Causal proportional hazards estimator Loeys and Goethebeur present a method for calcu lating the true treatment efficacy in situations where all patients take their allocated treatment in one arm and compliance is all or nothing in the other arm. This means that if a patient in this arm switches, the switch is assumed to have sellckchem happened at time zero, and the patient is assumed to have only received the treatment they switched onto and none of their allocated treat ment. The method and its implementation in the Stata package are described further by Kim and White. The authors consider a clinical trial in which patients are randomised to receive either a control treatment or an experimental treatment. The method works on the assumption that all patients in the control arm comply fully, and patients in the experimental arm may either comply fully or not at all.

Patients in the control arm are also classed Inhibitors,Modulators,Libraries as either being a complier or non complier depending on how they would Inhibitors,Modulators,Libraries have behaved if they had been randomised to the experimental arm. The proportion of non compliers is assumed to be the same in both arms due to randomisation. The method then makes use of Kaplan Meier survival estimates and the assumed relationship between control and experimental compliers to find an estimate of the hazard ratio. Loeys and Goetghebeur give full details of the methodology used. The method was applied in this investigation using the Stata program stcomply as described by Kim and White.

The all or nothing compliance assumption is a very important limitation of this method Inhibitors,Modulators,Libraries as this type of compli ance is only likely Inhibitors,Modulators,Libraries to occur in very specific scenarios, such as a trial to investigate a new screening program where alternative treatment. Consider a randomised trial with two arms, a control arm receiving no treatment, and an experimental arm. Each patient i has an observed time to event or censoring Ti. Ri A or B is the patients randomised treatment arm. Each patient also has a counterfactual event time Ui which is the event time which would have been observed if no treatment had been received. Patients in the control arm who do not switch treatment will have Ti Ui, so their counterfactual event time will be observed. Ui is unobserved for all other patients. The assumption is made that Ui is independent of Ri due to randomisation balance.

Consider the observed event time Inhibitors,Modulators,Libraries Ti as being made up of a patients time on the control treatment TAi and their time on the experimental treatment TBi, so Ti TAi TBi. For patients who did not switch treatments, either TAi or TBi will be equal to zero. Ti is related to the counterfactual event time Ui by the following causal model patients may be allocated to attend screening but may not attend. As mentioned previously any other enquiries the method also makes the important exclusion restriction assumption, which although untestable, we feel is likely to hold in most set tings.