Stein's Paradox and Audit Sampling
探讨审计师在多元估计中面临的问题，指出传统最大似然估计在多元正态分布均值向量估计中不可接受，并介绍斯坦因悖论及改进方法。
The auditor is interested in estimating many variables in performing his/her attest function. These estimates include such items as error rates, confidence intervals, maximum overor understated amounts, and account balances. These estimates along with other collateral evidence comprise a multivariate information set upon which the auditor concludes that a set of financial statements fairly presents the financial condition and operating results of the firm. In this multivariate context, the auditor should consider the efficiency of the procedures used to estimate the parameter set upon which the decisions are made. Traditional procedures for obtaining these estimates include the maximum likelihood estimation (MLE) procedure and Bayesian approaches. Bayesian techniques require a considerable amount of judgment and training. Moreover, Stein [1955] proved that the MLE was inadmissible (could be improved upon over some portion of parameter space without worsening over the remainder of the space) as an estimator for the mean vector of a multivariate normal distribution. James and Stein [1961] and Efron and Morris [1971; 1972; 1973; 1975; 1977], among many others, generalized and extended Stein's original proof.' The result