Counterfactuals and age

This is Joseph.

I was reading Andrew Gelman's post on potential outcomes and it reminded me of the classic problem with this parameter in epidemiology, namely chronological age. I like potential outcomes for a lot of questions -- what happens if we change an exposure and create a counterfactual outcome is a very logical way to think about things like prescription drugs, milk in the diet, and water purity. There is a reasonable causal contrast between exposed and unexposed that could be estimated (e.g., it is true in expectation in a randomized controlled trial).

It gets trickier with other variables. Sex (as opposed to gender) is difficult to imagine being anything but a theoretical contrast (gender, on the other hand, is a lot more complicated). Similarly, disentangling the effects of race as a genetic piece (usually very small) versus racism is a very challenging causal inference problem. You can imagine changing the racist nature of society but not really genotypes. Smart people work on this problem, but I think of it as quite hard (and thus I am glad for the smart people).

But the biggest challenge is age. Age is a very important etiological factor in driving key diseases like cardiovascular disease and cancer. You probably won't understand risks for these diseases very well should you not control for age. But what is the counterfactual for age? If I say taking drug X doubles your risk of a stroke, that is easy -- the contrast is taking/not taking the drug. But if I say aging 15 years doubles your risk for a stroke what is the potential outcome, as age is not in any sense a fully controlled variable. 

Now, some people say age is a correlated for a whole mess of biological processes and if we measured them all then we could create this contrast, piece by piece. But that simply sidesteps the blunt fact that age is a super powerful predictor and we don't have a great sense of all of the pieces of aging. 

It seems easy because the first thing that we consider when studying (for example) mortality, is age. But what is the counterfactual -- a person being 15 years older or younger with the same life-course exposures. It is an area where the (extremely useful) tool works . . . awkwardly. And that is a good lesson -- models (even conceptual models) are useful tools but need to be carefully thought about in the context of the specific question under study.