Nonstationary environmental and community relationships in the North Pacific Ocean

Common approaches for summarizing multivariate environmental or community data assume that relationships among variables are stationary over time, and this assumption is often not tested. Here we test the hypothesis that relationships among environmental and community time series are nonstationary in the Gulf of Alaska ecosystem (North Pacific Ocean) over multidecadal time scales. Dynamic factor analysis (DFA) is applied to environmental and community data from before and after 1988/1989, corresponding to the timing of an abrupt decline in temporal variance of the Aleutian Low atmospheric pattern, a leading driver of Gulf of Alaska climate. Results show that covariance among local atmosphere and ocean environmental variables weakened simultaneous to the decline in Aleutian Low variance. At the same time, community‐wide responses of 14 fish and crustacean populations to physical forcing weakened, as indicated by nonstationary environment–biology regression coefficients. In line with theoretical predictions, this loss of a shared response to environmental variability was accompanied by weakening community covariance. Individual populations also showed nonstationary relationships with shared trends of community variability. We conclude that assumptions of fixed environmental and community relationships are likely to produce mistaken inference in this ecosystem. Similar concerns may apply in other ecosystems subject to changing climate patterns. A process is stationary if it follows a probability density that is constant over time and nonstationary if the probability density is time dependent (Kolmogorov 1991). Ecologists increasingly expect environmental variables such as temperature or precipitation to be nonstationary over ecologically relevant time scales because of climate change, but nonstationarity may also exist in relationships among environmental variables (e.g., the covariance between temperature and precipitation may change over time). These sorts of nonstationary relationships are rarely considered in ecological studies, and relationships among environmental variables are typically modeled with fixed probability densities (Wolkovich et al. 2014). Ecological understanding may be highly vulnerable to violations of this assumption for reasons having to do with correlation among predictor variables (collinearity). The physical variables operating in an ecosystem are typically collinear, so that the effect of individual variables is difficult or impossible to estimate independently (Zuur et al. 2010). But when patterns of collinearity are