Answers are grounded in this paper and cite their sources.
Restricted, preliminary-test and Stein-type estimators incorporate uncertain linear restrictions into the multiple linear regression model, but their risk theory has been developed almost exclusively for symmetric (Gaussian or elliptical) errors. We let the error vector follow a scale mixture of skew-normal (SMSN) distribution with a common mixing variable, a class that contains the skew-normal, skew-t and skew-slash laws. Conditionally, the F statistic for the restriction follows a doubly non-central F law. Under symmetry it is exactly central F for every mixing law and working covariance, whereas under skewness the test is size distorted because the asymmetry changes the error covariance. We give exact bias and risk expressions for the five estimators under a balanced loss with weight ω, and prove that the Stein rule dominates the unrestricted estimator for every symmetric scale mixture of normals, with a dominance interval proportional to 1−ω. Under skewness we derive the exact conditional risk difference and sufficient conditions for unconditional dominance. When the model contains an intercept that the restriction does not involve, the exact dominance interval at the null hypothesis does not depend on the mixing law, and asymptotically, with V=In, the Stein rule dominates for every value of the skewness. The asymptotic conclusion extends to independent mixing variables. Simulations and two applications (Australian Institute of Sport; 1985 Current Population Survey) show that skew-t errors are preferred to Gaussian ones and that, in repeated cross-validation, the Stein and positive-rule estimators are on average never worse than least squares, whereas the restricted estimator is markedly worse.
No takes yet. Share an insight, caveat, or question.
Mohammad Arashi (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: