Intention-to-treat analysis: Who is in? Who is out? - Brief Report

Journal of Family Practice, Nov, 2002 by Robin L. Kruse, Brian S. Alper, Carin Reust, James J. Stevermer, Scott Shannon, Randy H. Williams

We could determine the proportion of randomized participants in the primary analysis in 93 studies; it varied from 49% to 100%, with a median of 98.7%. Ten of the 93 studies (11%) excluded more than 20% of participants from the primary analysis. In 16 of the 93 studies (17%), a non-ITT analysis (eg, "per protocol") was presented as the primary analysis. In these studies, an average of 80.1% (median, 82.4%; range, 49.0% to 92.4%) of randomized patients were included in the primary analysis.

Fifty-six studies included a definition of the fit population, primarily within the methods (38) and results (18) sections. Of the 42 studies where all randomized subjects were analyzed, 20 included definitions of ITT. Diagrams showing the flow of participants through each trial were present in 41 of 100 articles, including 1 on a journal's web site. An additional 8 articles had diagrams that showed patient flow without giving the number of patients. Presence of a flow diagram was not related to whether or not all randomized subjects were included in the ITT analysis (36% vs 45% respectively, P = .37). Of the 31 articles from journals that participate in CONSORT, 29 included flow diagrams, compared with 12 of the 69 articles from journals that do not participate in CONSORT (P < .0001).

DISCUSSION

The hallmark of ITT analysis is that all randomized subjects are analyzed. (7) In more than half of the articles we examined, this was not the case. Analysis of only certain subgroups of patients is sometimes appropriate, but an explanation should be provided whenever subjects are left out of any analysis. For example, we examined a report of a trial that was stopped based on the results of an interim analysis, thus excluding subjects who were randomized after the interim analysis. (11) This type of exclusion, based on an a priori decision rather than individual characteristics or behavior, is less likely to bias results.

While all the articles in our sample reported analysis by ITT, many authors did not define the term, even when they excluded some randomized subjects from the ITT analysis. In these cases, the reader is left to infer which subjects were excluded based on information given in the text, figures, and tables.

Despite numerous recommendations for detailed reporting of RCT methods, (1-4) many articles were vague and lacked detail. We could not determine which categories of participants were excluded from the ITT analysis in 13 articles. In 8 of the 100 articles we examined, we could not determine how many subjects were randomized or included in the ITT or primary analysis. Four of these 8 articles were in journals that endorsed the CONSORT statement. All were published well after the initial CONSORT statement was released in 1996. (1)

The number of randomized subjects excluded from the ITT analysis was usually small. It is unlikely that excluding up to 1% of subjects had a major effect on the results. In 11% of our sample, however, more than 10% of randomized subjects were excluded. Exclusions of this magnitude have significant potential to alter the findings. When outcome data can't be determined and the outcome is categorical (eg, alive/dead), it can be helpful to produce best-case and worst case scenarios in which patients lost to follow-up are arbitrarily ascribed good or bad outcomes. These extremes delimit the potential effect of the exclusions on results. (12) Similarly, missing continuous outcomes (eg, weight change) can he assigned specific values to determine the potential impact on the results.