First, a method is presented for the study of the conditional prevalence rate of a sensitive attribute, as a function of covariates. Then this method is also applied to estimation of the unconditional prevalence rate. It is demonstrated that this permits recovery of some of the usual loss in accuracy due to randomization.