Featured Post

Is the new professionalism and ACP's new ethics really just about following guidelines?

The Charter ( Medical Professionalism in the New Millennium.A Physician's Charter) did not deal with just the important relationship of ...

Tuesday, April 05, 2016

Statistical independence does not mean causal

   An article in the Jan 24, 2005 issue of the Archives of Internal Medicine(vol. 165 p138-145) made important points regarding the concept of "independent risk factors". Basically, Dr. Brotman and co-authors remind  (or more likely inform) the readers that ( my bolding): statistical independence does not mean causality, is context dependent ( ie in that particular data set) and risk factors may be causal even if not statistically independent. Independence is a statistical concept relying on a particular statistical model.

  I once downplayed the significance of elevated triglycerides as a risk factor for coronary disease because  I had read triglycerides were not an independent risk factor. Now, or course, clinical studies have shown the opposite- at least for now. The point is that a risk factor can be "significant" i.e. important whether or not a medical publication's analysis indicates that is an "independent risk factor" .

Articles like this can be  are important antidotes to the faith that we tend to have in the black box magical output of multivariate analysis as well as less familiar techniques ( eg propensity score matching  and,instrumental variables methods which purport to make observational studies more like randomized trials).  Few physicians have plowed through the pen and paper process of doing a multivariate analysis or even understand generally what it all about.I don't claim to. That exercise might give one a real sense of what is being done and perhaps how small variations in data input can alter the answer- changing an independent risk into one that is not and vice versa.

In regard to heart disease, the authors assert that as more variables are linked to disease, no study will be able to properly model all the risk factors to enable them to say that X is an independent risk factor. This problem of residual confounding limits medicine's search for the causes and might make us more circumspect when we make pronouncements to patients about what causes what and what we should do about it and we cannot just rely on whether a  particular study did or did not show statistical independence.

William Barrett
In his book "Illusion of Technique"(Anchor Books, 1979) says that Logic is the only modern science that has shown its own limits by showing the limits of formal systems(through the work of Godel and others).

We might tend to forget when we read " X is an independent risk factor for disease Y" that we are dealing with "provisional conclusions " extracted from "fragmentary" data and working within a particular statistical  model with a  particular data set.

 Some have make a distinction between "faith based" medicine with "evidence based medicine". Considering the faith required to believe the output of mysterious mathematical models about which most physicians readers know little , this distinction begins to fade away.Maybe we need a medical version of Godel's theorem as an antidote to hubris or faith based believe in technique. I eagerly await more articles,such as Brotman's,pointing out the limits of medicine's knowledge generating techniques.

note: This theme appeared in  comments  in a different venue by me about ten year ago and have not appeared  before on  RDT and has been lightly revised and edited.

No comments: