Via @betsylevyp, I came across a new paper by Maya Sen and Omar Wasow on race and causal inference. It’s excellent, well worth a read.
The title of this post is not a critique, not exactly. It’s a reflection on the problem that Sen and Wasow confront. The issue that the currently dominant theory of causality in most social science—which is really an epistemology of causality rather than a theory of it—defines causality in terms of differences between outcomes in a unit that occupies both treatment and control states. This entails that the unit could in principle occupy both treatment and control states. In other words, manipulability is the foundation of studying causal effects. This “causation requires manipulation” framework is usually associated with the Rubin causal model but has a much longer pedigree in philosophy.
Our problem is that race—and other things like gender, height, etc.—is not manipulable. That means that no matter how perfect the design, the difference in outcomes between two different racial groups cannot be interpreted as “the effect of race.” This bothers many people, mostly philosophers, but also me. It suggests that our epistemology of causality must be somehow mistaken. We should be able to ask questions of the type “what is the effect of race on Y?” Reality demands it. I can’t count the number of times I’ve stayed up late debating this point.
Sen and Wasow solve the problem by appealing to a constructivist conception of race, which renders some aspects of race manipulable. It’s a smart move and one that I endorse. But I am just as eager to read a thorough critique too. One preliminary thought is that their constructivist understanding of race is situationalist, so what we manipulate is not race itself or the sticks the constitute it, but the conditions under which race is understood, encountered, produced, etc. That strikes me as somehow different than manipulating the un-manipulatable.
tompepinsky.com 
Rich Maass November 21, 2014
This is a fascinating subject, and one that we should not expect to come to a conclusion on quickly, but I am nevertheless dismayed that I remain troubled by it after further reflection sparked by your post here, Tom.
On one hand, in social science we can easily dodge the problem by appealing, in this case, to the social perception of race as something that is manipulable and hence fits in a causality-requires-manipulability epistemology. That feels helpful for proceeding in addressing some social-science questions about effects of race on human behavior, but it feels like a cop-out in that it denies that race could have physical effects. What about when we try to answer questions about the effect of race on skin cancer, for example? Are we saying that physical sciences cannot produce knowledge about causality in cases where the key causal variable is not manipulable?
The real hurdle seems to be the conviction that, not only can we not physically change race in the real world, but comparing people of different races involves so many other related differences that it is severely problematic to consider them as “all else being equal” as required by dominant views on epistemology. This potentially creates lots of problems for causal inference in political science… e.g., trying to determine the effect of democracy on foreign policy… not only is it physically impossible to change a state’s democratic characteristics without changing a ton of other characteristics as well (many of which likely developed interdependently with its existing democracy), but so much varies across democracies that comparing them as “all else being equal” might often be extremely problematic.
I’m starting to make my head hurt, and I’m not sure that it’s necessary, since I do still think it is possible to answer a lot of questions about causality by testing causal processes within cases to evaluate e.g. whether particular characteristics of democratic government played a critical role in producing the outcome. Do these concerns bear more on some research methodologies than others?