The conditional compliance model (Rubin model) is compared with the conditional observable (Robins model) model to determine under what assumption the two estimands are equal. One important assumption is "no defiers" (there are no subjects who, if assigned to treatment take control and, if assigned to control, take the treatment). In this example -- a randomized study to improve primary care physicians' adherence to best practice guidelines in treating depressed patients -- both expert opinion and empirical data suggested defiers might be present.

• "Causal Linear Models for Non-Compliance under Randomized Treatment with Univariate Continuous Response," Ten Have, TR, Elliott, MR, Joffe, M, Zanutto, E (2004), Journal of the American Statistical Association, 99, 16-25.