DISTRIBUTIONAL TESTING OF DATA FROM MANUFACTURING PROCESSES

A common characteristic of classical variance investigation models includes the assumptions that the in‐control cost observations from manufacturing processes are independently and normally distributed. The purpose of this paper is to empirically assess the appropriateness of the above assumptions in an accounting setting. Several sets of data generated from manufacturing processes were examined, with the following results. First, normality and independence assumptions appeared to be violated with the daily data. In addition, several transformations commonly employed in the statistics literature did not alter this conclusion. Second, the above assumptions were not violated with the weekly data. It was then demonstrated that these findings will, in general, hold. These results, subject to the inherent limitations of a case study approach, led to the following inferences and recommendations. Using classical statistical techniques and daily data to control a process incurs costs that have been previously overlooked, costs resulting from distributional assumption violations. If these costs exceed the benefits of using daily data, a manager should (1) use classical techniques to monitor weekly data, (2) use Tchebycheff inequalities to monitor daily data, and/or (3) use variance investigation models that more formally incorporate the serial dependencies of daily data.