The influence of taxonomic level on the performance of a predictive model for water quality assessment

Abstract

Predictive models developed to assess water quality in the Mondego River basin (Portugal), based on the BEnthic Assessment of SedimenT (BEAST) model, were compared at three identifications levels: order, family, and genus (species) of macroinvertebrates. Fifty-five reference sites were originally selected for building the model, but this number was reduced to 51 (lowest level), 52 (family), and 53 (order), after the grouping procedures (CLUSTER, MDS, and SIMPER; Primer 5.2.6, Primer-E Ltd., Plymouth, UK). The discriminating variables (stepwise discriminant analysis) stream order, current velocity, pool quality, and substrate quality were common to the genus (species) and family models. Substrate quality was the only discriminating variable of the order model. The model performances, based on their ability to correctly predict reference site membership (complete MDS with jackknifed cross-validation), ranged from 78% (lowest level) to 81% (family and order levels). Twenty test sites were used to compare site assessments using each of the models. We concluded that the lowest-level model of identification provides the best evaluations of water quality assessment and performed well, that the family-level model reacted similarly and could be a good alternative for bioassessment programmes, and that a greater effort toward improving our knowledge of aquatic macroinvertebrates in Portugal is recommended as species are important in assessing water quality.