Correlated Metrics Yield Multimetric Indices with Inferior Performance

Multimetric indices (MMIs) are widely used to assess the ecological health of freshwater ecosystems. An MMI is a sum of several standardized numeric variables or metrics, each representing a different attribute of a biological assemblage. Many researchers believe that highly correlated metrics should not be included in the same MMI because they convey redundant information. To seek evidence for or against this belief, I compared the performance of 1,000 MMIs created for each of eight existing data sets by randomly resampling metrics from sets of previously identified candidates. An MMI's performance was measured by its precision and its ability to detect differences between assemblages sampled in independently assessed reference and impacted streams. Across the 1,000 MMIs, precision decreased with increasing mean correlation magnitude for seven of the eight data sets. For seven of the data sets, multiple linear regressions fitted to each set of 1,000 MMIs predicted a decrease in MMI detection ability as the mean correlation magnitude between metrics increased, after adjusting for the average responsiveness of individual metrics to the difference between reference and impacted conditions. However, similar regressions showed that the size of the largest correlation between any two metrics in an MMI had little or no effect on its detection ability. Thus, minimizing the mean of metric correlations is more effective than the widespread practice of setting an upper correlation limit when optimal MMI performance is desired. Finally, an MMI had originally been built for each data set by selecting one set of individually best metrics. In 23 of 24 assessments, 5–100% of randomly selected MMIs outperformed the original MMIs. Because individually best metrics rarely yielded a best‐performing summed index, I recommend assessing multiple candidate MMIs, rather than just multiple candidate metrics, when developing a new MMI.