The two parts of this paper form a critique of a variety of statistical techniques of actual or potential use in quantitative structure‐activity relationship (QSAR) studies and related fields. Part I explores the statistical thinking that is needed to underpin those techniques. Emphasis as placed on (a) the role of ‘exchangeability’ as an alternative to unrealistic statistical modelling and (b) the use of cross‐validation to limit self‐deception in the use of any particular technique. The problem of the almost unlimited range of molecular descriptors is seriously addressed. (Part II provides a concise critical review of methods‐some well‐established and some new.) Copyright © 1993 John Wiley & Sons, Ltd.