Modeling natural environmental gradients improves the accuracy and precision of diatom-based indicators

摘要

Diatom-based indicators can contribute significantly to comprehensive assessments of stream biological conditions. We used modeling to develop, evaluate, and compare 2 types of diatom-based indicators for Idaho streams: an observed/expected (O/E) ratio of taxon loss derived from a model similar to the River InVertebrate Prediction And Classification System (RIVPACS) and a multimetric index (MMI). Modeling the effects of natural environmental gradients on assemblage composition is a key component of RIVPACS, but modeling has seldom been used for MMI development. Diatom assemblage structure varied substantially among reference-site samples, but neither ecoregion nor bioregion accounted for a significant portion of that variation. Therefore, we used Classification and Regression Trees (CART) to model the variation of individual metrics with natural gradients. For both CART and RIVPACS modeling, we restricted predictors to natural variables unaffected by or resistant to human disturbances. On average, 46% of the total variance in 32 metrics could be explained by CART models, but the predictor variables differed among the metrics and often showed evidence of interacting with one another. The use of CART residuals (i.e., metric values adjusted for the effect of natural environmental gradients) affected whether or how strongly many metrics discriminated between reference and test sites. We used cluster analysis to examine redundancies among candidate metrics and then selected the metric with the highest discrimination efficiency from each cluster. This step was applied to both unadjusted and adjusted metrics and led to inclusion of 7 metrics in MMIs. Adjusted MMIs were more precise than unadjusted ones (coefficient of variation similar to 50% lower). Adjusted and unadjusted MMIs rated similar proportions of the test sites as being in nonreference condition but disagreed on the assessment of many individual test sites. Use of unadjusted MMIs probably resulted in higher rates of both Type I and Type 11 errors than use of adjusted metrics, a logical consequence of the inability of unadjusted metrics to distinguish the confounding effects of natural environmental factors from those associated with human-caused stress. The RIVPACS-type model for diatom assemblages performed similarly to models developed for invertebrate assemblages. The O/E ratio was as precise as the adjusted MMI, but rated a lower proportion of test sites as being in nonreference condition, implying that taxon loss was less severe than changes in overall diatom assemblage structure. As previously demonstrated for O/E measures, modeling appears to be an effective means of developing more accurate and precise MMIs. Furthermore, modeling enabled us to develop a single MMI for use throughout an environmentally heterogeneous region.