Estimating effects from randomized trials with discontinuations: the need for intent-to-treat design and G-estimation

Background Randomized trials provide pivotal evidence for evaluation and approval of therapies. Nonetheless, such trials are often plagued by noncompliance, especially in the form of premature discontinuation of treatment. While intent-to-treat (ITT) analysis can provide valid tests of no-effect hypotheses, some trials may make ITT analysis impossible by ceasing follow-up when patients go off assigned treatment. Furthermore, estimates based on ITT, on-treatment, or per-protocol comparisons can seriously understate harm or benefit. Purpose To show how g-estimation based on randomization status is a natural generalization of ITT null testing to estimating efficacy from trials with important discontinuation or noncompliance. Methods We contrast with an analysis of the effect of a tiotropium inhaler on the occurrence of chronic obstructive pulmonary disease (COPD) events in a six-month double-blind placebo-controlled trial of 1829 patients with good but imperfect compliance. Results The covariate-adjusted point estimates, 95% confidence limits (CL), and null P-values comparing expected COPD event times in placebo versus tiotropium patients were: ITT, 1.21, CL=1.02, 1.43, P=0.027; on-treatment, 1.27, CL= 1.06, 1.52, P=0.009; per-protocol, 1.36, CL=1.13, 1.63, P=0.001; and g-estimation, 1.31,CL=1.03,1.72, P=0.027. Thus g-estimation preserved the ITT test of the null, but exhibited more uncertainty about the size of the tiotropium effect than the other methods. In particular, it allowed for a much larger potential effect than did ITT analysis, but produced a much larger null P than exhibited by per-protocol analysis. Limitations Like ITT analysis, g-estimation requires all patients be followed to the end of the trial protocol, regardless of whether they comply with the protocol. Like on-treatment and per-protocol analyses, it also requires accurate compliance information be recorded. Conclusion G-estimation should become a standard procedure for the analysis of trials with noncompliance. Software to do so is available in major packages, and the procedure is easily coded for other packages.