Estimating treatment effect in the presence of non-compliance measured with error: precision and robustness of data analysis methods.

Non-compliance with the nominal prescribed dosage causes unintended variability in actual drug exposure during clinical trials. In the ideal case that compliance is not a confounder, and it is known--hence actual dosage is known--true dose-response can be validly estimated. Measuring compliance presents a challenge, however. A simulation study of the case that dosage history questionnaires (C(Q)--usually over-optimistic estimates of actual compliance) are available in all subjects enrolled in a clinical trial, but accurate compliance measurements (C--e.g. from electronic medication event monitors), are only available in a (random) fraction of subjects is reported. It reveals that a 'Maximum Penalized Marginal Likelihood' (MPML) method which uses all compliance data, effectively calibrating C(Q) to C, is superior to other methods which use only one compliance measure, or both, or neither (neither = ITT, intention to treat, which assumes actual dosage equals nominal dosage), but do not calibrate. MPML yields the most precise estimates of dose-response over widely varying clinical trial designs, extremes in quality and quantity of compliance information, and a range of drug effect sizes. It is most beneficial when compliance data are sparse and maintains good performance even when its key assumptions are somewhat violated.