[113]. ANNs have shown substantial improvement over traditional RIVPACS-type
models for predicting the richness of macroinvertebrate and fish assemblages
[97, 114].
Prediction tree approaches such as classification and regression trees account for
the complex effects of both continuous and categorical predictors by recursively
bisecting the dataset into groups that are increasingly similar with respect to the
response variable after each division [52, 53, 115]. Ensemble prediction tree
approaches such as random forests and boosted regression trees combine the results
of hundreds to thousands of trees to reduce prediction error. Random forests have
been used to model assemblage metrics directly [21, 101, 116] and to define
relationships between environmental variables and predefined biotic classes, effectively replacing MDA as used in RIVPACS [12, 101]. Comparisons of random
forests to boosted regression trees, a related ensemble tree method, indicate that the
latter may provide superior performance [108, 117].
In the absence of suitable reference sites, investigators have used whole-set
approaches that employ all sites in the dataset, rather than only reference sites, to
control for the effects of natural environmental variables. Most whole-set
approaches involve the use of regression techniques to model the responses of
metrics to stressors and then to estimate metric values at the point where the model
estimates that no impairment occurs [118, 119]. Because few to no minimally
impaired sites are included in these analyses, they are effectively estimates by
extrapolation of a stressor–response gradient and therefore may be subject to
greater prediction errors than models for which reference sites are available. Such
errors, however, may be unavoidable when test sites cannot be matched with
comparable reference sites. As an alternative whole-set approach, Chessman and
Royal [120] estimated the tolerance limits and preferences of macroinvertebrates to
substrate, temperature, and flow conditions across an extensive dataset of
Australian rivers. These limits were then used to predict the presence of taxa and
derive O/E values at test sites, which exhibited stronger correlations with stressor
gradients than O/E values derived using the AUSRIVAS method.
A case for using the whole-set approach as a replacement for the RCA was
recently presented [121]. Data simulations were conducted to model scenarios in
which natural environmental variables and stressors affected biotic metrics independently and also interactively. Metrics that were model-adjusted using the wholeset approach exhibited more accurate and precise relationships with the simulated
stressor gradient than metrics adjusted using the RCA. The difference in performance was greatest when stressors and natural environmental variables interacted,
as the RCA cannot account for such interactions. While the authors present a
compelling case, additional field-based empirical comparisons of the whole-set
approach to the RCA are needed.
248
A.L. Garey and L.A. Smock
models for predicting the richness of macroinvertebrate and fish assemblages
[97, 114].
Prediction tree approaches such as classification and regression trees account for
the complex effects of both continuous and categorical predictors by recursively
bisecting the dataset into groups that are increasingly similar with respect to the
response variable after each division [52, 53, 115]. Ensemble prediction tree
approaches such as random forests and boosted regression trees combine the results
of hundreds to thousands of trees to reduce prediction error. Random forests have
been used to model assemblage metrics directly [21, 101, 116] and to define
relationships between environmental variables and predefined biotic classes, effectively replacing MDA as used in RIVPACS [12, 101]. Comparisons of random
forests to boosted regression trees, a related ensemble tree method, indicate that the
latter may provide superior performance [108, 117].
In the absence of suitable reference sites, investigators have used whole-set
approaches that employ all sites in the dataset, rather than only reference sites, to
control for the effects of natural environmental variables. Most whole-set
approaches involve the use of regression techniques to model the responses of
metrics to stressors and then to estimate metric values at the point where the model
estimates that no impairment occurs [118, 119]. Because few to no minimally
impaired sites are included in these analyses, they are effectively estimates by
extrapolation of a stressor–response gradient and therefore may be subject to
greater prediction errors than models for which reference sites are available. Such
errors, however, may be unavoidable when test sites cannot be matched with
comparable reference sites. As an alternative whole-set approach, Chessman and
Royal [120] estimated the tolerance limits and preferences of macroinvertebrates to
substrate, temperature, and flow conditions across an extensive dataset of
Australian rivers. These limits were then used to predict the presence of taxa and
derive O/E values at test sites, which exhibited stronger correlations with stressor
gradients than O/E values derived using the AUSRIVAS method.
A case for using the whole-set approach as a replacement for the RCA was
recently presented [121]. Data simulations were conducted to model scenarios in
which natural environmental variables and stressors affected biotic metrics independently and also interactively. Metrics that were model-adjusted using the wholeset approach exhibited more accurate and precise relationships with the simulated
stressor gradient than metrics adjusted using the RCA. The difference in performance was greatest when stressors and natural environmental variables interacted,
as the RCA cannot account for such interactions. While the authors present a
compelling case, additional field-based empirical comparisons of the whole-set
approach to the RCA are needed.
248
A.L. Garey and L.A. Smock