assumed unstratified MPU sampling. In practice, however, population skewness often necessitates stratification. Application of Bayesian stratified methods requires the assessment of a vector of means and k(k+l)/2 elements of the variance- covariance matrix (see Ericson [1965]). But, in the special case in which the stratum means are independent (i.e., the variance- covariance matrix is diagonal), extremeness and calibration can be computed with respect to a PPD sequence for each stratum. This case represents the direct Bayesian analogue of the classical stratified sampling model in which the sample mean estimates are assumed to be distributed independently. In such Bayesian stratified sampling models, posterior esti- mates of both the mean and variance are computed for each stratum and then aggregated linearly to obtain a stratified posterior estimate for the population as a whole. As a result, if our assumed hypothesis test (decision rule) were employed to evaluate the stratified posterior estimates, we would obtain comparative statics results similar to those reported above. For example, if the stratum PPD mean sequences are displaced in the same direc- tion, the posterior means and stratified mean estimate for the population also will be displaced and in the same direction. If, in addition, the auditor is overconfident (underconf ident ) in -17- assessing the PPDs for one or more strata, the expected magnitude of posterior mean displacement will be increased (decreased) and the expected posterior variance also will be increased (decreased) Unfortunately we are unable to provide general results when the strata means are correlated or if calibration deficiencies differ across strata. 4.0 Audit Risk Implications We now identify the risk implications of extremeness deficien- cies and miscalibration for the audit evaluation decision. An ex post perspective is implicitly adopted in which we focus on a representative trial (t) and separately analyze Case I in which \i <_ X - M and Case II in which u > X - M. Separate risk analy- ses are required, because audit efficiency and effectiveness errors depend jointly upon the audit decision and whether the account balance actually contains a material (overstatement) error. 4.1 Case I: Book Value Materially Overstated Since the account book value is assumed to be materially overstated (i.e., u^ _<_ X - M) , we focus on the auditor's risk of committing an effectiveness error in the four subcases below. ^.1.1 No Mean Displacement We initially simplify by considering the risk effects of mis- 6 - calibration in the absence of PPD mean displacement. Figure 2 -18- presents two normal posterior probability distributions which, for expositional purposes, are assumed to be based upon PPDs assessed by auditors A and B, respectively. Since the PPDs and posterior distributions correspond to the same u outcome, we henceforth replace the t subscript on all PPD parameters with the letters A and B, so that each distribution can be associated with a par- ticular auditor-assessor. While both auditors are assumed to assess PPDs which are properly centered vis-a-vis the actual account balance (i.e., m = ol = u ) , Auditor A is more under- confident than Auditor B (i.e., „ > 0). Therefore, con- A a av r "i sistent with our assumption that — > 0, distribution A is more diffuse than distribution B and, thus, can be interpreted as reflecting greater underconfidence. Insert Figure 2 Here The area under each distribution to the right of X-M repre- sents the probability that the actual account balance mean (u ) exceeds X-M and is denoted by Pr(H»), while that to the left of X-M represent Pr(H ) = 1 - Pr(H ). Given the symmetry property of normal distributions and the further assumption that m ■ V , the posterior means, m , nL _<_ X - M, so the computed Pr(H~) _< .5 for both distributions A and B. The optimal decision, for either auditor, is to reject H~ when the computed likelihood ratio, Pr(H )/(l-Pr(H )) < L /L . Since -19- the costs of effectiveness and efficiency errors depend upon fac- tors specific to the decision-making context, general conclusions cannot be drawn about the loss ratio (L /L ) . Typically, however, the assumption is that losses from audit effectiveness errors are larger than those from audit efficiency errors, implying that L /L > 1.0 (see Elliott and Rodgers [1972] and AICPA [1981, 1983]). This means that, in the present subcase in which Pr(H )/(l-Pr(H )) _<_ 1.0, H Q would be (correctly) rejected by either Auditor A or B. In fact, varying the magnitude of the pos- terior variance would not alter the rejection decision, because the computed Pr(H_) < .5 when m , til = u <_ X - M. Therefore, in the absence of mean displacement, underconf idence does not affect audit risk exposure (see Table 1). But the audit planning decision would be adversely affected by underconf idence. That is, ceteris paribus , an underconf ident auditor would choose a larger sample size than a properly cali- brated auditor when ^ n /dv' > 0* This would result in an oppor- tunity cost due to insufficient reliance upon subjective knowledge. The next two subcases extend the preceding analysis by incorporating PPD mean displacement. Insert Table 1 Here 4.1.2 Moderate Upward Mean Displacement We now assume that both Auditors A and B assess PPDs whose means are equally displaced upward (i.e., m = hl > M._ )> but -20- Auditor A is more underconf ident than B. From our previous analy- sis, it is apparent that both posterior means will be displaced upward (i.e., m , itl > \i ), so that the computed Pr(H~) and like- lihood ratio will be increased. Here, we consider the risk effects when u < ra = nu < X - M, but in the next subsection we consider more serious upward mean displacements wherein \i L , both auditors will correctly reject H„, so changes in the PPD variance due to miscalibration would not alter the rejection decision. Therefore, upward mean displacement and/or miscacalibration do not affect risks provided that = nu < X - M (see Table 1). » i m 4.1.3 Serious Upward Mean Displacement In practice, the magnitude of the upward PPD mean displacement may be so great that ra = hl > X - M. If so, both likelihood ratios could exceed the loss ratio, which would result in the com- mission of effectiveness errors (see Table 1). The predisposi- tion to commit an effectiveness error, however, is influenced by Q the severity and nature of the auditor's miscalibration. Figure 3 presents two posterior probability distributions for Auditors A and B. As in 4.1.1, m = nu, but the posterior mean displace- ment is less for distribution A, because the (displaced) PPD mean -21- is weighted less heavily vis-a-vis the sample mean due to more serious underconf idence. Insert Figure 3 Here Reducing the upward posterior mean displacement, ceteris pari- bus , increases the computed Pr(H ) and reduces Pr(H ). Since the area under the lower tail of the distribution also is increased by underconf idence, the computed Pr(H ) is further increased, thereby 3. reducing the computed likelihood ratio. Thus, underconf idence reduces the auditor's predisposition to commit an effectiveness error due to upward mean displacement (see Table 1), but at the cost of suboptimal (larger) sample sizes. The risk effects of overconf idence can also be determined from Figure 3 by reinterpreting distribution B as exhibiting greater overconf idence than A. Predictably, the effects of overconf idence on the posterior distribution are contrary to those described above for underconf idence, because the weight accorded the PPD mean is increased and the audit sample size is simultaneously reduced. As a result, the posterior mean displacement is increased and the posterior variance is decreased. Both effects reduce the area to the left of the point X-M thereby increasing the computed Pr(H ). Hence, when ra = iil > X - M, overconf idence increases the auditor's predisposition to commit an effectiveness 9 error resulting from upward PPD mean displacement (see Table 1). -22- 4.1.4 Downward Mean Displacement We now assume that the PPD means are displaced downward (i.e., » i . < u ) . Since by assumption, ra = u <_ X - M, our previous analysis indicates that the posterior means, m , itl < X - M. Therefore, the computed Pr(H ) < .5 for both distributions which ensures that both auditors will correctly reject H„. Furthermore, as discussed above, changes in the PPD variance due to miscalibra- tion would not increase the computed Pr(H n ) above .5. Hence, no risk consequences are associated with downward PPD mean displace- ment and/or miscalibration (see Table 1). 4.2 Case II; Account Book Value NOT Materially Overstated We now assume that the account book value is not materially overstated (i.e., u > X - M) . This allows us to focus on the risk of committing an efficiency error due to PPD extremeness deficiencies and miscalibration. 4.2.1 No Mean Displacement We first consider a subcase in which there is no PPD or sample mean displacement (i.e., m = nu = m,= M ), but the auditor is underconf ident . Figure 4 presents two posterior probability distributions which are based upon the PPDs assessed by Auditors A and B and are analogous to those of Figure 2, except we now assume that U > X - M. As in 4.1.1, the computed Pr(H ) is higher for the more diffuse distribution (A) which reduces the likelihood -23- ratio. However, in contrast with the earlier effects, the more underconf ident auditor has a greater predisposition to commit an efficiency error by incorrectly rejecting H_. Insert Figure 4 Here 4.2.2 Upward Mean Displacement Upward PPD mean displacement (i.e., m. = m_ > u ), ceteris paribus results in upward posterior mean displacement which, as in cases 4.1.2 and 4.1.3, increases the computed Pr(H n ). Here, however, increasing the likelihood ratio reduces the risk of effi- ciency errors (see Table 1). Figure 5 presents two posterior distributions depicting the joint effects of upward PPD mean displacement and miscalibration. Posterior distribution A is cen- tered to the right of distribution B and is less diffuse to reflect greater overconf idence. As in subsection 4.1.3, both the displacement and dispersion effects upon the posterior distribu- tion increase the computed Pr(H ). However, the effect of increasing the computed Pr(H_) now is to reduce the probability of efficiency errors. Insert Figure 5 Here Unfortunately, when there is underconf idence, the risk effects are more difficult to discern. Figure 5 can be reinterpreted to facilitate an analysis of underconf idence in conjunction with -24- upward PPD mean displacement. The more diffuse distribution B now represents greater underconf idence than distribution A. Note that greater PPD underconf idence reduces the upward posterior mean displacement while simultaneously increasing the posterior vari- ance. This effect tends to mitigate the effect of upward PPD mean displacement. As a result, the ultimate risk consequences are indeterminate unless specific assumptions are made both about the magnitude of the PPD mean displacement and the severity of the auditor's underconf idence (see Table 1). 4.2.3 Downward Mean Displacement We now analyze the case of downward PPD mean displacement in which m = nu < y . Since the posterior mean is displaced in the same direction as the PPD mean, the computed Pr(H_) is decreased which, ceteris paribus , predisposes the auditor to commit effi- ciency errors (see Table 1). Such predisposition, however, is also influenced by the auditor's miscalibration. Figure 6 pre- sents two posterior distributions which depict the joint effects of downward PPD mean displacement and miscalibration. Auditor A's posterior distribution is displaced further than B's and also is tighter than distribution B due to greater overconf idence. Note that the areas to the left of X - M are approximately the same for both distributions, given the opposing mean and variance effects on the computed Pr(H ). Hence, the efficiency risk consequences cannot be determined. Finally, note that, while we have assumed -25- implicitly that X - M < m. * m_ < U , the same indeterminate risk effects also would arise when m=m_ < 0, distribution A would be more diffuse than distribution B, so that the effects of overconf idence would be analogous to those described above for underconf idence. Similarly, the effects of underconf idence when dE[v"]/d<}> < 0, would be similar to those for overconf idence when dE[v"]/d<£ > 0. 10. As discussed above (see fn. 8), if a two sided hypothesis test were employed, upward posterior mean displacement could decrease P r (H()) and create a predisposition toward effi- ciency errors. -31- Decisions: States : H > X - M y < X - M Accept H o ° L n Reject H o L i y = actual population mean X = average value of subsidiary account reported by the client L = loss from audit inefficiency L TT = loss from audit ineffectiveness Auditor's Loss Matrix Figure 1 -32- Frequency m A , m B Monetary Value Fig. 2-;Underconfidence-No Mean Displacement -33- Frequency " 1 T ' il il fi t x-m m A m B Monetary Value Fig. 3 - Miscalibration , Upward Mean Displacement -34- Frequency TW'B Monetary Value Fig.4- Underconfidence - No Mean Displacement -JD- Frequency x-m m B fj. t m A Monetary Value Fig.5 - Miscalibration, Upward Mean Displacemef -36- Frequency ii ii x-m m A m B fi t Monetary Value Fig. 6 - Miscalibration - Downward Mean Displacerr ►» o >> o o >, < ■ 3= B = — • MX 3 3 « ■ 3 V — « *-> H ii «4 tl — * *J u < o < u a c u «-t i a > «-i -3 u v *J O a -* »j ii — o a u os — . ... uct ». It v -3 y v C «*>» •- >< v v <** >\ U Bl <-> CU u 0. — U u b (I U M U «j j< « -* u -o -tf u ■O O L. ■ b « TI ■ b C — » «-> u -< e u e b o b ■ <-» x M ■ «-* >» o M >» o ■ s •* a a 3 f-4 4J O V — *J < u a. B ■< o V « ■ V w w "^ — > ^^ *J « ■o ~* u « u as e w u as a • b u • , e «M ^< » 44 U u c • 9 V fcl •^ ■»« « 55 C b / X fc> a » 41 X ■ X a 31 c 1 ■ B. V « a 1 a. X V 0. ft. "■« M b -■"' o I TJ V v l 4 1 a. — 3 o. V s u a. s m CL o 2 a c Z H X 1 CI w I 3 (J — « X ■ — > & A M — *■ o o a o ■a e ■ ■ u n . t c x ■ 3 CO! U 0.