Hollow Flynn Effect in Two Developing Countries and A Further Test of the Debatable Black-White Genetic Differences

Studies of the nature of the Flynn Effect are usually done in developed countries (e.g., Rushton, 1999; Wicherts, 2004; Nijenhuis, 2007; for an ‘Overview of the Flynn Effect’, see Williams, 2013). There are two recent data on two developing countries (Khaleefa, 2009; Liu, 2012). The reported numbers on subtests gains can be studied using either MCV or PC analysis. Next, we will see that shared (c²) and non-shared (e²) environments, as measured by Falconer’s formula, are unrelated to heritability (h²) of the WAIS and WISC subtests. Culture load, heritability, g-loadings, and black-white differences tend to form a common cluster (on PC1) that is different from the pattern of loadings shown by shared and non-shared environment.

The data I am using here is summarized in this XLS. It contains the references of the relevant data and calculations. Below is the SPSS table showing the column vectors of the variables.

Column vectors - culture, h2, c2, e2, g-loadings, BW gap, FE gains in WAIS

Using these numbers, a correlation matrix and a component matrix are produced. For the latter, I use the correlation matrix from Spearman correlation because it is known to be more robust to outliers than Pearson. But, in fact, PC analysis produce more or less the same pattern of loadings if I use Pearson.

Correlation Matrix - culture, h2, c2, e2, g-loadings, BW gap, FE gains in WAIS

Component Matrix - culture, h2, c2, e2, g-loadings, BW gap, FE gains in WAIS (Spearman)

While the vector of shared environment is correlated with the vector of culture load and somewhat with g-loadings, it has its highest loadings on a different component, meaning that they behave in different ways. As can be seen, culture load, heritability, g-loadings and BW differences tend to form a cluster on PC1 with significant loadings. Shared, non-shared environment and Sudan gain (from Khaleefa 2009) are loaded positively on a different component. Dutch gain (from Wicherts 2004) is loaded almost equally on PC1 and PC2. With regard to the measurement bias question, Wicherts (2004) said about the WAIS Dutch gains :

The model with identical configuration of factor loadings in both cohorts (Model 1: configural invariance) fits poorly in terms of χ². However, the large χ² is due to the large standardization sample (χ² is highly sensitive to sample size; Bollen & Long, 1993), and the RSMEA and CFI indicate that this baseline model fits sufficiently. In the second model (Model 2: metric invariance), we restrict factor loadings to be equal across both cohorts (i.e., Λ1=Λ2). All fit indices indicate that this does not result in an appreciable deterioration in model fit, and therefore, this constraint seems tenable. However, the restriction imposed on the residual variances (Model 3: Θ1=Θ2) is not completely tenable because AIC and Δχ² indicate a clear deterioration in fit as compared with the metric invariance model. However, RMSEA, CFI, and CAIC indicate that this restriction is tenable. In a formal sense, residual variances are unequal across groups, although the misfit due to this restriction is not large. More importantly, both models (4a and 4b) with equality constraints on the measurement intercepts (ν1=ν2) show insufficient fit. The RMSEA values are larger than the rule-of-thumb value of 0.05, and (C)AICs show larger values in comparison with the values of the third model. Although the CAIC values of Models 4a and 4b are somewhat lower than the CAIC of the unrestricted model (reflecting CAIC’s strong preference for parsimonious models), the difference in χ² comparing Models 4a and 4b to less restricted models is very large. [3] Both strong and strict factorial invariances therefore appear to be untenable.

Next is the data from Liu (2012) for WPPSI subtest gains. I used the g-loadings reported by Roid & Gyurke (1991) and Liu (2011). The latter reports the rotated factor loadings, not the unrotated loadings. As Jensen (1980, p. 209) noted, a first factor after rotation is no longer a general factor. The numbers are shown below :

Column vectors - g-loadings and FE gains in WPPSI

Correlation Matrix - g-loadings, FE gains in WPPSI

Once again, from the above correlation matrix, the negative correlation between g-loadings and secular gains appears to be crystal clear, in line with the other studies.

For the last analysis, now, I will replicate the previous PC analysis using this time the data from WISC subtests. Below is the SPSS table showing the column vectors of the variables.

Column vectors - culture, h2, c2, e2, g-loadings, BW gap, inbreeding depression in WISC

With these numbers, I produced the following correlation matrix and component matrix again using Spearman rho.

Correlation Matrix - culture, h2, c2, e2, g-loadings, BW gap, inbreeding depression in WISC

Component Matrix - culture, h2, c2, e2, g-loadings, BW gap, inbreeding depression in WISC (Spearman)

The pattern of correlations and components appear similar to what is shown using WAIS data (with the exception of BW differences, loaded equally well on PC1 and PC2), namely, the total absence of relationship between c² and e² with h² and g. At the same time, the cuture-loadedness of the tests is related to their heritabilities and g-loadedness. This however looks surprising.

Now, given all these numbers, what need to be said ? At first glance, the positive correlation between culture load and g and heritability is difficult to explain from the hereditarian viewpoint, if Kees-Jan Kan words were to be believed. This result, among other things, lead him to conclude in his PhD thesis (2011, ch. 3 & 4) that the genetic origin of the black-white differences, as posited (e.g.) by Rushton and Jensen, could be nothing more than “Weak inferences based on ambiguous results”. But at the same time, the above PC analyses showed that c² and e² are loaded significantly on a different component than PC1, which comprises the highest loadings for h² and cultural load. This pattern, of course, cannot be explained by the cultural hypothesis.

And because we haven’t discussed it yet. What are the characteristics of a culture-loaded test ? Kan (2011, Table 3.1, Figure 3.3) considers the culture-loaded tests as the crystallized ones, and the culture-reduced tests as the fluid ones. The reason is that, as Kan said, “Fluid tests minimize the role of individual differences in prior knowledge, whereas crystallized tests maximize it”, thus meaning that fluid tests are more reflective of cognitive processes while crystallized tests are more reflective of acquired skills and knowledge. Kan cites Jensen’s own definition of a culture-biased test :

“Tests and test items can be ordered along a continuum of culture loading, which is the specificity or generality of the informational content of the test items. The narrower or less general the culture in which the test’s information could be acquired, the more cultural loaded it is. A test may contain information that could only be acquired within a particular culture. This can actually be determined simply by examination of the test items. The specificity or generality of the content corresponds to its cultural loading.” (Jensen, 1976, p.340)

More on this can be read in Jensen (1980, pp. 133, 234-235, 374, 538-539, 635-637; 1998, pp. 123-126). Also important to keep in mind is that a culture-reduced test should maintain the same psychometric properties, or parameters, meaning that the tests behave the same way across groups (e.g., race, gender, cohort). As Jensen says, in Bias in Mental Testing :

Culture-reduced items are nonverbal and performance items that do not involve content that is peculiar to a particular period, locality, or culture, or skills that are specifically taught in school. Items involving pictures of cultural artifacts such as vehicles, furniture, musical instruments, or household appliances, for example, are culture loaded as compared with culture-reduced items involving lines, circles, triangles, and rectangles.

Operationally, we can think of the degree of “culture reducedness” of a test in terms of the “cultural distance” over which a test maintains substantially the same psychometric properties of reliability, validity, item-total score correlation, and rank order of item difficulties. Some tests maintain their essential psychometric properties over a much wider cultural distance than others, and to the extent that they do so they are referred to as “culture reduced.” Certain culture-reduced tests, such as Raven’s Progressive Matrices and Cattell’s Culture Fair Test of g, have at least shown equal average scores for groups of people of remotely different cultures and unequal scores of people of the same culture and high loadings on a “fluid” g factor within two or more different cultures. Such tests apparently span greater cultural distances than do tests that involve language and specific informational content and scholastic skills.

If we accept this premise, then I would not reach the same conclusion as Kan (2011, p. 56) did, when he affirms that the black-white differences are larger on the more cultural-influenced tests. It remains to be seen that crystallized tests, as compared with fluid tests, really behave differently across groups. While Jensen argued (1980, p. 133) that culture-loaded tests are not necessarily culture-biased, he has made it clear that a culture-influenced test should be manifested through between-group differences in the meanings of the tests/items. Given some tests of measurement equivalence done using the WISC tests (Dolan, 2000; Lubke, 2003), there is no clear evidence of psychometric bias across racial groups. Furthermore, a positive BW difference correlation with culture load is not always found, far from it. Using a different procedure for classifying culture loaded tests, Jensen and McGurk (1987, p. 298) showed that the BW difference is larger on culture-reduced tests while holding constant item difficulty (see also, Jensen, 1977, pp. 56-62; Reynolds & Gutkin, 1981, pp. 178-179; Jensen & Reynolds, 1982, pp. 426-427; Reynolds & Jensen, 1983, pp. 210-213). Given the extensive literature on this subject reviewed by Jensen (1973, ch. 4, 12, & 17; 1980, ch. 10, 11 & 12; 1998, ch. 11), I must maintain that the BW differences are larger on culture-reduced tests.

References :

Dolan (2000). Investigating Spearman’s hypothesis by means of multi-group confirmatory factor analysis.
Lubke (2003). On the relationship between sources of within- and between-group differences and measurement invariance in the common factor model.
Jensen (1973). Educability and Group Differences.
Jensen (1977). An Examination of Culture Bias in the Wonderlic Personnel Test.
Jensen (1980). Bias in Mental Testing.
Jensen (1998). The g Factor.
Jensen & McGurk (1987). Black-white bias in ‘cultural’ and ‘noncultural’ test items.
Jensen & Reynolds (1982). Race, social class and ability patterns on the WISC-R.
Kan (2011). The nature of nurture: the role of gene-environment interplay in the development of intelligence.
Kaufman (1988). Sex, race, residence, region, and education differences on the 11 WAIS-R subtests.
Khaleefa (2009). An increase of intelligence in Sudan, 1987–2007.
Liu (2011). Factor structure and sex differences on the Wechsler Preschool and Primary Scale of Intelligence in China, Japan and United States.
Liu (2012). An increase of intelligence measured by the WPPSI in China, 1984–2006.
Reynolds & Gutkin (1981). A multivariate comparison of the intellectual performance of Black and White children matched on four demographic variables.
Reynolds & Jensen (1983). WISC-R Subscale Patterns of Abilities of Blacks and Whites Matched on Full Scale IQ.
Roid & Gyurke (1991). General-Factor and Specific Variance in the WPPSI-R.
Rushton (1989). Japanese Inbreeding Depression Scores: Predictors of Cognitive Differences Between Blacks and Whites.
Rushton (1999). Secular gains in IQ not related to the g factor and inbreeding depression – unlike Black-White differences: A reply to Flynn.
te Nijenhuis (2007). Score gains on g-loaded tests : No g.
Wicherts (2004). Are intelligence tests measurement invariant over time? Investigating the nature of the Flynn effect.
Williams (2013). Overview of the Flynn effect.

SPSS syntax :

MATRIX DATA VARIABLES=Culture_Load Heritability shared_E nonshared_E g_WAIS BW_diff_WAIS Dutch_gain_WAIS Sudan_gain_WAIS
.491 1
.373 .097 1
-.529 -.993 -.174 1
.954 .514 .151 -.547 1
.414 .634 .000 -.589 .428 1
.532 .119 .142 -.091 .445 .433 1
-.575 -.409 .073 .388 -.592 -.203 -.506 1


MATRIX DATA VARIABLES=Culture_Load Heritability shared_E nonshared_E g_WISC BW_diff_WISC Inbreeding_Depression
.455 1
-.281 -.569 1
-.312 -.476 -.333 1
.936 .615 -.393 -.364 1
.257 .133 .161 -.308 .373 1
.613 -.027 .057 -.091 .632 .365 1


8 thoughts on “Hollow Flynn Effect in Two Developing Countries and A Further Test of the Debatable Black-White Genetic Differences

  1. Kids undoubtedly had a lot more toys to play with in China in 2006 than in 1984, and probably in Sudan in 2007 than in 1987. Kids really like to play with toys, especially new toys, so I suspect having a wide variety of toys is pretty good for them (much as I would tell my kids, “When I was your age, all we had were Lincoln Logs and we liked it.”)

  2. Generally, studies on the nature of the Flynn effect do not provide an explanation of the trend. Some subtests may have seen a rise in scores, and some a decline. Olev Must has made some speculations on the nature of the secular changes in Estonia (here) :

    The fact that the subtest Information (B2) has got more difficult may signal the transition from a rural to an urban society. Agriculture, rural life, historical events and technical problems were common in the 1930s, such as items about the breed of cows or possibilities of using spiral springs, whereas at the beginning of the 21st century students have little systematic knowledge of pre-industrial society. The fact that tasks of finding synonyms–antonyms towords (A4) is easier in 2006 than in the 1930s may result from the fact that the modern mind sees new choices and alternatives in language and verbal expression. More clearly the influence of language changes was revealed in several problems related to fulfilling subtest A4 (Synonyms–Antonyms). In several cases contemporary people see more than one correct answer concerning content and similarities or differences between concepts. It is important that in his monograph Tork (1940) did not mention any problems with understanding the items. It seems that language and word connotations have changed over time.

    The sharp improvement in employing symbol–number correspondence (A5) and symbol comparisons (B5) may signal the coming of the computer game era. The worse results in manual calculation (B1) may be the reflection of calculators coming in everyday use.

    More explanation like this might be needed.

    • What happens when give modern students a culture-loaded old test, like, say, a 1940 Wechsler IQ test? Are there questions about boiler rooms and “Where is Danzig?” that just totally throw contemporary youths?

  3. I just wanted to delurk and say that I hope this site continues to turn out these wonderful blog posts. I’ve found them very informative and useful.

    • I admit that the activity on this blog is slowing considerably recently. I’m not sure about the others and can’t speak for them. But I have a good couple of posts in preparation. One is a criticism of Kees-Jan Kan’s master thesis. It’s a very tough one. Probably the most substantial criticism of hereditarian I have ever seen since a while. Not finished (less than 50%), and need more time. But just to let you know, I will talk about the g-cultural loading correlation. That thing is quite suspicious. As my post made it clear, that correlation is near 100% but on the other hand it does not correlate with c² or e². One possible reason for this is that the Wechsler’s scale is somewhat biased towards crystallized g, so that factor analysis tends to yield higher g-loadings for more crystallized tests (and by the same token, more cultural influenced ones). Jensen, in The g Factor tells us that when we factor analyze Raven along with Wechsler’s subtests, the g-loading for the Raven is amongst the highest, if not the very highest one. If you remember, also, Rushton et al. (2007) did this kind of analysis on the Raven, using the item-total score correlation as an approximation of g-loadings. In this way, the g-loading is not a function of cultural or informational loadings. The g-h² correlation in Rushton, as I show here, is not trivial, about 0.30 or so, although there was a higher correlation between g and e² (about 0.40) with a correlation of g and c² approaching zero. You see in the present post that I found no such g-c2 or g-e2 correlation, at least when using Spearman. This is quite surprising because Wechsler, as opposed to Raven’s test, is supposed (in my understanding, at least) to reflect higher environmentality effect, but it seems that the opposite was true, if I compare with Rushton’s result. In any case, his (twin) sample size was quite small compared to the ones I collected (essentially taken from K.J. Kan, ch. 3). KJK by the ways talks a lot about the mutualism model (van der Maas 2006) that, as he believes, refutes g. Using Google Scholar, I know of only two criticisms : Rushton and Woodley. They make the same argument, namely, that g correlates with heritability and this counters van der Maas model. On the other hand, Maas and KJK say it does not. So, I need to disentangle this. As I said, it’s tough.

      I am working also on the issue of test (racial) bias in the GSS Wordsum. More than 50% finished. The least I can say is that it is not implausible that some items (perhaps Word F) are biased, even slightly although the correlation between blacks and whites with regard to the relative item difficulty is approaching 1.00 (Pearson) or 0.90 (Spearman) which is very good.

      Finally, the meta-analysis of Jensen effect on heritability and environmentality of cognitive tests (90% finished). I will surely post it tomorrow. The least I can say now is that it does not support the cultural hypothesis. The correlation between g-loadings and h2 is about 0.55 or 0.65 (after correction for artifacts such as range restriction and deviation from perfect construct validity). Conversely, the g-c2 and g-e2 are near zero or strongly negative, after and before corrections. The sample size is extremely large (more than 20 000) but there were a lot of rubbish studies. Remove them, and you have an N of about 5000-6000. It’s quite good, still. After that, I will rework on some of my previous article on the genetics of intelligence (here) and Flynn effect (here) and post it on the blog.

  4. Thanks. I’m looking forward to seeing your new posts. The one on KJK’s master thesis sounds particularly interesting, even if it’s a tough nut to crack.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s