Abstract
Measurement error is a serious problem in various scientific areas. The subjects of interests are suffered from inaccurate, imperfect and unobservable measurement. In addition, continuous exposure variables are frequently partitioned into categorical variables and those categorized exposure variables are fitted in the regression model. Both measurement error and dichotomization can lead to considerable loss of power and relative efficiency. This thesis adopts the hypothesis testing on the association between response and exposures with a specified power at fixed significance level in the linear regression model in which the explanatory exposures are subject to either measurement error or dichotomization.