|
Abstract:
|
The best evidence regarding the impact of a drug (intervention) on a disease (condition)is obtained from Randomized Clinical Trials. The benefit of the design is that randomization renders all groups in the experiment equivalent with respect to observed and unobserved characteristics and the only difference between groups is the intervention. Despite efforts to ensure compliance with the assigned treatment and follow-up protocol, some trials are prone to cross-over, drop-out, or use of a non-trial treatment. The Intention to Treat (ITT) analysis estimates treatment effects as if the randomization was maintained. This results in an effectiveness estimate but not an efficacy estimate.The Rubin Causal Model(RCM) is one way to get an efficacy estimate, i.e. a causal effect of the intervention based on actual exposure. The disadvantage of the RCM is that in most applications, estimation is not possible if the exclusion restriction assumption is not met. We present a Bayesian analysis of simulated and actual RCT data that does not need the exclusion restriction assumption, for dichotomous and ordinal (fully compliant, partially compliant & non-compliant) compliance variables
|
