Despite its long history (Luce, 1986) the study of sequential effects has mostly been confined to simple binary tasks such as two-alternative forced choice tasks (2AFC). Here we present experimental results from a choice task with three rather than two alternatives (3AFC) as well as a novel model that can explain them. We find that humans change the statistics they use to analyse a sequence depending on the task constraints, relying on first-order transition probabilities in a 2AFC but event relative frequencies (i.e., zeroth-order transition probabilities) in a 3AFC.