Randomized Intervention Analysis and the Interpretation of Whole-Ecosystem Experiments

Randomized intervention analysis (RIA) is used to detect changes in a manipulated ecosystem relative to an undisturbed reference system. It requires paired time series of data from both ecosystems before and after manipulation. RIA is not affected by non-normal errors in data. Monte Carlo simulation indicated that, even when serial autocorrelation was substantial, the true P value (i.e., from nonautocorrelated data) was <.05 when the P value from autocorrelated data was <.01. We applied RIA to data from 12 lakes (3 manipulated and 9 reference ecosystems) over 3 yr. RIA consistently indicated changes after major manipulations and only rarely indicated changes in ecosystems that were not manipulated. Less than 3% of the data sets we analyzed had equivocal results because of serial autocorrelation. RIA appears to be a reliable method for determining whether a nonrandom change has occurred in a manipulated ecosystem. Ecological arguments must be combined with statistical evidence to determine whether the changes demonstrated by RIA can be attributed to a specific ecosystem manipulation.