Date of Completion

7-30-2013

Embargo Period

1-26-2014

Keywords

MM-estimation Winsorization Truncation Robust estimator outliers

Major Advisor

Gim Seow

Associate Advisor

Yonghong An

Associate Advisor

Amy Dunbar

Field of Study

Business Administration

Degree

Doctor of Philosophy

Open Access

Campus Access

Abstract

Outliers can lead to incorrect inferences in least squares regression. In this study, I evaluate outlier treatment methods: MM-estimation as named in Yohai (1987), winsorization, and truncation, providing more evidence for outlier treatment in accounting research. Simulation results where I seed known “outliers” suggest that MM-estimation has the best performance. In particular, in the presence of the most influential type of outliers, theory-based MM-estimation provides inferences similar to the ordinary least squares (OLS) estimator, i.e. the best linear unbiased estimator, whereas winsorization and truncation fail to provide right inferences over 99 percent of the time. I also show that MM-estimation provides the most accurate predicted earnings and estimated discretionary accruals: with truncation creating more bias and inference problems than not treating the outliers, the means (medians) of the absolute earnings prediction errors and discretionary accruals using winsorization are higher than those using MM-estimation by more than 21% (48%) on average. In sum, MM-estimation provides more reliable results, which should be an important consideration for future accounting research employing datasets containing outliers.

Download
COinS