Excellent summary of the basic problem, Eric: "there are no natural experiments, only natural quasi-experiments." Period. Throughout my life I have been increasingly amazed that so few college graduates understand the full import of the fact that correlation doesn't prove causation. They can recite this mantra, but they haven't even begun to unpack its implications. It's not that genes don't "cause" myriad psychological phenomena. They do; but they do so via causal pathways that are so deeply and inextricably entangled with so many powerful environmental factors, and interactions involving such factors, that, in the end, correlation tells us very little, and often nothing. "Race science" thrives, and will continue to thrive, on science devoid of real, determinative controls--which actually, isn't science at all.
What you’re pushing back against isn’t the use of clever designs it’s the inflation of what those designs can legitimately claim. The point about “randomization” doing more rhetorical than scientific work really resonates. Segregation at meiosis solves one narrow problem and leaves a whole universe untouched. Once effects run through families, meanings, and responses, calling them “direct” feels like a category error, not a clarification.
> Within-family genetic associations are confounded in exactly the same way those twin correlations were confounded in our twin divorce studies.
But the difference is that genetic variants, unlike fathers, are randomly allocated between siblings.
You’re arguing for environmentally-mediated causation - like the red hair story - being a confound. I think it’s important to think through these mechanisms, but I don’t think calling them “confound” is the right way to frame it. The issues they raise aren’t “this isn’t real causality!” They are “what do we do about this?” In other words, policy questions.
Suppose society discriminates against red-haired kids. Here the answer is (in principle) simple: stop discriminating!
Suppose parents treat siblings with variant X differently, pushing them to work harder (say). Should they stop doing that? You need to know the counterfactual ie the gene-environment interaction. Maybe parents are responding optimally - given their preferences - to what they know about their child’s talents.
We are now in the world of evaluating possible interventions. (The social science will indeed be hard.) But if we just say “oh, parental behaviour (or similar evoked rge) is a confound, let’s go back to looking at the environment” we will miss the crucial interplay between genes and environment.
he is confused about causal mediators. which are not confounds in the usual sense. if a specific genetic variant causes parents to give one of the twin/sibling more books, and that book causes them to reach their potential, that's not a confounding variable! because that genetic variant also probably will cause more interactions with society at at large. you don't know the counterfactual of what happens when you give books to the other twin, maybe it will make them smarter, or maybe it won't do anything. that's exactly the mechanism that he laments that we are lacking! this is why i hate the within family between family distinction people make, they are not different estimators they are different estimands! they are not comparable estimates where you can easily say one number is more "biased" it's not that one is "better" than the other, even if you want to know what direct genetic effects are! what makes you think that direct genetic effects can not operate through between family differences that are CAUSED BY SAID GENES.
when you err on the side of rigorous causal inference, you should be careful! unless you want people to think you have some apriori reason for rejecting one position over another. scientific theories are in theory undecidable, we can not reject every possible permutation of every possible causal variable, but we don't need to. some sets of mechanisms are plausible and some are not! if you have a theory that within family results are still missing some potential causal variable that isn't controlled for, you can test for them with the interventions exactly in the way you are suggesting(and we have tested for tons of ideas about environmental interventions)
the possibility of error is not a sufficient reason, you have to at least show the the exact error
I’ve been hearing that for decades. First they are not null, then, upon further review, they are null and it was just confounding, which I believe was Turkheimer’s point. The broader point is that, in the interim, certain people will jump on the “non null” results then move on to the next “non null” results in an endless shell game to keep a childish ideology alive.
Much ado about a null finding. There is no gene environment interaction for education attainment, because they couldn’t find any genes. This gets in the realm of astrology, which often produces positive results without any clear causality. If your trait is even “near” null, which is the best they get in these genetic studies when the dust clears, then show me actual, agreed upon, null traits that do worse or you are just looking for a Mars effect.
That was my whole point: I don't expect the statistical results to always be null, but I don't expect those results, in the face of myriad uncontrolled confounds, to justify strong substantive causal conclusions. That's the problem.
Well done, Eric, a calm and clear plea to remain healthily sceptical. If you have already dealt with it, or might do so, I would be interested in your view of the following: suppose its the case that no greater heritabilities emerge from the Human Genome Project and therefore, that genes are playing very little role in causing individual differences in human psychological traits (which is what the several hundreds of studies have demonstrated again and again, with no exceptions). In that case, what do you make of the variations in heritabilities between traits in twin studies - why are there are higher MZ vs DZ concordances for intelligence, say, compared with pattern of attachment or most neuroses? Keep on keeping on, Best, Oliver James
Excellent summary of the basic problem, Eric: "there are no natural experiments, only natural quasi-experiments." Period. Throughout my life I have been increasingly amazed that so few college graduates understand the full import of the fact that correlation doesn't prove causation. They can recite this mantra, but they haven't even begun to unpack its implications. It's not that genes don't "cause" myriad psychological phenomena. They do; but they do so via causal pathways that are so deeply and inextricably entangled with so many powerful environmental factors, and interactions involving such factors, that, in the end, correlation tells us very little, and often nothing. "Race science" thrives, and will continue to thrive, on science devoid of real, determinative controls--which actually, isn't science at all.
What you’re pushing back against isn’t the use of clever designs it’s the inflation of what those designs can legitimately claim. The point about “randomization” doing more rhetorical than scientific work really resonates. Segregation at meiosis solves one narrow problem and leaves a whole universe untouched. Once effects run through families, meanings, and responses, calling them “direct” feels like a category error, not a clarification.
> Within-family genetic associations are confounded in exactly the same way those twin correlations were confounded in our twin divorce studies.
But the difference is that genetic variants, unlike fathers, are randomly allocated between siblings.
You’re arguing for environmentally-mediated causation - like the red hair story - being a confound. I think it’s important to think through these mechanisms, but I don’t think calling them “confound” is the right way to frame it. The issues they raise aren’t “this isn’t real causality!” They are “what do we do about this?” In other words, policy questions.
Suppose society discriminates against red-haired kids. Here the answer is (in principle) simple: stop discriminating!
Suppose parents treat siblings with variant X differently, pushing them to work harder (say). Should they stop doing that? You need to know the counterfactual ie the gene-environment interaction. Maybe parents are responding optimally - given their preferences - to what they know about their child’s talents.
We are now in the world of evaluating possible interventions. (The social science will indeed be hard.) But if we just say “oh, parental behaviour (or similar evoked rge) is a confound, let’s go back to looking at the environment” we will miss the crucial interplay between genes and environment.
he is confused about causal mediators. which are not confounds in the usual sense. if a specific genetic variant causes parents to give one of the twin/sibling more books, and that book causes them to reach their potential, that's not a confounding variable! because that genetic variant also probably will cause more interactions with society at at large. you don't know the counterfactual of what happens when you give books to the other twin, maybe it will make them smarter, or maybe it won't do anything. that's exactly the mechanism that he laments that we are lacking! this is why i hate the within family between family distinction people make, they are not different estimators they are different estimands! they are not comparable estimates where you can easily say one number is more "biased" it's not that one is "better" than the other, even if you want to know what direct genetic effects are! what makes you think that direct genetic effects can not operate through between family differences that are CAUSED BY SAID GENES.
when you err on the side of rigorous causal inference, you should be careful! unless you want people to think you have some apriori reason for rejecting one position over another. scientific theories are in theory undecidable, we can not reject every possible permutation of every possible causal variable, but we don't need to. some sets of mechanisms are plausible and some are not! if you have a theory that within family results are still missing some potential causal variable that isn't controlled for, you can test for them with the interventions exactly in the way you are suggesting(and we have tested for tons of ideas about environmental interventions)
the possibility of error is not a sufficient reason, you have to at least show the the exact error
I’ve been hearing that for decades. First they are not null, then, upon further review, they are null and it was just confounding, which I believe was Turkheimer’s point. The broader point is that, in the interim, certain people will jump on the “non null” results then move on to the next “non null” results in an endless shell game to keep a childish ideology alive.
Much ado about a null finding. There is no gene environment interaction for education attainment, because they couldn’t find any genes. This gets in the realm of astrology, which often produces positive results without any clear causality. If your trait is even “near” null, which is the best they get in these genetic studies when the dust clears, then show me actual, agreed upon, null traits that do worse or you are just looking for a Mars effect.
It sounds as if Professor Turkheimer expects future results on education not to be null. I think that is also quite likely.
That was my whole point: I don't expect the statistical results to always be null, but I don't expect those results, in the face of myriad uncontrolled confounds, to justify strong substantive causal conclusions. That's the problem.
Sure, but mediators are not the same as confounds - on which now see my latest post.
Siblings ma have the same parents but not the same parenting.
They have advanced past the ideologue stage to the conflict-of-interest stage and people like Alex Young shouldn’t be conducting these studies.
Well done, Eric, a calm and clear plea to remain healthily sceptical. If you have already dealt with it, or might do so, I would be interested in your view of the following: suppose its the case that no greater heritabilities emerge from the Human Genome Project and therefore, that genes are playing very little role in causing individual differences in human psychological traits (which is what the several hundreds of studies have demonstrated again and again, with no exceptions). In that case, what do you make of the variations in heritabilities between traits in twin studies - why are there are higher MZ vs DZ concordances for intelligence, say, compared with pattern of attachment or most neuroses? Keep on keeping on, Best, Oliver James