A framework for estimating probabilistic errors in identification of dynamic characteristics of structures is proposed in this paper. From the view of probability theory observed data are random samples of random variables. Therefore the evaluated parameters obtained by applying least squares method to the observed data are also random variables which have their own variances and correlation between each other. To estimate these probabilistic errors the derivation of the covariance matrix of identified parameter vector is formulated in both frequency and time domain cases. The probabilistic feature of modal parameters and the characteristics of each identification method are discussed.