Treatment, time, treatment-time and baseline effects (if recorded) can all be fitted as fixed. Estimates of the overall treatment effect will differ depending on whether treatment-time effects are included in the model. When they are, treatment effects are calculated as the average of the treatment estimates obtained from each time point. When they are omitted, a weighted average of estimates from each time point is obtained. The weights are related to the variances of estimates at each time point, which are in turn related to the number of observations at each time point. The decision as to which estimate to use should rest with the desired interpretation and whether the treatment-time effects are significant. The unweighted estimate may be more appealing if there are missing data at later time points, so that bias towards earlier time points is reduced. On the other hand, if a treatment effect appears relatively constant over time, then the weighted estimate has the advantage of being less influenced by potentially inaccurate estimates at time points with fewer observations. If treatment-time effects are significant, it would be wise to present additional treatment estimates at each time point. Less importance should then be attached to the overall treatment estimate.
The above discussion has assumed that we are interested in comparing treatments, though of course the applications can be much wider. In studies where patients are grouped by a variable other than treatment (e.g. an epidemiological study to compare patients suffering from different disease types), the comparator variable should simply be substituted for the treatment effects above.
Was this article helpful?