For some domains, panel surveys collect information about the period between interviews. Such data are typically affected by seam effects: transition rates from one month (or week) to the next are typically far higher if the months were covered in two different interviews, than if they were covered in the same interview. The causes of seam effects are not well understood. As a result, data collection methods designed to reduce the problem appear to work for some types of items, but not for others. This paper presents a theoretical framework of the causes of seam effects that unifies existing theories and evidence. The predictions from the framework are tested using data from the British Household Panel Survey and find support.