Macrobenthic data from samples taken in 1980, 1983 and 1985 along a pollution gradient in the Southern California Bight (USA) were analyzed at 5 taxonomic levels (species, genus, family, order, phylum) to determine the taxon and sample size sufficient for assessing pollution impacts on 5 measures of community structure. Two statistical designs were compared: a t-test for differences between reference and impacted stations where the error term was (1) among-year variation at the reference station (impact effects design), (2) replicate (within-station) sampling error (location effects design). The estimated statistical power (1-Beta) to detect impacts was a function of type and magnitude of impact, level of taxonomic identification, the statistical design, and the sample size (ni = number of sampling years at the reference station for the impact effects design, and nl = number of replicate samples per station for the location effects design). Four replicate 0.1 sq m van Veen grabs per station were needed to ensure community-wide, unbiased estimates of Shannon's, 1-Simpson's and McIntosh's Index. Family-level identification appeared to be a good choice for assessing pollution impacts at the study site as it ensured a high probability (1-Beta > or = to 0.80) of detecting intermediate or larger impacts on most (impact effects design) or all (location effects design) of 5 measures of community structure when ni and nl > or = to 4. The level of taxonomic identification and sample size should be considered along with other sampling variables (e.g. sample unit size, sieve mesh size) when seeking a statistically rigorous, cost-effective study design sufficient to meet pollution assessment objectives.