Recently, the GSS released the survey results for the year 2012. And a skin color variable has been included. But rather than using the SDA program, available here, I used the GSS cumulative datafile 1972-2012 for SPSS, available here. This allows more complex analyses to be performed than what is possible with the SDA.

(full results at the end of the post)

First, let’s look at the cognitive gap between light-skinned blacks and dark-skinned blacks. With the appropriate SPSS syntax [1], I recoded the skin color variable into a two-categories variable, to make comparison easier.

Skin Color, Verbal IQ, and Test of the Colorism Hypothesis in the GSS (Table 1)

The gap, using the following formula would be :

5.67-4.69/SQRT((69 x 1.9² + 99 x 1.7²)/(69 + 99))
1/SQRT((249 + 286)/(168))

This gives us 0.56*15 = 8.4 verbal IQ points. By way of comparison, the mean Wordsum score for whites in 2012 is 6.22, and for the cumulative years (1972-2012) it was 6.26 when I re-ran the comparison of means analysis without filter. The black-white wordsum gap for the last year of survey, that is, 2012, is :

6.22-5.13/SQRT((947 x 1.9² + 191 x 1.8²)/(947 + 191))
1.1/SQRT((3418 + 619)/(1138))

Or, 0.58*15 = 8.7. And the wordsum gap between whites and light skinned blacks is :

6.22-5.67/SQRT((947 x 1.93² + 69 x 1.87²)/(947 + 69))
0.55/SQRT((3527 + 241)/(1016))

Or, 0.28*15 = 4.2 verbal IQ points. In other words, light skinned blacks fall intermediate between whites and blacks. Using this calculator, which divides the black-white mean difference by the averaged SD of the two groups, produces similar results.

And, to see whether or not there is a colorism effect, due to a so-called color-based discrimination, I perform a multiple regression, with skin color (ratetone), age, gender, and Wordsum as independent variables and SQRT or LN real income as dependent variable. RATETONE variable is a 10-point scale variable, going from 1 (lightest) to 10 (darkest). A negative sign means that when skin color becomes darker, the respondent’s income decrease. But the negative coefficient is only about -0.029 (SQRT) or -0.011 (LN) in Model 2, when Wordsum has been held constant.

The reason for including respondent’s wordsum score, is that a colorism effect, as the discrimination hypothesis predicts, is universal and affect all people, regardless of their race and/or gender. If, after controlling for education, IQ, parents background, and all other socio-economic background characteristics possible, the coefficients of skin color among blacks fall to near zero, the discrimination hypothesis is nullified. Because, according to colorism, the effect of such discrimination is independent of characteristics such as skills, experience, competence, intelligence, and so forth. It argues that colorism effect has to do with physical appearance, and absolutely nothing more than that. In other words, probably the best way to test colorism is to look at the effect of skin color, net of SES and/or IQ. Here, verbal IQ (i.e., Wordsum) alone explains the lion’s share of skin color-income association.

Now, if we conduct a regression of degree on skin color, holding wordsum constant, something unexpected happens : skin color among blacks is positively associated (0.102) with education level. A positive sign, here, means that the darker the skin and the higher the educational attainment. Just the opposite of what colorism would have predicted. Overall, the color-outcome association in the GSS data is consistent with what I have previously found in the Add Health and NLSY97 data, and as others did too. Of course, all the above analyses are weighted [2].

What happens when Wordsum is being predicted by skin color, age, gender, degree, and real income ? The unstandardized B for skin color was -0.185, which means for one-unit gain in the skin color variable, black verbal IQ is reduced by a mere 0.185 Wordsum point. However, a 5-point gain in the skin color variable produces 0.185*5=0.925 point loss. Without the covariates, the unstandardized B is a mere -0.223, meaning that each one-unit increase in the 10-point skin color ratings is associated with a loss of 0.223 Wordsum point, which also means that 5 points gain in skin color is associated with a loss of -0.223*5=-1.115 Wordsum point, or exactly the same amount of IQ gap separating blacks and whites for the year 2012. As we can see, the same regression analysis using the black-white dichotomized variable yields an unstandardized coefficient of 1.113.

We can compare this result with that for blacks in year 1982 for which the “color” variable was available for 437 blacks. This 5-point scale variable has the following values : 1 “very dark brown”, 2 “dark brown”, 3 “medium brown”, 4 “light brown”, 5 “very light brown”. The effect of a lighter skin is positive for income (0.018) and degree (0.030) but apparently small, after partialing out age, gender, and wordsum. If Wordsum is predicted by skin color, age, gender, real income, degree, the unstandardized B for skin color is 0.305, or 0.305*3=0.915 Wordsum point higher when blacks become slighter by 3 points. And yet the standardized B was only 0.131. Without the covariates, the unstandardized B for skin color was 0.384 with standardized B of 0.165. Again, in examining the real-world impact of skin color, one should not look solely at the standardized coefficients, as they may be highly misleading (Sackett et al., 2008, p. 216).

To some extent, the color gap among blacks with regard to verbal IQ is certainly under-estimated because we suspect the color ratings not to have a high reliability. This lowers correlations and effect sizes. On the other hand, one would also think that the minor effect of color on wages and education is also under-estimated. However, Wordsum is not reliably high (~0.70), and controlling for additional SES-related variables is expected to reduce the color-income association even further.

Finally, a technical note about the multiple regressions I conducted, is that the residuals were initially not normally distributed. Generally, the cause behind this is that the dependent variable is not normally distributed. This was the case for income, and to a lesser extent degree. Those two variables showed some skewness towards the left (we can check this out using histogram). So I simply used the square root or natural log transformation, in order to get a more normally distributed pattern of residuals.


[1] For a short description of the original variables used in the below syntax, go to the SDA program there, and click on “Question Text” before running the analyses with the relevant variables. The syntax is available here.

[2] In the General Social Survey codebook, we read :

Due to the adoption of the non-respondent, sub-sampling design described above, a weight must be employed when using the 2004-08 GSSs. One possibility is to use the variable PHASE and weight by it so that the sub-sampled cases were properly represented. If one wanted to maintain the original sample size, one would weight by PHASE*0.86258 in 2004 and PHASE*.80853 in 2006. This weight would only apply to 2004-08 and would not take into account the number of adults weight discussed above. As such, it would be appropriate for generalizing to households and not to adults.

A second possibility is to use the variable WTSS. This variable takes into consideration a) the sub-sampling of non-respondents, and b) the number of adults in the household. It also essentially maintains the original sample size. In years prior to 2004+ a one is assigned to all cases so they are effectively unweighted. To adjust for number of adults in years prior to 2004, a number of adults weight would need to be utilized as described above. WTSSALL takes WTSS and applies an adult weight to years before 2004.

A third possibility is to use the variable WTSSNR. It is similar to WTSS, but adds in an area nonresponse adjustment. Thus, this variable takes into consideration a) the sub-sampling of nonrespondents, b) the number of adults in the household, and c) differential non-response across areas. It also essentially maintains the original sample size.

As with WTSS, WTSSNR has a value of one assigned to all pre-2004 cases and as such they are effectively unweighted. Number of adults can be utilized to make this adjustment for years prior to 2004, but no area non-response adjustment is possible prior to 2004.

For all these reasons, I obviously use WTSSNR. To note, the sample size is not exactly the same, as far as I can see, but it is very, very close to the original sample size, so that is of no concern.


The full result has to be found in the spreadsheet.