Detection-Controlled Prediction of Accounting Irregularities: Channel Stuffing as an Illustrative Case

• In Event: Understanding Real Earnings Management

Abstract

A number of prior studies use a sample of detected accounting irregularities, e.g., SEC’s enforcement
actions and class-action lawsuits, to predict the likelihood of accounting violations. The ability of these
models to correctly identify potential violations is inhibited by the fact that all violations that occur are
not observed. We demonstrate that using a partial-observability bivariate probit model to jointly estimate
the probability of commission and the probability of detection of an accounting violation helps to improve
the classification accuracy of the prediction model when not all violations are detected. As an illustrative
example, we use a sample of firms that engaged in accounting irregularities associated with a particular
form of revenue manipulation – channel stuffing. Channel stuffing is an ideal setting for this purpose
because it is difficult to detect without the help of whistle-blowers and provides multiple ex post
indicators that can help validate the model’s identification of potential violators. Our results show that the
power and specification of the bivariate probit prediction model is superior to that of the simple probit
model. We find that a sub-sample of the population of firms identified as having a high likelihood of
channel stuffing by the bivariate probit model has similar industry representation and exhibits future
reversals in sales, production and profitability that closely parallel those of the detected channel stuffing
sample. The same is not true for the sub-sample identified by the simple probit model. In fact, the subsamples
identified by the two models do not exhibit significant overlap. Our results highlight the need to
control for the probability of detection to minimize misclassification in studies predicting accounting
irregularities that are hard to detect.

Authors

• Haiwen Zhang, The Ohio State University

• Pervin K Shroff, University of Minnesota, Twin Cities Campus

• Somnath Das, University of Illinois at Chicago