Friday, June 09, 2006

Why those studying humans must know population genetics   posted by agnostic @ 6/09/2006 08:30:00 PM

NB: Extensively updated

Hell, at this blog, the point is probably taken for granted, but often the non-population geneticists (including me) can lose sight of just how confusing the realm of human beings is when left unilluminated by its basic insights -- and therefore, just how easily fooled you can be if you don't know squat. Elsewhere, DavidB and Greg have pointed out that when one crunches the numbers, sexual selection generally takes a hell of a lot longer than your garden variety directional selection (by roughly an order of magnitude). So, if the change is recent, then most sexual selection accounts of it will be more dubious than those whose premise is that the cause affects males & females roughly equally. But I'd like to take a closer look at a different concept of pop gen that's widely misunderstood and inappropriately applied in my own area of interest, which is psychology -- namely, assortative mating as an account of autism. [1]

The theory was developed by Simon Baron-Cohen, an eminent autism researcher who has advanced our understanding of the condition considerably. In brief, the idea is that there is an essential difference b/w male and female brains: males tend to be more systemizing, while females tend to be more empathizing. Still, there is variance in both groups, so you can find females who score high on systemizing; if they mated w/ fellow high systemizing males, then surely their children would be on average high systemizers as well. On Baron-Cohen's view that autism represents an "extreme male brain" (i.e., highly systemizing), children w/ such a brain are likely to result from two high systemizing parents, say two technology-oriented people from Silicon Valley. As more and more high systemizers pair up w/ each other (i.e., as we see more positive assortative mating), we should observe more and more children w/ autism.

Now, as far as it goes, the argument is fine (not to mention clever). However, it makes a crucial prediction about the pattern of autism at the population level which turns out to be just wrong. Assortative mating only means that "like pairs up with like," which is to say that w/ time the population becomes increasingly segregated into two groups of homozygote genotypes, thereby reducing the proportion of heterozygotes. Now, if we view an individual allele (like A or a) as a marble, and a genotype (like AA, Aa, or aa) as a bin, then assortative mating only moves the A marbles from the hetero bin to the AA bin, and the a marbles from the hetero bin to the aa bin. If we label the aa bin as the "autism" bin, then we've produced proportionally more autistic individuals by assortative mating, since it now contains more marbles -- but we've also produced proportionally more individuals who are the opposite of autistics, whatever they're to be called. So, assortative mating doesn't tell us anything about the change in frequencies of the alleles (number of marbles of one type divided by all marbles), only of the genotypes (how full the bins are, from 0-100%). Assuming there is no difference in fitness b/w homozygotes of type AA and of type aa, then the allele frequencies remain unaltered.

Importantly, when two AA (anti-autistic?) individuals pair up, they won't have an aa child, so we should see autism decreasing among the sub-population whose marbles are drawn from the AA bin. But in reality, autism is on the rise across the board, not just among Silicon Valley nerds; lots of relevant info available on the CDC website for autism. Under assortative mating, that can only happen if aa individuals were more fit -- i.e., got more copies of their genes into the next generation -- compared to AA. But given how crippling autism is in the search to find a mate & raise children, or to raise one's nieces & nephews, this is not a plausible explanation. Less severe but still high levels of nerdiness / systemizing in the latter half of the 20th C West would have resulted in fewer children than average. So, if there is assortative mating on the dimension of empathizing-systemizing, this should have tended to purge alleles implicated in autism from the genepool.

Therefore, we can safely conclude that autism cannot be a genetic condition in the sense that Sickle Cell is. The purported mechanism (assortative mating) would've had the opposite of the observed effect (prevalence increasing across the board). The lower bound of the estimate of current prevalence is about 1 per 500, which is greater than what could be produced by random mutation (about 1 per 10,000 at most), so it cannot be a freak mutation. Thus, we have two broad classes of explanations: 1) the diagnostic checklist has changed, artificially raising the estimates, or perhaps the detection methods have gotten better; and/or 2) environmental insults. If it's 1), then we have nothing more to worry about than before, and if 2), then we've narrowed down the list of where to look for the culprit. True, this wouldn't tell us which environmental insults were the culprit, but still, if you're a policeman looking for a burglar on some street, just knowing which side of the street he's hiding out on will significantly reduce the time, effort, and resources you expend in apprehending him. So, lack of knowledge of the most basic concepts in pop gen can not only lead one astray scientifically, but can also have very real consequences as far as public health and individual suffering are concerned.

[1] I noticed this post from the GNXP archives on this exact topic, but I felt some more flesh on the skeleton was necessary, especially for those of us who aren't practicing population geneticists. Also, the idea is not interred yet, so it bears reiterating.

Extensive Update: There's a good bit of discussion in the comments, the three main points of which I'll try to distill. The first is the complaint that my model is Mendelian, whereas a more realistic model would be polygenic. The reason I chose the simplified Mendelian model is that it preserves the relevant property of the polygenic model: namely, under assortative mating, we expect an increase in the phenotypic variance, not necessarily an increase in the mean or median. This is the relevant property b/c I claimed that empirical findings suggest that autism is on the rise across groups, not just among high systemizers / nerds. More on that in one second. Now, there are additional wrinkles that come w/ the more realistic polygenic model -- namely, a change in the mean or median could happen due to the non-additive effects of the genes involved. The effect that assortative mating has on the dominance variance is typically small enough that we can ignore it. That would mean, then, that though we wouldn't expect autism to rise among the left half of the "nerd curve" given the additive effects of the separate genes, nevertheless the interaction between these genes (i.e., epistatic effects) might produce something unexpected.

I searched PubMed for 'assortative epistasis' and 'assortative epistatic' and found nothing on unusual conditions, diseases, etc. -- just the expected articles on how assortative mating complicates calculating heritability, and so on. I also Googled ' +"assortative mating" +disease +epistasis' and likewise with 'epistatic' and again found no data points. So, while possible, there appears to be little precedent for such a scenario -- although I'm not a student of diseases and welcome any corrections to the effect that assortative mating plus epistatic effects result in a mean / median increase in some human genetic disease. By contrast, there is ample documentation of diseases similar to autism in prevalence & crippling-ness that are due to environmental factors.

The second big issue is: is autism really on the rise across all groups? The relevant group here is social class. There is an obstacle to overcome: studies which show that risk of autism rises as SES increases could reflect better access to health care services (diagnoses & treatment) among the more affluent, their higher IQ allowing them to better spot potential problems in their children, higher IQ making them more frantically concerned over their child's future, and so on. So what we want are data sets that can tease these variables apart, or that at least are exhaustive enough that selection bias in reporting will be minimal. Here follow six extensive enough studies that cast doubt on the putatively high link b/w parental SES and autistic children.

1) The most recent large, national study on the risk factors of autism analyzed data from Denmark (done by researchers at University of Aarhus in Denmark, and at Johns Hopkins and CDC in the US). Because health care is free and easy for all in Denmark, because only publicly hired people can diagnose autism, and because these data along w/ other demographic data on patients are collected into national databases, these researchers had access to essentially all cases of diagnosed autism in the country where the child was born after 1972 and was at risk of developing autism before 2000 -- a nice longitudinal study, rather than a snapshot.

Analyses showed no statistically significant association between risk of autism and weight for gestational age, parity, number of antenatal visits, parental age, or socioeconomic status. Results suggest that prenatal environmental factors and parental psychopathology are associated with the risk of autism. These factors seem to act independently.

2) Staying within Scandinavia, this study looked at SES factors of parents of autistic vs non-autistic children in Goteburg, Sweden, who were born from 1962 to 1973 -- so, before the autism epidemic, and before the expanded circle of definitions that included less severe forms, and thus representing what are now called "classic autistic" cases.

The distributions of social class were almost identical in the infantile autism group and in the random group. With respect to some other social circumstances the two groups were very similar. Thus, the present results lend no support for the view that autistic children tend to come from high social classes.

3) Moving next to France, this 1997 study identified autistic children born between 1976 and 1985, so also probably classic autism pre-epidemic.

One hundred seventy-four children (mean age: 11.6 years) with autism were identified. The prevalence rate was 5.35/10,000 (16.3/10,000 if other pervasive developmental disorders are included), with no difference according to geographical area or social class.

4) Suspecting a selection bias in studies claiming parental SES is a good predictor of bearing autistic children, Lorna Wing (who popularized the research on Asperger's Syndrome) examined the profiles of parents of autistic children in London in 1980, so as w/ the Swedish study, probably representing classic autism pre-epidemic.

Children with typical autism, other early childhood psychoses and severe mental retardation without autistic behaviour were identified in an epidemiological study in an area of South East London. The social class distribution of their fathers was examined and no significant differences were found between the groups, nor in a comparison with the general population of the area. Fathers of children with autism and related conditions referred to an out-patient clinic with a special interest in autism, mostly at their own request, and fathers joining the National Society for Autistic Children, were of higher social class than both the average for England and Wales and the fathers of the study children. Joining the NSAC during its early years, and keeping up membership were also linked with higher social class. The findings supported the view that reports of a social class bias in autism may be explained by factors affecting referral and diagnosis.

5) Not long afterward, in 1982, Tsai, et al. did the same thing w/ parents of autistic children in Iowa.

The social class distribution of fathers with autistic children attending a locally well-known and state-supported modern autism program was examined and was compared to the social class distributions observed in a nonautistic, mentally retarded population, in children with other psychiatric disorders, and in the general population from which the present autistic sample was drawn. No significant differences were found among the groups. The findings supported the view that if studies are not biased by certain selection factors outside the autistic child's clinical picture and diagnosis, and if services become better known and readily available, then no differences in social class distribution between autistic and nonautistic groups occur. The results suggest that social class is not an important factor in the origin of autistic syndrome.

6) A 1999 meta-analysis concludes that, "Social class and immigrant status did not appear to be associated with autism."

I think these studies suggest that the link between parental SES and autism is exaggerated, almost certainly in the case of classic autism -- the crippling, from-another-planet version -- and likely in the slightly less severe forms as well, if not in perhaps the highest functioning versions where the child appears to be simply "really nerdy." The perception that autistics are more likely to come from nerdy parents is likely mostly a reflection of such parents having better access to health care to diagnose and treat their child, having the requisite higher IQ to recognize problems w/ their child and to obsess over its well-being, and having better access to the popular media by which to broadcast their concerns.

Third, and lastly, is the contention that, even if ultimately due to an environmental insult of some kind, the fact that genetics play a strong role argues in favor of considering autism a (perhaps partially) genetic condition. See the references here, but most estimates ballpark the heritability of autism at around 0.7, which is pretty high even if less than 1. One puzzle to ponder first is the heritability of known infectious diseases. The following are the monozygotic and dizygotic twin concordance rates for such diseases, taken from Vogel & Motulsky's Human Genetics (p.237):

Measles - MZ = 97, DZ = 94
Scarlet Fever - MZ = 55, DZ = 47
Pneumonia - MZ = 32, DZ = 18
Tuberculosis - MZ = 53, DZ = 21
Leprosy - MZ = 59, DZ = 10

We can see that, aside from the presence of the pathogen known to cause the disease, there are also genetic factors involved (as shown by differences b/w MZ and DZ twins), as well as chance factors (as shown by MZ rate far less than 1 in most cases). Presumably, the role of genes is that of susceptibility to initial infection, weakness in attacking the infection, allowing an easier-to-navigate route to the site where the pathogen does its thing, and so on. Given the complexity of such a role, there are presumably many genes involved. So, the picture is quite like that of autism (and the MZ & DZ concordance rates of autism are just like those of leprosy). However, no one classifies measles, scarlet fever, pneumonia, tuberculosis, or leprosy as genetic. The reason is that classification of diseases follows a different modus operandi from other empirical inquiries. For example, philosophers view establishing "causation" as determining what the necessary and sufficient conditions are for X -- that is, a potentially multi-item checklist so that when all items are checked off (no more, no less), then X happens. Statistical scientists talk about which variables account for what portion of the total variance among outcomes, e.g., what variables (race, sex, IQ, etc.) account for individual differences in educational outcomes?

That's not how it's done for diseases, though. In this case, a disease is classified as infectious if the infection by pathogen(s) is a necessary condition -- it doesn't have to be sufficient. It seems a strange departure from the stricter defintions of philosophy and statistics, but think of it from the germ's point-of-view: you're trying to infect some host in order to reproduce. All sorts of other things have to line up in order for you to succeed -- they have to inhale the right gust of air or have sex w/ the right person, while you have to turn left instead of right in their gut, dodge their immune system response, etc. Hell, you may even have to rely on the host having a particular genetic profile for the job to get done right. But all of these things are just one or another potential obstacle that threatens to thwart your best efforts to exploit the host's body for your reproductive purposes.

There is another, pragmatic reason why the definition is relaxed when classifying a disease as infectious -- on a purely analytical level, there may be various factors involved that interact in complex ways, but in the real world, we want to isolate the factors that we can easily change in order to prevent disease. Complex suites of genes are notoriously more difficult and expensive to alter here-and-now on a population level compared to vaccination, antibiotics, sanitizing the water supply, spraying DDT, and so on. That's why a disease is classed as infectious if removing the pathogen prevents the disease (i.e., is just a necessary condition) -- it answers the pressing question: Is it possible to get rid of this stuff by the aforementioned cheap & easy methods?

So that's about it for the wrap-up. I'm sure this is longer than the original post, but I try to strike a balance between being too wordy and too vague, since I neurotically obsess over how much of a blabbermouth I come off as! From now on, though, I'll err more on the side of wordiness & depth.