Causality, genes, and the law

Ewen Callaway in New Scientist reports:

In 2007, Abdelmalek Bayout admitted to stabbing and killing a man and received a sentenced of 9 years and 2 months. Last week, Nature reported that Pier Valerio Reinotti, an appeal court judge in Trieste, Italy, cut Bayout’s sentence by a year after finding out he has gene variants linked to aggression. Leaving aside the question of whether this link is well enough understood to justify Reinotti’s decision, should genes ever be considered a legitimate defence?

Short answer: probably not.

Long answer: This reminds me of an issue I have with the Rubin Causal Model. In Holland’s 1986 paper on the RCM, he has a section titled “What can be a cause?” He introduces the notion of potential exposability – basically the idea that something can only be a cause if you could, in principle, manipulate it. He contrasts causes with attributes – features of individuals that are part of the definition of the individual. He uses as an example the statement, “She did well on the exam because she is a woman.” Gender can be statistically associated (correlated) with an outcome, but it cannot be a cause (according to Holland and I believe Rubin as well), because the person who did well on the exam would not be the same person if “she” weren’t a woman.

From a scientific/philosophical level, I’ve never liked the way they make the cause/attribute distinction. The RCM is so elegant and logical and principled, and then they tack on this very pragmatic and mushy issue of what can and cannot be manipulated. If technology changes to where something becomes manipulable, or if someone else thinks of a manipulation that escapes the researcher’s imagination (sex reassignment surgery?), things can shift back and forth from being classed as causes versus as attributes. Philosophically speaking: Blech. Plus, it leads to places I don’t really like. What about: “Jane didn’t get the job because she is a woman.” Is Holland saying that we cannot say that an applicant’s gender affected the employer’s hiring decision?

I think we just need to be better about defining the units and the nature of the counterfactuals. If we are trying to draw inferences about Jane, as she existed on a specific date and time and location, and therefore as a principled matter of defining the question (not as a pragmatic concern) we take as an a priori fact that Jane for the purposes of this problem has to be a woman, then okay, we’ve defined our problem space in a particular way that excludes “is a man” as a potential state of Jane. But if we are trying to draw inferences in which the units are exam-takers or job applicants, and Jane is one of many potential members of that population of units, then we’re dealing with a totally different question. In that case, we could have had either a man or a woman take the exam or apply for the job. Put another way: what is the counterfactual to Jane taking the exam or Jane applying for the job? If Jane could have been John for purposes of the problem that we are trying to solve, then it makes perfectly good sense to say that “Jane did well on the exam because she is a woman” is a coherent causal inference. It goes back to a principled matter of how we have defined the problem. Not a practical question of manipulability.

So back to the criminal… Holland (and Rubin) would make the question, “Is the MAOA-L variant a cause or an attribute?” And then they’d get into questions of whether you could manipulate that gene. And right now we cannot, so it’s an attribute; but maybe someday we’ll be able to, and then it’ll be a cause.

But I’d instead approach it by asking: what are the units, and what’s the counterfactual? To a scientist, it makes perfect sense to formulate a causal-inference problem in which the universe of units consists of all possible persons. Then we compare two persons whose genomes are entirely identical except for their MAOA variant, and we ask what the potential outcomes would be if one vs. the other was put in some situation that allows you to measure aggressive behavior. So the scientist gets to ask questions about MAOA causing aggression, because the scientist is drawing inferences about how persons behave, and MAOA is a variable across those units (generic persons).

But a court is supposed to ask different kinds of causal questions. The court judges the actual individual before it. And the units are potential or actual actions of that specific person as he existed on the day of the alleged crime. The units are not members of the generic category of persons. Thus, the court should not be considering what would happen if the real Abdelmalek Bayout had been replaced by a hypothetical almost-Bayout with a minutely different genome. A scientist can go there, but a court cannot. Rather, the court’s counterfactual is a different behavior from the very same real-world Abdelmalek Bayout, i.e., a Bayout who didn’t stab anybody on that day in 2007. And if Bayout had not stabbed anybody, there’d be no murder. But since he did, he caused a murder.

Addendum: it’s a totally different question of whether we want to hold all persons to the same standards. For example, we have the insanity defense. But there, it’s not a question of causality. In fact, defendants who plead insanity have to stipulate to the causal question (e.g. in a murder trial, they have to acknowledge that the defendant’s actions caused the death of another). The question before the court basically becomes a descriptive question — is this person sane or insane? — not a causal one.