The stochastic estimation of coherency is perceived, in some studies, to be significantly influenced by the time window length under consideration. For the engineering purposes, usually, coherency is estimated from the strong motion, i.e., the shear (S-) wave segment of the recorded time series. However, the identification of a purely S-wave dominated window could be challenging in case of complex wave mixing on the seismograms. Moreover, there has been relatively little research on how much variation, in reality, is introduced in the coherency estimates owing to the choice of signal length. Therefore, the debate about the procedure for selecting a representative time window has remained inconclusive: different authors keep their different practices. The current article is an effort to shed light on the sensitivity of the coherency estimates to the choice of various time windows. The research draws on the dataset gathered from a dense seismic-array deployed at the small size, shallow alluvial valley of Koutavos-Argostoli, situated in Cephalonia Island, Greece. The lagged coherency over interstation distance ranges from 5 to 80 m has been estimated from a set of 46 earthquakes having magnitude 2–5 and epicentral distance 3–200 km, considering different lengths of time windows. The findings revealed that the statistics of coherency estimates derived from many events is only weakly sensitive to the selection of time window lengths provided that the segments include the same energetic pulse.