Hello There, Guest!  
 Previous 1 2 3 4 5 9 Next   

[ODP] The Elusive X-Factor: A Critique of J. M. Kaplan’s Model of Race and IQ

#11
Quote:White flight and the resulting urban decay are such well known phenomena in 20th century US history that I regard what I wrote as common knowledge not requiring a source citation.

Recall that not all readers are from the US. I am a case in point (Danish). Recall also that for most of recent history, north European countries have been rather racially homogeneous countries, so concepts of white flight and deterioration of cities due to immigration of lower g peoples are entirely new here. To be sure, it is common now also in Denmark.

Rangvid, B. S. (2009). "School Choice, Universal Vouchers and Native Flight from Local Schools". European Sociological Review 26 (3): 319–335. doi:10.1093/esr/jcp024

Denmark has traditionally lacked the kind of ethnically self-aware groups that would produce histories of this kind of thing. They (white nationalist groups) are emerging now as a reaction to the massive immigration.

The comment will need a cite for a general history of this phenomena, a review or at least a study of it. The inflammatory remarks must be rewritten to neutral language. This is the only part of the 40-ish page article that I think must be changed before I can approve it.
 Reply
#12
(2014-Aug-12, 22:10:55)Dalliard Wrote: What published study would you suggest? R&J 2010 is a sort of review article, so I think it's appropriate.


Unfortunately, there is still nothing, unless te Nijenhuis decides to publish the entire thing. He told me before (one year ago) that he still needs to collect more sample. I don't know where this is going. You can cite R&J 2010, but precise it's about an unpublished study. And I'll be ok.

Quote:Of course, genetic correlations can be spurious (e.g., they may not reflect shared causal influences but common ancestry) but as they have been confirmed for g in multivariate GCTA with unrelated individuals, too, that's extremely unlikely.

I'm not talking about confoundings. Just that g must be modeled explicitly. What the studies are doing here is just a model with 2, 3 or 4, correlated factor models. Not g models.

Quote:Yeah, I think that study is more interesting for its methods than its results because of sampling issues. It should be replicated now that better data are available. I will remove the reference because it's really a rather equivocal study.

The sampling bias can be an explanation, but I remember several times I saw something like that. See Jensen's Educability & Group Differences, page 182, he reports some weird numbers on black twin heritabilities (ranging from 0.02 to 1.76). He said it's possibly due to low sample size. But too many times I read Rushton, Jensen, and Lynn, and others arguing about sampling issues in ethnic samples when white-group data differ from minority-group data. Of course, they are indeed smaller. But another possibility could be that the pattern of relationships found among whites are not necessarily generalizeable. This latter possibility seems to be somewhat obscured, more often than it deserves.

Quote:For example, the various Jensen and anti-Jensen effects can be readily explained by the g model, but not by non-g models.

Ok with that. But usually, the model fit has a purpose. It helps us to select the best model. So unless they are wrong in doing it like this, I'm not convinced that a g model should be selected. But if g model cannot be selected as the best model, how can we make claims such as "MCV shows that g1-loading correlate positively with group difference and non-g loading correlate negatively with group differences, and thus it proves that a g model fits better" ? Of course, if anything, this finding is more supportive of Jensen's view than otherwise. Like I said, I agree with that. But the selection of the best model is not a question for which i have found an answer already. But if I should trust what they (e.g., Dolan, Wicherts and others) said, then we still don't have a proof that g model is superior than non-g model, even though the parameter values of the g model confirm a Spearman-Jensen effect.

That said...
Model fit is sometimes a difficult concept to grasp. I saw many times people calling "model" a path analysis (or SEM) with a full model versus reduced model. In the full, they have all structural path coefficients freely estimated. In the reduced model, they constrain one or two path corelations to be equal to zero because model fit indicates not decrement. And they call it "hypothesis testing". I'm not sure about that. I think it's more appropriate to call "competitive models" such as a model with 4 correlated factors vs 4 correlated factors + second-order g, or a bifactor g model, etc. These are what I can truly call models: it is how you build your latent variables (how many indicators per latent var, etc.) and whether or not you include higher-order factors, etc. Of course, we can say the only difference between a correlated 4 factors vs non-correlated 4 factors is that only the (structural) paths correlations are removed, and yet I can consider them as models because they are theoretically-based (Jensen talked a lot about them). But I doubt many practioners see it like this. Each time they remove a path, they immediately call it "alternative model". And yet, Dolan did something like that too. In a 2nd-order g model, he has specified different freely estimated parameters, so that for g model, you can have 5 or 6 submodels, and for non-g models, you can also have 5 or 6 submodels.

Quote:MI comprises many stages, and it's still an analysis of MI even if not all stages are tested. But I can remove those two if you insist on it.

It really depends on what you're referring to. Intercepts means that 2 groups equalized on total score should have equal probability of answering correct. If not, that means they have different knowledge levels as required by the test(s). On the other hand, violation of loading invariance means the 2 groups use different abilities to resolve the same subtests/items. Which one is crucial for the discrimination/racism hypothesis, in your opinion ? I think loading invariance is more relevant here, because it has less (perhaps not at all) to do with knowledge. So, depending on the context (i.e., what you are attempting to show by citing them), you may cite them.

Quote:I take them to mean that because of measurement non-invariance cohort differences cannot be attributed to latent factors, whereas the b-w gap can be attributed to latent factors, given that MI is not violated. Do you disagree with this?

Wicherts makes no distinction (at least, not clearly) between subtest bias and test bias. The first does not imply the second, if there is DIF or DBF cancellation (DBF stands for differential bundle functioning, bundle is sometimes referred to subtest; yes, I know, it's truly annoying that people use different words and names to talk about the same thing). So, reading Wicherts, it says something like subtests 1-4 are biased against old cohorts, but 5-8 are biased against recent cohorts. Thus the subtests biases cancel out at the test level.

It seems to me that a lot of people using MGCFA are not aware of DBF cancellation/amplification. See here.

Roznowski, M., & Reith, J. (1999). Examining the measurement quality of tests containing differentially functioning items: Do biased items result in poor measurement?. Educational and psychological measurement, 59(2), 248-269.

They are not the first who make this argument. That said, it's only when I learned and read about DIF studies that I come to understand how users of MGCFA poorly understand the concept of bias. These are different things :

1. loading non-invariance
2. intercept non-invariance (difference statistically significant, through model fit indices)
3. intercept non-invariance (magnitude of the difference, needs to be calculated given the observed and expected mean difference)
4. intercept non-invariance (direction or "sign" of the bias)

Most of the time, when reading MGCFA studies, people just stop at 1. Sometimes, they consider 2. But almost never 3 and 4. They don't care or just don't know that "significant" difference is not equivalent to "meaningful" difference. Let alone the direction of the bias and the concept of amplification and cancellation DIF/DBF that is almost never discussed in their papers. But this concept is usually discussed in DIF studies.

This being said, things can be still different if Wicherts (2004) discovered in most studies that loading invariance is violated. But except for one study (the estonian data) it's usually the intercept that is violated (sometimes with residual invariance as well). See the models 2 with the lambda greek letter "Λ" (factor loadings). It's clear that it is not "Λ" that is source of problem.
http://wicherts.socsci.uva.nl/wicherts2004.pdf

I would say that it is only if most studies show violation of MI at factor loading level, that this claim can be more or less justified. Because it means that BW differences are comparable (groups use same ability to resolve the same test), but cohort differences are not (groups don't rely on the same ability), and thus any attempt to link BW gap and FE gains sits on a shaky ground.

Indeed, when they say stuff like "cohort differences cannot be attributed to latent factors" it's extremely confusing. They probably understand it as Lubke (2003) explained, i.e., it's about differences of the factors at work, as you also say in your paper. But the sentence seems to be more inclusive than just this idea. Still, I can agree if it concerns loading invariance. But if it's about intercept invariance, I cannot make sense of their sentence.

To illustrate, just look at Beaujean/Osterlind (2008). When "adjusting" for item bias, PPVT-R gains vanish but PIAT-math gain has been only cut in half. In the first case, you would say bias is an important factor, and in the latter case, you can say that bias can account for 50% of the FE gains. In the series of analyses performed by Wicherts, it's more like bias accounts for 0% of the FE gains, because the biases cancel out. And yet in both cases, because of intercept bias, the claim that "cohort differences cannot be attributed to latent factors" is entirely correct. For practical purposes however, it's completely irrelevant. It falsely gives the impression that bias fully or meaningfully accounts for Flynn gains.

Nonetheless, you can still ignore this comment here. It's not sufficient for disapproving your paper. I just wanted to make clear my views on this problem here. Perhaps it's me who's wrong. If you agree and add a note, that's excellent. Otherwise, no problem.
 Reply
#13
I will ask te Nijenhuis about the situation with that study. He presented the findings at the London conference that I attended.
 Reply
#14
A new version of the article is attached. I made the following changes:

- Added reference to Rowe et al. (1998) (Herrnstein's syllogism)
- Specified that Phillips et al. (1998) used a verbal IQ test.
– Rewrote the part of the housing discrimination section that Emil objected to. I don't think it needs a source though.
- Removed reference to Rowe & Cleveland's (1996) biometric SEM study.
- Removed references to the two MI studies where equal intercepts were not tested.
- Plus some minor rewordings.

(2014-Aug-13, 01:30:30)menghu1001 Wrote:
Quote:Of course, genetic correlations can be spurious (e.g., they may not reflect shared causal influences but common ancestry) but as they have been confirmed for g in multivariate GCTA with unrelated individuals, too, that's extremely unlikely.

I'm not talking about confoundings. Just that g must be modeled explicitly. What the studies are doing here is just a model with 2, 3 or 4, correlated factor models. Not g models.


The genetic correlations indicate that the same genes explain most of the heritability of seemingly unrelated different abilities, e.g. verbal and perceptual ability. This is consistent with the g model. It does not prove the validity of the model, just increases its plausibility. Modelling g explicitly would not prove the g model, either, just potentially increase its plausibility even more.

Quote:
Quote:Yeah, I think that study is more interesting for its methods than its results because of sampling issues. It should be replicated now that better data are available. I will remove the reference because it's really a rather equivocal study.

The sampling bias can be an explanation, but I remember several times I saw something like that. See Jensen's Educability & Group Differences, page 182, he reports some weird numbers on black twin heritabilities (ranging from 0.02 to 1.76). He said it's possibly due to low sample size. But too many times I read Rushton, Jensen, and Lynn, and others arguing about sampling issues in ethnic samples when white-group data differ from minority-group data. Of course, they are indeed smaller. But another possibility could be that the pattern of relationships found among whites are not necessarily generalizeable. This latter possibility seems to be somewhat obscured, more often than it deserves.

There's certainly a need for more data, but what we do have is consistent with there not being race interactions, see the meta-analysis by John and yours truly. Also, if the genetic and environmental determinants were very different across races, I don't think the factor structures could be so similar.

Quote:
Quote:For example, the various Jensen and anti-Jensen effects can be readily explained by the g model, but not by non-g models.

Ok with that. But usually, the model fit has a purpose. It helps us to select the best model. So unless they are wrong in doing it like this, I'm not convinced that a g model should be selected. But if g model cannot be selected as the best model, how can we make claims such as "MCV shows that g1-loading correlate positively with group difference and non-g loading correlate negatively with group differences, and thus it proves that a g model fits better" ? Of course, if anything, this finding is more supportive of Jensen's view than otherwise. Like I said, I agree with that. But the selection of the best model is not a question for which i have found an answer already. But if I should trust what they (e.g., Dolan, Wicherts and others) said, then we still don't have a proof that g model is superior than non-g model, even though the parameter values of the g model confirm a Spearman-Jensen effect.

You will never have a single test that will determine what the correct model is. The are always alternative models in CFA that fit equally well. You will have to look at the big picture, all the evidence.

I'll reply to the rest of your comments later.


Attached Files
.pdf   The Elusive X-Factor A Critique of J. M. Kaplan’s Model of Race and IQ.pdf (Size: 518.92 KB / Downloads: 502)
 Reply
#15
What about Piffer's method for checking population differences in g? Surely, the current known SNPs for IQ fit rather well with the world phenotypic IQ scores. The good thing about this method is that it does not actually require that someone conducts an admixture study which is expensive and extremely politically incorrect. One merely has to use the SNPs found in GWAS studies which seem more politically palatable.

Surely, if we found that after we have 100 SNPs for g, use the Piffer method or some descendant, and find that there is no such correlation across-races with frequencies, this would be a severe blow to a global hereditarianism i.e. the genetic model about group differences at the world level.
 Reply
#16
Comments about the second revision.

The Flynn 1980 is missing from the reference list.

The PDF file is strange in that copying text from it is somewhat buggy. What is the PDF creation procedure?

“Kaplan’s epistemological double standard cannot be explained away by the fact that he only briefly and cursorily reviews research on racism in America.”

This reminds me of this Gottfredson paper. Gottfredson, L. S. (2007). Applying double standards to "divisive" ideas. Perspectives on Psychological Science, 2(2), 216-220. http://www.udel.edu/educ/gottfredson/rep...ndards.pdf

“One of the most important discoveries made by Arthur Jensen in his research on the black-white IQ gap was the finding that its magnitude is not invariant across different tests but tracks their g loadings, or correlations with the latent general factor of intelligence. In a meta-analysis of 149 tests, he found an average correlation of 0.63 between the magnitudes of black- white gaps and g loadings (Jensen, 1998, pp. 377–378). 9 What this means is that the better a measure of the g factor a given cognitive test is, the greater the black-white gap on it usually is.”

This is somewhat unclear. There were 149 subtests in total fifteen studies combined into one study.

“However, this has not been universally found to be the case (e.g., Murray, 2007). It is
conceivable that insufficient sampling of black individuals from the full range of ability or
floor effects in tests have artificially lowered the variance of black IQ in many studies.”

Someone should do a meta-analysis. Chuck?

Aside from the missing reference, I approve of publication.
 Reply
#17
Perhaps I am egotistic but I cannot approve of a paper on race differences in intelligence which doesn't even cite Piffer's work. This paper is not in touch with the most recent developments in this field such as Piffer, 2013 (https://drive.google.com/file/d/0B7hcznd...sp=sharing)
or Piffer, 2014 (http://biorxiv.org/content/early/2014/08/14/008011).
 Reply
#18
I agree (and I wish I had remembered that myself). This paper ought to cite Piffer's work. Also, though I forgot to mention this, Jensen's estimate for r (B-W difference x g loading) = 0.63 is much too low. Applying meta-analytical corrections raises the figure substantially.
 Reply
#19
Jensen partialled out reliabilities though.

And yes, when it comes to the testability of HH. The obvious options are:
- Admixture studies, perhaps the strongest possible type is the within siblings version because it controls for shared environment.
- SNP counting á la Piffer's method

All the other methods I can think of are more indirect and have been used, e.g. cross-racial adoptions.
 Reply
#20
(2014-Aug-15, 17:22:46)Emil Wrote: The Flynn 1980 is missing from the reference list.


No, it isn't:

Flynn, J. R. (1980). Race, IQ and Jensen. London: Routledge.

Quote:The PDF file is strange in that copying text from it is somewhat buggy. What is the PDF creation procedure?

The PDF was made with Adobe PDF Maker from a Word document, so I don't think that's the problem. The culprit is probably the font I use, Cambria, which I rather like but which does not always come out properly when converting from Word to PDF. However, I don't notice any problems when copy-pasting from the file.

Quote:“One of the most important discoveries made by Arthur Jensen in his research on the black-white IQ gap was the finding that its magnitude is not invariant across different tests but tracks their g loadings, or correlations with the latent general factor of intelligence. In a meta-analysis of 149 tests, he found an average correlation of 0.63 between the magnitudes of black- white gaps and g loadings (Jensen, 1998, pp. 377–378). 9 What this means is that the better a measure of the g factor a given cognitive test is, the greater the black-white gap on it usually is.”

This is somewhat unclear. There were 149 subtests in total fifteen studies combined into one study.

I'll rewrite it.

(2014-Aug-15, 18:30:30)Duxide Wrote: Perhaps I am egotistic but I cannot approve of a paper on race differences in intelligence which doesn't even cite Piffer's work. This paper is not in touch with the most recent developments in this field such as Piffer, 2013 (https://drive.google.com/file/d/0B7hcznd...sp=sharing)
or Piffer, 2014 (http://biorxiv.org/content/early/2014/08/14/008011).


I don't refer to your work because my article is not an exhaustive review of arguments in favor of hereditarianism, and I don't think your arguments are among the most important and convincing of the various lines of evidence that exist, certainly not in the US context.

(2014-Aug-15, 23:37:06)Philbrick Bastinado Wrote: Also, though I forgot to mention this, Jensen's estimate for r (B-W difference x g loading) = 0.63 is much too low. Applying meta-analytical corrections raises the figure substantially.


I pointed this out in footnote 9.

(2014-Aug-13, 01:30:30)menghu1001 Wrote: Wicherts makes no distinction (at least, not clearly) between subtest bias and test bias. The first does not imply the second, if there is DIF or DBF cancellation (DBF stands for differential bundle functioning, bundle is sometimes referred to subtest; yes, I know, it's truly annoying that people use different words and names to talk about the same thing). So, reading Wicherts, it says something like subtests 1-4 are biased against old cohorts, but 5-8 are biased against recent cohorts. Thus the subtests biases cancel out at the test level.

It seems to me that a lot of people using MGCFA are not aware of DBF cancellation/amplification. See here.

Roznowski, M., & Reith, J. (1999). Examining the measurement quality of tests containing differentially functioning items: Do biased items result in poor measurement?. Educational and psychological measurement, 59(2), 248-269.

They are not the first who make this argument. That said, it's only when I learned and read about DIF studies that I come to understand how users of MGCFA poorly understand the concept of bias. These are different things :

1. loading non-invariance
2. intercept non-invariance (difference statistically significant, through model fit indices)
3. intercept non-invariance (magnitude of the difference, needs to be calculated given the observed and expected mean difference)
4. intercept non-invariance (direction or "sign" of the bias)

Most of the time, when reading MGCFA studies, people just stop at 1. Sometimes, they consider 2. But almost never 3 and 4. They don't care or just don't know that "significant" difference is not equivalent to "meaningful" difference. Let alone the direction of the bias and the concept of amplification and cancellation DIF/DBF that is almost never discussed in their papers. But this concept is usually discussed in DIF studies.

This being said, things can be still different if Wicherts (2004) discovered in most studies that loading invariance is violated. But except for one study (the estonian data) it's usually the intercept that is violated (sometimes with residual invariance as well). See the models 2 with the lambda greek letter "Λ" (factor loadings). It's clear that it is not "Λ" that is source of problem.
http://wicherts.socsci.uva.nl/wicherts2004.pdf

I would say that it is only if most studies show violation of MI at factor loading level, that this claim can be more or less justified. Because it means that BW differences are comparable (groups use same ability to resolve the same test), but cohort differences are not (groups don't rely on the same ability), and thus any attempt to link BW gap and FE gains sits on a shaky ground.

Indeed, when they say stuff like "cohort differences cannot be attributed to latent factors" it's extremely confusing. They probably understand it as Lubke (2003) explained, i.e., it's about differences of the factors at work, as you also say in your paper. But the sentence seems to be more inclusive than just this idea. Still, I can agree if it concerns loading invariance. But if it's about intercept invariance, I cannot make sense of their sentence.

To illustrate, just look at Beaujean/Osterlind (2008). When "adjusting" for item bias, PPVT-R gains vanish but PIAT-math gain has been only cut in half. In the first case, you would say bias is an important factor, and in the latter case, you can say that bias can account for 50% of the FE gains. In the series of analyses performed by Wicherts, it's more like bias accounts for 0% of the FE gains, because the biases cancel out. And yet in both cases, because of intercept bias, the claim that "cohort differences cannot be attributed to latent factors" is entirely correct. For practical purposes however, it's completely irrelevant. It falsely gives the impression that bias fully or meaningfully accounts for Flynn gains.

Nonetheless, you can still ignore this comment here. It's not sufficient for disapproving your paper. I just wanted to make clear my views on this problem here. Perhaps it's me who's wrong. If you agree and add a note, that's excellent. Otherwise, no problem.


I'm still not sure I understand your point. My point of citing Wicherts on the Flynn vs. b-w gap is that the causal processes behind these gaps are different, as indicated by MI analyses. Whether the Flynn effect can be explained by somehow adjusting for the bias is beside the point for my concerns.

BTW, te Nijenhuis et al. have <a href="http://www.sciencedirect.com/science/article/pii/S0160289614001007">published</a> a new meta-analysis of g x h2 in Japan. The correlation is on the low side (0.38), perhaps reflecting the smallness of the test batteries analyzed, but the results don't contradict those from Western studies. They also review Western studies, so I'll cite this one for the link between g loadings and h2.
 Reply
 Previous 1 2 3 4 5 9 Next   
 
 
Forum Jump:

Users browsing this thread: 3 Guest(s)