Nova Acta Leopoldina Band 110 Nummer 377

1. Introduction The age old quest for the golden grail of causal answers has been at the heart of science for centuries (COX and WERMUTH 2004). Extracting causal answers from observed data, has time and again proven to be harder than anticipated. The key difficulty stems from the fact that ob- served exposure to a compound of interest tends to be correlated in realistic populations with a host of other factors which themselves affect outcome: confounders. Measured confounders, pre-exposure variables which are associated with outcome as well as exposure as illustrated in Figure 1, can be accounted for through regression models, but the existence of unmeasured confounders can rarely be excluded. When important unmeasured confounders are suspected to play a role, an instrumental vari- ables approach is an alternative. Due to the digital revolution and exploding computer capacity we have been able to incorporate ever more data in our evaluations of causal effects. But more data does not automatically mean better and more reliable answers. Pitfalls are many in this field and seriously misleading conclusions can result. Encouragingly, progress continues to be made, and over the past decades breakthrough insights into methods for extracting causal information from empirical data have emerged. This has lead to Nobel prizes in Economics (HECKMAN 2008) and a new area of statistics. That progress is still possible after decades of hard thinking by the brightest minds we owe to ever more sophisticated insight into mathematical and statistical models, and in no lesser part to the development of computer technology that allows fitting complex models to data with relatively little time and effort. The latter encourages two things: (1.) new theoretical developments and (2.) new flexible implementation. The ease of implementation is a great facilitator but not without danger. While causal models may seem relatively simple and reminiscent of (linear) association models, the abstract level of thinking needed and assumptions involved have an added dimension beyond standard sta- tistical association analysis. At the heart of the causal question is the desire to transform the world and understand how outcomes would change if the exposure of interest were changed. Causal Inference: Sense and Sensitivity versus Prior and Prejudice Nova Acta Leopoldina NF 110, Nr. 377, 47–64 (2011) 49 Fig. 1 Disease severity confounding the relationship between treatment and mortality