In order to evaluate authentication methods for keystroke dynamics, it is necessary to create new datasets. We present in this paper an analysis study of the factors, involved during the acquisition process, which affect the performance of keystroke-based authentication systems. More generally speaking, we are looking for the optimal keystroke data acquisition scenario which may deeply affect the performance of such systems. Results show that (1) it is better to choose passwords based on short, simple and known words; and (2) during the acquisition process of a new keystroke dynamics dataset, there is no significant preference for displaying or telling to a user the text to type.