Abstract:
|
Consider a regression model E(Y) = XB having p predictors. Suppose in a sample of n subjects, complete information is available on the response Y. However, certain number of values are unavailable on a specific predictor. Fitting a response based on (p-1) predictors is not recommended for two reasons: one, it results in the wastage of data and second, a response equation based on (p-1) variates may not be appropriate. The literature offers several approaches for imputing the missing values (mostly in the case missing response Y), such as the mean substitution approach, regression approach, principal component approach, and nearest neighbor rule. The effectiveness of these different approaches is measured by using various criteria. In this paper, the authors will compare various imputation approaches by investigating the powers of certain tests for the regression coefficients. In addition, another imputation technique based on a different measure of nearness will be discussed, and its performance will be compared with that of some existing methods on the basis of the power of the regression tests.
|