S. ARIMA
The logit of the probability that subject i with latent trait 𝜃
i
answers with category h or higher than h
for item j is modelled as
logit
p
+
ijh
= log
P(Y
ij
⩾ h𝜃)
P(Y
ij
< h𝜃)
= 𝛾
j
𝜃
i
− 𝛽
jh
h = 1, … , H − 1(4)
where 𝛾
j
denes the discrimination power of the item j and 𝛽
jh
is the difculty of response category h
of the item j. The GRM is therefore a nonlinear mixed model in which the probability that a subject i
responds with category h or higher for item j is modelled as a function of both the subject’s latent trait
and the item characteristics, that is, difculty and discrimination power [20]. The difculty parameter is
usually modelled as
𝛽
jh
= 𝛽
j
+ 𝛿
h
j = 1, … , M, h = 1, … , H − 1(5)
where 𝛽
j
denotes the difculty of item j and 𝛿
h
is the difculty of response category h for all the items.
Equation (5) is not valid when items have different numbers of response categories. In this case, we can
assume that the term 𝛿
h
should vary according to response categories, that is,
𝛽
jh
j
= 𝛽
j
+ 𝛿
h
j
j = 1, … , M, h
j
= 1, … , H
j
− 1(6)
where H
j
is the number of response categories of the j-th item. Because all items have the same number
of response categories (H = 5) in the QOL-Dys questionnaire, we use (5) hereafter. Equations (5) and
(6) are not true in general because the difference between response levels could change from question
to question (i.e. 𝛿
jh
j
). However, we prefer to approximate 𝛿
jh
j
with 𝛿
h
j
to reduce the model parameters.
The discrimination parameter is constrained to assume positive values [17, 21] because the probability of
answering correctly is deemed to be increasing as an examinee’s latent trait increases. Figure 1 shows the
ICCs of the GRM in Equations (4) and (5) tted using our QOL-Dys data. In the gure, we show the rst
ten items. Because the response is categorical with H = 5 possible categories, we have H − 1 = 4 ICCs.
Each panel shows how the probability (y-axis) of responding in the h-th category, for each item, changes
based on the value of the latent variable that denes the symptom severity (x-axis). When the patient’s
symptom severity is higher, the probability that the patient frequently experiences the difculty situation
that is dened by the item is higher. For example, in the rst panel, we show the probability that a patient
never experienced a particular situation. For a patient with very severe symptoms (large value on the x-
axis), the probability that he or she will never experience a particular situation is very low; on the other
hand, a patient who feels good is very likely to answer that he or she has never experienced the difculty
situation that is investigated in the selected item. In Figure 2, we show the information function, which
indicates the relative ability of an item to discriminate between contiguous trait scores that are at various
locations along the trait continuum.
For the GRM, the information function of the j-thitemisdenedas
I
j
(𝜃)=
H
h=1
𝛾
j
p
+
jh
(𝜃)
1 − p
+
jh
(𝜃)
− p
+
jh+1
(𝜃)
1 − p
+
jh+1
(𝜃)
2
p
jh
(𝜃)
where p
+
jh
(𝜃)=
exp
(
𝛾
j
(
𝜃−𝛽
jh
))
1+exp
(
𝛾
j
(
𝜃−𝛽
jh
))
and p
jh
(𝜃)=p
+
jh
(𝜃)−p
+
jh+1
(𝜃).
Very easy items are usually more informative at low latent trait levels, whereas very difcult and
discriminating items may be more informative for higher latent trait levels. The information function
characterises the precision of measurement for people who are at different levels of the underlying latent
trait, with larger values of I
j
(𝜃) denoting more precision. Item information is also usually summarised
by the area that is under the item information curve: when the area is larger, the item is considered to be
more informative.
4. Item selection methods
In the classical test theory, the item-test correlation is widely used as a tool for selecting items. The item-
test correlation corresponds to the Pearson correlation coefcient between the question score and the
overall assessment score. It is expected that participants who answer any questions correctly should have
higher overall assessment scores. Items with the largest correlation coefcients can then be selected as
Copyright © 2014 John Wiley & Sons, Ltd. Statist. Med. 2015, 34 487–503
491