Does intention-to-treat analysis answer all questions in long-term mortality trials? Considerations on the basis of the ANZ trial

Aims: Intention-to-treat (ITT) analyses are the gold standard in endpoint a nalyses of randomized clinical trials. Valuable information, however, may b e lost in this approach as the actual time on trial medication is not accou nted for in patients who withdraw early. Since compliance per se can be a p rognostic factor and the actual treatment time is a variable likely to infl uence clinical outcome, this information should be added to an ITT analysis concept. Thus, the aim is to elucidate the influence of actual treatment t ime on the results of ITT analysis using available data from a well-conduct ed, large-scale clinical trial. Methods and results: The ANZ trial, a carve dilol study in heart failure, is characterized by a considerable number of early withdrawals in the carvedilol group. Using the ANZ trial database, th e percentage of withdrawals was calculated as well as the number of days on treatment. Fourteen out of 21 deaths (67%) in the carvedilol group occurre d after discontinuation of carvedilol. At the time of death, 14 out of 29 ( 48%) patients in the placebo group had withdrawn from treatment. Mean time of patients on medication who withdrew and died later were 301 days for pla cebo, but only 204 days for carvedilol. We performed a conventional ITT ana lysis, and an ITT-based Cox-proportional hazards model (modified ITT) in wh ich the actual time on trial medication was entered into the model as an ex planatory variable. The risk ratio in conventional ITT analysis was 0.71 (9 5% confidence interval (CI) 0.41 to 1.24, p = 0.228) and 0.50 (95% Cl 0.28- 0.90, p < 0.001) in modified ITT analysis. Conclusions: Incorporation of th e actual treatment time substantially influences the result of this exempla ry ITT analysis. The resulting risk ratio is clearly in accordance with cli nical implications and in close agreement with those of other P-blocker tri als in heart failure. Entering the actual treatment time into a Cox-proport ional hazards model as an explanatory variable is reasonable if not necessa ry from a clinical point of view and thus may strengthen the validity of IT T analysis.