Seismic hazard analyses require an estimate of earthquake ground motions from future events. These predictions are achieved through Ground Motion Prediction Equations, which include a prediction of the median and the standard deviation of ground motion parameters. The differences between observed and predicted ground motions, when normalized by the standard deviation, are referred to as epsilon (𝜖). For spectral accelerations, the correlation structure of normalized residuals across oscillator periods is important for guiding ground motion selection. Correlation structures for large global datasets have been studied extensively. These correlation structures reflect effects that are averaged over the entire dataset underlying the analyses. This paper considers the effects of site response, at given sites, on the correlation structure of normalized residuals. This is achieved by performing site response analyses for two hypothetical soil profiles using a set of 85 rock input motions. Results show that there is no significant difference between correlation coefficients for rock ground motions and correlation coefficients after considering the effects of site response for the chosen sites.