Inferring About the Average Value of Audit Errors from Sequential Ratio Tests Authors: Grzegorz Sitek, Mariusz Pleszczyński, Boris Ryabko Published: 2024 ABSTRACT AbstractThe book amounts are modeled as values of a random variable, represented by a mixture of distributions of both the correct and error-contaminated amounts. The mixing coefficient represents the proportion of items with non-zero error amounts. This study addresses the problem of determining the sample size needed for testing statistical hypotheses regarding mean accounting errors. The average sample size is estimated using the Sequential Probability Ratio Test (SPRT), applying the Monte Carlo method. Estimating average audit errors is a common challenge in economic research.Keywords: sequential likelihood ratio test, mixture of probability distribution, Monte Carlo method 1. Introduction Let Xi represent the book amount of the ith item in an account, with the total population book amount denoted as XU=∑i=1Nxi. At regular intervals, an auditor samples n line items from the account and to compare against their correct values. Therefore, let Yi denote the audited amount for the ith line item, and let τi=Xi−Yi denote the error amount. Note that the total book amount is known to the auditor. The fundamental issue is constructing confidence limits for means or totals in finite populations where the underlying distribution is highly skewed and contains a substantial proportion of zero values. This situation is often encountered in statistical applications such as statistical auditing, reliability assessments, and insurance. The most distinctive feature of accounting data is the large proportion of line items without error, while an audit sample may not yield any nonzero error amounts. For data with mostly zero observations, classical interval estimation based on asymptotic normality is not reliable. In auditing practice, auditors are often more interested in obtaining lower or upper confidence limits than in obtaining two-sided confidence intervals. Specifically, independent public accountants are focused on estimating the lower confidence bound for the total audited amount to avoid overestimation and potential legal liability. Stringer and Kaplan (see [1,2]) have demonstrated that accounting populations are highly positively skewed, with considerable diversity in the characteristics of error amounts across subsystems. There are several distributions that also exhibit the same form of the distribution observed in accounting populations. These include the Gamma, Log-normal, Weibull, and Beta distributions. The error rates are usually very low, which render many existing statistical procedures inappropriate for estimating and hypothesis testing of error rates and error amounts. There are two main types of audit tests where statistical sampling is advantageous. The first is a compliance test, used to determine the procedural error rate in a population of transactions. The second is a substantive test of details, aimed at evaluating the aggregate monetary error in a stated balance. Inferences about the total error amount are usually made using confidence intervals, which are closely related to hypothesis testing. The decision-making process in auditing is treated as a problem of testing statistical hypotheses about admissibility of the total or the mean accounting errors. This approach allows auditors to control both the significance level (risk of incorrect rejection) and the type II error probability (risk of incorrect acceptance). Substantive tests of details verify the accuracy of recorded monetary values in financial statements, providing direct evidence regarding the accuracy of the total recorded values. Auditors may apply substantive tests extensively or use compliance tests to evaluate the efficiency and effectiveness of internal controls in mitigating material errors. In compliance tests, the variable of interest is the error rate (proportion of transactions for which the internal control operates wrongly). Samples of transactions are used to infer this error rate. 2. A Mixture of Probability Distributions as a Model for Generating Accounting Values Wywiał in [3] proposed the following model. Let U be the population of accounting documents of size N with given accounting totals. Some of the documents contain errors. In the population U, there are given accounting totals (values) xi for each element i∈U. Let xT=[x1,⋯,xN]∈R+N be the observation of a random vector XT=[X1,⋯,XN]. We denote the true book values (without errors) by yi, i∈U and let yT=[y1,⋯,yN]∈R+N be the random vector observation YT=[Y1,⋯,YN]. Vector of accounting values contaminated by errors wT=[w1,⋯,wN] will be an observation of the random vector WT=[W1,⋯,WN]. Finally, let ZT=Z1,⋯,ZN, where Zi=0(Zi=1) if Xi=Yi(Xi≠Yi). In practice, all values of X are known before auditing process. Observations x of X are treated as a specific auxiliary data. Auditing process leads to observation of values Zi, Yi and Wi, i∈U. Let X¯=1N∑i∈UXi, Y¯=1N∑i∈UYi, W¯=1N∑i∈UWi. Their values will be denoted x¯, y¯, w¯. Let an auditor arbitrarily select a sample s of size n from U. Hence, Xs is the subvector of X, n≤N. The random vector Xs is observed in s where the objects are controlled. After the auditing process, the sample s is split into two disjoint sub-samples s0 and s1 where s=s0∪s1. The set s1 is of size n1=k and the set s0 is of size n0=n−k. In the sub-sample s0, there are observed accounting amounts without errors. Before the auditing process, we have observations of the following data: where After the auditing process, we have observations of the following data: Values T, TU, X, Xs, XU−s, Ys0 and Ws1 are denoted, respectively, by t, tU, x, xs, xU−s, ys0 and ws1. In the following work, we assume that ys0=ys and ws1=ws. Let τ=E(X¯−Y¯) be the expected mean accounting error. Audit purpose is inference on τ or on the expected total accounting error Nτ=E(∑i∈UXi−∑i∈UYi). Let F0(y|θ0) be the probability distribution function of the random variable Y, whose values are true accounting and θ0∈Θ0 where Θ0 is the parameter space. The distribution function of W is denoted by F1(w|θ1), where θ1∈Θ1. Moreover, let Θ=Θ0∪Θ1. We assume that an accounting errors appears with probability p. We can write Z=1 when an accounting error occurs P(Z=1)=p and Z=0 when it does not occur P(Z=0)=1−p. According to the well-known total probability theorem we have: Fx=Fx|Z=0PZ=0+Fx|Z=1P(Z=1) and finally where θ=θ0∪θ1 and θ∈Θ=Θ0∪Θ1 is the parameter space. Hence, the probability distribution of the observed accounting amounts is a mixture of the distribution function F0(x|θ0) of the true amounts and the distribution function F1(x|θ1) of the amounts contaminated by errors. When the random variables Y and W are continuous, by differentiating both sides of Equation (1) we have Therefore, the probability density of the observed accounting amounts is a mixture of density f0(x|θ0) of the true amounts and density f1(x|θ1) of the amounts contaminated by errors. Let R and Y be independent and R is the accounting error. Hence W=Y+R, X=Y+ZR, X=(1−Z)Y+ZW. The basic moments of the random variable X are: 2.1. A Mixture of Gamma Probability Distributions as a Model for Generating Accounting Values The well-known gamma probability distribution we denote by G(α,β), where parameters α>0 and β>0 are called scale and shape parameters. The shape of gamma density distribution does not depend on the scale parameter because its skewness and kurtosis coefficients are equal to 2β and 6β, respectively. Wywiał in [3] considered the model based on a mixture of gamma distributions. Let Y∼G(a,c) and R∼G(b,c) be independent random variables. The advantage of this model is that the density function for the sum of gamma distributions can be determined. Based on the above assumption, the random variable W=Y+R∼G(a+b,c). Using the previous considerations, we obtain where From Formulas (3) and (4) we obtain For more on the use of gamma decomposition to model accounting values, see the articles [4,5]. The book amounts are treated as values of a random variable which distribution is a mixture of the distributions of correct amount and the distribution of the true amount contaminated by error. Distributions of correct amount and true amount contaminated by error are right-skewed because small book amounts are more frequent than large book amounts. It is convenient to assume that the book values are additive function of true accounting amounts and accounting errors. Hence, we can expect that the above proposed quite simple model describes accounting data well. 2.2. A Mixture of Poisson Probability Distributions as a Model for Generating Accounting Values We assume that the variable Y∼Pois(a) (true values) and R∼Pois(b) (accounting error), where Pois(λ) is a Poisson distribution with probability function If variables Y and R are independent, then the random variable W=Y+R∼Pois(a+b). From Formulas (3) and (4) we obtain 3. Sequential Tests Based on the Ratio of the Likelihood Function The set of issues involved in proceeding sequentially in verifying statistical hypotheses is called sequential analysis (see [6]). Sequential analysis was created by Wald (see [7]). The sequential approach to statistical inference was the subject of systematic research during the Second World War. This research was concerned with the quality of munitions and was conducted by the Statistical Research Group at Columbia University. Data in sequential analysis are used both to decide when to end an observation (data collection) and to draw actual conclusions (regarding the parameter being estimated or hypotheses being tested). The sequential method is applicable in auditing when assessing the internal control system in compliance testing when verifying hypotheses regarding the proportion of accounting errors. In their doctoral thesis, Byekwaso in [8] used a polynomial Dirichlet model to estimate and test hypotheses about error rates in accounting populations using Bayesian methods with sequential stopping rules. 3.1. Sequential Ratio Test Let f(x,θ) be the density (probability) function with θ=[τ,γ], where τ is subject to verification, and γ is the set of parameters indifferent from the point of view of hypothesis testing H0(τ=τ0) against H1(τ=τ1). We propose a hypothesis H0(θ=θ0). Hypothesis H0 is verified by comparing with the Hypothesis H1(θ=θ1), where θ1 is some specific value different from θ0, and θj=[τj,γ], j=0,1. During the Sequential Probability Ratio Test (SPRT), we progressively sample one or more items at each stage of the sequential hypothesis verification. Depending on the predetermined errors of the α and β, the numbers A=1−βα and B=β1−α satisfying the condition 0