This JAMA Guide to Statistics and Methods describes various methods to handle trial participant nonadherence to study interventions.
Randomized clinical trials allocate individuals to different treatments, or, more generally, to interventions and comparators, to determine whether one is better than another. However, after having been randomized to a given intervention, some study participants may not adhere to the assigned protocol. Treatment nonadherence may result from study participants crossing over to the other randomized treatments, taking nontrial medications, or not adhering to the study protocol. All of these situations introduce postrandomization problems that may have to be accounted for in the analysis of data from the trial.1
By convention, clinical trials are evaluated with analyses that are based on an intention-to-treat (ITT) principle; this means comparing the outcomes of study participants according to the treatment to which they are assigned, not the treatment they ultimately receive. The ITT principle acknowledges that treatment outcomes depend not only on the effect of the treatment, but also on whether participants or clinicians adhere to the treatment protocol.2 However, when treatment nonadherence occurs, an analysis based on the ITT principle will not capture the effect that would have been observed had people adhered to the new treatment.
WHY USE METHODS TO ADJUST FOR TREATMENT NONADHERENCE?
To adjust for treatment nonadherence, statistical adjustments of clinical trial results are necessary. Some methods are simple and others more complex, but because simple methods can bias results, more sophisticated modeling is required.3 For example, treatment nonadherence occurred in the Women's Health Initiative Randomized Controlled Dietary Modification Trial,4 which assessed the effects of a low-fat dietary pattern on breast cancer incidence in postmenopausal women without breast cancer. The trial included 48 835 postmenopausal women, aged 50 to 79 years, and randomized 19 541 to the dietary modification intervention group and 29 294 to the comparison group, who were not asked to make dietary changes. The trial showed that over an average of 8.1 years, women who were randomized to the intervention group consumed less dietary fat and also had a lower incidence of breast cancer compared with women randomized to the control group (0.42% yearly vs 0.45% yearly; hazard ratio, 0.91 [95% CI, 0.83-1.01]) not adjusted for adherence. Analyses estimated a lower hazard ratio had all women adhered to their assigned intervention.
DESCRIPTION OF ADJUSTMENT METHODS
When analyzing trial data, some researchers may deal with nonadherent participants by ignoring some or all data from these individuals in a process called censoring. One method for doing this is per-protocol analyses, in which participants contribute data to the analysis until they stop following the treatment protocol. However, study participants who do not adhere to treatment may have fundamentally different risks of developing the study outcome than those who are ...