Statistical techniques for the analysis of change in longitudinal studies of the menopause

Critical review of statistical techniques used to analyze data from longitudinal studies.

Literature search and classification of statistical methods and their utilization evaluated against known underlying assumptions.

One hundred and twenty-three papers found: 1. cross-sectional reduction of data (56%); 2. multifactorial techniques (26%); 3. repeated measurement analysis of variance (14%); 4. other (time series and structural equation modelling) (4%).

Cross-sectional reduction violates underlying statistical assumptions. A simple and powerful technique is to mean values prior to and following an event such as the final menstrual period. To allow for the influence of multiple factors, linear regression is preferred to logistic regression where continuous data are available. For more information about evolution in time, more complex techniques are needed. A suitable technique is repeated measurement multivariate analysis of variance using a number of contrasts to estimate various effects. Split plot or randomized block designs cannot be recommended as they violate compound symmetry assumptions. For series involving more than 100 observations for each subject, time series and spectral analysis techniques should be considered. Structural equation modelling is recommended for examination in detail of a range of factors that may influence the studied end-point, the presence of feedback and of latent or non-measurable variables.