Effects of sampling error on bioassessments of stream ecosystems: application to RIVPACS-type models

We evaluated the influence of various sources of sampling error on the precision, accuracy, and sensitivity of bioassessments based on River Invertebrate Prediction and Classification System (RIVPACS)-type models. We used data from 98 minimally altered streams in western Oregon and Washington to generate 18 models representing 2 field collection techniques (fixed-area riffle and multiple-habitat collections) and 9 levels of subsampling (50, 100, ...450 fixed counts). For each model, we generated 2 observed-to-expected taxa ratios (O/E). The 1(st) O/E was based on all predicted taxa and the 2(nd) excluded rare taxa (i.e., those taxa with predicted probabilities of occurrence <0.5). We then compared O/E values to determine the extent to which subsampling effort, field collection method, different field personnel, and individual site characteristics altered model performance. We also generated O/E values for 63 streams with varying amounts of watershed and channel alteration (test sites) to quantify the extent to which sampling error influenced the sensitivity of these models in detecting biological impairment. Model precision improved with increased sampling effort. However, neither collection method consistently led to more precise models. Model accuracy generally was not affected by subsampling effort, sample collection techniques, or different sample collectors. However, similar to50% of the error in predictions was associated with unexplained characteristics of individual sites. Average assessment values across all test sites were robust to both collection method and subsampling effort. However, inferences about the biological condition of some test sites varied among models. Precision, accuracy, and sensitivity all improved with the exclusion of rare taxa, although O/E values for individual test sites were more variable among models based on different subsample counts when rare taxa were excluded from model calculations. Overall, we found that the effects of sampling error on RIVPACS model performance can be minimized by constructing models from subsamples of greater than or equal to350 individuals and by excluding rare taxa from calculations of O/E.