Data Truncation Bias, Loss Firms, and Accounting Anomalies
研究发现，事后剔除极端收益观测值（非数据错误）会导致市场效率检验和异象行为解释出现虚假推断，尤其在亏损公司多的子样本中，截断偏差使收益均值被低估，并人为制造出收益与预测变量间的虚假关系。
ABSTRACT Ex post trimming of extreme returns observations that are not data errors causes spurious inferences in tests of market efficiency and behavioral explanations for anomalies. Trimming causes a downward truncation bias in estimated mean returns that is stronger in ex ante subsamples with more loss firms and in which return distributions are more right-skewed. There is an asymmetric U-shaped relation between return right-skewness and loss frequency across deciles of negative return predictors (Accruals, ΔNOA, and NOA), and a downward sloping relationship for positive return predictors (CFO and FCF). Consequently, a least-trimmed square (LTS) 1 percent deletion of returns induces a spurious inverted-U-shaped relation between returns and negative predictors, and an exaggerated positive relation for positive predictors. Thus, the resulting trimmed relations do not reject behavioral explanations for these anomalies. Trimming also induces a spurious loss anomaly. These findings highlight that in return prediction studies, observations should not be deleted based upon the values of the dependent variable, only based upon clearly identified data errors.