Frequency-dependent equivalent linear (FD-EQL) site response analysis methods, developed as potential substitutes for the EQL procedure to better simulate the nonlinear soil response, have been reported to overpredict the high-frequency wave propagation. Modified procedures proposed to overcome this limitation have demonstrated to improve the fits with the nonlinear analysis results. The method has not yet been applied to perform a deconvolution analysis, where a conventional EQL analysis often fails to converge or provide reliable estimate. The conventional EQL procedure and one of the modified FD-EQL method, which uses an empirical factor f to interpolate the strain spectrum between the EQL and FD-EQL outputs, are used to perform a series of deconvolution analyses using an idealized 1000 m profile and twelve Kik-net downhole arrays. The residuals of recorded and deconvoluted motions are shown to increase with strain for the EQL method even when using the recommended frequency cut-off. For the FD-EQL method, the range of f recommended for convolution analyses is shown to provide unrealistic responses in performing deconvolution analyses. A new range that produces unbiased residuals for all strain amplitudes is proposed. Comparisons highlight that the modified FD-EQL yields reliable predictions of the within motion for all profiles and motion intensities, automatically suppressing the amplification of high-frequency noise typically accompanied in a deconvolution analysis.