Sample Variability Influences on the Precision of Predictive Bioassessment

The rapid bioassessment technique we investigate (AUSRIVAS) requires a nationally standardized sampling protocol that uses a single collection of macroinvertebrates (without replication) taken from 10 m of specific habitats (e.g. stream edge and/or riffle) and sub-samples of 200 animals. The macroinvertebrate data are run through predictive models that provide an assessment of biological condition based on a comparison of the animals found in the collection (the observed) and those expected to be there given the site-specific characteristics of the stream (the O/E taxa score). The important questions are related to the conclusions regarding river condition that can be drawn from the biological assessment. Rapid bioassessment studies are generally of two types: those for assessment of individual sites and those where many sites are selected to collectively assess the potential impacts of some human activity such as forestry or agriculture. We wanted to identify the effects of sample variability on the outputs of this predictive bioassessment technique. We found that a single collection of benthic macroinvertebrates was sufficient for bioassessment when taken from a site that had a large area of nearly uniform substrate and was in good condition. Also, collections taken from a larger and smaller area of substrate (1.75, 3.5 or 7 m2) gave the same bioassessment. In other sites, not in such good condition, the variability in bioassessment from different collections could result in different interpretations of biological condition. For all sites, regardless of condition, much of the variation in bioassessment was derived from sub-sampling the macroinvertebrates. We develop a statistical sub-sampling and solver algorithm that provides a measure of variability and a statistically valid probability of impairment for a single site, without the need to actually collect the hundreds of replicated collections needed for this study. We found that assessment at impaired sites, where only 1 collection and 1 sub-sample are taken (a common situation in rapid assessment), the 95% confidence level for O/E taxa scores is estimated to be as much as ±0.22. At sites in reference condition, the 95% confidence interval may be much narrower (~±0.1 O/E units). Therefore, assessments of sites at, or near, reference condition will be more precise than for impaired sites. Power analysis revealed that where single sites are being assessed we recommend a sample collected from 3.5 m2 of habitat, but replicate collections should be taken at a site (rather than one only) and we recommend replicate sub-samples of each collection (total of six sub-samples from a site). However, this would remove a ‘rapid’ component of the bioassessment. We recommend the addition of sub-sampling and solver algorithms to the predictive models such as AUSRIVAS to provide a statistical measure of probability of impairment. An adaptive sub-sampling regime could then be used to optimize sampling effort. For example, a single sub-sample may be sufficient for screening or the agency could use the sub-sample and solver algorithms to sub-sample the parent sample for a more precise estimate of the biological condition. Replication should be maximized at the spatial scale required for reporting: site, or regional. But as a general rule, catchment or land-use scale studies should maximize replicate sites, and site-scale assessments should maximize replication within sites.