24 Correlated metrics generally yield a multimetric index with inferior performance

Monday, May 18, 2009: 2:45 PM
Imperial Ballroom
John Van Sickle , NHEERL/ Western Ecology Division, U.S. Environmental Protection Agency, Corvallis, OR
Most developers of multimetric indices (MMI’s) believe that highly correlated metrics carry redundant information and hence should not be included in the same MMI. To seek evidence for (or against) this belief, I created 1000 comparable MMI’s from 1000 random subsets of 55 candidate macroinvertebrate metrics for assemblages sampled at 152 independently-determined reference sites and 47 impaired sites, during the 2000-2004 EPA/EMAP survey of Western USA streams and rivers.  Metric redundancy of a subset was measured by the average correlation among all metric pairs in the subset. An MMI was created for each subset, and its performance was measured by the percentage of impaired sites with MMI scores significantly below the distribution of reference-site scores. Across the 1000 metric subsets, MMI performance declined steadily with increasing metric redundancy. For example, at any fixed level of metric responsiveness, a 0.10 increase in average metric correlation led to a predicted decrease of 9.9 (+/- 0.3 CI) points in the percentage of impaired sites that were declared impaired by MMI’s.  Reduced MMI performance for metric subsets having higher redundancy was also seen for 7 other MMI development data sets. Results suggest improved strategies for selecting MMI metrics.
See more of: Bioassessment I
See more of: Contributed Sessions