4.3 Metric Redundancy
Redundant metrics may respond in the same manner to stress and if used together
result in overly complex indices that include metrics that do little to increase (and
may decrease) accuracy. The problem of redundancy has long been recognized,
although the best approach for minimizing it remains unclear. Some investigators
prefer to focus on ecological redundancy by including metrics from different
ecological categories [122, 128], while others focus on reducing statistical redundancy by evaluating pairwise correlations among metrics and choosing only one
metric within each pair that is correlated [124]. Combinations of these approaches
may be employed, which consider both ecological and statistical redundancy [138].
Correlations among metrics generally reduce MMI precision and accuracy, but
these characteristics appear to be most related to the mean pairwise correlation
rather than the maximum correlation among metrics in an index [139]. To best
Fig. 5 A graphical comparison of the accuracy and precision of two O/E indices (after methods in
[72]). Figure panels depict frequency distributions of O/E values for two indices at reference sites.
The data were simulated for example purposes only. The O/E index in the top panel is relatively
accurate (mean ¼ 1.01) and precise (SD ¼ 0.08). In contrast, the index depicted in the bottom panel
is less accurate (mean ¼ 0.70) and less precise (SD ¼ 0.16, also note the greater spread of the
distribution). This approach can be modified to assess the value distribution of any metric or index
at reference sites (e.g., [125])
252
A.L. Garey and L.A. Smock
Redundant metrics may respond in the same manner to stress and if used together
result in overly complex indices that include metrics that do little to increase (and
may decrease) accuracy. The problem of redundancy has long been recognized,
although the best approach for minimizing it remains unclear. Some investigators
prefer to focus on ecological redundancy by including metrics from different
ecological categories [122, 128], while others focus on reducing statistical redundancy by evaluating pairwise correlations among metrics and choosing only one
metric within each pair that is correlated [124]. Combinations of these approaches
may be employed, which consider both ecological and statistical redundancy [138].
Correlations among metrics generally reduce MMI precision and accuracy, but
these characteristics appear to be most related to the mean pairwise correlation
rather than the maximum correlation among metrics in an index [139]. To best
Fig. 5 A graphical comparison of the accuracy and precision of two O/E indices (after methods in
[72]). Figure panels depict frequency distributions of O/E values for two indices at reference sites.
The data were simulated for example purposes only. The O/E index in the top panel is relatively
accurate (mean ¼ 1.01) and precise (SD ¼ 0.08). In contrast, the index depicted in the bottom panel
is less accurate (mean ¼ 0.70) and less precise (SD ¼ 0.16, also note the greater spread of the
distribution). This approach can be modified to assess the value distribution of any metric or index
at reference sites (e.g., [125])
252
A.L. Garey and L.A. Smock