Accurately predicting surface ground motions is critical for many earthquake engineering applications. Equivalent-linear (EQL) site response analysis is a numerical technique used to compute surface ground motions from input motions at bedrock using the site-specific dynamic soil properties. The purpose of this study was to investigate the accuracy of EQL site response analysis for stiff soil sites by comparing computed and observed transfer functions and response spectral amplification. The Kiban Kyoshin network (KiK-net) in Japan is a seismograph network consisting of downhole array sites with strong-motion accelerometers located at the ground surface and at depth. Recorded motions and shear wave velocity profiles are available for most sites. Observed transfer functions and response spectral amplification were computed for 930 individual seismic recordings at 11 stiff soil KiK-net sites. Computed transfer functions and response spectral amplification were calculated from EQL site response analysis by specifying the KiK-net base sensor motion as the input motion. Sites were characterized using the measured shear wave velocity profiles and nonlinear soil properties estimated from empirical models. Computed and observed transfer functions and response spectral amplification were compared at different levels of strain for each site. The average difference between the observed and computed response spectral amplification across the 11 sites were compared at different levels of strain. Overall, there is reasonable agreement between the computed and observed transfer functions and response spectral amplification. There is agreement between the computed and observed site periods, but with over-prediction of the computed response at the observed site periods. Higher modes often computed by the theoretical model were not always observed by the recordings. There is very good agreement between the computed and observed transfer functions and response spectral amplification for periods larger than the site periods. There is less agreement between the computed and observed transfer functions and response spectral amplification for periods less than the site periods. There is mostly over-prediction of the response spectral amplification at these periods, although some under-prediction also occurred. Across all 11 sites the predicted spectral amplification is within +/-20% at shear strains less than 0.01%. At shear strains between approximately 0.01 and 0.03%, the spectral amplification is over-predicted for these sites, in some instances by as little as 5% and in other instances by a factor of 2 or more. / text