Abstract:
|
Statistical models approximate natural phenomena by summarizing signals in data and, beyond non-stochastic mathematical models, they also attempt to quantify identified sources of uncertainty using probability. Competing hypotheses about a system can be reflected through a set of carefully specified statistical models, representing model uncertainty. Conditioning inference on one model is often described as naïve because model uncertainty is left “unquantified.” Model averaging (MA) is a popular technique accounting for some model uncertainty, but can also lead to unintended complications in interpretations, including oversold faith in the uncertainty quantified. For example, in many ecological applications, interest lies in explaining relationships between species indicators and covariates, and complicated interpretations resulting from MA can affect the utility of MA results for management decisions. We demonstrate cases where the “protection against naïve inference” may not be realized, and encourage more critical thought around what sources of uncertainty are actually quantified through MA, and multimodel inference in general.
|