Quantitative Psychology Research The 78th Annual Meeting

of the Psychometric Society

Visit the link below to download the full version of this book:

https://medipdf.com/product/quantitative-psychology-research-the-78th-annual-mee
ting-of-the-psychometric-society/

Click Download Now

Preface

This volume represents presentations given at the 78th annual meeting of the
Psychometric Society, organized by Cito and held at the Musis Sacrum in Arnhem,
the Netherlands, during July 22–26, 2013. The meeting attracted 334 participants
from 28 countries, with 242 papers being presented, along with 49 poster pre-
sentations, five pre-conference workshops, three keynote presentations, six invited
presentations, six state-of-the-art lecturers, and three invited symposia. We thank
the local organizer Anton Béguin and his staff and students for hosting this very
successful conference.
After the 77th meeting in Lincoln, Nebraska, the idea was presented to publish a
proceedings volume from the conference so as to allow presenters to quickly make
their ideas available to the wider research community, while still undergoing a thor-
ough review process. Because the first volume was received successfully, it was sug-
gested that we publish proceedings more regularly. Hence, this is the second volume,
and a third volume following the 79th meeting in Madison, Wisconsin, is expected.
We asked authors to use their presentation at the meeting as the basis of their
chapters, possibly extended with new ideas or additional information. The result is a
selection of 29 state-of-the-art chapters addressing a diverse set of topics, including
classical test theory, item response theory, factor analysis, measurement invariance,
test equating and linking, mediation analysis, cognitive diagnostic models, marginal
models, and multi-level models.
The joy of editing these proceedings was overshadowed by the tragic news
that Roger E. Millsap had passed away suddenly on May 9, 2014. As editor of
Psychometrika and former president, Roger played an important role in the Psy-
chometric Society. He was also the initiator and principal editor of the proceedings.
He passed away shortly after finalizing these proceedings. We will always remember
him fondly as the driving force of this project, and we will miss the friendly, helpful,
and competent advice of this well-seasoned editor. May you rest in peace Roger.

Amsterdam, The Netherlands L. Andries van der Ark

Madison, WI, USA Daniel M. Bolt
Hong Kong Wen-Chung Wang
v
Contents

1 What Do You Mean by a Difficult Item?

On the Interpretation of the Difficulty
Parameter in a Rasch Model . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
Ernesto San Martín and Paul De Boeck
2 Thurstonian Item Response Theory and an Application
to Attitude Items . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
Edward H. Ip
3 Robustness of Mixture IRT Models to Violations of Latent
Normality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
Sedat Sen, Allan S. Cohen, and Seock-Ho Kim
4 An Option-Based Partial Credit Item Response Model . . . . . . . . . . . . . . . . 45
Yuanchao (Emily) Bo, Charles Lewis, and David V. Budescu
5 A General Saltus LLTM-R for Cognitive Assessments . . . . . . . . . . . . . . . . . 73
Minjeong Jeon, Karen Draney, and Mark Wilson
6 Multidimensional IRT Models to Analyze Learning
Outcomes of Italian Students at the End of Lower
Secondary School .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 91
Mariagiulia Matteucci and Stefania Mignani
7 Graphical Representations of Items and Tests That are
Measuring Multiple Abilities . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113
Terry A. Ackerman and Robert A. Henson
8 New Item-Selection Methods for Balancing Test Efficiency
Against Item-Bank Usage Efficiency in CD-CAT . . .. . . . . . . . . . . . . . . . . . . . 133
Wenyi Wang, Shuliang Ding, and Lihong Song

vii
viii Contents

9 Comparison of Linear, Computerized Adaptive and Multi

Stage Adaptive Versions of the Mathematics Assessment
of Turkish Pupil Monitoring System . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
Semirhan Gökçe and Giray Berberoğlu
10 Optimal Sampling Design for IRT Linking with Bimodal Data .. . . . . . 165
Jiahe Qian and Alina A. von Davier
11 Selecting a Data Collection Design for Linking
in Educational Measurement: Taking Differential
Motivation into Account . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
Marie-Anne Mittelhaëuser, Anton A. Béguin,
and Klaas Sijtsma
12 Vertical Comparison Using Reference Sets . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 195
Anton A. Béguin and Saskia Wools
13 A Dependent Bayesian Nonparametric Model for Test Equating .. . . . 213
Jorge González, Andrés F. Barrientos,
and Fernando A. Quintana
14 Using a Modified Multidimensional Priority Index for
Item Selection Under Within-Item Multidimensional
Computerized Adaptive Testing . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 227
Ya-Hui Su and Yen-Lin Huang
15 Assessing Differential Item Functioning in Multiple
Grouping Variables with Factorial Logistic Regression .. . . . . . . . . . . . . . . 243
Kuan-Yu Jin, Hui-Fang Chen, and Wen-Chung Wang
16 MTP2 and Partial Correlations in Monotone
Higher-Order Factor Models . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 261
Jules L. Ellis
17 A Comparison of Confirmatory Factor Analysis
of Binary Data on the Basis of Tetrachoric Correlations
and of Probability-Based Covariances: A Simulation Study . . . . . . . . . . 273
Karl Schweizer, Xuezhu Ren, and Tengfei Wang
18 On Cronbach’s Alpha as the Mean of All Split-Half Reliabilities . . . . . 293
Matthijs J. Warrens
19 An Empirical Assessment of Guttman’s Lambda 4
Reliability Coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301
Tom Benton
20 A Test for Ordinal Measurement Invariance .. . . . . . .. . . . . . . . . . . . . . . . . . . . 311
Rudy Ligtvoet
Contents ix

21 Model Selection Criteria for Latent Growth Models Using

Bayesian Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 319
Zhenqiu (Laura) Lu, Zhiyong Zhang, and Allan Cohen
22 A Comparison of the Hierarchical Generalized Linear
Model, Multiple-Indicators Multiple-Causes, and the Item
Response Theory-Likelihood Ratio Test for Detecting
Differential Item Functioning . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343
Mei Ling Ong, Laura Lu, Sunbok Lee, and Allan Cohen
23 Comparing Estimation Methods for Categorical
Marginal Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 359
Renske E. Kuijpers, Wicher P. Bergsma,
L. Andries van der Ark, and Marcel A. Croon
24 Evaluating Simplicial Mixtures of Markov Chains
for Modeling Student Metacognitive Strategies .. . . .. . . . . . . . . . . . . . . . . . . . 377
April Galyardt and Ilya Goldin
25 Partitioning Variance Into Constituents in Multiple
Regression Models: Commonality Analysis . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 395
Burhanettin Ozdemir
26 Multilevel Random Mediation Analysis: A Comparison
of Analytical Alternatives .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 407
Fang Luo and Hongyun Liu
27 Mediation Analysis for Ordinal Outcome Variables.. . . . . . . . . . . . . . . . . . . 429
Hongyun Liu, Yunyun Zhang, and Fang Luo
28 Comparison of Nested Models for Multiply Imputed Data .. . . . . . . . . . . 451
Yoonsun Jang, Zhenqiu (Laura) Lu, and Allan Cohen
29 A Paradox by Another Name Is Good Estimation . .. . . . . . . . . . . . . . . . . . . . 465
Mark D. Reckase and Xin Luo
Chapter 1
What Do You Mean by a Difficult Item?
On the Interpretation of the Difficulty
Parameter in a Rasch Model

Ernesto San Martín and Paul De Boeck

Abstract Three versions of the Rasch model are considered: the fixed-effects
model, the random-effects model with normal distribution, and the random-effects
model with unspecified distribution. For each of the three, we discuss the meaning of
the difficulty parameter starting each time from the corresponding likelihood and the
resulting identified parameters. Because the likelihood and the identified parameters
are different depending on the model, the identification of the parameter of interest is
also different, with consequences for the meaning of the item difficulty. Finally, for
all the three models, the item difficulties are monotonically related to the marginal
probabilities of a correct response.

1.1 Introduction

In the Rasch model, the probability of success in an item is defined on the basis of a
contribution from the part of the person (person ability) and a contribution from the
part of the item (item difficulty), while the contribution from the part of the persons
does not depend on the item and neither does the effect of the items depend on the
person. The Rasch model is, therefore, a main-effect model. The basic formula is
the following:

Ypi ∼ Bern [Ψ (θ p − βi )], (1.1)

E. San Martín ()

Faculty of Mathematics and Faculty of Education, Pontificia Universidad
Católica de Chile, Vicuña Mackenna 4860, Santiago, Chile
Center for Operations Research and Econometrics CORE, 30 Voie du Roman Pays,
Louvain-la-Neuve, Belgium
e-mail: [email protected]
P. De Boeck
Department of Psychology, The Ohio State University,
1835 Neil Avenue Columbus, Columbus, OH 43210, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 1

R.E. Millsap et al. (eds.), Quantitative Psychology Research, Springer Proceedings
in Mathematics & Statistics 89, DOI 10.1007/978-3-319-07503-7__1
2 E. San Martín and P. De Boeck

where Ψ (x) = exp(x)/(1 + exp(x)), θ p is the effect of the person on the probability,
also called ability, and βi is the effect of the item on the probability, also called the
difficulty.
Different choices can be made for how the effects of the persons are considered.
Either the persons are modeled with fixed-effects (FE) or with random-effects (RE),
and for these random-effects one can either specify the distribution, for example,
the normal distribution (RE-N), or one can leave the distribution unspecified
(RE-U). The three models have led to three different likelihood functions (or
sampling probabilities) and, accordingly, to three different ways to estimate the
corresponding parameters: joint maximum likelihood (JML) in the case of the FE
model, parametric marginal maximum likelihood (MML) in the case of the RE-N
model, and semi-parametric MML in the case of the RE-U model.
It is our purpose to infer the consequences these choices have for the meaning of
the other parameter of the model, βi , the item difficulty. We will make this inference
from the likelihood for each of the three types of models. This approach is justified
by the fact that the likelihood function is supposed to generate the responses patterns
and, therefore, it provides the statistical meaning of the parameters indexing it; for
details, see Bamber and Van Santen (2000), Cox and Hinkley (1974), Fisher (1922),
and McCullagh (2002). Consequently, the inference proceeds in two steps:
1. One step for the identified parameterization, which is the parameterization as far
as possible on the basis of the likelihood.
2. Another step for the further identification of the parameter of interest, which
will require to establish an injective relationship (under constraints, if necessary)
between the parameter of interest and the identified parameterization.
It will be shown that the βi parameters have a different meaning in the three Rasch
models. The identified parameterization differs and also the identification of the
parameter of interest is different. Furthermore, we will also discuss a condition
based on marginal probabilities under which the difficulty of an item i is larger
than the difficulty of an item j; this empirical condition can also be related to the
empirical difficulty of an item (that is, the proportion of persons correctly answering
the item).
In order to differentiate between the three models, item difficulty parameters of
the three models will be denoted with different symbols: βiFE , βiRE-N , and βiRE-U . It will
be shown that the meaning of these three parameters depends on the choice that is
made for how to treat the effects from the part of the persons and the assumptions
one is making regarding these effects.

1.2 Fixed-Effects Specification

1.2.1 Assumptions

The specification of a Rasch model rests on the following two assumptions:

1 What Do You Mean by a Difficult Item? On the Interpretation of the. . . 3

Assumption 1: {Ypi : p = 1, . . . , N; i = 1, . . . , I} are mutually independent.

Assumption 2: For each person p and each item i, Ypi ∼ Bern(π pi ), where π pi =
Ψ (θ p − βiFE ) and Ψ (x) = exp(x)/(1 + exp(x)).

1.2.2 Likelihood and Identified Parameters

These assumptions induce the following likelihood function:

N I
P(θ 1:N ,β 1:I ) (Y1 = y1 . . . YN = yN ) = ∏ ∏ π pipi (1 − π pi)1−y pi
FE y

p=1 i=1

N I exp[y pi (θ p − βi )]
= ∏ ∏ 1 + exp(θ p − βi) ,
p=1 i=1

where Y p = (Yp1 , . . . ,YpI ) ∈ {0, 1}I , θ 1:N = (θ1 , . . . , θN ), and similarly for β 1:I .
The parameter of a Bernoulli distribution is identified. This fact, together with
Assumption 1, implies that the identified parameters are {π pi : p = 1, . . . , P; i =
1, . . . , I}.

1.2.3 Parameters of Interest

The problem now is to identify the parameter of interest (θ 1:N , β FE 1:I ), which means to
write them as functions of the identified parameters. From Assumption 2, it follows
that
 
π pi
θ p − βi = ln
FE
, p = 1, . . . , N; i = 1, . . . , I;
1 − π pi
 
1 − π pi π p j
βi − β j = ln , for all person p and i = j.
π pi 1 − π p j

These relationships show that {θ p − βiFE : p = 1, . . . , N; i = 1, , . . . , I} as well as

{βiFE − β jFE : i = 1, . . . , I, j = i, . . . , I} are identified since they are written as
functions of identified parameters. Therefore, the parameters of interest (θ 1:N , β FE 1:I )
are identified if one identification restriction is imposed. Two possibilities can be
considered:
1. To restrict one person parameter, namely θ1 = 0. Under this restriction, the
difficulty parameter becomes
4 E. San Martín and P. De Boeck

 
1 − π1i
βi = ln
FE
,
π1i

that is, the logarithm of the ratio between the probability that person 1 incorrectly
answers item i and the probability that person 1 correctly answers the item.
2. To restrict one item parameter, namely β1FE = 0. Under this restriction, the
difficulty parameter becomes
 
1 − π pi π p1
βiFE = ln , (1.2)
π pi 1 − π p1

that is, the logarithm of the odd ratio between item 1 and item i for each person p.
The first restriction depends on a specific person who is present in one application
of the test. Therefore, the second identification restriction is more convenient since
we may apply the same test to various sets of persons; see Andersen (1980).

1.2.4 Relationship of Item Difficulty with Empirical Difficulty

Regarding the relationship between item difficulty and empirical difficulty, from
(1.2) it follows that

βiFE > β jFE ⇐⇒ P(θ p ,β 1:I ) (Ypi = 1) < P(θ p ,β 1:I ) (Yp j = 1) for all persons p. (1.3)
FE FE

Thus, item i is more difficult than item j if the probability that the person correctly
answers item i is less than the probability that a person correctly answers item j.

1.2.5 Comments

The previous considerations lead to the following comments:

1. The fixed-effects specification of the Rasch model is a rather easy model from
the perspective of identification, easier than the other two specifications, and it is
therefore often implicitly used to interpret the parameters of the Rasch model,
even when one is interesting in is the random-effects specification; for more
discussion, see San Martín and Rolin (2013).
2. On the other hand, for an estimation of the model, mostly the assumption of a
random-effects model is made, because the maximum likelihood estimator of
the difficulty parameters is inconsistent due to the presence of the incidental
parameters. For details, see Andersen (1980), Ghosh (1995), and Lancaster
(2000).
1 What Do You Mean by a Difficult Item? On the Interpretation of the. . . 5

1.3 Random-Effects Specification

The random-effects assumption for the persons leads to consider the ability
parameters as realizations of an iid process. Using the statistical jargon, the person’s
abilities are now considered as random-effects.

1.3.1 Assumptions

A random-effects specification rests on the following assumptions:

Assumption 1: {Y p : p = 1, . . . , N} are mutually independent conditionally on
θ 1:N .
Assumption 2: For each person p, the conditional distribution of Y p given θ 1:N
only depends on θ p and it is parameterized by β RE-N
1:I .

Assumption 3: For each person p, {Ypi : i = 1, . . . , I} are mutually independent

conditionally on θ p . This is the so-called axiom of local independence.
Assumption 4: For each item i, (Ypi | θ p ) ∼ Bern [Ψ (θ p − βiRE-N )].
Assumption 5: θ p ’s are mutually independent and identically distributed, with a
common distribution N (0, σ 2 ).

1.3.2 Likelihood and Identified Parameters

These assumptions imply that the response patterns Y p ’s are mutually independent
and identically distributed. To describe the likelihood function, it is enough to
describe the probability of each of the 2I response patterns, namely

q12···I = P(β 1:I ,σ )

RE-N
(Yp1 = 1,Yp2 = 1, . . . ,Yp,I−1 = 1,YpI = 1),

q12···I = P(β 1:I ,σ )

RE-N
(Yp1 = 1,Yp2 = 1, . . . ,YpI, − 1 = 1,YpI = 0),
.. ..
. . (1.4)
q12···I = P(β 1:I ,σ )
RE-N
(Yp1 = 0,Yp2 = 0, . . . ,YpI, − 1 = 0,YpI = 0),

and
 ∞ I
exp[yi (σ θ − βiRE-N )]
P(β 1:I ,σ )
∏ 1 + exp(σ θ − β
RE-N
(Y p = y) = φ (θ ) d θ ,
−∞ i=1 i
RE-N
)
6 E. San Martín and P. De Boeck

with φ (·) as the density of a standard normal distribution. Therefore, the

likelihood function corresponds to a multinomial distribution Mult (2I , q), where
q = (q12···I , q12···I−1,I , · · · , q12···I ). Consequently, the parameter q is the identified
parameter.

1.3.3 Parameters of Interest

It is possible to prove that β RE−N

1:I and σ can be written in terms of the identified
parameter q without restrictions on the item parameters. The proof rests on the
following arguments:
1. Let
 ∞
. .
αi = P(β 1:I ,σ ) (Ypi = 1) =
RE-N
Ψ (σ θ − βiRE-N )φ (θ ) d θ = p(σ , βiRE-N ).
−∞

The parameter αi is an identified parameter because it is a function of q.

2. The function p(σ , βiRE-N ) is a strictly decreasing continuous function of βiRE-N .
Therefore it is invertible and consequently

βiRE-N = p−1 (σ , αi ). (1.5)

3. For i = j, let
 ∞
.
αi j = P(β 1:I ,σ ) (Ypi = 1,Yp j = 1) =
RE-N
Ψ (σ θ − βiRE-N) Ψ (σ θ − β jRE-N ) φ (θ )d θ .
−∞

The parameter αi j is also an identified parameter because it is a function of q.

Using (1.5), it follows that
 ∞
αi j = Ψ (σ θ − p−1 (σ , αi )) Ψ (σ θ − p−1 (σ , α j )) φ (θ ) d θ .
−∞
.
= r(σ , αi , , α j ).

It can be shown that r(σ , αi , α j ) is a strictly increasing continuous function of σ ;

for details, see San Martín and Rolin (2013). It follows that

σ = r−1 (αi j , αi , α j ). (1.6)

1.3.4 Relationship of Item Difficulty with Empirical Difficulty

Regarding the relationship between item difficulty and empirical difficulty, from
(1.5) it follows that
1 What Do You Mean by a Difficult Item? On the Interpretation of the. . . 7

βiRE-N > β jRE-N ⇐⇒ P(β 1:I ,σ )

(Ypi = 1) < P(β 1:I ,σ )
RE-N RE-N
(Yp j = 1) for all person p. (1.7)

Thus, item i is more difficult than item j if probability that a person correctly
answers item i is less than the probability that the person correctly answers item j.

1.3.5 Comments

The previous considerations lead to the following comments:

1. For each person p, the responses are positively correlated, that is,

cov(β 1:I ,σ )
RE-N
(Ypi ,Yp j ) > 0

for i = j. This is a marginal correlation and it follows from both Assumption 3

and the strict monotonicity of Ψ (θ p − βiRE-N ) as a function of θ p for all βiRE-N .
2. According to equality (1.6), σ represents the dependency between items i and
j induced by both the marginal probabilities αi and α j and the joint marginal
probability αi j . Furthermore, this dependency is the same for all pairs of items
since equality (1.6) is valid for all pairs of items i and j.
3. The item difficulty βiRE-N is not only a function of the marginal probability αi of
correctly answering the item i, but also of terms of the common dependency.
4. The previous identification analysis is valid in the case π pi = Φ (θ p − βiRE-N ),
where Φ is the distribution function of a standard normal distribution; see San
Martín and Rolin (2013). In this case, it is possible to show that
 
βiRE-N
αi = Φ − √ .
1 + σ2

Therefore, the difficulty parameter βiRE-N can be written as


βiRE-N = − 1 + σ 2 Φ −1 (αi ). (1.8)

It follows that the difficulty parameter βiRE-N is decreasing with σ . In other words,
the larger the individual differences, the more extreme the difficulty parameters
become.
5. There is not an explicit function as (1.8) for the logistic model, √ but approxi-
mately the same equation applies with σ 2 premultiplied by 16 3/(15π ); see
Molenberghs et al. (2010), Zeger et al. (1988).
6. The distribution of θ p can be specified as a N (μ , σ 2 ). In this case, the identified
.
parameters are (β 1:I , σ ), where βiRE-N = βiRE-N − μ . In order to identify the
RE-N

difficulty parameters β RE-N

1:I and the scale parameter μ , it is enough to introduce a
linear restriction on the item parameters β RE-N
1:I .
8 E. San Martín and P. De Boeck

1.4 Semiparametric Specification

As pointed out by Woods and Thissen (2006) and Woods (2006), there exist
specific fields, such as personality and psychopathology, in which the normality
assumption of θ p is not realistic. For instance, psychopathology and personality
latent variables are likely to be positively skewed, because most persons in the
general population have low pathology, and fewer persons have severe pathology.
However, the distribution G of θ p is unobservable and, consequently, though a
researcher may hypothesize about it, it is not known in advance of an analysis.
Therefore, any a priori parametric restriction on the shape of the distribution G could
be considered as a mis-specification.

1.4.1 Assumptions

These considerations lead to extend parametric Rasch models by considering the

distribution G as a parameter of interest, and thus specifying semi-parametric Rasch
models. These models rest on the following assumptions:
Assumptions 1–4 as in the random-effects specification.
Assumption 5: θ p ’s are mutually independent and identically distributed, with a
common unspecified distribution G.

1.4.2 Likelihood and Identified Parameters

As in the random-effects specification, these assumptions imply that the response

patterns Y p ’s are mutually independent and identically distributed, with a common
multinomial distribution Mult(2I , q), with q = (q12···I , q12···I−1,I , · · · , q1 2···I ), where

q12···I = P(β 1:I ,G)

RE-U
(Yp1 = 1,Yp2 = 1, . . . ,Yp,I−1 = 1,YpI = 1),

q12···I = P(β 1:I ,G)

RE-U
(Yp1 = 1,Yp2 = 1, . . . ,YpI, − 1 = 1,YpI = 0),
.. ..
. .
q12···I = P(β 1:I ,G)
RE-U
(Yp1 = 0,Yp2 = 0, . . . ,YpI, − 1 = 0,YpI = 0),

and
 I
exp[yi (θ − β RE-U )]
P (β RE-U
1:I ,G) (Y p = y) = ∏ 1 + exp(θ − βi G(d θ ). (1.9)
i )
RE-U
i=1
1 What Do You Mean by a Difficult Item? On the Interpretation of the. . . 9

Therefore, the likelihood function is parametrized by q, which corresponds to the

identified parameter.
Following San Martín et al. (2011), Equation (1.9) can be rewritten as follows:
for all J ⊂ {1, . . . , I} \ 0,
/
⎛ ⎞

P(β 1:I ,G) ⎝

{Yp j = 0}⎠ =
RE-U
{Yp j = 1} ∩
j∈J j∈J c
   ∞
e|J |θ
= exp − ∑ βj RE-U
× G(d θ ). (1.10)
∏ (1 + eθ −βi
RE-U
j∈J −∞ )
1≤i≤I

By taking (1.10) with J = {1} and after (1.10) with J = {i}, we identify ( β2RE-U −
β1RE-U , . . . , βIRE-U − β1RE-U ) because
⎡  ⎤
⎢ P(β 1:I ,G) {Yp1 = 1} ∩
RE-U
{Ypi = 0} ⎥
⎢ ⎥
⎢ 
2≤i≤I
 ⎥
β j − β1 = ln ⎢
RE-U RE-U
⎥. (1.11)
⎢ ⎥
⎣ P(β RE-U
1:I ,G)
{Yp j = 1} ∩ {Ypi = 0} ⎦
i= j

Not only the item differences can be identified, but also some characteristics of the
distribution G. As a matter of fact, working with the identified parameters β jRE-U −
β1RE-U leads to a shift of θ which we express with u = θ + β1RE-U . Thus, for all J ⊂
{1, . . . , I}, (1.9) can be rewritten as
⎛ ⎞
,G) ⎝
P(β 1:I {Yp j = 0}⎠ =
RE-U
{Yp j = 1} ∩
j∈J j∈J c

 ∞
−β1RE-U )] e|J |u
= e[− ∑ j∈J (β j
RE-U
  Gβ RE-U (du),
∏ 1 + eu−(βi −β1 )
RE-U RE-U
−∞ 1

1≤i≤I

.
where Gβ RE-U ((−∞, x]) = G((−∞, x + β1RE-U ]). Therefore, the functionals
1

 ∞
eku
mGβ RE-U (k) =   Gβ RE-U (du),
∏ 1 + eu−(βi −β1 )
RE-U RE-U
1 −∞ 1

1≤i≤I

for k = 0, 1, . . . , I, are identified. Note that mGβ RE-U (0) = mGβ RE-U (|0|)
/ corresponds
1 1
to P(β 1:I ,G) (Yp1 = 0, . . . ,YpI = 0). Summarizing, the following I + 1 relationships
RE-U

follow: For all J ⊂ {1, . . . , I} such that |J | = k,