Influential Observation Detection For Model Averaging Estimator In Linear Models

Model averaging is an important method for multi-model inference.More robust estimators and better predictions are obtained by combining different candidate models.The theoretical development of model averaging has attracted increasing research interest Moreover,model averaging techniques have been widely used in comuter science,economics,medical studies and biology.Like many traditional estimators,model averaging estimators are vulnerable to the presence of influential observations.However,existing literature on statistical diagnostics rarely considers the multi-model inference and very few studies have been considered for influential observation detection in the context of model averaging.Therefore,it is utterly important to study the influential observation detection within the framework of model averaging.Current thesis studies the influential observation detection for model averaging estimator in linear models.The weight choice methods in model averaging is reviewed and Cook distance for model averaging estimator is developed based on case-deletion method.The proposed method is assessed by simulation study.Finally,the diagnostic instrument is employed to analyze the Boston housing price data and car seats data,the results are reasonable in general.