Journal of Statistical Planning and Inference 143 (2013) 116130

Contents lists available at SciVerse ScienceDirect

Journal of Statistical Planning and Inference

journal homepage: www.elsevier.com/locate/jspi

Identication of parametric Rasch-type models

Ernesto San Martn a,b,c,n, Jean-Marie Rolin d
a

lica de Chile, Chile

Faculty of Mathematics, Ponticia Universidad Cato
lica de Chile, Chile
Faculty of Education, Ponticia Universidad Cato
c
lica de Chile, Chile
Measurement Center MIDE UC, Ponticia Universidad Cato
d
Institut de Statistique, Biostatistique et Sciences Actuarielles, Universite Catholique de Louvain, Belgium
b

a r t i c l e in f o

abstract

Article history:
Received 17 January 2011
Received in revised form
13 June 2012
Accepted 16 June 2012
Available online 26 June 2012

In modern Item Response Theory, the Rasch model is viewed as a Generalized Linear
Mixed Model, where the item parameters correspond to the xed-effects, whereas the
person specic parameters are the random-effects. The statistical model, bearing on the
observable variables only, is obtained after integrating out the random-effects.
Although it is widely accepted that the parameters of this model are identied, it is
hard to nd a correct justication. Furthermore, the meaning of the parameters of the
Rasch model as well as of its extensions is typically based on the xed-effects
specication of the model, that is, when the person specic parameters are also treated
as xed-effects. The contribution of this paper is to provide an explicit proof of the
identication of the random-effects Rasch model. The proof is valid for a large class of
Rasch-type models. It is also shown that such a proof can be applied to analyze the
identication of Explanatory Rasch Models. Finally, the meaning of the parameters of
interest with respect to the different data generating process is discussed.
& 2012 Elsevier B.V. All rights reserved.

Keywords:
Fixed-effects Rasch model
Random-effects Rasch model
Location-scale distributions
Generalized Linear Mixed Model
LLTM
MIRID
Latent regression models

1. Introduction
Generalized Linear Mixed Models (GLMM) are widely used in many elds, notably in psychometrics and biometrics.
The comprehensive books by De Boeck and Wilson (2004) and Verbekee and Molenberghs (2009) are relevant references.
This class of models is formulated in a hierarchical way as follows: let Ypi denote the i-th measurement available for the
p-th person, and let Y p denote the corresponding vector of all measurements. It is assumed that
Y p 9yp , b  F b ,

1:1

where F is a pre-specied distribution, parameterized through a vector b of unknown parameters, common to all persons.
Further, yp is a unidimensional variable of person specic abilities, called random effect, assumed to follow a so-called
mixing distribution G which is typically considered to be a member of a parameterized family, namely
iid

yp 9s  Gs :

1:2

In psychometrics, the Rasch model is an important example of this type of models (De Boeck and Wilson, 2004).
Dening b6b1 , . . . , bI , the conditional model (1.1) is specied as
Y pi 9yp , bi  BernCyp bi ,

p 1, . . . ,N; i 1, . . . ,I,

Corresponding author at: Faculty of Matehmatics, Ponticia Universidad Catolica de Chile, Vicuna Mackenna 4860, Macul, Santiago, Chile.
E-mail address: esanmart@mat.puc.cl (E. San Martn).

0378-3758/$ - see front matter & 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jspi.2012.06.014

1:3

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

117

where, for each person p, fY pi : i 1, . . . ,Ig are mutually independent conditionally on yi , b and Cx expx=expx 1.
In the psychometric literature, the conditional independence of the Ypis given yi , b is known as the assumption of local
independence (Lazarsfeld, 1954) and it is related with the hypothesis of common cause already used by Laplace (1774) in
the context of the rules of succession, and by Reichenbach (1956, Section 19) in quantum eld theory. For the marginal
model (1.2), it is typically assumed that the random effects yp s are iid N 0, s2 ; these random effects reect the betweenperson heterogeneity in the population with respect to the distribution of Y p (Molenbergs and Verbeke, 2004, p. 120).
This hierarchical structure is intended to model the measurements Y p through a statistical model induced after
integrating out the random effects, namely
PY p y9b, s

Z Y
I
Ri1

Csybi yi 1Csybi 1yi fy dy,

1:4

where y 2 f0; 1gI and f corresponds to the density function of a standard normal distribution.
It is widely accepted that the parameters b, s of the statistical model (1.4) are identied, although it is hard to nd a
correct justication. Most of the standard justications reduce to the following statement: the identication restriction
needed to identify the parameters of (1.3), where the person specic ability yp is treated as a xed effect, is sufcient to
identify the parameters of (1.4), where yp is treated as a random-effect; see, among others, Adams et al. (1997), Smits and
Moore (2004), Rijmen and De Boeck (2005), Adams and Wu (2007). Moreover, the statistical meaning of the item
parameters in the marginal Rasch model (1.4) is typically based on its meaning in the conditional Rasch model (1.3); see,
for instance, Embretson and Reise (2000). It should be said that this interpretation is coherent with estimation procedures
such as the conditional maximum likelihood or joint marginal likelihood, but not with the marginal maximum likelihood
which is based on (1.4); for a summary on these estimation procedures, see Molenaar (1995). Taking into account that the
distinction between models (1.3) and (1.4) stems from the way in which person specic abilities are treated, in what
follows (1.3) is called xed-effects Rasch model and (1.4) is called random-effects Rasch model.
In spite of those heuristic justications, San Martn et al. (2011, Section 3) have recently shown that there not exist a
general relationship between the identication of the xed-effects Rasch model and the random-effects Rasch model. This
suggests to explore how different are the meaning of the parameters in both specications. Since parameter interpretation
should be based on an identication analysis (Rasch, 1966a; San Martn et al., 2009), this paper is rst focused in providing
an explicit proof of the identication of the random-effects Rasch model. The proof is done with respect to a large class of
random-effects Rasch-type models in which the logistic distribution in (1.3) is replaced by a general cumulative
distribution function (cdf) F, and the marginal model generating the person specic abilities corresponds to a distribution
G known up to the scale s and location m parameters. Thus, the central question to be discussed in this paper can be
phrased as follows: under which conditions on the cdfs F and G, the item parameters b and the location-scale parameters
m, s are identied by the observations? Once this question is answered, the interpretation of the parameters is explored.
This parameter interpretation is compared with both the meaning of the parameters in a xed-effects Rasch model.
A second topic to be discussed in this paper is the extension of the identication analysis of Rasch-type models to two
particular cases, namely when the item parameters and the parameters of the distribution generating the random effects
are restricted. Following De Boeck and Wilsons (2004) terminology, these models are termed explanatory IRT-models at the
item side and explanatory IRT-models at the person side, respectively. It is shown how different is the identication analysis
if it is done either in the context of a xed-effects specication or in the context of a random-effects specication.
This paper is organized as follows: in Section 2, the identication strategy used in this paper is explained. The
identication analysis of random-effects Rasch-type models is developed in Section 3. In Section 4, these results are
extended to some explanatory IRT-models, which are of interest in psychometrics. The paper concludes with a discussion.

2. Identication strategy
2.1. Random-effects Rasch-type models
Let Y p Y p1 , . . . ,Y pI 0 denote the vector of all measurements available for person p, where Y pi 1 if person p correctly
answers item/probe i, and Y pi 0 otherwise. Rasch-type models assume that the distribution of Ypi conditionally on yp , bi
is given by
PY pi 19yp , bi  Fyp bi ,

p 1, . . . ,N; i 1, . . . ,I,

2:1

where F is a known continuous strictly increasing cumulative distribution function (cdf). If, for instance, F F, then (2.1)
corresponds to the probit model; if F is the logistic distribution, (2.1) corresponds to the Rasch model; see Rasch (1966a). It
is also assumed that (H1) for each person p, his/her measurements are mutually independent conditionally on yp , b,
where b b1 , . . . , bI 0 (the assumption of local independence); (H2) the person specic random effects yp are mutually
independent and identically distributed, with a common distribution Gs dened as

x i
,
2:2
Pyp r x9s Gs 1,x6G 1,

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

118

where G is an arbitrary cumulative distribution function dened on R, known up to the scale parameter s 4 0. The model
is completed by assuming that (H3) Y 1 , . . . ,Y N are mutually independent conditionally on y1 , . . . , yN , b.
The statistical model, bearing on the observable variables only, is obtained after integrating out the person specic
random effects. It can be shown that H1, H2 and H3 jointly imply that the measurements Y 1 , . . . ,Y N are mutually
independently given b, s, with a common probability distribution
Z Y
I
Fsybi ypi 1Fsybi 1ypi Gdy,
2:3
PY p yp 9b, s
Ri1

where yp yp1 , . . . ,ypI 2 f0; 1gI . This statistical model (hereafter called random-effects Rasch-type model) is based on a
hierarchical specication in which each person has his/her own unique effect (hypothesis H2), unexplained by person
properties; and each item has its own unique effects (namely bi ), unexplained by item properties. A relevant characteristic
of the statistical model (2.3) is that, for each person p, the measurements Y p1 ,Y p2 , . . . ,Y pI are correlated provided that the
link function F is a strictly increasing monotonic function (Sijtsma and Molenaar, 2002). This property is valid for each
(non-degenerate) cdf G generating the random effects.
0

2.2. Identication strategy

The identication of b, s by one observation Y p is entirely equivalent to their identication by an innite sequence of
measurements Y 1 ,Y 2 , . . . . This is due to the fact that, in the statistical model (2.3), the observations are mutually
independent and identically distributed. The identication problem we are dealing with corresponds, therefore, to the
injectivity of the mapping b, s/P9b, s, where P9b, s is given by (2.3). To solve this problem, we follow an
identication strategy based on two steps:
1. A distinction between parameters of interest and identied parameters should be made. Such a distinction is based on the
fact that a statistical model always involves an identied parametrization; for details, see San Martn et al. (2009, 2011,
in press).
2. An injective relationship (under restrictions, if necessary) between the parameters of interest and the identied
parameters should be established. By so doing, the parameters of interest become identied by the observations.
In the case of the statistical model (2.3), the parameters of interest are b, s. Now, the probabilities of the 2I different
possible patterns are given by
q12I PY p1 1, . . . ,Y p,I1 1,Y pI 19b, s,
q12I PY p1 1, . . . ,Y p,I1 1,Y pI 09b, s
^
q12I PY p1 0, . . . ,Y p,I1 0,Y pI 09b, s:
The statistical model (2.3) corresponds to a multinomial distribution Y p 9q  Mult2I ,q, where q q12I , q12I1,I , . . . ,
q1,2,...,I . It is known that the parameter q of a multinomial distribution is identied since the mapping q/Mult2I ,q is
injective. It corresponds, therefore, to the identied parameter indexing the statistical model (2.3); the qs with less than I
subscripts are linear combinations of them and, consequently, are also identied. Thus, the identiability of the
parameters of interest b, s is obtained after establishing an injectivity relation between them and functions of q. By
this way, the restrictions under which the identiability of b, s is obtained are not only sufcient conditions, but also
necessary conditions.
This strategy is followed in the rest of the paper. Let us mention that this identication strategy has already been used
in psychometrics (Muthen, 1979; San Martn et al., in press) and in biometrics (Rabe-Hesketh and Skrondal, 2001).
3. Identication of random-effects Rasch-type models
3.1. Strategy of the identication analysis
In this section, we establish the identication of b, s in the random-effects Rasch-type model (2.3) for arbitraries cdfs
F and G. The identication analysis is based on the following two steps:
Step1: It is shown that the item parameters b are a function of both the scale parameter s and the identied parameters
PY pi 19b, s for i 1, . . . ,I.
Step2: It is shown that the scale parameter s is a function of the identied parameters PY p1 19b, s, PY p2 19b, s and
PY p1 1,Y p2 19b, s.
Combining both steps, b, s becomes identied by one observation.

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

119

3.2. Proof of Step 1

Assume that F is a strictly increasing cdf. For all i 1, . . . ,I, dene
Z
ai 6PY pi 19b, s Fsxbi Gdx,

3:1

which is an identied parameter because it is a function of the identied parameter q dened in Section 2.1. Clearly, the function
Z
3:2
ps, b6 FsxbGdx
R

is a continuous function in s, b 2 R0  R that is strictly decreasing in b since F is a strictly increasing continuous

function; here R0 denotes the positive real line. Therefore, if we dene
ps, a6inffb : ps, b o ag,

3:3

it follows that
ps,ps, b b:

3:4

Using (3.1), this implies that, for each i 1, . . . ,I, bi ps, ai ; that is, for each i 1, . . . ,I, the item parameter bi is a function
of both the scale parameter s and the identied parameter ai .
3.3. Proof of Step 2
Suppose that I Z 2 and let

a12 6PY p1 1,Y p2 19b, s

Z
R

Fsxb1 Fsxb2 Gdx:

Using Step 1, the identied parameter a12 can be written as a function of s, a1 and a2 , namely
Z
a12 6qs, a1 , a2 Fsxps, a1 Fsxps, a2 Gdx:

3:5

3:6

If F is a strictly increasing cdf with a continuous density function strictly positive on R, then it can be proved that the
function a12 qs, a1 , a2 is a strictly increasing continuous function of s. It follows that s qa12 , a1 , a2 , where
qa, a1 , a2 inffs : qs, a1 , a2 4 ag:
In other words, s becomes a function of identied parameters and, therefore, it is identied by the observations. The
details are provided in Appendix A.
3.4. Main result
Summarizing, we obtain the following theorem:
Theorem 3.1. Consider the class of random-effects Rasch-type models (2.3), where F is a continuous strictly increasing cdf, with
a continuous density function strictly positive on R. If at least two items are available, then the item parameters b and the scale
parameter s of the distribution G generating the person specic abilities are identied by one observation Y p .
This theorem deserves the following comments:

1. Theorem 3.1 is valid for all distribution functions Gs known up to the scale parameter s, including the discrete
distributions.
2. Theorem 3.1 establishes the identication of both the item parameters and the scale parameter when F is a continuous
strictly increasing distribution function, with a continuous density function strictly positive on R. This is the case when,
for instance, the link function F is either a standard normal distribution, a logistic distribution or a complementary
loglog distribution.
3. In particular, for the probit model, the identication of the scale parameter of the normal N 0, s2 is well established,
proving thus a conjecture due to Chen and Dey (1998, p. 352). For the logit link function, Theorem 3.1 offers a rigorous
proof of the identiability of both the item parameters and the scale parameter in a random-effects Rasch model; it is
particularly shown that, when G is a normal distribution, the identication restriction corresponds to x at zero the
mean of the distribution generating the random effects.
4. Theorem 3.1 does not apply when an explanatory IRT-model is dened with a complementary-log link function dened as
(
FZ

1expZ,
0,

Z 40,
Z r0:

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

120

In contrast to the above-mentioned link functions, this one may not be differentiable at Z 0 (see Piegorsch, 1992; Fahrmeir
and Tutz, 2001), and its density function is positive on R only.
5. It should be remarked that Theorem 3.1 is still valid if all the items parameters are equal between them, that is, if bi b
for all i 1, . . . ,I and I Z2. This is not the case when the identication analysis is done in the context of a xed-effects
specication of the Rasch model, that is, when yp is treated as a xed effect and the statistical model is characterized by
both (2.1) (with F the logistic distribution) and the mutual independence of fY pi : p 1, . . . ,N; i 1, . . . ,Ig. In this case,
the identied parameters are fyp bi : p 1, . . . ,N; i 1, . . . ,Ig, whereas the parameters of interest are y1 , . . . , yN , b.
Under the linear restriction
a0 b 0

for a known a 2 RI such that a0 1I a0,

3:7

where 1I 1, . . . ,10 2 RI , the parameters of interest become identied by the observations. This condition implies that
b 2 /aS? and 1I2
= /aS? , where /aS denotes the linear space generated by the vector a. In other words, b belongs to a
linear space which excludes equal difculties for all the items. Thus, in a xed-effects specication of the Rasch model,
the identication restriction not only xes the scale of the parameters of interest, but also imposes a specic design on
the multiple-choice test in the sense that it should contain at least two items with different difculties.

3.5. Identication when G is known up to both the location and the scale parameters
Random-effects Rasch-type models can also be specied assuming that the distribution generating the random effects
corresponds to a location-scale distribution, that is,
Pyp r x9m, s Gm, s 1,x6G


xmi
:
1,

It is assumed that G is known up to both the location parameter m 2 R and the scale parameter s 4 0. In this case, the
parameters of interest are b, m, s and, consequently, the identication problem consists in identifying these parameters by
the observations. This problem can be solved as a corollary of Theorem 3.1. As a matter of fact, the statistical model is now
characterized by
PY p yp 9b, m, s

Z Y
I
Ri1

Fsyb~i ypi 1Fsyb~i 1ypi Gdy,

3:8

where yp yp1 , . . . ,ypI 0 2 f0; 1gI and b~i 6bi m. In particular, it holds that
Z
a~i Fsxb~ i Gdx, i 1, . . . ,I,
R

a~ 12

Fsxps, a~ 1 Fsxps, a~ 2 Gdx,

3:9

where ps, a~ is given by (3.3). These equations are similar to Eqs. (3.1) and (3.6), which were derived in the context of a
random-effects Rasch-type model with G known up to the scale parameter s. Theorem 3.1 ensures, therefore, that
b1 m, . . . , bI m, s are identied by one observation. Under a linear restriction of the form (3.7), an injective relationship
between those identied parameters and the parameters of interest b, m, s is established. Summarizing, we obtain the
following corollary:
Corollary 3.1. Consider the class of random-effects Rasch-type models, where F is a continuous strictly increasing cdf, with a
continuous density function strictly positive on R, and the person specic random effects are distributed according to a locationscale distribution Gm, s . If at least two items are available, then b, m, s are identied by one observation Y p provided that a0 b 0
for a known I-dimensional vector a such that a0 1I a0.
It should be remarked that if the random-effects Rasch-type models are specied assuming that yp 9m, s  Gm, s , then
the multiple-choice test should include at least two items with different difculties. It should also be mentioned that in the
psychometric literature there are contributions stating that, when the distribution generating the random-effects is an
N m, s2 , the parameters of the random-effects Rasch model (1.4) are identied provided that m is xed at 0 or a particular
item parameter is xed at 0; see Bechger et al. (2003, p. 328) and De Boeck and Wilson (2004, pp. 5354). Theorem 3.1 and
Corollary 3.1 make precise those statements.

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

121

3.6. Parameter interpretation for different specications of the Rasch model

3.6.1. Statistical meaning of a parametrization
A statistical model corresponds to a family of sampling distributions indexed by a parameter (see Fischer, 1922; Basu,
1975; Cox and Hinkley, 1974; McCullagh, 2002). The data generating process is characterized by those sampling
distributions. In agreement with the likelihood principle, the statistical meaning of a parametrization should be based
on the sampling information which in turn is captured by the data generating process (Basu, 1975). Consequently, the
relationship between sampling information and parametrization should be one-to-one (that is, the parametrization should
be identied) and, consequently, the likelihood surface as a function of the parameters is not constant. This mean not only
that the statistical model does not include redundant parameters (Catchpol and Morgan, 1997, Theorem 2), but also that
unidentied parameters cannot be statistically interpretable.
Let us illustrate these considerations with the following two examples:
Example 1. Consider an ANOVA One Way Model specied as follows:
X pi 9mi , s2  Nmi , s2 ,

mi a yi ,

with p 1, . . . ,ni and i 1, . . . ,I. The statistical model is indexed by x6a, y1 , . . . , yI , s2 2 R  Rm  R . Without
identifying restrictions, these parameters are not statistically interpretable. Their statistical interpretation depend, as it
is well known, on specic identifying restrictions. Thus, if y1 0, then a m1 EX 1 9x and yi mi m1 EX i 9xEX 1 9x.
That is, a corresponds to the mean of the rst population, whereas yj corresponds to the difference between the mean
of j-th population and the mean of the rst population.
P
P
However, if we consider the restriction Ii 1 yi 0, then a m 1=I Ii 1 EX i 9x and yi mi m EX i 9xm . That is,
a corresponds to the general mean, whereas yi corresponds now to the mean of the i-th population corrected by the
general mean. It can therefore be seen that the interpretation of the parameters depends on the identifying restrictions;
once a specic restriction is adopted, the sampling information provides the meaning to the parametrization. &
Example 2. Consider a random-effect ANOVA One Way model specied as follows:
c:iid

X pi 9yi , x  N a yi , s2 ,
iid

yi 9x  N 0, t2 ,

p 1, . . . ,ni , i 1, . . . ,I,

i 1, . . . ,I,

3:10

where x a, s2 , t2 and c.iid is described through the hypotheses H1, H2 and H3 introduced in Section 2.1. The sampling
process corresponds to a IN-variate normal distribution of mean equal to a1IN and a variancecovariance matrix dened as
8 2
0
2
0
>
< s t , p p , i i ,
0
pap0 , i i ,
covX pi ,X p0 i0 9x t2 ,
>
:
0,
pap0 , iai0 ,
P
where N Ii 1 ni . The parameter x is actually indexing the sampling process. Its identiability can easily be veried.
The statistical meaning of x is based on the sampling process. Thus, for instance, the parameter t2 corresponds to the
covariance of observations Xpi and X pi0 nested into the p-th group, and not to the variance of the random-effects yj (Snijders
and Bosker, 1999). In order to emphasize the fact that the meaning of the parameters (in this example, the meaning of t2 )
depends on the sampling process, consider the following alternative specication:
c:iid

X pi 9yi , x  N a tyi , s2 ,
iid

yi 9x  N 0; 1,

p 1, . . . ,ni , i 1, . . . ,I,

i 1, . . . ,I,

3:11

with t 4 0. This specication induces the same statistical model as specication (3.10). The meaning of t2 is, therefore, still
ensured by the data generating process. Specications (3.10) and (3.11) are equivalent models since they induce the same
statistical model; for a denition of equivalent models, see Maris and Bechger (2004). &
3.6.2. Application to the Rasch model
As mentioned above, in the psychometric literature, the Rasch model is specied either as a xed-effects model or as a
random-effects model. Considering the identication analysis in both contexts, it is possible to understand how different is
the statistical meaning of the corresponding identied parameters. More precisely,

1. In a xed-effects Rasch model, the parameters of interest b, y1 , . . . , yN are identied by the observations if and
only if b1 0. Under this restriction, yp represents the logarithm of the betting odds of a correct answer to the

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

122

standard item 1, that is,

"
#
PY p1 19yp , b1 0
yp ln
:
PY p1 09yp , b1 0
Similarly, the item parameter bi corresponds to the logarithm of the odds ratio between item 1 and item i for each
person p, that is,
"
#
PY p1 19yp , b1 0 PY pi 19yp , bi

:
bi ln
PY p1 09yp , b1 0 PY pi 19yp , bi
The statistical meaning of bi depends, therefore, not only on the item i, but also on the standard item 1. For a discussion,
see Rasch (1966a,b) and San Martn et al. (2009).
2. However, the statistical meaning of the parameters of interest b, s in a random-effects Rasch model is quite different.
As a matter of fact,
(a) As derived in Section 3.3, the scale parameter s is given by qa12 , a1 , a2 , where ai (with i1,2) corresponds to the
marginal probability that each person p answers item i correctly (see Eq. (3.1)), and a12 corresponds to the marginal
probability that each person p answers both items 1 and 2 correctly (see Eq. (3.5)). Thus, in a random-effects Rasch
model, s represents the dependency between items 1 and 2 induced by the marginal probabilities a1 and a2 , and
the joint probability a12 . As the reader can recognize, this is very similar to the interpretation of the variance of the
random effects in a hierarchical linear mixed model; see Example 2.
(b) As derived in Section 3.2, the item parameter bi is given by bi ps, ai and, therefore, represents a function of both
the marginal probability ai and the dependency between items 1 and 2 as measured by s. In other words, in a
random-effects Rasch model, the item parameter not only captures information on the item itself, but also on items
1 and 2 through their dependency, whereas in a xed-effects Rasch model, the item parameter provides
information on both the item itself and the standard item without any type of dependency between them.
4. Identication of explanatory-IRT-Models
If the parameters of a random-effects Rasch-type model are restricted, we obtain explanatory Rasch-type models. Three
types of restrictions can be introduced: (i) explanatory Rasch-type models at the item side, which are obtained by restricting
the item parameters; (ii) explanatory Rasch-type models at the individual side, which are obtained after restricting the
parameters of the distribution generating the person-specic abilities; and (iii) doubly explanatory Rasch-type models,
which are obtained after combining the previous two specications. In this section, we show how the identication of the
parameters of interest of explanatory Rasch-type models can be obtained from the identication of random-effects Raschtype models. Furthermore, differences between identication results obtained under a random-effects specication and
under a xed-effects specication are discussed.
4.1. Identication of explanatory Rasch-type models at the item side: LLTM-type models
Following Fischer (1983) and De Boeck and Wilson (2004), in the linear logistic test model (LLTM), item properties are
used to explain differences between items in terms of the effect they have on the probability to answer correctly the items.
It is actually assumed that the difference between the item difculty parameters bi and bj can only be due to structural
characteristics of the respective items, that is, to the cognitive structure of the tasks underlying the solution of the item.
This leads to restrict the item parameters bi as

bi Q 0i c c, i 1, . . . ,I,

4:1

where Q i is a K-dimensional design vector. The entries of the vector Q i can be 0 and 1 (denoting the hypothetical absence
or presence of an operation or task in solving item i), integers (denoting the hypothetical minimum number of times an
operation or task has to be used in solving item i), or in general real numbers. The constant c is viewed as an intercept
(De Boeck and Wilson, 2004, p. 62) and corresponds to the differences
X
bi bj
Q i k Q j k gk 6c,
1rkrK

where Q i k is the k-coordinate of the vector Q i . This equality means that the difference of each of the two item parameters
is explained as the sum of the difculty parameters of those cognitive operations which have to be performed in solving
item i but not in solving item j, and vice versa; see Fischer (1983, p. 5). The specication of the model is completed by
assuming that the person abilities are distributed according to a distribution Gs dened by (2.2).
In this context, the induced explanatory Rasch-type model is parameterized by c,c, s; these are the parameters of
interest. Theorem 3.1 ensures, however, that b, s are identied by the observations, where b b1 , . . . , bI 0 . Consequently,
the parameters of interest c,c, s become identied if an injective relationship between them and the identied
parameters b, s is established. To do it, let Q be an I  K matrix, where its i-th row is the vector Q 0i . Equality (4.1) can

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

equivalently be rewritten as
 
c
Q 91I
b,
c

123

4:2

where 1I 1, . . . ,10 2 RI . Therefore, an injective mapping between b, s and c,c, s can be dened if the rQ 91I  K 1.
This condition is equivalent to the following two conditions: (i) rQ K and (ii) 1I2
= ImQ , that is, 1I does not belong to the
linear space generated by the columns of Q . Summarizing, we obtain the following corollary:
Corollary 4.1. Consider explanatory Rasch-type models as specied in Section 2.1, where the link function F is a continuous
strictly increasing cdf, with a continuous density function strictly positive on R; the person specic abilities are distributed
according to a scale distribution Gs ; and the item parameters b are restricted according to (4.2). Then
1. A necessary condition to identify c,c, s by one observation is that K r I.
2. If at least two items are available, c,c, s are identied by one observation provided that rQ K and 1I2
= ImQ .
Example 3. The arguments underlying Corollary 4.1 can be illustrated through the following example: consider K 3 tasks
and I4 items related according to the design matrix
0
1
1 1 1
B0 1 0C
B
C
Q B
C:
@0 0 1A
1

Theorem 3.1ensures that b1 , b2 , b3 , b4 , s are identied by Y p . Now, the task parameters g1 , g2 and g3 , as well as the
intercept c are the parameters of interest and, therefore, should be identied by the observations. It can be veried that
rQ 3 and that 142
= Im Q ; therefore, the matrix Q 914 is non-singular and the parameters of interest g1 , g2 , g3 ,c can be
written as a function of the identied parameters b1 , b1 , b3 , b4 :
1
0 1 0
10 b 1 0
b2 b4
0 1 0
1
g1
1
B
C
B
Bg C B 1
b2 C B b1 b2 b3 b4 C
C
1 1 1 C
B 2C B
CB
CB
C:
B CB
CB
B
C
B
C
b
b

b
@ g3 A @ 1
A
0
0 1 @ 3 A @
1
4
A
b

b

b
c
1 0
1
1
4
1
3
4
Thus, g1 , g2 , g3 ,c are identied because they are functions of identied parameters.

&

In the hierarchical specication considered in Corollary 4.1, the mean of the person specic random effect is 0. An
alternative specication is to consider the location parameter of the distribution G equal to c and, therefore, to omit the
contribution of the constant predictor in the explanatory part of the model (see, e.g., De Boeck and Wilson, 2004, p. 62);
that is, to assume that
iid

i yp 9c, s  Gc, s ,

ii b Q c,

4:3

where Q is an I  K design matrix.

This hierarchical specication as well as the specication considered in Corollary 4.1 are equivalent models in the sense
that they induce the same explanatory Rasch-type model; for a denition of equivalent models, see Maris and Bechger
(2004). However, the identication analysis is not the same. As a matter of fact, if at least two items are available, Corollary
3.1 ensures that b,c, s are identied by one observation provided that a linear restriction of the form a0 b 0, with a a
known I-dimensional vector such that a0 1I a0. By (4.3)(ii), this identication restriction is rewritten as a0 Q c 0. It remains
to identify c. From (4.3)(ii), it follows that c is identied if rQ K. Summarizing, we obtain the following corollary:
Corollary 4.2. Consider explanatory Rasch-type models as specied in Section 2.1, where the link function F is a continuous
strictly increasing cdf, with a continuous density function strictly positive on R; the person specic abilities are distributed
according to a location-scale distribution Gc, s ; and the item parameters b are restricted as in (4.3)(ii). Then
1. A necessary condition to identify c,c, s by one observation is that K r I.
2. If at least two items are available, c,c, s are identied by one observation provided that rQ K and a0 Q c 0 for a known
I-dimensional vector a such that a0 1I a0.
Example 4. Let us illustrate this corollary considering the design matrix Q of Example 3. In this case, the identied
parameters are b~ 6b1 c, b2 c, b3 c, b4 c0 , whereas the parameters of interest are g1 , g2 , g3 ,c. Combining the linear
restriction under which the mapping b~ /b,c is injective, along with restriction (4.3)(ii), we obtain an equation from

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

124

which the parameters of interest can be identied, namely

! 
 Q 0 ! 
 
c
c
I4 14
4
b~
6AQ~
,

0
0
03 1
c
c
a
0
0
where 04 0; 0,0; 00 , 03 0; 0,00 and a 2 R4 is a known vector such that a0 14 a0. The latter condition ensures that A is a
full rank matrix. Furthermore, rQ~ rQ 1. Therefore,
!
 
~
c
1 b
~
Q A
,
c
0

where Q~ is MoorePenrose inverse of Q (Marsaglia, 1964). For instance, if a 1; 0,0; 00 , then

0
1
01
1
b~
0 1 0 2
1
2
1 1B 1 C
2b1 3b2 b3 2b4
g1
1


3
3
3
3
B b~ 2 C B 3
C
B g C B 1 1
B
C
b1 b2
C
0
0
1 C
B 2C B
CB ~ C B
C:
&
B CB 1
CB b 3 C B
1
B
2
1
2
b1 2b3 b4 C
@ g3 A @  3 0
A
B
C

3
@
A
3
3
3
B~ C
1
b
@
A
4
 2 b1
c
1 0
0
0
1
0

Corollaries 4.1 and 4.2 deserve the following comments:

1. The choice between using Corollary 4.1 or 4.2 is in principle arbitrary. However, if the design matrix Q is dened in
such a way that 1I 2 ImQ , as for instance
0
1
1 0 0
B0 1 1C
B
C
Q B
C,
@0 0 1A
1

then Corollary 4.1 does not apply and the identication of c,c, s follows from Corollary 4.2.
2. In Corollary 4.2, if the vector a is chosen in such a way that a 2 KerQ 0 (i.e., a belongs to the null space of matrix Q 0 ),
then the identication restriction a0 Q c 0 is automatically satised provided that 1I2
= KerQ 0 ? ImQ (Halmos, 1987,
Section 49). As an example, consider the design matrix Q of Example 3. We know that 142
= ImQ . Furthermore, the
KerQ 0 is generated by the vector 1; 0,1,10 . Thus, if the vector a is chosen equal to 1; 0,1,10 , then
0 1
1 b1
0b
1
0 1 01
b3
2b4
1
2
1
g1
1

B C
3 b2  3 3
3
3
3
CB b2 C B
C
Bg C B
C:
B C B
b2
0
0 C
@ 2AB
@0 1
AB b C @
A
3
@
A
1
2
1
b1
2b3
b4
g3
0



3
3
3

b4

3. Let us consider a xed-effects specication of LLTM-type models. The parameters of interest y1 , . . . , yN , g1 , . . . , gK ,c are
identied under two identication restrictions: (i) the item parameters b satisfy a linear restriction of the type (3.7) and
(ii) rQ 91I  K 1 (which is equivalent to rQ K and 1I2
= ImQ ); see Fischer (1983, 1995, 2004). However, in a
random-effects specication, Corollary 4.1 shows that it is possible to identifying the parameters of interest without
introducing a linear restriction on the item parameters b.
4. Under a xed-effects specication, Bechger et al. (2002) consider the identication problem of a LLTM-type model
when the item parameters b are restricted as b Q c. In this context, they establish that c is identied by the
observations if the rank of the matrix II 1I a0 Q is equal to K, where the I-dimensional known vector a satises the
equality a0 1I 1. Under a random-effects specication, an argument similar to that of Corollary 4.1 leads to prove that c
is identied provided that rQ K only.

4.2. Identication of explanatory Rasch-type models at the item side: MIRID-type models
Butter et al. (1998) (see also Smits et al., 2003) described the model with internal restrictions on item difculties (MIRID)
as a componential model for binary data. In the MIRID, the parameters of some items are dened to be a linear
combination of the parameters of other items. Two sets of items should be dened: composite items and component items.
A composite item is an item that measures a concept that can be decomposed into components. A component item is an
item that measures one of these components. The item parameters of the composite items are decomposed into parts
attributed to the component items (the item parameters of the component items). As a simple example, Smits et al. (2003)

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

125

propose the following one: 10  (53) as a composite item has two component items: 5 3 and 10  8. The rst
component item is of the addition type; the second is of the multiplication type. For relevant psychological applications of
the MIRID model, see Smits and De Boeck (2003).
In general, J item families j 1, . . . ,J are dened so that, within each family, there is one composite item, to be
conceived as a dependent variable, and K components items, to be conceived as the independent variables. For the
composite items, subscript k is set to zero. The total number of items is JK 1. The crucial assumption of the MIRID is that
the item parameters of a composite item is a linear function of the item parameters of the associated component items:

bj0

K
X

ok bjk t for j 1, . . . ,J,

4:4

k1

where bj0 is the item parameter of the composite item from item family j, bjk is the item parameter of the component item
of type k from item family j, ok is the weight of the component item parameters of type k in determining the composite
item parameters, and t is a normalization constant.
Following Smits and De Boeck (2003), in an IRT-type model, each item (composite and component) has its own item
parameter, so that the conditional probability that person p will give a correct answer to item jk is given by
PY pjk 19yp , bjk  Fyp bjk ,

j 1, . . . ,J, k 0, . . . ,K:

4:5

Hypotheses H1, H2 and H3 as introduced in Section 2.1 are assumed to be valid for this class of models, and the random
effects yp s are iid Gs .
The corresponding explanatory Rasch-type model is indexed by b~ , x, t, s, where b~ is a JK 1-dimensional vector
containing the item parameters (composite and component), and x o1 , . . . , oK . From Theorem 3.1 it follows that b~ , s
are identied by one observation if at least two items are available. It remains to identify x, t. But (4.4) can equivalently
be rewritten as
 
x
b0 B91K
,
4:6

where b0 b10 , . . . , bJ0 0 and B is a J  K matrix dened as

0
1
b11    b1K
B

^ C
B@ ^
A,
bJ1    bJK
that is, a matrix of component item parameters. Therefore, x and t are identied if rB91K K 1. In this case, x and t can
be written as a function of the item parameters which in turn are identied by one observation. The rank identication
restriction is equivalent to the following two conditions: rB K and 1K2
= ImB. This implies that J Z K 1. Now, Theorem
3.1 requires that at least two items are available. It follows, therefore, that J Z2 and accordingly K Z 1.
Summarizing, we obtain the following corollary:
Corollary 4.3. Consider explanatory Rasch-type models as specied in Section 2.1, where the conditional probability that a
person p correctly answers the item jk is given by (4.5); the link function F is a continuous strictly increasing cdf, with a
continuous density function strictly positive on R; the person specic abilities are distributed according to a scale distribution Gs ;
and the item parameters b~ are restricted according to (4.4). Then
1. A necessary condition to identify b~ , x, t, s by one observation is that J Z K 1, with J Z2 and K Z1.
2. If J Z2 and K Z 1, b~ , x, t, s are identied by one observation provided that rB K and 1K2
= ImB.

This corollary deserves the following comments:

1. Corollary 4.3 shows that in identifying b~ , x, t, s by one observation, it is not sufcient to x the location of G at zero
(when G F, this is equivalent to x the mean of the distribution of the person abilities at zero), but two additional
identication restrictions should also be considered, namely rB K and 1K2
= ImB. Thus, Corollary 4.3 complements
the identication restriction suggested by Smits and Moore (2004, p. 272).
2. For the minimal case J2 and K 1, the sufcient identication restrictions reduce to ensure that b11 ab21 . In this case,
o1 and t can easily be written as a function of the item parameters:

o1

b10 b20
b b b21 b10
, t 11 20
:
b11 b21
b11 b21

As expected, o is actually a slope: if o 4 1 (resp. o o 1), then the distance between the two composite items is larger
(resp. smaller) than the distance between the two component items.

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

126

3. The k-th column of matrix B is a J-dimensional vector whose coordinates correspond to the k-th component item of
each family of composite items. Thus, for instance, if rB K1, then the rst column b1 can be written as a linear
P
combination of the remaining columns of B, that is, b1 Kk 2 lk bk , where at least one lks is different from zero. In
such a case, for each family of composite items, it follows that

bj0

K
X
k1

ok b k t

K
X

ok lk bk t:

k2

Therefore, the identication restriction rB K ensures that each composite item is decomposed into K component
items only. This is also the meaning of the restriction 1K2
= ImB. In other words, the information provided by each
component item is exhaustive in the sense that it is not recovered from the other component items, neither from the
intercept.
4. In a xed-effects specication of the MIRID, the parameters of interest are h, b~ , x, t, where h y1 , . . . , yN 0 . In this case,
the item parameters b~ fbjk : j 1, . . . ,J; k 0; 1, . . . ,Kg are identied under a linear restriction of the type (3.7); see
Butter et al. (1998, p. 51). Taking into account the relationship (4.6), the identication of o1 , . . . , oK , t follows if
rB K and 1K2
= ImB. Thus, the identication of the parameters of interest is different if it is considered under a xedeffects specication of MIRID-type models, or under a random-effects specication. In the former case, there also exists
a problem dealing with the scale of t, which is due to the linear restriction of the type (3.7) imposed on the item
parameters; for details, see Butter et al. (1998, p. 52), Smits and De Boeck (2003, pp. 168169) and Smits and Moore
(2004, p. 272). In a random-effects specication of MIRID-type models, this problem does not exist because the
identiability of the parameters of interest does not require a linear restriction of the type (3.7).
5. Maris and Bechger (2004) consider the identiability of xed-effects MIRID models. They specify relationship (4.5) in a
way similar to LLTM-type models, namely b Q xg, where x represents the regression weights and g contains both
the parameters of the component items and the intercepts. More specically, Maris and Bechger (2004) (see also
Bechger et al., 2001) analyze the identication problem when Q x can, after permuting rows and/or columns, be
written as
!
IJK
0JK,mJK
,
4:7
ArJ,JK T J,mJK
here, we are considering J families of composite items, K families of component items, and mJK intercepts. In this case,
= ImQ x; (ii) Q x is a matrix of full rank; and (iii)
they show that the parameters of interest are identied if (i) 1JK J2
Q x1 g1 Q x2 g2 implies that x1 , g1 x2 , g2 ; see Maris and Bechger (2004, Theorem 3 and Lemmas 1 and 2). It
should be remarked that, when mJK 1, there exist specications of the form b Q xg that can be rewritten in the
form of Eq. (4.6). For instance, with six items consisting of four component items and two composite items, Butter et al.
(1998, p. 51) and Maris and Bechger (2004, p. 628) consider the following design matrix:
0
1
1
0
0
0 0
B
C
1
0
0 0C
B 0
B
C
B o1 o2 0
0 1C
B
C
Q x B
4:8
C,
0
1
0 0C
B 0
B
C
B 0
0
0
1 0C
@
A
0
0 o1 o2 1
where g b11 , b12 , b21 , b22 , t, x o1 , o2 0 , b11 , b12 , b21 and b22 are the parameters of the component items, o1 and o2
are the regression weights, and t is the intercept of the regression. The relationship between component items and
composite items as summarized in matrix Q x can be rewritten as follows:
0
1
! o
!
b10
b11 b12 1 B 1 C
o :

b20
b21 b22 1 @ 2 A

Cases as the previous one can be analyzed under a random-effects specication; by so doing, the identication of the
parameters of interest follows provided that rB K and 1K2
= ImB. Thus, parameter identiability under a randomeffects specication is less restrictive than under a xed-effects specication. Let us nally mention that, in the example
we are dealing with, the parameters of interest are not identied under a random-effects specication, neither under a
xed-effects specication of the model (as shown by Maris and Bechger, 2004, Section 6.3).

4.3. Identication of explanatory Rasch-type models at the person side

Explanatory Rasch-type models at the person side can be specied by assuming that the location parameter is not only
person-dependent, but also characterized by explanatory variables. This type of models are also known as latent regression

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

127

IRT-models (De Boeck and Wilson, 2004, Chapter 2) and are typically used in the context of large-scale surveys as the PISA
test (Wu, 2005). Although their identication analysis is a straightforward consequence of Corollary 3.1, it is necessary to
make explicit the hypotheses satised by the covariates. Let us, therefore, review the specication of the process
generating Y 1 , . . . ,Y N , y1 , . . . , yN ,Z 1 , . . . ,Z N , k, b, s, where Z p is a L-dimensional vector of covariates specic to person p:

C1. For each person p, the distribution of Y p 9yp ,Z p , b, k, s depends on yp , b only.

C2. For each person p, Y p1 , . . . ,Y pI are mutually independent conditionally on yp , b; this corresponds to hypothesis H1 or
the condition of local independence.
C3. The distribution of Y pi 9yp , b only depends on yp , bi through (2.1).
C4. Y 1 , . . . ,Y N are mutually independent conditionally on y1 , . . . , yN , b; this corresponds to hypothesis H3.
C5. For each person p, the distribution of yp 9Z p , b, k, s depends on Z p , k, s only.
0
C6. For each person p, yp 9Z p , k, s  GZ p k, s ; that is, the location of the distribution generating the person specic abilities is
restricted as Z 0p k, where k is a L-dimensional vector of regression parameters.
C7. y1 , . . . , yN are mutually independent given Z 1 , . . . ,Z N , k, s.
C8. The distribution of Z 1 , . . . ,Z N does not depend on b, k, s. This condition not only denes Z p as an exogenous variable
in the conditional model yp 9Z p , k, s, but also that Z 1 , . . . ,Z N is exogenous in the conditional model
Y 1 , . . . ,Y N 9Z 1 , . . . ,Z N , b, k, s; for details on exogeneity, see Engle et al. (1983), Hendry (1995), and Florens et al.
(1990, Section 3).
It should be remarked that the previous specication assumes that both observations, random effects and parameters are
endowed with a unique probability measure. The object of such a specication is to examine the construction of a model
rather than to compute a posterior or a predictive distribution relative to a particular prior specication on the parameters.
Therefore, the reader is only asked not to object to the idea that a prior distribution might be specied in a very large class
of distributions, involving only weak restrictions such as the class of null sets (that is, events of probability zero) or some
independence properties (as the previous one).
This specication implies that the statistical model is characterized by the following properties:
S1. Y 1 , . . . ,Y N are mutually independent given Z 1 , . . . ,Z N , b, k, s.
S2. The conditional distribution of Y p given Z 1 , . . . ,Z N , b, k, s depends on Z p , b, k, s only.
S3. For each person p, the covariates Z p are exogenous with respect to the parameters b, k, s; that is, the distribution
generating Z p is uninformative for the choice of the parameters k, b, s, and reciprocally.
Taking into account condition C6 above, Corollary 3.1 implies that b, mp , s, where mp Z 0p k, is identied by Y p for each
p 1, . . . ,N, provided that a linear restriction on the item parameters of the type (3.7) is imposed on the item parameters b.
Under conditions S1 and S2, Theorem 2 in Mouchart and San Martn (2003) ensures that the previous N identication
relations imply the identication of b,Z 01 k, . . . ,Z 0N k by Y 1 , . . . ,Y N given Z 1 , . . . ,Z N .
It remains to identify k. But Z 01 k, . . . ,Z 0N k can equivalently be rewritten as
Z k l,

4:9
0

where l m1 , . . . , mN , and Z is an N  L matrix whose p-th row is

Therefore, if rZ L Z-a:s:, b, m1 , . . . , mN and b, k, s
are related through an injective mapping and, therefore, b, k, s is identied by the observations. Summarizing, we obtain
the following corollary:
Z 0p .

Corollary 4.4. Consider explanatory Rasch-type models as specied by conditions C1C8, where the function F is a continuous
strictly increasing cdf, with a continuous strictly positive density function on R. Then
1. A necessary condition to identify b, k, s by one observation is that Lr N.
2. If at least two items are available, b, k, s are identied by the observations conditionally on the covariates Z, provided that
the item parameters satisfy a linear restriction of the type (3.7) and that rZ L Z-a:s:

This corollary deserves the following comments:

1. The identication restriction rZ LZ-a:s: leads to impose restrictions on the distribution generating Z. A standard
distributional assumption for the exogenous variables Z is to take its distribution dominated with respect to the
Lebesgue measure on RNL . Let Z l be the l-th column of matrix Z, that is, the l-th covariates across the N persons. It
follows that the set
fZ : Z 1 , . . . ,Z L are linearly dependentg
has Lebesgue measure zero (see Farrell, 1985, Section 8.4) and, therefore, rZ LZ-a:s:

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

128

2. The previous identication result is still valid provided that PfrZ Lg 4 0; this type of conditions can be satised if the
distribution of (some) columns of Z is discrete. In this case, it is said that b, k, s are identied in the conditional model;
for details, see Oulhaj and Mouchart (2006, Section 4).
3. Following De Boeck and Wilson (2004, Chapter 2), the conditional component of a Rasch-type model can be written as
PY pi 19Zpi  FZpi

where Zpi yp bi ,

mp , s

. A doubly explanatory model or latent regression LLTM model corresponds, therefore, to restrict
where yp 9mp , s  G
both the location of the person specic ability or random-effect yp and the item parameters or xed-effect bj as follows:
L
X

mp

ll Z pl 6Z 0p k,

l1

bi

K
X

gk X ik 6X 0i c,

k1

where Zpl is the value of person p on person property l and ll is the xed regression weight of person property l.
Similarly, Xik is the value of item i on item property k and gk is the regression weight of item property k. Therefore,
PY p y9yp ,Z p ,X, c, k

I
Y

F ypi yp Z 0p kX 0i c1Fyp Z 0p kX 0i c1ypi ,

4:10

i1

where X is an I  K matrix whose i-th row is X 0i and y 2 f0; 1gI . For each person p, the statistical model is given by
Z Y
I
F ypi syp Z 0p kX 0i c1Fsyp Z 0p kX 0i c1ypi Gdyp :
PY p y9Z p ,X, c, k, s
Ri1

The likelihood function corresponds to the product over p 1, . . . ,N of the above marginal probabilities. The
identication of c, k, s by one observation Y p follows from both Corollaries 4.1 (or 4.2) and 4.9.
4. It should be mentioned that the identication of semi-parametric binary models has been considered by Butler and
Louis (1997, Theorem 4.1), where the parameters of interest are the xed effects and the distribution generating the
random effects. The structure of the conditional component of this type of models is similar to (4.10) (see Butler and
Louis, 1997, Eq. (2.2)), although the nested structure itemspersons of IRT models is absent of those models.
Consequently, the identication results of Butler and Louis (1997) are not immediately applied to semi-parametric
doubly explanatory IRT models. Other recent contribution on identication of semi-parametric binary models is Fox
et al. (2012), which consider the nested structure already mentioned. How these results can be used to identify semiparametric explanatory IRT models will be addressed by the authors in a future work.
5. Discussion
The identiability problem is basic to the problem of statistical inference. Unless the parameters in a statistical model
are identiable, there is no meaning of estimability or estimation of such parameters as several combinations of different
values for the parameters may lead to the same statistical model. Although this kind of statements are widely known in
statistics and psychometrics, there does not exist a systematic study of parameter identication of random-effects Raschtype models. In particular, parameter identication in the statistical model (2.3) is typically justied from parameter
identication in the xed-effects Rasch model (2.1). However, as it was shown by San Martn et al. (2011), there are no
general relationships between the identication of the conditional model (2.1) and the identication of the statistical
model (2.3).
This paper provides an explicit proof of the identication of both the item parameters and the parameters indexing the
distribution generating the random effects. This identication is established in the statistical model (2.3). Based on it, the
meaning of the parameters is explored and compared with the meaning of the parameters in a xed-effects specication of
the Rasch model; it is concluded that the meaning of the item parameters is quite different. This discussion is extended to
specic explanatory Rasch-type models, particularly LLTM models and MIRID models. It is shown how different are the
identication restrictions when they are obtained in a random-effects specication of the model, or in a xed-effects
specication. Finally, the identication of latent regression models was discussed, showing how the parameter
identication depends on specic conditions on the covariates.

Acknowledgments
The rst author was partially supported by the CEPPE CIEO-01 grant from the National Commission of Science and
Technology of the Chilean Government, and by the PUENTE Grant 08/2009 from the Ponticia Universidad Catolica de
Chile. The nal version of this paper was completed while the rst author was visiting the Center for Operational Research

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

129

and Econometrics CORE, Universite Catholique de Louvain, Belgium, during January 2012. The authors are grateful to the
reviewers for their comments leading to improve the nal version of this paper.
Appendix A
To prove that the function a12 given by (3.5) is a strictly increasing continuous function of s, we need to study the sign
of its derivative with respect to s. This requires not only to use the Implicit Function Theorem (see Spivak, 1965), but also
to assume regularity conditions allowing to perform derivatives under the integral sign. We accordingly assume that the
cdf F has a continuous density function f strictly positive on R. Furthermore, to prove that a12 is a strictly increasing
continuous function of s, we need to obtain the derivatives under the integral sign of the function ps, b as dened in (3.2)
with respect to s and to b. Consequently, it is assumed that, 8s 2 R0 and 8b 2 R, there exist E 4 0 and Z 40, such that
Z
0
sup
sup f s0 xb Gdx o1,
R 9s0 s9 r E9b0 b9 r Z

Z
R

9x9 sup

sup f j0 xb Gdx o1:

9s0 s9 r E9b0 b9 r Z

Thus, under these regularity conditions, the function ps, b is continuously differentiable under the integral on R0  R
and, therefore,
Z
@
i D2 ps, b6 ps, b  f sxbGdx,
@b
R
ii D1 ps, b6

@
ps, b
@s

xf sxbGdx:

A:1

Thus, ps, a as dened by (3.3) is also continuously differentiable on R0  0; 1 and from (3.4), we obtain that
@
ps,ps, b D2 ps,ps, b  D2 ps, b,
@b
@
ii 0
ps,ps, b D1 ps,ps, b D2 ps,ps, b  D1 ps, b,
@s

i 1

A:2

where
D1 ps, a6

@
ps, a,
@s

D2 ps, a6

@
ps, a:
@a

Combining (A.1) and (A.2), we obtain that

i D2 ps, a

1
1
R
,
D2 ps,ps, a
R f sxps, aGdx

ii D1 ps, a 

R
xf sxps, aGdx
D1 ps,ps, a
RR
6Es, a X,
D2 ps,ps, a
R f sxps, aGdx

A:3

where
f sxps, aGdx
:
R f sxps, aGdx

Ps, a X 2 dx6Gs, a dx6 R

A:4

Thanks to the regularity conditions allowing to perform derivatives of ps, b, and to the fact that F r 1, it can be shown
that a12 is continuously differentiable under the integral sign in s, b1 and b2 ; therefore the function qs, a1 , a2 is
continuously differentiable under the integral sign with respect to s. It remains to show that the derivative w.r.t. s is
strictly positive. Now, using (A.3)(ii), we obtain that
@
Fsxps, a xEs, a Xf sxps, a:
@s
But

Z
R

A:5

xEs, a1 Xf sxps, a1 Fsxps, a2 Gdx

Z
f sxps, a1 Gdx  C s, a1 fX,FsXps, a2 g:
R

Now, since Fsxps, a2  is a strictly increasing function of x, the covariance between X and FsXps, a2  (with respect to
Gs, a1 ) is strictly positive (if X is not degenerate). Furthermore,
Z
f sxps, a1 Gdx
R

130

E. San Martn, J.-M. Rolin / Journal of Statistical Planning and Inference 143 (2013) 116130

is clearly strictly positive since the density function f is strictly positive on R. The two terms of the derivative of qs, a1 , a2
are, therefore, strictly positive. &
References
Adams, R.J., Wilson, M., Wang, W., 1997. Multidimensional random coefcients multinomial logit model. Applied Psychological Measurement 21, 123.
Adams, R.J., Wu, M.L., 2007. The mixed-coefcients multinomial logit model: a generalization form of the Rasch model. In: von Davier, M., Carstensen,
C.H. (Eds.), Multivariate and Mixture Distribution Rasch Models, Springer, pp. 5775.
Basu, D., 1975. Statistical Information and Likelihood (with Discussion). Sankhya 37, 171.
Bechger, T.M., Maris, G., Verstralen, H.H.F.M., Beguin, A.A., 2003. Using classical test theory in combination with item response theory. Applied
Psychological Measurement 27, 319334.
Bechger, T.M., Verstralen, H.H.F.M., Verhelst, N.D., 2001. Identiability of Nonlinear Logistic Test Models. Psychometrika 66, 357372.
Bechger, T.M., Verstralen, H.H.F.M., Verhelst, N.D., 2002. Equivalent linear logistic test models. Psychometrika 67, 123136.
Butler, S.M., Louis, T.A., 1997. Consistency of maximum likelihood estimators in general random effects models for binary data. Annals of Statistics 25,
351377.
Butter, R., De Boeck, P., Verhelst, N., 1998. An item response model with internal restrictions on item difculty. Psychometrika 63, 4763.
Catchpol, E.A., Morgan, B.J.T., 1997. Detecting parameter redundancy. Biometrika 84, 187196.
Chen, M.-H., Dey, D., 1998. Bayesian modeling of correlated binary responses via scale mixture of multivariate normal link functions. Sankhya:
The Indian
Journal of Statistics 60, 322343.
Cox, D.R., Hinkley, D.V., 1974. Theoretical Statistics. Chapman and Hall, London.
De Boeck, P., Wilson, M., 2004. Explanatory Item Response Models. A Generalized Linear and Nonlinear Approach. Springer, New York.
Embretson, S.E., Reise, S.P., 2000. Item Response Theory for Psychologists. Lawrence Erlbaum Associates, Publishers, New Jersey.
Engle, R.E., Hendry, D.F., Rixhard, J.F., 1983. Exogeneity. Econometrica 51, 277304.
Fahrmeir, L., Tutz, G., 2001. Multivariate Statistical Modelling Based on Generalized Linear Models, second edition Springer, New York.
Farrell, R.H., 1985. Multivariate Calculation. Use of Continuous Groups. Springer, New York.
Fischer, G., 1983. Logistic latent trait models with linear constraints. Psychometrika 48, 326.
Fischer, G.H., 1995. Derivations of the Rasch model. In: Fischer, G.H., Molenaar, I.W. (Eds.), Rasch Models. Foundations, Recent Developments and
Applications, Springer, New York. (Chapter 2).
Fischer, G.H., 2004. Remarks on Equivalent Linear Logistic Test Models by Bechger, Verstralen, and Verhelst (2002). Psychhometrika 69, 305315.
Fischer, R.A., 1922. On the mathematical foundations of theoretical statistics. Philosophical Transactions of the Royal Society of London, Series A 222,
309368.
Florens, J.P., Mouchart, M., Rolin, J.M., 1990. Elements of Bayesian Statistics. Marcel and Dekker, New York.
Fox, J.T., Kim, K., Ryan, S., Bajari, P., 2012. The random coefcients logit model is identied. Journal of Econometrics 166, 204212.
Halmos, P., 1987. Finite-Dimensional Vector Spaces. Springer, New York.
Hendry, D.F., 1995. Dynamic Econometrics. Oxford University Press, Oxford.
Laplace, P.S., 1774. Memoire sur la probabilite des causes par les evenements. Academie Royal des Sciences de Paris, Paris.
Lazarsfeld, P.F., 1954. A conceptual introduction to latent structure analysis. In: Lazarsfeld, P.F. (Ed.), Mathematical Thinking in the Social Sciences, Russell
& Russell, New York, pp. 349387.
Maris, G., Bechger, T., 2004. Equivalent MIRID models. Psychometrika 69, 627639.
Marsaglia, G., 1964. Conditional means and covariance of normal variable with singular covariance matrix. Journal of the American Statistical Association
59, 12031204.
McCullagh, P., 2002. What is a statistical model? (with Discussion). Annals of Statistics 30, 12251310.
Molenaar, I.W., 1995. Estimation of item parameters. In: Fischer, G.H., Molenaar, I.W. (Eds.), Rasch Models. Foundations, Recent Developments and
Applications, Springer, New York. (Chapter 3).
Molenbergs, G., Verbeke, G., 2004. An introduction to generalized (non)linear mixed model. In: De Boeck, P., Wilson, M. (Eds.), Explanatory Item Response
Models. A Generalized Linear and Nonlinear Approach, Springer, New York, pp. 111153.
Mouchart, M., San Martn, E., 2003. Specication and identication issues in models involving a latent hierarchical structure. Journal of Statistical
Planning and Inference 111, 143163.
Muthen, B., 1979. A structural probit model with latent variables. Journal of the American Statistical Association 74, 807811.
Oulhaj, A., Mouchart, M., 2006. The role of the exogenous randomness in the identication of conditional models. MetronInternational Journal of
Statistics 64, 253261.
Piegorsch, W., 1992. Complementary log regression for generalized linear models. American Statistician 46, 9499.
Rabe-Hesketh, S., Skrondal, A., 2001. Parametrization of multivariate random effects models for categorical data. Biometrics 57, 12561264.
Rasch, G., 1966a. An individualistic approach to item analysis. In: Lazarsfeld, P.F., Henry, N.W. (Eds.), Readings in Mathematical Social Sciences, MIT Press,
Massachusetts, pp. 89107.
Rasch, G., 1966b. An item analysis which takes individual differences into account. British Journal of Mathematical and Statistical Psychology 19, 4957.
Reichenbach, H., 1956. The Direction of Time. University of California Press, Los Angeles.
Rijmen, F., De Boeck, P., 2005. A relation between a between-item multidimensional IRT model and the mixture Rasch model. Psychometrika 70, 481496.
San Martn, E., Gonzalez, J., Tuerlinckx, F., 2009. Identied parameters, parameters of interest and their relationships. Measurement: Interdisciplinary
Research and Perspective 7, 97105.
San Martn, E., Jara, A., Rolin, J.-M., Mouchart, M., 2011. On the Bayesian nonparametric generalization of IRT-type models. Psychometrika 76, 385409.
San Martin, E., Rolin, J.M., Castro, L.M. Identication of the 1PL model with guessing parameter: parametric and semi-parametric results. Psychometrika,
in press.
Sijtsma, K., Molenaar, I.W., 2002. Introduction to Nonparametric Item Response Theory. Sage Publications, Thousand Oaks.
Smits, D.J., De Boeck, P., 2003. A componential IRT model for guilt. Multivariate Behavioral Research 38, 161188.
Smits, D.J., De Boeck, P., Verhelst, N., 2003. Estimation of the MIRID: a program and a SAS-based approach. Behavior Research Methods, Instruments, and
Computers 35, 537549.
Smits, D.J.M., Moore, S., 2004. Latent item predictors with xed effects. In: De Boeck, P., Wilson, M. (Eds.), Explanatory Item Response Models.
A Generalized Linear and Nonlinear Approach, Springer, New York, pp. 267287.
Snijders, T., Bosker, R., 1999. Multilevel Analysis: An Introduction to Basic and Advanced Multilevel Modeling. Sage, London.
Spivak, M., 1965. Calculus on Manifolds: A Modern Approach to Classical Theorems of Advanced Calculus. Perseus Book Publishing, Cambridge.
Verbekee, G., Molenberghs, G., 2009. Linear Mixed Models for Longitudinal Data. Springer, New York.
Wu, M., 2005. The role of plausible values in large-scale surveys. Studies in Educational Evaluation 31, 114128.