Problems with Instrumental Variables Estimation When the Correlation Between the

Instruments and the Endogeneous Explanatory Variable is Weak

Author(s): John Bound, David A. Jaeger and Regina M. Baker
Source: Journal of the American Statistical Association, Vol. 90, No. 430 (Jun., 1995), pp. 443-
450
Published by: American Statistical Association
Stable URL: http://www.jstor.org/stable/2291055 .
Accessed: 12/02/2014 13:24

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 [email protected].

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

http://www.jstor.org

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
 ProblemsWithInstrumental Variables Estimation
When the CorrelationBetween the Instruments
and
the Endogenous ExplanatoryVariable Is Weak
and Regina M. BAKER*
John BOUND,David A. JAEGER,

We drawattentionto twoproblems associatedwiththeuseofinstrumental variables(IV), theimportance ofwhichforempirical

workhasnotbeenfully appreciated.
First,theuseofinstruments thatexplainlittleofthevariation intheendogenous explanatory
variables in theIV estimates
canleadto largeinconsistencies evenifonlya weakrelationship existsbetween theinstruments and
theerrorinthestructural Second,infinite
equation. samples, IV estimatesarebiasedinthesamedirection as ordinary
leastsquares
(OLS) estimates.
Themagnitude ofthebiasofIV estimates approaches thatofOLS estimates as theR2 between theinstruments
andtheendogenous variable
explanatory approaches 0. To illustrate
theseproblems, wereexamine theresults
ofa recent paperby
AngristandKrueger, whousedlargesamplesfrom theU.S. Censusto estimate wageequations inwhichquarter ofbirth is usedas
aninstrumentforeducational Wefindevidence
attainment. that,despitehugesamplesizes,their IV estimates
maysuffer from finite-
samplebiasandmaybeinconsistentaswell.Thesefindingssuggestthatvalidinstruments maybemoredifficult tofindthanpreviously
imagined.Theyalsoindicatethattheuseoflargedatasetsdoesnotnecessarily insulate researchersfromquantitatively
important
biases.We suggest
finite-sample thatthepartialR2 andtheF statistic oftheidentifyinginstruments inthefirst-stage
estimation are
useful
indicators
ofthequalityoftheIV estimatesandshouldbe routinely reported.
KEY WORDS: Compulsory
attendance;
Finite-sample Weakinstrument.
bias;Inconsistency;

1. INTRODUCTION atoryvariable intoexogenous andendogenous components.

Theexogenous component isthenusedinestimation. More
Empirical researchers often wishtomakecausalinferencesspecifically, theIV estimator usesone ormoreinstrjuments
abouttheeffect of one variableon another.Doing so in topredict thevalueofthepotentially endogenous regressor.
nonexperimental settings isfrequently difficult,becausesome Thepredicted valuesarethenusedas a regressor intheorig-
oftheexplanatory variables areendogenous; thatis,theyare inalmodel.Undertheassumptions thattheinstruments are
influenced by someof thesameforcesthatinfluence the correlated withtheendogenous explanatory variable buthave
outcomeunderstudy.Forexample,economists examining no directassociation withtheoutcomeunderstudy, theIV
theeffect ofeducation onearnings havelongbeenconcerned estimates oftheeffect oftheendogenous variablearecon-
abouttheendogeneity ofeducation. Itseemsquiteplausible sistent. BowdenandTurkington ( 1984)andAngrist, Imbens,
thatthesameunobserved factors mightinfluence bothin- andRubin(forthcoming) haveprovided useful introductions
dividuals'educational attainment and theirearnings (.see, to IV estimation.
forexample, Griliches 1977)."Ability" is often citedas one Whensearching forplausible instruments fora potentially
factor possibly correlated withearnings (thosewithhigher endogenous explanatory variable, itis commonto findthat
ability earnmore)andeducation (thosewithhigher ability thecandidates areonlyweaklycorrelated withtheendoge-
obtainmoreeducation). nousvariablein question.It is wellrecognized thatusing
Whenexplanatory variables areendogenous, ordinary least suchvariables as instruments is likelyto produceestimates
squares(OLS) givesbiasedandinconsistent estimates ofthe withlargestandard errors.In thisarticlewe drawattention
causaleffect ofan explanatory variableon an outcome.A to twootherproblems associatedwiththeuse ofsuchin-
commonstrategy fordealingwiththisendogeneity is touse struments. First, iftheinstruments areonlyweakly correlated
instrumental variables(IV) estimation, usingas "instru- withtheendogenous explanatory variable, thenevena weak
ments"variables thought to haveno directassociation with correlation betweentheinstruments and theerrorin the
the outcome.The exogenousinstruments allow the re- originalequationcan lead to a largeinconsistency in IV
searcher topartition thevariance oftheendogenous explan- estimates. Second,in finite samples, IV estimates arebiased
inthesamedirection as OLS estimates, withthemagnitude
* JohnBoundis Associate Professor,Department ofEconomics, Uni- of thebias approaching thatof OLS as theR2 between the
versityofMichigan,AnnArbor, MI 48109,andFaculty ResearchFellow, instruments and theendogenous explanatory variableap-
NationalBureauofEconomic Research,Cambridge,MA 02138.DavidA. proaching 0. Thoughtheseresults areknown, theirgeneral
Jaegerisa Ph.D.candidate,Department ofEconomics, ofMich-
University
igan,AnnArbor, MI 48109.ReginaM. Bakeris a Ph.D. candidate, De- importance for empirical work has not been fully appreciated.
partment ofPolitical
Science,University AnnArbor,
of'Michigan, MI 48109. To illustrate theseissues,we reexamine theresultsofa
Computations forthisarticle
werefunded bythePopulation StudiesCenter provocative paperbyAngrist andKrueger ( 1991;henceforth
andtheComputational Sciences Information
Division, TechnologyDivision
at theUniversityofMichigan. The authorsthankChrisCavanagh, Gary denoted by AK-9 1). This paper used the large samplesavail-
Chamberlain,ZviGriliches,GuidoImbens, DavidLam,KevinM. Murphy, able in theU.S. Censusto estimate wageequationswhere
andseminar ofMichigan
attheUniversity
participants forhelpful
discussions.quarterof birthis used as an instrument foreducational
TheauthorsalsothankJoshuaAngrist,Eli Berman, Alan
CharlesBrown,
Krueger,Jeffrey
Grogger,KevinLang,Lung-Fei Lee,Marianne Page,two
anonymous GarySolonforhelpful
andespecially
referees, comments on
drafts
earlier andCynthiaGlovinsky in preparing
forassistance themanu- ? 1995American Association
Statistical
The workwas stimulated
script. in partbya questionaskedofthefirst JournaloftheAmerican Association
Statistical
authorbySandersKorenman. and Case Studies
June1995,Vol.90,No.430,Applications
443

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
444 Journal of the American Statistical Association, June 1995

attainment.Although quarter ofbirth is onlyweakly related to includecommon exogenousvariableswould complicate

to educational attainment-the R2 in theregression ofed- thenotationbut would not otherwisechangetheresults.
ucationalattainment on quarterof birthrangedbetween It is straightforward to showthat
.0001and.0002intheirsamples-theauthors obtained rea-
sonablestandard errorson theirestimates oftheeffect of plim FOoir = A X2 (4)
education on weekly earnings,duetothelargesamplesthey a-x
used. and
We present evidencesuggesting thata weakcorrelation
between quarter ofbirthandwages(independent oftheeffect
ofquarterofbirthon education)existsand is sufficiently
plim Oiv = A + (y2 (5)
largeto havequantitatively significant on AK-9I's
effects
estimates.
Wealsopresent resultsthatindicate thatthefinite-whereair is thecovariancebetweeni andj, a j is thevariance
samplebias maybe quantitatively significantforsomeof of i, and x representsthe (population) projectionof x
theestimates thatAK-91reported. Together theseresults onto z.
suggestthatthe"naturalexperiment" afforded bytheinter- Equation (4) indicatesthat ?x mustbe nonzeroand x
actionbetweencompulsory schoolattendance laws and mustbe uncorrelatedwithc for,%is to be a consistentesti-
quarterofbirthdoesnotgivemuchusableinformation re- matorforA. Similarly,Equation (5) makes clear that ?x
gardingthecausaleffect ofeducation on earnings. mustbe nonzero(i.e. thattheremustbe an associationbe-
Ourresults illustrate
thegeneral significanceoftheseissues. tween x and z) and thatz (and therefore x) mustbe uncor-
In particular,
theysuggest thatit maybe evenharderthan related with c forfivto be a consistent estimatorof A. That
manyhavethought to findlegitimate instruments forpo- is, fivis consistentonly ifthere is no directassociationbe-
tentially
endogenous variables.Although researchersmay tween z and y.
a priori,
believe, thatthevariation inaninstrument islargely 2.1 Inconsistency
unrelatedtotheprocessunderstudy, thisis notsufficient
to
implythatIV estimates willbe lessbiasedthanthoseesti- Equations(4) and (5) implythattheinconsistency ofIV
matesproduced usingOLS. Inaddition, theseresultsindicate relativeto OLS is
thatevenresearchers working withverylargedatasetsneed
to be moreconcerned aboutthefinite-sample propertiesof plim AV- ?xe/( e
IV estimators. plim &OS - Rx,z v

PROBLEMSUSINGAN INSTRUMENT whereR 2z is thepopulationR2 fromtheregressionofx on

2. POTENTIAL
THATIS WEAKLYCORRELATED
WITHTHE z. WhenEquations(1) and (2) includecommonexogenous
ENDOGENOUSEXPLANATORYVARIABLE variables,therelevantparameteris thepartialR2, thepop-
ofx on z once thecommon
ulationR2 fromtheregression
We areinterested in estimating
A,thecausaleffect
ofx exogenous variables have been partialledout of both x
on y in Equation(1) fromthefollowing
system (in which, and z.
forsimplicity,we assumethatall randomvariableshave WhenK = 1, Equation (6) can be rewritten
as
mean0): - plim iV Pz,e/Px,e (7)
y=Ox+c (1) plim $01S -f Px,z
x = ZIl + v (2) wherePi,j is thecorrelationbetweeni andj. It is clearfrom
wherex, c, andv areN X 1 vectors ofindependentrealiza- Equation(7) thata weakcorrelationbetweenthepotentially
tionsoftherandomvariables y is endogenousvariable,x, and the instrument,
x, c, and v,respectively, z1, will exac-
an N X 1 vector, Z is an N X K matrix in whichtherows erbateany problemsassociatedwitha correlationbetween
are independent realizations ofthevectorz, composedof the instrument and the error,c. If the correlationbetween
random variablesz, ... ., Zk, I isa K X 1vectorofconstants,theinstrument and the endogenousexplanatoryvariableis
and A is a scalarconstant. Notethat( 1) differsfromthe weak,theneven a smallcorrelationbetweentheinstrument
formulation familiartostatisticians inthatwedo notassume and the errorcan producea largerinconsistency in the IV
thatx is uncorrelated withtheerrortermc, and thereforeestimateofA thanin theOLS estimate.
Ox may notbe the conditional meanofygivenx. Weassume It is instructive to examinethe specialcase whereK = 1
E(v Iz) = 0. The IV estimator ofA is z
and is dichotomous,partitioning thesampleintotwo dis-
i
tinctsubpopulations.Let 1, 2,xl, and x2 representthe
fiv= (x'Pzx) x PZy, (3)
subpopulationmeans of y and x, and defineA\y -
. The IV estimatorofA can thenbe written
wherePz = Z(Z'Z)-1Z', theprojectionmatrixforZ. This and &A
k 2-

is numerically (1) and(2) bytwo- as

to estimating
equivalent
stageleastsquares,wherethefirst stage[equation(2)] is
estimatedbyOLS and thepredicted valuesfromthisesti- Fiv=l'\X-' ~~(8)
mation,x = ZIl, areused in placeof x in thesecond-stage
estimationofEquation( 1) byOLS. Expanding ( 1) and(2 ) whichis theWald estimatorof13
( Durbin 1954;Wald 1940).

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
Bound, Jaeger, and Baker: Problems WithInstrumentalVariables Estimation 445

It is easyto see thepossibleinconsistency

oftheWald it is possibleto deriveapproximationsto
Alternatively,
estimator. Ifwetakethedifferenceacrossthetwogroupsof thefinite-sample
bias oftheIV estimatorwithoutassuming
z thenEquation(1) becomes normalityusingpowerseriesapproximationmethods.Buse
-= (1992), buildingon earlierworkby Nagar (1959), derived
0Ax+lk , (9)
an expressionforthe approximatebias of fl1in samplesof
whereA-c - byAJuandtakingtheproba- size N,
1. Dividing
limitgives
bility
l'v (K- 2) (11)
plim?-c
plimfliv=+ - (10)
plim /x
whenEquations(1)
whereK is thenumberofinstruments;
i3 willbe inconsistent
ifplimA-c* 0. Themagnitude ofthe and (2) includecommonexogenousvariables,K is thenum-
inconsistencywilldependon theextent to whichPI and-Y2 ber of excludedinstruments. A littlerearranging givesthe
differforreasonsunrelatedtodifferences
betweenxl andx-2 approximatebias as
(i.e.,plimlzc* 0) andon themagnitude ofplimAM~. Even
verysmalldirecteffectsofz on y willmatterifplim&A7 is u2,v
a v
(K-2). (12)
small.
Although theseresults
arenotnew(see Angrist,Imbens, Note that
av/ a v is approximately equal to the asymptotic
and Rubin1993;Bartels1991,and Shea 1993 forrecent whenz explainslittleofthevariationofx. Define
bias ofFO1,
discussionsthatmirrorourown),theirimportance hasbeen Ir2, the concentration parameter (Basmann1963), as 2
largely ignored
byempirical researchers. Equation ( 12) impliesthatforK > 2, the
(ll'ZZII)/ aov.
2.2 Finite-Sample Bias bias oftheIV estimatorofA relativeto OLS is approximately
inverselyproportional to r2/K.This is thepopulationanalog
We nowassumethatE(e Iz) = 0, implying thatthein- totheF statistic on theinstruments, Z, in theOLS estimation
struments z arelegitimate and thatAvis a consistent esti- ofEquation (2). WhenEquations(1) and (2) includecom-
matorofA.In finite samples,however, oivis biasedin the mon exogenousvariables,therelevantstatisticis analogous
directionoftheexpectation oftheOLS estimator of 3. The to theF statisticon the excludedinstruments. It shouldbe
magnitude ofthisbiasdependson boththesamplesize(as notedthattheF statisticestimatedfromany particular(fi-
thesamplesizeincreases, thebiasis reduced)andthemul- nite) sample will tend to overestimater2/K forthe same
tiplecorrelationbetween theinstruments andtheendogenous reasonthatthe sampleR2 is an upward-biasedestimateof
explanatory variable(as R2 increases, thebias ofoivde- thepopulationR2. Even so, Equation ( 12) suggeststhatex-
creases).The finite-sample bias arisesbecausewe do not aminingthe F statisticon the excludedinstruments in the
knowthefirst-stage coefficients, II, but insteadmustuse first-stage regressionof IV is usefulin gaugingthe finite-
estimates. Intuitively, this impliesa certainamountof samplebias of IV relativeto OLS.
overfittingofthefirst-stageequation, leadingtoa biasinthe It is possibleto call into question the validityof using
direction ofOLS. Considerthespecialcase wherethetrue power series approximationmethodsto studythe finite-
valueofeachelement ofII is zero;thatis,theinstruments,samplepropertiesofIV. Atissueis thepotentialimportance
z, arecompletely unrelated to theendogenous explanatoryof the higher-order termsin the expansion.Under the as-
variable,x. Butin anyfinite sample,theestimates ofthe sumptionof normality,however,the exact distributionof
elements ofII willnotbe exactly equaltozero.Thedecom- theIV estimatorcan be derived.Drawingon Sawa's ( 1969)
positionofx intocomponents x and vi(wherev - x-x) is work,we derivethe relativebias of the IV estimatorusing
arbitraryinthiscase.Itthusseemsquitenatural thatinthis this assumptionin the Appendix.When K = 1 (i.e., the
case,theexpectation ofAvwouldequaltheexpectation of system is just identified), the expectation of ,3jv does not
0B1, The sampling variability of the two estimators will not exist.When K > 1, it again turnsout to be truethatmag-
be thesame,ofcourse.Interesting andintuitive discussions nitudeofthebias dependson theparameterr2/K.It is worth
ofthefinite-sample properties oftheIV estimator forthe notingthatunliketheresultsbased on powerseriesmethods,
specialcase ofexactidentification and one stochastic dis- the exact finite-sample resultsdo not show a knifeedge at
turbancehave been presentedby Nelson and Startz K = 2. For moderatelylargeK's and small valuesof r 2/K,
(1990a,b). thepowerseriesmethodsshowsomewhatlargerbiasesthan
Resultson themagnitude ofthefinite-sample biasofIV the resultsbased on the assumptionof normality.The two
estimates extend backtotheworkofR. L. Basmann.Under methodsshowrelativebiasesofsimilarmagnitudeformod-
theassumption ofjointnormality, Richardson (1968) and eratelylargeK's and values of r 2/K largerthan 2. Details
Sawa( 1969)independently derived expressions fortheexact can be foundin theAppendix.
finite-sample distributionoftheIV estimator in thecaseof In a recentpaper,Staigerand Stock( 1994) tooka different
onlyoneendogenous explanatory variablebutmultiple in- approachto analyzingthefinite-sample properties of IV es-
struments. In particular,Sawashowedthatthefinite-sampletimates.They developedan asymptoticdistribution theory
biasofIV is in thesamedirection as theOLS biasand,in thatdoes not relyon approximationor on the assumption
thelimitas R2, approaches 0, is ofthesamemagnitude as of normalityforIV estimateswithweak instruments. Their
theOLS bias. approach holds the first-stage II, in an -/
coefficiefits,

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
446 Journalofthe AmericanStatisticalAssociation,June 1995

neighborhood ofzeroas thesamplesizeincreases. In this pending onwheneachstaterequires children tostartschool.

context,theyshowed that F - 1 is an asymptotically unbiased Each ofthese patterns is consistent with the assertion that
estimator ofX2/K and that 1 /(1 + 9 /K) approximates the compulsory school attendance laws are responsible for the
magnitude ofthefinite-sample bias ofIV relative to OLS. association between quarter ofbirthandeducational attain-
Thisimplies that1/Fisanapproximate estimate ofthefinite-ment.
samplebiasofFivrelative to /3ols. GiventheevidencethatAngrist and Krueger presented,
All threeapproaches suggest thatthebias ofIV relative weareleftwithlittle doubtthatcompulsory attendance laws
to OLS is a function ofr2/K.Iftherelationship between areworking toinducea correlation between quarter ofbirth
theinstruments andtheendogenous explanatory variable is andeducational attainment. Wequestion, however, whether
weakenough, evenenormous samplesdo noteliminate the theselawsaretheonlyreasonforthiscorrelation, and,there-
possibility ofquantitatively important finite-sample biases. fore,whether quarterofbirthis a legitimate instrument in
Eachapproach suggeststhatthefirst-stage contains estimating
F statistic either wageoreducational attainment equations.
valuableinformation aboutthe magnitude of the finite- Therelationship between quarter ofbirthandageat school
samplebiasandthatF statistics closeto 1 shouldbe cause entrymustbe theonlyreasonfortheassociation between
forconcern. quarterofbirthand educational attainment forquarterof
birthtobe a legitimate instrument forageat schoolentry in
3. A REEXAMINATIONOF ANGRIST AND educational attainment equations.Similarly, therecannot
KRUEGER'SRESULTS be anydirect association between quarter ofbirth andwages
for quarterofbirthto be a legitimate instrument foredu-
3.1 Inconsistency
cationalattainment in wageequations.Becausetheassoci-
AK-91 usedquarter ofbirthas an instrument foreduca- ationbetween quarterofbirthand educationis veryweak,
tionalattainment inwageequations. In a subsequent article Equation(6) indicates thatevena smalldirectassociation
(Angrist andKrueger 1992,henceforth denoted byAK-92), between quarter ofbirth andwagesislikely tobadlybiasthe
theyusedquarter ofbirthas an instrument forageatschool estimated coefficient on education in wageequations.
entryin educational attainment equations.For quarterof Although we knowof no indisputable evidenceon the
birthtobe a legitimate instrument inthefirst case,itseffect directeffect ofquarter ofbirthon educationorearnings, it
on educational attainment mustbe theonlyreasonforits seemsquiteplausible thatsucheffects exist.First, quarter of
association withearnings. Similarly, forquarterofbirthto birthmayaffect a student's performance in school.Thereis
bea legitimate instrument inthesecondcase,theonlyreason someevidencethatquarterofbirthis relatedto schoolat-
foritsassociation witheducational attainment mustbe its tendance rates(Carroll1992),thelikelihood thata student
effecton ageat schoolentry. willbeassessedas having behavioral difficulties(Mortimore,
AK-91and AK-92documented significant associations Sammons, Stoll,Lewis,andEcob 1988),thelikelihood that
between quarter ofbirth andageatschoolentry, educational a student willbe referred formentalhealthservices (Tar-
attainment, and earnings forcohortsof menbornduring nowski, Anderson, Drabman,andKelly1990),andperfor-
the1930sand1940s.Individuals bornduring thefirst quarter manceinreading, writing, andarithmetic (Mortimore etal.
oftheyearstartschoollater,havelowereducational attain- 1988;Williams,Davies,Evans,and Ferguson1970 fora
ment,and earnlessthanthosebornduringtherestofthe summary ofearlier literature). Second,thereareidentifiable
year. in thephysical
differences andmentalhealthofindividuals
Angrist and Krueger arguedthatcompulsory schoolat- bornat different timesoftheyear.Thereis substantial evi-
tendancelawsaccountfortheseassociations. The typical dencethatindividuals bornearlyintheyeararemorelikely
law requires a student to startfirst gradein thefallofthe tosuffer fromschizophrenia (see,forexample, O'Callaghan
calendar yearinwhichheorsheturnsage6 andtocontinue etal. 1991,Shametal. 1992,andWatson, Kucala,Tilleskjor,
attending schooluntilheorsheturns16.Thusan individual and Jacobs1984). Thereis also evidenceof variation by
bornin theearlymonths oftheyearwillusuallyenterfirst quarterof birthin the incidenceof mentalretardation
gradewhenhe orsheis closeto age7 andwillreachage 16 (KnoblockandPasamanick 1958),autism(Gillberg1990),
inthemiddleoftenth grade.Anindividual borninthethird dyslexia(Livingston, Adam,and Bracha1993), multiple
orfourth quarter willtypically startschooleither justbefore sclerosis (Templeretal. 1991),andmanicdepression (Hare
orjustafter turning age6 andwillfinish tenthgradebefore 1975),as wellas somewhat mixedevidenceregarding dif-
reaching age 16. ferencesinIQ amongchildren bornatdifferent timesofthe
AK-91andAK-92presented several tabulations tosupport year(Whortonand Karnes1981). Third,thereare clear
theirclaimthatcompulsory attendance lawsarepartofthe regional patterns inbirth seasonality (Lam andMiron1991).
mechanism generating a relationship between quarter ofbirth Finally, thereis someevidence suggesting thatthoseinfam-
andeducational attainment. First, therelationship between ilieswithhighincomes(Kestenbaum1987) arelesslikely
quarterofbirthand educational attainment is weakerfor tobe bornin thewinter months. Although theevidencefor
morerecent cohortsthatwouldhavebeenlesslikely tohave someoftheseassociations isweaker thanforothers, andnot
beenconstrained bythelaw. Second,therelationship be- all oftheeffects arein thesamedirection, theexisting lit-
tweenquarterofbirthand educationis weakerforbetter- erature castsdoubton thenotionofnegligible directasso-
educated individuals. Third, therelationship between quarter ciationsbetween quarter ofbirthandeither educational at-
ofbirthandeducational attainment variesacrossstates, de- tainment orearnings.

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
Bound, Jaeger, and Baker: Problems WithInstrumentalVariables Estimation 447

Areanyoftheseseasonaleffects largeenoughto cause smalldifferences intheseasonalbirth pattern havesubstantial

largebiasesonthecoefficient ofinterest intheestimation of effects
on theestimated coefficient
on education.
eithereducational attainment orwageequations? As noted Theearlier calculation usingdifferences inpercapitafam-
earlier,thefamily incomeofthosebornearlyin theyear ilyincomeacrossquarter ofbirthsuggests thatomitted factors
tendstobe lowerthanthefamily incomeofthosebornlater couldeasilyaccountforall ofthesmallassociation between
in theyear.Our calculations usingthe 1980U.S. Census earnings andquarter ofbirth.Asnotedpreviously, wefound
showthatthedifference inmeanlogpercapitafamily income thatthedifference in meanlog percapitafamilyincome
between thoseborninthefirst quarter oftheyearandthose amongchildren age0 to 3 yearsbetween thosebornin the
bornin thesecondthrough fourth quarters oftheyearis first quarter oftheyearandthoseborninthesecondthrough
-.0238 amongchildren age0 to 3 years.Ideally,we would fourth quarters oftheyearis -.0238. Solon(1992) andZim-
likean estimate ofthisdifference forthesamplesof men merman (1992) bothfoundan intergenerational correlation
bornin the1930sand 1940sthatAngrist and Krueger an- in long-run incomeof at least.4. Giventheirresults, the
alyzed.Butbecauseseasonalvariation infertilityhasdeclined differences infamily incomesbetween thoseborninthefirst
overthelast50 years(Lam andMiron1991;Seiver1985), quarter andtherestoftheyearwouldleadus toexpectthat
ourestimate ofdifferences in family incomebychildren's menborninthewinter toearnabout.95%lowerwagesthan
seasonofbirthis likelyto underrepresent thedifference for thosebornduring therestoftheyear.Data fromthe1980
menbornin thatperiod.UsingdatafromthePanelStudy U.S. Censusshowthatamongmenbornduring the1930s,
of IncomeDynamics(unreported resultsusingthesame thosebornduringthefirst quarterearn 1.1%lowerwages
sampleas in Solon1992)foundthatin a regression ofchil- (AK-91). Thusdifferences infamilyincomeattimeofbirth
dren'seducational attainment onthelogoffathers' earnings,wouldseemto accountforvirtually all oftheassociation
a 1% riseinfathers' earnings wasassociated witha .014-year between quarterofbirthandwages.
risein childrens'educational attainment. The difference in
family incomebetween thosebornin thefirst quarterand 3.2 Finite-Sample Bias
thosebornin subsequent quarters thuscanexplainapprox- Becausequarterofbirthand educational attainment are
imately a .03gradedifference ineducational attainment be- onlyveryweakly correlated,AK-9I's estimates maybe sub-
tweenthesegroups.AK-91reported thatformenbornbe- jectto finite-sample biasevenwithitsenormous samples.
tween1930and 1939,thoseborninthefirst quarter hadon To investigate thispossibility,we reexamined thesedata
average.1 yearlessof educational attainment thanthose drawnfromtheU.S. Census.AK-91usedthree10-year co-
borninthesecondthrough fourth quarters. Thusdifferenceshorts, emphasizing resultsformenbornbetween1930and
in familyincomeacrossthosebornin different quarters 1939.We usethissampleto illustrate ourpoints.We were
wouldseemto be capableofexplaining aboutone-third of ableto replicateexactly AK-91'ssamplesand results using
theassociation between quarter ofbirthandeducational at- theinformation in itsAppendix1,and we refer thereader
tainment. therefordetailsregarding sampleandvariablecreation.
In termsofwages,theweakassociation betweeneduca- Table 1 presents estimates oftheeffects ofeducationon
tionalattainment andquarter ofbirthindicates thatevenif the logarithm of men's weeklyearnings.Columns(3)
otherseasonaleffects areweak,theycouldstillhavelarge through AK-9I's TableV, columns
(6) replicate (5) through
effectson theestimated coefficients. AK-91 actuallypre- (8). Columns(3) and (5) present OLS estimates, and col-
sentedresultssuggesting theeffects ofquarterofbirthon umns(4) and(6) present IV estimates, withquarter ofbirth
wagesindependent oftheeffect ofquarterofbirthon edu- andquarter ofbirthinteracted withyearofbirthas theex-
cationalattainment. Itreported IV estimates thatcontrolledcludedinstruments. Columns(3) and (4) differ fromcol-
and did notcontrolforrace,urbanstatus,maritalstatus, umns(5) and(6) intheagecontrols used.Columns(3) and
and regionofresidence. In eachcase,including thesevari- (4) use singleyearofbirthdummies, whereasin columns
ablesas controlsreduces theIV estimates substantially more (5) and(6) ageandagesquared,measured inquarter years,
thantheir inclusion reduces theOLS estimates. Forexample, areadded.In additionto thesereplications, in columns(1)
inAK-91whenthesecontrols wereaddedtotheOLS results and (2) we report resultsfroma simpler specification than
reported in column(3) ofTableV, thecoefficient on edu- thoseusedin AK-91. Thesemodelsuse onlyage and age
cationdropped11%,from .071to.063.In comparison, when squaredas controls forage,andcolumn(2) usesquarter of
thesamecontrols wereaddedto theIV estimates reported birthdummieswithout year-of-birthinteractions as instru-
incolumn4,thecoefficient oneducation dropped 21%,from ments.We report thecoefficientand standard erroron ed-
.076 to .060. This resultimpliesan associationbetween ucationfromeachOLS and second-stage IV regression. In
quarterofbirth andthecontrol variables. Forexample, blacks addition, we reporttheF statistic forthetestofthejoint
are .7% morelikelythanwhitesto havebeenbornduring statistical significanceoftheexcludedinstruments and the
thewinter quarter.Becauseblackson averagehavelower partialR2 oftheexcludedinstruments fromthefirst-stage
educational attainment andearnings thanwhites, racepar- regression ofeachIV specification.Wealsoreport Basmann's
tiallyaccountsforthelowereducationalattainment and (1960) F testforoveridentificationforeachIV specification.
earnings amongindividuals bornduring thewinter quarter. The F statistic on the excludedinstruments decreases
In thesamplesused in AK-91, the associations between acrossthedifferent IV specificationsin Table 1. In thesim-
quarterofbirthand othercovariates (race,maritalstatus, plestspecification thatincludesonlythreequarter-of-birth
andlocationofresidence ) areallquitesmall.Buteventhese dummiesas instruments, theF statistic suggests negligible

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
448 Journal of the American Statistical Association, June 1995

Table 1. EstimatedEffectof CompletedYears ofEducationon Men's Log WeeklyEarnings

in parentheses)
(standarderrorsof coefficients

(1) (2) (3) (4) (5) (6)

OLS IV OLS IV OLS IV

Coefficient .063 .142 .063 .081 .063 .060

(.000) (.033) (.000) (.016) (.000) (.029)
F (excluded instruments) 13.486 4.747 1.613
PartialR2 (excluded instruments,
X100) .012 .043 .014
F (overidentification) .932 .775 .725

Age ControlVariables
Age, Age2 x x x x
9 Year of birthdummies x x x x
Excluded Instruments
Quarterof birth x x x
Quarterof birthx year of birth x x
Numberof excluded instruments 3 30 28

NOTE: Calculated fromthe 5% Public-UseSample of the 1980 U.S. Census formen born 1930-1939. Sample size is 329,509. Allspecificationsinclude
Race (1 = black),SMSA (1 = centralcity),Married(1 = married,livingwithspouse), and 8 Regionaldummiesas controlvariables.F (firststage) and partial
is thatsuggested by Basmann (1960).
inthe firststage of IV estimation.F (overidentification)
R2 are forthe instruments

finite-sample bias.Becausequarter ofbirthisrelated,bydef- andTable2,columns (2) and(4), usingrandomly generated

inition,to agemeasured in quarters withina singleyearof information inplaceoftheactualquarter ofbirth, following
birth,andbecauseageis an important determinant ofearn- a suggestion byAlanKrueger. The meansoftheestimated
ings,wefindthespecification usingwithin-year agecontrols standard errors reported inthelastrowarequiteclosetothe
[column(6) ] tobe moresensible thanthespecification that actualstandarddeviations of the 500 estimates foreach
doesnot[column(4) ]. TheF statistic on theexcludedin- model.Moreover, thedistribution oftheestimates appears
struments in column(6) indicates thatquantitatively im- tobe quitesymmetric. In thesecases,therefore, theasymp-
portant finite-sample biasesmayaffect theestimate. Com- toticstandard errors givereasonably accurateinformation
paringthepartialR2 in columns(2) and (6) showsthat onthesampling variability oftheIV estimator. Thisisspecific
adding25 instruments does not changethe explanatorytothesecases,however. NelsonandStartz(I 990a) showed,
poweroftheexcluded instruments byverymuch,explaining in thecontext ofa different example, thatasymptotic stan-
whytheF statistic deteriorates so muchbetweenthetwo darderrors cangiveverymisleading information aboutthe
specifications. actualsampling distribution oftheIV estimator whenthe
Compulsory attendance laws,and thedegreeto which correlation betweenthe instrument and the endogenous
theselawsareenforced, varybystate.In AK-91 theauthors variableis weak.
usedthiscross-state variation tohelpidentify thecoefficient
on education byincluding stateofbirthX quarterofbirth Table2. Estimated Effect ofCompleted YearsofEducation on
interactions as instruments in someoftheirspecifications. Men'sLog Weekly Earnings, Controlling forStateofBirth
Besidesimproving theprecision oftheestimates, usingvari- (standard errors ofcoefficients inparentheses)
ationacrossstateofbirthshouldmitigate problems ofmul- (1) (2) (3) (4)
ticollinearitybetweenage and quarterofbirth.In Table2 OLS IV OLS IV
we reportreplications ofAK-9I's Table VII, columns(5) .063 .083 .063 .081
Coefficient
through (8). Thesemodelsuse quarterofbirthX stateof (.000) (.009) (.000) (.011)
birthinteractions inaddition toquarter ofbirthandquarter 2.428 1.869
F (excludedinstruments)
ofbirthX yearofbirthinteractions as instruments fored- PartialR2(excludedinstruments, X100) .133 .101
ucationalattainment. F (overidentification) .919 .917
Including thestateofbirthX quarter ofbirth interactions AgeControl Variables
reducesthestandard errors on theIV results bymorethan x x
Age, Age2
a factoroftwoandstabilizes thepointestimates considerably.9 Yearofbirth dummies x x x x
TheF statistics ontheexcluded instruments inthefirst stage ExcludedInstruments
ofIV do notimprove, however. TheseF statisticsindicate
birth x x
thatalthough including stateofbirthX quarterofbirthin- Quarterof ofbirth x yearofbirth x x
Quarter
teractions improves theprecision I reduces
arn theinstabilityQuarterofbirthx stateofbirth x x
oftheestimates, thepossibility thatsmall-sample biasmay Number ofexcludedinstruments 180 178
be a problem remains. NOTE: Calculatedfromthe5% Public-UseSample ofthe 1980 U.S. Census formenbom 1930-
To illustratethatsecond-stage resultsdo notgiveus any 1939. Sample size is 329,509. AllspecificationsincludeRace (1 = black),SMSA (1 = central
Maried (1 = maried,livingwithspouse), 8 Regionaldummies,and 50 State of Birthdummies
indication oftheexistence ofquantitatively important finite-city),
as controlvariables.F (firststage) and partialR2 are forthe instrumentsin the firststage of IV
samplebiases,wereestimated Table1,columns(4) and(6), estimation.F (overidentification) is thatsuggested by Basmann (1960).

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
Bound, Jaeger, and Baker: Problems WithInstrumentalVariables Estimation 449

It is strikingthatthesecond-stage resultsreportedin Table Table 3. EstimatedEffectof CompletedYears ofEducationon Men's

Log WeeklyEarnings,UsingSimulatedQuarterofBirth
3 lookquitereasonable evenwithno information aboutedu-
(500 replications)
cationalattainment in thesimulated instruments. Theygive
no indication thattheinstruments wererandomly generated. Table (column) 1 (4) 1 (6) 2 (2) 2 (4)
Astheanalytic resultsinSection 2.2wouldsuggest, themean
EstimatedCoefficient
oftheestimated coefficientsin eachcaseis closetothecom-
parableOLS estimate. Theestimated standard looken-
errors Mean .062 .061 .060 .060
couraging andareonlysomewhat largerthantheonesreportedStandarddeviationof mean .038 .039 .015 .015
forthecomparable IV estimates reported in Tables1 and 2 5th percentile -.001 -.002 .034 .035
where actualquarter ofbirth isusedas theinstrument. On the Median .061 .061 .060 .060
other hand,theF statistics ontheexcluded instruments inthe 95th percentile .119 .127 .083 .082
first-stage regressionsarealwaysneartheirexpected valueof EstimatedStandardError
essentially 1 anddo givea clearindication thattheestimatesMean .037 .039 .015 .015
ofthesecond-stage coefficientssufferfrom biases.
finite-sample
Calculatedfromthe5% Public-UseSample ofthe 1980 U.S. Census formenborn1930-
We concludethatitis likelythatsomeoftheresults re- NOTE:
1939. Sample size is 329,509.
portedin AK-91 areaffected byquantitatively largefinite-
samplebiases.Ourresults implythatifthecorrelation be-
tweentheinstruments andtheendogenous issmall, largecross-sectional
variable samplesshouldbe cautiousaboutadding
theneventheenormous samplesizesavailablein theU.S. instruments to increaseprecision.
Censusdo notguarantee thatquantitatively importantfinite- Our resultsemphasizetheimportanceof examiningchar-
samplebiaseswillbe eliminated fromIV estimates. They acteristics of thefirst-stage estimates.The standarderrorsre-
alsoindicate thatthecommonpractice ofaddinginteractionportedin AK-91 appearreasonable, and overidentification tests
terms as excluded instruments mayexacerbate theproblem, giveno indicationthatthe authors'modelsare misspecified.
evenwhilereducing thestandard errorofthecoefficient on But despitetheseobservations, we have shownthatthereis
theendogenous explanatory variable. goodreasonto doubtthevalidity ofinferences theydrewabout
4. CONCLUSION theeffectofeducationalattainment on earningswhenquarter
ofbirthis used as an instrument.More generally, our results
Theseresultsillustrate thattheuse of instruments that suggestthatthepartialR2 and F statisticon theexcludedin-
jointlyexplainlittleofthevariation intheendogenous vari- struments inthefirst-stage areusefulas roughguides
regression
ablecan do moreharmthangood.The examplewe chose to thequalityof IV estimates.We suggestthatbothstatistics
toanalyzeisnoteworthy, becauseAngrist andKrueger would be routinelyreported whenIV estimates are presented.
haveseemedto be on stronggroundin choosinga valid
instrument. Theyproduced evidencesupporting thenotion APPENDIX:THEEXACTFINITE-SAMPLE BIAS OF IV
thatcompulsory attendance lawsinducea correlation be-
IV estimatesare biased in finitesamples.As we have shownin
tweenquarterofbirthand educational attainment.More- the text,the magnitudeof thatbias can be approximatedusing
over,itseemsimplausible thattherewouldbeanyverystrong powerseriesexpansionmethods.It is also possibleto derivethe
directassociation betweenquarterofbirthand wages.We exactmagnitudeofthebias undertheassumptionofnormality.In
haveshown,however, thattheseconclusions arenotsuffi-thisappendixwe expandon Sawa's ( 1969) workon theexactmag-
cientto ensurethattheuse ofquarterofbirthas an instru- nitudeofthefinite-sample bias ofIV to showthatthetwomethods
mentforeducational attainment willreducethemagnitude yieldcomparableresults.
oftheinconsistency inherent in theuse ofan endogenous We areinterested in estimatingequation(A. 1) fromthefollowing
variableas a regressor. system(in which,forsimplicity,we assumethatall randomvariables
Havingbecomeacutely awareoftheendogeneity ofmany have mean 0):
ofthevariables whoseimpactwewishto study, wetendto y Ox +e, (A.1)
believethattheuse ofplausibleinstruments willimprove
x =ZII + v, (A.2)
thevalidity ofourinferences. Although standarderrorsmay
be large,we imaginethatwe haveeliminated mostofbias wherex, e, and v are N X 1 vectorsofindependentrealizationsof
inherent in theOLS estimates. ButEquation(6) indicates therandomvariablesx, e, and v,respectively,y is an N X 1 vector,
thatthismaynotbe trueevenwithinstruments thatseem Z is an N X K matrixin whichtherowsareindependentrealizations
reasonably exogenous totheprocessunderstudy.Ifa setof ofthevectorz, composedof randomvariablesz1, . . ., Zk, II is a
potential instruments is weaklycorrelated withtheendog- K X 1 vectorof constants,and ,Bis a scalarconstant.We assume
enousexplanatory variable(as is oftenthecase),theneven thatE(e Iz) = 0 and E(v Iz) = 0. In addition,because E(&i3) does
not existwhenK = 1, we assumethatK > 1.
a smallcorrelation betweenthepotential instrumentsand
Let thereducedformof equation(A. 1) be
theerrorcan seriously biasestimates.
We also showthatworking withlargesamplesdoesnot y = ZIIO + vO. (A.3)
insulateus fromquantitatively important finite-sampleDefine
biases.Although wehaveno wayofknowing towhatextent
thisissueis empirically important forthoseworking with (A.4)
2r
suchdata,ourresults suggest thateventhoseworking with

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions
450 Journal of the American Statistical Association, June 1995

TableA.1. Bias ofIVEstimates

Relative
toOLS Estimates Bowden,R. J.,and Turkington, D. A. (1984), Instrumental Variables,
Cambridge, U.K.: Cambridge UniversityPress.
T2/K Buse,A. (1992), "TheBiasofInstrumental VariableEstimators," Econo-
metrica, 60, 173-180.
K .5 1.0 2.0 4.0 10.0 100.0 Carroll, H. C. M. (1992),"SeasonofBirth andSchoolAttendance," British
Journal ofEducational Psychology, 62,391-396.
2 .61 .37 .14 .02 .00 .00 Durbin, J.(1954), "Errors in Variables," ReviewoftheInternational Sta-
3 .62 .41 .21 .09 .03 .00 tistical
Institute, 22,23-32.
10 .65 .47 .30 .17 .08 .01 Gillberg,C. (1990), "Do Children withAutism HaveMarchBirthdays?,"
20 .66 .49 .32 .19 .08 .01 ActaPsychiatrica Scandinavica, 82, 152-156.
100 .67 .50 .33 .20 .09 .01 Griliches,Z. (1977), "Estimating theReturns to Schooling: SomeEcono-
200 .67 .50 .33 .20 .09 .01 metric Problems," Econometrica, 45, 1-22.
Hare,E. H. (1975), "Manic-Depressive Psychosis and SeasonofBirth,"
NOTE: K is the numberof excluded instruments and T2IK is the populationanalog to the F ActaPsychiatrica Scandinavica, 52,69-79.
statisticforthejointstatisticalsignificanceofthe instruments inthefirst-stage
regression.Entries Kestenbaum, B. (1987), "Seasonality ofBirth:TwoFindings from theDe-
are [1 - (ir2/K)IFI(1, (K + 2)/2; -T2/2)], whichis theapproximatebias offi. relativeto4. when cennialCensus,"SocialBiology, 34,244-248.
the R2 between the instruments and the endogenous explanatoryvariableis small. Details are
containedinthe text.
Knoblock, H.,andPasamanick, B. (1958),"SeasonalVariation intheBirths
oftheMentally American
Deficient," Journal ofPublicHealth, 48, 1201-
1208.
Lam,D., andMiron,J.A. (1991), "Seasonality ofBirths in HumanPop-
and ulations" SocialBiology, 38,51-78.
Livingston, R., Adam,B. S., and Bracha,H. S. (1993), "SeasonofBirth
and Neurodevelopmental Disorders: SummerBirthis Associated with
p 2 (A.5) Dyslexia,"Journal oftheAmerican Academy ofChildandAdolescent
Psychiatry, 32,612-616.
Mortimore, P., Sammons, P., Stoll,L., Lewis,D., and Ecob,R. (1988),
Sawa showedthatundertheassumptions thatvand Poaredistributed SchoolMatters: TheJunior Years,Somerset, U.K.: OpenBooks.
as jointlynormal,theOLS and IV biasescan be writtenas Nagar,A. L. (1959),"TheBiasandMoment Matrix oftheGeneral k-Class
Estimators oftheParameters inSimultaneous Equations," Econometrica,
E(3ols)- = -IP)[y 1 1FI( I, N+ 1
; - I
27,575-595.
Nelson,C. R.,andStartz,
Small-Sample
R. (1990a),"SomeFurther
Properties oftheInstrumental
Results
Variable
ontheExact
Estimator," Econ-
ometrica, 58,967-976.
(A.6) "TheDistribution oftheInstrumental Variable Estimator
J(1990b),
anditst-Ratio whentheInstrument isa PoorOne,"Journal ofBusiness,
_ ] (A.7 63 (Part2), S125-S140.
E(Oiv)-,B = (/3-P)[ KIF,( 1, 2
O'Callaghan, E.,Gibson, T.,Colohan, H. A.,Walshe, D., Buckley, P.,Larkin,
C., andWaddington, J.L. (1991), "SeasonofBirthin Schizophrenia:
where1F1 , is the confluenthypergeometric function.Note Evidence forConfinement ofanExcessofWinter BirthstoPatients With-
thatr2/(N-1) I 2 the outa FamilyHistory ofMentalDisorder," British Journal ofPsychiatry,
R populationR2 fromtheregression of
x on z, and thatT21K is populationanalog to F statisticforthe 158,764-769.
Richardson, D. H. (1968),"TheExactDistribution ofa Structural Coefficient
regression ofx on z. For largevalues ofN and smallvalues of T2, Estimator," Journal oftheAmerical StatisticalAssociation, 63, 1214-
1F1(l,(N+ 1)/2;-92/2) 1.ForlargeNand smallR ,therefore, 1226.
theOLS biasapproachesj3-p and therelativebiasofIV approaches Sawa,T. (1969), "The ExactSamplingDistribution of Ordinary Least
theexpressionin squarebracketsin (A.7). SquaresandTwo-Stage LeastSquares Estimators,"Journal oftheAmerical
Statistical
Association, 64,923-937.
Althoughthe implicationof Equations (A.6) and (A.7) is far Seiver,D. A. (1985), "Trendand Variation in theSeasonality ofUnited
fromobvious,itis possibleto approximate1F1( 1, y; N) forvarious StatesFertility, 1947-1976", Demography, 22,89-100.
values of oyand t. In Table A. 1 we presentthe magnitudeof the Sham,P. C., O'Callaghan, E., Takei,N., Murray, G. K., Hare,E. H., and
bias ofIV relativeto OLS forvariousvaluesofT2/KandK. Clearly, Murray, R. M. (1992), "Schizophrenia Following Prenatal Exposure to
Influenza Epidemics Between1939and 1960,"British Journal ofPsy-
thebias ofIV relativeto OLS dependson theT2/K,thepopulation chiatry,160,461-466.
analog to theF statisticon theexcludedinstruments. Shea,J.(1993), "Instrument Relevance inLinearModels:A SimpleMea-
sure,"unpublished manuscript, University ofWisconsin-Madison.
[Received June1993.RevisedJune1994.] Solon,G. (1992),"Intergenerational IncomeMobility intheUnitedStates,"
American Economic Review, 82,393-408.
D., and Stock,J.H. (1994), "Instrumental
Staiger, Variables Regression
REFERENCES withWeakInstruments," Technical Working Paper151,National Bureau
ofEconomicResearch, Cambridge, MA.
Angrist, J.D., Imbens, G. W.,andRubin,D. B. (forthcoming), "Identifi- Tarnowski, K. J.,Anderson, D. F.,Drabman, R. S.,andKelly,P. A.(1990),
cationofCausalEffects UsingInstrumental Variables," Journal ofthe "Disproportionate Referrals forChildAcademic/Behavior Problems:
American Statistical Association, forthcoming. Replication and Extension," Journal ofConsulting and ClinicalPsy-
Angrist, J.D., andKrueger, A. B. (1991), "Does Compulsory SchoolAt- chology, 58,240-243.
tendance Affect Schooling andEarnings?," Quarterly Journal ofEconom- Templer, D. I., Trent,N. H., Spencer, D. A., Trent,A., Corgiat, M. D.,
ics,106,979-1014. Mortensen, P. B.,andGorton, M. (1991), "SeasonofBirthin Multiple
(1992),"TheEffect ofAgeatSchoolEntry onEducational Attain- Sclerosis," ActaNeurologica Scandinavica, 85, 107-109.
ment:An Application ofInstrumental Variables WithMoments From Wald,A. (1940) "TheFitting ofStraight LinesifBothVariables areSubject
TwoSamples," Journal oftheAmerical Statistical Association, 87,328- to Error,"AnnalsofMathematical 11,284-300.
Statistics,
336. Watson, C. G., Kucala,T., Tilleskjor, C., andJacobs, L. (1984), "Schizo-
Bartels, L. M. (1991),"Instrumental and'Quasi-Instrumental' Variables," phrenic BirthSeasonality inRelation totheIncidence ofInfectious Diseases
American Journal ofPolitical Science,35,777-800. andTemperature Extremes," Archives ofGeneral Psychiatry, 41,85-90.
Basmann, R. L. (1960), "On Finite-Sample Distributions ofGeneralized Whorton, J.E.,andKarnes, F. A. (1981),"SeasonofBirth andIntelligence
ClassicalLinearIdentifiability TestStatistics," Journal oftheAmerical in SamplesofExceptional Children," Psychological Reports, 49, 649-
Statistical Association, 55,650-659. 650.
(1963), "RemarksConcerning theApplication of ExactFinite Williams, P., Davies,P., Evans,R., and Ferguson, N. (1970), "Seasonof
SampleDistribution Functions ofGCL Estimators in Econometric Sta- BirthandCognitive Development," Nature,228,1033-1036.
tistical Inference," Journal oftheAmerical Statistical Association, 58, Zimmerman, D. J.(1992), "Regression TowardMediocrity in Economic
943-976. Stature,"American Economic Review, 82,409-429.

This content downloaded from 128.189.191.239 on Wed, 12 Feb 2014 13:24:52 PM

All use subject to JSTOR Terms and Conditions