Abstract Regression analysis is a useful audit tool, but regression-based decision rules rely on several underlying assumptions. The assumption of error-term independence is frequently violated when auditors use small-interval, timeseries data. The violation of this assumption, referred to as autocorrelation, biases the standard error of the regression coefficient and the standard error of the estimate. These biases influence precision calculations as well as tests of significance for the regression coefificients. This study compares the efficiency and effectiveness of models with and without autocorrelation, to evaluate whether transformed models increase efficiency and/or effectiveness, and to discuss the advantage of transformed models in terms of improved precision. This study simulates 200 data series, some of which result in models exhibiting autocorrelation, and others resulting in models with no violation. Material errors are randomly seeded into the audit years (the most recent twelve months), and three regression-based decision rules (STAR, KA, KS) are used in an attempt to identify these errors. The number of incorrect rejections (signaling a month for investigation when no error exists) and the number of incorrect acceptances (not signaling a month for investigation when a material error exists) are tabulated for each model and each decision rule. The number of incorrect decisions is compared for models with and without autocorrelation. All the autocorrelated models are transformed by the Cochrane-Orcutt Iterative Least Squares method, and the number of incorrect decisions is tabulated for the transformed models. The transformed models are compared to the corresponding original autocorrelated models in evaluating relative efficiency and effectiveness. When raw data were used, models with and without autocorrelation revealed little difference in efficiency and effectiveness. However, when the data of the autocorrelated models were transformed, the decision models resulted in fewer incorrect rejections and fewer incorrect acceptances. Related to this increased efficiency and effectiveness was improved precision as a result of the transformation. Improved precision addresses a primary concern of auditors as it results in fewer unexplained dollars. The transformation lowered the standard error, which reduced the dollars associated with the confidence interval. In terms of improved efficiency and effectiveness (including improved precision), this study lends support to the use of transformed data in regression-based models where raw data are autocorrelated. Such transformation is increasingly available in commercial software.
Arlette C. Wilson (Sun,) studied this question.