Thoughts on the BGI IQ study

Share on FacebookShare on Google+Email this to someoneTweet about this on Twitter

I’ve been following the development of the BGI study on IQ pretty closely. I wanted to note two main caveats people should be aware of with regard to its methodology.

First, as with any case-control study, volunteer bias will be an issue. If the cases are a certain class of very smart people, rather than a representative sample, then genes peculiar to that class of smart people will show up as hits. The BGI study is choosing people who are more math than verbal-oriented; will math-specific genes show up as general intelligence genes? Other confounds along these lines are possible– PhD genes, Ashkenazi genes, curiosity in new study genes, etc..

Second, because the study doesn’t completely control for family environments (possible only by comparing siblings to each other), gene-environment correlations and interactions can cause problems as well. For example, suppose that high IQ parents also confer better environments for their children. Then the IQ gene effects will get an extra “boost” from that environment.

None of this is to downgrade the awesomeness of the BGI study. It should be viewed as an important step in resolving the nature vs nurture controversy. Overeager journalists and bloggers are urged to wait a few more years before we finally resolve the IQ debate.

4 Comments

  1. When is someone going to do a study on Ashkenazim to find their high intelligence genes?

  2. @James

    Greg Cochran on the Straight Dope board discussed some attempts by scientists to test his theory on Ashkenazi intelligence. Unsurprisingly, politics tends to get in the way of research.

  3. I think that their initial phase was not that ambitious. They were expected to only find a few of these alelles. They expect the total number to be in the hundreds of genes involve in various aspects of IQ. Once they have enough members sequenced along with their biometric data, they can comb the data looking at specific mental traits and find genes involved in these traits. As you said, we need to wait a few years.

  4. I tried posting on Steve’s blog but apparently it was deleted. I commend Steve on his efforts with such an endeavour, certainly more of these types of studies will be carried out in the near future to better reinforces the genetic associations with intelligence, as it is intuitively obvious. The subject of intelligence is rather important and I felt compelled to post my comments since I detect a slight incongruence with the title of the study which seems to be on “intelligence” and the screening for the cohort sought which seems to be based on scholastic achievement;

    What is also stated in the video clip is that the subject are selected for intelligence above the stochastic mean as follows
    > 3SD to
    > 4SD in math

    The study is named “Gene-Trait Association Study of intelligence”. The accepted operational definition (really, up to this point the only definition that makes any sense) of intelligence is essentially the g-factor and tests which “load” on g are best measures of “intelligence”.

    [[Tests]]

    Since the tests or competitions stated as the requirements will screen for the cohort on the desired “intelligence” range then the tests itself are important and by definition should be acceptable measures of g but not only that, be able to measure g at > 3SD.

    Let’s look at the SATs. If you look at the study carried out by Detterman & Frey on the SATs post and pre 1994 the result of the regression analyses are 2 equations to convert SAT scores to IQ (for both pre and post 1994). You will notice several things;

    i. The Authors used test subjects who sat for both the ASVAB and the SAT (pre-1994) and carried out a regression analysis. They determined the correlations of about 0.82 with the IQ test and also derived an equation to convert the SAT scores to IQ.
    ii. The Authors administered the RAPM to subjects who sat for the SAT (post-1994) and carried out a regression analysis. They determined the correlations of about 0.483 which was a bit low due to range restriction, but corrected it the inclusion of other subjects and ended up with r= 0.72. The authors then derived equation 2 to convert the SAT scores to IQ.

    Point 1: g-factor

    In both studies no factor analyses were carried out on the SAT test items to extract 1st order factors (e.g. from the math-only items) or 2nd order factors (e.g. between the math and verbal items) before extracting a higher order facto. So, the g-load of the test is inferred by correlations to standardized IQ tests (Ravens and ASVAB) and is indirect.

    Point 2: ceiling

    Eq. 2 (after 1994) = 0.095 x SAT-M + 0.003 X SAT-V + 50.24

    = 0.095 x 800 + 0.003 X 800 + 50.24= 128.64

    Std Error = 9.79 so the ceiling is 138.4.
    3SD = IQ 145
    The re-centered SAT does not have enough ceiling. In fact even American Mensa which has a relatively low cut-off of 2SD does not accept the SAT.

    Point 3: math

    If you look at equation 2, the SAT-V hardly loads on g at all as the constant for SAT-V is 0.003. This means that only SAT-M seem to be g-loaded and counts towards IQ.

    Point 4: Competitions:

    Physics – are there any studies carried out with high repeatability and rigorous multivariate analyses to show that the material used for physics Olympiads is g-loaded? Ditto for Informatics, the Putnam Competition?

    [[Cohort & Apparent Math Focus]]

    “g” (other subscribe to 2-factors gf and gc) is the general factor that explains the largest variance in many types of sub-test scores/abilities not narrow abilities like math only. If some of the subjects are selected using the SAT, Math competitions, they inadvertently have a math tilt and the study will be doing the GWAS on a cohort with very specific abilities and in the end may be identifying specific alleles (or a polygenic association e.g. thousands of SNPs which could explain the difference with the control group instead of several marker genes) that are common for mathematical ability but not “g” since “g” is a general factor that explains the largest variance in many types of sub-test scores/abilities.

    [[Measurement of g at > 3SD & Spearman’s Law of Diminishing Returns (SLODR)]]

    SLODR manifests as we go higher up the ability scale, i.e. the inter-correlations between sub-test scores drops dramatically (as studies by Evans, Brand, and Detterman have shown and numerous studies carried out by others have repeated these findings) since at higher levels the specificity differentiation “kicks” in, and hence less and less of psychometric g and consequently more of specificity is measured, no matter what the test (SAT, GRE e.t.c). For instance using the Otis Lennon Evans found that at the top, the average correlation was 0.35; at the low end, the average correlation was 0.50. In the slides (5) RT or ITs was briefly mentioned. In fact IQ test correlations with IT are about 0.5 for an average ability group but can drop to as low as 0.3 for a higher ability groups (about IQ115-see Chris Brand). So, even for absolute measures (IT/RT) Spearman’s Law seems to hold. I haven’t seen a paper on what the IT correlations are at 3SD or even 4SD but am pretty sure it will be very small of which the statistical significance will be brought into question.

    I surmise that the test and competition requirements do not seem to have a proper psychometric basis as screening tools for “intelligence”. I will be happy to accept an alternative view if supported. A more accurate name of the study perhaps should be “Gene-Trait Association Study of Mathematical Ability” or “Gene-Trait Association Study of Scholastic competition” not “Intelligence”.

    Sorry a rather long missive.

Leave a Reply

a