Abstract Traditional environmental studies have employed sampling at different times, but based on re-randomized ‘replicate’ samples taken at each time. For example, in a 4 year monitoring study of near-shore marine benthic communities there might be three box cores collected annually at each of three depths along each of three transects. Repeated measures designs, long used in medicine and the social sciences, are based on resampling replicates (e.g. sites) at a series of times. In such designs spatial sampling variability is not used for tests of environmental impact. Error for such tests is based on variability of time trends among similar sites (similar with respect to impact). For example in a tropical oil spill study five oiled and five unoiled coral reefs were studied over 5 years. Error for tests of oil impact was based on variability among reefs (within degree-of-oiling category) in the year-to-year trends of biological response variables. It was not based on variability among field samples within reefs at given times.
The two approaches (univariate and multivariate) to repeated measures analysis of variance are described. The pros and cons of each are discussed, as are the assumptions and consequences of their violations. Emphasis is especially placed on the adequacy of error degrees of freedom in the two approaches, and some exploration of power to detect impact is presented. Examples of application of repeated measures designs to various impact and monitoring studies are presented and discussed, including (i) interpretation of significant effects; (ii) decomposition of effects by contrasts (e.g. before vs after impact); and (iii) modelling time trends by polynomial and cosine functions.