Causal Analysis

Does smoking cause lung cancer? It is hard to believe that this was once a question in some dispute. Yet despite the fact that there has been no randomized controlled trial (RCT) in which research subjects were randomly assigned to smoking or nonsmoking conditions with subsequent long-term follow-up to ascertain differences in health outcomes, there has long been a consensus that smoking does indeed cause lung cancer, although it is certainly not the only cause. However, although smoking-and-health is certainly not the only case where a consensus has been reached about causality, asbestos exposure being another, the research literature and the popular press are full of cases where causal impacts are in hot dispute. For example, currently bisphenol A, a chemical found in baby bottles and many other plastic products, has been tentatively associated with various health conditions. However, the extent to which the association is causal and the strength of the effect, if any, remain in dispute, and a long series of investigations will need to be conducted to resolve the matter.

Why is causal inference so difficult? Even in cases where RCTs are possible, the results are often open to challenge. In cases where randomized studies are not possible, due to ethical or other reasons, establishing causality is far more difficult. The concept of cause itself is famously elusive. Apart from definitional problems, attempts to elucidate sets of causal criteria, from David Hume to John Stuart Mill to Austin Bradford Hill, have not provided necessary and sufficient conditions for concluding that an observed association between two variables results from the causal impact of one on the other. From the standpoint of social science research, at least three issues are problematic. First, many philosophical discussions of cause begin with a deterministic relationship. If X changes, Y changes, by the same amount and for all cases under study. But in health services research relationships are usually probabilistic and heterogeneous. A change in X may or may not result in a change in Y, the amount of change may vary across units of the population, and changes in X may not be the only source of variation in Y. While statistical models are designed to cope with probabilistic outcomes, they are often based on assumptions that are difficult to defend (e.g., that the source of random noise in the data is uncorrelated with systematic sources of variation). A second problem, related to the first, is that variation in many outcomes is multicausal. For example, a teenager's proclivity to commit violent acts may have its origins in a variety of genetic and environmental factors, any one of which may be sufficient to cause violent behavior in some but not all persons exposed to the risk. Finally, in health services research, researchers are often interested in a causal sequence such that at a particular attribute, say race, puts an individual at varying levels of risk for some outcome, say discrimination, which in turn is reflected in a subsequent outcome such as access to healthcare. Demonstrating the validity of the mediational assumption is often difficult.

...