The Bell Curve: Intelligence and Class Structure in American Life (94 page)

To illustrate how troubling the results have been, let me turn to two studies postdating Richard Lynn’s review that we cite on p. 289. One was a South African study led by Kenneth Owen published in the refereed British journal
Personality and Individual Differences.
¹⁷
Its sample consisted of enrolled seventh-grade students: 1,056 whites, 778
coloureds (mixed race), 1,063 Indians, and 1,093 blacks. The SPM was administered without time limits. Except for the Indians, subjects were tested by school psychologists of the same ethnic group. Owen presents the full psychometric profile for the test results (distributional characteristics, reliability, item difficulty, item discrimination, congruence coefficients, and discriminant analysis), demonstrating that the test was measuring the same thing for the various ethnic groups. The differences in test means, expressed in standard deviations, were as follows: Indian-white: −.52; coloured-white: −1.35; black-white: −2.78.

The second example of a recent, careful study was conducted by a black scholar, Fred Zindi, and published in the
Psychologist.
¹⁸
It matched 204 black Zimbabwean pupils and 202 white English students from London inner-city schools for age (12—14 years old), sex, and educational level, both samples being characterized as “working class.” Despite the fact that the white sample was well below average for the whites, with a mean IQ measured by the Wechsler Intelligence Scale for Children-Revised (WISC-R) of only 95, the black-white difference was 1.97 standard deviations on the SPM and 2.36 standard deviations on the WISC-R. Professor Zindi expressed the SPM results as IQ scores. The means for the Zimbabwean sample were 72 for the SPM and 67 for the WISC-R, consistent with Richard Lynn’s estimates. There is reason to think that the WlSC-R score was somewhat depressed by language considerations but not much: The (nonverbal) performance IQ score of the Zimbabwean sample was only 70.

What should one make of these results? Above all, one must proceed cautiously in drawing conclusions, for all the reasons that kept us from presenting these results in detail in
The Bell Curve.
The problem is not, as often alleged, that such studies are written by racists (in the two instances just cited, a charge belied by Owen’s scholarly reputation and by Zindi’s race) but that the African story is still so incomplete. Our view was that the current differences will narrow over time, probably dramatically, as nutrition and the quality of schools for black Africans improve. Changes in black African culture may provide an environment more conducive to cognitive development among young children. But the current differences as measured through these samples as of the 1990s are not figments of anyone’s imagination. Lane, Kamin, and others who have attempted to discredit
The Bell Curve
by focusing on our “tainted sources” have ensured that the African data will get more attention.

The third line of attack on
The Bell Curve
that I predict will have an unintended outcome is the attempt to dismiss the statistical power of the book’s results. Perhaps the most important section of
The Bell Curve
is Part II, the series of chapters describing the relationship of IQ to poverty, school dropout, unemployment, divorce and illegitimacy, welfare, parenting, crime, and citizenship, using non-Latino whites from the National Longitudinal Study of Youth. The eight chapters in this part deal with questions like, “What role does IQ play in determining whether a woman has a baby out of wedlock?” Or: “What are the comparative roles of socioeconomic disadvantage and IQ in determining whether a youngster grows up to be poor as an adult?” These are fascinating questions. But you will have a hard time figuring out from the published commentary on
The Bell Curve
that such questions were even asked, let alone what the answers were.

Instead, the main line of attack has been that no one really needs to pay any attention to those chapters because Herrnstein and I are confusing correlation with causation, IQ really does not explain much of the variance anyway, and even if that were not true, our measure of socioeconomic background is deficient. On all three counts, the critics are setting up a reexamination of the existing technical literature on social problems that will be embarrassing to them in the end.

First, regarding correlation and causation, read pp. 122-124 of the Introduction to Part II. Reduced to its essentials: The nonexperimental social sciences cannot demonstrate unequivocal causality. In trying to explain such social problems as poverty, illegitimacy, and crime, we use statistics to show what independent role is left for IQ after taking a person’s age, socioeconomic background, and education into account. When there are other obvious explanations—family structure, for example—we take them into account as well. Apart from the statistics, we describe in commonsense terms what the nature of the causal link might be—why, for example, a poor young woman of low intelligence might be more likely to have a baby out of wedlock than a poor young woman of high intelligence. At the end of this exercise, repeated in similar form for each of the eight chapters in Part II, there will still be unanswered questions, and we point out many of those unanswered questions ourselves. But readers will know more than they knew before, and the way will be opened for further explorations by our colleagues.

The statistical method we used throughout is the basic technique for discussing causation in nonexperimental situations: regression analysis, usually with only three independent variables. We interpret the results according to accepted practice. To enable readers to check for themselves, the printout of all the results is shown in Appendix 4.

The assault on this modest but useful analysis has been led by Leon Kamin in his
Scientific American
review. He argues that we cannot disentangle the role of IQ from socioeconomic background and suggests that in our database, the children of laborers have such uniformly low IQ scores that we cannot possibly tell whether the low IQ or the disadvantaged background is to blame for the higher rates of crime, unemployment, and illegitimacy that afflict such youngsters. “The significant question,” Kamin writes, “is, why don’t the children of laborers acquire the skills that are tapped by IQ tests?”
¹⁹

The answer to his significant question is that they
do
acquire such skills often enough to permit a good look at the comparative roles of socioeconomic background and IQ. Of the non-Latino whites used in the analyses throughout Part II, 1,589 came from families in the bottom quartile on our SES index. Of these, 451 had above-average IQs and 147 were in the top quartile of IQ. As we report throughout Part II, the results are encouraging: In America, bright children of laborers tend to do well in life, despite their humble origins. Herrnstein and I suggest that such a pattern points to causation. This is indeed an inference—a sensible inference.

We approach the correlation-causation tangle in other sensible ways as well. Consider the vexing case of education. People with high IQs tend to spend many years in school; people with low IQs tend to leave. Does the IQ cause the years of education, or the years of education the IQ? As we also discuss in the Introduction to Part II (pp. 124-125), it is unwise, for various technical reasons, to enter years of education as an additional independent variable, so instead we define two subsamples, each with homogeneous education: one of adults who had completed exactly twelve years of school and obtained a high school diploma, no more and no less; the other of adults who had completed .exactly sixteen years of school and obtained a bachelor’s degree, no more and no less, enabling us to report the independent effect of IQ for people with identical education.

Our procedure irritated a number of academic critics, who grumble
that the state of the art permits much more. Yes, it does, and in the book we mention periodically how much we look forward to watching our colleagues apply more sophisticated techniques to unanswered questions. But more sophisticated modeling techniques would also have opened a wide variety of technical questions that Herrnstein and I wanted to avoid. The procedure we chose gives an excellent way of bounding the independent effects of education, and that was our purpose.

But let us say a critic grants the existence of independent relationships between IQ and social outcomes after holding other plausible causes constant. How important are these “independent relationships”? Trivially so, says Stephen Jay Gould in his
New Yorker
review.
The Bell Curve
can safely be dismissed, he says, because IQ explains so little about the social outcomes in question—just a few percent of the variance, in the statistician’s jargon.

Here is the truth: The relationships between IQ and social behaviors that we present in the book are so powerful that they will revolutionize sociology. They are not only “significant” in the standard statistical sense of that phrase but are as well powerful in a substantive sense—often much more powerful than the relationships linking social behaviors with the usual suspects (education, social status, affluence, ethnicity). Not only are the attacks on these relationships unwarranted, but Herrnstein and I actually understate the strength of the statistical record. The story is complex but worth recounting because it tells so much about the academic response to
The Bell Curve.

In ordinary multiple regression analysis, two statistics are of special interest. The first is the set of regression coefficients, one for each independent variable, explaining the magnitude of the effect each independent variable has on the dependent variable after taking the role of all the other independent variables into account. Each coefficient has a standard error, which may be used to determine whether the coefficient is statistically significant (i.e., unlikely to have been produced by chance). The second statistic of special interest is the square of the multiple correlation,
R
2, which tells how much of the variance in the dependent variable is explained by all the independent variables taken together.

One of the early topics about multiple regression that graduate students study is the different uses of regression coefficients and
R
2. If a coefficient is both large in a substantive sense and statistically significant, it is typically the statistic of main interpretive importance, while
R
2 is of secondary and sometimes trivial importance. Such is the case with
the kind of analysis in
The Bell Curve,
for reasons we explain in Appendix 4. In all this, we treat our data as our colleagues around the country treat regression results every day. There is nothing controversial here, as evidenced by the fact that none of the quantitative social scientists who reviewed this part of the manuscript for
The Bell Curve
raised a question about our methods.

But that is not the end of the story. Herrnstein and I refer to the
R
2s in the introduction to Part II and in Appendix 4 as if they represent “explained variance”—and thereby we commit a technical error, falsely
understating
the overall explanatory power of our statistics. In logistic regression analysis with binary dependent variables, the kind of analysis we used throughout Part II, the statistic labeled
R
2 is an ersatz and unsatisfactory attempt to express the model’s goodness of fit. Most statisticians to whom I have talked since say we should have ignored it altogether. Stephen Jay Gould fell into the same error.

Once again, Gould’s criticism has been picked up by many others. It would be nice if a few respected professors would publicly point out that whatever else one might think about
The Bell Curve,
the criticisms about the small
R
2s in
The Bell Curve
are wrong. But this is unlikely to happen. Probably the allegation will quietly fade away as the academics who know the true story discreetly impart the news to those who do not.

The unfounded criticisms of the statistics in
The Bell Curve
that I have discussed so far will cause merely embarrassment among a few who both understand the issues and have the decency to be embarrassed. The real potential for backfire in the statistical critique of
The Bell Curve
comes from the attack on our use of socioeconomic status (SES).

Measures of SES are a staple in the social sciences. Leaf through the dozens of technical articles in sociology and economics dealing with issues of success and failure in American life, and you will frequently find a measure of SES as part of the analysis. A major purpose of
The Bell Curve
was to add IQ to SES as an explanatory variable. To avoid controversy, we deliberately constructed an SES index that uses the same elements everybody else does: income, occupation, and education. We did not have any reason for weighting any of these more heavily than the other, so we converted them to what are called “standard scores” and added them up to get our index, all of which would ordinarily have caused no comment.

But when it comes to
The Bell Curve,
a standard SES index suddenly becomes problematic. James Heckman notes ominously that we did not
have income data for a large part of the sample.
²⁰
Arthur Goldberger looks suspiciously on the idea of standardizing the variables.
²¹
Leon Kamin figures that probably the self-reports of income, education, and occupation are exaggerated in ways that falsely produce the relationships we report.
²²

My response to such criticisms is, “Fine. Let’s test out these potential problems.” Compare the results for the subsamples with and without income data. Do not standardize the variables; create some other scales, and use some other method of combining them. Examine the validity of the self-report data. If one does not like the idea of using an index at all, there is a simple solution: Enter the constituent variables separately, and ask directly how parental education, income, and occupation compete individually with the independent role of the child’s IQ.