Nova Acta Leopoldina Band 110 Nummer 377

With non-parametric curves for the entire survival distribution, there is not a single obvious parameter summarizing the causal effect. To achieve this, one can propose a model for the structural relationship between survival time Ti and potential treatment-free survival time Ti0. A simple semi-parametric accelerated failure time model is: Ti0 (Ψ) = exp(–Ψ Xi) Ti [7] The procedure outlined above, then estimates that the pump implant has reduced survival time to 70 % times its original value with a 95 % confidence interval for exp(Ψ) ranging from 0.35 to 2.06 (LOEyS and GOETGHEBEUR 2003). 5.2 Instrumental Variables in the Observational Setting Since randomization enables such powerful and reliable causal conclusions in experimental designs, one has sought to identify instrumental variables that allow to mimic this approach in observational studies. A recent study of the effect of COX-2 inhibiting rather than non-se- lective non-steroidal anti-inflammatory drugs on the risk of gastro-intestinal (GI) bleeding for rheumatoid arthritis patients identified physician-specific prescribing preference for the COX-2 inhibitors (coxibs) as a candidate instrumental variable (BROOKHART et al. 2006). We offer some background. Randomized Clinical Trials found COX-2 inhibitors to reduce the risk of GI complications. Even so, several observational studies remained unable to find a significantly negative associ- ation between the use of COX-2 inhibitors and GI bleeding – even after adjustments. Because observational studies typically involve broader populations treated under less controlled cir- cumstances, it is important to verify how benefits seen in trials stand up in the practical setting. Since COX-2 inhibitors have a reputation of provoking fewer GI problems, they are likely prescribed more to patients deemed at higher risk.Adjusting the relationship between GI bleeds and COX-2 inhibitors for all confounders is then complicated, especially due to poor meas- urement of established confounders in databases typically available for pharmaco-epidemiol- ogy. Residual positive confounding could leave the group of observed users of COX-2 inhibitors at higher risk, even if this class of drugs effectively reduces natural risk. Lacking reliable adjustments, RASSEN et al. (2008, 2009a, b) developed an instrumental variables ap- proach. RASSEN et al. (2009b) see “physician prescription preference” as an instrument Z, satisfying the two conditions: – the prescription of COX-2 inhibitors (exposure X) depends on individual physician prefer- ence (Z); – a different physician preference (Z) does not come with a different patient population or ref- erence risk mix (Y0) or: Z affects the outcome Y through X and only through X, often labeled “the exclusion restric- tion”. Clearly, whether or not one’s physician has a preference for the COX-2 inhibitors will change one’s chance of receiving this drug; but the physician preference itself need not predict the natural GI risk levels of his or her patients. If the IV assumptions hold, no matter how se- lectively the drug has been prescribed, one can recover its causal effect as the effect that back- transforms observed risks to levels that are comparable for different values of the instrument. For their study of COX-2 inhibitors, SCHNEEWEISS at al. (2009) considered a large population- Causal Inference: Sense and Sensitivity versus Prior and Prejudice Nova Acta Leopoldina NF 110, Nr. 377, 47–64 (2011) 57