Introduction: The 'gold-standard' for analysing clinical trial data individually and in reviews is the intention-to-treat approach whereby patients are analysed in the group to which they were allocated, irrespective of the treatment they actually received. It can be argued that although this reduces the risk of certain biases, it does not represent the real situation in which patients do not follow a treatment for the time initially prescribed. We will use the example of fluoxetine treatment for major depression to illustrate this problem.

Objective: To examine the effect of the different analytical approaches for data from a clinical trial on the results of a meta-analysis.

Methods: Using the Eli Lilly Fluoxetine database, we selected trials in which fluoxetine was compared with tricyclic antidepressants (TCA) performed in the United States (US) and those performed outside the US (non-US) that satisfied our predefined inclusion criteria. The primary outcome was defined as a binary variable; partial remission, i.e. at least 50% reduction compared with the baseline score on the HDRS-17 instrument. Three types of analyses were performed using 'a last observation carried forward' technique; 1) intention-to-treat; 2) patients that had completed at least 4 weeks therapy (efficacy analysis); and 3) patients that had at least one post-baseline visit (endpoint analysis). The logarithm of the odds ratio method was used.

Results:

Type of analysis OR [95% CI] p-assoc p-homo

US vs TCA (11 trials; n = 1370)
Intention to treat 1.18 [0.94; 1.48] 0.149 0.886
Efficacy 0.93 [0.71; 1.22] 0.599 0.593
Endpoint 1.10 [0.88; 1.37] 0.402 0.377

Non-US vs TCA (13 trials; n = 643)
Intention to treat 0.80 [0.58; 1.12] 0.196 0.576
Efficacy 0.62 [0.42; 0.92] 0.019 0.608
Endpoint 0.73 [0.53; 1.02] 0.063 0.303

Discussion: The efficacy and endpoint analyses gave similar results, whereas the efficacy analysis resulted in a lower value for the point estimate of the treatment effect.