The method of choice for multivariate representation of community structure is often non-metric multi-dimensional scaling (MDS). This has great flexibility in accomn~odating biologically relevant (i.e. non correlation-based) definitions of similarity In species composition of 2 samples, and in preserving the rank-order relations amongst those similarities in the placing of samples in an ordination. Correlation-based techniques (such as Canonical Correlation) are then inappropriate in linking the observed biotic structure to measured environmental variables; a more natural approach is simply to compare separate sample ordinations from biotic and abiotic variables and choose that subset of environmental variables which provides a good match between the 2 configurations. In fact, the fundamental constructs here are not the ordination plots but the (rank) similarity matrices which underlie them: a suitable measure of agreement between 2 such matrices is therefore proposed and used to define an optimal subset of environmental variables w h ~ c h 'best explains' the biotic structure. This simple technique is illustrated wlth 3 data sets, from studles of macrobenthic, meiobenthic and diatom communities in estuarine and coastal waters.