I introduce a new class of non-ignorable non-monotone missing data models. These models are useful for investigating the sensitivity of one's estimates to untestable assumptions about the missing data process. I use the new models to analyse data from a case-control study of the effect of radiation on breast cancer.