A full likelihood method is proposed to analyse continuous longitudina
l data with non-ignorable (informative) missing values and non-monoton
e patterns. The problem arose in a breast cancer clinical trial where
repeated assessments of quality of life were collected: patients rated
their coping ability during and after treatment. We allow the missing
ness probabilities to depend on unobserved responses, and we use a mul
tivariate normal model for the outcomes. A first-order Markov dependen
ce structure for the responses is a natural choice and facilitates the
construction of the likelihood; estimates are obtained via the Nelder
-Mead simplex algorithm. Computations are difficult and become intract
able with more than three or four assessments. Applying the method to
the quality-of-life data results in easily interpretable estimates, co
nfirms the suspicion that the data are non-ignorably missing and highl
ights the likely bias of standard methods. Although treatment comparis
ons are not affected here. the methods are useful for obtaining unbias
ed means and estimating trends over time.