The physical oceanographic literature contains many instances of cross spectral and coherence calculations in which statistical stability is achieved through band averaging of the cross spectrum in frequency space. 'Constant Q' averaging is often used, sometimes with unexpected results. We show how 'constant Q' frequency smoothing of the cross spectrum can lead to erroneous estimates of coherence. We propose that the well known technique of segment averaging using a number of raw spectral estimates computed from a sequence of time-domain windowed data be used for routine coherence (or cross-spectrum) calculations. The effects of window shape on coherence estimation are also considered.