Causal modeling with multivariate species data

Recent advances in causal modeling have made it possible to build and test structural equation models without any restriction on the functional forms or error distributions of the structural equations. We propose here a method for building and testing causal models that uses ordination axes arising from multivariate species data. This is demonstrated through the analysis of macrobenthic species abundance data observed at multiple times before and after the 1978 Amoco Cadiz oil spill (Dauvin, 1982). The available data consist of 21 quarterly observations on 257 species during the period 1977-1982. A causal model of the impact and subsequent recovery was built and tested using distance-based redundancy analysis (dbRDA). In addition, to predict the time required for recovery of the community, nonlinear models were fitted to the first two PCO axes, and the fitted nonlinear models were used to generate predictions for 20 years beyond the last observation in the data set. These predictions were found to compare favorably with the results from longer term studies carried out by Dauvin (1998). The methods described here are sufficiently well established to be used in ecological research, and will allow ecologists to move towards plausible causal models and generate stronger inferences from observational multivariate community data than has been achieved to date.