specialized for the analysis of assemblage data. McCune and Grace [52] and
Legendre and Legendre [53] present additional details on most of the analytical
methods presented here. Software for conducting most of the techniques is available free of cost in the R statistical programming language [54].
The objective of classification is to maximize assemblage similarity within classes, while maintaining sufficient replication to allow for statistical comparisons of
test sites with the reference classes. Many measures of assemblage similarity exist
[52, 53]. The Bray–Curtis coefficient [55] for abundance data and the Sorenson
coefficient [56, 57], the equivalent of Bray–Curtis for presence-or-absence data, are
most commonly used. Both are well suited to the numerical structures of assemblage
datasets. Similarity may also be calculated using bioassessment metric values in place
of taxonomic data (a sound but under-used technique, [58]).
Similarity is summarized as the mean within-class similarity (W, the mean of all
pairwise similarities of sites within classes) and the mean between-class similarity (B,
the mean of all pairwise similarities of sites not in the same class). The precision of a
classification is described by the relationship of W to B, referred to as the classification strength [58]. High classification strength is indicated by a large positive
difference or large ratio of within- to among-class similarity (i.e., high W–B or
W/B). Predictions regarding assemblage conditions are most reliable when classification strength is high. Multivariate techniques such as MEANSIM [58], analysis of
similarity [59], multiresponse permutation procedure [60], and nonparametric, multivariate analysis of variance [61] are used to test the hypothesis that classification
strength is higher than expected by chance, providing an indication of whether the
classification improves the reliability of predictions regarding reference conditions.
2.3.1 Typological Site Classification
Typological site classifications are based on a priori judgments regarding the
site conditions that best group reference sites with similar assemblages. Early
typological approaches focused on coarse-scale, map-based classifications (e.g.,
ecoregions) [62, 63]; however, typological classifications that do not account for
the effects of local-scale variables typically exhibit much lower classification
strength than biotic classifications [43, 64, 65]. Typological classifications are a
convenient and useful tool that should be at least considered as an initial step
toward site classification [65, 66]. Like all classifications, the effectiveness of a
priori-defined typologies should be assessed by a posteriori, quantitative evaluations of the assemblages of interest [66]. For example, investigators in Virginia
(USA) observed a striking difference in stream macroinvertebrate assemblage
structure between low-gradient coastal plain sites and upland piedmont and mountain sites, requiring the use of separate bioassessment indices for coastal and
non-coastal sites (Fig. 1) [67, 68]. Because assemblages are affected by both
regional- and local-scale environmental factors, typological classifications that
consider smaller spatial-scale variables as well as large-scale zones may provide
comparable, or greater, classification strength than biotic classifications [64, 69].
Principles for the Development of Contemporary Bioassessment Indices for. . .
241
Legendre and Legendre [53] present additional details on most of the analytical
methods presented here. Software for conducting most of the techniques is available free of cost in the R statistical programming language [54].
The objective of classification is to maximize assemblage similarity within classes, while maintaining sufficient replication to allow for statistical comparisons of
test sites with the reference classes. Many measures of assemblage similarity exist
[52, 53]. The Bray–Curtis coefficient [55] for abundance data and the Sorenson
coefficient [56, 57], the equivalent of Bray–Curtis for presence-or-absence data, are
most commonly used. Both are well suited to the numerical structures of assemblage
datasets. Similarity may also be calculated using bioassessment metric values in place
of taxonomic data (a sound but under-used technique, [58]).
Similarity is summarized as the mean within-class similarity (W, the mean of all
pairwise similarities of sites within classes) and the mean between-class similarity (B,
the mean of all pairwise similarities of sites not in the same class). The precision of a
classification is described by the relationship of W to B, referred to as the classification strength [58]. High classification strength is indicated by a large positive
difference or large ratio of within- to among-class similarity (i.e., high W–B or
W/B). Predictions regarding assemblage conditions are most reliable when classification strength is high. Multivariate techniques such as MEANSIM [58], analysis of
similarity [59], multiresponse permutation procedure [60], and nonparametric, multivariate analysis of variance [61] are used to test the hypothesis that classification
strength is higher than expected by chance, providing an indication of whether the
classification improves the reliability of predictions regarding reference conditions.
2.3.1 Typological Site Classification
Typological site classifications are based on a priori judgments regarding the
site conditions that best group reference sites with similar assemblages. Early
typological approaches focused on coarse-scale, map-based classifications (e.g.,
ecoregions) [62, 63]; however, typological classifications that do not account for
the effects of local-scale variables typically exhibit much lower classification
strength than biotic classifications [43, 64, 65]. Typological classifications are a
convenient and useful tool that should be at least considered as an initial step
toward site classification [65, 66]. Like all classifications, the effectiveness of a
priori-defined typologies should be assessed by a posteriori, quantitative evaluations of the assemblages of interest [66]. For example, investigators in Virginia
(USA) observed a striking difference in stream macroinvertebrate assemblage
structure between low-gradient coastal plain sites and upland piedmont and mountain sites, requiring the use of separate bioassessment indices for coastal and
non-coastal sites (Fig. 1) [67, 68]. Because assemblages are affected by both
regional- and local-scale environmental factors, typological classifications that
consider smaller spatial-scale variables as well as large-scale zones may provide
comparable, or greater, classification strength than biotic classifications [64, 69].
Principles for the Development of Contemporary Bioassessment Indices for. . .
241