Estimating baseline values of the variable of intervention in a clinical trial

Parametric empirical Bayes methodology is suggested for determining estimators of individual baseline values of the variable of intervention in a clinical trial, when the variable is measured twice--once for subject selection, and again, without selection, just before randomization. The resulting compromise estimator is seen to have more precision than the baseline estimator employing only the second value and less bias than the estimator that simply averages the two values. Construction of such an estimator is illustrated using data from the recruitment phase of the Lipid Research Clinics Coronary Primary Prevention Trial. Generalizations to other designs are also suggested. In all cases, however, an estimate of the intraindividual variance of the variable of intervention is required.