CAN ARMA MODELS BE USED RELIABLY IN ECOLO GY

Several characteristics in ecological time series (limited size, non-n ormal distribution, missing data) are usually put forward to justify t hat statistical methods as ARMA modeling (AutoRegressive Moving Averag e) are not used unlike in economics or in hydrology. However, none stu dy of sensitivity has been published on applying ARMA method to ecolog ical data. We tested its robustness by simulation and considered four factors: the nature of the generating process (AR, MA or ARMA), the le ngth of the time series, the error model and the criterion used to sel ect a model. We studied the probability of correctly identifying the g enerating process in the simulation for all combinations of the four f actors. This probability is high (about 90%) only if time series lengt h exceeds 50 points, if the generating process is simple (pure AR or M A) with a parameter having a high modulus (about 0.8) and if SEC is th e selecting criterion; a mixed process (ARMA) was correctly identified at best in 10% to 20% of the simulations. Among the three criterions tested, AIC(C) leads to the best selection of a model, especially with low time series length. All models are not sensitive to asymmetry of the error model in our simulations. The results are discussed accordin g to the goals of the ecologists when analysing time series. Limitatio ns of an automatic identification of a model are underlined.