Ground motion attenuation relationships for spectral acceleration have been developed in Japan. Although these prediction models can usually provide a mean value and a standard deviation for each specified period, they do not characterize covariance property of spectral acceleration between two different periods. This leads to less accurate prediction of response of structural systems such as multi-degree-of-freedom systems and nonlinear systems. Some studies have paid attention to this shortcoming and tried to model the correlation of spectral accelerations by statistical method based on the small number of records. In the present paper, a large amount of ground motion records, which were observed from the recent earthquakes by K-NET and KiK-NET, were used in the covariance analysis. The covariance structure of two spectral values were investigated through principal component analysis. The covariance characteristics were fully investigated by examining the effects of the different source, path and soil conditions. And response spectra are simulated using principal components as an application of principal component analysis.
