OTIC FILE COPYJ RR-88-54-ONR -- 3

AD-A200 179 ---

RANDOMIZATION-BASED INFERENCES ABOUT -

LATENT VARIABLES FROM COMPLEX SAMPLES

Robert J. Mislevy

Nov o 8 1988 • .

This research was sponsored in part by the

Office for Educational Research and Improvement
Center for Education Statistics, under
Grant No. NIE-G-83-001 1; and the "
Cognitive Science Program
Cognitive and Neural Sciences Division
Office of Naval Research, under
Contract No. N00014-88-K-0304
R&T 4421552

Robert J. Mislevy, Principal Investigator

# Educational Testing Service

Princeton, New Jersey

September 1988 •

Reproduction in whole or in part is permitted

for any purpose of the United States Government.

Approved for public release; distribution unlimited. ,

807
 ie

![ n ol: s si fie d .

SECURITY CLASSIF'CA7ION OF THIS PG-E '

Form Approved
REPORT DOCUMENTATION PAGE OMB No 0704 0188

'a REPORT SECuRITY CLASSIF.CATION lb RESTRICTIVE MARKINGS

Inc lass if led
2a SECURITY CLASSFICAT ON AUTHORITY 3 DISTRIBUTION AVA!LABILIJY OF REPORT
Approved for public release;
"-2t DECLASSIFICATION DOWNGRADING SCHEDULE dist ribution unlimited.

* 4 PERFORMING ORGANIZATION REPORT NUMBER(S) 5 MONITORING ORGANIZATION REPORT NuMBER(S)

-RR-88-54-IjNR

6a NAME OF PERFORMING ORGANIZATION 6b OFFICE SYMBOL 7a NAME OF MONITORING ORGANIZATION Cognitive

* ducati.)na1 lesting Service (if applicable) Science Pro gram Office of Naval Research
I (Code 1142CS), 800 North Quincy Street
6c ADDRESS City State, and ZIP Code) 7b ADDRESS (City. State, and ZIP Code)

I'rinceton, N.i 08541 Arlington, VA 22217-5000

8a NAME OF F1JNDNG SPONSORING 8b OFFICE SYMBOL 9 PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER

ORGANIZATiON (If applicable) N00014-88-K-0304

Bc ADDRESS (City, State. and ZIP Code) 10 SOURCE OF FUNDING NUMBERS

PROGRAM PROJECT TASK WORK UNIT
ELEMENT NO NO NO ACCESSION NO
61153N RR04204 RR04204-01R&'1'4421552

11 TITLE (Include Security Classification)

ci:Jomization-Based Inferences about Latent Variables from Complex Samples
I:s i f iec 1
L'.lc
12 PERSONAL AuTHOR(S)
Robert J. Mislevv
13b TIME COVERED 14 DATE OF REPORT (Year, Month Day) 15 PAGE COUNT
13a TYPE OF REPORT
FROM TO September 1988 64
[eiin iiII
T
16 SLPP-EMEN ARY NOTATION

17 COSATI CODES 18 SUBJECT TERMS (Continue on reverse if necessary and identity by block number)
FIELD GROUP SUB-GROUP Complex samples, item response theory, latent structure,
Jj i0 missing data, multiple imputation, National Assessment of
Educational Progress, sample surveys
'9 ABSTRACT )Continue on reverse if necessary and identify by block number)

Standard procedures for drawing inferences .from complex sampl-ks do not

apply when the variable of interest 9 cannot 4 observed directly, but-must
be inferred from the values of secondary random variables that depend on 9
stochastically. Examples are examinee profic'iency variables in item response
theory models and class memberships in latehft class models. This paper uses
Rubin's "multiple imputation"-approach toipproximate sample statistics that
would have been obtained, had 0 been obse vable. Associated variance
estimates account for uncertainty due to-both the sampling of respondents
from the population and the latency of i. The approach is illustrated with
artificial examples and with data from the 1984 National Assessment for
'
Educational Progress reading survey. ( .

AVAILABILITY OF ABSTRACT 21 ABSTRACT SECURITY CLASSIFICATION

20 DISTR;BuTION
"J NCLASSIFIED IJN[IMITED E[ SAME AS RPT E DTIC USERS Unclassified
22b TELEPHONE (Include Area Code) 22c OFfICE1 SYMBOi
22a NAME OF RESPONS.BLE INDIVIDUAL
Dr. Charles E. Davis 202-696-4046 ONR 1142CS

DD Form 1473, JUN 86 Previous editions are obsolete SECURITY CLASSIFICATION OF THIS PAGE

S/N 0102-LF-014-6603 Unclassified

 Randomization-Based Inferences about Latent

Variables from Complex Samples

Robert J. Mislevy

Educational Testing Service

September 1988

CPY

lt4 SPfEC1 --

This research was supported by Grant No. NIE-G-83-O011 of the

Office for Educational Research and Improvement, Center for
Education Statistics, and Contract No. N00014-88-K-0304, R&T
4421552 from the Cognitive Sciences Program, Cognitive and Neural
Sciences Division, Office of Naval Research. It does not
necessarily reflect the views of either agency. I am grateful to
R. Darrell Bock for calling my attention to the applicability of
Rubin's multiple imputations to the assessment setting; to Albert
Beaton and Eugene Johnson for enlightening discussions on the
topic; and to Henry Braun, Ben King, Debra Kline, Paul Rosenbaum,
*Don Rubin, John Tukey, Kentaro Yamamoto, and Rebecca Zwick for
comments on earlier drafts. Example 3 is based on the analysis of
the 1984 National Assessment for Educational Progress reading
survey, carried out at Educational Testing Service through the
tireless efforts of too many people to mention by name, under the
direction of Albert Beaton, Director of NAEP Data Analyses.
E
rto

Copyright Q 1988. Educational Testing Service. All rights reserved.

IL

Randomization-based Inferences about Latent

Variables from Complex Samples

Abstract

Standard procedures for drawing inferences from complex

samples do not apply when the variable of interest 9 cannot be

observed directly, but must be inferred from the values of

* secondary random variables that depend on 0 stochastically.

Examples are examinee proficiency variables in item response

theory models and class memberships in latent class models. This

paper uses Rubin's "multiple imputation" approach to approximate

sample statistics that would have been obtained, had 9 been

observable. Associated variance estimates account for uncertainty

due to both the sampling of respondents from the population and

the latency of 9. The approach is illustrated with artificial

examples and with data from the 1984 National Assessment for

Educational Progress reading survey.

Key words: Complex samples, item response theory, latent

* structure, missing data, multiple imputation,

National Assessment of Educational Progress, sample

surveys.
 Randomization-Based Inference

Introduction

Latent-variable models are used in the social sciences to

provide parsimonious explanations of associations among observed

variables in terms of theoretical constructs. Practical benefits

can accrue as well, as when examinees who have been presented

different test items are compared using item response theory (IRT)

psychometric models, or when consumer satisfaction levels are

tracked over time with different survey questions by means of a

latent class model.

0 This paper addresses the problem of estimating the

distributions of latent variables in finite populations, when the

data are obtained in complex sampling designs. The solution it

offers is to apply Rubin's (1987) "multiple imputations"

procedures for missing data in sample surveys. This approach

provides consistent estimates of population characteristics,

* supports statements of precision that account for both the

sampling of subjects and the latency of variables, and produces

filled-in pseudo datasets that are easy for secondary researchers

* to analyze correctly.

The following sections briefly review randomization-based

inference about manifest variables from complex samples, and

* multiple imputation for nonrespons.. The next sections apply

these ideas to latent-variable measurement models and illustrate

them with an example using classical test theory. Computing

 Randomization-Based Inference

approximations are then discussed. The paper concludes by

sketching the implementation, the results, and the lessons learned

in applying the procedures to the 1984 reading survey of the

National Assessment of Educational Progress (NAEP).

Drawing Inferences from Complex Samples

Standard analyses of sample survey data (e.g., Cochran, 1977)

provide estimates of population characteristics and associated

sampling variances when the values of survey variables are

observed from each respondent in the realized sample. This

section gives background and notation for the analysis of survey

data, first when all responses are in fact observed, then when

some are missing.

Randomization-Based Inference with Complete Data

Consider a population of N identifiable units, indexed by the

subscript i. Associated with each unit are two possibly vector-

valued variables Y and Z. The values of the design variables, Z,

are known for all units before observations are made, but the

values of the survey variables, Y, are not. Let (X,Z) denote the

population matrix of values. Interest lies in the value of a

function S - S(Y,Z) of the population values; examples include a

population total for an element of Y, a subpopulation mean for an

element of Y given specified values of elements of Z, and the

linear regression coefficients of some elements of Y on others.

 Randomization-Based Inference

Any of these functions could be calculated directly if values of Y

and Z were observed for all units in the population.

There are 2N possible subsets of units, and a sample design

assigns a probability to each. A simple random sample of size

100, for example, assigns equal probability to all subsets of 100

units, and zero to all other subsets. A sample design yields a

"complex sample" when it exhibits one or more of the following

features: unequal probabilities of selection for different units;

stratification, which ensures prespecified rates of representation

in the sample according to values of Z; or clustering, which uses

values of Z to link selection probabilities of units when their

joint occurrence facilitates gathering the data (e.g., it is

easier to interview two respondents in the same town than two in

different towns).

Randomization-based inference about S is based on the

distribution of a statistic s - s(y) calculated using X, the Y

values of a subset of units selected in accordance with a

prespecified sample design. In practical work, sample designs are

usually constructed so as to yield a nearly unbiased statistic s

and an estimate of its variance in the form of another statistic U

- U(X). Inferences are then drawn using the normal approximation

(s - S) - N(0,U)
4

Randomization-Based Inference

Under the randomization approach to inference in sample

surveys, population values Y are taken as fixed unknown

quantities, and the notion of randomness enters only in the

selection of a sample in accordance with the sample design. Under

the model-based approach, in contrast, Y is viewed as a realized

sample from a hypothetical "superpopulation" under which variables

are distributed in accordance with some model p(ylz) (Cassel,

Sarndal, and Wretman, 1977). The randomization approach dominates

current practice, and will be used in the sequel to deal with

uncertainty due to sampling subjects. It will be seen that

superpopulation concepts are required nonetheless to handle

missingness and latency.

Randomization-Based Inference with Incomplete Data

In practice, values of one or more survey variables will not

be observed from some subjects in the realized sample, for

reasons that may or may not be related to the values that would

have been observed. For any respondent, let the partitioning

y=(YmisYobs) indicate the elements of the survey variables that

were missing and observed. Much progress has been made extending

inferential procedures to survey data with missing responses when

those responses are missing at random (MAR); that is, the

probability that the elements in y are missing may depend on

I
the values of yobs and z, but beyond that, not on the value of

ymis (Little and Rubin, 1987; Rubin, 1976).

I

Randomization-Based Inference

Let (-mis'yobs) be the matrices of missing and observed

survey variables in a realized sample. The values of lobs- -which

may comprise different elements of y for different subjects--are

now known, but the values of Ymis are not. It is not possible to

calculate s directly, but it may be possible to calculate its

conditional expectation:

E[s() 'Yobs L(mis',yobs) p(XmislobsE) dmis (i)

I

The predictive density p( mis1 obsz) expresses the extent of

knowledge about what the missing responses might have been, given

the observed responses and the survey variables. If MAR holds,

the predictive distribution of a missing element for subject i--

say yi,mis-- is approximated by the responses to that element from

subjects who did respond to it and have the same z values as

subject i and the same responses on the elements in y i,obs* In

large surveys with few missing values, one can use these empirical

distributions directly. The Census Bureau's "hot deck" imputation

procedure (see Ford, 1983), for example, assumes that predictive

densities are independent over respondents--i.e.,

p(misLobsz) = p(yi,mislyi,obszi) --
i

and calculates s with each person's observed responses and, for

those he or she is missing, draws from the actual responses of

suitably similar respondents.

I

Randomization-Based Inference

Alternatively, one can posit a functional form for

p(YmisYobs, z), such as a regression model with parameters 6. It

is usually reasonable to assume independence over respondents, but

now conditional on the unknown parameters 6; that is,

P(YmisIYobsZP) = T p(yi,mislyi,obszi,) (2)

In this case, the appropriate predictive distribution is obtained

after marginalizing with respect to P:

)
P(Xmislyobs' ) = f P(Xmis'Xobs'Z'O) p('obs di , (3)

where p(Olyobs ) is the posterior distribution of / after Xobs has

been observed.

Multiple Imputation for Missing Responscs

As with the hot deck, one can obtain a rough numerical

approximation of (1)--an unbiased estimate of the expectation of

s--by filling in each missing response with a random draw from its

4 predictive density, then calculating s as if these imputations

were the true response values. An analogous approximation of U

using these single imputations underestimates the variability of

the resulting estimate of S, however. It accounts for uncertainty

due to sampling subjects, but not uncertainty due to imputing

values for missing responses. To remedy this deficiency, Rubin

suggests multiple imputations. When the statistic s is scalar and

 Randomization-Based Inference

the predictive distribution for ymis is model-based, one proceeds

as follows:

1. Determine the posterior distribution p(Olob

2. Produce M distinct "completed" datasets Y(m), m=l,...,M.

Each looks like a complete dataset; it has the values Xobs

for the responses that were observed, and, for those that

were not, a draw from (3). Two steps complete x(m)"

a. Draw a value P(m) from p(Plyobs ) .

b. For each respondent with one or more missing responses,

draw a value from p(y i,mislYi,obs'i,3=0(m)). Taken

togethe-, the resulting imputation for y imis and the

observed value of Yi,obs constitute Yi(m)' the

"completed" response of subject i.

3. Using each completed dataset, calculate s(m)S(Y (m) ) and

U () Y(m))

4. The final estimate of S is the average of the M estimates

from the completed datasets:

M
sM = Z s (m)/ M (4)

5. The estimated sampling variance of sM as an estimate of S,

say V M , is the sum of two components:

H

Randomization-Based Inference

VM U M + (l+M - ) BM , (5)

where

M
UM Z U / M
m=l(i)

quantifies uncertainty due to sampling subjects, and

M
BM - mH
E l(in)
(s - SM) 2 (-i
(M-1)
M= 1

the variance among the estimates of s from the M completed

datasets, quantifies uncertainty due to missing responses

from the sampled subjects.

6. If inferences about S would have been based on (s-S) - N(O,U)

had all responses been observed, inferences are now based on

S MS) - t (O,VM , (6)

a t-distribution, with degrees of freedom given by

= (M-1) (l+rM
-1 2
,

where rM is the proportional increase in variance due to

* missingness:

rM - (1+M ) BH / UM
 Randomization-Based Inference

When BM is small relative to UM, v is large and the normal

approximation uo the t-distribution suffices.

For k-dimensional s, such as the vector of coefficients in a

multiple regression analysis, each U(m ) and UM are covariance

matrices, and BM is an average of squares and cross products

rather than simply an average of squares. Then the Quantity

(S-SM) V (S-S ) (7)

is approximately F-distributed with k and v degrees of freedom,

with v defined as above but with a matrix generalization of rM:

-1
rM = (+M ) Trace(BMUM -1)/k

By the same reasoning as used for the normal approximation for

scalar s, a chi-square distribution on k degrees of freedom often

suffices.

Example 1: A Numerical Illustration

* Consider a large population in which the scalar y is

distributed N(iI), and it is desired to estimate p from an SRS of

size 12. If y values for all 12 units in the realized sample were

* observed, inferences about A would be based on y, the sample mean,

and U, the squared standard error of the mean (SEM2). U is I/N

since the population variance is known to be one. In particular,

 Randomization-Based Inference

10

(y-g) - N(0,1/12). Suppose, though, that the values for the last

two sampled units are missing (at random). The first ten are

observed to take the values shown below:

-.797, -.176, 1.419, .029, -1.107,

1.794, -1.619, 1.104, .418, .161.

These observations comprise Xobs' while ym is (YllYl2).

Using multiple imputations to estimate the population mean is

accomplished as follows.

Since we are assuming MAR and there are no collateral or

design variables, the distributions of yll and Y 1 2 are simply

N(jI) too. What we know about M from the first ten observations

is conveyed by their mean and the SEM2; with an indifference

prior, p(Alyobs ) is N(.123,.100).

A completed dataset consists of the ten observed y values and

an imputation for (yllyl2 ). For the sake of illustration, five

completed datasets will be constructed. The procedure for each

consists of two steps:

a. A value u() is drawn from p or N(.123,.I).

b. Imputations Yll W)and Y12(m) are drawn from N(p(m),).

A table of random normal deviates gave the following results:

 Randomization-Based Inference

11

m (m) Yll(m) Yl2(m)

1 .306 -.095 -.801

2 .238 -.599 .644

3 .033 -.029 .510

4 .089 .386 -1.248

5 .001 .205 2.106

From each completed dataset, an estimate ytm) is calculated,

the mean of ten observed y values and two imputations. The

results are .028, .106, .143, .031, and .295. For each m,

U()/12, an SEM appropriate for twelve observations from a

normal population with a variance known to be one.

The estimate of u based on the imputed values y is the

M
average of the five completed-data estimates, or .121. Not

surprisingly, this is close to the estimate .123 based on Xobs'

since that is its expected value. Indeed, yM would converge to

.123 as M increased. The estimated sampling variance of yM is

obtained by (5) as

Z (m)/5 + (1+51) Z (y(m)-yM)2/4

1/12 + 1.2 x .012

.098.

This value is very close to 1/10, the SEM 2 that corresponds to a

 Randomization-Based Inference

12

sample of size 10--as it should, since this is what the data

actually are.

Randomization-based inferences about p using the multiple

imputations are based on

(.121 - p) - t1 8 4 (0,.098)

a t-distribution with degrees of freedom obtained by (6) with rM,

the proportional increase in variance due to the missingness of

Yll and Y 1 2 , equal to .173.

Latent Variables and Sample Surveys

This section provides notation and background for latent-

variable models, and shows how they can be handled in complex

samples with multiple imputations.

Latent-Variable Models

Latent variables in educational and psychological measurement

models account for regularities in observable data; for example,

examinees' tendencies to give correct responses to test items, or

respondents' inclinations toward liberal responses on social

questions. The probability that a subject with latent parameter 9

will make the response x W to Item j is modeled as a function of

0 and a possibly vector-valued item parameter Pj, as p(x(j), j

The assumption of local or conditional independence posits that

the latent variable accounts for associations among responses to

 Randomization-Based Inference

13

various items in a specified domain; i.e., if x - (x .. X(n )

is a vector of responses to n items, then

n
p(xiOA = H p(x (j) 0j)' , (8)

where 6 = (1 1.... on). Moreover, the latent variable is typically

assumed to account for associations between response variables and

collateral subject variables such as demographic or educational

standing. Denoting collateral variables by y and z to anticipate

developments below, the extension of local independence posits

wiP(xio,p,y,z) - p(xJO',l) (9)

When such a model is found to provide an adequate fit to

data, observing the responses to any subset of items induces a

likelihood function for 0 through (8), and it becomes possible to

draw inferences about individual values of 9 or about their

4 distributions in populations even though different subjects

respond to different items. This capability is particularly

attractive in educational measurement, as when an IRT model can be

used to customize tests to examinees adaptively, or to update the

item pools in educational assessments over time.

If the focus is on measuring individuals for placement or

selection decisions, enough items can be administered to each to

make the likelihood function for his or her 0 peak sharply. A

satisfactorily precise point estimate of each 0, such as the MLE 9

4q or the Bayes mean estimate B - E(O x), can then be obtained. If

__ _ _ _ __ - _
 Randomization-Based Inference

14

the focus is on the parameters a of a population distribution of

0, however, these locally optimal point estimates can be decidedly

nonoptimal. For a fixed test of fixed length, their distributions

do not converge to the true distribution of 0 as examinee samples

increase. It becomes necessary to estimate a directly, bypassing

the intermediary stage of calculating point estimates for

individuals.

Suppose the data matrix x consisted of response vectors

(xI .... xN) of an SRS of respondents. The starting point for the

direct estimation of the parameters a of the density function

p(Ola) in this context would be the marginal probability:

N
p(xI,fl) - H f P(xi1O,P) p(Ola) dO (10)
i-i

A number of recent papers in the statistical and psychometric

literature show how to obtain Bayesian or maximum likelihood

estimates of a and/or 6 using (10) (e.g., Andersen and Madsen,

1977; Bock and Aitkin, 1981; Dempster, Laird, and Rubin, 1977;

Laird, 1978; Mislevy, 1984, 1985; Sanathanan and Blumenthal, 1978;

and Rigdon and Tsutakawa, 1983). As they are presented, however,

these methods are poorly suited to general analyses of survey data

involving latent variables. As well as being limited to SRS data,

*they require advanced statistical concepts and iterative computing

methodologies that render them inaccessible to the typical

secondary analyst.

I1 ...-
 Randomization-Based Inference

15

Latent Variables as Missing Responses

A key insight for dealing with latent variables in sample

surveys is to view them as survey varf ibles whose responses are

missing for every respondent (e.g., Dempster, Laird, and Rubin,

1977, on factor analysis models). Their missingness satisfies MAR

because they are missing regardless of their values, and as such

are amenable to the (relatively) simple procedures for MAR missing

data described above. In essence, knowledge about subjects'

latent variables 0 will be represented in the form of predictive

distributions conditional on what can be observed- -background

survey variables y, design variables z, and item responses x whose

distributions are assumed to be governed by 0.

Suppose the object of inference is the scalar S - S(,,,)

some function of the population values of latent variables, item

response variables, background survey variables, or design

variables of all units. Suppose further that a sample design has

been specified. Three assumptions are central to drawing

randomization-based inferences about S. The first is that one

would know how to proceed if values of 0 were observed rather than

latent:

Assumption 1. If values of 0 could be observed from sampled

respondents, along with values of x and y, a sample statistic s=

s(,xZ) and an associated variance estimator U - U(O,x,Z) would

be available for randomization-based inference about S, via

(s-S) - N(O,U).
 Randomization-Based Inference

16

Inferences about S cannot be based on a direct calculation of

s because all values of 0 are missing. As before, however, one

can base inferences on its conditional expectation given what is

known--the sample values x and X and the population values Z--by

adapting (1) as follows:

E(slx,X) = f s(O,x,y) p(I d.

,lx,Z) (11)

The second and third assumptions are needed to define the

predictive distribution for 0 that appears in (11), namely

p(Ojx,q,Z). These assumptions are embodied in the latent variable

model and a superpopulation structure for distributions of 0 given

background survey variables and design variables.

Assumption 2. Item responses x are governed by the latent

variable 0 through a model of known functional form, p(xlOf),

characterized by possibly unknown parameters 6 and satisfying the

local independence properties (8) and (9). Independence is

assumed over subjects, so that

P(XI6,fi,,Z) = P(XIO,P)

N
= H P(Xii,p) (12)
i

Assumption 3. The distribution of latent variables given

collateral survey variables y and design variables z follows a

 Randomization-Based Inference

17

known functional form, p(Oly,z,a), characterized by possibly

unknown parameters a. Independence is assumed over subjects, so

that

N
p(O[y,Z,a) = H p(ilYi,zi,a) (13)
i

It is important to note that these distributions are

conditioned on the design variables z employed in the sampling

design. While the sample design for selecting units from the

existing population may be complex, sampling this population from

the hypothetical superpopulation is SRS given z. This is the

essence of the model-based approach to sampling-survey inference.

It is used in this presentation not to handle uncertainty due to

sampling examinees, but to build conditional distributions for

latent variables, since it opens the door to the aforementioned

methods of estimating the parameters of latent-variable

distributions from SRS data.

To see how Assumptions 2 and 3 lead to p(OIx,y,Z), first

consider some relationships conditional on a and P. Using Bayes

theorem, then (12) and (13), gives

p(OIx,X,Z,a,#) = Ka p(XIO',X,Z,afi) p(yI,Z,afi)

 Randomization-Based Inference

18

N
H Kip P(xilOi,fl) p(ilyi,zi,a) , (14)
i

where the normalizing constants K = P(xi yi,zi,a,,) depend on a

and f but not on 6..

1 Given a and 6, then, the predictive

distribution for the latent variable 0.

1
of Subject i, or

p( ilxiYizia,,), would be obtained by normalizing the product

of (i) the likelihood function induced for 0 by x.1 via the latent-

variable model (8), and (ii) the conditional distribution for 0

implied by his or her background and design variables yi and z.

Multiple Imputations for Latent Variables

The preceding paragraphs give the framework for

randomization-based inference with latent variables in sample

surveys. To operationalize the approach with multiple imputations

requires specializing the procedure outlined above as follows:

1. Obtain the posterior distribution of the parameters P of the

latent-variable model and a of the conditional distributions

of 0, namely p(a,fIx,q,Z), by the methods discussed in

connection with (10)--for example, a large sample normal

approximation based on the MLE (a,#) and asymptotic

covariance matrix Za.

th
2. Produce M "completed" datasets (9(m),Xy). For the mth
 Randomization-Based Inference

19

a. Draw a value (af) from P( ,plx,yZ). (If a and

have been estimated very precisely, it may be acceptable

to use (arB) for each completed dataset in what is

commonly known as an empirical Bayes approximation.

This expedient introduces a tendency to underestimate

the uncertainty associated with final estimates of S.)

b. For each respondent, draw a value from the predictive

distribution p(Oxiyizi,(Czf)-(at)()). Taken

together, the resulting imputation for 0.

i
and his or her

observed values of x,, y, and z. constitute the

"completed" response of Subject i.

3. Using each completed dataset, calculate s(m)=s(0(m),xy) and

U m=U( (m),x,y).

4. The final estimate of S is the average of the M estimates

from the completed datasets, or sM = Z s(M ) / M

5. The estimated sampling variance of sM as an estimate of S,

namely VM, is the sum of two components:

VM - UM + (I+M " 1 ) BM ,

where U M and B M are defined as for (5), to quantify

uncertainty due to sampling subjects and the latency of 0,

respectively.
 Randomization-Based Inference

20

6. If inferences about S would have been based on (s-S) - N(O,U)

had all responses been observed, inferences are now based on

(sM-S) - tI(O,VM), a t-distribution with degrees of freedom v

given in (6).

These procedures apply to vector-valued latent variables 9 as

well as scalars. Extensions to vector-valued S are as discussed

4previously.

Steps 1 and 2 above produce M completed datasets that can be

used to draw inferences about any number of sample statistics by

applying Steps 3-6 repeatedly. An attractive feature of the

approach is that the sophisticated methodologies and heavy

computation are isolated in Steps 1 and 2, which can be carried

out just once--probably by the institution held responsible for

primary data analysis, where the necessary expertise and resources

are more likely to be available. The completed datasets are then

provided to secondary researchers, who need only apply standard

routines for complete data M times and combine the results in

simple ways.

Example 2: Multiple Imputation under Classical Test Theory

This example lays out imputation procedures for when the

latent-variable model is the classical true-score test model with

normal errors, and there are two collateral survey variables. In

order to focus on the construction and the nature of imputations,

it is assumed that a large SRS will be drawn from a multivariate

 Randomization-Based Inference

21

normal population with known parameters. There are four variables

for each subject:

0, the latent variable, is examinee "true score;"

x is examinee observed score; and

y, and Y 2 are collateral examinee variables.

Consistent with true-score test theory with normal errors,

assume that x = 6 + e, where the residual or error term e is

2
distributed N(O,a ) independently of 0, yI, and Y2. The latent
e Y
variable model is thus

p(xJO) - N(O, 2 ) (14)

e

Assume further that (6,yly 2) follows a standard multivariate

normal distribution in the population, so that jointly,

(x,6,yly 2) - MVN(O,Z) with

l+u2 (sym)
e
1 1

r 1r 911

r62 r0 2 r1 2 1

Z will refer to the covariance matrix for y=(yly 2 ).

-y

The conditional distribution of 0 given y is obtained as

p(Oly) - 2 (15)

where
 Randomization-Based Inference

22

'Y = I112 Yl + 2 11 Y2 E(O1y), and

2 2
°Oly

where R 2 is the proportion of variance of 6 accounted for by y.

In this example, f 1 1 2 =(r0 1 -r02 r 1 2 )/(l-r 1 2 2) and P211-

2 2
(r 2 -r 1 r1 2 )/(l-r . It follows that p(xly) - N(fi'y,a ), with
2 a =a 2 +u. 2
axly a 6 ly e
In addition to the multiple regression coefficients there are

simple regression coefficients for 0 or x on yl or Y 2 alone; e.g.,

1
E(O1Yl) = E(xlyl) = Y =-(112 + r 1 2 f2l1 ) Yl

In this example, 8lr 01 and P 2 -r02 .

Define the "conditional rcliability" p of x as a measure of

0 given y as

2 / 2 a2
PC = ly / (a 1y + Oe)

2 2
Note that Op l and pCaxly - a~y.

An imputation for a sampled subject will be drawn from a

predictive distribution of the form p(6lx,yl which is

proportional to p(xJ6) p(Oly). The first factor in this product

is the likelihood, which in this example is N(x,a ); the MLE is in

e
fact simply x. The second factor is the conditional density of 6

given y, which is N(P'y,a 2). By Kelley's (1947) formula,

0
0 .

Randomization-Based Inference

23

p(Olx,y) - N(O,a 2
OIXY

where

i E(01x,y) Pcx + (1-p C ) f'y

and
2 2 22
2 x Var(Olx,y) - (1-P ) a (I-P )(1-R )
OIxy c Oly C

An imputation 8 - 8(x,y) is thus constructed as

9 - 0 + f

where f is drawn at random from N(O,a 2 ).

For a given individual, an imputation is not an optimal point

estimate of 0. It is neither unbiased nor efficient, as is the

MLE O-x; nor does it minimize mean square error over the

population, as does 0. But it can be shown that the distribution

of (6,y) is multivariate normal with the same mean vector and

* covariance matrix as that of (O,y). For the population mean, for

example,

E(O) - E[p cx + (i-P C) f'y +f]

- PcE(X) + (1-p f' E(y) + E(f)

- Pc 0 + (1-p) ' 0 + 0

0 ...
 Randomization-Based Inference

24

-E(O)

For the covariance between 6 andy,

Gov(9,yl) - E([pcx + (1-p) P'y +f] y1

- E[pcxy1 ] + E[(l-p C)(p 11 2 y1 +p2 lly 2) y11 + E[fyl]

- p r01 + (l-pc)L01, 2 E(y1 2 + 21 1 E (yly2 )] + 0

-p r6 + (l-P)+p, r12

p r81 + (l-p) r81

- ov(6,y1 )

By similar calculation,

Var(6) - 1 - Var(6)

E(Ojy) - '6'y - E(Ojy)

Var(Oly) - a - VrOy

E(Oly 1 ) - 161 '1 - E(O1y 1 ) ,and

*Var(P1y,) - 1 r01 - Var(Plyl)

Given that the mean of 0 is an unbiased estimate of the mean

of 6, how much does uncertainty increase when x and Xare observed

!2

Randomization-Based Inference

i2 25
IIX
rather than 0 itself? For an SRS of size N, the SEM 2for N

observations of 6 is U-l/N. For large M, the variance among

estimates of the population mean from the completed datasets is

the variance of the average of N realizations of f, or c2 I/N.

The total variance for an estimate of the mean based on many

imputations is the sum of these components, representing an

inflation from 1/N of (1-p )(1-R ) percent--the proportion of

variance in 6 unexplained by x and y.

It has been shown that treating Os as Os, one obtains the

correct expectations for estimates of population attributes such

as percentile points, conditional distributions, and proportions-

of-variance-accounted-for. In contrast, treating either Os or s

as Os one obtains estimates for some attributes that have the

cor-ect expectations for some attributes, but not for others. For

estimating the mean of 8, all three turn out to have the correct

expectation of zero. For the variance, the expectation using

imputations is the correct value of one, but the expected

* -2 2
variances of 0 and 0 are +ae and 1-a respectively.

Table 1 gives the expectations of estimates of some

population attributes that result when various point estimates of

o are treated as if they were 0. Imputations 0 appear there, as

well as the MLE 0 (-x). The Bayes estimate referred to above as

poplaio atrbtstaxeutwevrospitetmtso
is denoted x to indicate that it is conditional on both x a y;

another Bayes estimate often used in practice, denoted in the

table as Ox, ignores y. The variance of is known to be less

 Randomization-Based Inference

26

than that of 0, so it is sometimes suggested that 6x values be

inflated by the factor needed to bring their variance up to that

of 0. The resulting rescaled Bayes estimate is denoted by 0x (r)

While this rescaling corrects the variance of the population as a

whole, it does not completely remove the biases associated with

attributes of conditional distributions involving y.

Table 1 about here

Note that the distortions in secondary analyses of all these

point estimates depend on test reliability. Reliability, and

therefore the magnitudes of biases, will vary if test lengths

differ over time or across respondent groups. Tables 2 and 3

illustrate this point by giving numerical values for the

expressions in Table 1 that are obtained from a test with pT=* 50

and a test with pT=' 9 1 , in both cases with r0 1 - r0 2 - r 1 2 - .50.

The first test has a reliability that might be expected with ten

items on a particular topic that appear on an educational

assessment instrument. The second has a reliability more like

q1-- that of a 60-item achievement test. The biases that occur when

using any of the "optimal" point estimates instead of the

imputations, are readily apparent for the short test but are less

serious for the long test.

 Randomization-Based Inference

27

Tables 2 and 3 here

Along with moments of population and subpopulation

distributions, cumulative probabilities are sometimes required--

e.g., P(B I) or P(B 1yl=-l). Statistics of this type are

important in NAEP, where selected points along the proficiency are

14 anchored by behavioral descriptions, and the proportions of

populations and subpopulations above these points are tracked over

time (NAEP, 1985). P(B I), in this example, is the proportion of

the total population above one standard deviation in the standard

normal distribution, or .1587. P(BLllYl--l) is the proportion of

the subpopulation defined by yl--i with B-values higher than one

standard deviation above the total population mean, which, in this

example, is .0418, the proportion of N(-.5,J.75) above 1. Table 4

gives the expectations of these values that obtain when point

estimates of B are treated as B. Even though the regression

estimates for B1 from B and 0 are unbiased, the estimated

population proportions above the cut point are distorted. Again

4the distortion is less serious with the long test.

Insert Table 4 here

-- - - -- - -
 Randomization-Based Inference

28

Computing Approximations and Secondary Biases

Obtaining consistent estimates of population attributes with

multiple imputations requires drawing imputations from consistent

estimates of the correct predictive distributions p(Ox,y).

Assuming the latent variable model p(xI0) is specified correctly,

it is possible to obtain detailed nonparametric approximations of

p(0ly), and therefore of p(81x,y), when the dimensionalities of 0

and y are low--say less than five latent variables and five

collateral variables (Mislevy, 1984). When the dimensionalities

of 9 and y are high, however, as in NAEP with its hundreds of

background and attitude items, simplifications and computing

approximations cannot be avoided. This section lays out a general

framework for the problems entailed by using simplified

approximations of p(Oly), derives some explicit results for the

true-score example introduced above, and offers guidelines for

practical applications.
E

The Nature of the Problem

The imputation-based estimate sM(x,y) approximates the

4 expectation of s(0,x,y) defined in (11) by evaluating s with draws

9 from

p(OIx,Y) c

If p(xIO,) p(lIy,E,r) p(a,flx,X,E) do dfl (16)

 Randomization-Based Inference

29

As noted above, however, procedures for characterizing p(9I_,z,a)

are neither numerous nor easily applied. It becomes necessary to

approximate these conditional distributions by some tractable

form, such as multivariate normal with a common residual variance

(Mislevy, 1985). Rather than conditioning on all elements of y

and z and their interactions, which may number into the millions,

approximations based on perhaps fewer than a hundred effects must

* suffice. The correct conditional distribution is replaced by the

* *
computing approximation p (Bjy,z,a), which, when combined with

the latent variable model as in (16), yields the computing

approximation from which imputations 8 are drawn:

P (O{5,X,E)

ff p(X {e,) p (OIX,E,a) p(a ,fjx,X,E) da d*f . (17)

While the expectation of sM based on imputations 0 has the

correct value s, the expectation of sM based on imputations 0 may

S
not. Its expectation is

E(sM) Jf s(O,x,y) p (O x,y,z) dO (18)

For a fixed observed sample (x,y), the bias in estimating s caused

* by using p (B1y,z,a) rather than p(O{y,z,a) is thus

Bias = E(sM Ix,) - E(slx,y)

* - ~ s(O,x,y) [p*(OIx,X,z)-p(OIx,X,z)] dO (19)

 Randomization-Based Inference

30

Rubin raises the possibility of biases in secondary analyses

when he mentions that the imputer's model may differ from the

analyst's (Rubin, 1987, pp. 107, 109-112, 151). His analyses

suggest that biases for population means and variances may be mild

when filling in a modest number of missing responses in standard

surveys, particularly if the imputer's model is more inclusive

than the analyst's. It will become apparent that thik conclusion

need not hold in the context of latent variables models, however,

especially when practical constraints force the imputer to use a

less inclusive model than the analyst.

Example 2 (continued)

Consider again the multinormal example introduced above, with

its true-score latent variable model xIB - N(,a 2 ) and population

e
model Oly - N(#'y,a 2y). Suppose now that the imputer conditions

on Yl but not Y 2 when building the model from which imputations

are drawn. That is, the correct predictive distribution for

imputations is

,4 2
W y ,a xy
p(Oix,y) = N[px + (-P

but the imputer draws from

* * * 2
y 201xy ]
p (Oix ,y) - N [p x + (l-P ) I ,

I,' c 1y
 Randomization-Based Inference

31

where p is the conditional reliability of x given yl but not

or

* 2 2 2 2 2 2
PC) = (l(ral) / (l-r6 1 + ae)

and

2 * 2 * 2
a0 xy = Var(O1x,y) (l-p) a2 (-P )(l-r2 1 )
I

An imputation for Subject i now takes the form

0i = PcXi + (l-pc)plYil + g
I 1 1

where g is a draw from N(O,x2

How do the attributes of the distribution of imputed values

0 fare as estimates of the attributes of the distribution of 6?

By calculations similar to those in the first part of the example,

*
it can be shown that (9 ,y) is normal with mean vector 0, as is

(O,y), and its covariance matrix agrees with that of (9,y) for all

elements except the one for 6 with Y2 . Rather than r02 , one

. obtains

Cov(O 'Y2 ) r 2 c r0
02 - (lpc) r81 r1 2

It follows first that characteristics of the joint

*
distribution of 0 and y, are identical to those of 0 and yl. For

instance,

E(6 ) - 0 - E(6)
S Randomization-Based Inference

32

and

Var(0 ) = 1 = Var(O)

This corresponds to the case in which the imputer's model is more

inclusive than the analyst's, and, as Rubin suggests, the result

is satisfactory. Also, corresponding to the case in which the

imputer and the analyst use the same model, one obtains

*
E(O 1yl) = ily I - E(O-yl)

and

Var(O lyl) = 1-aOIxylI Var(01yl)

The secondary an-j,,- thus obtains the correct expectations for

both marginal a:,alyses of 0 and conditional analyses involving

only 0 ar" the collateral variable that was conditioned on.

Less salubrious are results for analyses involving Y2' the

collateral variable that was not conditioned on. Whereas

4 E(Oly) = E(Ojy) - 01 12 y I + f2 11 Y2

one finds that

E(O ly) - [P112 + (l-P )1r

- 01 12 y 1 + f 2 1 1 [pcy 2 + (l-pc)E(y 2 Yl)

1 ] (20a)

= 0 1 12 y1 + 0 2 1 1 Y2 - (l-p* 2 11 (Y2 -rl2 Yl) (20b)

A bias is thus introduced into the imputations, the nature of

which is to attenuate the contribution from Y 2 , the omitted

 Randomization-Based Inference

33

variable. Equation (20a) shows that the contribution associated

with Y 2 is a weighted average of (i) the correct contribution--

fi2 1 1 y2 --to the degree that x is a reliable indicator of 0, and

(ii) to the degree that x is unreliable, a contribution associated

with the expectation of the omitted variable given the observed

value of the conditioned variable. Equation (20b) shows that the

bias can be driven to zero in three ways:

a. As pc+l; i.e., x is a perfectly reliable measure of 0;

b. As P2,1-0; i.e., there is no contribution from Y 2 anyway;

c. As r l2yl'y 2 for all yl; i.e., Y 2 is perfectly predictable

from yI.

Some consequences for regression analyses involving y2 are now

considered.

Whereas the regression coefficient for Y 2 in the multiple

regression of 0 on y is 21 the corresponding coefficient for

211' corsodigcefiin2o
*
in the multiple regression of 6 on y is

P211 = Pc0211 (21)

The expected regression coefficient has been shrunken by the

,
factor (1-p c), the complement of test reliability given yI.

Whereas the regression coefficient for yl in the multiple

regression of 6 on y is obtained as

6"
19. ~.

Randomization-Based Inference

34

112 = (r01-r02r12)/(l-r122

*
the corresponding coefficient in the multiple regression of 0 on

y is

(r 1 -r 2 r 1 2 )/(l-r1 2 (22)

The bias can be expressed as

0

16112 11 2 - (l-Pc) 0 2 11 r1 2

Thus, in the analysis of conditional 9 distributions given both Yl

and Y 2 , bias exists for the coefficient of yl even though it has

been conditioned on. Since r 1 2y I - E(Y 2Iyl), the character of the

bias is to absorb a portion of the unique contribution of the

nonconditioned variable, to the extent that x is unreliable.

Whereas the coefficient P2 for the simple regression of 6 on

Y2 is r02 , the corresponding coefficient for 8 on Y 2 is

+ (23)
02 r0 2 pc r02 ( r81 r12
c-Pc)

The bias with which f2 is estimated from 0 can be expressed as

S - (-r12

* 2
* This bias is reduced as either P, the test reliability, or r2

the propor-;un of Y 2 predictable by y, approaches one, or as the

unique contribution of Y 2 in predicting 0 approaches zero.

0I
 Randomization-Based inference

35

To summarize, the degree of biases in secondary analyses

involving variables not conditioned on involves (i) the

reliability of x, (ii) the association between the conditioned and

the nonconditioned collateral variables, and (iii) the unique

contribution of the nonconditioned variable to predicting 0.

Higher values of p are unequivocally helpful, as they reduce

2
biases of all types. Higher values of r12 on the other hand,
12'
mitigate bias in simple regression involving only the

nonconditioned variable, but exacerbate the bias for the

coefficient of the conditioned variable in multiple regression

involving both. These conclusions extend in natural ways to sets

of conditioned and nonconditioned variables (Beaton and Johnson,

1987; Mislevy and Sheehan, 1987).

Table 5 gives values for expected regression coefficients in

*
analyses of 0 and 0 , using the numerical values employed in

Tables 2 and 3. For simple regression, the results for the

conditioned variable are unbiased, and the results for the

nonconditioned variable are comparable in accuracy to those

obtainable from "optimal" point estimates (bearing in mind the

fact that MLEs yield unbiased regression coefficients but biased

conditional variances and percentile points). It can also be seen

that the results of multiple regression are more sensitive to

omitting variables than those of simple regression. These

findings underscore the importance of choosing conditioning

 Randomization-Based Inference

36

variables wisely if practical considerations preclude conditioning

on all the background variables that have been surveyed.

Insert Table 5 here

Implications and Recommendations

The foregoing results indicate that care is required to

impute values for latent variables in a way that leads to

acceptably accurate results in secondary analyses. Precise

determination of 0 by item responses x alleviates the problem of

biases, but to do so in the context of educational or

psychological measurement requires large numbers of item responses

from every subject--a design that is inefficient for drawing

inferences about only population attributes. When testing time

for individuals is limited, the imputer must build a computing

approximation p (91y,z) that gives good results for a broad range

of potential statistics s involving 0.

If there are only a small number of values that y and z can

take, and a large number of subjects at each combination of

values, one can obtain a nonparametric estimate for each (y,z)

combination by the methods of Laird (1978) or Mislevy (1984).

This leads to imputations that are free from specification error

in p(Oly,z), and secondary analyses will not suffer from biases

from this source. This approach is simply not possible, though,

 Randomization-Based Inference

37

for surveys such as NAEP with large numbers of background and

attitude items.

As mentioned above, one possibility is to assume normal

conditional distributions with structured means and a common

residual variance (Mislevy, 1985). In addition to assuming this

tractable distributional form, further simplification becomes

necessary when there are large numbers of design and background

variables. In ANOVA terms, conditioning on the full joint

distribution of (y,z) could involve millions of effects, while

currently available computing procedures can handle up to about a

hundred. Specifying p wisely means choosing contrasts so as to

optimize the accuracy of potential secondary analyses. Based on

the results of Example 2, the following advice can be offered:

Determine 0 as well as is practical. In the context of

measuring latent proficiencies by test items, recall the

decreasing rate at which reliability increases--and potential

4
biases in secondary analyses thereby decrease--with additional

test items. Trade-offs arise among potential item-sampling

designs. Compared to a design that gives five items to each

subject, a design that gives ten items yields less efficient

estimates of statistics involving variables yc that have been

conditioned on, but less biased estimates of those involving

I
variables yn that have not been conditioned on.
 Randomization-Based Inference

38

Borrow information from related scales. As noted earlier,

imputation methods apply to vector-valued latent variables. In

such applications, one estimates multivariate conditional

distributions p(Oy,z), combines them with multidimensional

likelihoods p(xJ6), and draws vector-valued imputations from joint

predictive distributions p(Olx,y,z). This was done in the NAEP

survey of Young Adult Literacy (Kirsch and Jungeblut, 1986), where

e a each respondent was presented five to fifteen items in each of

four IRT literacy scales. The population correlations of about .6

among scales sharpened the predictive distribution of each scale

for an individual: while the information available directly about

the scale was worth ten items, the thirty items from the other

scales indirectly contributed information worth about another ten.

The biases in secondary analyses were thus reduced as much by

using multivariate imputations for the four scales jointly as they

would have been under separate univariate imputations with twice

* as many items.

Condition explicitly on contrasts that are particularly

K important, such as treatment group in a survey designed expressly

to compare treatment effects in a program evaluation. By doing

so, one ensures that the marginal subpopulation means or

r0 regression coefficients involving key variables are estimated as

accurately as possible. Note that y can include interactions as

well as main effects.

 Randomization-Based Inference

39

Condition on well-chosen combinations of variables. Given

current computational limits, it will often be impossible to

condition on the main effects of all background variables, let

alone two-way or higher interactions among them. One can reduce

biases for a large number of contrasts of interest, beyond those

that can be conditioned on explicitly, by conditioning on linear

combinations of contrasts--for example, the first h component

scores from a principal components decomposition of the covariance

matrix among effects. The results of Example 2 imply that the

variation these partially-conditioned-upon variables share with

the explicitly-conditioned-upon component scores will have

salutary effects in secondary analyses. The degree of bias for an

effect will be limited to the proportion of its variance

unaccounted for by the conditioned-upon components, times the

complement of conditional test reliability.

Example 3: The 1984 NAEP Reading Assessment

During the 1983-84 school year, the National Assessment of

Educational Progress (NAEP) surveyed the reading and writing

skills of national probability samples of students at ages 9, 13,

and 17, and in the modal grades associated with those ages, namely

4, 8, and 11. Beaton (1987) gives details of assessment

procedures and analyses. This section of the present paper

highlights the multiple-imputations procedures used in the

analysis of the reading data.

 I]

Randomization-Based Inference

40

The Student-Sampling Design

A multistage probability sampling design was employed to

select students into the NAEP sample, with counties or groups of

counties as the primary sampling units (PSUs). Schools served as

second-stage sampling units. The assignment of testing sessions

of different types to sampled schools was the third stage of

sampling, and the selection of students within schools was the

fourth. A total of 64 PSUs appeared in the sample, and

assessments were administered at 1,465 schools. About 20,000

students were assessed in reading at each grade/age cohort. For

convenience, grade/age cohorts will be referred to below simply by

their age designations.

Sampling was stratified at the first stage according to

geographic regions, Census Bureau "Sample Description of

Community" (SDOC) classes, and, within urban and rural SDOC

classes, a measure of SES; the latter two criteria comprise Size

* and Type of Community (STOC) classes. Selection probabilities of

sampling units were inversely proportional to estimated population

size, except that extreme rural and low-SES urban areas were

*" oversampled by a factor of two. Neglecting minor adjustments for

nonresponse and poststratification, the design variables Z were

therefore region, STOC, PSU, and school membership.

*Population means and totals of survey variables were computed

as weighted sample means and totals, with a student's weight

essentially inversely proportional to his or her probability of

I

Randomization-Based Inference

41

selection. Uncertainty of such a statistic s due to student

sampling was approximated with a multiweight jackknife procedure.

Thirty-two pairs of similar PSUs were designated. Approximating

the uncertainty of s required computing it 33 times: once in a run

with the total sample, and once with a run corresponding to each

[ PSU pair, with one of its members left out of the sample but with

the sampling weight of its partner doubled. The variance of the

[
32 jackknife estimates around the total value is U, the estimated

sampling variance of s around S.

6

The Survey Variables

L
Each student responded to a number of survey items (Y)

tapping demographic status, educational background and reading

practices, and attitudes about reading and writing. About 50

were common to all assessment forms. Examples are gender,

parents' education, ethnicity, and time spent watching television.

Another 300, of which a given student would receive between about

10 and 30 under the assessment's balanced incomplete block (BIB)

item-sampling design, addressed reading activities in the home and

school.

A total of 340 multiple-choice and free-response reading

* exercises were used in the assessment, although a student who

received any reading exercises received between 5 and 50 of them

under the BIB design. About 80 percent of the students received

some reading exercises. A few of the exercises appeared at all

three ages, but most appeared in only one or in two adjacent ages.
 Randomization-Based Inference

42

The Latent-Variable Model

A priori considerations and extensive dimensionality analyses

supported summarizing responses to a subset of 228 of the 340

items by an IRT model for a single underlying proficiency variable

0 (Zwick, 1987). Responses to these items will be denoted by x.

The 3-parameter logistic (3PL) IRT model was used. Under the 3PL,

the probability of a correct response to Item j from a student

with proficiency 0 is given by

P(x(j)=l0,ab j , b j ,c ) - c + (l-c.)/(l+exp[-l.7a.(0-b)] ,

where x-l indicates a correct response and x-0 an incorrect one,

while (aj,bj,c.) are parameters that characterize the regression

of x M on 0. Coupled with this model for a single item, the

assumptions of local independence embodied by (8) and (9) give the

likelihood function p(xlO,8) [-p(xiB,P,y,z)] for a response vector

x to any subset of the 228 items. Here P denotes the vector of

item parameters of all 228 items in the scale.

Students receiving any of the 228 reading items in the scale

at all received between 5 and 40 of them, with the average about

17. The responses of a sample of 10,000 students were used with a

modified version of Mislevy and Bock's (1983) BILOG computer

program to estimate 3PL item parameters. The modifications

allowed the c-parameters of free-response items to be set to zero

a priori, and distinguished age cohorts when computing marginal

probabilities of students' response patterns. The latter

 Randomization-Based Inference

43

extension is necessary to achieve consistent item parameter

estimates since items were assigned to cohorts based on prior

knowledge about the proficiencies of the cohorts--younger students

were generally administered easier items--but that consistency

holds whether or not sampling weights within cohorts are used

(Mislevy and Sheehan, in press). The resulting item parameter

estimates were placed on a scale in which the unweighted

calibration sample of students was standardized. This scale is

arbitrary up to a linear transformation, so any other convention

would have served equally well to set the scale. The resulting

estimates were taken as fixed throughout the remainder of the

study, thereby fixing the origin and unit-size of the 6 scale.

The Imputation Model

The ideal population model p(Oly,z) to use in conjunction

with p(Olx,B) is a joint distribution involving hundreds of survey

variables and background and attitude items. The computer program

available to estimate a latent population distribution was a

prototype developed for Mislevy's (1985) 4-group example, with

ANOVA-type structures on the means and normal residuals with a

common within-group variance. In the time available to meet NAEP

reporting deadlines, it was possible to extend the program to a

design with 17 effects. A main effects model was chosen that

focused accuracy on traditional NAEP reporting categories: sex,

ethnicity, STOC, region, and parental education, along with

indicators for at, above, or below modal grade and age. A

 Randomization-Based Inference

44

"miscellaneous" cell was included in the model for the small

fraction of students whose values on the aforementioned items were

unrecoverable. Altogether, these variables comprised the vector

(y,z) c . The following model was assumed for conditional

distributions:

91y,z - N[(y,z)c'r,a 2

j 2
where r is a vector of seventeen main effect parameters and a is
2
the residual variance. Together, r and a comprise the

superpopulation parameter a referred to in preceding sections.

Note that the design variables Z are captured largely but

not completely, as STOC and region have been included but PSU,

school membership, and interaction terms have not been. Because

PSU and school-level variation are largely explained by the

conditioned-upon region and SES indicators STOC and parental

education, however, biases caused the omission of PSU and school

c
membership in (y,z) are largely mitigated.
2
Maximum likelihood estimates of r and a were obtained

separately within ages using all available reading data--over

20,000 respondents per age. Separate age analyses implicitly

allow for all two-way interactions between cohort and each of the

other effects. The results are shown in Table 6. Like item

parameter estimates, these estimates were also taken as known true

values thereafter.
I

Randomization-Based Inference

45

Table 6 about here

The posterior distribution of each student i was approximated

by a histogram over 40 equally spaced points 8 between -4.785 and

th
+4.785. The weight of the qth bar in the histogram for the 1

student was obtained as follows:

I

^ A

cc P(x il0= q,"6=0) P(0=8 lY Ci,ZiC ,a=Q)

P(e q IX y,z.) - * c c ^ , (24)
q P(X i 0=8s ,0=0) P(0=es lyi,zia=)

c c
where P(0=8 S ly1i z 1i,ct=a) is the density at 85 of the normal pdf

with mean (y,z)C'rF

1
and the residual variance for the age group to

which Student i belongs. Each imputation was drawn from this

distribution in two steps. First, a bar was selected at random in

accordance with the weights determined in (24). Second, a point

was selected at random from a uniform density over the 0 range

spanned by the selected bar. (Logistic interpolation would have

been better.) Five imputations were drawn in this manner for each

student in the sample.

Illustrative Results

The results of primary interest in this first analysis of the

1984 data were the mean proficiencies of the subpopulations

determined by the traditional NAEP reporting categories. A given

weighted mean was calculated five times, once from each completed
 Randomization-Based Inference

46

data set. The average of the five was the reported estimate. A

jackknife variance was also calculated from each completed

dataset, with the imputations in the set treated as known true

values of 0; these values were also averaged, to estimate U. The

reported variance was the sum of this average jackknife variance

and 1+M- =1.2 times the variance of the five aforementioned means.

In order to avoid negative numbers and fractional values, the

original 0 scale was linearly transformed by 500 + 250.5. Table 7

illustrates some results for Age 17. The full public report is

available as The Reading Report Card: Progress toward Excellence

in our Schools (NAEP, 1985).

Table 7 about here

The proportional increase in variance due to the latency of

6, denoted earlier as rM, varies from 2 percent up to nearly 30

percent. The largest increase is associated with a lower-than-

average scoring group, for whom the test items were relatively

more difficult. The proficiencies of low-scoring individuals were

determined less precisely by their item responses, so that the

K9 likelihoods induced for 6, and the consequent posterior

distributions p(O1x,y,z) from which their imputations were drawn,

were more dispersed. Estimated means from such subpopulations

tended to vary more widely across completed datasets than would

r,
 Randomization-Based Inference

47

means for groups of similar size whose typical members were

measured more accurately by their item responses.

Biases in Secondary Analyses

The completed datasets described above were constructed so as

to focus their accuracy on the marginal subpopulation means

featured in The Reading Report Card. As discussed in a previous

section, however, analyses involving survey variables that were

not conditioned on are subject to bias. An opportunity soon arose

to examine such biases. A report on reading proficiency levels

among pupils whose primary language was not English came due, the

analysis plan for which specified multiple regression analyses

involving both variables that had and variables that had not been

conditioned on when imputations for the original analysis were

constructed. Aware that the proposed analyses were sensitive to

failure to condition, the NAEP staff created new completed

datasets in which all the variables required in the analyses were

conditioned upon. This was made possible by Sheehan's (1985) M-

2
GROUP computer program for estimating r and a in larger models.

Some fifty effects were included in the recalculations for each

age. The same multiple regression analyses were carried out

twice, once with the original completed datasets and once with new

ones created with extended conditioning vectors. The results of

one such comparison are summarized in Table 8. Baratz-Snowden and

Duran (1987) give the final results of all runs.

 Randomization-Based Inference

48

Table 8 about here

It may be seen that multiple regression coefficients for the

two effects which were conditioned upon in both analyses were

least affected in the recalculation, differing by amounts only 4

and 9 percent of their new estimated values. Coefficients for

significant effects not originally conditioned on, however,

differed by between 10 and 40 percent. The direction of the

difference, in nearly all of these cases, was that the original

estimates were shrunken toward zero. Performing the analysis on

the original completed datasets would correctly inform the

researcher about the directions of effects, but would tend to

underestimate their magnitudes by an average of 30-percent.

Extensions

The experience with multiple imputation procedures gained

with the 1984 reading assessment led to a number of insights on

how to extend or improve the procedures. Four are mentioned

below.

Multivariable Imputation. The preceding discussions have

concentrated on the case of a single latent variable. While this

proved adequate for summarizing reading data, both empirical and

theoretical evidence demonstrate the need for multiple scales in

broader content areas such as mathematics and science. NAEP

 Randomization-Based Inference

49

extended multiple imputations procedures to the case of four

variables in the Young Adult Literacy Assessment (Kirs' ai.d

Jungeblut, 1986), and later applied these procedures to five

subscales each to analyze the 1986 mathematics and scipn,

assessments (Beaton, 1988).

In the multivariable case, each latent variable--a different

aspect of, say, literacy skill--is defined through an IRT model.

Assuming conditional independence, the four-dimensional likelihood

p(xI ..... x4611 ... 64) is simply the product of the four

univariate IRT likelihoods, or ITP(xklok). The conditional

distributions p(6 1 ... 4 1y,z) are generally not independent,

however, their associations reflecting population correlations

among skills. The predictive distributions

P(1 0 4 1Xl .... x 4 ,Yz) from which imputations are drawn

reflect these associations. Compared to carrying out imputation

procedures separately within each scale, the multivariable

solution exploits information from all scales to strengthen

inferences about each, and yields consistent, rather than

attenuated, estimates of association among the scales.

Conditioning on Principal Components. An aspect of multiple

imputation procedures that requires improvement is the accuracy of

multiple regression analyses that include nonconditioned

background variables. Conditioning on more background variables

with Sheehan's (1985) improved M-GROUP program certainly increases

the number of secondary analyses whose accuracy will be improved,

 Randomization-Based Inference

50

but research is currently under way to determine how to choose

them wisely. As mentioned above, conditioning on well-chosen

linear combinations of large numbers of variables holds promise.

Analyses of the 1988 data will examine gains in accuracy

attainable with different combinations of numbers of effects

conditioned upon explicitly and effects conditioned on partially,

though principal components.

Accounting for Uncertainty in a and Q. Analyses to date have

taken estimates of item parameters P and conditional distribution

parameters a as known, thereby neglecting the component of

uncertainty associated with them. Present indications are that

the resulting overstatement of precision is negligible because of

the huge sample sizes from which these effects are estimated, but

research is under way to develop efficient methods for

incorporating uncertainty about them as well as about 0. One

approach to doing so is leans on asymptotic results, drawing from

multivariate normal (aB) distributions with means given by MLEs

and variances given by inverses of information matrices. An

alternative approach is to draw from multivariate distributions

whose mean and variance matrix were obtained by a jackknife

procedure. The latter is more intensive computationally, but

captures variation due to lack of model fit as well as due to

sampling.
 Randomization-Based Inference

51

The Average Response Method (ARM). To analyze the data from

the 1984 NAEP survey of writing, Beaton and Johnson (1987) worked

out multiple imputation procedures for the setting of general

linear models. They address the problem of characterizing the

distribution of the average of ratings over all writing exercises-

-a straightforward problem when every examinee is presented all

exercises, but effectively a latent variable problem under an

item-sampling design in which each examinee takes only a few

exercises from the pool. Computational procedures are simpler

under the ARM than with IRT models because the assumed linearity

of relationships permits noniterative unweighted least squares

solutions. Expressions for estimation, imputation, and

expectation in secondary analyses offer insight into the problem

for those familiar with the theory of general linear models.

Conclusion

At the beginning of the decade, Bock, Mislevy, and Woodson

(1982) hailed item response theory as a cornerstone of progress

for educational assessment. Assuming that one can manage the

challenges of control and consistency that arise in any study that

extends over time, IRT does indeed make it possible to solve many

practical problems in assessment, such as allowing item pools to

evolve over time, providing results on a consistent scale in the

face of complex item-sampling designs, and reducing the numbers of

items students are presented.

9
.
i~7

Randomization-Based Inference

52

Possible, but not necessarily easy. Familiar IRT procedures,

based on obtaining point estimates for individual examinees, break

down in efficient assessments that solicit relatively few

responses from each student. This paper and others on IRT in

assessment (e.g., Mislevy and Bock, 1988) make it clear that

higher levels of theoretical and computational complexity are

required to realize the benefits IRT offers.

This paper argues that Rubin's (1987) multiple imputation

procedures provide a suitable theoretical framework for latent

variables in sample surveys, and illustrates how the procedures

can be applied. This method has the advantage of placing the

burden of the problem on the primary analyst, who must create

completed datasets. With them, the secondary analysts can carry

out their research using standard routines for complete data.

0O
 Randomization-Based Inference

53

References

Andersen, E.B., and Madsen, M. (1977) Estimating the parameters

of a latent population distribution. Psychometrika, 42, 357-

374.

Baratz-Snowden, J.C., and Duran, R. (1987). The Educational

Progress of Language Minority Students: Findings from the

1983-84 NAEP Reading Survey. Princeton: National Assessment

of Educational Progress/Educational Testing Service.

Beaton, A.E. (1987). The NAEP 1983/84 Technical Report (NAEP

Report 15-TR-20). Princeton: Educational Testing Service.

Beaton, A.E. (1988). The NAEP 1985/86 Technical Report.

Princeton: Educational Testing Service.

Beaton, A.E., and Johnson, E.J. (1987). The average response

method (ARM) of scoring. In A.E. Beaton, The NAEP 1983/84

Technical Report (NAEP Report 15-TR-20). Princeton:

Educational Testing Service.

Bock, R.D., and Aitkin, M. (1981). Marginal maximum likelihood

estimation of item parameters: An application of an EM

algorithm. Psychometrika, 46, 443-459.

Bock, R.D., Mislevy, R.J., and Woodson, C.E.M. (1982). The next

stage in educational assessment. Educational Researcher, 11,

4-11, 16.

Cassel, C-M, Sarndal, C-E., and Wretman, J.H. (1977). Foundations

of Inference in Survey Sampling. New York: Wiley.

Cochran, W.G. (1977). Sampling Techniques. New York: Wiley.

 Randomization-Based Inference

54

Dempster, A.P., Laird, N.M., and Rubin, D.B. (1977). Maximum

likelihood from incomplete data via the EM algorithm (with

discussion). Journal of the Royal Statistical Society,

Series B, 39, 1-38.

Ford, B.L. (1983). An overview of hot-deck procedures. In W.G.

Madow, I. Olkin, and D.B. Rubin (Eds.), Incomplete Data in

Sample Surveys, Volume 2. Theory and Bibliographies. New

York: Academic Press.

Kelly, T.L. (1947). Fundamentals of Statistics. Cambridge:

Harvard University Press.

Kirsch, I.S., and Jungeblut, A. (1986). Literacy: Profile of

America's Young Adults (Final Report, No. 16-PL-01).

Princeton, NJ: National Assessment of Educational Progress.

Laird, N.M. (1978). Nonparametric maximum likelihood estimation

of a mixing distribution. Journal of the American

Statistical Association, 73, 805-811.

Little, R.J.A., and Rubin, D.B. (1987). Statistical Analysis with

Missing Data. New York: Wiley.

Mislevy, R.J. (1984). Estimating latent distributions.

Psychometrika, 49, 359-381.

Mislevy, R.J. (1985). Estimation of latent group effects.

Journal of the American Statistical Association, 80, 993-997.

Mislevy, R.J., and Bock, R.D. (1983). BILOG: Item Analysis and

Test Scoring with Binary Logistic models [computer program].

Mooresville, IN: Scientific Software, Inc.

L
Randomization-Based Inference

55

Mislevy, R.J., and Bock, R.D. (1988). A hierarchical item

response model for educational assessment. In R.D. Bock

(Ed.), Multilevel Analysis of Educational Data. New York:

Academic Press.

Mislevy, R.J., and Sheehan, K.M. (1987). Marginal estimation

procedures. In A.E. Beaton, The NAEP 1983/84 Technical

Report (NAEP Report 15-TR-20). Princeton: Educational

0 Testing Service.
Mislevy, R.J., and Sheehan, K.M. (in press). The role of

collateral information about examinees in the estimation of

item parameters. Psychometrika.

National Assessment of Educational Progress (1985). The Reading

Report Card: Progress toward Excellence in our Schools.

Princeton: Educational Testing Service.

Rigdon, S. E., and Tsutakawa, R.K. (1983). Parameter estimation

in latent trait models. Psychometrika, 48, 567-574.

6 Rubin, D.B. (1976). Inference and missing data. Biometrika, 63.

581-592.

Rubin, D.B. (1987). Multiple imputation for nonresponse in

surveys. New York: Wiley.

Sanathanan, L, and Blumenthal, N. (1978) The logistic model and

estimation of latent structure. Journal of the American

Statistical Association, 73, 794-798.

Sheehan, K.M. (1985). M-GROUP: Estimation of Group Effects in

.. ..
. ......
 Randomization-Based Inference

56

Multivariate Models [computer program]. Princeton:

Educational Testing Service.

Zwick, R. (1987). Assessing the dimensionality of NAEP reading

data. Journal of Educational Measurement, 24, 293-308.

0
 Randomization-Based Inference

57

C W --
I,- Z
0 ,- t

ii -4

ZL

C4II C lb+
I 1 -p3 ,, I€1

iiii I3 - - (N

0 %-4
0) ) I

CO N-4
M C t Q)-
-

'-4 a) -44 q --

b .-4
-4 4 C1 - -4J

.-
o1-
'
4 00N
~ .Jo ,

a) IC4 r) - ) , )Q
a-
_44 4J
...
-4 1 _ $ . r
4) 4.. Z 4- C O 0 M :

-4 C4a -4C') 4
o H + 0. 4 a)M- 4M (
H CL- 1 c . o u C w > C

I 0) .I0 M
04> E
 Randomization-Based Inference

58

Table 2

Numerical Values for a Short Assessment Instrument

Dependent Variable
Population - A

Attribute 0 and 6 6 (r)

xy x x
Mean .000 .000 .000 .000 .000

Variance 1.000 2.000 .600 .500 1.000

*I Simple
Regression

Coefficient .500 .500 .500 .250 .354

Residual
Variance .750 1.750 .350 .438 .875

% variance
accounted for .250 .125 .417 .125 .125
.................................................................

Table 3

Numerical Values for a Long Test

Dependent Variable
4 Population - A

6 0 0 (r)
Attribute 0 and 6
xy x x
Mean .000 .000 .000 .000 .000

Variance 1.000 1.100 .940 .910 1.000

Simple
Regression

Coefficient .500 .500 .500 .455 .477

Residual
Variance .750 .850 .690 .703 .773

% variance
accounted for .250 .227 .266 .227 .227
.................................................................
 Randomization-Based Inference

59

Table 4

Estimated Proportions Above Cut Point

Dependent Variable
Population - *

Attribute 0 and 0 9 9 (r)

xy x x
Short Test

P(9:I) .1587 .2611 .0985 .0778 .1587

P(6Ollyl=-l) .0418 .1292 .0055 .0294 .0735

Long Test

P(9 I) .1587 .1711 .1515 .1469 .1587

P(9O11yl=-i) .0418 .0516 .0351 .0409 .0465

Table 5

Expected Regression Coefficients for a Short and a Long Test,

with Complete and Incomplete I Conditioning for Imputations

Dependent Variable
Population * *
Attribute 0 and 0 6 (pT* 50) 0 (pT=.91)

Simple regression

O1 .500 .500 .500

02 .500 .357 .471

Multiple regression

0112 .333 .429 .353

0211 .333 .143 .294

I Imputations constructed by conditioning on y1 but not Y 2 "

 Randomization-Based Inference

60

Table 6
Estimates of Conditional Distribution Parameters

Grade 4/ Grade 8/ Grade 11/

Effect Level Age 9 Age 13 Age 17
.................................................................

Intercept All subjects -1.351 -.433 .159

Gender Male .000 .000 .000

Female .096 .139 .160
,
Ethnicity Black .000 .000 .000
White & other .460 .403 .405
Hispanic .076 .113 .135

STOC Rural or Low metro .000 .000 .000

High metro .490 .308 .230
Other .245 .122 .148

Region Northeast .000 .000 .000

Central -.133 -.042 .028
Southeast -.009 -.021 .023
West -.087 -.042 .005

Parent Ed. Less than HS .000 .000 .000

High school grad .209 .140 .082
Beyond HS .395 .404 .379
Don't know/missing .120 -.017 -.075

Grade/Age < M age, = M grade .000 .000 .000

S= M age, < M grade -.672 -.433 -.617
= M age, = M grade -.065 -.013 -.084
= M age, > M grade .338 .549 .077
> M age, = M grade -.307 -.260 -.533

Misc. Subjects with

* unrecoverable
missing values .510 -.329 .810

Residual variance .464 .386 .457

Sample size 22,950 23,553 23,932

-----------------------------------------------------------------

Effect fixed at zero

"M" denotes "modal"; e.g., "> M age, - M grade" means "above

the modal age and at the modal grade in one's age/grade
cohort."
 Randomization-Based Inference

61

Table 7

Estimating Age 17 Means from Completed Datasets

Total Males Black West Rural

.................................................................

(1) Imputation 1 288.005 282.644 266.195 287.177 282.990

(2) Imputation 2 288.258 283.201 265.104 288.338 283.499
(3) Imputation 3 288.208 282.869 265.259 288.018 283.285
(4) Imputation 4 288.135 282.554 264.832 287.745 282.854
(5) Imputation 5 287.819 282.314 264.241 287.196 282.360

(6) Average (l)-(5) 288.085 282.718 265.126 287.695 282.997

(7) Variance (l)-(5) .025 .092 .406 .208 .152

(8) Average jackknife

variance 1.248 1.225 1.742 4.333 9.218

(9) Total variance

1.2 x (7) + (8) 1.278 1.335 2.229 4.583 9.400

(10) Proportional increase

[(9)-(8)]/(8) .024 .090 .280 .058 .020

I.
 Randomization-Based Inference

62

Table 8

Multiple Regression Estimates based on Imputations

Constructed with Partial and Full Conditioning

Partial Full Conditioning %-attenuation

Effect 0 f SE(p) t all significant

White; language
minority 6.08 4.23 3.96 1.07 -43.74

White; language
non-minority 12.22 13.72 2.94 4.67 10.93 10.93

Hispanic; lang.
non-minority - .76 1.22 3.25 .38 162.30

Asian; language
minority -2.90 -6.25 4.39 -1.42 53.60

Asian; language
non-minority 9.54 17.34 4.66 3.72 44.98 44.98

Black; language
non-minority -8.64 -10.82 2.95 -3.67 20.15 20.15

Sex = male -8.55 -9.35 .80 -11.69 8.56 8.56

Parent education 6.03 5.80 .38 15.26 -3.97 -3.97

Home language
* minority -9.41 -13.78 2.78 -4.96 31.71 31.71

Study aids 2.63 3.89 .43 9.05 32.39 32.39

Homework 2.78 3.82 .30 12.73 27.23 27.23

* Hours of TV -1.22 -2.04 .24 -8.50 40.20 40.20

Pages read 6.36 10.59 1.01 10.49 39.94 39.94

Years academic
courses .91 1.25 .14 8.93 27.20 27.20
E

Effect included in partial conditioning set

 1988/07/06

Educational Testing Service/Mislevy

Dr. Terry Ackerman Dr. R. Darrell Bock

American College Testing Programs University of Chicago
P.O. Box 168 NORC
Iowa City, IA 52243 6030 South Ellis
I Dr. Robert Ahiers Chicago, IL 60637

Code N711 Cdt. Arnold Bohrer

Human Factors Laboratory Sectie Psychologisch Onderzoek
Naval Training Systems Center Rekruterings-En Selectiecentrum
Orlando, FL 32813 Kwartier Koningen Astrid
Bruijnstraat
Dr. James Aigina 1120 Brussels, BELGIUM
1403 Norman Hall
University of Florida Dr. Robert Breaux
Gainesville, FL 32605 Code 7B
Naval Training Systems Center
Dr. Erl ing B. Andersen Orlando, FL 32813-7100
Department of Statistics
Studiestraede 6 Dr Robert Brennan
1455 Copenhagen American College Testing
DENMARK Programs
P. 0. Box 168
Dr. Eva L. Baker Iowa City, IA 52243
UCLA Center for the Study
of Evaluation Dr. James Carlson
145 Moore Hall American College Testing
University of California Program
Los Angeles, CA 90024 P.O. Box 168
Iowa City, IA 52243
Dr. Isaac Bejar
Mail Stop: 1O-R Dr. John B. Carroll
Educational Testing Service 409 Elliott Rd., North
Rosedale Road Chapel Hill, NC 27514
Princeton, NJ 08541
Dr. Robert M. Carroll
Dr. Menucha Birenbaum Chief of Naval Operations
School of Education OP-01B2
Tel Aviv University Washington, DC 20350
Ramat Aviv 69978
ISRAEL Dr. Raymond E. Christal
UES LAMP Science Advisor
Dr. Arthur S. Blaiwes AFHRL/MOEL
Code N712 Brooks AFB, TX 78235
Naval Training Systems Center
Orlando, FL 32813-7100 Dr. Norman Cliff
Department of Psychology
' Dr. Bruce Bloxom Univ. of So. California
Defense Manpower Data Center Los Angeles, CA 90089-1061
550 Camino El Estero,
Suite 200
Monterey, CA 93943-3231
 1988/07/06

Educational Testing Service/Mislevy

Director, Dr. Fritz Drasgow

Manpower Support and University of Illinois
Readiness Program Department of Psychology
Center for Naval Analysis 603 E. Daniel St.
I 2000 North Beauregard Street Champaign, IL 61820
Alexandria, VA 22311
Defense Technical
Dr. Stanley Collyer Information Center
Office of Naval Technology Cameron Station, Bldg 5
Code 222 Alexandria, VA 22314
800 N. Quincy Street Attn: TC
Arlington, VA 22217-5000 (12 Copies)

Dr. Hans F. Crombag Dr. Stephen Dunbar

Faculty of Law 224B Lindquist Center
University of Limburg for Measurement
P.O. Box 616 University of Iowa
Maastricht Iowa City, IA 52242
The NETHERLANDS 6200 MD
Dr. James A. Earles
Dr. Timothy Davey Air Force Human Resources Lab
Educational Testing Service Brooks AFB, TX 78235
Princeton, NJ 08541
Dr. Kent Eaton
Dr. C. M. Dayton Army Research Institute
Department of Measurement 5001 Eisenhower Avenue
Statistics & Evaluation Alexandria, VA 22333
College of Education
University of Maryland Dr. John M. Eddins
College Park, MD 20742 University of Illinois
252 Engineering Research
Dr. Ralph J. DeAyala Laboratory
Measurement, Statistics, 103 South Mathews Street
and Evaluation Urbana, IL 61801
Benjamin Bldg., Rm. 4112
University of Maryland Dr. Susan Embretson
College Park, MD 20742 University of Kansas
Psychology Department
Dr. Dattprasad Divgi 426 Fraser
Center for Naval Analysis Lawrence, KS 66045
4401 Ford Avenue
P.O. Box 16268 Dr. George Englehard, Jr.
Alexandria, VA 22302-0268 Division of Educational Studies
Emory University
Dr. Hei-Ki Dong 210 Fishburne Bldg.
Bell Communications Research Atlanta, GA 30322
6 Corporate Place
PYA-1K226 Dr. Benjamin A. Fairbank
Piscataway, NJ 08854 Performance Metrics, Inc.
5825 Callaghan
Suite 225
San Antonio, TX 78228

4t
* 1988/07/06

Educational resting Service/Mislevy

Dr. P-A. Federico Dr. Robert Glaser

Code 51 Learning Research
NPRDC & Development Center
San Diego, CA 92152-6800 University of Pittsburgh
3939 O'Hara Street
Dr. Leonard Feldt Pittsburgh, PA 15260
Lindquist Center
for Measurement Dr. Bert Green
University of Iowa Johns Hopkins University
Iowa City, IA 52242 Department of Psychology
Charles & 34th Street
Dr. Richard L. Ferguson Baltimore, MD 21218
American College Testing
P.O. Box 168 DORNIER GMBH
Iowa City, IA 52243 P.O. Box 1420
D-7990 Friedrichshafen 1
Dr. Gerhard Fischer WEST GERMANY
* Liebiggasse 5/3
A 1010 Vienna Dr. Ronald K. Hambleton
AUSTRIA University of Massachusetts
Laboratory of Psychometric
Dr. Myron Fischl and Evaluative Research
U.S. Army Headquarters Hills South, Room 152
DAPE-MRR Amherst, MA 01003
The Pentagon
Washington, DC 20310-0300 Dr. Delwyn Harnisch
University of Illinois
Prof. Donald Fitzgerald 51 Gerty Drive
University of New England Champaign, IL 61820
Department of Psychology
Armidale, New South Wales 2351 Dr. Grant Henning
AUSTRALIA Senior Research Scientist
Division of Measurement
Mr. Paul Foley Research and Services
Navy Personnel R&D Center Educational Testing Service
San Diego, CA 92152-6800 Princeton, NJ 08541

Dr. Alfred R. Fregly Ms. Rebecca Hetter

AFOSR/NL, Bldg. 410 Navy Personnel R&D Center
Boiling AFB, DC 20332-6448 Code 63
San Diego, CA 92152-6800
Dr. Robert D. Gibbons
Illinois State Psychiatric Inst. Dr. Paul W. Holland
Rm 529W Educational Testing Service, 21-T
1601 W. Taylor Street Rosedale Road
Chicago, IL 60612 Princeton, NJ 08541
Dr. Janice Gifford Prof. Lutz F. Hornke
University of Massachusetts Institut fur Psychologie
School of Education RWTH Aachen
Amherst, MA 01003 Jaegerstrasse 17/19
D-5100 Aachen
WEST GERMANY
 1988/07/06

Educational Testing Service/Mislevy

Dr. Paul Horst Dr. G. Gage Kingsbury

677 G Street, #184 Portland Public Schools
Chula Vista, CA 92010 Research and Evaluation Department
501 North Dixon Street
Mr. Dick Hoshaw P. 0. Box 3107
OP-135 Portland, OR 97209-3107
Arlington Annex
Room 2834 Dr. William Koch
Washington, DC 20350 Box 7246, Meas. and Eval. Ctr.
University of Texas-Austin
Dr. Lloyd Humphreys Austin, TX 78703
University of Illinois
Department of Psychology Dr. James Kraatz
603 East Daniel Street Computer-based Education
Champaign, IL 61820 Research Laboratory
University of Illinois
* Dr. Steven Hunka Urbana, IL 61801
3-104 Educ. N.
University of Alberta Dr. Leonard Kroeker
Edmonton, Alberta Navy Personnel R&D Center
CANADA T6G 2G5 Code 62
i Dr. Ruynh Huynh San Diego, CA 92152-6800

College of Education Dr. Jerry Lehnus

Univ. of South Carolina Defense Manpower Data Center
Columbia, SC 29208 Suite 400
1600 Wilson Blvd
Dr. Robert Jannarone Rosslyn, VA 22209
Elec. and Computer Eng. Dept.
University of South Carolina Dr. Thomas Leonard
Columbia, SC 29208 University of Wisconsin
Department of Statistics
Dr. Douglas H. Jones 1210 West Dayton Street
Thatcher Jones Associates Madison, WI 53705
* P.O. Box 6640
10 Trafalgar Court Dr. Michael Levine
Lawrenceville, NJ 08648 Educational Psychology
210 Education Bldg.
Dr. Milton S. Katz University of Illinois
European Science Coordination Champaign, IL 61801
* Office
U.S. Army Research Institute Dr. Charles Lewis
Box 65 Educational Testing Service
FPO New York 09510-1500 Princeton, NJ 08541-0001

Prof. John A. Keats Dr. Robert L. Linn

Department of Psychology Campus Box 249
University of Newcastle University of Colorado
N.S.W. 2308 Boulder, CO 80309-0249
AUSTRALIA
-t 1988/07/06

Educational Testing Servi ce/Mislevy

Dr. Robert Lockman Dr. William Montague

Center for Naval Analysis NPRDC Code 13
4401 Ford Avenue San Diego, CA 92152-6800
P.O. Box 16268
Alexandria, VA 22302-0268 Ms. Kathleen Moreno
Navy Personnel R&D Center
Dr. Frederic M. Lord Code 62
Educational Testing Service San Diego, CA 92152-6800
Princeton, NJ 08541
Headquarters Marine Corps
Dr. George B. Macready Code MPI-20
Department of Measurement Washington, DC 20380
Statistics & Evaluation
Col lege of Education Dr. W. Alan Nicewander
University of Maryland University of Oklahoma
Col lege Park, MD 20742 Department of Psychology
0Dr. Gary Marco Norman, OK 73071

Stop 31-E Deputy Technical Director

Educational Testing Service NPRDC Code 01A
Princeton, NJ 08451 San Diego, CA 92152-6800

Dr. James R. McBride Director, Training Laboratory,

The Psychological Corporation NPRDC (Code 05)
1250 Sixth Avenue San Diego, CA 92152-6800
San Diego, CA 92101
Director, Manpower and Personnel
Dr. Clarence C. McCormick Laboratory,
HQ, USMEPCOM/MEPCT NPRDC (Code 06)
2500 Green Bay Road San Diego, CA 92152-6800
North Chicago, IL 60064
Director, Human Factors
Dr. Robert McKinley & Organizational Systems Lab,
Educational Testing Service NPRDC (Code 07)
16-T San Diego, CA 92152-6800
Princeton, NJ 08541
Library, NPRDC
Dr. James McMichael Code P2O1L
Technical Director San Diego, CA 92152-6800
Navy Personnel R&D Center
* San Diego, CA 92152-6800 Commanding Officer,
Naval Research Laboratory
Dr. Barbara Means Code 2627
SRI International Washington, DC 20390
333 Ravenswood Avenue
Menlo Park, CA 94025 Dr. Harold F. O'Neil, Jr.
0 School of Education - WPH 801
Dr. Robert Mislevy Department of Educational
Educational Testing Service Psychology & Technology
Princeton, NJ 08541 University of Southern California
Los Angeles, CA 90089-0031
 1988/07/06

Educational Testing Service/Mislevy

Dr. James B. Olsen Department of Operations Research,

WICAT Systems Naval Postgraduate School
1875 South State Street Monterey, CA 93940
Orem, UT 84058
Dr. Mark D. Reckase
Office of Naval Research, ACT
Code 1142CS P. 0. Box 168
800 N. Quincy Street Iowa City, IA 52243
Arlington, VA 22217-5000
(6 Copies) Dr. Malcolm Ree
AFHRL/MOA
Off~ce of Naval Research, Brooks AFB, TX 78235
Code 125
800 N. Quincy Street Dr. Barry Riegelhaupt
Arlington, VA 22217-5000 HumRRO
1100 South Washington Street
Assistant for MPT Research, Alexandria, VA 22314
Development and Studies
OP 01B7 Dr. Carl Ross
Washington, DC 20370 CNET-PDCD
Building 90
Dr. Judith Orasanu Great Lakes NTC, IL 60088
Basic Research Office
Army Research Institute Dr. J. Ryan
5001 Eisenhower Avenue Department of Education
Alexandria, VA 22333 University of South Carolina
Columbia, SC 29208
Dr. Jesse Orlansky
Institute for Defense Analyses Dr. Fumiko Samejima
1801 N. Beauregard St. Department of Psychology
Alexandria, VA 22311 University of Tennessee
310B Austin Peay Bldg.
Dr. Randolph Park Knoxville, TN 37916-0900
Army Research Institute
5001 Eisenhower Blvd. Mr. Drew Sands
Alexandria, VA 22333 NPRDC Code 62
San Diego, CA 92152-6800
Wayne M. Patience
American Council on Education Lowell Schoer
GED Testing Service, Suite 20 Psychological & Quantitative
4 One Dupont Circle, NW Foundations
Washington, DC 20036 College of Education
University of Iowa
Dr. James Paulson Iowa City, IA 52242
Department of Psychology
Portland State University Dr. Mary Schratz
P.O. Box 751 Navy Personnel R&D Center
Portland, OR 97207 San Diego, CA 92152-6800

Dept. of Administrative Sciences Dr. Dan Segall

Code 54 Navy Personnel R&D Center
Naval Postgraduate School San Diego, CA 92152
Monterey, CA 93943-5026
o1988/07/06
Educational Testing Service/Mislevy

Dr. W. Steve Sellman Or. Hariharan Swaminathan

IDASD MRA&L) Laboratory of Psychometric and
2B269 The Pentagon Evaluation Research
Washington, DC 20301 School of Education
University of Massachusetts
Dr. Kazuo Shigemasu Amherst, MA 01003
7-9-24 Kugenuma-Kaigan
Fujisawa 251 Mr. Brad Sympson
JAPAN Navy Personnel R&D Center
Code-62
Dr. William Sims San Diego, CA 92152-6800
Center for Naval Analysis
4401 Ford Avenue Dr. John Tangney
P.O. Box 16268 AFOSR/NL, Bldg. 410
Alexandria, VA 22302-0268 Boiling AFB, DC 20332-6448

Dr. H. Wallace Sinaiko Dr. Kikumi Tatsuoka

Manpower Research CERL
and Advisory Services 252 Engineering Research
Smithsonian Institution Laboratory
801 North Pitt Street, Suite 120 103 S. Mathews Avenue
Alexandria, VA 22314-1713 Urbana, IL 61801

Dr. Richard E. Snow Dr. Maurice Tatsuoka

School of Education 220 Education Bldg
Stanford University 1310 S. Sixth St.
Stanford, CA 94305 Champaign, IL 61820

Dr. Richard C. Sorensen Dr. David Thissen

Navy Personnel R&D Center Department of Psychology
San Diego, CA 92152-6800 University of Kansas
Lawrence, KS 66044
Dr. Paul Speckman
University of Missouri Mr. Gary Thomasson
* Department of Statistics University of Illinois
Columbia, MO 65201 Educational Psychology
Champaign, IL 61820
Dr. Judy Spray
ACT Dr. Robert Tsutakawa
P.O. Box 168 University of Missouri
Iowa City, IA 52243 Department of Statistics
222 Math. Sciences Bldg.
Dr. Martha Stocking Columbia, MO 65211
Educational Testing Service
Princeton, NJ 08541 Dr. Ledyard Tucker
University of Illinois
* Dr. William Stout Department of Psychology
University of Illinois 603 E. Daniel Street
Department of Statistics Champaign, IL 61820
101 Illini Hall
725 South Wright St.
Champaign, IL 61820
 1988/07/06

Educational Testing Service/Mislevy

Dr. Vern W. Urry Dr. Douglas Wetzel

Personnel R&D Center Code 51
Office of Personnel Management Navy Personnel R&D Center
1900 E. Street, NW San Diego, CA 92152-6800
Washington, DC 20415
Dr. Rand R. Wilcox
Dr. David Vale University of Southern
Assessment Systems Corp. California
2233 University Avenue Department of Psychology
Suite 440 Los Angeles, CA 90089-1061
St. Paul, MN 55114
German Military Representative
Dr. Frank L. Vicino ATTN: Wolfgang Wildgrube
Navv Personnel R&D Cente- Streitkraefteamt
San Diego, CA 92152-6800 D-5300 Bonn 2
4000 Brandywine Street, NW
Dr. Howard Wainer Washington, DC 20016
Educational Testing Service
Princeton, NJ 08541 Dr. Bruce Williams
Department of Educational
Dr. Ming-Me! Wang Psychology
Lindquist Center University of Illinois
for Measurement Urbana, IL 61801
University of Iowa
Iowa City, IA 52242 Dr. Hilda Wing
NRC MH-176
Dr. Thomas A. Warm 2101 Constitution Ave.
Coast Guard Institute Washington, DC 20418
P. 0. Subs ation 18
Oklahoma City, OK 73169 Dr. Martin F. Wiskoff
Defense Manpower Data Center
Dr. Brian Naters 550 Camino El Estero
HumRRO Suite 200
12908 Argy.e Circle Monterey, CA 93943-3231
* Alexandria, VA 22314
Mr. John H. Wolfe
Dr. David . Weiss Navy Personnel R&D Center
N660 Elliott Hall San Diego, CA 92152-6800
University of Minnesota
75 E. Rive Road Dr. George Wong
Minneapoli., MN 55455-0344 Biostatistics Laboratory
Memorial Sloan-Kettering
Dr. Ronald A. Weitzman Cancer Center
Box 146 1275 York Avenue
Carmel, CA 93921 New York, NY 10021

Major John Welsh Dr. Wallace Wulfeck, III

AFHRL/MOAN Navy Personnel R&D Center
Brooks AFB, TX 78223 Code 51
San Diego, CA 92152-6800
91988/07/06
Educational Testing Service/Mislevy

Dr. Kentaro Yamamoto

03-T
Educational Testing Service
Rosedale Road
Princeton, NJ 08541

Dr. Wendy Yen

CTB/McGraw Hill
Del Monte Research Park
Monterey, CA 93940

Dr. Joseph L. Young

National Science Foundation
Room 320
1800 G Street, N.W.
Washington, DC 20550

Mr. Anthony R. Zara

National Council of State
Boards of Nursing, Inc.
625 North Michigan Avenue
Suite 1544
Chicago, IL 60611

Dr. Peter Stoloff

Center for Naval Analysis
4401 Ford Avenue
P.O. Box 16268
Alexandria, VA 22302-0268