Seminars

Study on adaptive model selection through generalized degrees of freedom in nested linear regression models

106
reads

Chuan-Fa Tang

2010-06-11
10:00:00 - 12:00:00

504 , Freshman Classroom Building

Various model selection criteria have been proposed to fit linear regression models to data, such as AIC (Akaike 1974), BIC (Schwarz 1978) and Mallows' Cp (1973). Those criteria is the sum of residual sum of squares and the complexity of model. The complexity of model is the number of unknown parameters time a tuning constant lambda. For AIC and Cp, lambda is 2 while lambda is n for BIC. Here n is the number of observations. Shen and Ye (2002) proposed the adaptive model selection by determining pr lambda through general final prediction error. However, it is not clear whether their proposal leads to a better model selection method in terms of consistency in selecting correct model. In this talk, we will introduce the adaptive models selection criterion through generalized degrees of freedom proposed by Shen and Ye (2002). Show that proposal is not fully adaptive unless the range of lambda is properly restricted in nested linear regression models and the sample size is much larger than the number of covariates.