Students are taught the perils of inferring causality from observational studies, and the shortcomings of nonrandomized clinical trials. In his seminal text book on modern epidemiology, Rothman et al.  dedicate considerable discussion to causal inference and even goes so far as to try to present and critique criteria necessary to consider or establish when concluding a causal relationship between two variables. Although formal conceptualization of causal inference began early in the last century, there remains disagreement concerning the ability to discovery novel causal effects from all but the most rigorous of controlled clinical trials and mechanistic experiments. Causality is connected to probability by some experts (e.g. , and ), whereby an attempt is made to quantitate the probability that A causes B, with assumptions about the mechanism by which individuals were assigned to levels of A. If that probability exceeds some threshold, a causal relationship is claimed. However, interpreting probabilities as causal quantities in the absence of clear knowledge about the assumptions underlying, this interpretation can lead to confusion. To avoid such confusion, Pearl promoted a deterministic interpretation of causal inference at the population level using structural equation modeling [4-6]. Rubin also defines the causal effect deterministically, but at the individual level ; for discussions on causal effect definition see also [8-11] and for causality in genetics effects see [12-13].
The difference between the “individual” and “population” causal effect has meaning that transcends esoteric or theoretical considerations. As an example, let’s consider a drug, a desired outcome and an adverse event. The policy arm of health care wants to know whether prescribing the drug to the population of patients will increase the frequency of the desired outcome (and presumably reduce disease incidence) without undo increase in the frequency of the adverse event. The physician, on the other hand, wants to know whether prescribing the drug to the patient in his/her office at that time will elicit the desired outcome without leading to the adverse event in that patient. Typically, analyses and inference are done on a large sample from the population and then the results are used to make inference about whether the next individual sampled from the same population will respond or not. In its simplest form, inference about the response of the next individual sampled from the population is the average response in the population. Personalized medicine connotes the idea that treatment has been tailored to specific characteristics of the individual. In practice, treatment is not tailored to each individual, but rather is tailored to groups of individuals based on the results of specific diagnostic information, such as the level of a biomarker or genetic information. The term “Decision Medicine” has recently been suggested, which indicates a more immediate translational perspective .
The question we ask is whether we should approach causal inference including the assumptions and data analysis task differently depending on whether our primary interest is the average response of treatment in the population or the ability to characterize the response for an individual or a subgroup. Regardless of the term, the field of personalized medicine has much to benefit from advances in causal inference. This perspective provides a tutorial of modern causal inference and then provides suggestions how application of specific kinds of causal inference would promote advances in translational applications of personalized or decision medicine. The example application is carried out pragmatically using a graphical approach followed by Structural Equation Modeling (SEM).
Citation: Yazdani A, Boerwinkle E (2014) Causal Inference in the Age of Decision Medicine. J Data Mining Genomics Proteomics 6:163. doi: 10.4172/2153-0602.1000163