Statistical approaches to trial durations in episodic affective illness

In light of the high variability in illness characteristics and patterns among patients with bipolar illness, parallel group designs present severe methodologic difficulties. Crossover, off-on-off-on (B-A-B-A), and other individualized designs may be a useful substitute, but no consensus exists about how to estimate the individual trial durations required in these instances. Several methods for determining optimum trial lengths in crossover designs are presented, illustrated, and discussed. These include: chi-square (χ²) for the expected versus observed number of either episodes or days well; exceeding two standard deviations for average duration of episodes or euthymic intervals; or the Sequential Probability Ratio Test (SPRT), which detects when mean values differ from prior statistical expectations. Each method was applied to three demonstration cases using data from actual clinical trials of three patients with different patterns of recurrent affective illness. Each method detected changes in illness severity, although different tests appeared to be sensitive to differing cycle patterns in the patients illustrated. We suggest that these types of analyses and others can be used as indicator statistics to augment global impressions and clinical judgment, and to assist in determining individualized trial durations, both in formal clinical trials and in clinical treatment settings. Once individual responsivity is confirmed with an appropriate interplay of trial design and statistical analysis, the percentage response in a given population can then be compared to other agents or in other populations. Moreover, meta-analytic techniques based on addition of z scores from individuals' effect sizes can then be used to assess overall significance of a drug effect in a given population or subpopulation. The need for further development of appropriate and alternate study designs and analysis methods for bipolar illness is highlighted. Approaches to estimating required trial durations in individuals with different cycle frequencies in crossover and B- A-B-A designs constitute one element of that exploration.