Kees Jan can't

Kees-Jan Kan, a young Dutchman, has recently rediscovered one of the most basic facts in IQ testing: That it's easiest to detect IQ differences if the people you are studying (Ss) have a common background.  So if the Ss are all in the same class at school, for instance, a vocabulary test (finding out how many hard words they know) will give you a quick and easy way to sort them out.  And you will find that the guys who know lots of words are also good at a whole range of puzzles, even mathematical ones.

So a common background optimizes your chances of assessing IQ accurately. And to be a bit technical, vocab loads highly on 'g' (the general factor in intelligence), meaning that, where it can be used, it is a powerful predictor of other abilities.  Vocab is however convenient rather than essential in IQ measurement.  Tests designed for use among people who do not have a common background (such as the Raven PMs) don't use it but still work perfectly well.

On those basic facts, KJK has erected an elaborate theory, which comes to the conclusions that IQ is mostly cultural, with a genetic component much smaller that is generally thought.  And it is the cultural part which is hereditary.

To arrive at that, KJK goes via the concept of the "cultural load" of each IQ question -- which he assesses by looking at how often a question has to be altered when you are adminstering it to a new and different population.  And he finds that by removing (statistically) the influence of cultural load, all other correlations are much reduced.

When we look more closely at his data, however (e.g. Table 3.1 in KJK's doctoral dissertation) we find that only two out of 11 question types have a high cultural load:  Vocab and general knowledge.  And the cultural dependency of those two question types has been obvious to everyone since the year dot.

What is interesting however is that the remaining 9 question types have low to negligible cultural load.  In other words, we could remove the vocab and knowledge subtests from the overall test and still have a robust test.  So my conclusion is that what KJK should have done from the beginning is to remove those two flawed item types from his calculations altogether.  Once you do that all his exciting findings melt away.  His findings rely on items that he himself knows to be flawed.

There is a summary of KJK's dissertation at  The Unscientific American -- JR

1 comment:

  1. You confuse HIGH heritability (the topic you address here) and HIGHEST heritability (the topic addressed in the original paper and in Scientific American).

    Within g theory, HIGH heritability of crystallized is indeed easily explained if you assume the underlying variables are also highly heritable and educational environment also being a (possible) source of variance in crystallized intelligence is homogeneous. But that is not the problem here.

    The problem is that crystallized abilities have HIGHEST heritability, meaning higher than fluid abilities. A higher heritability for crystallized intelligence than for fluid intelligence cannot be explained by assuming that (educational) environment is homogeneous.

    Proof:

    We introduce the following theoretical variables:
    General intelligence (g), Fluid intelligence (F), Crystallized intelligence (C), Educational environment (E_edu)

    And a scaling variable k (which is not really necessary).

    We have the following formulas, which describe theoretical relations in g theory, and, in addition, which are according to factor models of intelligence:

    1) F = g + ξ_f

    (in Jensen's view the residual ξ_f is 0, because he considers g and fluid intelligence to be one and the same variable, but leave this an open question).


    2) C = k*F + E_edu + ξ_c (= k*(g + ξ_f) + E_edu + ξ_c)

    (This is the investment hypothesis. Note that the concept of C does not depend on whether C is modeled as a latent factor or not)

    When education is assumed homogeneous, variance in E_edu equals/approaches 0. Note that this implies that the covariance between fluid intelligence and educational influences will also be 0. It also implies that the heritability of educational measures is 0.

    Residual variance is by definition independent from g, so there are no other covariance terms that contribute to variance in crystallized intelligence, so the expression for the variance in C becomes:

    σ(C) = σ(k*F) + σ(E_edu) + σ(ξc) = σ(k*F) + 0 + σ(ξ_c)

    Heritability of crystallized intelligence will thus be somewhere in between the heritability of fluid intelligence and heritability of the residual. Now, in g theory the residuals are less heritable than fluid intelligence (this is assumed in order to explain the positive correlation that exists between g-loading and heritability coefficients of fluid tasks). Two remarks. First, in this educational homogeneous environment, heritability can indeed be high, especially if k (the impact of F, hence of g) is relatively large as compared to the residual. This is in line with your position.

    Second, it also implies that the heritability of crystallized intelligence may be high, yet it is expected not to exceed the heritability of fluid intelligence. Still it does. The latter is the actual problem and is all but 'a basic fact' or a fact that (supposedly) has been rediscovered.

    ReplyDelete

All comments containing Chinese characters will not be published as I do not understand them