Workshop

Polytomous IRT models

(# 144, Remo Ostini and Michael L. Nering)

Jorge Tendeiro

16 April 2014

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 1/49
Literature
Presentation based on the book:
Ostini, R., & Nering, M. L. (2006). Polytomous item
response theory models. Sage University Paper Series
QASS.
(“Little green book” # 144)

I also used a classic book:

Embretson, S. E., & Reise, S. P. (2000). Item response
theory for psychologists. Chapter 5.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 2/49
Overview

1 Introduction

2 (Some) Polytomous IRT models

Nominal response model (NRM)
Partial credit model (PCM)
Generalized partial credit model (GPCM)
Rating scale model (RSM)
Graded response model (GRM)

3 Model selection

4 Software

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 3/49
Introduction

Introduction

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 4/49
Introduction

Item response theory (IRT): Main idea

Modeling the relationship item↔person by means of a mathematical
function:
P(Xi = c|θ) = f (θ)
| {z }
Pic (θ)

X Xi = Item i with discrete response categories.

X c = Coded response category:
• If X is dichotomous, c = 0, 1;
• If X is polytomous, c = 0, 1, . . . , m (m > 1).
X θ = Person trait parameter.

This is the item response function (IRF).

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 5/49
Introduction

IRT: Important property

Item location (to be defined shortly) and person trait are indexed on the
same metric.
Example: Dichotomous item
1.0

0.8
P(Xi = 1)

0.6
0.5
0.4

0.2

0.0
b
−4 −2 0 2 4

θ (person ability)

• θ > b −→ person is more likely to answer Xi = 1.

• θ < b −→ person is more likely to answer Xi = 0.
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 6/49
Introduction

IRT: Dichotomous models recap.

• Dichotomous items:
Xi = 0 (incorrect, false) or Xi = 1 (correct, true).
• Most common models (logistic): 1PLM, 2PLM, 3PLM
• These models typically relate θ and Pi1 (θ):

Pi1 (θ) = f (θ).

[Pi0 (θ) ≡ 1 − Pi1 (θ)].

We usually simplify notation in the dichotomous case:

Pi (θ) = Pi1 (θ).

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 7/49
Introduction

IRT: Dichotomous models recap.

1PLM
1
Pi (θ) =
1 + exp[−(θ − bi )]

• bi = difficulty param.
1.00
b1= −1
b2= 0.5
0.75 b3= 1.5
P(Xi = 1)

0.50

0.25

0.00
b1 b2 b3
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 8/49
Introduction

IRT: Dichotomous models recap.

2PLM
1
Pi (θ) =
1 + exp[−ai (θ − bi )]

• bi = difficulty param., ai = discrimination param.

1.00
a1= 1, b1= −1
a2= 2, b2= 0.5
0.75 a3= 0.5, b3= 1.5
P(Xi = 1)

0.50

0.25

0.00
b1 b2 b3
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 9/49
Introduction

IRT: Dichotomous models recap.

3PLM
1
Pi (θ) = ci + (1 − ci )
1 + exp[−ai (θ − bi )]

• bi = difficulty param., ai = discrimination param., ci = guessing param.

1.00
a1= 1, b1= −1, c1= 0
a2= 2, b2= 0.5, c2= 0.2
0.75 a3= 0.5, b3= 1.5, c3= 0.25
P(Xi = 1)

0.50

0.25

0.00
b1
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 10/49
Introduction

IRT: Polytomous models

In this case Xi = 0, 1, . . . , m, where m > 1.
Example of items with multiple response items:
• Rating scale
(e.g., Likert-type items: ‘Strongly disagree’, ..., ‘Strongly agree’).
• Ability test items awarding partial credit.

Now we need to define models which allow estimating each Pic (θ),
c = 0, 1, . . . , m: 
 Pi0 (θ) = f1 (θ)

··· .


Pim (θ) = fm (θ)
These are the item category response functions (ICRFs).

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 11/49
Introduction

IRT: Polytomous models – Why?

Polytomous items. . .
• are extensively used in applied psychological measurement.
• measure across a wider range of the trait continuum θ.
• are related to an increase of statistical information when
compared to dichotomous items.
• (in some settings) may help reducing test length
(time&, costs&, respondents’ motivation%).

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 12/49
(Some) Polytomous IRT models Nominal response model (NRM)

Nominal response model

(NRM)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 13/49
(Some) Polytomous IRT models Nominal response model (NRM)

NRM (Bock, 1972)

• Type of items: Polytomous with two or more nominal categories.

• Here, nominal categories = unordered in terms of the trait being
measured.
• E.g.: Multiple choice items (namely the distractors).

The NRM is a “divide-by-total”, or “direct” model:

The ICRFs are modeled directly.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 14/49
(Some) Polytomous IRT models Nominal response model (NRM)

NRM (Bock, 1972)

The ICRF for category c (c = 0, 1, . . . , m) is

exp(λic θ + ζic )
Pic (θ) = Pm .
h=0 exp(λih θ + ζih )

• λih = slope associated to category h of item i.

• ζih = intercept associated to category h of item i.

To identify the model (i.e., to estimate parameters), one of two

constraints is typically imposed:
•
Pm Pm
h=0 λih = h=0 ζih = 0, or
• λi0 = ζi0 = 0.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 15/49
(Some) Polytomous IRT models Nominal response model (NRM)

NRM (Bock, 1972): Example

Item measuring student mathematical achievement (N ' 2, 000).
Response options P
A B C D
λi −.30 .81 −.31 −.20 .000
ζi .21 .82 −.09 −.94 .000
1.0

0.8
Pi (X = c)

0.6
Response A
Response B
0.4 Response C
Response D
0.2

0.0
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 16/49
(Some) Polytomous IRT models Nominal response model (NRM)

NRM (Bock, 1972): Example

Interpretation:
• Response B is the most popular for the more able respondents.
• Response A is the most popular for the less able respondents
(followed by Response C).
• Response D was not popular across the entire trait scale.

In general, for the NRM:

• The popularity of response categories across the entire trait scale
is associated to the order of the intercepts ζic .

For the example, in increasing order of popularity:

Response D < Response C < Response A < Response B.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 17/49
(Some) Polytomous IRT models Partial credit model (PCM)

Partial credit model

(PCM)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 18/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982)

• Type of items: Polytomous with two or more ordinal categories.

• Ideal when the answer to an item consists of an ordered
sequence of steps.
• Partial credit can be given if the respondents only answered
correctly to the first (but not all) steps.
• Varying number of categories across items is possible.
• PCM = Applying the 1PLM to each pair of adjacent item
response categories.
• The PCM is an extension of the 1PLM.

The PCM is a “divide-by-total”, or “direct” model:

The ICRFs are modeled directly.
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 19/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982)

The ICRF for category c (c = 0, 1, . . . , m) is
hP i
c
exp j=0 (θ − δij )
Pic (θ) = Pm hP
h
i.
h=0 exp j=0 (θ − δij )

• δij (j = 1, . . . , m): Item step difficulties, also known as

• category boundaries;
• category intersections.
• Notation: 0j=0 (θ − δij ) = 0.
P

Xi = 0 Xi = 1 Xi = 2 Xi = 3

δi1 δi2 δi3

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 20/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982)

• δij = θ-value at which two consecutive ICRFs intersect:

Pi(j−1) (δij ) = Pij (δij ).

• The higher the δij , the more difficult a particular step is.
• The δij ’s aren’t necessarily ordered in the same sequence as the
categories (reversals; such a case indicates that the item is
probably not functioning as intended).

Special restriction of the PCM:

There must exist responses in every response category.
(Problematic for sparse data.)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 21/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982): Example

Item from a survey of morality (N ' 1, 000).
Five-point Likert-type rating scale.
Step Difficulties
δi1 δi2 δi3 δi4
−1.618 −0.291 0.414 2.044
1.0
Category 1
0.8 Category 2
Category 3
Category 4
Pi (X = c)

0.6 Category 5

0.4

0.2

0.0
δi1 δi2 δi3 δi4
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 22/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982): Example

Interpretation:
• In this case the δij ’s are ordered, so adjacent ICRFs intersect at
locally optimal trait values.
• In particular, each answer option has the highest probability in
some subinterval of the θ-scale.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 23/49
(Some) Polytomous IRT models Partial credit model (PCM)

PCM (Masters, 1982): Example

Interpretation:
• In this case the δij ’s are ordered, so adjacent ICRFs intersect at
locally optimal trait values.
• In particular, each answer option has the highest probability in
some subinterval of the θ-scale.

1.0
Category 1
0.8 Category 2
Category 3
Category 4
Pi (X = c)

0.6 Category 5

0.4

0.2

0.0
δi1 δi2 δi3 δi4
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 23/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

Generalized partial credit model

(GPCM)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 24/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992)

• The GPCM is a generalization of the PCM.

• Idea: Add discrimination parameter (one per item).
• So, in a way, PCM→GPCM just like 1PLM→2PLM.

The GPCM is a “divide-by-total”, or “direct” model:

The ICRFs are modeled directly.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 25/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992)

The ICRF for category c (c = 0, 1, . . . , m) is
hP i
c
exp j=0 αi (θ − δij )
Pic (θ) = Pm hP
h
i.
h=0 exp j=0 αi (θ − δij )

• δij (j = 1, . . . , m): Item step difficulties (category intersections).

• αi : Item discrimination (slope parameters).
• Notation: 0j=0 αi (θ − δij ) = 0.
P

Xi = 0 Xi = 1 Xi = 2 Xi = 3

δi1 δi2 δi3

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 26/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992)

• δij = θ-value at which two consecutive ICRFs intersect.

• αi — Intuitive interpretation:
• Small values (say, ≤ 1) → ‘flatter’ ICRFs.
• Large values (say, ≥ 1.5) → more ‘peaked’ ICRFs.

In Muraki’s (1992, p. 162) words:

“[The αi ’s] indicate the degree to which categorical
responses vary among items as θ level changes.”

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 27/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992): Example

• Items from the Neuroticism Extraversion Openness Five-Factor

Inventory (NEO-FFI; Costa & McCrae, 1992).
• Five-point Likert-type rating scale.
(0 = strongly disagree;. . . ; 4 = strongly agree.)
• N = 350.

Let’s see three items.

Response category
Item Content 0 1 2 3 4
5 Feels tense and jittery 17 111 97 101 24
6 Sometimes feels worthless 72 89 52 94 43
9 Feels discouraged, like giving up 27 128 66 95 34

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 28/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992): Example (slope ' 1)

Item 6 ‘Sometimes feels worthless’.
(0 = 72, 1 = 89, 2 = 52, 3 = 94, 4 = 43).

Slope Step Difficulties

α6 δ61 δ62 δ63 δ64
1.073 −0.873 0.358 −0.226 1.547
1.0

0.8
P6 (X = c)

Category 0
0.6 Category 1
Category 2
0.4 Category 3
Category 4
0.2

0.0
δ61 δ63 δ62 δ64
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 29/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992): Example (slope < 1)

Item 5 ‘Feels tense and jittery ’.
(0 = 17, 1 = 111, 2 = 97, 3 = 101, 4 = 24).

Slope Step Difficulties

α5 δ51 δ52 δ53 δ54
0.683 −3.513 −0.041 0.182 2.808
1.0
Category 0
0.8 Category 1
Category 2
Category 3
P5 (X = c)

0.6 Category 4

0.4

0.2

0.0
δ51 δ52 δ53 δ54
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 30/49
(Some) Polytomous IRT models Generalized partial credit model (GPCM)

GPCM (Muraki, 1992): Example (slope ' 1.5)

Item 9 ‘Feels discouraged, like giving up’.
(0 = 27, 1 = 128, 2 = 66, 3 = 95, 4 = 34).

Slope Step Difficulties

α9 δ91 δ92 δ93 δ94
1.499 −1.997 0.210 0.103 1.627
1.0

0.8
Category 0
P9 (X = c)

0.6 Category 1
Category 2
Category 3
0.4
Category 4

0.2

0.0
δ91 δ93 δ92 δ94
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 31/49
(Some) Polytomous IRT models Rating scale model (RSM)

Rating scale model

(RSM)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 32/49
(Some) Polytomous IRT models Rating scale model (RSM)

RSM (Andrich, 1978)

• Type of items: Polytomous with two or more ordinal categories.

• Requirement: All items of the measurement instrument have the
same consistent structural response form.
E.g.: When the set of responses is the same for all items.
• As a consequence, the response format is intended to function in
the same way across all items.
• The RSM is an extension of the 1PLM.
Moreover, the RSM can be seen as a special case of the PCM.

The RSM is a “divide-by-total”, or “direct” model:

The ICRFs are modeled directly.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 33/49
(Some) Polytomous IRT models Rating scale model (RSM)

RSM (Andrich, 1978)

The ICRF for category c (c = 0, 1, . . . , m) is
nP o
c
exp j=0 [θ − (λi + δj )]
Pic (θ) = Pm nP
h
o.
h=0 exp j=0 [θ − (λi + δj )]

• λi : Item location parameter.

• δj (j = 1, . . . , m): Category threshold parameters.
• Notation: 0j=0 [θ − (λi + δj )] = 0.
P

δ1 δ1
δ2 δ2
δ3 δ3

λ1 λ2 θ scale
Item 1 (4 cats.) Item 2 (4 cats.)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 34/49
(Some) Polytomous IRT models Rating scale model (RSM)

RSM (Andrich, 1978)

• Two consecutive categories intersect at θ = (λi + δj ):

Pi(j−1) (λi + δj ) = Pij (λi + δj ).

• RSM is a special case of the PCM:

Corresponding (across items) category intersections are equally
spaced.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 35/49
(Some) Polytomous IRT models Rating scale model (RSM)

RSM (Andrich, 1978): Example (NEO-FFI)

Thresholds: δ1 = −1.600, δ2 = 0.224, δ3 = −0.184, δ4 = 1.560.
Item 1: λ1 = −0.44 ('easiest' of the 12 items in the scale)
1.0

0.8
Category 0
P1 (X = c)

0.6 Category 1
Category 2
Category 3
0.4
Category 4

0.2

0.0
−4 −3 −2 −1 0 1 2 3 4

Item 1: λ11 = 0.47 ('hardest' of the 12 items in the scale)

1.0

0.8
P11 (X = c)

0.6

0.4

0.2

0.0
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 36/49
(Some) Polytomous IRT models Graded response model (GRM)

Graded response model

(GRM)

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 37/49
(Some) Polytomous IRT models Graded response model (GRM)

GRM (Samejima, 1969)

• Type of items: Polytomous with two or more ordinal categories.

• Varying number of categories across items is possible.
• GRM = Applying the 2PLM at each category boundary
(i.e., between two consecutive category responses).
• The GRM is an extension of the 2PLM.

The GRM is a “difference”, or “indirect” model:

The ICRFs are modeled indirectly.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 38/49
(Some) Polytomous IRT models Graded response model (GRM)

GRM (Samejima, 1969)

The ICRF for category c (c = 0, 1, . . . , m) is
∗
Pic (θ) = Pic∗ (θ) − Pi(c+1) (θ),
where
1
P∗
ic = (the 2PLM).
|{z} 1 + exp[−αi (θ − βic )]
P(Xi ≥c|θ)
∗ ∗
(And Pi0 ≡ 1, Pim ≡ 0.)

For example, if m = 4 (i.e., c = 0, 1, 2, 3):

∗
Pi0 (θ) = 1 − Pi1



∗ ∗
Pi1 (θ) = Pi1 − Pi2


∗ ∗

 Pi2 (θ) = Pi2 − Pi3
∗

Pi3 (θ) = Pi3 − 0.


Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 39/49
(Some) Polytomous IRT models Graded response model (GRM)

GRM (Samejima, 1969)

• αi : Item slope parameter (one per item).

• βic : Category threshold parameters
(one set {βi1 , . . . , βim } per item).
These are the θ-values of transition between response categories.
• The βic ’s are necessarily ordered.

Xi = c − 1 Xi = c
Category responses
βic
Category thresholds
(θ scale) P(Xi ≤ c) = .50 P(Xi ≥ c) = .50

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 40/49
(Some) Polytomous IRT models Graded response model (GRM)

GRM (Samejima, 1969): Example (NEO-FFI)

Item 4 ‘Rarely feels lonely, blue’.
(0 = 20, 1 = 90, 2 = 68, 3 = 125, 4 = 47).

Slope Category thresholds

α4 β41 β42 β43 β44
1.31 −2.72 −0.81 0.04 1.85
1.0
Category 0
Category 1
0.8
Category 2
Category 3
P4 (X = c)

0.6 Category 4
0.5
0.4

0.2

0.0
β41 β44
−4 −3 −2 −1 0 1 2 3 4
θ (person ability)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 41/49
Model selection

Model selection

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 42/49
Model selection

Model selection

• There are plenty of polytomous IRT models available

(models + variants > 10).
• Choosing one model may be a hard enterprise.

Criteria to help choosing the ‘best’ model:

1 Data characteristics

2 Measurement philosophy

3 Mathematical approaches to check fit

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 43/49
Model selection

Model selection

1 Data characteristics
• Dichotomous vs polytomous item scores.
• Nominal vs ordinal categories.
• Number of response categories.
E.g.: The RSM requires the same number across items.

2 Measurement philosophy
• Does the model reflect the the psychological reality that
produced the data?
E.g.: Can one conceptualize the answer to an item as being an
ordered sequence of subtasks for which awarding partial credit
to each is meaningful (i.e., PCM)?

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 44/49
Model selection

Model selection

3 Mathematical approaches to check fit

• Check plots
,→ Compare model-predicted vs empirical response functions.
,→ Plot residuals.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 45/49
Model selection

Model selection

3 Mathematical approaches to check fit

• Statistical fit tests
These may vary depending on their level of generality.
(Assessing fit of all items, of a specific group of items, or of individual
items.)
,→ Residual-based measures.
Based on differences between observed and expected item scores.
,→ Multinomial distribution-based tests.
Based on differences between observed and expected frequencies
of response patterns.
,→ Response function-based tests.
Based on differences between observed and expected
log-likelihood of response patterns.
,→ Guttman error-based tests
Nonparametric approach based on the number of Guttman
errors.
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 46/49
Model selection

Model selection

3 Mathematical approaches to check fit

• Goodness of fit
Consider model fit ⊕ number of estimated parameters.
,→ Akaike’s information criterion (AIC; Akaike, 1977).
,→ Procedures based on likelihood ratio of two comparing models.

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 47/49
Model selection

Model selection
Some problems of statistical fit tests:
• The sampling distributions are often unknown.
• Some tests require very large sample sizes (on the hundreds),
specially for χ2 -based tests.
• Unknown influence of using estimated parameters or of mild
model violations on the performance of the tests.
• Too large sample sizes invariably lead to rejections of the null
hypothesis (effect size?).
A final reassurence:
Some comparative studies of polytomous IRT models suggest that
results don’t vary much between models.
(E.g., Dodd, 1984; Maydeu-Olivares et al., 1994; Ostini, 2001; van Engelenburg,
1997; Verhelst et al., 1997.)
Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 48/49
Software

Software

• IRTPRO
• R: Several packages worth checking
(see http://cran.r-project.org/web/views/Psychometrics.html)
ltm, eRm, TAM, mcIRT, pcIRT,. . .

Workshop Polytomous IRT models, (# 144, Remo Ostini and Michael L. Nering) 49/49