Performance of experimental dynamic system identification depends on sensor placement especially when the number of sensor locations is relatively small and measurement noise is significant. We propose a method to select robust measurements based on the estimation error in modal analysis, which is one of the most popular system identification methods for dynamic systems, by defining a measure of the estimation. The measure is developed to deal with various types of sensors and with general damped systems. It is calculated from the eigenvector information obtained by finite element analysis or preliminary experiments and represents the weighted covariance of estimation error induced by measurement noise as well as the orthogonality of the projection matrix which is the corresponding submatrix of eigenvector-matrix by partially measured outputs. Effectiveness of the proposed measurement selection schemes is demonstrated by computer simulations and experiments.