The output-only damage detection in a composite cantilever beam is investigated under varying temperature conditions via a recently introduced novel vibration-based statistical time series method. The method is based on global, Functionally Pooled, vector autoregressive models and statistical decision making. Two versions are investigated: a modal parameter based version using the natural frequencies as characteristic quantity, and a model parameter based version using the model parameters as characteristic quantity. The method is tested via hundreds of laboratory experiments with damages consisting of progressive saw-cuts on the beam and under temperatures in the -20 to 20°C range. The results of the study demonstrate the method's high effectiveness and potential for practical use. The modal parameter based version achieves excellent results with very few false alarms or missed damage cases, whilethe model parameter based version appears somewhat less effective for the lowest level of damage due to temperature dependence of the employed test statistic.