On the problems of using linear models in ecological manipulation experiments: lessons learned from a climate experiment

Abstract

Manipulation experiments are often used to investigate ecological and environmental causal relationships and to understand and forecast impacts of anthropogenic pressures on ecosystem functioning. Such manipulation experiments often use factorial designs, and the data are analyzed using factorial linear models. Factorial designs build on the fundamental assumption that the treatment factors are independent and orthogonal. This assumption is, however, often violated because of variation within and in particular covariation between the performed experimental manipulations. For example, manipulation of temperature and precipitation in factorial setups has been widely applied in climate experiments, but manipulating soil temperature will likely have a strong impact on soil water content. Such dependency among environmental state variables will violate the assumed orthogonality in a factorial linear model and may lead to erroneous conclusions. Here, we demonstrate the importance of the assumption of orthogonality using simulated ecological responses that act on observed soil state variables from a large climate experiment with an apparent orthogonal design. More specifically, we explore the problematic consequences of analyzing ecological treatments as categorical variables in a linear model. Suitable alternative methods for the statistical analysis of manipulated ecological experiments are suggested. The key recommendation is to use the observed effects of the manipulations on the state variables directly in the analysis instead of the categories of treatments. For example, if soil water content and temperature are manipulated, then it is essential to measure the water content and temperature in the soil of all the manipulated plots.