Using the strong ground motion dataset of Mw 4.0–6.0 earthquakes in Japan from K‐NET and KiK‐net, we studied empirically the 5% damped spectral acceleration amplification for periods from 0.05 to 5 s at 75 seismic stations located in the Iwate–Miyagi and Niigata regions. The method of study is based on estimating the empirical site effect from the spectral ratio of observed and predicted ground motions based on the ground‐motion prediction equations (GMPEs). Our results show a large amplification at most analyzed stations at periods of 0.1 s and shorter; at periods of 0.3 s and longer, large amplifications generally dominate at basin stations, but deamplifications generally dominate at mountain stations. Empirical amplification factors of spectral acceleration were used to correct the observed strong ground motion data of two crustal earthquakes: the 2008 Iwate–Miyagi Nairiku earthquake (Mw 6.9) and the 2004 Chuetsu earthquake (Mw 6.6). The site‐corrected data fit better with GMPEs inferred by less standard error. We took advantage of a large body of empirical data to develop a mean empirical site‐response model for each region. These empirical models are superior to models predicted from GMPEs using only VS30 site corrections. Those relations have shown a low correlation at short periods such as 0.1 s. We also found the relations are different for the mountain and basin stations. The method is efficient for calculating site effects at observation stations where insufficient velocity structure information and a history of earthquake recordings are available.