The Statistical Analysis of Behavioural Latency Measures

This article concerns two important problems with the statistical analysis of behavioural latency measures: they typically have severely skewed distributions, and are often censored (truncated). These problems, however, were not generally recognised by animal behaviour researchers: most people either allot an arbitrary score to all censored values or simply ignore them. Yet, such treatments could easily lead to dubious conclusions because of reduction of power and spuriously significant p-values. Thus, one should always use specially devised survival analysis methods whenever the study involves the measurement of censored latencies. The present article provides a short catalogue of some appropriate references, concentrating on the methods which are not “standard” for the common biomedical applications of survival analysis, but may be crucial in many behavioural studies. The statistical analysis of uncensored latencies is also discussed, with a particular attention to the analysis of variance.