User Tools

Site Tools


Causal Accounting

Innate: in the child implies not malleable, but there's a range:

  • developmental chance
  • genetic
  • epigenetic

This fades into factors that would normally be considered environment:

  • prenatal environment
  • ubiquitous environmental factors
  • self selected environment

Nurture implies action by parent, teacher, etc. Randomness isn't nurture, though it may be someone's responsibility to protect from random effects, if possible. Acts of the child himself or of other children also aren't nurture, since they aren't properly responsible.

Is heritability interpretation tricky because it has to do with accounting for variability, and not the mean? Caring about difference is intuitive.

Environmental Multiplier

What is the right way to do causal accounting when causes act in parallel?

  • intuition is that there is no clearly right answer, but this might not be true if we concede that practical causality is always about differences (manipulation) and not “cause of existence” (sufficient cause).
  • We do have an emergent (non additive) interaction, but we can still say how much output variance is contributed by the two terms.
  • Other intuition is that I prefer to attribute to genetic. One story is that the kind of flat environment that could completely undo the multiplier is unreasonable or undesirable in modern life. If we're serious about reducing wealth inequality then we need social control via norms and laws. Competition is unavoidable without huge social control, and the resulting society would be outcompeted by others more tolerant of inequality.

- if the crucial property of causality is the possibility of other outcomes in the absence of these factors, and all are preconditions, (necessary) then it does make sense to say that both are causes. - I'd say that the ease of modification of a precondition isn't usually relevant to whether it's a cause, though that's an obvious practical consideration.

The “gene for basketball” What happens if you make the environment uniform? Heritability increases and all causation becomes innate. Yet variation in outcomes decreases (?) because there is no longer any environmental multiplier (GxE interaction). But if the environment is uniform then there is only one niche. Then there *is* a best kind of person. The winners will be those adapted to the particular chosen environment. In fact, the assumption that flattening the environment will reduce outcome variation presupposes that the GxE term is smaller than the additive E. Isn't that a contradiction? How can GxE interaction both increase and decrease heritability, both increase and decrease outcome variation? I think that this might depend on the nature of the interaction, eg. convexity. Simulate? Flynn is definitely saying that environment variation is increasing heritability, magnifying small differences. So flattening would have to reduce. Competition can certainly magnify small differences.

What is the significance of shared vs. non shared environment for the multiplier effect? Shared environment is defined by family, but is larger. If the environment is flat, then it's all shared? You could say that, but the shared environment seen by family studies is the variance of families from the norm. If something is constant then it's part of the “all other things” that are equal, and is invisible. Specifically in twin studies, the multiplier says that MZ non shared is in fact partially shared because MZ end up selecting into similar environments. This increases their similarity. But if environment were flat, then DZ would have no option of different environments, so would be more similar too. Plausibly this would reduce MZ DZ difference. It does seem likely that E flattening would reduce outcome variation too. I guess that is the Flynn theory. How do we square this with Turkheimer's observation that poverty reduces heritability? What does that look like? It seems that poor kids have greater variation in shared effects. Possibly not that variation in shared environment is greater, but that the mean has decreased so that variation dips below “good enough”. This increases correlation of DZ (and presumably MZ to a lesser degree). At the extreme it swamps the genetic effect. Presumably also greater total variation. The paper I found was 7 year olds. It seems likely heritability will be greater for older poor.

So increased variation in the shared environment reduces heritability (Turkheimer), but increased variation in the non shared environment could increase heritability if MZ can select into similar environment (Flynn). Presumably non shared could reduce heritability too by swamping genetic effect.

Though we may think of self selection as being voluntary, that isn't necessary. MZ could select into jail. All that it requires is a reproducible GxE interaction.

Suppose 1/2 of people are given lobotomy. If this is random, reduces heritability. If based on hair color, increases. Either way, this is a non shared effect. Note that a capricious environment (either shared or non shared) reduces heritability and increases outcome variation above flat.

Effect of a responsive environment depends on how it responds to individual variation. Responsive doesn't mean nurturing, just consistent. In the west, environment likely responds to increase heritability and outcome variation. But consider a reverse dominance culture with strong conformity pressures. Outcome variation and individual differences are both suppressed. Every individual is constantly being pressured according to their actions and traits (responsive), but the socially controlled outcome variation is far less than the flat environment. Even for outcomes that aren't socially controlled, the flatness of the environment will minimize any multiplier effect. One way to look at the responsive environment is whether it is positive feedback or negative feedback. Social control is negative feedback, trying to maintain behavior at a desired setpoint. In contrast, social reinforcement is a positive feedback, trying to maximize a desired behavior. All cultures control some behavior, but there is a lot of variation in the encouragement given to high achievers and the degree to which behavior can be neutral and voluntary.

WEIRD cultures maximize variation by minimizing social control and tolerating positive feedbacks. Competition creates positive feedback, both rich get richer and also the “negative” cycle of failure perpetuating itself.

There is a question here of whether self selection into niches is a zero-sum game. Is all life merely competition for the desirable niches? But if people are innately diverse, then they don't entirely agree on what the desirable niche is. IMO this tends toward sophistry, but it is good to provoke thought about non zero sum and diversity. The flat environment may be a useful thought experiment, but as a policy proposal it's a straw man. Not only would we have to impose uniform material wealth, we'd need strict social controls to prevent behavior variation from causing nonuniformity in the social environment. There are cultures that do that, but most people in the modern world wouldn't really want to live there.

So would increasing environmental homogeneity increase heritability or not? Because the whole environmental multiplier theory assumes that inhomogenous environment creates illusory heritability increase. Or at least insofar as the GxE term dominates it may be hard to assign cause. But this isn't true with significant differences, I think. One way to think about the GxE term is as an emergent effect. Pearl's rules allow a non causal correlation with emergence.

Consider movie stars. Why do they even exist? All value is socially constructed, so it isn't possible to truly level outcomes without social control.

What does the environment need to be like to have a multiplier? Competition, or even just being nonuniform. The only thing that would clearly prevent a multiplier is an environment where the outcome would be completely unaffected by any sort of behavior. In other words, behavior has no consequence at all, at least w.r.t. the outcome.


f(g, e) = Ag +  Be + d(g, e)

Mathematically, we'd be happy to say that if one term dominates, then f is mainly that term. We might drop the others, as an approximation. In a sensitivity analysis, given particular levels of variance in g and e, we could attribute output variance to one factor. Of course with more general nonlinearities it gets complicated, but with g * e it's just variation / mean, for each factor. We are really talking about the variance for the additive terms too, but the A/B ratio comes into play then. Of course models like these are ludicrously simple compared to the thing modeled, but that doesn't mean they can't give some insight.

How is the environmental multiplier different from the search for authenticity? It could be that increasing heritability with age isn't so much a matter of locating a supportive environment (in the face of competition) as simply figuring out what pleases you, despite the environment. A more supportive environment would just speed this process. Of course these two would both operate to some degree. A question is whether they would make different predictions. This might be in the case of environment interventions, but wouldn't you do different interventions?

This is the flat vs. responsive environment again. But suppose you can eliminate scarcity without eliminating diversity of options. Or at least some practical constraints are eliminated. Free college, all school teachers good, etc. So this question is whether the outcome variation is due to unfair or avoidable scarcity, or just individual diversity. (Both)

The answer is that there *are* different predictions. If diversity of outcome is purely due to individual diversity, then increasing education availability would have no effect. That is a classic naturist argument. Of course no effect is hardly likely. Reducing the user cost of education will increase usage, and this will surely have multiple effects, some of which even resemble social capital theory.

Of course if everyone is the same, then a responsive environment would respond the same to all, and the environment would be flat.

analysis/nature/causal_accounting.txt · Last modified: 2014/01/31 11:15 by ram