When did it really start or stop: the impact of censored observations on the analysis of duration.

Behaviour is often described in terms of bout lengths. Because of censoring, some of these bout lengths may only be observed partially. For instance, when observation is finished after a fixed period the end moment of the last bout remains unknown. The only available information on such a bout length is that it exceeds a certain value. This value is the censored observed bout length. Censored data are quite common in ethology, but the problem is often not recognized. Therefore, the well established statistical methods that account for censoring are rarely used in ethology. We report on the consequences of using standard methods instead of methods adjusted to account for censoring. We demonstrate that the usual methods of dealing with censored observations, such as treating them as uncensored observations or omitting them altogether, leads more often to erroneous conclusions. When an unadjusted test is used for testing the equality of two censored samples of bout lengths, the probability of rejecting the null hypothesis when the samples are different is much lower than when an adjusted test is used. Moreover, especially when censoring patterns differ between samples, the probability of wrongly rejecting the null hypothesis may be increased.