American Economic Association

Review Article: The Bell Curve by Herrnstein and Murray The Bell Curve: Intelligence and Class Structure in American Life. by Richard J. Herrnstein; Charles Murray Review by: Arthur S. Goldberger and Charles F. Manski Journal of Economic Literature, Vol. 33, No. 2 (Jun., 1995), pp. 762-776 Published by: American Economic Association Stable URL: http://www.jstor.org/stable/2729026 . Accessed: 06/08/2013 10:43
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Economic Literature.

http://www.jstor.org

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Journal of Economic Literature Vol. XXXIII(June 1995), pp. 762-776

Review

Article: Herrnstein

The and

Bell Curve Murray

By ARTHUR S. GOLDBERGER and CHARLES F. MANSKI University of Wisconsin-Madison

We are grateful to Glen Cain, Steven Durlauf, Robert Hauser, Leon J. Kamin, Jacqueline Macaulay, and Alexandra Minicozzi for their comments.

1. Introduction
25 YEARS AGO, Arthur Jensen (1969) kicked off the Great IQ Debate, arguing that intelligence is a highly heritable trait, that differences in intelligence across races are quite possibly genetic, and that the high heritability of IQ could account for the failures of compensatory education programs. Herrnstein (1971) added to the analysis by asserting that success in equalizing opportunity would make socioeconomic achievement increasingly dependent on genetic factors, and hence stratification would become increasingly rigid. Ten years ago, Murray (1984) argued that the War on Poverty was failing because the poor, responding to the incentives generated by government programs, were reducing their work effort and increasing their dependency. These several themes have now merged to produce The Bell Curve, a long and complex four-part book by the psychologist Herrnstein and the political scientist Murray (henceforth HM). According to HM, intelligence is a highly heritable trait, plays a critical role in socioeconomic achievement and social pathology, and is becoming increasingly unVER The Bell Curve: Intelligence and Class Strucin American By RICHARD J. Life. HERRNSTEIN AND CHARLES MURRAY. New York: Free Press, 1994. 845 pp. $30.00. ISBN 0-02914673.
4

equally distributed. This leads to an increasingly stratified society, a trend which compensatory interventions cannot halt. A. Sorting The Bell Curve begins with an introductory chapter on the nature and measurement of intelligence or IQ or cognitive ability, terms which HM use interchangeably through most of the book. Following this introduction, HM proclaim in Part I, entitled "The Emergence of a Cognitive Elite," that people have always been sorted by talent as well as by social class, but now we approach a world "in which cognitive ability is the decisive dividing force" (p. 25). The story begins in an exploratory, often anecdotal, mode. The enormous increase in college enrollment in the United States ordinarily would be thought of as decreasing rather than increasing stratification. But in Chapter 1, HM assert that "At the same time that many more young people were going to college, they were also being selected ever more efficiently by cognitive ability" (p. 33). Even more significant than the cognitive sorting between college and noncollege groups is the cognitive sorting within the college population. For the university system "became radically more efficient at sorting the brightest of the bright into a handful of elite colleges" (p. 38). 762

ture

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldberger and Manski: The Bell Curve

The sorting continues as people enter occupations. Consider the dozen occupations that make up HM's set of "high-IQ professions" (pp. 54-55): accountants, architects, chemists, college teachers, computer scientists, dentists, engineers, lawyers, mathematicians, natural scientists, physicians, and social scientists. HM speculate that, having grown so rapidlyin recent years, these professionsalong with senior management-must by now be absorbing most of the high-IQ population. (This is necessarily speculation because they present no time-series data on the association between occupation and IQ.) And so Chapter 2 concludes that occupational choice reinforces cognitive stratification. After digressing in Chapter 3 to discuss the relationship of IQ and job performance, HM use Chapter 4, which closes Part I, to carry further their argument that cognitive stratification is increasing. Along with many scholars, HM are intrigued by the increasing inequality in the American wage distribution. Indeed: mostof the increasing wage inequality during the past two and a half decades is due to changesin the demandfor the residualcharacteristics of workers rather thanto changesin the demandfor educationor experience. . . What then is this residual, this X factor, that increasingly commands a wage premium over and above education?. . . It could be rooted in diligence, ambition,or sociability. It could be associatedwith different industries or different firms . . . Or it could be cognitive ability . . . we believe that it includescognitiveability.(p. 97) For HM, increasing inequality is an irremediable consequence of a permanent shift in the intellectual requirements of jobs: Cognitivepartitioning througheducationand occupationswill continue, and there is not muchthatthe government or anyoneelse can do aboutit. Economicswill be the mainreason . . . the value of intelligencein the marketplaceis rising.(p. 91) They do not entertain the possibility that advanced technology may simplify tasks and so increase the demand for unskilled labor. Further, HM see increasing stratification

763

due to increasing assortative mating by cognitive ability: "Intermarriageamong people in the top few percentiles of intelligence may be increasing far more rapidly than suspected" (p. 113). As qualitative evidence, they offer a calculation about the offspring of a hypothetical marriage between an average Harvard man and an average Radcliffe woman, and an anecdote about a hypothetical elite New York law firm. As quantitative evidence, they cite a study by Robert Mare (1991) which in fact says that "educational homogamy increased from 1940 to 1970 and may have stabilized or decreased somewhat in the 1980s" (Mare 1991, p. 24). B. IQ and Job Performance In Chapter 3, HM argue that cognitive stratification in the labor market makes good economic sense because cognitive ability is a good predictor of job performance. Reviewing the literature on the association between measures of intelligence and job productivity, they conclude that "the overall correlation, averaged over many tests and many jobs, is about 0.4" (p. 72).1 They argue that this correlation has significant economic implications. The two key claims are: an employerthat is free to pick amongapplicants can realize large economicgains from hiring those with the highest IQs. An econwith omy that lets employerspick applicants the highest IQs is a significantly more efficient economy.(p. 64) Readers of this Journalwill recognize a potential fallacy of composition in moving from the first claim to the second one, as all employers cannot simultaneously pick applicants with the highest IQs. HM are aware of this but nonetheless speculate that the macro1 Actually, this "overall correlation" is not an average of the correlations reported in different studies, but rather reflects upward adjustments for unreliability and "restriction of range." The unreliability correction arises because both the tests and the performance measures are subject to random measurement error. The "restriction-of-range" correction (a version of the selection bias correction familiar to economists) arises because job performance measures are available only for those who were hired. For discussion, see John Hartigan and Alexandra Wigdor (1989, pp. 117-71).

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

764

Journal of Economic Literature, Vol. XXXIII (June 1995)

screening job applicants but prohibited the use of job-specific information. They are not relevant to assessment of the actual decision in Griggs. C. Genes Near the end of Chapter 4, HM raise the question "How much is IQ a matter of genes?" and proceed to answer it: In fact IQ is substantially heritable . . . the genetic component of IQ is unlikely to be smaller than 40 percent or higher than 80 percent . .. For purposesof this discussion, we will adopt a middlingestimateof 60 percent heritability, which, by extension,means that IQ is about40 percent a matterof environment.The balance of the evidence suggests that 60 percentmayerr on the low side. (p. 105) They go on to assert that As a general rule, as environments become moreuniform,heritability rises. Whenheritability rises, children resemble their parents more, and siblings increasingly resemble each other; in general, family membersbecome more similarto each other and more different from people in other families. (p. 106) They proceed to refer to "the limits that heritability puts on the ability to manipulate intelligence," asking us to imagine "a United States that has magically made good on the contemporary ideal of equality. Every child in this imaginary America experiences the same environmental effects, for good or ill, on his or her intelligence. How much intellectual variation would remain?" (p. 109). Their answer is that 60 percent of the current IQ variance would remain. Their story is tainted by two misconceptions. Heritability is not a measure of parentchild resemblance in IQ, nor is it a biological parameter that sets limits on the effectiveness of policy. The heritability of an observable characteristic is the proportion of its variance that is associated with variance in genetic factors. Imagine that an individual's observed IQ test score Y is the sum of her "genotype" Z and her "environment" U, so Y = Z + U. Imagine that Z and U are uncorrelated,

economic benefits of using IQ tests to select among job applicants could be huge "if testing tended to place the smartest people in the jobs where the test-job correlations are large, the spread of the productivity distributions is broad, the absolute levels of output value are high, and the proportions hired are small" (p. 86). They do not flesh out this brief verbal argument. As for the first claim, it may be at odds with HM's assertion that high IQ employees are paid a wage premium in the labor market. Furthermore, it is based on the "overall correlation"of 0.4 between measures of intelligence and job productivity. At best, this correlation (or rather the corresponding coefficient in a regression of productivity on test score) tells a prospective employer hiring at random from the labor force what gain in output to expect if he were instead to hire by test score. But surely employers are not now hiring randomly from the labor force, but rather from those who apply. And within that pool, employers who seek references, consider past experience, hold interviews, administer skills tests, etc. are not selecting randomly with respect to applicant ability. HM are critical of the 1971 U.S. Supreme Court decision in Griggs v. Duke Power Co., which prohibited American employers from using hiring criteria that produce disparate outcomes but do not have a "manifest relationship to the employment in question." In practice, this has inhibited employers from using IQ tests to screen job applicants but has allowed them to use tests of job-specific skills. How much would predictions of job performance be improved if employers were to add IQ scores to the job-specific test scores and other information that employers are allowed to use under Griggs? HM offer no empirical evidence on this central question. Instead, on pages 75-76 they present evidence, in the form of stepwise regression findings, on the reverse question: How much would predictions of job performance be improved if employers using IQ scores to predict job performance were to add job-specific test scores and other information as predictors? These findings would be relevant if the Supreme Court had allowed the use of IQ scores in

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldberger and Manski: The Bell Curve

so the variance of Y equals the variance of Z plus the variance of U: V(Y) = V(Z) + V(U). Dividing through by V(Y), we can write h2 + e2 = 1, where h2 =V(Z)/V(Y) defines heritability, and e2 =V(U)/V(Y)is its complement. In this classical biometric model, the genotype is a scalar index of the individual's genome or genetic constitution. The environment U is just the residual defined by U Y Z. Neither "genotype"nor "environment"are observed; they are latent variables, or statistical abstractions. What does heritability say about resemblance of relatives? With a few more assumptions, it turns out that for paired relatives denoted persons 1 and 2, corr(Y1,Y2) = h2 corr(Z1,Z2) + e2 corr(U1,U2). See Goldberger (1979). That is, the relatives' correlation on IQ is a weighted average of their genotypic correlation and their environmental correlation. Clearly heritability does not measure parent-child resemblance in IQ. The genotypic correlation for parent and child is about 0.5, and parents transmit environment as well as genes to the child: corr(U1,U2) is not zero. The fact that observed parent-child IQ correlations run about 0.4 to 0.5 is consistent with
h2= 1, and also with h2= 0.

765

2. Cognitive Ability or Socioeconomic Status: Which Is "More Important"?

Part II of The Bell Curve presents a series of empirical analyses of data on non-Latino white respondents to the National Longitudinal Survey of Youth (NLSY). Observing statistical associations between various social behaviors in 1990 and IQ measured in 1980, HM initially ask whether these associations hold up after controlling for (i.e., conditioning on) measured parental socioeconomic status (SES). But their analyses soon degenerate into the calling of horse races-discovering whether IQ or SES is "more important" as a determinant of social behaviors. They lead off in Chapter 5 by examining the probability of being in poverty. In the key section titled "Socioeconomic Background versus Cognitive Ability," HM conclude that "Cognitive ability is more important than parental SES in determining poverty," and offer this figurative advice to unborn children: "If a white child of the next generation could be given a choice between being disadvantaged in socioeconomic status or disadvantaged in intelligence, there is no question about the right choice" (p. 135). Subsequent chapters analyze other outcomes. Chapter 6 examines the determination of schooling. Here they state To raise the chances of getting a college degree, it helps to be in the upper half of the distribution for either IQ or socioeconomic status. But the advantage of a high IQ outweighs that of high status. Similarly, the disadvantage of a low IQ outweighs that of low status. (p. 143) Chapter 7 examines the determination of labor supply. Here they assert Most men, whatever their intelligence, are working steadily. However, for that minority of men who are either out of the labor force or unemployed, the primary risk factor seems to be neither socioeconomic background nor education but low cognitive ability. (pp. 15556) In Chapter 8, we learn that "Low cognitive ability is a much stronger predisposing factor for illegitimacy than low socioeconomic background" (p. 167). In Chapter 9, the message

If heritability is high, does that mean that policies are ineffective? HM's thought experiment called for equalizing environments, making V(U) = 0. Suppose instead that we preserve V(U) at it current value, but make U perfectly negatively correlated with Z by introducing an extreme compensatory policy. Then IQ variance would fall from V(Y)(h2 + e2) to V(Y)(h2+ e2- 2 h e) = V(Y)(h - e)2. So with h2= 0.6 and e2 = 0.4, this intervention would reduce IQ variance to ({G- 74)2 = 2 percent of its current value V(Y). Of course, such calculations are fatuous. Using a somewhat more substantial model, John Conlisk (1974) gave a thorough critique of Herrnstein's (1971) proposition that equalizing opportunity necessarily increases stratification. As far as we know, no geneticist, or animal breeder, uses heritability to set limits on the elCfectiveness of environmental change. Such use seems to be an affectation of some social scientists.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

766

Journal of Economic Literature, Vol. XXXIII (June 1995)

part caused by IQ and by SES; education may have a discontinuous effect on outcomes; education and IQ may be collinear; and education and IQ may have an interactive effect on outcomes. As we see it, none of these reasons has any force. Correlation among explanatory variables does not affect the interpretation of regression coefficients as descriptions of how mean outcomes vary with each regressor, holding the others fixed. Discontinuous and interactive effects are routinely captured through the use of flexible functional forms and, more generally, by nonparametric regression analysis. As for chains of causation, the possibility that education may be partly determined by IQ and SES does not preclude the possibility that, after controlling for IQ and SES, education still affects outcomes. Their main concession to the possible role of education as an intermediate step in the determination of adult outcomes is to rerun the logistic regressions separately for the subsamples of those whose education ended with a high school diploma and those whose education ended with a bachelor's degree. HM report those results in Appendix 4, but say
little about them in the text.3 In each chapter, Y is a binary outcome of interest. In Chapter 5, for example, Y=1 if a respondent's family was below the poverty line in 1989 and Y = 0 otherwise. In Chapter 8, Y = 1 if a woman's first child was born out of wedlock and Y = 0 otherwise. In Chapter 12, Y = 1 if a respondent scored "Yes" on the Middle Class Values index and Y = 0 otherwise.

is that "Socioeconomic background and IQ are both important in determining whether white women become chronic welfare recipients" (p. 198). In Chapter 10, on parenting, it is that "A white mother's IQ is more important than her socioeconomic background in predicting the worst home environments" (p. 222). In Chapter 11, it is that "On two diverse measures of crime, the importance of IQ dominates socioeconomic background for white men" (p. 249). In Chapter 12, on civility and citizenship, they find that the probability of scoring "Yes"on their "Middle Class Values" index rises with both socioeconomic status and IQ, but that the former relationship "was not as significant" as the latter (p.
265).2

A. The Analysis Behind the Conclusions

HM draw their conclusions on the relative importance of IQ and SES primarily from their estimates, on NLSY data, of a series of logit models for conditional probabilities, of the form Prob(Y = llIQ, SES, A), where A denotes the respondent's age. They refer to these logistic regressions as their "basic analysis." Many analysts have reported estimates of models relating behavior to measured IQ, SES, and other attributes (e.g., Christopher Jencks et al. 1979; Manski and David Wise 1983). Most analysts have considered education to be an intervening variable in the chain that runs from child backgroundto adult outcomes. But HM deliberately omit any measure of education as a conditioning variable in their basic analysis. On pages 124-25, they write that "the role of education versus IQ as calculated by a regression equation is tricky to interpret, for four reasons":education is in
2 Herrnstein and Murray define the Middle Class Values index as follows (p. 263): "It has scores of 'Yes' and 'No.' A man in the NLSY got a 'Yes' if by 1990 he had obtained a high school degree (or more), been in the labor force throughout 1989, never been interviewed in jail, and was still married to his first wife. A woman in the NLSY got a 'Yes' if by 1990 she had obtained a high school degree, had never given birth to a baby out of wedlock, had never been interviewed in jail, and was still married to her first husband. People who failed any one of the conditions were scored 'No.' "

The IQ variable is a normalized transformation of the respondent's score on the Armed Forces Qualifying Test (AFQT), which was administered in 1980 to virtually all NLSY respondents. The SES variable is an
3 Also, for some 1,400 of the NLSY respondents, some sort of early IQ score is available in school records. In Appendix 4 (pp. 590-92), HM develop a simple recursive modef in which years of schooling intervenes between that early score and the later IQ score obtained in the AFQI,T.The date of taking the first test appears as an explanatory variable, but it is unclear whether age does. They calculate that each year of schooling increases the later IQ score by about one point, which they say "is in line with other analyses," without citing any.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldbergerand Manski:The Bell Curve

Probability of being in poverty 30%

767

20% 1'

As IQ goesfrom low to high

10%

that the curve marked "As IQ goes from low to high" is sloped more steeply downward than the one marked "As parental SES goes from low to high." This is the primaryempirical finding on which HM base their conclusion, cited above, that "Cognitive ability is more important than parental SES in determining poverty." They present analogous graphs in Chapters 6 through 12 and interpret them in the same way.5 B. Concepts and Measures
Veryhigh (+2 SDs)

0% Verylow (-2 SDs)

As parentalSES goes from low to high

roles of IQ and parental Figure 1. The comparative SES in determining whether young white adults are below the povertyline
Source: Herrnsteinand Murray (1994, p. 134) Note: For computingthe plot, age and either SES

(for the solid curve) or IQ (for the dashed curve) were set at their mean values. index combining information on parental income, education, and occupation. The respondent's IQ, SES, and age are all measured in the standardized form (x - m)/s, where x is the raw value of the variable, m is its sample average, and s is its sample standard deviation.

If one defines concepts in a particularway, HM's conclusion about the determination of poverty follows immediately from their empirical analysis. Define a respondent's "cognitive ability"to be his or her AFQT score and, similarly, define "parental SES" to be HM's SES index. Define "more important"to mean that one curve in Figure 1 is steeper than the other. With these definitions, about the only reason to question HM's conclusion is that Figure 1 does not indicate the statistical precision with which the two curves are estimated, but this turns out not to be much of an issue (see the standard errors in footnote 4). But the link between HM's logistic regressions and their conclusions becomes frayed if
Over this range of SES values, the probability of poverty falls from P(Y = 110, -2, 0) = 0.12 to P(Y = 110, 2, 0) = 0.04. Analogously, the poverty-IQ curve sets SES = A = 0 and computes P(Y = 1IIQ, 0, 0) for -2 < IQ < 2. Over this range of IQ values, the probability of poverty falls from P(Y = 11-2, 0, 0) = 0.27 to P(Y = 112,0, 0) = 0.01. 5 How well do HM's logistic regressions fit the data? HM (pp. 593-94) downplay, quite properly in our opinion, the role of goodness-of-fit measures in assessing the validity of a regression. Indeed, in the text they pay no attention at all to goodness-of-fit measures. Still, their logistic regression computer output in Appendix 4 does include an R2, which they describe as "the square of the correlation between the set of independent variables and the dependent variable expressed as the logarithm of the odds ratio." But that description cannot be correct; after all, the log-odds ratio, or logit, is not even defined for individual observations, which have Y = 1 or Y = 0. What their computer output calls R2 is in fact one minus the ratio of the unconstrained and constrained log-likelihoods (constrained meaning setting all slopes at zero). As noted by William Greene ( 1993, pp. 651-53) it is by no means clear how that statistic measures fit.

HM collect the coefficient estimates for the various logistic regressions in an Appendix. In the text, they present their findings through a series of graphs that show how the probability of each outcome of interest varies over a four standard-deviation range of IQ or SES values, holding the other covariates fixed at their sample average values. Our Figure 1 reproduces the HM graph (p. 134) of the probability of being in poverty.4 Observe
4 The probability of being in poverty conditional on (IQ, SES, A) has the form P(Y = lIIQ, SES, A) = exp(x'p)/[1 + exp(x'P)].

In Appendix 4 (p. 596), HM give these coefficientestimates,with asymptotic standard errorsin parentheses: - 2.65 - 0.84 IQ - 0.33-SES - 0.024 A. x' (0.08) (0.09) (0.09) (0.07) HM reporttheirestimatesto eightsignificant digits.We haveroundedto the nearesthundredth forsimplicity. Recall that the regressorsare in standardized form. The poverty-SES curvegraphedin Figure1 sets IQ = A = 0 and computesP(Y= 110,SES, 0) for -2 < SES < 2.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

768

Journal of Economic Literature, Vol. XXXIII (June 1995)

insist that (p. 130) "their IQ scores were already as deeply rooted a fact about them as their height," and on page 691 assert that using IQ test scores from age 7 or 8 would not change the qualitative conclusions about the effect of IQ scores, but merely show weaker effects. Thus they consider the difference between early and late scores to be purely random error, rather than having an experiential component.

one is unwilling to define a concept simply by its measure. Then one must ask at least these questions: What is "cognitive ability," and how well is it measured by the AFQT? What is "parentalSES" and how well is it measured by the HM SES index? What does it mean to say that cognitive ability is "more important" than SES in determining poverty and other outcomes? We now consider these questions in turn.

C. Cognitive Ability and the AFQT

HM have much to say about the nature and measurement of cognitive ability. They devote their entire introductory chapter to this subject and summarize their perspective as follows: Here are six conclusions regarding tests of drawn fromtheclassical cognitive tradition, ability, that are by now beyond significanttechnical dispute: 1. There is such a thing as a general factor of cognitiveabilityon which humanbeings differ. or aptitude 2. All standardized tests of academic achievementmeasurethis generalfactorto designed some degree,but IQ tests expressly it mostaccurately. measure forthatpurpose it 3. IQ scoresmatch,to a firstdegree,whatever is that people meanwhen they use the word orsmartin ordinary language. intelligent notperfectly so, 4. IQ scoresarestable,although life. overmuchof a person's 5. Properly administered IQ tests are not biasedagainst social,economic, demonstrably groups. ethnic,or racial heritable, 6. Cognitive ability is substantially andno more nolessthan40 percent apparently than80 percent.(pp.22-23) Later in the book, they argue at length that the AFQT meets the criteria for a good IQ test; they devote all of Appendix 3 to this matter. The AFQT was administered to NLSY respondents when they were 15-23 year-old adolescents, whose educational experiences were in some cases completed. And so it is possible to view the test scores as measuring educational attainment, rather than as an inborn endowment of cognitive ability. But HM

D. Socioeconomic Environment and the HM SES Index

Whereas HM dwell at length on the concept and measurement of cognitive ability, their treatment of respondents' socioeconomic environment is cavalier at best. In the text, they do not even define the term "socioeconomic status;" they just use it. In Appendix 2, they define their SES index as follows:
The SES index was created with the variables that are commonly used in developing measures of socioeconomic status: education, income, and occupation. Since the purpose of the index was to measure the socioeconomic environment in which the NLSY youth was raised, the specific variables employed referred to the parents' status: total net family income, mother's education, father's education, and an index of occupational status of the adults living with the subject at the age of 14. (pp. 573-74)

They offer no opinion on the adequacy of their SES index as an expression of respondents' socioeconomic environment.
To us, the casualness with which HM treat socioeconomic environment is astonishing. At the beginning of Part II, HM observe that low cognitive ability is statistically associated with various socially undesirable behaviors and say: "We will argue that intelligence itself, not just its correlation with socioeconomic status, is responsible for these group differences" (p. 117). In practice, they simply take it for granted that their SES index-a rather ad hoc concoction of information on parental attributes-adequately captures the social environment within which a child grows up. This single variable carries the burden of expressing all aspects of the child's

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldbergerand Manski:The Bell Curve

upbringing, from family structure to sibling relationships to neighborhood characteristics.6

769

E. What Does "MoreImportant" Mean?

Let us now set aside all qualms about HM's use of the AFQT to measure cognitive ability and the SES index to measure a child's social environment. We must still confront HM's interpretation of the relative slopes of the two curves in Figure 1 as measuring the "relative importance" of cognitive ability and SES in determining poverty status. An immediate issue is that the normalized AFQT score and raw SES index are measured in qualitatively different units. Hence, comparison of slopes with respect to unit changes in the raw variables is not a well-posed question. HM's resolution of this most basic problem is to transform each variable x into the standardized form (x - m)/s where, as described earlier, m is the sample average of x, and s is its sample standard deviation. Given these transformations,HM measure cognitive ability and socioeconomic status in an apparently common unit-standard deviations. Thus the two curves in Figure 1 graph the probability of poverty status over four standard-deviationranges of IQ and SES values. Standardizationin the manner of HM-essentially using "beta weights"-has been a common practice in sociology, psychology, and education. Yet it is very rarely encountered in economics. Most econometrics textbooks do not even mention the practice. The
6 HM measure the four components of their index in standardized form and then average them. The burden on the index is further increased by the fact that many respondents are missing data on some components, in which 6ase HM simply use the available components to form the index. Of the 12,000 NLSY respondents (including nonwhites and the supplementary samples), only 7,500 reported data on all four components of the index (pp. 574-75). The parental income component is automatically missing for youth who are already living in their own households at the beginning of the survey; reported income then refers to the respondent rather than to his or her parents. As for the other components, the very fact that one is missing may indicate something about family structure. For example, if father's education is missing, that might indicate the father was not present during the respondent's childhood.

reason is that standardization accomplishes nothing except to give quantities in noncomparable units the superficial appearance of being in comparable units. This accomplishment is worse than useless-it yields misleading inferences.

Consider Figure 1. The slope of the poverty-SES curve depends on the standard deviation of raw SES-index values within the population of non-Latino white respondents to the NLSY; the larger the standard deviation of SES, the steeper the slope. If SES values were to vary more across the population, then the poverty-SES curve would be more steeply sloped than the one in Figure 1.7

We find no substantively meaningful way to interpret the empirical analysis in Part II of The Bell Curve as showing that IQ is "more important"than SES as a determinant of social behaviors. How might the phrase "more important" be given policy-relevant content? The answer was given years ago by Cain and Harold Watts (1970) in their critique of the Coleman Report on equality of educational opportunity (for a related discussion, see Goldberger 1991, pp. 240-41). The Coleman Report sought to measure the "strength" of the relationship between various school factors and pupil achievement through the percent of variance explained by each factor, an approach similar to that of HM. Cain and Watts write (p. 231): "this measure of strength is totally inappropriate for the purpose of informing policy choice, and cannot provide relevant information for the policy maker."They go on to offer an alternative approach:
7 Assume that the estimated logistic regression presented in footnote 4 provides a correct description of P(Y = lIIQ, SES, A). Consider how the poverty-SES curve in Figure 1 would appear if, all else equal, the standard deviation of the SES index were to grow to 2.5 times its current magnitude (as it might, if HM are correct about the upward trend in cognitive stratification). The coefficient on the standardized SES variable would change from -0.33 to -0.825, and the other coefficients would remain unchanged. Over the range -2 < SES < 2, the quantity P(Y = 110, SES, 0) would fall from P(Y = 110,-2, 0) = 0.27 to P(Y = 110,2, 0) = 0.01. Thus, the new poverty-SES curve would essentially coincide wit the poverty-IQ curve shown in Figure 1.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

770

journal of EconomicLiterature, Vol. XXXIII(June1995)

SES index. Earlier, we expressed our strong doubts about the adequacy of their SES index as a measure of a child's socioeconomic environment. HM have their doubts, too, but of a quite different sort. They worry that: socioeconomic statusis also a result of cognitive ability,as people of high and low cognihigh and tive abilitymove to correspondingly low places in the socioeconomiccontinuum. The reasonthat parentshave high or low socioeconomicstatus is in part a function of their intelligence,and their intelligencealso affectsthe IQ of the childrenvia both genes andenvironment. (pp. 286-87) At this point (and elsewhere), HM seem to have lost sight of the fact that the classical model on which they relied for heritability estimation-the Y = Z + U scheme-requires that genes and environment, Z and U, be uncorrelated. This requirement is difficult or impossible to reconcile with their story of parent-child transmission. HM acknowledge that on academic aptitude and achievement tests (e.g., the NAEP, SAT, and ACT), the gaps between the average scores of American blacks and whites have narrowed substantially in the past 20 years (pp. 290-92, 637-42). This they attribute to improved social, economic, and educational conditions for blacks, but they express some doubt that the trend will continue.

it would seem evident that our interest lies in purposive manipulation of the x's in order to effect an improved performance in terms of y. We can, and should, ask for the expected change in y induced by spending some specific amount of money (or political capital, man hours, etc.) on working a change in x2, say, as compared with the alternative of spending the same sum on X3.Budgetary cost is not necessarily the only basis of comparability. But unless some such basis is defined and its relevance to policy explained, the question of "strength"has no meaning. (Cain and Watts 1970, p. 231) To apply this approach in the context of The

Bell Curve, one could contemplate allocating some fixed sum to improve IQ or to improve SES. It would be meaningful to say that IQ is more important than SES if spending the sum on IQ improvement rather than SES improvement were to yield a larger expected change in some outcome of interest. 3. Ethnic Differences With Part III, entitled "The National Context," ethnicity, or race, comes to the fore. Recognizing the sensitivity of this topic, HM "believe that the best way to keep the temperature down is to work through the main facts carefully and methodically"(p. 271). In Chapter 13, on "Ethnic Differences in Cognitive Ability," they begin with the wellestablished empirical finding of an approximately one standard-deviation difference in IQ scores between American whites and blacks. They report that the average AFQT score of black respondents to the NLSY is 1.21 standard deviations below that of white respondents (p. 278). Drawing heavily on Jensen (1980), they argue in the text and in Appendix 5 that the AFQT and other ability tests are not biased against blacks, in the sense of underpredicting black achievement. They go on to argue that black-white differences in IQ cannot be accounted for by differences in SES. Their reasoning is that, when they regress AFQT score on a race dummy and their SES index, the coefficient on the race dummy implies a 0.76 standard deviation difference between the average IQ of blacks and whites who are matched on the

A. Genetics and Racial Differences in IQ

Now HM confront "the flashpoint of intelligence as a public topic: the question of genetic differences between the races" (p. 295). Their discussion begins wisely enough: in indiThata trait is geneticallytransmitted vidualsdoes not meanthat groupdifferences in that trait are also geneticin origin.Anyone who doubtsthis assertionmaytake two handfuls of genetically identical seed corn and plant one handfulin Iowa, the other in the MojaveDesert, and let nature(i.e., the environment)take its course.The seeds will grow and th9 resultwill in Iowa,not in the Mojave, have nothingto do with genetic differences.
(p. 298)

But in the very next paragraph they use the within-group heritability estimate of 0.6 as if

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldbergerand Manski:The Bell Curve

it were a between-group measure. To pursue their own analogy, they use information on the variance' of corn yields within Iowa and the variance of corn yields within the Mojave Desert to estimate the role of environment in accounting for differences in corn yields between Iowa and the Mojave. They apply their 0.6 estimate to calculate how much of a mean environmental difference there would have to be between ethnic groups to provide a purely environmental explanation of the mean IQ difference. Their conclusion: "Environmental differences of this magnitude and pattern are implausible" (p. 299). So they are encouraged to look for genetic sources of the ethnic differences. Relying primarily on work by Jensen, they note that black-white differences are largest on the more g-loaded tests (this refers to Spearman's g, or "general intelligence" measure), and they point out that "g or other broad measures of intelligence typically have relatively high levels of heritability" (p. 303). They summarize data on African blacks (pp. 288-89); for an evaluation, see Leon Kamin (1995). They cite two studies of mixed-race children matched in one or another way with shows no mean IQ difwhite children-one ference, the other shows some differences (pp. 309-10). It is in this context that they introduce us to the "Flynn effect"-the worldwide phenomenon of rising test scores. When past versions of IQ tests are given nowadays, average performance is higher than it was in the past. This tendency for IQ scores to move upward over time is quite strong, perhaps as much as one point per year. HM say, To put it another way, on the average, whites today may differ in IQ from whites, say, two generations ago as much as whites today differ from blacks today. Given their size and speed, the shifts in time necessarily have been due more to changes in the environment than to changes in the genes.
(p. 308)

771

tion in the distribution that has the most significant implications for social and economic life as we know it and also for the position of one's children" (p. 309). Further, they suggest in a footnote that the rise in scores may not reflect a rise in intelligence, but merely in test-taking skills: "A shifting link between IQ and intelligence is not only possible but probable under certain conditions" (p. 728). This sudden skepticism about the trustworthiness of IQ tests as measures of true intelligence reappears in Chapter 15, where they say that "Comparisonsbetween successive generations tested with the same instrument . . . were contaminated by the Flynn effect, whereby IQ scores (though not necessarily cognitive ability itself) rise secularly over time" (p. 346). After this extensive discussion of the possible role of genetic sources of group IQ differences, HM announce that "it matters little whether the genes are involved at all" (p. 312). Then they inform us that the distinction between genes and environment as determinants of IQ is logically irrelevant to the treatment of individuals socially or educationally, that this distinction says nothing about the malleability of IQ, and that environmental differences, just like genetic differences, may persist across generations. Their conclusion? "If tomorrow you knew beyond a shadow of a doubt that all the cognitive differences between races were 100 percent genetic in origin, nothing of any significance should change" (p. 314). To us, HM's treatment of genetics and race is akin to standing up in a crowded theater and shouting, "Let's consider the possibility that there is a FIRE!" The authors of The Bell Curve appear to lack the wisdom of an earlier writer on intelligence and race:
Virtually every commentator on what it is like to grow up black in America, whether novelist or sociologist or memoirist, has reflected on the devastating effects of racism on selfconfidence . . . When the real difficulties are compounded by the fears engendered by centuries of white propagandizing that white is smarter (and by elements of self-denigration by blacks), the result can be immobilization of even the most able and ambitious. (Murray 1984, p. 187)

Lest we be encouraged, they tell us that because the Flynn effect applies to both blacks and whites, it will yield no reduction over time in the black-white IQ gap. And, of course, "at any point in time, it is one's posi-

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

772

Journal of Economic Literature, Vol. XXXIII (June 1995)

lor's degree to be 0.27 for whites and 0.11 for blacks. Controlling also for IQ, and evaluating at IQ = 114 (the sample mean IQ for college graduates), they estimate the probability of attaining a bachelor's degree to be 0.50 for whites and 0.68 for blacks. So HM conclude that "after taking IQ into account, blacks have a better record of earning college degrees than . . . whites" (p. 320). Later in the book (Chs. 19 and 20), they use this and a similar finding on the probability of being in a high-IQ occupation to attack affirmative action policies in higher education and the workplace. Drawing strong
policy conclusions from primitive empirical analyses is characteristic of The Bell Curve. Whatever one thinks of affirmative action, a serious analyst would recognize that HM's logistic regressions of educational attainment on IQ and age are consistent with a wide variety of structural stories. For example, the HM findings are consistent with the hypothesis that, holding IQ and age fixed, blacks tend to be more motivated than whites, and also with the hypothesis that the AFQT test is, after all, an ability test that is biased against blacks. Still, the main point of the analysis is that IQ is a powerful variable in accounting for observed racial differences in outcomes. HM conclude Chapter 14 saying "Racial and ethnic differences in this country are seen in a new light when cognitive ability is added to the picture. Awareness of these relationships is an essential first step in trying'to construct an equitable America" (p. 340).

B. Ethnicity, IQ, and Behavior

In Chapter 14, HM resume their analysis of the NLSY. Many of the same outcome variables are taken up as in Part II, but now the sample is split into three racial, or ethnic, Latinos, and non-Latino groups-blacks, whites. Separate logistic regressions are run for each group, using only IQ, SES, and age as explanatory variables. Some linear regressions with wage and income as the dependent variables are also run (with occupational dummies added to the list of explanatory variables). Estimated coefficients on IQ and SES are reported in an Appendix (but not those for the age variable). In the text, the emphasis is entirely on the "shortened" regressions that use only age and IQ as explanatory variables, so the notion of controlling for SES has been abandoned, without explanation. For each binary outcome, the text presents a bar chart answering two questions for each ethnic group: First, what is the probability of the outcome conditional only on age? Second, what is the probability of the outcome conditional on age and IQ? HM often report the latter probabilities at the average age and IQ of all NLSY respondents, but sometimes choose another reference point. HM use these bar charts to show that, in comparing outcomes across ethnic groups, it matters whether one conditions on IQ. They "find that Latinos and whites of similar cognitive ability have similar social behavior and economic outcomes." But for blacks and whites, the story is more complicated. On two vital indicators of success-educational attainment and entry into black-white prestigious occupations-the discrepancy reverses. After controlling for IQ, larger numbers of blacks than whites graduate from college and enter the professions . . . In contrast, the B/W gap in annual family income or in persons below the poverty line narrows after controlling for IQ but still remains sizable. (p. 317) HM refer to the reversal in the black-white discrepancy in educational attainment as a "black advantage" (p. 320). Controlling only for age, and evaluating at age = 29 (the sample mean age of all NLSY respondents), they estimate the probability of attaining a bache-

C. Apocalypse Coming
With Chapter 15, entitled "The Demography of Intelligence," HM's excursion toward an apocalyptic vision of the American future begins in earnest. On page 341, they assemble evidence that "demographic trends are exerting downward pressure on the distribution of cognitive ability in the United States and that the pressures are strong enough to have social consequences." They report that "blacks and Latinos are experiencing even more serious dysgenic pressures than whites, which could lead to further divergence between whites and other groups in future gen-

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldberger and Manski: The Bell Curve

erations." And they warn that immigration is exacerbating the problem: "the self-selection process that iised to attract the classic American immigrant-brave, hard-working, imaginative, self-starting, and often of high IQhas been changing, and with it the nature of
some of the immigrant population." In brief, the lower IQ classes and races are outbreeding the upper ones. HM report that, among NLSY respondents, the lower the IQ of the mother, the earlier she has her first child; moreover, black and Latino women tend to have their first child earlier than do whites. HM recognize that the number of children ultimately born to NLSY respondents cannot be accurately assessed using data through 1990 because the respondents were still relatively young (age 25 to 33) at that time, and because "Presumably the new births will be weighted toward more highly educated women with higher IQ" (p. 350). Nevertheless, they use IQ data contained in the Child Supplement to the NLSY to generate worrisome forecasts of the future distribution of IQ in America.8 Of course, there is the Flynn effect-the observed secular increase in IQ scores. But on page 346, HM remind us that IQ score is not necessarily the same thing as cognitive ability itself.

773

cient lines of separation based on hereditary rank were being erased, replaced by a more complicated set of overlapping lines. Social standing still played a major rofe, if less often accompanied by a sword or tiara, but so did out-and-out wealth, educational credentials, and, increasingly, talent. Our thesis is that the twentieth century has continued the transformation, so that the twenty-first will open on a world in which cognitive ability is the decisive dividing force. The shift is more subtle than the previous one but more momentous. (p. 25) To develop this thesis, HM make two key empirical assertions: the increasingly technological character of modern society places an increasing premium on intelligence, and, mating patterns in America show increasing stratification by intelligence. In Part IV, HM push their thesis much further. Here they express concern that the United States is experiencing
* *

An increasingly isolated cognitive elite. A merging of the cognitive elite with the affluent. A deteriorating quality of life for people at the bottom end of the cognitive ability distribution. (p. 509)

4. Vision of the American Future A. The Coming of the Custodial State

Following the introductory chapter on the nature and measurement of intelligence, Part

They predict that the country is being transformed into "something resembling a caste society" (p. 509) in which the cognitive elite "will implement an expanded welfare state for the underclass that also keeps it out from underfoot" (p. 523). They label this the custo-

dial state.
HM develop their vision of a cognitively stratified custodial state in the last chapter of the book, entitled "A Place for Everyone." They pose this question: "How should policy

I of The Bell Curve begins this way:

The twentieth century dawned on a world segregated into social classes defined in terms of money, power, and status. The an8The Child Supplement, begun in 1986, provides data on the children of the NLSY respondents, and includes the Peabody Picture Vocabulary Test, a test said to be highly reliable and g-loaded. HM find that the mean IQ of the children was only 92, fully half a standard deviation below the national mean, and below the mean score of their mothers. They state "If we take the NLSY results at face value, American intelligence is plunging" (p. 355). They go on to say that the black-white IQ gap for the children is larger than the gap separating their mothers.

deal with the twin realities that people differ in intelligence for reasons that are not their fault and that intelligence has a powerful bearing on how well people do in life?" (p. 535).
Their answer has two parts. First, they assert that no practical policy instrument can raise cognitive ability. Early in Part IV they write: "For the foreseeable future, the problems of low cognitive ability are not going to be solved by outside interventions to make children smarter," (p. 389) and toward the end, they declare: "Cognitive partitioning will continue. It cannot be

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

774

Journal of Economic Literature, Vol. XXXIII (June 1995)

time t, individual output Y is determined by (IQ, X) through a production function Y = ft(IQ, X). Assume that individual output and earnings are identical. Then the distribution of earnings at time t is determined by the distribution of (IQ, X) and by the form of

stopped, because the forces driving it cannot

be stopped" (p. 551).9 Regarding schooling, a channel for intervention that comes readily to mind, they say this: "Formal schooling offers little hope of narrowing cognitive inequality on a large scale in developed countries, because so much of its potential contribution has already been realized with the advent of universal twelve-year systems" (p. 389). According to HM, the American educational system is already teaching most students as much as they are capable of learning: "The problem with American education is confined mainly to one group of students, the cognitively gifted" (p. 417). The second part of HM's answer is that society must adapt to the inevitability of increased cognitive stratification. But how? HM declare "The answer turns us back to the ancient concern with place" (p. 535): society will function effectively if everyone knows and accepts his or her appropriate place. The text of the book concludes with this: Inequality of endowments, including intelligence, is a reality. Trying to pretend that inequality does not really exist has led to disaster. Trying to eradicate inequality with artificially manufactured outcomes has led to disaster. It is time for America once again to try living with inequality, as life is lived: understanding that each human being has strengths and weaknesses, qualities we admire and qualities we do not admire, competencies and incompetencies, assets and desits; that the success of each human life is not measured externally but internally; that of all the rewards we can confer on each other, the most precious is a place as a valued fellow citizen. (pp. 551-52)

Jft(*,*).
Assume that, holding X fixed, ft(., X) is an increasing function of IQ. Also assume that endowments of IQ and X are statistically independent within the population. Then it follows that the economy at time t is characterized by a positive association between IQ and earnings. Now consider the economy at time t + 1. Assume that technological change makes ft+l(*, X) increase more with IQ thanft(o, X). Assume that the distribution of X remains unchanged between t and t + 1 but that increased stratification in mating patterns makes the distribution of IQ more dispersed at t + 1. Then the economy at time t + 1 is characterized by a stronger positive association between IQ and earnings than the economy at t. Moreover, the economy at t + 1 is characterized by greater earnings inequality than the economy at t. Finally, assume that there is no feasible intervention that can alter the distribution of IQ at time t + 1 or change the form of the production function ft+l(*,.). Also assume that it is not possible to use taxes or other means to substantially alter the identity of output and earnings. Then increasing cognitive stratification is inevitable.

B. Visions and Empirical Analysis

HM do not write in the standard language of economic analysis, but we may recast their vision of the American future in these terms. Here is a simple model that seems to us to capture their main assertions: Assume that each individual is endowed with an IQ and with other attributes X. At
9 There is only one type of intervention that HM are willing to endorse (p. 389): "The one intervention that works consistently is adoption at birth from a bad family environment to a good one."

While we find it relatively easy to interpret HM's vision in standard economic terms, we cannot similarly interpret the empirical analysis that HM use to support their vision. In Part I, they offer only scattered anecdotes, hypothetical vignettes, and selective citations of serious empirical studies to justify their assertions of increasing demand for intelligence, and increasing assortative mating by intelligence. In Part II they are obsessed with using the NLSY data to show that IQ is "more important" than SES in determining social behavior. The gist of the empirical analysis in Part III is that, in comparing outcomes across ethnic groups, some findings .depend on whether one controls for IQ. Whatever one makes of the NLSY regressions, these regressions offer no meaningful empirical evidence on the dynamic of American society.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

Goldberger and Manski: The Bell Curve

In Part IV, HM revert to the loose style of Part I to make their case that no practical policy instrumnent can raise cognitive ability. Where they do cite evaluations of policy interventions, they systematically slant their interpretation of the findings in two ways. First, whenever evaluations yield ambiguous results, HM take the position that, in the absence of firm empirical evidence for a treatment effect, one should maintain the null hypothesis of no effect. For example, on the effects of nutrition on IQ, their assessment is: "Many studies that seem to be wellconducted variations of the successful ones have failed to demonstrate any effect on IQ at all" (p. 393). "Failed to demonstrate any effect on IQ at all" must mean that, using classical statistical significance tests, these studies do not reject the null hypothesis of no effect. But every empirical researcher should know that "failure to reject" the no-effect hypothesis does not establish its truth. It simply implies that the empirical findings are consistent with the no-effect hypothesis; it may well be that the findings are also consistent with alternative hypotheses which assert substantial treatment effects. Second, HM repeatedly discount evaluations that do yield clearcut findings of treatment effects. Consider, for example, the Perry Preschool Project undertaken in Ypsilanti, Michigan in the early 1960s.10 Here is what HM have to say: Although this intensive attempt to raise intelligence failed to produce lasting IQ gains, the Ypsilanti group believes it has found evidence for a higher likelihood of high school graduation and some post-high school education, higher employment rates and literacy scores, lower arrest rates and fewer years spent in special education classes as a result of the year or two in preschool. The effects are small and some of them fall short of statistical significance. They hardly justifying investing
10 In this experiment, intensive educational and social services were provided to a random sample of about sixty black children aged three and four. No special services were provided to a second random sample of such children, drawn to serve as a control group. The treatment and control groups have been followed into adulthood. See John Berrueta-Clement et al. (1984).

775

billions of dollars in run-of-the-mill Head Start programs. (p. 405) The rhetoric of this assessment is revealing. HM begin by portraying the Perry Preschool Project as a failed attempt to raise IQ. In fact, the investigators have long been concerned with the effect of the intervention on a range of social behaviors, such as those that occupied HM in Parts II and III. They conclude by stating that the findings of the Perry Preschool Project do not justify "investing billions of dollars in run-of-the-mill Head Start programs," when they should instead ask what the findings reveal about the benefits of intensive interventions. En route, they belittle the Perry Preschool findings by saying that the group of investigators "believes it has found evidence" for various outcomes and by saying that "The effects are small and some of them fall short of statistical significance." We find it difficult to reconcile these statements with the wellknown findings of the project. For example, the Ypsilanti investigators have reported that 67 percent of the treatment group and 49 percent of the control group were high school graduates by age 19 (see Berrueta-Clement et al. 1984). This effect is neither small nor statistically insignificant by conventional criteria. We conclude that The Bell Curve is driven by advocacy for HM's vision, not by serious empirical analysis. America may or may not be on the way towards a custodial state. Policy interventions may or may not be effective. We know no more after studying The Bell Curve than we did before.

C. Lost Ground
That antipoverty programs are ineffective, indeed counter-productive, is not a new theme for Murray. In an earlier book, Losing Ground, his critique emphasized the rationality of the poor, unwed, dropouts, and criminals: I will suggest that changes in incentives that occurred between 1960 and 1970 may be used to explain many of the trends we have been discussing. It is not necessary to invoke the Zeitgeist of the 1960s, or changes in the work ethic, or racial differences, or the com-

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions

776

Journal of EconomicLiterature, Vol. XXXIII(June1995)

ternal scientific scrutiny, but with beautifully orchestrated initial publicity. A vast stream of reactions in the general media followed immediately (e.g., New York Times Magazine, October 9, 1994; Newsweek, October 24, 1994; Time, October 24, 1994; The New Republic, October 31, 1994). Through essays like ours, a process of scientific review is now under way. But, given the process to date, peer review of The Bell Curve is now an exercise in damage control rather than prevention.
REFERENCES BERRUETA-CLEMENT, JOHN ET AL. Changed

plexities of postindustrial economies, in order to explain increasing unemployment among the young, increased dropout from the labor force, or higher rates of illegitimacy and welfare dependency. All were results that could have been predicted (indeed, in some instances were predicted) from the changes that social policy made in the rewards and penalties, carrots and sticks, that govern human behavior. All were rational responses to changes in the rules of the game of surviving and getting ahead . . . I begin with the proposition that all, poor and not-poor alike, use the same general calculus in arriving at decisions; only the exigencies are different. (Murray 1984, pp. 154-55) In contrast, Part III of The Bell Curve ends with this passage: The lesson of this chapter is that large proportions of the people who exhibit the behaviors and problems that dominate the nation's social policy agenda have limited cognitive ability. Often they are near the definition for mental retardation . . . When the nation seeks to lower unemployment or lower the crime rate or induce welfare mothers to get jobs, the solutions must be judged by their effectiveness with the people most likely to exhibit the problem: the least intelligent people. And with that, we reach the practical questions of policy that will occupy us for the rest of the book. (p. 386)

lives: The effects of the Perry preschool program on youths through age 19. Ypsilanti, MI:
High/Scope Press, 1984. CAIN, GLEN G. AND WATTS, HAROLD W. "Problems in Making Policy Inferences from the Coleman Report," American Sociological Review, Apr. 1970, 35(2), pp. 228-42. CONLISK, JOHN. "Can Equalization of Opportunity Reduce Social Mobility?" Amer. Econ. Rev., Mar. 1974, 64(1), pp. 80-90. GOLDBERGER, ARTHUR S. "Heritability," Economica, Nov. 1979, 46(184), pp. 327-47. - A course in econometrics. Cambridge, MA: Harvard U. Press, 1991. GREENE, WILLIAM H. Econometric analysis. 2nd ed. New York: Macmillan, 1993. HARTIGAN, JOHN A. AND WIGDOR, ALEXANDRA K., eds. Fairness in employment testing: Valid-

ity generalization, minority issues, and the General Aptitude Test Battery. Washington, DC:
HERRNSTEIN,

While Murray'srationale has changed drastically, his policy conclusions have remained the same from 1984 to 1994. 5. A Concluding Comment on Process A serious scientific book should be the culmination of a program of research that has been subjected to external scientific scrutiny, revised appropriatelyin the light of that scrutiny, and iteratively honed into a well-rea-

National Academy Press, 1989. RICHARD J. "IQ," Atlantic Monthly, Sept. 1971, 228, pp. 43-64.

JENCKS, CHRISTOPHER

ET AL. Who gets ahead: The determinants of economic success in America. New York: Basic Books, 1979.

soned and credible final form. In this paradigm, research that purports to be scientific would first be reviewed on its scientific merits. Only if that review is passed successfully would society at large be concerned with the research. HM and their publishers have done a disservice by circumventing peer review. The Bell Curve was sprung full blown without ex-

JENSEN, ARTHUR. "How Much Can We Boost IQ and Scholastic Achievement?" Harvard Educational Review, Winter 1969, 39(1), pp. 1123. JENSEN, ARTHUR. Bias in mental testing. New York: Free Press, 1980. KAMIN, LEON J. "Review of The Bell Curve," Scientific American, Feb. 1995, 272(2), pp. 99103. MANSKI, CHARLES F. AND DAVID A. WISE. College choice in America. Cambridge, MA: Harvard U. Press, 1983. MARE, ROBERT. "Five Decades of Educational Assortative Mating." American Sociological Review, Feb. 1991, 56(1), pp. 15-32!' MURRAY, CHARLES A. Losing ground: American social policy, 1950-1980. New York: Basic Books, 1984.

This content downloaded from 186.18.32.91 on Tue, 6 Aug 2013 10:43:08 AM All use subject to JSTOR Terms and Conditions