This article presents a quantitative case study on the site amplification effect observed at Heathcote Valley, New Zealand, during the 2010-2011 Canterbury earthquake sequence for 10 events that produced notable ground acceleration amplitudes up to 1.4g and 2.2g in the horizontal and vertical directions, respectively. We performed finite element analyses of the dynamic response of the valley, accounting for the realistic basin geometry and the soil non-linear response. The site-specific simulations performed significantly better than both empirical ground motion models and physics based regional-scale ground motion simulations (which empirically accounts for the site effects), reducing the spectral acceleration prediction bias by a factor of two in short vibration periods. However, our validation exercise demonstrated that it was necessary to quantify the level of uncertainty in the estimated bedrock motion using multiple recorded events, to understand how much the simplistic model can over- or under-estimate the ground motion intensities. Inferences from the analyses suggest that the Rayleigh waves generated near the basin edge contributed significantly to the observed high frequency (f>3Hz) amplification, in addition to the amplification caused by the strong soil-rock impedance contrast at the site fundamental frequency. Models with and without considering soil non-linear response illustrate, as expected, that the linear elastic assumption severely overestimates ground motions in high frequencies for strong earthquakes, especially when the contribution of basin edge-generated Rayleigh waves becomes significant. Our analyses also demonstrate that the effect of pressure-dependent soil velocities on the high frequency ground motions is as significant as the amplification caused by the basin edge-generated Rayleigh waves.