The degree of non-compliance with a methadone programme is such that a
zero is often recorded as the response variable. Such outcome variabl
es cannot be transformed to normality in order to investigate a regres
sion relationship between the response variable and various risk or tr
eatment variables. The presence of a random effect as well in this reg
ression causes added difficulty in the analysis. We show here that use
of threshold models can provide a satisfactory solution to both these
problems.