Sound Confidence Intervals in the Heteroscedastic Linear Model Through Releveraging
针对异方差下普通最小二乘回归置信区间覆盖不足的问题，提出一种新的简单加权方案，虽可能牺牲效率，但在数据探索和回归诊断阶段有实用价值。
SUMMARY Under heteroscedasticity, ordinary least squares regression can fail to yield adequate inference on parameter coefficients with respect to hypothesis testing or confidence intervals. For example, confidence intervals can have coverage well below that claimed. This is especially the case in small or moderate-sized imbalanced samples and holds even if the ordinary least squares variance estimator is replaced by a robust-against-heteroscedasticity variance estimator. A new simple weighting scheme corrects this problem, although at possibly serious cost in efficiency. Inefficient methods have been rated as useful at the data exploration stage of analysis. The present method is also useful as an adjunct to least squares regression, at the stage of regression diagnostics. As such, it can sometimes replace the need for them altogether.