The 2010–2012 Canterbury earthquake sequence generated a large number of near‐source earthquake recordings, with the vast majority of large events occurring within 30 km of Christchurch, New Zealand’s second largest city. We utilize the dataset to estimate the site attenuation parameter, κ0, at seven rock and stiff‐soil stations in New Zealand’s GeoNet seismic network. As part of this study, an orientation‐independent definition of κ is proposed to minimize the influence of observed high‐frequency 2D site effects. Minimum magnitude limits for the traditional high‐frequency fitting method are proposed, based on the effect of the source corner frequency. A dependence of κ0 on ground‐shaking level is also observed, in which events with large peak ground accelerations (PGAs) have lower κ0 values than events with small PGAs. This observation is not fully understood, but if such a trend holds in future investigations, it may influence how κ0 is used in hazard assessments for critical facilities. κ0 values calculated from Fourier amplitude spectra of acceleration (κ0,AS) are compared with the native κ0 of local, empirical, ground‐motion prediction equations (GMPEs), calculated using the inverse random vibration theory method (κ0,IRVT). κ0,IRVT is found to be independent of magnitude and distance and agrees with the average κ0,AS for the region. κ0,IRVT does not scale strongly with VS30, indicating that current GMPEs may be capturing the average kappa effect through the VS30 scaling. The results from this study are of particular interest for site‐specific ground‐motion prediction studies as well as for GMPE adjustments between different regions or rock types.