Multi-level, random-effects pattern-mixture (MLREPM) models [17] were used to evaluate the age-related BP trajectories while accounting for non-ignorable dropouts/deaths. This approach has been described in more detail in a previous paper [18]. In brief, the variable defining three groups on the basis of the pattern of their missing data (completers, lost because of drop out, lost because of death) was included as a model covariate, and the overall estimates of the trend parameters were obtained by means of weighted averaging (mixing) the pattern-specific estimates. The weights were the estimated proportion of women in the three missing-pattern groups within each
WZ4002 × drinking × BMI group, and the standard error of the averaged overall estimates were obtained using the delta method. In a hierarchical (multi-level) structure, AHD use constituted the third level (treatment groups, subjects within treatment group and BP measurements within subjects) and, in order to avoid the problem of over-parameterisation, we addressed heterogeneity in terms of random slope parameter variations across the four treatment groups. Age at baseline was included in the models to account for birth year effects, and the individual contributions to the age-related trends were spread over a
neutron single age axis ranging from 60 to 90 years. As the inclusion of quadratic age terms did not significantly improve the likelihood of the models, only linear terms were fitted. Examination of the residual plots did not show any particular trend in their distribution, thus indicating the good fit of the models to the observed data.