In part one of this talk I reviewed the binary instrumental variable potential outcome model, motivated this model in the context of non-compliance in clinical trials, and showed that the posterior distribution over the average causal effect is sensitive to the prior.
This sensitivity motivates an analysis of identifiability. In part two of the talk I will provide a detailed graphical description of the set of distributions over patient types that give rise to the same likelihood for the observables. This also leads to a variation independent parameterization of the likelihood, which facilitates the incorporation of baseline covariates.
(Joint work with James Robins, Harvard School of Public Health)
