Influential observations and inference in accounting research

Accounting studies routinely encounter observations taking on extreme values. Such observations can influence statistical estimates (coefficient) and inferences. Our survey of the accounting literature documents that the two most common approaches used to address influential observations are winsorization and truncation. Although widely used, there is little systematic evidence in the literature as to their efficacy. While expedient, both approaches depend on researcher-selected cutoffs, are often applied on a variable-by- variable basis, and can alter legitimate data (i.e., valid) points. We compare winsorization and truncation with influence diagnostics (e.g. Cook's distance) and robust regression (RR), the latter two of which are suggested in the statistics and econometrics literatures. We replicate three published studies to show how the choice of method to account for influential observations impacts estimates and inferences. We also use simulations to compare the alternative approaches. We find that winsorization and truncation are largely ineffective in mitigating the effect of influential observations. Although both influence diagnostics and RR outperform winsorization and truncation, RR generally outperforms all other methods. RR, which focuses on model fit to deal with influential observations, is relatively straightforward to implement in accounting settings. While RR is not without limitations, overall, our findings lead us to recommend that future accounting studies consider using RR, or at least report sensitivity tests using RR