A COMPARATIVE STUDY OF AUDIT EVIDENCE AND AUDIT PLANNING MODELS USING UNCERTAIN REASONING

Auditing requires the aggregation of evidence. The Audit Risk Model proposed in professional standards is used to aggregate audit evidence, despite extensive criticism; nevertheless, it does not address interdependencies among audit objectives, and among audit evidence and audit objectives. Network solutions have been proposed, but pose significant computational difficulties. This study examines the role of evidence in the formation of audit opinions, and considers alternative paradigms for modeling audit planning.; Existing audit planning models have largely been based on probability theory, which has semantic limitations in representing ignorance. Dempster-Shafer belief functions offer an alternative; however, auditors have difficulty in articulating the required probabilities that the change to belief functions will not alleviate. Spohn's epistemic beliefs make fewer epistemological demands, and Zadeh's possibility theory is computationally equivalent; neither has been studied in auditing contexts. This dissertation explores the comparative advantages and disadvantages of these four systems for representing uncertainty in auditing.; Valuation-Based Systems provide an abstract framework generalizing these calculi, using local computation techniques. This study uses Valuation Networks based on such systems to represent interdependencies among audit objectives, and among audit evidence and audit objectives, and compares their different representations of audit assurance. Modeling issues examined include: representation of "AND" and discounted "AND" nodes; identification of joint valuations given marginals; representation of statistical evidence; and hierarchical aggregation of audit evidence at different levels.; This study develops new auditing postulates and an extended theory of audit evidence. It provides the first Bayesian probability model of audit planning to address the various modeling issues raised. Existing belief function models are extended by provision of: statistical methods for Monetary Unit Sampling and attribute sampling, discounted "AND" nodes, revised proposals regarding joint valuations, and alternative nodes for hierarchical aggregation. Models using epistemic beliefs and possibility theory are provided for the first time: "AND" nodes are proposed, and the impossibility of discounted "AND" nodes is proved; joint valuations are identified from marginals; and hierarchical aggregation is considered. A method is developed (and applied) for eliciting values that may be interpreted within the four uncertain reasoning paradigms. Finally, extensive proposals for future research projects are given.