This project will combine logic and probability theory to develop new methods and decision aids for causal reasoning. We will use a new axiomatization of the classical idea of a probability tree, which goes well beyond the current paradigm of Bayes nets in allowing modular representation of probabilistic information. Financial statement auditing will be used as a test bed for the methods, which will ultimately be applicable to many domains where expert systems are used.
The goal of an audit is the expression of an opinion on whether a set of financial statements presents a business's financial position and operations fairly in all material respects, in accordance with generally accepted accounting principles. In reaching this opinion, the auditor may need to consider a variety of subsidiary questions: whether items such as the value of inventory are materially correct; whether internal control is operating effectively; whether the business can continue as a going concern, etc. The evidence bearing on these questions evidence is usually persuasive but not conclusive, and its assessment must be based on a causal model of the business, its environment, and the audit.
Prior research has focused on representing and aggregating uncertainty within a model. Although this research has met its own goals reasonably well, the model itself is almost always either overly simplified or else very specialized, and hence the research has not produced decision aids with the flexibility to serve auditors going into new engagements. Such aids would have to help auditors build causal models, a task they find challenging.
Constructing adequate models means refining models. We always begin an audit with a simplified model, but the audit constantly raises new issues that must be incorporated into this model. The modularity allowed by our approach permits this constant refinement. We will develop an inference engine that permits such refinement, together with detailed casual models for an illustrative group of audit problems and prototype decision aids that may provide a test-bed for future empirical research on audit judgment.
Probabilistic models currently available to auditors are neither simple enough to be used effectively nor complex enough to fit practical auditing realities. This project will overcome these limitations by developing a clear, rigorous and usable concept of refinement. When new evidence reveals that factors previously deemed unimportant are, in fact, relevant, they can be taken into account by refining the existing model, rather than replacing it. This ability to reason simultaneously and consistently with causal analyses at different levels of detail will allow the auditor to avoid models that are hopelessly naive or impossible complex
