Stream macroinvertebrate response models for bioassessment metrics: Addressing the issue of spatial scale

Abstract

We developed independent predictive disturbance models for a full regional data set and four individual ecoregions (Full Region vs. Individual Ecoregion models) to evaluate effects of spatial scale on the assessment of human landscape modification, on predicted response of stream biota, and the effect of other possible confounding factors, such as watershed size and elevation, on model performance. We selected macroinvertebrate sampling sites for model development (n = 591) and validation (n = 467) that met strict screening criteria from four proximal ecoregions in the northeastern U.S.: North Central Appalachians, Ridge and Valley, Northeastern Highlands, and Northern Piedmont. Models were developed using boosted regression tree (BRT) techniques for four macroinvertebrate metrics; results were compared among ecoregions and metrics. Comparing within a region but across the four macroinvertebrate metrics, the average richness of tolerant taxa (RichTOL) had the highest R2 for BRT models. Across the four metrics, final BRT models had between four and seven explanatory variables and always included a variable related to urbanization (e.g., population density, percent urban, or percent manmade channels), and either a measure of hydrologic runoff (e.g., minimum April, average December, or maximum monthly runoff) and(or) a natural landscape factor (e.g., riparian slope, precipitation, and elevation), or a measure of riparian disturbance. Contrary to our expectations, Full Region models explained nearly as much variance in the macroinvertebrate data as Individual Ecoregion models, and taking into account watershed size or elevation did not appear to improve model performance. As a result, it may be advantageous for bioassessment programs to develop large regional models as a preliminary assessment of overall disturbance conditions as long as the range in natural landscape variability is not excessive.