instrumental variable methods for the estimation of test score reliability

advertisement
Journal of Educational Statistics
Fall 1983, Volume 8, Number 3, pp. 223-241
INSTRUMENTALVARIABLEMETHODSFOR
THE ESTIMATIONOF TEST SCORERELIABILITY
RUSSELL ECOB
Inner London Education Authority
AND
HARVEY GOLDSTEIN
University of London Institute of Education
EducaKey words:InstrumentalVariables,Errorsin Variables,Reliability,Longitudinal,
tionalAttainment
method of InstrumentalVariablesis suggestedas an alternativeto
traditionalmethods for estimatingthe reliabilityof mental test scores which avoids
certaindrawbacksof thesemethods.The consistencyand efficiencyof the instrumental
variablemethodare examinedempiricallyusing data from the BritishNational Child
DevelopmentStudyin an analysisof 16 year, 11 year and 7 year old scoreson tests of
mathematicsand reading.
ABSTRACT.The
In the simple regression model (with the usual assumptions about the error
term u)
y = a + fix + u,
it is well known (Goldstein, 1979) that if the observed independent variable x
contains errors of measurement and we wish to estimate the regression
coefficient on the "true" value of x, then the ordinary least squares (OLS)
estimator is inconsistent. The simplest and most common model relating the
true value to the observed value of x is
x = T+ e,
(1)
where T is the true value, e the random error of measurement, and Cov(T, e)
= 0. It is supposed, therefore, that we wish to estimate the parameters a, /Pin
y = a+ pT+ e
+ fixx + (e - Pie).
=a +
It is because x is correlated with (e - fpe) that the OLS estimator, b, is an
inconsistent estimator of ft. A consistent estimator is given by b/R, where R is
the reliability of x and defined as
R = Var(T)/Var(x),
where Var(x) = Var(T) + Var(e).
223
224
Ecob and Golstein
In many situations,the value of R is very close to 1, and any adjustmentto
the usualestimatecan be safely ignored.In other applications,for examplein
mental testing, R may be considerablyless than 1 so that an adjustment
becomesnecessary.In a linear model with severalfurtherindependentvariables, the estimators of the associated regressioncoefficients will also be
inconsistentif OLS is used. Consistentestimatesmay be obtainedby adjusting
the observed covariance matrix of the independentvariables so that the
observedvariancescorrespondingto variablescontainingmeasurementerror
have estimatesof their measurementerrorvariancesubtractedprior to inversion of the matrixand calculationof the coefficients(Warren,White,& Fuller,
1974).To do this, it is importantto have accurateand consistentestimatesof
the measurementerrorvariances,or alternatively,reliabilities.In this paperwe
exploresomenew proceduresfor obtainingsuch estimatesbasedon instrumental variabletechniques.
ReliabilityEstimation
A large body of literatureexists concerningproceduresfor obtaininggood
estimatesof reliabilityfor mental tests that consist of a fixed set of response
items. Details can be found in Lord& Novick (1968, Chapter4). The methods
that havebeen proposedcan be dividedusefullyinto "internal"and "external"
methods.
Internal Estimates of Reliability
In the internal methods, various relationshipsbetween the test item responses are used to providereliabilityestimates,of which the best known is
"coefficientalpha."Undergeneralassumptionsaboutthe conditionalindependence of item responses,and constancy of measurementerror variances,a
useful lower bound on the reliabilitycan be estimated,which under certain
furtherassumptionsleads to a consistentestimateof the reliability.We write,
similarlyto (1),
xji = Tji+ eii,
wherexi is the observedvalue of itemj for individuali, Tjiis the truevalue of
item j for individual i, and eji is the measurementerror. A conditional
independenceassumptionis usuallymade;thatis, for givenTj,,Ti,j :j'
Cov(ej,ie-,)
= 0.
(2)
The conditionsfor a consistentestimatorof the reliabilityare
ji -
aj+ >,
(3)
whereaj and T;can be interpretedas the difficultyof itemj and the abilityof
individuali. This implies that the items are not only "unidimensional"but
essentially equivalent apart from a location shift. If (2) does not hold, then
Instrumental VariableMethods
225
coefficientalpha is not necessarilya lower bound and may even overestimate
the truereliability.
We see, therefore, that there are two serious drawbacksto 'internal"
estimatesof reliability.First, thereis the difficultyof satisfying(3) to obtain a
consistent estimatorof the reliability,or equivalently,of the measurement
errorvariance.While there is a considerableliteratureon this and the more
in practiceit seemsextremelydifficultto
generalconceptof unidimensionality,
ensurethat item responsesare indeeddeterminedby a singlequantityfor each
individualsuch as givenby (3). For the types of educationaltests we deal with
in this paper,it seems even less likely that a unidimensionaltraitis operating.
A moredetaileddiscussionof this topic is givenby Goldstein(1980).Secondly,
assumption(2), often known as the "local independence"assumption,seems
somewhat unreasonablea priori. It is difficult to imagine that if a given
individualfails one item then the probabilitiesof successon lateritems are the
same as when the individualsucceedson the earlieritem. Nevertheless,there
seems to have been little, if any, serious study of this problem and the
consequenteffect of nonzero correlationson reliabilityestimates.A further
discussion of this point in this context of latent trait models is given by
Goldstein (1980). Thus, there is as yet no really satisfactorymethod for
obtaininga consistentestimateof reliabilityusing "internal"methods,nor of
even providing a lower bound. We suggest that estimates based on these
methodsshouldbe treatedwith somecaution.
External Estimates of Reliability
The most obvious method of estimatingreliabilityor measurementerror
varianceis to carry out repeat measurements.Thus, we have (droppingthe
suffix i), for two applicationsof a test,
X, = T + e,;
X2-= T+ e2;
and Var(X, - X2)= Var(e, - e2)=
2a,2 - Cov(e , e2)}, where a,2 denotes
the constantmeasurementerrorvariance.For many physicalmeasurementsit
is reasonableto assumeindependenceerrors,that is, Cov(e,, e2) = 0, so that
we havea,2 -= Var(Xi - X2). For mentaltests, however,this usuallywill not
be a reasonableassumptiondue to the presenceof memoryeffects, learning,
and so on. If morethanone test relatingto the sameattributeis available,then
by assumingsuitablerelationshipsbetween the true scores on the tests, it is
possible to obtain reliabilityestimates.The usual assumptionis that the tests
are congenericso that for a set of p tests,
X,= a + bjT+ e1,
j= 1,...,p.
Ecob and Golstein
226
The observedcovariancematrixof the Xi contains p( p + 1) elementsand if
we assumeCovj,i(ejej,)= 0 and Cov(Tjej)= 0, the matrixis a function of
the bj and errorvariancesoj, which gives 2p parameters.Hence, for threeor
more tests, unique estimates,based for exampleon maximumlikelihood,are
available.Details of this approachare given in Joreskog(1971).Althoughit is
not quite as seriousas in the simple test-retestcase, this method also has the
difficulty that the measurementerrors of the tests may be correlated,for
example,because of day to day fluctuationsamong examinees,and the like.
This immediatelyraises the questionof definitionof true score, but we shall
postponea discussionof that until a latersection.
In the next sectionwe proposea generalizationof congenerictests to include
any variablehavingnonzerocorrelationwith the test whose reliabilitywe wish
to measure.Such an "instrumentalvariable"does not requireany assumption
about unidimensionalityor independence;also, unlikethe simpletest-retestor
the congenerictest models,the possibilityof choosingany variablemeansthat
we can searchfor those that arelikelyto be uncorrelatedwith the measurement
error eP. The possibility of dropping both these restrictiveassumptionsis
attractiveand the remainderof the paper investigatesthis problemusing an
extensivelongitudinaldata set.
The Data
The data come from the National Child DevelopmentStudy (NCDS) that
followedup a cohort of 17,000childrenborn in one week of March 1958, at
the ages of 7, 11 and 16. The childrenbelonged to the first year-groupfor
whom the minimumschool-leavingage was 16 years. A descriptionof the
social and educationaldata (amongothers)collectedat these ages is given in
Fogelman(1976).
Testingand scoringin the NCDS was carriedout by the class teacherwho
also noted the child'sachievements.Becausethe studywas a nationalstudy of
all childrenborn in a particularweek, most childrenselectedwere testedin a
differentsituationand by a differenttester.
Four possiblesituationsgivingrise to responsevariationare as follows:
1. the environmentin which the test is administered;
2. the processof test administration;
3. the coding and scoring of the test (this includes the interpretationof the
correctnessof the response);and
4. day-to-dayvariationin individualtest performance.
Becauseonly one test of a given type was done by each child at each occasion,
the sourcesof variation1-4 aboveare confounded.It is important,however,to
distinguish"day-to-day"variationfrom changesin true score over time. We
can regardvariationover time as contributingeitherto measurementerroror
to true score variation or to both. A reasonable estimate of the true score at a
Instrumental VariableMethods
227
particularmomentwould be obtainedfrom a movingaverageof scores taken
at successivetimeintervalsbeforeand after.The continuouschangein truetest
scoreover,say, a weekis thereforeregardedas beingsupplementedby random
error to produce the observedday-to-dayvariation.The variouseducational
measurementsin the NCDS were completedwithin a week for each child, so
that any true scorechangesover not morethan a 1-weekperiodare effectively
regardedas part of day-to-day variation. In addition to these sources of
measurementerror,there will typicallyremainan unexplainedvariationthat
can be conceptualizedas the variationbetweenthe responseto an item and its
hypotheticalreplication.
Cronbach,Gleser,Nanda, & Rajartnan(1972) arguethat test evaluationor
studies,whichalso view a particulartest as a samplefrom a
"generalizability"
universeof tests and which use experimentaldesigns to estimateindividually
the above components of variation, should be carried out prior to test
administration.We use here a "test-specific"interpretationof true score that
treatstruescoreas relevantonly to the particulartest. A justificationfor this is
given by Goldstein (1979), althoughthe methods used in this paper can be
extended to a full "generalizability"approach.Goldstein (1979), using the
same NCDS data, also drew attentionto the use of instrumentalvariablesin
estimatingthe relation between mathematicsand reading attainmentswhen
measuredat different ages. He emphasizedthe potential usefulnessof this
method when impreciseprior knowledgeabout the reliabilityof the earlier
attainmentscoresis available,and pointedout that little was knownabout the
degree to which the instrumentalvariablesused satisfied the conditions of
consistency.
In this paper, the propertiesof a variety of instrumentalvariables are
examinedin the context of the regressionof 16 years attainmenton 11 and 7
years attainmentfor mathematicsand readingtest scoresseparately.Comparisons are made with the use of ordinaryleast squaresand also with the use of
the internalestimatesof the reliabilitycoefficientsfor the 11-yearattainment
givenin Goldstein(1979).
VariablesEstimation
Theoryof Instrumental
GeneralTheory
SupposeX1i,X2i are the observedvaluesof test score variablesmeasuredas
deviationsfrom theirmeans at the first and second occasionsand let thembe
the predictorand dependentvariables,respectively,in a simple linear regression model.Furthermore,denote theirtruevaluesas Tli and T2, and the errors
of observationor measurementerrors for the ith subject as eUi and e2i
(i = 1,...,n). Then,as before:
Xi = Tli + eli;
X2i T2i + e2i,
(4)
228
Ecob and Golstein
and a modelrelatingthe truevaluesat eachoccasionis
T2i Tli+ ui.
(5)
If Z1is theobservedvalueof anothervariable,calledtheinstrumental
variable,
then
n
n
bIv =
ZiX2i
(6)
ZiXi
is called the instrumentalvariableestimatorof the regressioncoefficient 3
(Johnston,1972,p. 279).
From(4), (5) we have
biv=-(fP ZTi +
Ziui+
Ze2i)(
Z,Xi)',
and bv - p = (2 Zu, + Zie2i - P I2Zieli)(I ZiXli)-'. In terms of the sample correlations,rze, 2Ze and rze, betweenthe instrumentalvariableZ and the
disturbanceand errors of measurementof X2, X,, respectively,and the
reliabilityR of X1,
bIv
-
Urzu +
Ze/2
e2rze2
e
zx,
rze
(7)
whereae,, ae2, u,are, respectively, the standarddeviationsof the errorson the
1st occasion,2nd occasion,and the disturbanceterm.As the samplesize tends
to infinity,the followingconsistencyconditionis obtained:
+
ouPZU
aezPZe2-
= 0.
floaePZe,
Thus, a generalcondition for consistencyis that the three correlationspzu,,
PZe29 PZe, are all zero. Equation (6) shows that if the predictorand dependent
variablesare interchanged,the instrumentalvariable estimatorbecomes its
reciprocal.
The efficiency of the instrumentalvariableestimatoris the squareof the
multiplecorrelationof the instrumentalvariableset with the predictor.This
assumesthat the estimatoris consistent.
The Use of Many Instrumental Variables
Whenwe havep instrumentalvariablesZ, j = 1,...,p,
blv
=
2
t ij
jZijX2i
1
(
c
i
ZijXli.
(8)
j
The combinationof Z, that gives the most efficient estimate of biv can be
found by choosing the constant cj so that Corr(YIc-Z.-, Xi) is a maximum.
This is the usual two-stage,least squaresestimator(Johnston, 1972, p. 380).
The ci are then the sample regression coefficients, bj, of X, on Z, j = 1,...,p.
Letting Xi =-- b,Z, , we obtain biv = (2 XliX2i)/E XIXi,.
Instrumental VariableMethods
229
S
The Use of Dummy Variablesas Instrumental Variables
The previousdiscussionhas assumedthe existenceof instrumentalvariables
that can be modelled as having simple linear relationshipswith the first
occasion variables.Two other cases can be distinguished.First, where an
intervalscaled instrumentalvariablehas a nonlinearrelationshipto the first
occasionvariable,and second,when the instrumentalvariableis categoric(e.g.,
measuredon an ordinalor nominalscale).
In the first case the nonlinear relationshipcan be modelled, say by a
polynomial function, or the instrumentalvariablecan be groupedinto categories.In the lattercase, eachcategorycan be representedin the usualway by
a dummy variable.This takes the value 1 for this categoryand 0 for every
other category. Let X,,,, XIrTbe two observations on the first occasion variable
belonging to the same instrumentalvariablecategory,r. Using the dummy
instrumentalvariablesto estimatethe first occasionvariablegives the estimate
r,,whichis the meanvalueof all observationsin categoryr.
Substitutingin (8) gives
biv =
r p,
X2
2r
X,
r
prSir)-2
where X2r is the mean of the X2r in the category r, and Pr is the proportion in
category r. This is essentiallythe "Method of grouping"as introducedby
Wald(1940).
The literatureon groupingmethods(e.g., see Madansky,1959; Neyman &
Scott, 1951;Wald, 1940)using observedvalues of X, has tended to focus on
conditionsfor consistentestimatesand on the relativeefficiencyof different
groupingsratherthan quantifyingthe inconsistencyof variousgroupingmethods. The resultson the NCDS data, given below, go some way to remedying
this situationfor a particulardata set.
VariableMethodsto the NCDS Data
Applicationof Instrumental
Selection of Variables
In all, 50 variablesare consideredas instrumentalvariables,being measured
at ages 7, 11 and 16. These consist of test scores, teacher ratings and
backgroundvariables.The test scoresare of readingand mathematicsat each
age, and in addition,of generalability and copyingdesigns scoresat age 11.
The teacherratingsareof readingand mathematicsat all ages,and in addition,
of oral abilityand creativityat age 7, of oral abilityand generalknowledgeat
age 11, and of practicalsubjectsat age 16. The "background"variablesare
socialclass and indicesof behaviorin the home at all threeages;the numberof
childrenin the household,birthorderof the child,an index of accommodation
facilities and childrens' heights at ages 11 and 16, overcrowding at 11; and
230
Ecob and Golstein
region, indices of school behavior,and a varietyof feelings towardschool at
16. The readeris referredto Davie, Butlerand Goldstein(1972)and Fogelman
(1976)for a morecompletedescriptionof thesevariables.
The variablesused as dependentvariablesin the regressions,that is, the test
scores at age 11 and 16 of mathematicsand reading,are transformedto have
standardnormal distributionsin the same way as in Goldstein (1979) who
showedthatnearlinearrelationshipsbetweenobservedscoresresulted.
As the relativeefficiencyof an instrumentalvariableestimatoris proportional to the squareof the correlationwith the predictor,variablesare only
retained for furtheranalyseswhen this correlationis greaterthan 0.3. This
eliminatesall the "background"variablessave socialclass,but only one of the
teacherratings(that of outstandingabilityin any area at age 11) and none of
the test scores,leaving25 variablesin all. All cases with missingvalueson any
of these 25 variablesare excluded,leaving5,371 cases with test scoresat each
age.
In the appendixto Fogelman(1976), Goldsteinshows that the attritionof
subjectsin the study does not affect to any markedextent the relationships
found and that test scores for subjectshaving missing values on the backgroundvariablesshowno significantdifferencesfromothersubjects.
In the resultsreportedhere, all instrumentalvariablesare treatedas sets of
dummyvariablesfor reasonsgiven earlierin that section. For the test scores
the dummyvariablecodinginto five roughlyequalsize groupsensuresthat the
relativeloss in predictiveefficiencyfromdummyvariablescomparedto simple
linearregressionis alwaysless than6 percent.
In fact, using these instrumentalvariablesas dummyvariablesor as interval
scale variablesgives very similar results, with the maximum differencein
regressioncoefficientsfor any instrumentalvariablebeing2 percent.
Forming Hypothesesof Error Structurein PredictionRelations
As in the section on ExternalEstimatesof Reliabilitywe assume that the
true scoreof a test comprisesthat componentof test scorethatis unaffectedby
day-to-dayvariationby the particulartestor or by the test situation,and is
specific to the particulartest used. We now examine the correlationof the
variablesto be consideredas instrumentalvariableswith measurementerrors
on the first occasionand secondoccasiontests and with the disturbanceterm.
These variablesare teacherratingson a varietyof attainments,test scoresand
social class. The teacher ratings, like the test scores, generallywill contain
measurementerror, thus reflectingand being reflectedby variationsin the
child'sinterestin subjectsand day-to-dayvariationsin the type of relationship
to the teacher.Thus, a teacherwho has very recentlyseen a child do a good
piece of work or show a keen interest may tend to rate him higher than
Instrumental VariableMethods
231
otherwise.If the same contributoryfactors affect test score, then a teacher
ratingmade at the same time as the test would be expectedto have a positive
correlationwith the test score measurementerror.Likewise,wherea different
attainementis ratedat the sametime as the test, similarcorrelationsmay exist,
althoughpresumablythey wouldbe smaller.
Thereare two variablesthat we hypothesizewill have a zero correlationwith
test score measurementerror.These are teachers'ratingstaken at a different
point in time and social class. We would expect none of the sources of
measurementerrorto relate to teacherratingat a point in time 4 or 5 years
away. Nor would we expect social class, which shows little variationfor an
individualover short time periods,to relateto any of the sourcesof measurement error.
Finally,we examinethe relationof the disturbancetermsto teacherratings
and social class. Either of these variables,particularlysocial class, may be
correlatedwith the disturbancesif they relate to the dependentvariableonce
the predictorvariablehas been controlledfor.
The hypothesesformulatedabovemay be summarizedas follows:
H1. Teacherratingson a test, wherethe ratingis at the same time as the test,
will be positivelycorrelatedwith test scoremeasurementerror.
H2. Teacher ratings on a different attainmentfrom that tested, where the
ratingis at the same time as the test, will be positivelycorrelatedwith test
scoreerrorbut to a lesserextentthan for the sameattainment.
H3. Teacherratingswhen the child is at a differentage from that of the test
will be uncorrelatedwith test score measurementerrorwhetheror not the
sameattainmentis tested.
H4. Teacher ratings are not correlatedwith disturbanceterms from the
regressionof second-occasionscoreon first-occasionscore.
H5. Socialclassis correlatedwith disturbanceterms.
H6. Socialclass if not correlatedwith test score measurementerror.
Generally,teacherratingsare held to correlatewith test score measurement
errorsonly when tested at the same time as the test and not to be correlated
with equationdisturbances.In contrast,socialclass is hypothesizedas correlating with disturbancesbut not with test score measurementerrorwhen measuredat the sametime as the test.
Examiningequation (7) we see that these six hypothesesgive rise to the
followingpredictions.The hypothesesgiving rise to each predictionare given
in bracketsafterthe prediction.
Pl. Comparingteacherratingsat 11 years,the lowest estimateof / will occur
for the teacherratingof the same attainment(from HI to H4, particularly
H2 affectingrze,), and at 16 yearsthe highestvalue will occurfor the rating
of the sameattainment(fromH, to H4, particularlyH2 affectingrze2).
232
Ecob and Golstein
P2. For a givenattainment,teacherratingswill give higherestimatesof 3 when
measuredat 16 than 11 years(fromH, to H4).
P3. For a givenattainment,teacherratingswill give higherestimatesof P when
measuredat 7 ratherthan 11 years(fromH, to H4).
P4. There is no differencein estimatesbetween teachersratings at 7 years
(fromH3, H4).
P5. Social class gives higherestimatesof P than teacherratingsat 7 and 11
years(fromH, to H6).
P6. Socialclass will give similarestimatesof 3 irrespectiveof the age at which
measured(fromH5, H6).
Note that these predictionsare not unequivocaltests of the hypotheses.For
instance,even if P6 holds,one could conceiveof differentcorrelationsof social
class with measurementerrorat differentages, these termsbeing counteracted
by different correlationswith equation disturbances.This, however, seems
unlikely.
Results
Tables I and II give estimated regressioncoefficients for reading and
mathematics,respectively,for 16-yearattainmenton 11-yearattainmentusing
a varietyof teacherratingsand socialclass as instrumentalvariablesat ages 7,
11, and 16. Using the ungrouped11-yeartest score as instrumentalvariable
gives the ordinaryleast squaresestimate,and assumingthe samplesize is large
enough to make use of the asymptotic properties,this enables reliability
estimatesto be calculatedfor each choice of instrumentalvariableby dividing
the ordinaryleast squaresestimateby the instrumentalvariableestimate.Each
predictionwill be examinedin turn.
P . This holds for mathematicsusing teacherratingsboth at ages 11 and 16
and for readingfor teacherratingsat 11 but not 16.
P2. This holds for comparableteacher ratings at ages 11 and 16 for both
readingand mathematicstest scores.
P3. This only holds for one out of the six possible comparisons,namely,
teacherratingof "number"for readingattainmentregression.
P4. Thisholds for both attainments.
P5. Thisholds in all cases.
P6. Thisholds at all ages for both attainments.
Note that predictionsP4 and P6 specify no differencesbetweenregression
coefficients,whereasthe other predictionsare of a differencein a specified
direction.In fact, for P4 and P6 the differencesbetweencoefficientsare small
in relationto the standarderrors,whichhave been estimatedin the usualway,
and are strictlyapplicableonly if the estimatesareconsistent.
The predictionsare all seen to hold generallywith the exceptionof P3. This
implies the rejection of H3 or H4 or both. Rejecting H3 implies that the
~
Instrumental VariableMethods
C13
233
"
,
"
-
C)
C)N
00
C)
0
C)
0
V
0000
cl-
r
N-M
-00
F
.
--.
C
;•,C-),C )
•
,--
C,3,
cd
bONO
=41
x,
ON
u
0
I-Sr
CA
00
cd
CA
C,
ZZ
•
r d d d
r-
4~
k
C1
bD
)
'Tt
C=
9
OON
\o
CS
'd
6666
I-~
Zt
C13
c -:
S0 -
•,
0t-o0
%f)itCI
0?I)
Cs
C~sC~
C~Cs
- 000
O)C
o
Pc
o•
"00 4
C13
cj
c~t
`Cd
u
13C1
C1
C13
Ee
Q,
tQt
1
C
234
Ecob and Golstein
0-
--
00
>cl
~3
~
cl
Cd
_
r-0)
_
_
_
_
000 qzt
r-
00
-
vs
\oO
00
00000
0ON
~
o
_
---a
l
c
000
"~"
0(
6
C6~
SC,
000000a-
O
r) r
0
cl
\9
00<
-U4
w \u
C~tj
c-?cd
clc
$-4
4t
- u ?5l
cl
dU
F-4
cl
F-4
cl c
CIA
00 r
Instrumental VariableMethods
235
differencesin coefficientsamong the differentteacherratingsat age 7 should
be similarto thoseusingthe threecorrespondingteacherratingsat age 11. This
is true with the one exceptionbeing the reversalin the relativemagnitudeof
the teacher ratings of reading and number between ages 7 and 11 for
mathematics.
The smallercoefficient estimatesusing 7-year ratherthan 11-yearratings
could be explained by a correlationof 11-year rating with errors in the
dependent variable, which counteracts the correlationwith errors in the
independentvariable,the 7-year ratings having lower correlationswith the
dependentvariable.
If H4 is the sole reason for the failureof P3, this suggeststhat given the
11-yeartest score,the partialcorrelationof teacherratingswith socialclass at
11, whichif H5 holds is correlatedwith the disturbances,would be higherfor
11-yearteacherratingsthan for 7-yearteacherratings.This is not the case.
It seems,then, that we shoulddiscardsocial class as a suitableinstrumental
variable due to its correlationwith the equation disturbances,and teacher
ratingsat 16 years are positivelycorrelatedwith test score errorat 16 (form
HI1,H2) and probablyalso with equationdisturbances.This leaves a choice
betweenteacherratingsat 7 and 11 years.As H3 does not hold, we cannotbe
completelycontentwith usingthe same-attainment,
7-yearteacherratings,and
in addition, it is not known how highly correlatedthese are with the disturbances.It was suggestedearlier,however,that we shouldexpectthe 11-year
ratingsto be more highly correlatedwith the disturbances.As H2 also holds,
the wisestchoicewould seem to be the ratingof a differentattainment(out of
mathematicsor reading) at age 7. For the reading attainmentthis gives a
reliabilityof 0.81 (using teacherrating of numberat age 7) and for mathematics attainmentgives a reliabilityof 0.89 using teacherratingof readingat
age 7. In fact, the choicebetween7 and 11 yearsfor the instrumentalvariables
makes little differencefor mathematicsattainmentgiving a reliabilityof 0.86
using readingrating at age 11, and for readingattainmentthe differencein
reliabilityestimateis only 0.005.
The instrumentalvariableestimatechosenhere is subjectto possibleinconsistenciesfrom two causes operatingin opposite directions.If H3 does not
hold and the errorsin the instrumentalvariablehave positivecorrelationswith
the test scoreerror,whicharehigherat age 11 that at age 16, then the expected
value of the estimateis higherthan the true value. On the other hand, if H4
does not hold and the errors in the instrumentalvariable have a positive
correlationwith the disturbanceterm,then the expectedvalueis lowerthanthe
true value. If both H3 and H4 do not hold, then these biases operatedin
opposite directionsand an evaluationof the meritsof the estimatedependon
furtheranalysisof the relativestrengthsof the inconsistenciesfrom the two
causes.
Ecob and Golstein
236
The questionnaturallyariseshere as to whetherthe use of tests of the same
attainmentat different ages is necessaryto obtain reasonableestimates of
reliabilitycoefficientsby this method. Estimatesof the reliabilityof li-year
readingtest scoreobtainedby regressing16-yearmathematicsscoreon 11-year
readingscore using as instrumentalvariablesteachers'ratingsof readingand
mathematicsattainmentsseparatelyat 11 years, gave reliabilityestimatesof
0.76 and 0.61, respectively,comparedwith the value of 0.81 give above. This
suggeststhat the disturbancetermsin this regressionare correlatedwith the
mathematicsteachers'ratings. For the regressionof 16-yearreading test on
11-yearmathematicstest using teachers'ratingsof mathematicsand reading
separatelyas instrumentalvariables,reliabilityestimatesof 0.85, 0.68, respectively,are obtained,comparedwith the value of 0.88 given above.Careshould
thereforebe taken when estimatingreliabilitiesby this method to use similar
attainmentsas dependentand predictorvariables.
Use of GroupedFirst-Occasion Variableas Instrumental Variable
Table III gives estimated regression coefficients for both reading and
mathematicswhen the first-occasionvariableis grouped into 2, 3, 5, or 7
groupsof equalsize.
The inconsistencyof the groupingestimator(bG)relativeto the ungrouped
(OLS)estimator(boLs)is givenby
k b=
b
- bG
(9)
bk - bOLS
TABLEIII
Estimated Regression Coefficientsand Standard Errors
Using the GroupedPredictoras Instrumental Variable
Ungrouped (OLS)
No. of groups
(of equal size)
7
5
3
Reading
0.797 (0.0082)
k
Mathematics
k
1.00
0.748 (0.0097)
1.00
0.808 (0.0085)
0.810 (0.0086)
0.94
0.93
0.755 (0.0100)
0.763 (0.0102)
0.93
0.86
0.777(0.0109)
0.73
0.84
0.780 (0.0121)
0.70
0.818(0.0092)
2
0.827 (0.0103)
Varyingposition of dichotomywith 2 groups
Proportionin lower test group
0.89
0.2
0.810(0.012)
0.93
0.687(0.014)
1.58
0.4
0.6
0.8
0.823 (0.010)
0.822 (0.010)
0.789 (0.012)
0.87
0.87
1.04
0.765 (0.012)
0.802 (0.012)
0.805 (0.014)
0.84
0.49
0.46
Note. Standarderrorsin brackets;k is definedin equation9.
Instrumental Variable Methods
237
whereb1v is the instrumentalvariableestimator(using an appropriateteacher
rating) and is assumedto be consistent.As suggestedin the Use of Dummy
Variablessection,substantialinconsistenciesare indicatedin thesedata, which
are greaterwith a largernumberof groups.Thus, thereis a trade-offbetween
inconsistencyand efficiency,the latterbeing greateras the numberof groupsis
increased.For reading attainmentthe lowest estimate of reliability,arising
from the divisioninto two groups,is 0.97. This is higherthan any of the values
derivedfrom the regressionestimatesby instrumentalvariablesmethodsgiven
in Table I. Furthermore,the estimatedregressioncoefficientis seen to vary
with the point of dichotomy.Whereasfor readingthe lowest estimateoccurs
for both extremedivisions,for mathematicsthe lowest estimateoccurs when
divisionis at the lowerend of the scale and the highestestimatewhen division
is at the higherend.
If the assumptionis made that the correlationsof the groupedfirst-occasion
variablewith the disturbancesand errorin the second-occasionvariableare
both zero, then using the reliabilityestimatesfrom the previoussectionwe can
substitutein (7) to obtain the correlationswith the errorin the first occasion
variable,rze,. These are given in Table IV, assumingthe reliabilityestimates
givenin the Resultssection(0.81 and 0.89) arecorrect.
Thus, the correlations,while reasonablyconstantfor reading,are systematically decreasingfor mathematics.We have no good explanationfor this but
possible causes are nonhomogeneityof errorsin the mathematicstest or a
nonzerocorrelationbetweentruescoreand errorsof measurement.
The Use of GroupedTest Score as an Instrumental Variable
Test scores of readingand mathematicsat 7, 11, and 16 years, a score of
GeneralAbility at 11 years (with Verbaland Nonverbalcomponents)and a
CopyingDesign Test are consideredas instrumentalvariables,and the results
are givenin TableV.
Sinceshort-termfluctuationsin attainmentwill be correlatedto some extent
over all attainments,we would expect the argumentsand predictions(given
previously)in relationto teacherratingsto apply to test scores.P1 is satisfied
trivially in the light of the results on the use of a grouped predictor as
TABLEIV
Correlationsof DichotomizedInstrumental Variableand
First-occasionMeasurementErrorfor Different Division Points
belowdivisionpoint
Proportion
Reading
Mathematics
0.2
0.4
0.5
0.6
0.8
0.280
0.390
0.302
0.226
0.329
0.233
0.305
0.128
0.324
0.120
Ecob and Golstein
238
TABLEV
RegressionEstimates of 16-yearAttainment Test on 11-yearAttainment Test in Reading
and Mathematics Using GroupedTest Scores as Instrumental Variables
Instrumentalvariable
At 7 years:
ReadingTest
MathematicsTest
At I1 years:
ReadingTest
MathematicsTest
GeneralAbilityTest:Verbal
:Nonverbal
:Overall
CopyingDesignsTest
At 16years:
ReadingTest
MathematicsTest
Reading
Mathematics
0.920 (0.015)
1.004 (0.020)
0.822 (0.017)
0.850 (0.018)
0.810 (0.009)
0.995 (0.012)
0.942 (0.012)
0.982 (0.014)
0.957 (0.012)
0.989 (0.032)
0.866 (0.014)
0.763 (0.010)
0.826 (0.013)
0.901 (0.014)
0.860 (0.013)
0.933 (0.032)
1.197 (0.013)
1.053 (0.015)
0.949 (0.015)
1.315 (0.017)
Note. Standarderrorsin brackets.
instrumental variable. P2 is satisfied, but P3 is again contradicted in one of
four cases by the behavior of the reading tests at 7 and 11 when used as
instrumental variable for mathematics attainment.
Generally, the behavior of test scores is similar to the teacher ratings and the
standard errors are similar, giving little indication for preference of one set of
variables over the other.
Discussion
Our resultssuggestthat differingcorrelationsof instrumentalvariableswith
measurementerrors account for the observeddifferencesin regressionand
reliabilityestimates, although social class has a negligible correlationwith
measurementerrorbut a nonnegligiblecorrelationwith the errorof prediction.
The estimatedcorrelationcoefficient between true scores on reading and
mathematicstests at 11 and 16 years is respectively-0.96and 0.90. The
estimatedreliabilitiesusing the selectedinstrumentalvariables(teacherratings
of an unrelatedabilityat age 7 as the predictor)are 0.81 and 0.89 for reading
and mathematics, respectively. These compare with the values of 0.82 and 0.94
given in Goldstein (1979) by split-halfitem analysis on a subsampleof 300
cases. While the values for reading are similar, the value obtained by item
analysis for mathematicsis somewhat higher than any obtained for the
instrumentalvariablesused here, althoughthe differenceis of the same order
as the standarderrorof the separateestimates.
Instrumental VariableMethods
239
For reading attainment the estimated standard error of the reliability
estimate obtained by item analysis is 0.030, while the instrumentalvariable
methodgives 0.012. A split-halfestimateusing all availabledata wouldhave a
standarderrorof about 0.007. For mathematicsthe relevantstandarderrors
are 0.020,0.014 and 0.005.
Using a groupingof the predictorvariableitself as an instrumentalvariable
gives estimatesof the reliabilitythat are higherthan any obtainedusing other
variablesas instrumentalvariables,irrespectiveof the numberof groupsused.
Theseestimators,we suggest,are not to be recommended.
Whileinstrumentalvariableestimationhas had a long history(earlypapers
on theoryand applicationin the economicfield includeDurbin, 1954;Madansky, 1959; Reiersol, 1945; Sargan, 1958; Wald 1940),it has not yet become
generally accepted as an estimation method in the social and educational
fields. Sargan(1958), in discussingthe (unknown)correlationof instrumental
variableswith measurementerrorin an economiccontext states:
It is not easy to justify the basic assumptionconcerningthese errors,
namelythat they are independentof the instrumentalvariables.It seems
likely that they will varywith a trendand with the tradecycle. Insofaras
this is true,the methoddiscussedhere will lead to inconsistentestimates
of the coefficients.Nothing can be done about this since presumablyif
anythingwere known about this type of error,better estimatesof the
variablescould be produced.It must be hoped that the estimatesof the
variablesare sufficientlyaccurate,so that systematicerrorsof this kind
are small.(p. 396)
We have arguedthat comparisonsof differentinstrumentalvariables,considered separately,can throw some light on the errorstructurein the data, and
thus lead to better knowledgeof the consistencyof the estimatesproduced.
Furthermore,it is also our view that this approachprovidesa flexibletool for
an empirical study of the various assumptions needed to produce good
estimates.
Finally,fourissuesseemparticularlyworthyof attention:
1. Obtaining estimates of the standard errors of the difference between
different instrumentalvariable estimates (these will be lower than those
obtainedusing the individualstandarderrorsand assumingindependence).
This wouldenablea morecarefulanalysisof the hypothesesof the paper.
2. Obtaininggood estimatesof the standarderrorsof the reliabilityestimates
producedby instrumentalvariablesmethods.
3. Examinationof the use of more than one instrumentalvariablein connection with a single predictorin terms of the efficiency and consistencyof
estimates.
Ecob and Golstein
240
4. The study of differing reliabilities and measurement error variances in
different groups such as social classes, to incorporate these into linear model
estimates.
Acknowledgements
We aregratefulto the NationalChildren'sBureaufor permissionto use the data and
to Dougal Hutchinsonfor his commentson an earlydraftof the paper.The workwas
carriedout largelyon a grant from the National Instituteof Education,Washington
(NIE-G-77-0065).
References
Anderson,E. B. Comparinglatentdistributions.Psychometrika,
1980,45, 121-134.
L.
structural
relation.
R.
84-90.
Bivariate
Brown,
Biometrika,1957,44,
Cronbach,L. J., Gleser,G. C., Nanda, H., & Rajartnan.Thedependability
of behavioural measurements:Theoryof generalisabilityfor scores and profiles. New York: Wiley,
1972.
Davie, R., Butler, N. R., & Goldstein, H. From birth to seven: A reportof the National
Child DevelopmentStudy. London: Longmans, 1972.
Durbin, J. Errors in variables. Review of Institute of International Statistics, 1954, 22,
23-54.
Fogelman,K. Britain'ssixteenyear olds.London:NationalChildren'sBureau,1976.
Goldstein,H. Somemodelsfor analysinglongitudinaldata on educationalattainment.
Journal of the Royal Statistical Society, 1979, A, 142, 402-442.
Goldstein,H. Dimensionality,bias, independenceand measurementscale problemsin
latent trait test score models. British Journal of Mathematical and Statistical Psychology, 1980, 33, 234-246.
Johnston, J. Econometricmethods.New York: McGraw-Hill, 1972.
1971,36,
Jireskog,K. G. Statisticalanalysisof sets of congenerictests. Psychometrika,
109-133.
Kendall,M. G., & Stuart,A. Theadvancedtheoryof statistics(Vol. 2). London:Griffin,
1977.
Layard,M. W. J. Robustlargesampletests for homogeneityof variances.Journalof the
American Statistical Association, 1973, 68, 195-198.
Lord, F. M., & Novick, M. R. Statistical Theories of Mental Test Scores. Reading,
Mass.: Addison-Wesley, 1968.
McDonald, R. The dimensionality of tests and items. British Journal of Mathematical
and Statistical Psychology, 1981, 34, 100-117.
Madansky,A. The fitting of straightlines when both variablesare subject to error.
Journal of the American Statistical Association, 1959, 54, 173-205.
Nair, K. R., & Banerjee.A note of fittingstraightlines if both variablesare subjectto
error. Sankya, 1942, 6, 331-333.
Neyman, J., & Scott, E. L. On certain methods of estimatingthe linear structural
relation. Ann. Math. Statist., 1951, 22, 352-355.
Reiersol, O. Confluence analysis of means of instrumental sets of variables. Arkiv. for
Mathematik, AstronomiOch Fysik, 1945, 32.
Instrumental Variable Methods
241
Sargan,J. D. The estimationof economicrelationshipsusing instrumentalvariables.
Econometrica, 1958, 26, 393-415.
Wald, A. Fitting of straightlines if both variablesare subject to error.Ann. Math.
Statist., 1940, 11, 284-300.
Warren,R. D., White, J. K., & Fuller, W. A. An errors-in-variables
analysis of
managerial role performance. Journal of the American Statistical Association, 1974,
69, 886-893.
Authors
ECOB,RUSSELL,Researchstatisticianin the Researchand Statisticsdivisionof the
Inner LondonEducationAuthority,England.Specializations:Longitudinalstudies,
educationalmeasurement,
effectsof schooling.
GOLDSTEIN, HARVEY, Professor of statistical methods in the departmentof
Mathematics,Statisticsand Computingat the Universityof London Institute of
Education.Specializations:Longitudinalstudies,test score theory,educationalmeasurement.
Download