Tradeoffs in the choice between logit and OLS for accounting choice studies.
通过模拟和实际会计数据，比较了Logit和OLS在小样本、偏态预测变量下的统计性能，发现Logit在样本量低至50时仍优于OLS，为会计选择研究的方法选择提供依据。
Abstract Many accounting studies examine dichotomous choices (e.g., qualify/do not qualify an audit opinion or capitalize/do not capitalize a cost). These studies often involve small overall sample sizes, disparate response group sizes, and predictor variables that are skewed and collinear. These factors can cause distributional problems in test statistics for logit (or probit) regression models, which can lead to incorrect inferences. However, few empirical benchmarks exist for assessing the effect of these factors. The current paper determines how response group size and the number, distribution, and correlation of predictor variables affect empirical error rates and the minimum required sample size for using logit. Comparisons are made with the error rates obtained from an ordinary least squares (OLS) linear probability model, an alternative that has been suggested for small sample studies. Because accounting researchers choosing between logit and OLS may be concerned with more than the calibration of the models' test statistics, comparisons of the sensitivity of logit and OLS parameter estimates to the range of data sampled for the predictor variables and of the models' classificatory ability also are made. Both simulated and real accounting data are used. The results of Monte Carlo simulations show that logit test statistics are biased when the sample size is small However, much of the bias is attributable to the skewness of the predictor variables, a problem that is characteristic of accounting research and that also affects OLS test statistics. In such settings, OLS may result in test statistics that are minimally better calibrated. The parameter estimates of the model, however, will be more sensitive to the sampling frame. Furthermore experimentation with data on auditors' Statement No. 87 consistency judgments indicates that 01$ also may result in higher Type I error rates when it is used for prediction or classification These results are interpreted as indicating that, even for sample sizes as small as 50, log it rather than OLS still may be the preferable model for accounting choice studies.