Parenting and the Non-Shared Environment

October 14, 2013 by · 9 Comments

There is a major thread in the parenting literature that claims, in short, that parenting (short of outright abuse) has little to no effect on adult outcomes that we care about. Two exemplars in this category include The Nurture Assumption and Selfish Reasons to Have More Kids. This claim is largely based on twin and adoption studies, which allow us to attribute the observed variance in traits to genetic, shared environment, and non-shared environment factors. The general pattern is that lots of traits are about half genetic and half non-shared environment, with little contribution from shared environment. There is a major embedded assumption in this line of reasoning: that parenting effects are mainly in the shared environment. It turns out that this assumption is not a particularly good one.

A review article on research into the non-shared environment was released about two years ago, and it provides some fascinating data on this subject. I will quote some sections at length, but the entire paper is worth reading. Here is the first excerpt of interest:

For example, parental warmth and control have been assessed in hundreds of studies as the two ‘super-factors’ of parenting which are then correlated with children’s outcomes. These studies have traditionally assumed that parenting is a shared environmental factor in the sense that only one child per family was considered and parent–child associations were analyzed ‘between’ families. However, by studying more than one child per family and by targeting parenting that is specific to each child, it is possible to investigate the extent to which parents’ warmth and control differ for children ‘within’ families. Research of this kind has shown that parents do treat their children differently. If you ask parents about their differential parenting they report only modest differential parenting (sibling correlations of ~0.70) but if you ask children about it you might think they were raised in different families (sibling correlations of ~0.25). Observations of parent–child interactions support not the parents’ but the children’s view (sibling correlations of ~0.20).

In short, parents think they treat their kids the same… but the kids think the parents treat them differently, and outside observations would support this claim. If anything, the outside observer sees slightly more unequal treatment than the kids themselves do. This indicates that the vast majority of parenting effects would show up in the non-shared environment.

That said, modern study designs have indeed allowed us to decompose the known sources of non-shared environmental influences. Here is the relevant data from the paper:

The proportion of total variance accounted for in outcomes such as adjustment, personality and cognition was 0.01 for family constellation, 0.02 for differential parental behaviour, 0.02 for differential sibling interaction and 0.05 for differential peer or teacher interaction. Moreover, these effects are largely independent and they add up to account for 13% of the total variance. If non-shared environment accounts for ~40% of the variance in these domains, we could say the cup is already more than one quarter full.

The most important takeaway here is that we really don’t know where most of this variation comes from yet, and it’s likely that some of the remainder is still unobserved parenting/family/peer/teacher effects. Of the one-third that we can explain, it seems like about 20% of that effect is directly from parenting, a full half of that effect is from inside-family dynamics, and the remainder is peer and teacher effects. These statistics of course come from a large number of people from the general public, and it’s not clear how much we can generalize from this. Do parents have any effect on how siblings treat each other, for instance? What happens if the child is homeschooled? Do parents influence the kids’ choice of peers? This seems like a lower bound on what is possible from parenting techniques, in my opinion.

I will also note that I’m relatively optimistic that genetics play such a strong role – I picked Divia for good reasons. Insofar as we are largely an expression of our genes, these genes have led to good results. A child that is mostly the average of the two of us would be someone I very much want to exist.

This does, of course, leave quite a lot of variance unexplained. The author actually concludes by saying:

The message is not that family experiences are unimportant but rather that the relevant experiences are specific to each child in the family, not general to all children in the family. However, my main conclusion has to be that the key question largely remains unanswered: why are children in the same family so different?

This non-shared environment factor by construction includes basically everything that happens in life. That means all of the random perturbations that happen over a lifetime might account for much of this variance:

It is also possible that non-systematic factors, such as accidents and illnesses and other idiosyncratic experiences, initiate differences between siblings. Compounded over time, small chance differences in experience might lead to large differences in outcome, as discussed in the next section. The 1987 paper concluded that ‘one gloomy prospect is that the salient environment might be unsystematic, idiosyncratic, or serendipitous events such as accidents, illnesses and other traumas, as biographies often attest’. It is striking how often biographies and autobiographies point to chance as a tipping point to explain why siblings are so different. Illnesses and accidents also feature in interviews with parents of discordant identical twins in explaining why they thought their children differed. Support for chance as a source of non-shared environmental influence comes from quantitative genetic research that suggests that non-shared environmental effects are trait specific and age specific. That is, non-shared environmental effects on one trait are largely uncorrelated with such effects on other traits and non-shared environmental effects at one age are largely uncorrelated with such effects at other ages.

Clearly the role of random chance and idiosyncratic events plays a large role in our narrative of these individual differences. The statistical evidence, that the non-shared environmental factors are largely uncorrelated with one another, is an interesting finding. This doesn’t necessarily mean it’s entirely random chance, it means that any intervention we do is unlikely to fully generalize. Depending on your perspective, this could be a good or a bad thing. It does suggest that everything we does has lots of little, cumulative effects. There may be a lot of independent factors to get right, each of which has a small but positive impact on our kids’ lives. But looking on the bright side, that also means that if you screw one or two things up, you still get another chance, and the net result can turn out quite good.

I definitely wonder how much research like this affects people’s willingness to invest more effort in their children. I am already quite motivated to provide Lydia with a good life, as I best understand it. It seems to me like my interactions with her at the very least impact her experience right now, which is something I definitely care about. Intuitively, it seems crazy to me that we might not have a major impact on our kids. Reading this paper does make me think that the research is more in line with my beliefs than I felt like was portrayed in the aforementioned literature. I don’t believe that parents are the only effect on a child’s life by any means, but it certainly seemed like one in mine, and from working with many people it seems to be quite important. Realizing that parents do treat their kids differently, and that this produces different results, along with the combination of a huge number of individual life events, is my main takeaway from this paper. We can’t completely determine the lives of our children – and maybe we shouldn’t be able to do so – but we clearly have an impact, and I want that impact to be as positive as possible.

Filed Under: Blog · Tagged: , ,
  • razibkhan

    interesting post.

  • pnin1957

    A substantial proportion of the “unshared environment” is probably not “environment” but developmental noise, i.e., it cannot be controlled.

  • Paul Conroy

    Will, I agree with a lot of what you say, and coming from a family of 6 kids, can say that parents do treat their kids significantly differently.

    Additionally, one of the things parents can do is help influence their kid’s choice of friends, something I consciously do.

  • There is a major embedded assumption in this line of reasoning: that parenting effects are mainly in the shared environment. It turns out that this assumption is not a particularly good one…

    This indicates that the vast majority of parenting effects would show up in the non-shared environment.

    I wouldn’t say that. This idea was investigated and rejected by Steven Pinker in The Blank Slate, and covered extensively by Judith Harris No Two Alike. The key problems are these:

    1. It requires that there are no across the board differences from one set of parents to another. To see why, imagine parents did have some effect. Even if there were differences in the exact treatment each child received, there are going to be systematic similarities with the way each set of parents treat all their children. If such an effect existed, it would turn up in the shared environment, since kids growing up together would be impacted by these across the board similarities. But the shared environment influence is in fact negligible.

    2. It requires perfect crossover interaction. Let’s say you assume that parental effects – whatever their across the board similarity for a given set of parents – had totally idiosyncratic effects on children. Then in order to explain the null effect of the shared environment, the sum of these effects on children’s traits had to be exactly zero. Any number of children exhibiting one sort of effect would have to be balanced by an equal number of children exhibiting the exact opposite effect. That is a rather large stretch.

    Occam’s razor and….

    The most important takeaway here is that we really don’t know where most of this variation comes from yet, and it’s likely that some of the remainder is still unobserved parenting/family/peer/teacher effects. Of the one-third that we can explain, it seems like about 20% of that effect is directly from parenting

    No, actually not quite. Ignoring the methodological problems of trying ascribe such correlations to the effects of parenting (no way to prove causation), the simple fact is that correlations that low are well below the level statistical significance in these studies. In short, they are overwhelmed by noise. As such, they don’t really tell you anything.

    It is possible, if indeed likely, that a good bit of what makes up the unshared “environment” – beyond which that’s not simply due to measurement error – may stem from developmental noise. This is addition to a variety of factors out of our “control”, like de novo mutations (which is technically genetic, if not exactly “heritable”), and infectious agents and other physiological insults. It’s very hard to stick parents into that equation despite the wishes of many.

    • WilliamEden

      Thanks for posting! I haven’t read those books yet, and I am curious to hear their take on this.

      I will note that depending on what traits you’re measuring, the effect of shared environment is small, but not zero. So clearly something is being picked up there. I don’t think that #1 is much of a stretch, given the lack of correlation in treatment of different kids, as I quoted from the paper.

      I don’t think I quite follow you on #2. If their impacts are totally idiosyncratic, that would show up in non-shared environment, that’s the entire point of how it is constructed. I think you’re still assuming there is some correlation in how the parents are treating their kids. In that case, for an across the board difference to show up as a null result, the kids would indeed have to have opposite effects to the same kind of parenting technique. Maybe you can point me to a specific section of the books you cite to clarify the statistics involved here?

      Another interesting thought: we might be able to back out an order of magnitude estimate of how much non-shared environment is parenting, by attributing the impact of the correlated parenting behaviors to the shared environment component, and then using that effect size to very roughly estimate the effects of the uncorrelated behaviors…

      • I will note that depending on what traits you’re measuring, the effect of shared environment is small, but not zero. So clearly something is being picked up there

        Studies that turn up an effect of the shared environment tend to fall under one or more of the following categories:

        1. They have small sample sizes. The problem here is obvious.

        2. They are MZT/DZT (identical twins raised together vs fraternal twins raised together) studies which are by their nature incapable of actually measuring the shared environment. They can only estimate it. Their estimate can be thrown off by a variety of things, such as assortative mating between parents or the degree of non-additive genetic variance that contributes to the trait.

        3. The are on young children, where shared environment effects do appear (but disappear in samples of older individuals).

        Large behavioral genetic studies/meta-analyses converge on a zero shared environment.

        I don’t think I quite follow you on #2. If their impacts are totally idiosyncratic, that would show up in non-shared environment, that’s the entire point of how it is constructed. I think you’re still assuming there is some correlation in how the parents are treating their kids. In that case, for an across the board difference to show up as a null result, the kids would indeed have to have opposite effects to the same kind of parenting technique. Maybe you can point me to a specific section of the books you cite to clarify the statistics involved here?

        Yes, you’re right, it does make the assumption of across the board similarities. However, there is no question that such across the board similarities in how each set of parents treat all their children must exist – because to say otherwise would imply that there are no systematic differences between different sets of parents.

        As such, in order for the unique environment to be boiled down to interactions, such interactions would need to be perfect cross-over interactions, that is, with no main effects. As Steven Pinker put it (The Blank Slate, p. 388):

        Here is what would have to happen if the effects of the unique environment are to be explained by an interaction between parents and children … A given practice would have to affect some children one way, and other children another way, and the two effects would have to cancel out. For example, sparing the rod would have to spoil some children (making them more violent) and teach others that violence is not a solution (making them less violent). Displays of affection would have to make some children more affectionate (because they identify with their parents) and others less affectionate (because they react against their parents). The reason the effects have to go in opposite directions is that if a parenting practice had a consistent effect, on average, across all children, it would turn up as an effect of the shared environment. Adopted siblings would be similar, sibs growing up together would be more similar than sibs growing up apart — neither of which happens. And if it was applied successfully to some kinds of children and was avoided, or was ineffective, with other kinds, that would turn up as an effect of the genes.
        The problems with the parent-child interaction idea now become obvious. It is implausible that any parenting process would have such radically different effects on different children that the sum of the effects (the shared environment) would add up to zero. If hugging merely makes some children more confident and has no effect on others, then the huggers should still have more confident children on average (some becoming more confident, others showing no change) than the cold fish. But, holding genes constant, they don’t. (To put it in technical terms familiar to psychologists: it is rare to find a perfect crossover interaction, that is, an interaction with no main effects.)

        The notion that parental influences exist, but the factors that vary among parents in how they treat their children across the board must either have no consistent effect or must perfectly cancel across children such that it gives the appearance in the data that it has no effect at all requires us to entertain many more causal entities that the simpler idea that they just have no effect at all.

        It’s true that to reliably detect a small, but real effect of the shared environment, you would need a large study with many twin/adoptee pairs (>7,000). So let’s entertain for a moment that the shared environment impact is real, but small. The moral of the story would then be that parental impact is either weak or has largely cancelling effects across many children. That information is of little use to parents, who are the people who want to know about what they can do for their kids. It would only have an academic interest – which is indeed reason enough to investigate it – but it would have little practical significance.

        In any case, there is additional trouble for the idea of significant parental impact: it has to contend with the absence of birth order effects. Thorough analysis has failed to find any systematic differences in children due to birth order. Birth order is one case of systematic, non-genetic differences in the home environment of children. It’s hard to reconcile the existence of parental effects with the failure to find anything when looking at a reliable systematic difference in the parental environment.

        Now, to be completely fair to parenting, some commenters on the matter, such as Steve Sailer, posit that parental effects are largely unimportant when looking at higher-SES, Western parents. The difference from one set of middle-class Western parents to another set may not matter much, but the difference might matter if we compared White Western parents to say poorer Latin American parents. On that we have to say the data don’t yet rule such a difference out. My own suspicion is that even in those cases, we will find that the difference – if any – stems from environmental impacts that only act in a negative way; e.g., we know that childhood malnourishment can stunt IQ, but nutrition beyond what’s adequate won’t raise IQ. In the same way, children from poorer families might miss out on critical developmental inputs. As well, they may miss out on key opportunities to achieve, which itself is perhaps more a function of the outside the home environment in which these poor children happen to dwell.

        You raise some good points in your post, but these were considered and generally ruled out. :)

  • WilliamEden

    Adding a somewhat tangential point that says a lot about the dangers of interpreting these kinds of studies:

    http://wholehealthsource.blogspot.com/2013/10/sleep-and-genetic-obesity-risk.html

    This plots the decomposition of genetic, shared, and non-shared environmental factors of obesity… plotted across hours of sleep/night. And the ratios change drastically across this factor!

    The moral of the story is that lots of factors can potentially affect this decomposition, and if we’re not measuring the relevant ones we will completely miss these relationships.

    • I’m just seeing this today. I left a comment over at Guyenet’s blog explaining the problem here:

      Yeah, I wouldn’t put too much faith in this study, yet. For one, the main reason their shared environment term seems to have increased in the long sleepers is because their sample size got pretty small there (47 pairs). In general, unless you have a huge number of subjects, the confidence intervals around the shared environment are too large to for your study to be useful. In this case, as per usual in behavioral genetic studies, their confidence intervals on the shared environment, even in their “interaction” hypothesis, went down to near 0.1, or insignificance.

      Of course, it’s worth mentioning that you can’t actually measure the shared environment from MZT-DZT studies – the extraction of the shared environment term in these studies is largely a fingle factor, and is itself confounded by assortative mating.

      I’d be more impressed if these results held up in a much larger sample.

      EDIT: I also added:

      In addition, and I should have said this before, you can’t actually do what this study is attempting. Lumping subjects by phenotypic similarity is really taking the magic out of behavioral genetic studies: if you try to decompose ACE from twins for which you’ve already sorted by phenotype, you are going to confuse heritable effects (say, shared non-additive variance) with shared environment; the study will lose the power to distinguish the two.