You are on page 1of 23
MERITOCRACY AND ECONOMIC INEQUALITY Kenneth Arrow, Samuel Bowles, and Steven Durlauf, Editors PRINCETON UNIVERSITY PRESS PRINCETON, NEW JERSEY 2000 Seven__ A Reanalysis of The Bell Curve: Intelligence, Family Background, and Schooling SANDERS KORENMAN AND CHRISTOPHER WINSHIP ‘Siice 11s PuBLicaTION at the beginning of October 1994, The Bell Curve by the late Richard Hermstein and Charles Moray has been discussed in more than one thousand articles inthe public and acedemic press Initial commen- tary focused primarily on the book's treatment of race. The majority ofthese essays were negative, with many denouncing the book as racist, More recent reviews (e.g Heckman 1995; Hunt 1995; Goldberger and Manski 1995) fave focused on the disjunction between the evidence presented and the strong conclusions drawn by the authors Hermstein and Murray srgue in The Bel! Curve that intelligence is the most important determinant of socal and economic success in present-day ‘America. They support this conclusion with statistical analyses that suggest that a youth's intelligence (measured at ages fifteen to twenty-three by the ‘Armed Forces Qualifications Test (AFQT) is considerably more important than his or her parents" sociat and economic status (SES) in determining social and economic status in adulthood, the well-being of a woman's chil- dren, and the avoidance of antisocial behaviors. In their analysts, the effect of AFQT score is more than twice as lage asthe effect of parents" SES in predicting whether, 2t ages twenty-five to thirty-two, someone (1) is poor, (2) dropped out of high school, (3) is unemployed, (4) had a child out of wedlock, (5) had been on welfare, (6) had a low-binh-weight baby, oF (7 had a child with low 1Q scores. In this chapter, we reanalyze Hermstein and + We patel acknowledge the nccal support ofthe Russel Sage Foundation. We hank CChares Momay fee providing» copy ofthe data ued in The Bell Curve and for anewering umes questions nbout the aralycs therein, We thaak John Heund Chrstopher Seacks, Lany Kat, Cares May, Doog Stger and sear puricipants atthe NBER, he Has ‘School of Public Poy Sea at Univesity of Chicago, RAND Corprain, UC Santa Bar tara, Univesity of Mines, ad the University of Wisconsin, Mads, fr tice wggesons ‘We a resonebe fray er. An ear vesion of thi pape pened at NBER Working Paper 9. 5250, Angst 195, T Se alio Gliberger and Mansi 1985, p. 765-166, fora suciet summary of Herasteln and Murray's conlsons regading tbe importance of AFQT sore minve to pares SES. 138 KORENMAN AND WINSHIP ‘Murray's data in order fo assess whether their principal conclusions are warranted. ‘Our analyses address three questions, The first two are related to whether Hermstein and Murray's estimates of the effects of AFQT score and parents’ SES are biased by measurement error (especially in parents’ SES) or by ‘unmeasured or omitted parental SES or other family background characters tics. The third involves their treatment of schooling. (1) Ts Hermstein and Murray's index of parents’ socioeconomic status adequate for isolating the effects of AFQT score on economic and social success in adulthood? Ideally, in order to isolate the effeets of AFQT on ‘adult outcomes, one Would like to bold constant—witk perfect measures all aspects of family background that influence both adult outcomes and ‘AFQT scores. Hermstein and Murray employ a narrowly conceived and poorly measured index of parenta] SES, s Heckman (1995) notes: “The statistical methods used by Herastein and Muray are vulnerable 19 measure. ment ecor, It would be ineredibe if 15 10 23 years af environments inftueaces, incfusing the nurturing of parents, the esources they spent on a child tei cultura tnvironment, thei Intractons with ther children andthe influence ofthe larger ‘community on the children could be summarized by a single megure of edvcaton ‘cupation and family income in one yea. I environnent is poorly measured bat affects the test score—and thee is slid evidence of environmental impacts on test scores—the Murray and Herastein fining that 1Q uta stronger impact on socio= conomie outcomes than measured environment may simply arse from the poor quality oftheir measure of the envionment. Their measure of IQ proxies the mis ‘measured envircninental variable. (p. 21) In the first part of our analysis we use comparisons of siblings in order to control more completely and broadly for differences in family background Characteristics that may in@aence APQT scores and adult outcomes. We esti- mate the eects of AFQT score net of family background by relating differ- fences between siblings in adult putcomes to differences in their AFQT ‘sores (controlling forage and gender). In effect, a youth's siblings) acts as his or her “control group.” Incredible as it may seem, our sibling analyses suggest that, even though Hermstein and Murray's parental SES index is poorly measured, «appears to Be adequate for producing unbiased estimates Of the effects of AFOT score on Socioeconomic outcomes. 2) Is Hermstein and Murray's measure of parents’ socioecanomic status adequate for estimating the effects of either parental SES ot family socio- ‘conomie backround more broadly conceived on social and economic suc- cess in atulthood? Here there are two problems. (1) Random measurement terror in the parental SES index will bias downward the estimated effects of, parental SES. (2) Hermstein and Murray's index of parents’ SES may fail to Capture important components of parents’ SES and other environmental ine BACKGROUND, AND SCHOOLING 139 ‘Auences shared by family members (such as neighborhoad and school chay- acteristics). Regarding te fist point, because Herrastein and Murray's index of parental SES is highly correlated (0.58) with APQT score, and because, as wwe shall see, parental SES is less reliably measured than is AFQT score, Hermstein and Murray's estimates may substantially understate the effects of parents’ SES and overate the effects of AFQT score on adult outcomes, ‘This is che classical ervors-in-variables problem, and the potential for bias is easily demonstrated for a subset of Hermstein ang Murray's adult outcomes that are continuous (as opposed to binary) variables. For these outcomes, we adjust estimates for measurement error using a range of values for the re- liability of APQT score and parents’ SES, We find evidence of substantial downward bias in their estimates of the effects of parents” SES. Next, we investigate more directly the consequences of Hermstein and Murray's narrow conceptualization of parents’ SES. As reviewers have noted, Herrstein and Muray’s index of parental SES covers an important but limited range of socioeconomic attributes of the parental family. For ‘example, Goldberger and Manski (1995, pp. 768-769) remark: “In practice they simply take it for granted that their SES index —a rater ad hoc concoc- tion of information on parental attibutes—adequately captures the socio- economic environment within which a child grows up. This single variable carries the burden of expressing al} aspects of the child's upbringing from Family structure to sibling relationships to neighborhood characteristls.’ ‘We find evidence that Hermstein and Murray's index of parents’ SES produces substantially misleading estimates ofthe effects of parental family socioeconomic status on social and economic outcomes of youths. Herrns- fein and Murray's index of parents’ SES fails to capture components of socioeconomic family berkground that are demonstrably important determi rants of adult outcomes. We illustrate this point in two ways that together form upper and lower bounds for the effects of family socioeconomic background. ‘To obtain lower-bound estimates we first add to Herrnstein and Murray's models a variety of socioeconomic family background controls including indicators of parenta) fomily arrangement (e.g., single-parent family, step- parent) at age founeen, family structure (e.g., number of siblings), urban/ rural residence at age fourteen, as well as other aspects of the home eaviron- ment at age fourteen. We combine the effects of Herrnstein and Murray's parental SES index withthe effects of this richer set of socioeconomic back- aground controls to form a single standardized composite effect of family socioeconomic background. The composite effect is dramatically larger than the effect of parental SES alone, and is sometimes larger than the effect of 2.4 alr pot ried by Fischer et a, (1996). Tey ao emphasize he importance of race und peer in the determination of il and econo sat adlbood,. 40 KORENMAN AND WINSIE |AFOT score. Nonetheless, these composite effets are lower bounds forthe Sifees of family sociosconomic background because they ae based on the necessarily limited set of imperfecdy measured family background charac- teristics avilable in the National Longitudinal Sudy of Youu ‘Our upper-bound estimates of the effects of family socioeconomic back- ground are based on analyses of siblings. These estimates are upper bounds Eecause they are derived under the assumption that anything common to ings other than AFQT score, age, and gender is atihbutable to fenily background. These residual effects of family background are far larger than Hernstein and Murray's estimated effects of parental SES, and are at tines to to four times as large asthe effects of AFQT score, (9) Are Hermstein and Murray's estimates ofthe effets of AFQT score sensitive to ther tratment of education? Does schooling have an effect on diferent outcomes, controling for AFQT score? For a vaiety of reasons, Fermstein and Muray Were reluctant to include education controls in their models of vious outcomes (pp. 124-125). Hermstein and Murray do esi- inate their models for edveationa'y homogenous subsamples (i. high schoo! graduates; college graduates). Nonetheless, itis difficult wo get a sense from their analyses either of the sensitivity ofthe effects of AFQT to the inclusion of education controls, or, pechaps more important, of the size of the schooling effects, contaling for AFQT score. It may be important to txamine more carefuly the rote of eduaton, however, given the potential for public policy to change educational atainment and thus, possibly, indi- vida outcomes. Tn our analytes we find that for many oucomes the effects of AFQT are substantially reduced by the inclusion of education controls, Furthermore, for many outcomes the standardized effect of scholing is larger than the cffectof AFQT. This suggests that even if Hermstein and Murray are corect that AFOT is largely immutable and unaffected by schooling, attempts to raise educations atzinmeat may nonetheless be important, due to its posi tive partial effect on a variety of socal and economic outcomes. Outline of The Bell Curve ‘The Bell Curve is divided into four sections. In part one Hermstein and ‘Muray argue that America is becoming increasingly dominated by a cogri tive eite. They discuss the increasing selectivity of elite universities and colleges and the rising educational credentials of top managers. In part two they present an extensive set of original analyses aimed at demonstrating that inelligence is the principal determinant of a variety of social and eco- ‘nomic outcomes. We are concemed with this portion ofthe book. The third section examines previous work on racial differences in intelligence and pre- INTELLIGENCE, BACKGROUND. AND SCHOOLING “ sents new analyses of the importance of AFQT score as a determinant of different outcomes across racialethnic groups. The final section of the book discusses 2 variety of policy issues, most notably affirmative action, ‘Much of The Bell Curve reviews and interprets the analyses and data of others, The exceptions are the chapters in section Il, and chapters 14 and 16 of section Ill, in which Hermstein and Murray present originat analyses of the National Longitudinal Survey of Youth ‘The analysis methods used by Hermstein and Murray are those commonly ‘employed in the sacal sciences, and theit approach to the data i straightfor- ‘ward and clearly explained? As noted, Hermstein and Musray’s principal conclusion is that for al! raciaf and ethnic groups and across a variety of social and economic outcomes, an individual's AFQT scare is a more impor tant determinant than is the social and economic status of his parents, Methods and Data ‘The original analyses in The Bel! Curve all use the Department of Labor's National Longitudina) Survey of Youth (NLSY). The NLSY. is an ongoing longitudinal study of approximately 12,000 youths aged fourteen t0 twenty fone as of January 1, 1979 (Center for Human Resource Research 1994), In section one of their book, Hermstein and Murray restrict their analyses to whites. In chapter 14 and appendix 6 they repeat these analyses for blacks and Latinos. In our analyses, we estimate models for the entice sample and enter controls for rae, ethnicity, and sex. Most of the time we have done 80 to ensure that we had the largest sample possible. As explained below, saz- ple size becomes an important concern in analyses of sibling differences, We have also repeated the analyses for black, Latino, and white subsamples for continuous outcomes (income, wages, and years of schooling) where sample sizes are sufficient to permit analysis of sibling differences, The results of these analyses, which we preseat ia appendix A, parallel those for the full sample. Table 7.1 provides a description of the outcome variables and samples from The Bell Curve that we have used in our analysis. Our approach to the analysis is to use Hermstein and Murray’s daca (supplied to us by Murray) ‘and estimate models analogous to their models. We then report alternative estimates based on different assumptions or modeling strategies. Because we Present some models that include controls for years of schooling completed 2 They we ier an opt epresion, nd estimate the eect of ARQT sore and pen ‘SS on diferent outcomes whey age i conte, In models for some outcomes ey cde ‘one aditonl oeuolvarible or rest the analysis sample as way 1 "conta for one imporant career (eg they sty sample of poor ether Inte analyses of Welive tae pp. 122-125 fora desmpton of hel modeing sey) 1 KORENMAN AND WINSHIP ‘TABLE 7.1 Unweighted Sample Means, (SDs), (Analysis Sample Sizes), and Descriptions of ‘Analysis Variables ‘Analysis Variables ‘Mean (so) Descriptions of Variables Lots} ‘and Sample ‘BAA Peal net Family income In 1989, (27080) 1990 dollars. Excludes persons not 7977 working because of school in Family income (19908) in 1989) 1989 or 1990, a povey in 1989 115 Tne frly income below x " (0.36) U.S. Census poverty line, Ex- ‘T9T7 clues persons not working be- ‘eause of schoo! in 1989 or 1990. ‘Annual eamings (19908) in 1989 24,225 Yearround wavkers «iso ‘one Year ot tooling completed OL 1550 some) “om ass = idm ot Spa, ine ae (0.39) ing those who later eamed a GED ais Bade 218 Obed a bachelor's deer or (0.38) higher. Excludes persons enrolled 3585s undrrauntes 190. 004 Bele persons eld i ol eee (0.20) Jege or graduate schoo! in 1990 tos gf LF 1+ in 1989,men 01S om or os aus templet 1-4 omy in 1989, O10. Btclades persons not working be ae (60) of iol 5 F580 bas ect by 630 72 Exes pens nde 30 : 7 (0.45) 1990 interview (H&M exclude age fan sacomred ivr Sean of marige “020 a0) se pics aes inde OST Men: HS gad 4 in LE fl yar Lome (0.48) + never in jail + married to first 2 wit, Womens HS gd oc abl cmsinser) INTELLIGENCE, BACKGROUND, AND SCHOOLING 3 TABLE 7.1 (continued) ‘Mean (sD) Descriptions of Variables 08s) and Sample ‘of-wedlock births + never in jail “+ married to fist husband. Ex- cludes single persons who met other conditions and men who were disabled or enrolled in school, ver interviewed in jil, men on 025) 4309 ‘utp ovTCOMES.RSTBORN CHILDREN “legitimate” (out of-wedlock) 036 bith 048) 3488 arly AFDC use 0.24 Mothers poor in year prior to bint oa) 2083, Mosher smoked during pregnancy 0.30 046) 3333, Low binh weight 0.05 Below 5.5 pounds. Excludes (024) LBW-premature babies whose 3.325 weight was appropriate for gest tional age. ver in foster or relative care? 0.05 Ever lived in foster care or with (023) nonpareexatreatives 3415 ‘cum ovrcoMES, ALL CHILDREN HOME score (percentiles) 462. Home Observation for Measure- (255) ment ofthe Environment (hort 6711 form. Test year and age of child entered 28 controls ‘Motor and social development SLL Children aged 0-4, Test year and index (percentiles) (26.1) age of child entered as controls 406 PVT (standardized score) 859 Peabody Pitre Vocabulary Test. (20.7) Receptive vocabulary for standard 4707 American English. Test year and age of child entered as controls. (lable continues) KORENMAN AND WINS: ry ‘TALE 7.3 (continued a ‘Mean (sD) Descriptions of Variables [0bs) tand Sample 875. Sample resticiel 0 children age 6 Gr yore alr ee (16.9) years and over, ston) asp vie probes inex Tir chien aged 412. Maem (Gana 0) C132) prs ef avr pls Tet SG) Saran ag of ch ened enclose nds mow lens re acsin nd Maray 1% panda Seuss cp recs nea ample aes tcc anon yah sansmising ales for AFQT sore, pres SES scr ge, and 1990 education. Sv coment rest cre) I valle fra given dul in tore han one yea (1986, 1968, oe ovo then te oucone se average (5 years) of Be assent oF that chi ‘as of 1990, we have restricted the samples to respondents who have valid information oa Henmstein and Murray's (standardized) schooling attainment ‘Variable. This restriction results in the loss of about 1 percent of the sample In 1980 the ASVAB (Armed Services Vocatimial Aptitode Bastery) was ‘administered (0 neatly the entre sample s0 that the Deparment of Defense ould renorm the tests based on a national population. The AFQT score Hermstein and Mortay use as their measure of 1Q is a weighted average of four of the ten components of the ASVAB. They provide arguments and idence that their measure Js one of the best available for general intel~ Tigence, We leave discussion and evaluation of this claim to future work in ‘which te intend t0 take up issues of endogenous determination of ARQT Srores (gee also Neal and Johnson 1995; Rogers and Spriggs 1995; Hunt 1995; Winship and Koreaman 1997 and forthcoming) : "The components of AFQT with ther factor loadings are (Herrnstein and ‘Murray. 1994, p. 583): Word Knowledge (87), Paragraph Comprehension (8D), Arithmetic Reasoning (87), and Mathematical Knowledge (82) Hermstein and Murray do not discuss whether, net of their measure, other ‘of the ASVAB might affect the different adult outcomes. For ‘Heckman (1995) notes that the numerical operations component is ‘of laborsmarket outcomes. Although the construction and ‘one that we do camp sone pedir ; ‘mete of be APQT sor an importa se, 06, od Mary's mess of petal sil and conomie satss tem age openers ers an oes erat, occu 4 ci ser mn he hows (be Hest vised Dune INTELLIGENCE, BACKGROUND, AND SCHOOLING ws Index score among the two parents or adults), and the natural log of income of the parental family (the average of available years, 1978 and 1979, for ‘youths who report iacome of the parental household). The components of ‘SES are each standardized to have variance one. A simple average of the available standardized measures is taken in order to create the SES index. The index is standardized to have variance one. Omitted Variable Bias in the Effects of AFQT Score: Family Fixed-Effect Estimates Many aspects of the fumily socioeconomic environment could conceivably bbe included in Hermstein and Murray's models because they may influence AFQT score at age fifteen to twenty-three and adult outcomes, and may not bbe captured adequately by the parental SES index. For example, growing up in a single-parent family has beea faked 10 a variety of social and economic disadvantages (e.g, Murray 1984; Garfinkel and McLanahan 1986). The ef- fects of 1Q score estimated by Hermstein and Murray may be exaggerated (iased upward) by omitted variables. One approach to this problem is to atlempt to measure and include in the models additional family background variables, We pursue this strategy below. ur initial approach isto carry out analyses of siblings. We compare the fect on various outcomes of differences in AFQT scores between siblings. In the case of continuous outcomes (dependent variables), fixed-effet analy- sis amounts to entering # dummy variable for each family of origin. For dichotomous (binary) outcomes, we estimate fixed-effect logit models for the ‘oldest pair of siblings from each baseline household, The estimation of fixed-effects logit models necessarily involves a substantial reduction in sample sizes because only sibling pairs shat have different values for an ‘outcome (e.g,, one graduated from high school and the other did not) con- tribute tothe likelihood function (Chamberlain 1980). ‘The fixed-effect analyses correct for bias due to both measurement error jn the parental SES index and omitted family characteristics (Le, charac- teristics that are common to siblings), This is @ broad notion of family back- ground that includes, for example, characteristics of the neighborhood and the surrounding geographic area (Griliches 1979), The advantage of this “ Hemnsein and Mary (196, 123) ane ba adlag vat of tional family eckground vaible may Be problema fee vaviabes are “intervening” variables (ie, ) ember of bs ate 1.096 1. Unemployed 14-044 -000 052-002 047 ‘mon in 189, men (OT) a) a9 co (bie conte) 148 “TABLE 72 (continued) ‘OES or Lgl Confit (SEs) INTELLIGENCE, BACKGROUND, AND SCHOOLING 19 TABLE 7.2 (continued) (OLS or Logit Coficient (SEs) at somo Slings SEO Slings FB Rae a aE Gag ws ao arses ara aol china: creak Tes cons 01 0m OL ~aNs ~035 023 (am wn md 1) I) ow ith weight 038-008 02-3 ons Number of os ams wo 6 oi Gy “ W035" 008 00-023 012 008 © 070 — 36, Marit by age 30-004 —007 on 010 om 3) oH) cit es a (as) (0) ap a aD Naber of obs ee ee ee! WGP om on 007-024 039-003 2 (mn) tl) DD Sein feree 042-022 =o ois a umber os sna ot 136 itive cae? un a “co on 00-026 -06 O16 008 da ~ ae TS) te Go “@ ee petal mariage gs, 005-0 09 -Oa8 047 OL a oa » (0 (es) oy) GD GO com ovrcomes, aL COREY umber of bs ie 1046 198 ome score 44 s2 39 Ee Misite class valves © 075023 om 020 067 resenties) a eo an ao ay Index op (a) (09 3 S28 eane sg eB ot osr 005,20 Naber of ob i a ec toy (oy CO) COB) (06) CB). CIS) 13 ‘on Number o bs 1692 2682 ‘20 Maweaeitde 2219 wo on ver interviewed in -091 ~006 Host 016 01 ‘pores 05 05) aya 19 James aH on 8) 9 PPVT, 6 year ods cs 665822 hee 6) 5) 6) (0.5) ate 403) 1s) 03) eas 04) 3) 4) (0.3) tet J nude APQT sure and SES SC, age (Sor) an Wee appre oe ere clepnicly @ dummy varie), ya, and c's age athe dine cae crete for ntindpendenc of Observations ang yOubs fom he ee Toho welgound contl ice nly engine a a 1 Gdn wl: aoe town ay cher ge en ens a orale Lo ay ce mapas repr nse ewspe rel ad» ee emmy way ge mtr bth fend (is ong en mie ead ee espe Brn oie he Unie Sa Ne re ne aed on eng a je) of enones ont Va et aE row Jn fe cst coy ie len ww econ ow repent: of evn 04 0) se. INTELLIGENCE, BACKGROUND, AND SCHOOLING 157 \we repeat the results presented in table 7.2. inthe third and fourth columns we present coefficients and standard errors for the AFQT and SES variables from models that include detsiled controls for family socioeconomic back- ‘ground, Finally, in the last two columns of the table we present twa “com- posite” estimates of the effects of family socioeconomic background (both in absolute values). The first is a standardized composite of the SES effect aad the effects of the various family background characteristics described in the previous paragraph. The second composite adds to the first the effect of racialethnic identification. Since AFQT score is controlled, the effects of racelethnicity may reflect, atleast in pat, additional effects of family socio- ‘economic hackground (also see Fischer et al. 1996). ‘The composite effects we have constructed may be unfamiliar to many readers, This procedure allows us to extend Hermstein and Murray's ‘methodology for comparing effects of AFOT and SES to compare the effects fof AFQT to a single, yet more compretensive measure of family socio- ‘economic background. The composite effects ae derived as follows. We first estimate a model for each outcome using the different controls for family socioeconomic background described above. For example, in a linear regres- sion with dependent variable Y, family background components X, and AFQT we would have: YY = by + Xb, + AFOTD: + w where by is a vector of coefficients representing the effects of different fam- lly background measures.7 Using our estimate of b, we then calculate the Predicted (inear) component of Y, F, due te family background factors: Pe xb; @ ‘Using the estimated form of the equation for ¥ (equation 1) we can rewrite (as Yuh +h+ AROTb +é @ We then standardize F to have standard deviation equal to one in the popula- tion, producing a new variable P, We can then rewrite (3) as: Y = by + fay + AFQTS: + @ ® 1 The model so incl ems for ape and ender contol (ot shows). The ary X io condes Hest and Marys penal SES ines 158 sation one in the population. Or sine AFT is sled 0 ave standard devin om 9 Er sng A i amir ee wee ce rie i mo Pc af fn oe Ss ‘wishes t0 isola ee leo oa ym slo 2 a iodels. Given Hermstein ‘and Murray’s positior ee across all models Phying eonsiuct of inalligence, their tents 7 eres! yropriate (although an area ‘of future research is re ne Pot age ty es ee medommacrane ee rccpacn 9 speaki family arrangement and the other family socio- Genel spestng es ae included in he model the es zona bck anged and te fet of SES als mal Come A game ea TON FC Lr lL abe oe EB vals pial fr oxen hat of SS er cae ‘column 2 to column 5 or 6). For ‘example, the effec oS 4 SCS (como es 0 pret ge han he et of sone ct 219 orate, et of abot fee, wo faa «30, middle class values index, and ever if jai ei mae tomes Sen ewe el SS fs sai ae aa index has no diter™itive size of the AFOT and composite. FB . ota rf ‘outcome is to schooling attainment, ao ie ratte comport i ot es tetany cgi elt yi te Tecnico roy oy ceed 1 i ma 995) examine ee “ne sn yes fe 8 C7. ters APT INTELLIGENCE, BACKGROUND, AND SCHOOLING 159 cof AFQT in predicting education and edveation-related outcomes further un- derscores the need to model carefully the joint determination of education, ‘AFOT score, and the various adult outcomes. Other than schooling out- ‘comes, the magnitude of the composite FB effect tends to be in the neigh- borhood of the AFQT effect, and the point estimate of the composite FB effect is larger than the AFQT effect for seven outcomes (oUt of the labor force, marriage, illegitimate bith, early AFDC use, foster care, HOME score, and motor and social development score) when race/ethnicity is ex- Cluded from the composite effect. The composite FB effect is larger than the [AFOT effect for thre other outcomes (divorce, low birth weight, and PPVT score forall children) when the effects of race/ethnicity are included in the composite. Biases in the Effects of Parents? SES: Residual Family Background Effects ‘The results presented in table 7.4 suggest that the combined family socio- economic background effect was considerably larger than the effect of the index of parents" SES alone. Its also possible to derive an omnibus estimate ‘of the family background effect implied by the fixed-effect models. This effect captures the effects ofall characteristics siblings have in corimon that fare not included in the model (such a AFQT, age, and gender). Thus, for example, it includes not only the effect of having grows up in the same household, but also the effect of having grown up in the same neighborhood or state, This effect potentially includes similarities in such things as person- ality, motivation, and effort. With continuous dependent variables, we e: rate directly the effect ofthe latent family background variable by conduct- ing a one-way ANOVA (analysis of variance) analysis (by household) of the residual from the fixed-effect model. Here the residual is constructed 10 i clude all variance in the dependent variable not due to the observed indepen- dent Variables. That is, it includes both the individval and family-specitic ‘components ofthe dependent variable, once we have removed the effects of AFQT and other observed variables that may differ among siblings. If we assume thatthe latent variable has variance one, then is coefficient is equal to the standard deviation ofthe household effect. These results are shown in table 7.5. With discrete outcomes, the same methodology is not available. Instead, wwe estimate a bivariate probit model. This model isis not as powerful since ‘is a random effects model, and so we must assume that any unobserved family component is uncorrelated with observed variables such as AFT. However, we noted in our discussion of the fixed-effect models that SES sppeared to be an adequate control for family background for the purpose of 10 x a ym Analyses of TABLE 7 ay mil Bako nt AFOT Se fo Aly sr on aces : Estimated Effect (SE) eae O15, Fiaed Effects ror ses aor ‘Family no 7296 Aas? «S558 (12482 ame ot ae eee (622) (S77) (390) (543) Sos to S317 18D ise) wo (ss I eee ng, 1990 (z-score) 059 O18 04s z a =e (02) (92) (02), (01) SoS as Track, Latina, other race, and, where appropriate, Trae: Ober online: 2ACE, ‘Annual earnings, 1989. XR er Background. Fay pn fhe aly bakin simaing the effets of AFQT. (The exceptions 10 this finding were cue ea ‘pildren of NLSY respondents. However, we do not eae Mate) AS inthe finedeffect model if we assume then is effet is the square root of the is of the bivariate probit analyses are f= comes forthe hil outcomes fnvnt viable has variance one, ering coat, The es in ble 7.5. : : : et 7.5 sows he ing tet offs Riaaen is con SES (64.487) or ether AFOT sa considerably Inger than that of 4 ‘Art Si er at ere are crnings (65,180) is far lage than that of SES alone ($1,169) some Sat age than the effect of AFT score (OLS: $5.58, o> sce Pally te implied effect of fay background on eueation (50) var hrgr than the OLS estimate of SES's effect (18), and is thnbe APT Hef eat (49, te “pots the resls from the bivasiate probit analyses. Res ble 7.5 po som in ble 15 a in alt al ess he a Sm yale an the combine effect of SES andthe Intent sf cg lf AFT, enone, Te Th the probability of teciving a BA degree, wher oe ete ress in ale 7S suggest here vey age foray eau! ground component that is erogonal © the pret fal effects on many outcomes. ne es 1 INTELLIGENCE, BACKGROUND, AND SCHOOLING fa TABLE 758 Standardized Effects of Family Background and AFQT Gro Bivariate Probit ‘Analyses of Siblings Estimated ects (SEs) “Absolute Value of —_ Leent Tova waror SES FB! FB In povery in 1989 =037 018 0650.68 (0) (08 (4) (01 HE dropout ~ae 029-0708. (op (04) (03) 1) BA degree ose = 028075 a79 (O03) (03) ON) Out of labor force 1+ mos.in 1989 021-008-054 Oss men) CO (03) (0H) .004) Unemployed 1+ mos. in 1989(men) ~020 0.11 0.76 0.75 (03) (03) (03)——(.004) Married by age 30 010 © 003037037 (0) OD) (OH) Divorced, fist 5 years of mariage 0), O.17 021027 (on (05) (26) 004) ver interviewed i al, men -019 013 096 096 (60 = OD 008 Middle class valves index oa a2 OSs O36 (0 03) (05) OH) High-1Q occupation 0 = ons SO Oy (060) (004) "Noes Ore cons jsut: 2AGE, block Latin. ahr ee and, where appropri gea- er. See ble 7 for varible nd ape dein FB: faily backround he net fect the square re of te cross-eiton cole for slags. ‘ye wil eet ite agar oo ofthe sum of he SES eet qiared pls te Isnt effect squat 7.5b. These estimates attribute to family background all common variance ‘among siblings inthe outcome variables that is independent of the effect of APQT score, gender, and age. For example, the total family background effects include genetic traits that are common to siblings and orthogonal to AFOT score. Similarly, if siblings have grows up in the same places, any effects of location on outcomes will be included in our estimates of the effect of family background. Nonetheless, our estimates do not simply reaf- ‘frm Hermstein and Murray's acknowledgment thatthe explanatory power of 18 komen their models is low. se rioen im has 2 very large eet on etependent of measured incelgece ae or sling moe tata et he fail aoe re ‘chances of success in adulthood, ‘The Role of Education erstein and Murray do not present estimates of the effects of schoo erin a a mm oe oe it pet an Ao hy examine in ely ied way the fs at tela Ce ra nO rae Sl ee eel eo ree er to chy ea fee na we ar fo ee ra law ornSt a ma Tee ke Tae ot eae TMret of intligence is indirect, through edcation. This objection oins © Cnc gern neg, eh eco ae ston et aoa ee te 3 "an outcome —that is, the effects of AFQT ‘and SES net of the effect 2 cesar the "Ya" eet of hee two vals (Dek est riicear and (@) te etn Between chong sens yay apd he org sive modeling my be eos of etuxinn ee cr ey lie ine tis wet te ee cope nd sae a COP ha wesc he a ed "tc fh tok”. a {ind uo Syeda by ali hosig to ave win oe. ate eset eatin sma hbo ae ae mh cn bn ef of cson nd AFT orn cl. we secs cg an Mary event apne od preison ont vended ren a gossty interacting eaionhip between inelige nce er eee on enn vet de a aes cooing cm Sie Shira INTELLIGENCE, BACKGROUND, AND SCHOOLING 163 Furthermore, if one wants fo contrast total effects and thus account for the indirect effects of ARQT through education, one should also account forthe possible indirect effects of SES through AFQT (and education) on diferent outcomes. ‘A cftical question i, therefore, whether effects of AFQT are direct or primarily indiect through education. tn the later situation, itis because individoats who have higher AFQT scores tend 1 get more education, and education directly affects an outcome, that outcomes differ by AFQT Score. In this cas, the relation Between AFQT and an outcome might be changed by policies chat alter the celationship between AFQT and sehool- ing. In fat, in The Bell Curve Heerstein and Murray recognize that the sslationship between schooling and 1Q is malleable when they argue that higher education as become increasingly selective with respect 1Q, and agsin in chaper 18 when they discuss the “dumbing dow education. Schooling atainment can potentially be manipulated by public policy. If elucation has a substantial effect on various outcomes, then Herastein and Muray’s pessimism about society's ability to change individual outcomes ray be unwaranted, That is, even if addional education has no effect on 1Q. an increase in an individual's education level may enhance his oF Ber hances of success (Hauser and Carter 1995; Jencks etal. 1979) Table 72 reports estimates when edvaton is incladed 3s an independent ‘arable. Ta eleven of twenty-three cases the inclusion of education reduces the elfet of AFQT by more than 25 percent. In many cases the standardized cifectof education is Jager then that of AFQT. In the OLS and standard logit analyses, education has a larger effect than AFQT for family income, ~2naat-carngs, highsIQ. occupations, the middle-class values index, sinter the mother smoked during pregnancy, HOME index, and child's ‘motor and social development index. Parallel changes are found inthe fixed- ertect model." ‘The inclusion of education controls also substantially changes the effect of zarontal SES, This result is hardly surprising, since previous research hs tepeaely shown that much ofthe effect of parental SES on status atain- ‘ment works indirectly through education. In six of twenty-three cases the toot of SES is reduced by more than 25 percent. Is notable that including ‘location has litle impact on the estimates of the effets of SES on the stcomss associated withthe children of NLSY respondents. ‘ne might argue that itis appropriate to exclude edveation controls be- sau Hmstein and Murray intend 0 compare the total effects of AFQT Aprenix C presents analyses offal income and anual earsngs where we havea say tebe estimates fo measurement eet AFOT scores and eduction In model tis some, the effect of AFQT score fall rst uy am 0540 54627 when we core for meatrenent ero Both effets ise SH dls fanaa ening. 164 KORENMAN AND WINSIIID score and parer in tion is problematic since educa J understate the effect of parental SES (because epee et of AFOT™ ae that education has a minimal effect al SES, Even in his nse, howe, he isin of aya AFT soe Ion ese wo cl SES aes soc seatet St emi and Mary ie one pend 3 of The Bll Care Hermcin snd Muay AF ti ote poole fleas of cation on AFT sng ar measures of 1Q as a contol variable, They find that an increase i enn 10 i ig sia oo 2 eo. TN Tare ria “rec i a ch ai of en fa ae FS a te inn Nao ero et inane cena 1) " Korean 95 malig effort is ned 0 srt ut the posible muta Frere on each te, and oun Fr he foc otesaion tA ete fle ees oan gh bef on AFT. We inde et ays mols 0 red papers (Winship and ad "Winship and Korean frcoming). Nonetiless Ce tepared here we fam temas ofthe det fet fom te ans Ptr een, tn subtly smelt SA ee ART repo nie Bell Cure Furermore,edcaton than ets on many otcomes, coming for AFQT 82 Conclusion rl Carve an chapter 14 sw dmenstate sre pups of section I of The Bel Cs mat assoc of AGT score in determining vane of outcomes. th mes suai esl te flloving ay Ta wie vot coal fer uly of don, Petnaly ing 0 WoUM fer i LRG Seaton to ce be acs of APOE INTELLIGENCE, BACKGROUND, AND SCHOOLING 16s child of the nest generation could be given 8 choice between being disad- ‘antaged in socioeconomic satus or disadvantaged in inligence, there is ro question about the right choice” (p, 138). Hermstein and Murray are Confident that innate inteligence isthe principal determinant of economic and social success, In their 1979 book, Who Gets Ahead? Christopher Jencks et al. using a large numberof dataset, analyze the importance of inteligence, education, family background, and noneogntve abilities in determining various eco: nomic outcomes. Jencks etal. concive that al four Sts of factors are im- Ponant, that n0 single factor dominates the other, and that their relative Importance differs aeross samples and outcomes. ‘Which conclusion is righ” Are Hermsten and Munay conect in asserting that inteligece is the dominant for in determining Social and economic sucess? Oras Jencks etal. aster, is intelligence just one of several impor tant factors including education and family background? “Although we ‘mostly confirm with sibling analyses Hermstein and Morry's finding that the effects of AFQT are substantial and robus, on balance our results are closer to Jencks et a's Although we have not replicated the Jencks etal. analyses, we do find evidence thatthe paral effects of Family beckground and schooling are 3 ange x, and in many cases larger than those of AFQT in predicting a varity of outcomes. ‘The large paral effects we find for ecucation (ast of AFQT score and family background) are particulaly im- Porant given Herstcin and Murray's pessimism abou the potential of so- Cll polcies to change outcomes. Tn addition, in models that exclude schooi- ing controls, the effers of family background are as large 88 or lager then the effects of AFQT sere. {In reaching these concisions we have ignored the potently serious problem of the endogenous determination of AFQT score. For example, family socioeconomic hackground and schooling quality are imporantdeter- minantsof AFQTT score tages fiftesn to twenty-three, then he estimates of [AFT score and parental SES that we have presented may exaggerate the importance of APOT scar rive to family background in influencing so- cioweonomic outcomes, Ths endogenity of AFQT scores is subjet of ongoing investigation ac, og, Neel and Joison 1995: Rogers and Spriggs 1995) "Aa exeepin 6 our Sing of rob fects of AFOY ae analyses ofthe developments coco of youn sion oF NLSY fone sme mem Lift of ae’ APT wae Sc smal nd ot signin whe he comparizon in made between he children of mater who ‘xe sie Ge,fit eousn). This ing al stds in cons the Rigs of Come and ‘Thomas (1995), who opr sbnatt ets of tars ARQ score ai aig onl or Iie econ nd pane Rene, Tey met ent aa of sng ‘ces however) 166 KORENMAN AND WINSHIP [APPENDIX A mand Execs from Models of Socioeconomic Sus, by Race (O15 Cofiint (SE) Pall Sompee Sibling XSEC™ ‘Sing FE ror ESD POT Ses ED SAPD Fay INCOME (8, 19 ites 66274146 7003 s208 6196 600) 8) as) 67) (36) 3S AMO 471 40M) 38K) S08 S910 58S 5) 6) G2) SH GI) RD). TH. GH ber fab, “68 wae sn ele eon 530. 71083586 8s ey om) du doa abe 4g 42015353 3.6389 aos 736 (Sa) 0H) Ss) (98) ASTD 399 Nae of 1931 fo esis suas are sero 3.038 as) GH (365) 63 soi. 15M. 389 G2 3384 5.90028 Gos HG) ICI) Gos 2m) Naber. Ins so ‘01 “ANNUAL EARNINGS, 1989 YEAR-ROUND WORKERS ies sos sm soo 110 ear am G75) 1) a) ag 290s BRS 3663 Ase 583647 3.061 38) (USTD Nutro obs 2 m8 6 Bucks sus 1302 4o9 1st 3086 a) GD es) @n sy ae 892909332057 2213 Las 009) SN COM), (96H) mbar i 1 ‘ot Latinos smi a sss 96 340 (os) 7D) ais) 67) (1570) eM LMR Ast 4535963) 7) GHAI) S— aR CATH cio) Nanterofots 76 = ae _reats OF SCHOOLING COMPLETED 190 SCORE) Wes om Oe ass 030 040 con cob (00) o Naber obs sisi nas 24 (ube cone) = S INTELLIGENCE, BACKGROUND, AND SCHOOLING 167 INDIX A (continued) (OLS Coficiens (SE) Fal Sample ‘Siege ASE wor Ses uED ROT SES ike oss 07 0570 (0) (oy (on, Nurberfobs 2908, mais Latinas on om os 002 on aD (05) (0) amber ao 1603 m Note Se bl 7. fo deacriton of depeent vars aad vamp XSEC: crogeazcoe: FE: Redeflecs YR: year-round ‘Maovdels conan conls forage (score) and pendec. Stand ees ae comet for non-independece of eservaions among Jul from the same baie ouch “sioing netfee modes are sibling differences existed by incloing in the repression mols army variable foreach aly of origin. rein te see (camuuos 2990) () » ‘ng we snow 28 ew oee~ ou 0 so cw san, ont, 0N po Amar snodedenon sourmdeyg apo opeusy apy ouoptsas gin uso oee~ song das yoo sao aD LNBON NEY ATURE oyeueg so 290 une a aove sase wave (oops) 2g + a+ Sas (Goren) a + 83S wo oro io dog Ws SH $1094 (38) su2pafo09 HBO 19 $10. “janie ponoidyoea Ane Fa wourdojnog PINE) POW HR NHIOUGSDDIIOG TO SPOR WH ADAH POUT enue ponostyone Aang poRMag wae WAND oe (comune 290) me ary snadedenon surety aviom opus py souapysar win $90 ovo. vag gov. sase abave (Gores sgn 20ra e pourgEIOD (nye sqp) 2 pouigwe nod oad¥ Suny “Aug 15) sHIVOD WBC 10 $0 WoL 905% oes Saws coos Lust. G00) 6) wD 900 160 ss0~ as) 590 30 sang, «@) 0g ao og oUPeRes sag By) nes se (eas) snanalfoog 1807 0 $10 (Goma) axa swe Stee eee ET ow) @ be wD va Buidé tnd (95) HOD HOT 10 S10 ewe eve ‘Os0n (comes 290) e290 cone ve 2 = 6008 80 S18 ‘one (Gores sae) ga + SS $90 30 UMN. 09-804 womens otAParebs snow 28y) ov so tg we smpou a8 we) s0 0030390 sont O- on 9) a) soo~ oro- ort~ at 3D MET FiOS Od—«COGIY wg or Soma sito 4mq “hurg tod (6S) saa0 WB} 50 Eee “ponuioo) @ XIGNSAY 390 Jo QUIN, 09 uBio, weg otras seqow 28y) a we sacgous oy sou 30 380, san wD «D 6D 90- or- $0 ont wo wo © ve I- ee 300N Go 0 wo- 0 pms Arar Go oD Toe co suodedenaN Go 60 go- £0 sou, GoD 60 00- apo opeu HHP Ce zo- -90- souoptsa wean 0 das oD 00 v0 2901 Tad 9ihdd tAdd _OSOM__3NOH (eas) swenefoog HOT 2 STO. ‘pamanos) @ GNA 16 kon ‘APPENDIX C Bifects of Reliability Comeetions on Fixed-etect Coeftcent Estimates Puaed-Efece Reliability Coeficiens (SEs) _—_—Ratios* wer uD ROT kD 1, Family income, 1989 5.558 00 (975) 3610430800 (653) oss, sar oes TT 1258) 2, Bamings, 1989 (erround 1.00 workers) 2M 106.00 (56) 086 26 ad 600) Tans co Fraga tay ass we tbs il aos ol ame The vals eit a askin we band om Mara pen commana ck ae Mee (938 te Sis i of cca rbd ohne ad Kreger (98), Teves 2 eae etn tutes emai ig tt inf eatin of tos 208 eatin, Se ate cope oxeoes References ent Oe, an Alan B. Kee 1994 “Etat of he zoom Retm 1 Se eal New Sump of vis” Americar Eamon Revie 8 05 eli a Pt and Fank Mot 189, LS Chl Handbook 1988 Counts, OF: er, vr Hanan Ree Retarh-The Ohio Ste Unies a tnd ae. Mave 186 hahnog and Daine:& tan Th ill 0 Lawrie Eby Asc et tape td Le A.B 198, Mesure ros Now "Dau et Lo Copan Pa cai aa Resous Rech 1994, MLS Handbook 194, Colas io fr nes ay, 880. ani of Covance with Quiave Da” Rew of emt Ses: 25.2% INTELLIGENCE, BACKGROUND, AND SCHOOLING an Curie, Janet, and Duncan Thomas. 1995. “Race, Children's Cognitive Achievement ‘and Tite Bell Carve." NBER, Working Paper no. 3240, August. Fischer, Claude S.. Michael Hout, Marin Sanchez Jankowski, Samuel R. Lucas, Ann ‘Swidler, and Kim Voss. 1996, Inequality by Design: Cracking the Bell Curve Myth Princeton, NI: Princeton Univesity Press Garfaket, vin, and Sara S. MeLanahan. 1986. Single Mothers and Their Children: 'A New American Dilenima. Washington DC: Ueban Insitute Press, Goldberger, Achur S, 1991. A Course in Econometrics. Cambridge, MA: Harvard Univesity Press. Goldberger, Arts S,, and Charles Manski 195, Review Arile: The Bell Curve by eatin and Muray, Jourmal of Econamic Literate 33 (une): 162-776. Griliches, Zvi. 1979, “Sibling Esimates: The Beginnings of a Survey." Jounal of Political Economy 81 (gat 1): 37-864 “Mauser, Robert Ma and Wendy ¥. Cater 1995. “The Bell Curve as a Study of Social ‘Stratication.” Unpublished piper, Sociology Department, University of Wiscon- sin, Madison, Heckman, James J 1995. "A review of The Bell Curve: Iuelligence and Class Struc: ‘ure tv American Life by Richard Hermstein and Charles Murray.” REASON (March), Reprised in the Harris School Report, University of Chicago (Summer 1995); 49-56. Hermstein, Richard J, and Chatles Muay: 1994 The Bell Curve: Inligence and ‘Cass Sreture in American Life. New York: Fre Pres ‘unt, Ear. 1995. "The Role of Imelignce in Modern Society" American Scientist uly/August): 355-368. Jencks, Christophe, et al. 1979, Who Getr Ahead” ie Determinants of Economic ‘Success én America. New York: Basic Books, orenman, Sanders D. Jane E. Miler, and John E.Sjastad, 1995, “Long-Term Poy- ety and Child Development in the United Stats: Evidence from the NLSY." Chil. diren and Youth Services Review 17,30. V2: 127-155. Miller, fae E., and Sanders D. Korenman. 1994, “Poverty and Children’s Nustiona ‘Status in the United Sites." American Joumal of Epidemiology 140, no. 3: 233~ 243, Moray, Chavles. 1984. Loving Ground: American Social Policy 1950-1960. New ‘York: Basie Books. Neal, Derek A., and William R. Johnson 1995. “The Role of Pre-Market Factors in Black-White Wage Differences.” NBER Working Paper no, $124, May. Rodgers, Willa M., and Wiliam E. Spriggs. 1995, "What Does AFQT Really Mea- sure: Race, Wages, Schooling andthe AFQT Scar.” Mimeo, College of Wiliam and Mary, Mar. Solon, Gary S. 1982. “Intergeneratonal Income Mobility inthe United States." Amer. lean Economie Review $2, 0. % 793-408. ‘STATA Comporation. 1993. STATA Reference Manual: Release 3.1. 6th ed. College Station, TX. Winship, Christopher, and Sanders D. Korenman. 1997, “Dees Staying in Schoo! Make You Smarter? The Efe of Eduction 00 10 in The Bell Curve” In Itel- ligence. Genes and Success: Scentins Respond to "The Bell Curve," ed. Bere 18 KORENMAN AND WINSHIP Devlin, Stephen Fenberg, Daniel Resnick, and Kathryn Roeder. New York Springer Vriag, pp. 215-234, ‘Winship, Christopher, and Sanders D, Koreaman, Fortheoming. “Economic Success and the Evolution of Schooling and Meatal Ability.” In When Schools Make # Difereace, 08. Susan E. Mayer and Paul E. Peterson. Washington, DC: Brookings Institution. Zimmerman, David. 1992. “Regression Toward Mediocrity in Economic Stature” ‘American Economic Review 82,90. 5: 49-429,

You might also like