The intention-to-treat approach in randomized controlled trials: Are authors saying what they do and doing what they say?

Background Intention-to-treat (ITT) is an approach to the analysis of randomized controlled trials (RCT) in which patients are analyzed as randomized regardless of the treatment actually received. Purpose To ascertain the proportion of RCT reporting the use of intention-to-treat and the accuracy of that report and to examine the distribution and analysis of missing data for the studies reporting an ITT analysis. Method We conducted a cross-sectional literature review of RCTs reported in 10 medical journals in 2002. All articles were assessed using a standardized form. Two evaluators independently reviewed a 10% sample of articles to assess reliability. Subsequently, one evaluator reviewed the remaining articles. The proportion of articles reporting the use of ITT was calculated. Among these, the proportion of articles that 'analyzed patients as randomized' and the proportion and analysis of missing data were evaluated using standardized definitions. Results Of the 403 articles, 249 (62%) reported the use of ITT. Among these, available patients were clearly analyzed as randomized in 192 (77%). Authors used a modified ITT in 23 (9%); clearly violated a major component of ITT in 17 (7%), and the approach used was unclear in 17 (7%). More than 60% of articles had missing data in their primary analysis. Few articles reported a strategy for missing data. The main reason for missing data was loss to follow-up. Limitations A single evaluator evaluated most articles, but the high concordance obtained during the inter-rater evaluation suggests that the assessments were consistent. In addition, the small spectrum of journals limits generalizability. Finally, there could be a difference between what was reported and what was performed. Conclusions This study emphasizes that authors use the label 'intention-to-treat' quite differently. The most common use refers to the analysis of all available subjects as randomized regardless of the missing data aspect.