Pharmacoepidemiological studies often use prescription registries to assess patients' drug episodes. The databases usually provide information on the date of the redemption of the prescription as well as on the dispensed amount, and this allows us to define episodes of drug use. However, when patients take less medication than prescribed, apparent gaps between prescriptions occur, and most studies handle this issue by allowing for small gaps when defining continuous drug use. This paper argues that it becomes crucial whether gaps are 'filled' prospectively or retrospectively. In the latter case the inferred exposure status depends on the patient's future dispensing behaviour and this can lead to severe bias in the findings of the study. In this paper we investigate this potential bias in a study of the risk of acute myocardial infarction (AMI) for women using hormone therapy (HT), and we show that the retrospective exposure definition introduces an artificially protective effect of HT.