[Epistemic status: I am not a behavioral geneticist, I don’t really understand the behavioral genetics papers, and I am fully expecting to have to run a correction on some if not all of this post.]
The conventional wisdom of behavioral genetics, as I understand it, is that approximately fifty percent of variance in personality is genetic, fifty percent is environmental factors not shared between siblings (often inaccurately rendered as “schools and peers”), and zero percent is environmental factors shared between siblings (often inaccurately rendered as “parenting”).
However, this argument proves far too much.
Consider lead. Lead is the biodeterminist’s favorite story: we poisoned our children and, right on schedule, they became criminals; we stopped poisoning our children and, once again right on schedule, crime dropped. However, lead varies a ton between neighborhoods: some of the evidence for the lead-crime hypothesis is that criminality is highly correlated with growing up in a lead-full neighborhood. One assumes that siblings generally do grow up in the same neighborhood, so lead exposure is shared environment. Why doesn’t that show up?
Or consider breastfeeding. An study of 14,000 infants suggests that an intervention to increase the rate of breastfeeding caused a remarkable six-point rise in IQ. One would expect that there is not literally zero correlation between whether you breastfeed your first child and whether you breastfeed your second child, since you’d expect women who are incapable of producing milk, have unsupportive spouses or employers, or simply don’t want to breastfeed would usually continue to be in those situations. Again, some of that IQ rise should show up in the data.
It also doesn’t really pass the sniff test. There are lots of ways that other people’s behavior can affect your outcomes. It seems to me to be well-established that people with borderline personality disorder behave in extremely obnoxious ways, unless we are taught certain emotion regulation and mindfulness skills, in which case we are no more unpleasant than anyone else. People who have been assaulted are at much higher risk of PTSD than people who have not been assaulted. A person with ADHD who has access to medication is likely to be more focused and productive than one who has been refused medication. Heck, a good boss inspires people to be more productive, while a bad boss encourages slacking. In theory, all of these ought to show up in shared environment: parents who teach their children emotion regulation skills, assault their children, give their children psychiatric medication, or are terrible bosses are likely to continue to have those traits across multiple children. There is no reason to believe that children are less prone to learning helpful life skills or being inspired. And while it’s possible that the effects don’t last until adulthood– people who weren’t on ADHD medication as children regularly get on it as adults, people may recover from PTSD quickly– you’d still expect flow-through effects. A child who maintained a C average because of their ADHD gets into a worse college and earns less money as an adult– even once they take medication.
The obvious explanation is that twin studies control for too much. Twin studies are typically conducted on economically homogeneous households and/or controlling for parental socioeconomic status. This explains both lead and breastfeeding: rich people are less likely to live in lead-full neighborhoods and more likely to breastfeed, so controlling for wealth makes lead and breastfeeding less obvious than they previously were. This could plausibly save some other parenting advice as well: rich people do parent differently from poor people, and if a parenting strategy is widely adopted among one class but not another, then its effects will be less obvious.
But I think there are some other things to bear in mind. Twin studies mostly concentrate on the broad strokes of personality: IQ, five-factor personality, and easy-to-study outcome measures like educational attainment and income. However, there are lots of people who have the same IQ, five-factor personality, educational attainment, and income, but are very very different people. I think that explains the borderline thing: borderlines are absurdly neurotic regardless of what skills you teach them, and they are always going to be neurotic, but learning emotion regulation is the crucial, life-altering difference between “neurotic” and “I just destroyed all my possessions because I love you”.
Furthermore, the studies really aren’t good at picking up on small trends. A study which used multiple measures of personality, instead of relying on a single self-reported measure, found large reductions in the amount of variation attributable to non-shared environment: a substantial amount of the difference between twins comes down to nothing more than “Alice pressed ‘agree’ when Eve pressed ‘strongly agree'”. That is a huge amount of noise. My understanding is that it is very hard to pick out small signals from very noisy measures. While parenting can’t get you from 25th percentile conscientiousness to 75th percentile, the measures aren’t good enough to suggest it can’t get you from 25th percentile to 35th percentile– which is actually a pretty large change in terms of how much work you get done. I admit that this is a sort of ‘parenting of the gaps’ theory: however, it does fit well with my understanding of how people work. People’s basic personality is fixed and unlikely to change; but there’s a lot of good you can do fiddling with the margins.
And, of course, there is always the final conclusion that– no matter how little effect you have on your children age 25– you are likely to have a great deal of effect on their happiness age 5, and perhaps the best parenting strategy is, while respecting your own boundaries and needs, to try to make that five-year-old’s life as pleasant and fulfilling as possible.
The counterpoint of “lead varies by neighborhood so why doesn’t lead show up” is extremely strong. I am usually a proponent of the “conventional wisdom of behavioral genetics.” But the lead example is extremely hard to respond to and I might need to re-think my views.
(The breastfeeding example s less slam-dunk since the science on breastfeeding is still unclear.)
It’s always seemed weird to me that studies get summed up as “X% of variation is genetic, X% is upbringing, X% is background noise” … when surely the amount of variation caused by “upbringing” and “personal experiences” is based almost entirely on the sample?
If you look at people on pre-Civil War plantations, you’re going to find that black people mysteriously do very poorly for genetic reasons.
If you look at the entire world population, you’ll find your odds of being shot by a drone are mostly shared environment; but if you look at a single neighborhood, it’s suddenly non-shared.
These studies don’t tell us deep truths about human nature; they tell us deep truths about how varied our society is.
Which is important!
It’s good to know if outcomes are mostly random or genetic in your society, so you know what to focus on!
But then you get people going “oh, outcomes are 50% genetic, 50% upbringing, you see, so you see no matter what you do kids have a 50% chance of ending up just as bad as their parents” and ARGH YOU’RE JUST USING NUMBERS THAT HAVE NOTHING TO DO WITH EACH OTHER. It’s like that thing where people see a 90% chance a test is correct, so clearly nine out of ten tests will correctly identify the one-in-a-million condition.
LikeLiked by 2 people
There’s a recently discovered genetics answer for this. Heritability studies are generally forced to treat genetic effectives as being purely additive, because otherwise the model has too many parameters (and it in general works okay anyway). However recent studies using whole genome similarity have the power to look at more complex genetic effects such as dominance heritability. Dominance effects tend to make siblings less than they otherwise would be, so if you have shared environment and dominance, the two cancel each other out in a simple additive model twin study. The recent study showed dominance effects to be much smaller than additive ones but still present, enough to mask roughly the level of shared environment you’re positing.
So, you already said “inaccurately rendered as “parenting””, but “environmental factors shared between siblings = parenting” thing seems really bizarre to me.
Like, for one thing, just from knowing people, it seems fairly clear that one thing that can have a pretty major effect on people is ‘how exactly did their parents abuse them’ etc. People being terrified of conflict, people who hide it when they like things because that was used against them, etc etc. But, from what I know of people who study that area of things, it is in fact very common for parents to abuse their children in *different* ways from each other.
Also, the thing you said about ‘too much shared environment’ seems important. Like – as far as I know, it’s generally accepted that people who aren’t taught how sex things actually worked are more likely to run into particular problems. But, that kind of thing tends to be very socially widespread, especially at any given time.
I also always wonder about, like, cultural things. Specifically with respect to how family vs peer things are arranged. Because we live in a society where peer environments are pretty major and families environments can be less so. But there’s other situations where people’s lives consist almost entirely of their families.
Finally, much like with the sex ed, this seems to fail on fairly obvious things. Like, I recently realized that I tend to have no idea how to deal with people apologizing to me a lot of the time, and a thing with that is that I’ve never really seen it modeled. Which is clearly related to my family. Saying that never has an effect sounds like saying that there’s no relation between the fact that I was never taught geography and that I tend to have awful knowledge of geography. Like – of course those things are related, that’s how knowledge works.
LikeLiked by 1 person
I wrote a piece in the same vein a couple years ago:
This is more about attribution to parenting, and less about the missing shared environmental term.
Carl Shulman said:
“The randomized trial in Belarus did evaluate IQ. Its results are mixed and a little confusing.
First, researchers looked at all the kids in the study. For this sample, the evaluation of IQ was done by evaluators who knew whether or not a child was in the breastfeeding-encouraged treatment group. There were no significant effects of breastfeeding on overall IQ. In addition, breastfeeding had no effect on teachers’ evaluation of the children’s school performance. But the researchers observed large effects of breastfeeding on verbal IQ.
Because the researchers were concerned about evaluator bias, they also had a subset of children evaluated by independent evaluators who did not know which children were breastfed. The differences in verbal IQ disappeared. This, in combination with the teacher evaluations, makes it seem likely that the overall effect was driven by the evaluators, not by true differences among children because of breastfeeding.
This explanation seems especially likely since the effects observed in the full sample are too large to be plausible. Taking into account the impact of the program on breastfeeding rates, the results suggest that nursing increases child IQ by about 24 IQ points, which is far outside of what any other study — even one seriously biased by differences across mothers — would suggest. Overall, as others have noted, this study doesn’t provide especially strong support for the claim that breastfeeding increases IQ.
Comparisons among siblings (i.e., this and this) also show no IQ impacts. Again, these studies make clear that if you ignore differences across mothers, you can find large impacts of breastfeeding on IQ. It is only when you compare within the same family that you reveal the fact that it really doesn’t seem to matter.”
Carl Shulman said:
The SSC post overstates in saying genes/shared/non-shared (A, C, and E in the standard model framework) is 50-50-0.
Turkheimer’s “Three Laws of Behavior Genetics” are:
1. All human behavioral traits are heritable. [That is, they are affected to some degree by
2. The effect of being raised in the same family is smaller than the effect of genes.
3. A substantial portion of the variation in complex human behavioral traits is not accounted
for by the effects of genes or families.
#2 doesn’t mean 0 shared environment.
Here’s a site with all the twin studies ever if you want to explore results for different traits:
For the cognitive categories the average reported ACE values are genes 0.468, shared environment 0.177, non-shared environment 0.355 (estimated by subtracting the other two from 1). [Genetic variance share goes up with age and less noisy measures.]
For ‘major life areas’ average reported ACE values are genes 0.297, shared environment 0.277, non-shared 0.426 (estimated as above).
If you use international data, or data from multiple time periods/cohorts, you will accordingly have more shared environmental variance.
Sorry for commenting on such an old thread, but I had a debate with a friend yesterday that reminded me of this post. So, I think Carl Shuman has exactly the right answer here; SSC overstated the case. The Nature paper here: http://www.nature.com/ng/journal/v47/n7/full/ng.3285.html…found an average shared environment effect of 17%, which I think is totally compatible with the lead/crime hypothesis. Later in the paper, they seem to do a test for model parsimony and conclude that by some formal criteria, they should prefer the simpler model that doesn’t include shared environment, but I don’t think that actually means shared environment has no effect.
Carl Shulman said:
Thread opened at SSC, with this:
“Fixing measurement error will boost shared environment effects alongside genetic effects.
Another thing to think about: these variance components are coming from squared correlations. Take the square root to get the correlation coefficient. If common environment accounts for 10% of variance on a trait, then 1 standard deviation of environment can get you a little less than 1/3rd of a standard deviation of the trait. If genes account for 50% of the variance, then 1 standard deviation of genetic factors will buy you a little more than 2/3rds of a standard deviation of the trait.
[Scott] talked about this in another post:
William Adams said:
Actually, evidence-based parenting training has a 0.5 SMD effect size for some conduct issues:
that could be interpreted as taking a kid at the 25 to the 75 percentile I think.
The randomized controlled trials that made the cut for this Cochrane review were all on forms of Parent Management Training:
William Adams said:
Correction: a 0.5 effect size is a increase in 0.5 standard deviations. For instance, it would would take the 50 to the 69.1 percentile.
Pingback: Links for the Week of June 11, 2018 – Verywhen