Modeling rates of trial attrition: an analysis of individual participant data from 90 randomized controlled trials of...

Trial attrition threatens the validity of randomized controlled trials (hereafter trials) and has implications for trial design, conduct, and analysis. Few studies have examined how attrition rates change over follow-up or the types of attrition reported. Therefore, we estimated attrition rates using individual participant data for a range of conditions.

We obtained the number of days participants spent in trials, completion status, and reported reasons for noncompletion. For consistency with the clinicaltrials.gov reporting guidelines, we categorized attrition into adverse event, lack of efficacy, lost to follow-up, principal investigator/sponsor decision, protocol violation, voluntary withdrawal, and other. For each trial, we estimated the cumulative incidence of attrition and fitted six parametric time-to-event models (exponential, generalized gamma, Gompertz, log-logistic, log-normal, and Weibull). Goodness of fit was evaluated graphically and using the Akaike Information Criterion (AIC). Attrition rates were obtained for each trial as instantaneous risk (ie, hazard rates) from the best-fitting model.

We included 90 trials (86,107 participants): type 2 diabetes (45.6%), chronic obstructive pulmonary disease (22.2%), and eight other conditions (32.2%). Attrition occurred for 14,572 (16.9%) participants, ranging from 3.4% to 43.7% among trials. Adverse event (43.5%) and voluntary withdrawal (24.1%) were the commonest categories of attrition. Gompertz and log-normal time-to-event models were the most frequent best-fitting models. Hazard rates typically peaked near the beginning of trials and decreased thereafter.

Attrition rates were generally highest near the beginning of trials, decreased thereafter, and were well-described by Gompertz and log-normal time-to-event models. These findings can inform the design, conduct, and analysis of clinical trials.