Note: Our analysis has been updated and extended. Refer also to: Color Differences: Corrections and Further Analysis. Part 1
(or Colorism: Game Over)
Following up with Dalliard’s pioneering post on colorism, we looked at the relationship between skin color, cognitive ability, and educational attainment in the full NLSY 97 sample (across all racial and ethnic groups). Our purpose was to test what we term “Strong Colorism”. According to “Strong Colorism”, color discrimination is a unifying explanation for sub-population differences in the US and beyond; it is a generalized phenomenon that can account for pigment related outcome differences. Lest readers suspect that a straw position is being constructed, below is an excerpt from an article, recently published in the journal of Industrial and Organizational Psychology, discussing this model (which the authors take to be established):
Although colorism theorists recognize that other facial characteristics covary with skin tone (e.g., eye shape, nose shape, and lip shape), and are thus also inextricably interwoven into the colorism process (Thompson & Keith, 2001), they nevertheless place primacy on the causal role of skin tone in engendering the colorism phenomenon. It is also important to note that while lighter- or darker-skinned individuals can be disfavored by the phenomenon, it is typically lighter skin tones that are idealized and favored in the colorism process (Russell, Wilson, & Hall, 1992).
Why Should I–O Psychology Be Concerned With Colorism? I–O psychologists should be concerned with the issue of colorism because the phenomenon has implications that are capable of cutting across categories such as race, religion, gender, age, sexuality, nationality, and occupation. That is to say, extant psychological research contains evidence for the preference and undue favoritism of lighter skin complexions among Black, White, Latino, and Asian populations from around the world (Glenn, 2009). Much of the most rigorous research concerning the nexus of colorism and labor market outcomes has been conducted on African American and Latino populations in the United States. This research has evinced that African Americans and Latinos possessing lighter skin tones attain higher incomes, attain employment in more prestigious occupations, and experience less discrimination compared to their darker-skinned counterparts (Espino & Franz, 2002; Keith & Herring, 1991; Klonoff & Landrine 2000). In fact, Hughes and Hertel (1990) noted that the magnitude of the difference in socioeconomic outcomes between light and dark-skinned Blacks in the United States is analogous to the magnitude of the difference in socioeconomic outcomes between Blacks and Whites. However, as stated previously, colorism does not simply affect African Americans and Latinos; rather, it is a global phenomenon that consistently privileges lighter skin tones over darker ones (Glenn, 2008). Thus, the pervasiveness of this form of discrimination and its impact on workplace and labor market related outcomes, both in the United States and abroad, dictate that I–O psychologists become more acquainted with this form of discrimination. (Marira, T. D., & Mitra, P. (2013). Colorism: Ubiquitous Yet Understudied. Industrial and Organizational Psychology, 6(1), 103-107.) [Emphasis added]
Four points are relevant here. Firstly, discrimination on the basis of color is said to be ubiquitous and contemporaneous. Secondly, this color discrimination is said to cut across racial, ethnic, and national boundaries — it’s generalized. Thirdly, this color discrimination is said to be the direct cause of practically significant outcome differences. Fourthly, the evidence for this ubiquitous, outcome-related, contemporaneous colorism primarily comes from research on African and Latino Americans in the US. It is notable that hitherto this model of colorism has not been tested. To date, it has only been demonstrated that color differences are associated with various outcome difference. Proponents of colorism seamlessly move from “This research has evinced that African Americans and Latinos possessing lighter skin tones attain higher incomes, attain employment in more prestigious occupations” to “However, as stated previously, colorism does not simply affect African Americans and Latinos.” Little to no evidence for “colorism” as a causal model, as opposed to colorism as a description of associations, is offered.
To start out, we identified plausible causal scenarios for the association between color and outcomes. The ones we felt worth investigating were as follows:
Colorism. Color discrimination leads to outcome differences (e.g., Marira and Mitra (2013)). If IQ differences, IQ differences are consequent to other outcome differences (e.g., education).
Pleiotropy-IQ. Direct genetic effect between alleles coding for color and IQ; IQ differences are antecedent to outcome differences (e.g., Jensen (2006)).
Shared Environment-IQ. Difference in shared family environment, which are correlated with differences in color, condition difference in IQ; IQ differences are antecedent to outcome differences.
Additive Genetic-IQ. Additive genetic differences, which are correlated with differences in color, condition difference in IQ; IQ differences are antecedent to outcome differences. (e.g., Lynn (2002)).
We then identified predictions. First, IQ based models predict that IQ measured antecedent to outcomes will statistically explain outcome differences. Colorism predicts otherwise, since IQ differences are consequent to (non IQ) outcome differences. Second, colorism and pleiotropy predict that the correlation between color differences and outcomes differences (at least between color and IQ in the case of IQ-pleiotropy) will be similar within and between families. In contrast, both a shared environmental and an additive genetic model predict that the differences will mainly be between families. Third, a shared environmental model predicts no differences within families regardless of genetic kinship. An additive genetic model, on the other hand, predicts mostly no differences only between full siblings. Half siblings and unrelated siblings would, by a genetic hypothesis, show color outcome associations.
These three sets of predictions allow us to disentangle our four hypotheses:
Colorism: Antecedent differences in IQ largely don’t mediate color-outcome associations. Similar association within and between families.
pleiotropy-IQ: Antecedent differences in IQ largely do mediate color-outcome associations. Similar association, at least for IQ, within and between families.
Shared Environment-IQ: Antecedent differences in IQ largely do mediate color-outcome associations. Association mostly between families. No difference within families regardless of genetic relatedness.
Additive Genetic-IQ: Antecedent differences in IQ largely do mediate color-outcome associations. Association mostly between families. Associations within families between non full siblings.
To test these models we followed the method discussed by Dalliard. Our sample is from the National Longitudinal Study of Youth 97. The primary variables used were color, cognitive ability measured between ages 12 and 18 (in 1997 for the AFQT), and self reported highest grade ever, measured more than a decade latter (about 2008-2010). Because Strong Colorism proposes a generalized color discrimination we looked at the full sample unsorted by race or ethnicity. Because the evidence for this model is said to come from research on African and Latino Americans, we also looked at those sub-populations.
Analysis and Results
For the first part of this analysis we looked at the correlations within and between families in the NLSY sample not decomposed by race or ethnicity. These are shown below along with descriptives. We included rho, since parametric assumptions were violated. Weighted and unweighted results are provided. Sibs were not sampled within families, so the appropriateness of using weights for this sample was unclear. The sibling weights were computed by scaling the original weights, integerizing these, and then computing the sum of the intergerized scaled weights for the sibling pairs.
(To note: correlations that were statistically significant at the .05 level, one tail, are in red. All other correlations did not reach this level of significance.)
As the correlational results show, there was no significant correlation between color and AFQT or color and HGE within families for full sibs (who were not missing color, HGE, or AFQT scores.) To confirm this, a non parametric ANOVA, Kruskal-Wallis, was run. It was found that while skin color was significantly associated with AFQT (H(9) = 183.16, p<.05) and HGE (H(9) = 64.25, p<.05) between families, within families there was no significant association between skin color and AFQT (H(5) = 3.23, n.s.) or between skin color and HGE (H(5) = 4.83, n.s.). Jonckheere's test showed a significant trend between color and AFQT and between color and HGE between families. As color darkness increased, AFQT decreased, J=91255, z = -13.372, r = -0.46. And as color darkness increased, HGE decreased J=114138, z = -7.75, r = -0.27. There were no significant trends within families for AFQT (J=5958, z = -0.324, r = -.02) or Highest Grade Ever (J= 5835, z = -.60, r = -0.04). Using method (7) discussed by Steiger (1980) for comparing correlations between dependent samples, we compared the between family correlations to those within (for the full sample, between full siblings). The color-AFQT and the color-HGE correlations between families were all significantly different from those within families.
Box Plots Between Families and Within Families between Full Siblings
To better understand the associations between color, AFQT, and HGE we ran linear regressions. As expected, AFQT – or, in this analysis, g – statistically explained the association between color and HGE between families. Indeed, after accounting for differences in general intelligence, the association between color and HGE turned positive.
The analysis above confounded a number of factors such as mean age with years of education attained. Nonetheless, the overall effect was robust. For example, when limiting the sample to those individuals who had attained 13 to 20 years (n(sib pair) = 836) and controlling for mean age (of the siblings), we found a similar effect. In the NLSY 97, the birth years ranged from 1980 to 1984 and the year of cognitive testing was 1997. As such, education attained after year 12 necessarily was consequent to cognitive ability measured in 1997. The results then bear out our main finding which should be of no surprise to anyone who is even remotely familiar with the literature on race differences in educational attainment and cognitive ability.
The absence of association within families indicates that color differences, per se, are not, to a practically significant extent, causally associated, either by way of discrimination (or pleiotropy), with IQ or HGE differences. Color and outcomes are largely correlated by third factors. One possibility is shared environment another is additive genetics. To disentangle these two models, we looked at the respective correlations between Unrelated Sibs, Half Sibs, and these groups combined. Contrary to an additive genetic hypothesis, as noted above, a pure shared environmental hypothesis predicts that there will be no correlation between color and outcomes between non full siblings.
Below shows the correlations. As can be seen, AFQT was generally correlated with color between unrelated siblings both within and between families. Interestingly, HGE and color was not associated within families.
We then explicitly pitted a shared environmental model against an additive genetic model. The results are shown below:
The patterns of association are consistent with a shared environmental hypothesis for HGE differences and an additive genetic hypothesis for AFQT differences.
The lack of association between color and HGE, despite the association between color and AFQT, between non full sibs within families suggested two possibilities. First: either the correlation between AFQT and color or the association between HGE and color could be spurious. Second: AFQT is correlated with color within families between non full siblings but HGE is not. The latter is not particularly inconsistent with an additive genetic hypothesis for IQ x color differences because the heritability of self reported educational attainment is significantly lower than that for IQ (e.g., Nielsen and Roos, 2011). (To note, in this sample, the correlation between HGE and AFQT is 0.56; if differences in g are driving differences in HGE, then the color-HGE correlations should be about one half the magnitude of the color-g correlation, assuming no moderators.) To explore the sample further, we looked at the correlation between color and AFQT and between color and HGE independently, using pairwise instead of listwise deletion so to maximize sample size and power. We also added averaged PIAT scores of siblings, when available, who were missing AFQT scores. (To note, in the full sample, the correlation between averaged PIAT scores and AFQT was >0.7.) The results are presented below:
As can be seen for non-full sibs, color and measures of cognitive ability (both AFQT and PIAT) were significantly negatively correlated within and between families, but color and HGE was not. The correlations for full siblings were similar to those discussed above, except that one of the color-HGE correlations between families, between full sibs reached statistical significance at the .05 level, one tail.
We then looked at the associations within the Black and Hispanic populations separately and combined. To expedite analysis we used a modified syntax which is given in the attached excel file. The results for Blacks were as expected:
For Hispanics, while there were no significant associations between color and outcomes within families (though the color-HGE correlations were trending towards significance), surprisingly there were also no significant associations between families, despite there being associations in the full sample between all Hispanic individuals. Further investigation showed that the lack of association was only found between families in which more than one individual participated in the study.
We then looked at the combined Black and Hispanic population, grouping them together, since jointly they represent “people of color” or “under-represented minorities.” Here, the color-AFQT correlations were significant between families but not within. The color-HGE correlations were not significant either within or between populations.
Next we looked at the association between color, g-factor scores, and t-factor scores. For a discussion of these variables, refer to Dalliard’s prior post. These results are shown below.
Within families, color was only significantly associated with g between unrelated siblings in the full sample. Between families, color was significantly associated with g for Blacks, with g for full sibs in the full sample, and with g for half sibs and unrelated sibs in the full sample. We also looked at the full sibling correlations for g and t. As we predicted the full sib correlation was higher for g, the more genetically influenced factor, than for t, the more environmentally influenced factor. (To note: using the same method, the full sib correlations for HGE was r =.561, rho =.548).
As a final analysis we used the method of correlated vectors to look at the association between color and g-loadings within and between families. We looked at the correlations (r and rho) between g-loadings and the color-subtest correlations (r and rho) within and between families using weighted and unweighted values. We limited ourselves to the full sample and to Blacks. This gave us 32 correlations. For both the full sample and for Blacks none of the within family correlations (between g-loading and magnitude of the color-test correlations) reached statistical significance. For the full sample, all of the between family correlations reached statistical significance. For Blacks, between families, five out of eight of the possible permutations reached significance.
This last analysis provided surprisingly ambiguous results. Previously, we had found a strong Jensen Effect on the IQ-Color correlations for Blacks when looking at all individuals in the sample. We had expected to find an equally strong effect between Black families. Instead the effect was equivocal, heavily depending on which statistics were used and how the data was treated. It is probable that our method contributed somewhat to the ambiguity.
We conducted this analysis to test what we term “Strong Colorism”, the theory that the associations between color and outcomes in the US are driven by color discrimination. To test this theory we employed a within/between family design. We first looked at the associations in the full NLSY 97 sample. We found that between families, color was associated with both AFQT scores, measured between ages 12 to 18, and Highest Grade Ever (HGE), reported by the individuals more than a decade later. We found that, in the full sample, AFQT/g differences completely statistically explained HGE differences. We also found that neither AFQT nor HGE were, for the most part, associated with color within families. The upshot of these results is that none but a very weak version of “Strong Colorism” (or, alternatively, pleiotropy) can be true. Not only did AFQT scores completely account for subsequent HGE difference between families but there was practically no association between color and AFQT or between color and HGE within families. The latter finding virtually rules out the possibility of any practically significant contemporaneous colorism. Given these findings, to salvage this paradigm, one would need to posit a historic reign of color discrimination that is now being intergenerationally transmitted (i.e., by way of shared environmental effect.) Theoretically, such a model is problematic because, as we have seen, IQ is the main mediating factor between color and outcomes and because IQ is moderately to poorly environmentally inter-generationally transmitted (as indexed by the correlation between the IQs of parents and those of their adopted offspring, by adulthood.)
Regardless, to test such a hypothesis we looked at the correlations between color, AFQT, and HGE between non full sibs. A shared environmental hypothesis predicts zero correlations for these sibs, while an additive genetic hypothesis predicts correlations. We found that, indeed, color and AFQT scores (along with color and PIAT scores) were significantly associated within families between non-FS. Color and HGE scores were not, though. When we pitted a shared environmental and an additive genetic hypothesis, we found support for an additive genetic IQ hypothesis and for a shared environmental HGE hypothesis. We take this as tentative evidence for an additive genetic contra shared environmental explanation for the color-AFQT association in the full Sample. Generally, we find it easier to reconcile the null HGE association with a color-IQ genetic hypothesis than to reconcile the strong AFQT association with a shared environmental everything hypothesis. But we are open to alternative interpretations.
We then looked at the associations within the Black and Latino populations and within the combined Blacks and Latino population, as these associations are said to furnish proof for Strong Colorism. For Blacks there were no significant associations within families but there were such associations between families. For Hispanics, anomalously, there were no significant associations between families, at least families in which there were more than one child in the study. Given the associations between all Hispanics in the study this is curious and it suggests a sampling issue.
Next, we looked specifically at the association between color and g, extracted from the 12 ASVAB sub-tests. All of the associations were consistent with what was said above, indicating that AFQT is a good measure of g. At this point we looked at the full sibling correlations for g, the principal factor, and what we have labeled as t, the second factor. We found that, as expected, the full sibling correlations for g were significantly higher than for t. Presumably this because while the full sibling correlations for both g and t are mediated by shared environment the correlations for g are also mediated to a substantial extent by genes.
Finally, we applied the method of correlated vectors (MCV) to the color-ASVAB correlations for the Black sample and for the full sample. As expected, there was a strong Jensen effect on the color effect in the full sample between families. The results for Blacks were equivocal, though the effects were trending in the predicted direction when not significant. Generally, we don’t give MCV results based on these ASVAB scores much weight for reasons discussed elsewhere.
Overall, we find substantial support against the model we have termed “Strong Colorism.” Claims of “ubiquitous colorism” are unwarranted and unjustified. Moreover, we find that the results are at very least consistent with a non-trivial additive genetic hypothesis which proposes that the association between color and outcomes are partially mediated by genetic g. In light of these results, and taken together with others, we argue that the nature (not just nurture) of the association between color and outcomes in the US deserves further investigation. As we have noted, a more detailed investigation can be conducted using the Add Health sibling sample. (Unfortunately, the requisite data is not publicly available and those individuals who have access to the sibling file whom we have contacted have been unwilling to assist us.)
Jensen, A. R. (2006). Comments on correlations of IQ with skin color and geographic–demographic variables. Intelligence, 34(2), 128-131.
Lynn, R. (2008). Pigmentocracy: Racial hierarchies in the Caribbean and Latin America. The Occidental Quarterly, 8(2), 25-44.
Marira, T. D., & Mitra, P. (2013). Colorism: Ubiquitous Yet Understudied. Industrial and Organizational Psychology, 6(1), 103-107
Nielsen, F., & Roos, J. M. (2011). Genetics of Educational Attainment and the Persistence of Privilege at the Turn of the 21st Century.
Steiger, J. H. (1980). Tests for comparing elements of a correlation matrix. Psychological Bulletin, 87, 245–251.