SAS Models Item Response Theory Manual

Bayesian Analysis
of Item Response Theory
Models Using SAS
®
Clement A. Stone and Xiaowen Zhu
From Bayesian Analysis of Item Response Theory Models
Using SAS®. Full book available for purchase here.
Contents
About this Book ...............................................................................................ix
About the Authors ......................................................................................... xiii
Acknowledgments .......................................................................................... xv
Chapter 1: Item Response Theory ................................................................... 1
Introduction ........................................................................................................................................... 1
Overview of IRT Models and Their Application.................................................................................. 2
Visualization of IRT Models ........................................................................................................... 3
Organization of the Chapter .......................................................................................................... 4
Unidimensional IRT Models for Dichotomously Scored Responses ............................................... 5
Unidimensional IRT Models for Polytomously Scored Responses ................................................. 5
The Graded Response Model ........................................................................................................ 6
Muraki’s Rating Scale Model ........................................................................................................ 8
The Partial Credit Model and Extensions..................................................................................... 8
The Nominal Response Model .................................................................................................... 11
Other Testing Effects in Unidimensional IRT Models ..................................................................... 12
Locally Dependent Item Sets (Testlets) ..................................................................................... 13
Rater Effects ................................................................................................................................. 13
IRT Models for Multidimensional Response Data ........................................................................... 14
Differential Item Functioning and Mixture IRT Models ................................................................... 16
Hierarchical Models—Multilevel and Random Effects IRT Models ............................................... 17
Evaluation of IRT Model Applications ............................................................................................... 19
Dimensionality .............................................................................................................................. 19
Local Independence ..................................................................................................................... 19
Form of the IRT Model ................................................................................................................. 20
Speededness................................................................................................................................. 20
Model Fit ........................................................................................................................................ 20
IRT Model Parameter Estimation ...................................................................................................... 22
Chapter 2: Bayesian Analysis ........................................................................ 25
Introduction ......................................................................................................................................... 25
Elements of Statistical Models and Inference ................................................................................. 25
Frequentist and Bayesian Approaches to Statistical Inference ..................................................... 26
Use of Bayesian Analysis to Estimate a Proportion—An Example ................................................ 28
Choosing a Prior Distribution ...................................................................................................... 29
Computing the Posterior and Drawing Inferences ................................................................... 30
Comparing Bayesian and Frequentist Estimates for Proportions .......................................... 30
iv
Bayesian Estimation and the MCMC Method .................................................................................. 31
SAS Code for Implementing a Metropolis Sampler to Estimate a Proportion ....................... 32
MCMC and the Gibbs Sampler.................................................................................................... 35
Burn-In and Convergence .................................................................................................................. 35
Informative and Uninformative Prior Distributions .......................................................................... 38
Model Comparisons............................................................................................................................ 39
Deviance Information Criterion ................................................................................................... 39
Bayes Factor ................................................................................................................................. 40
Model Fit and Posterior Predictive Model Checks .......................................................................... 41
Chapter 3: Bayesian Estimation of IRT Models Using PROC MCMC ............... 43
Introduction ......................................................................................................................................... 43
Summary of PROC MCMC Statements for IRT Model Estimation ................................................ 44
MCMC Sampling Algorithms ....................................................................................................... 45
Built-In Distributions .................................................................................................................... 46
Density Plots ................................................................................................................................. 47
Arbitrary Distributions .................................................................................................................. 48
PROC MCMC Template for IRT Model Estimation .......................................................................... 49
How PROC MCMC Processes the SAS Data Set ...................................................................... 51
More on the RANDOM Statement............................................................................................... 51
Model Identification...................................................................................................................... 52
Example Use of the Template for the 1-Parameter IRT Model................................................ 52
Example Output from Estimating the 1P IRT Model ................................................................. 54
Strategies to Improve Convergence of the Monte Carlo Chain ..................................................... 59
Chain Thinning .............................................................................................................................. 59
Parameter Blocking...................................................................................................................... 61
Model Re-parameterizing ............................................................................................................ 64
Analysis of Multiple Chains—Gelman-Rubin Test for Convergence ............................................. 67
Informative, Uninformative, and Hierarchical Priors ....................................................................... 70
Informative Priors ......................................................................................................................... 70
Uninformative Priors .................................................................................................................... 70
Hierarchical Priors ........................................................................................................................ 71
Effect of Different Priors .............................................................................................................. 72
Treatment of Item Parameters as Random Effects ......................................................................... 73
Model Fit and Model Comparisons ................................................................................................... 76
Comparison with WinBUGS ............................................................................................................... 76
The OUTPOST SAS Data Set ............................................................................................................. 77
Autocall Macros for Post-Processing PROC MCMC Data Sets .................................................... 78
Preliminary Item Analyses .................................................................................................................. 79
Chapter 4: Bayesian Estimation of Unidimensional IRT Models for
Dichotomously Scored Items ......................................................................... 81
Introduction ......................................................................................................................................... 81
The 1-Parameter IRT or Rasch Model .............................................................................................. 82
Use of a Normal Ogive Link Function ............................................................................................... 85
The 2-Parameter IRT Model ............................................................................................................... 85
v
A 2-Parameter IRT Model with a Hierarchical Prior ........................................................................ 91
The 3-Parameter IRT Model ............................................................................................................... 94
Comparison of Results Based on MML Estimation....................................................................... 105
Display of Item Response Functions .............................................................................................. 107
Adding Elements to the Graph: Uncertainty in Item Parameters .......................................... 109
Adding Elements to the Graph: Observed Results ................................................................. 111
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for
Polytomously Scored Items ......................................................................... 113
Introduction ....................................................................................................................................... 113
The Graded Response Model .......................................................................................................... 114
Program Template for the GR Model ....................................................................................... 114
Options for Specifying Prior Distributions for the Threshold and Intercept Parameters ...116
Estimation of the GR Model by Using Method 1 ..................................................................... 118
Comparison of Separate and Joint Prior Specifications ........................................................ 120
Output from Estimating the GR Model (Method 1 Prior Specification)................................. 121
Comparison of the Posterior Densities for the Three Prior Specifications .......................... 127
Computation of Transformations by Post-Processing Posterior Results ............................ 128
Specification of the Likelihood Model with Use of the Table Function ................................ 129
Estimation of the One-Parameter Graded Response Model ................................................. 129
Muraki’s Rating Scale Model ........................................................................................................... 131
Estimating the RS-GR Model .................................................................................................... 131
Output from Estimating the RS-GR Model .............................................................................. 132
The Nominal Response Model ......................................................................................................... 136
Estimating the NR Model ........................................................................................................... 136
Output from Estimating the NR Model ..................................................................................... 137
The Generalized Partial Credit Model ............................................................................................. 143
Estimating the GPC Model ........................................................................................................ 144
Output from Estimating the GPC Model .................................................................................. 146
Comparison of Results Based on MML Estimation....................................................................... 149
Graphs of Item Category Response Functions.............................................................................. 151
Graphs of Test Information Functions ............................................................................................ 153
Chapter 6: IRT Model Extensions................................................................. 155
Introduction ....................................................................................................................................... 155
The Bifactor IRT Model..................................................................................................................... 155
Description of the Model ........................................................................................................... 156
Estimation of the Model in PROC MCMC ................................................................................ 156
Output from PROC MCMC ........................................................................................................ 158
Other Multidimensional IRT Models ......................................................................................... 163
The Problem of “Label Switching”............................................................................................ 165
The Testlet IRT Model....................................................................................................................... 166
Description of the Model ........................................................................................................... 166
Estimation of the Model ............................................................................................................. 166
Output from PROC MCMC ........................................................................................................ 167
Hierarchical Models—Multilevel IRT Models ................................................................................. 172
vi
Description of the Multilevel IRT Model ................................................................................... 172
Estimation of the Model ............................................................................................................. 173
Output from PROC MCMC ........................................................................................................ 175
Differential Item Functioning—Multiple Group and Mixture IRT Models .................................... 180
Multiple Group Models for Detecting DIF ................................................................................ 180
Mixture IRT Models .................................................................................................................... 182
Chapter 7: Bayesian Comparison of IRT Models .......................................... 189
Introduction ....................................................................................................................................... 189
Bayesian Model Comparison Indices ............................................................................................. 190
Deviance Information Criterion ................................................................................................. 190
Conditional Predictive Ordinate ................................................................................................ 190
Computing Model Comparison Statistics in PROC MCMC ................................................... 191
Example 1: Comparing Models for Dichotomously Scored Items (LSAT Data) ......................... 193
DIC Results.................................................................................................................................. 193
CPO Results ................................................................................................................................ 194
Example 2: Comparing GR and RS-GR Models for Polytomously Scored Items (DASH Item
Responses) ........................................................................................................................................ 196
DIC Results.................................................................................................................................. 196
CPO Results ................................................................................................................................ 197
Example 3: Comparing a Unidimensional IRT Model and a Bifactor IRT Model ........................ 200
DIC Results.................................................................................................................................. 200
CPO Results ................................................................................................................................ 201
Chapter 8: Bayesian Model-Checking for IRT Models .................................. 205
Introduction ....................................................................................................................................... 205
Different Model-Fit Statistics ........................................................................................................... 206
Test-Level Fit .............................................................................................................................. 206
Item-Level Fit .............................................................................................................................. 206
Person Fit .................................................................................................................................... 207
Posterior Predictive Model Checking ............................................................................................. 207
The PPMC Method ..................................................................................................................... 207
The Posterior Predictive Distribution ....................................................................................... 208
Discrepancy Measures ..................................................................................................................... 212
Test-Level Measures .................................................................................................................. 213
Item-Level Measures.................................................................................................................. 213
Pairwise Measures ..................................................................................................................... 214
Person-Fit Measures .................................................................................................................. 214
Evaluation of Model Fit .............................................................................................................. 215
Example PPMC Applications ........................................................................................................... 216
Example 1: Observed and Predicted Test Score Distributions—LSAT Data........................ 216
Example 2: Observed and Predicted Item-Test Score Correlations ..................................... 220
Example 3: Item Fit Plots and Yen’s Q1 Measure ................................................................... 223
Example 4: Observed and Predicted Odds Ratio Measure .................................................... 231
Example 5: Observed and Predicted Yen’s Q3 Measure ........................................................ 236
Example 6: Observed and Predicted Person-Fit Statistic ...................................................... 239
vii
Example 7: Use of PPMC to Compare Models ........................................................................ 243
References .................................................................................................. 249
Index ........................................................................................................... 257
From Bayesian Analysis of Item Response Theory Models Using SAS®, by Clement A.
Stone and Xiaowen Zhu. Copyright © 2015, SAS Institute Inc., Cary, North Carolina,
USA. ALL RIGHTS RESERVED.
From Bayesian Analysis of Item Response Theory Models
Using SAS®. Full book available for purchase here.
Chapter 5: Bayesian Estimation of Unidimensional
IRT Models for Polytomously Scored Items
Introduction....................................................................................................... 113
The Graded Response Model ............................................................................. 114
Program Template for the GR Model ...............................................................................114
Options for Specifying Prior Distributions for the Threshold and Intercept
Parameters ..........................................................................................................................116
Estimation of the GR Model by Using Method 1 .............................................................118
Comparison of Separate and Joint Prior Specifications ...............................................120
Output from Estimating the GR Model (Method 1 Prior Specification) ........................ 121
Comparison of the Posterior Densities for the Three Prior Specifications ................. 127
Computation of Transformations by Post-Processing Posterior Results ................... 128
Specification of the Likelihood Model with Use of the Table Function........................ 129
Estimation of the One-Parameter Graded Response Model.........................................129
Muraki’s Rating Scale Model ............................................................................. 131
Estimating the RS-GR Model ............................................................................................131
Output from Estimating the RS-GR Model ...................................................................... 132
The Nominal Response Model ........................................................................... 136
Estimating the NR Model ...................................................................................................136
Output from Estimating the NR Model.............................................................................137
The Generalized Partial Credit Model ................................................................ 143
Estimating the GPC Model ................................................................................................144
Output from Estimating the GPC Model ..........................................................................146
Comparison of Results Based on MML Estimation ............................................ 149
Graphs of Item Category Response Functions ................................................... 151
Graphs of Test Information Functions ............................................................... 153
Introduction
The item response theory (IRT) models discussed in Chapter 4 apply to dichotomously scored items. When
items are polytomously scored, such as constructed-response items in educational assessments and rating or
Likert-type response items in health, social, and behavioural measures, polytomous IRT models are
required. These models similarly represent the relationship between a person’s level on the latent trait being
measured and the probability of a particular response. But now there are more than two response
categories. This chapter focuses on polytomous IRT models where items are also assumed to measure one
underlying latent trait.
There are various unidimensional polytomous IRT models. The most commonly used models include (1)
the graded response (GR) model (Samejima, 1969; 1996); (2) a modified GR model (Muraki, 1990), also
called Muraki’s rating scale (RS) model; (3) the partial credit (PC) model (Masters, 1982); (4) the
generalized partial credit (GPC) model (Muraki, 1992); and (5) the nominal response (NR) model (Bock,
1972).
114 Bayesian Analysis of Item Response Theory Models Using SAS
This chapter illustrates how to estimate a variety of IRT models for polytomous responses using PROC
MCMC. Although the models are briefly described in each section, the reader is referred to Chapter 1 for
more detail. The data used for the illustrations are a subset of cases and items from the DASH survey data
(Stone & Irrgang, 2004) and comprise responses of 1000 examinees to 10 five-category items. The DASH
survey was designed to measure physical functioning with items asking about the degree of difficulty, from
1 (“No Difficulty”) to 5 (“Unable”), in performing different tasks such as opening a jar, writing, making the
bed, and others. One latent trait was found to underlie the responses to this survey. In order for you to
execute the code provided in this chapter, it is assumed that that the SAS data set of item responses
(DASH_DATA) has been read into the SAS Work library. (See the SAS Press authors’ web page for this
SAS data set.)
The Graded Response Model
Samejima’s (1969; 1996) graded response (GR) model is used to model the cumulative probability of
responding in a particular response category or higher for graded or ordered response scales. Specifically,
for an item j with mj response categories (1, 2,… mj), the cumulative probability that a person receives a
‫כ‬
(Ʌ)) is modelled by the logistic deviate ‫ݖ‬௝(௞ିଵ) =
category score k (k = 2, … mj) or higher (ܲ௝(௞ିଵ)
‫ܽܦ‬௝ ൫Ʌ െ ܾ௝(௞ିଵ) ൯, where D is the scaling constant (1.7 or 1), Ʌ references the trait being measured for a
person, ܽ௝ is the discrimination (slope) parameter for item j, and ܾ௝(௞ିଵ) is a between category threshold
parameter (k−1 and k) for item j. Given the cumulative response probabilities, the probability of receiving a
category score k on item j is defined as the difference between the cumulative probabilities for two adjacent
‫כ‬
‫כ‬
(Ʌ) = 1 and ܲ௝(௠
(Ʌ) = 0. Note that within each item, the between
categories, with two constraints ܲ௝ଵ
ೕ ାଵ)
category threshold parameters ܾ௝௞ are necessarily ordered in the GR model. Also, the response function is
for an individual person, but the subscript for the person is excluded for convenience in this equation and
subsequent equations.
Program Template for the GR Model
A preliminary template of PROC MCMC code for estimating a GR model using the DASH data is provided
in Program 5.1. For each 5-category item, one slope parameter (ܽ௝ ) and four threshold parameters (ܾ௝ଵ ,
ܾ௝ଶ , ܾ௝ଷ , ܾ௝ସ ) are estimated. Therefore, for the DASH data and GR model, a total of 10 slope parameters
(ܽଵ ~ܽଵ଴ ) and 40 threshold parameters (ܾଵ,ଵ ~ܾଵ,ସ , …, ܾଵ଴,ଵ ~ܾଵ଴,ସ ) are estimated. As for the IRT model
specifications for dichotomously scored items in Chapter 4, the model is parameterized in a slope intercept
form where the logisitic ‫ݖ‬௝௞ = ܽ௝ Ʌ െ ݀௝௞ , D = 1 and ܾ௝௞ = ݀௝௞ /ܽ௝ . These bjk parameters are saved rather
than the djk parameters using the MONITOR option for the PROC MCMC command because the bjk
parameters are more directly interpretable.
The code in the template also presents the set of commands necessary to define the likelihood model for
IRT models given polytomously scored items. As discussed in Chapter 3, rather than program the response
probability model and use a MODEL statement to specify the conditional distribution of the data given the
parameters, you can program the log-likelihood function explicitly using SAS programming statements and
use the GENERAL option with the MODEL statement to specify the log of the joint distribution for the
observed item responses. If you have SAS/STAT 13.1 or earlier, this approach is necessary for IRT models
with polytomously scored items because there is no general categorical distribution function available in
the MODEL statement beyond the binary distribution specification. As will be shown in a later section, a
general categorical distribution function is available, starting in SAS/STAT 13.2, that can be used with the
MODEL statement for polytomously scored items. However, this function is less efficient than the code
presented in Program 5.1.
Notably, the program for estimating a GR model assumes that the response categories for each item in the
data are ordered sequentially from 1 to mj, where mj equals the number of response categories for each item
j. If the item data do not conform to this assumption, the item responses will need to be transformed prior to
your using any of the programs in this chapter.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 115
Program 5.1: Preliminary PROC MCMC Template for Estimating the GR Model for the DASH data
proc mcmc data=dash_data outpost=dash_post seed=23 nbi=5000 nmc=20000
nthreads=8 monitor=(a b1_ b2_ b3_ b4_) diagnostics=all plots=(trace
autocorr);
array
array
array
array
ix[10]; array a[10]; ߑ
b1_[10]; array b2_[10]; array b3_[10]; array b4_[10];
d1_[10]; array d2_[10]; array d3_[10]; array d4_[10];
p_star[10,4]; array p[10,5];
/* Declare model parameters and initial values */ ޾
parms a1 1 d1_1 -1 d2_1 -0.5 d3_1 0 d4_1 1;
parms a2 1 d1_2 -1 d2_2 -0.5 d3_2 0 d4_2 1;
parms a3 1 d1_3 -1 d2_3 -0.5 d3_3 0 d4_3 1;
parms a4 1 d1_4 -1 d2_4 -0.5 d3_4 0 d4_4 1;
parms a5 1 d1_5 -1 d2_5 -0.5 d3_5 0 d4_5 1;
parms a6 1 d1_6 -1 d2_6 -0.5 d3_6 0 d4_6 1;
parms a7 1 d1_7 -1 d2_7 -0.5 d3_7 0 d4_7 1;
parms a8 1 d1_8 -1 d2_8 -0.5 d3_8 0 d4_8 1;
parms a9 1 d1_9 -1 d2_9 -0.5 d3_9 0 d4_9 1;
parms a10 1 d1_10 -1 d2_10 -0.5 d3_10 0 d4_10 1;
/* Prior distributions for slope and intercept parameters */
prior a: ~ lognormal(0, var=16); ޿
prior
prior
prior
prior
d1_1
d2_1
d3_1
d4_1
~
~
~
~
normal(<distribution
normal(<distribution
normal(<distribution
normal(<distribution
specifications>); ߀
specifications>);
specifications>);
specifications>);
[repeat set of prior commands for other item threshold parameter sets]
/* Prior distribution for person ability parameters */
random theta ~ normal(0, var=1) subject=_obs_;
/*Specify the log-likelihood function based on the GR model */
llike=0;
do j=1 to 10;߁
p_star[j,1]=logistic(a[j]*theta-d1_[j]);
p_star[j,2]=logistic(a[j]*theta-d2_[j]);
p_star[j,3]=logistic(a[j]*theta-d3_[j]);
p_star[j,4]=logistic(a[j]*theta-d4_[j]);
b1_[j]=d1_[j]/a[j];
b2_[j]=d2_[j]/a[j];
b3_[j]=d3_[j]/a[j];
b4_[j]=d4_[j]/a[j];
p[j,1]=1-p_star[j,1] ;
do k=2 to 4;
p[j,k]=p_star[j,(k-1)] - p_star[j,k];
end;
p[j,5]=p_star[j,4] ;
llike=llike+log(p[j,ix[j]]);
end;
model general(llike);߂
run;
116 Bayesian Analysis of Item Response Theory Models Using SAS
Numerous arrays are designed to facilitate reference to different parameters and variables in the model.
For example, ARRAY b1_ [10] represents the parameters b1_1 to b1_10 or the set of first threshold
parameters for the 10 items. Similarly, b2_ [10], b3_ [10], and b4_ [10] represent the second, third, and
fourth sets of threshold parameters for the 10 items. Thus, the item parameters for Item 1 are defined
by a1, b1_1, b2_1, b3_1, and b4_1 in the program. The arrays P_STAR and P define two dimensional
arrays with rows indexing the items and columns indexing cumulative response probabilities
(P_STAR) and category response probabilities (P).
޾ In order to reduce autocorrelations in sampled parameter values, multiple PARMS statements are used
to block sets of parameters for each item (see Chapter 3, the section “Parameter Blocking”). In this
program template, initial values of 1 are given for slope parameters (aj), and the initial values of −1,
−0.5, 0, and 1 are given for the first, second, third, and fourth intercepts ( ݀௝ଵ , ݀௝ଶ , ݀௝ଷ , ݀௝ସ ). It should
be noted that the initial values of intercepts for this example should also be ordered to reflect the order
constraint for the thresholds in the GR model: ܾ௝ଵ < ܾ௝ଶ < ܾ௝ଷ < ܾ௝ସ .
ߑ
The PRIOR statements provide the prior distributions for parameters specified in PARMS statements.
As for the models in Chapter 4, the prior distributions for the slope parameters (aj) for each DASH
item are assumed to follow a lognormal distribution with the mean of 0 and the variance of 16. In
contrast to the priors in Chapter 4, no lower and upper bounds were necessary to define a reasonable
set of bounds on sampled values. Since all item slopes follow the same distribution, one PRIOR
statement is used with A: indicating the 10 slopes.
߀ The prior distribution specifications for the threshold (bjk), or intercept parameters (djk) which are
functions of the threshold parameters, must conform to the order constraint in the GR model. That is,
ܾ௝ଵ < ܾ௝ଶ < … < ܾ௝௞ or ݀௝ଵ < ݀௝ଶ < … < ݀௝௞ . Though the priors for these parameters also typically
follow normal distributions, one PRIOR statement for all intercepts PRIOR D: ~ NORMAL
(<distribution specifications>) cannot be used as for the dichotomous IRT models (see Chapter 4).
Instead, for each item, you have to specify four threshold or intercept parameters in four separate
PRIOR statements to reflect the order constraint. Thus, there are 40 prior statements for the threshold
or intercept parameters in total. Three different options for specifying the prior distributions are
discussed in the next section.
߁ The DO loop is used to compute cumulative response category probabilities for each item
(P_STAR[J,1] − P_STAR[J,4]), transform the intercept parameters (djk) to compute the threshold
parameters (bjk), and use differences between cumulative response probabilities to compute the
probability of responding in each of the j response categories (P[J,1] − P[J,5]). Note that the response
probability for response Category 1 (P[J,1]) is the difference between the cumulative probabilities for
response Category 1 (1) and for response Category 2 (P_STAR[J,1]). The response probability in
response Category 2 (P[J,2]) is the difference between the cumulative probabilities for response
Category 2 (P_STAR [J, 1]) and for response Category 3 (P_STAR [J, 2]). The probability associated
with the last response category (P[J,5]) is equal to cumulative probability (P_STAR [J, 4]). Lastly, if
threshold parameters are included in the response probability model, rather than intercept parameters,
the code calculating the threshold parameters as a function of the slope and intercept parameters would
be deleted.
߂ The MODEL statement specifies the log likelihood distribution. There is no need to specify the
dependent variable in the MODEL statement since the log likelihood function includes the dependent
variable (IX[J] array).
޿
Options for Specifying Prior Distributions for the Threshold and Intercept Parameters
The literature contains different approaches to specifying the prior distributions given the order constraint
for the threshold or intercept parameters for each item. Three different methods are discussed that reflect an
ordered constraint between four parameters, analogous to the four threshold parameters for each item in the
GR model. The code below is presented without reference to the item subscript j to simplify the
presentation.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 117
Method 1(L)—Use the sampled values for bk or dk as lower bounds on the distributions for bk+1 or dk+1 :
parms
prior
prior
prior
prior
d1
d1
d2
d3
d4
-1 d2 -.5 d3 .5 d4 1;
~ normal(0, var=16);
~ normal(0, var=16, lower= d1);
~ normal(0, var=16, lower= d2);
~ normal(0, var=16, lower= d3);
Method 2 (L/U)—Use sampled values for bk or dk as both lower and/or upper bounds on the distributions
for bk+1 or dk+1 :
parms
prior
prior
prior
prior
d1
d1
d2
d3
d4
-1 d2 -.5 d3 .5 d4 1;
~ normal(0, var=16, upper=
~ normal(0, var=16, lower=
~ normal(0, var=16, lower=
~ normal(0, var=16, lower=
d2);
d1, upper= d3);
d2, upper= d4);
d3);
Method 3 (Sort)—Sort the set of thresholds and use the sorted parameter values in the response probability
model (Curtis, 2010):
parms q1 -1 q2 -.5 q3 .5 q4 1;
prior q: ~ n(0, sd=4);
d1=q1;
d2=q2;
d3=q3;
d4=q4;
call sortn(d1, d2, d3, d4);
The use of Method 3 in PROC MCMC is complicated by the fact that no transformations can be performed
directly on model parameters specified in PARMS statements. Thus, a copy of the variables in the PARMS
statements may be obtained, and these copied variables may be sorted. These variables are then used in the
response probability model for the GR model to satisfy the order constraint. It should also be noted that a
direct sampling method cannot be used with the code as presented above for Method 2. Rather, the above
presentation is used to more easily describe the parameter constraints on the normal distributions. Instead, a
joint density function must be specified. This is discussed in the next section.
Output 5.1 presents the prior densities for the four threshold parameters (d1−d4) from the above three
methods where Method 1 is labeled L for lower bounds, Method 2 is labeled L/U for lower or upper
bounds, and Method 3 is labeled Sort. These graphs can be produced by modifying the code presented in
Chapter 3 (Program 3.1), stacking the data sets of sampled values from the distributions, and overlaying
graphs using DENSITYPLOT graphics statements.
As you can see, all three prior densities are different despite the same mean and variance specifications for
the normal distributions. While Methods 2 (L/U) and 3 (Sort) have similar locations, the spreads in the
densities are quite different. In contrast, you can see that the locations of the densities from Method 1 (L)
are very different from the locations of the densities for Methods 2 (L/U) and 3 (Sort). Given the
differences between the prior densities for the three methods, it may be useful to evaluate the use of all
three methods with the GR model. Use of all three prior distributions and the impact on the posterior
densities for the threshold parameters is explored further in a later section. Finally, the observation that
Methods 1 and 2 are different runs counter to the claim by Curtis (2010) that these approaches are
essentially equivalent.
118 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.1: Densities of the Four Parameters for the Different Prior Specifications
Estimation of the GR Model by Using Method 1
Program 5.2 provides code for estimating the GR model. The code in the program is based on the Method 1
prior specifications (lower bound specifications) because this method is commonly used with the GR
model. In these programs, changes were made to the template to make the program more concise. First,
two-dimensional arrays, B[4, 10] and D[4, 10], are used to represent the 40 threshold and 40 intercept
parameters across the 10 items. Second, a joint prior distribution for the intercept parameters is specified,
and all intercept parameters are sampled from this distribution. This is described in more detail below.
Finally, a BEGINNODATA−ENDNODATA block of statements are used. Any statements included within
the block are executed only at the first and the last observation of the data set. Thus a
BEGINNODATA−ENDNODATA block of statements reduces unnecessary observational-level
computations, and is appropriate for computations that are identical across observations, such as
transformation of parameters.
Although not presented, Programs 5.3 and 5.4 contain the code for estimating the GR model using the other
two methods and are included on the SAS Press authors’ web page for this book. As discussed above, given
the differences observed for the different methods of prior specifications, a comparison of results based on
the different methods is presented in a later section.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 119
Program 5.2: PROC MCMC Code for Estimating the GR Model (Method 1 Prior Specification)
proc mcmc data=dash_data outpost=dash_postGR seed=23 nbi=5000 nmc=20000
nthreads=8 monitor=(a b) diagnostics=all plots=(trace density);
array
array
array
array
ix[10]; array a[10];
b[4,10] b1_1-b1_10 b2_1-b2_10 b3_1-b3_10 b4_1-b4_10;
d[4,10] d1_1-d1_10 d2_1-d2_10 d3_1-d3_10 d4_1-d4_10;
p_star[10,4]; array p[10,5];
beginnodata;
lprior=0;
do j=1 to 10;
lprior = lprior
lprior = lprior
lprior = lprior
lprior = lprior
end;
+
+
+
+
lpdfnorm(d[1,j],0,5);ߑ
lpdfnorm(d[2,j],0,5,d[1,j]);
lpdfnorm(d[3,j],0,5,d[2,j]);
lpdfnorm(d[3,j],0,5,d[3,j]);
do j=1 to 10;
do k=1 to 4;
b[k,j] = d[k,j]/a[j]; ߒ
end;
end;
endnodata;
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
a1 1
a2 1
a3 1
a4 1
a5 1
a6 1
a7 1
a8 1
a9 1
a10 1
d1_1 -1 d2_1 -0.5 d3_1 0.5 d4_1 1;
d1_2 -1 d2_2 -0.5 d3_2 0.5 d4_2 1;
d1_3 -1 d2_3 -0.5 d3_3 0.5 d4_3 1;
d1_4 -1 d2_4 -0.5 d3_4 0.5 d4_4 1;
d1_5 -1 d2_5 -0.5 d3_5 0.5 d4_5 1;
d1_6 -1 d2_6 -0.5 d3_6 0.5 d4_6 1;
d1_7 -1 d2_7 -0.5 d3_7 0.5 d4_7 1;
d1_8 -1 d2_8 -0.5 d3_8 0.5 d4_8 1;
d1_9 -1 d2_9 -0.5 d3_9 0.5 d4_9 1;
d1_10 -1 d2_10 -0.5 d3_10 0.5 d4_10 1;
prior a: ~lognormal(0, var=25);ߓ
prior d: ~general(lprior); ߔ
random theta ~ normal(0, var=1) subject=_obs_;
llike=0;
do j=1 to 10;
do k=1 to 4;
p_star[j,k]=logistic(a[j]*theta-d[k,j]); ߕ
end;
p[j,1]=1-p_star[j,1];
do k=2 to 4;
p[j,k]=p_star[j,(k-1)]−p_star[j,k];
end;
p[j,5]=p_star[j,4];
llike = llike + log(p[j,ix[j]]);
end;
model general(llike);
run;
ߑ
The set of prior specifications for threshold or intercept parameters for each item may be more
concisely specified with use of a joint density function. In PROC MCMC a joint density function must
be specified in the log scale. Here, the LPDFNORM is a log-density function for the normal
distribution. PROC MCMC has a number of internally defined log-density functions for univariate and
multivariate distributions. These functions have the basic form of LPDFdist(<variable>, <distribution
parameters>), where ‘dist’ is the name of the distribution, the variable argument references the random
variable, and distribution arguments are used to specify distribution parameters. For a normal
120 Bayesian Analysis of Item Response Theory Models Using SAS
ߒ
ߓ
ߔ
ߕ
distribution, the form is LPDFNORM (variable, mu, SD, lower bounds, upper bounds). The
specification LPDFNORM(D[2,J], 0, 5, D[1,J]) in the DO loop indicates that the second intercept for
each item follows a normal distribution with a mean of 0, standard deviation of 5, and a lower bound
equal to the sampled value for first intercept parameter. The statement LPRIOR=LPRIOR +
LPDFNORM(<distribution specification> ) computes the sum of the log prior density values across all
intercept parameters. Therefore, lprior defines the log of the joint prior distribution for the intercept
parameters across the 10 items. Since these computations are the same across observations, the
statements are embedded within a BEGINNODATA−ENDNODATA block. The equivalence between
separate and joint prior specifications is illustrated in the next section.
The transformation function of B[J, K]=D[J, K] / A[J] is included in the BEGINNODATA and
ENDNODATA block also because the computations are the same across observations. Results are the
same, assuming the same SEED value, whether the transformations are included in a
BEGINNODATA and ENDNODATA block or as specified in Program 5.1.
No bounds are used to restrict the parameter space for estimating the ܽ௝ parameters or the ݀௝௞
parameters other than bounds required to satisfy the order constraint. As you will see in the posterior
summary output and trace plots, the sampled values did not include unreasonable values.
To declare the prior distributions for the intercept parameters (djk), one PRIOR statement is required:
PRIOR D: ~ GENERAL (LPRIOR). The GENERAL function indicates that the intercepts are sampled
from a joint prior distribution with the joint distribution lprior. Notably, any new distributions used
with the GENERAL function, either in the PRIOR statement or in the MODEL statement, have to be
specified on the logarithm scale.
The use of two-dimensional arrays simplifies the programming statements used to compute the
cumulative probabilities. As shown in the program, one statement: P_STAR [J, K] = LOGISTIC
(A[J]*THETA í D[K, J]) is used for computing all of the required cumulative probabilities.
Comparison of Separate and Joint Prior Specifications
Program 5.5 presents code to test the equivalence between separate and joint prior specifications for
Method 1. Output 5.2 compares the densities for the two sets of parameters. As you can see, there is
essentially complete overlap between the densities for the four parameters (d1− d4), indicating equivalence
between the approaches. Any trivial differences are because direct sampling is used to sample values in the
separate prior specifications, whereas the metropolis sampler is used to sample values in the joint prior
specifications.
Program 5.5: Comparing Separate and Joint Prior Specifications—Method 1
proc mcmc data=a nmc=100000 seed=1 outpost=o1 plots=(density);
parm d1 -1 d2 -.5 d3 .5 d4 1;
prior d1 ~ n(0, sd=5);
prior d2 ~ n(0, sd=5, lower=d1);
prior d3 ~ n(0, sd=5, lower=d2);
prior d4 ~ n(0, sd=5, lower=d3);
model general(0);
run;
proc mcmc data=a outpost=o2 seed=23 nbi=1000 nmc=100000 plots=(density);
parms d1 -1 d2 -.5 d3 .5 d4 1;
beginnodata;
lprior = lpdfnorm(d1,0,5);
lprior = lprior + lpdfnorm(d2,0,5,d1);
lprior = lprior + lpdfnorm(d3,0,5,d2);
lprior = lprior + lpdfnorm(d4,0,5,d3);
endnodata;
prior d: ~general(lprior);
model general(0);
run;
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 121
Output 5.2: Densities of the Four Parameters for Separate and Joint Prior Specifications
Output from Estimating the GR Model (Method 1 Prior Specification)
Output 5.3 presents selected posterior summary results from executing Program 5.2. These results were
obtained using the autocall macro %POSTSUM to select a subset of parameters (see Chapter 3 for a
discussion of the autocall macros that can be used with PROC MCMC OUTPOST data sets). As a result,
the summary statistics differ from the default output produced by PROC MCMC. All slope parameter
results are displayed along with a subset of threshold parameters due to the large number of threshold
parameters in the model (40). These particular items (1, 7, and 10) were selected because they reflect
measurement of the trait at varying locations on the θ scale. Point estimates for all parameters are available
in a later section that compares Bayesian and marginal maximum likelihood (MML) parameter estimates.
As you can see in the output, the slope point estimates (means) range from 1.83 to 3.49 indicating that the
DASH items are highly discriminating. The threshold estimates for the selected items reflect the trend for
all the items and do not spread out over the entire latent trait scale. A majority of the thresholds estimates
are larger than 0, indicating that the subset of DASH items is measuring the trait at the upper end of the
trait scale. Given the order constraint for the threshold parameters, you can see that the four thresholds
within each item are indeed ordered. In addition, the standard deviations (SDs) of the posterior distributions
for the parameters are quite small, indicating that all the slope and threshold parameter estimates were
estimated with relatively high precision.
122 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.3: Selected Posterior Summary Results—GR Model
Because of the volume of diagnostic plots that are available from this analysis, Outputs 5.4 to 5.12 present
selected trace, autocorrelation, and density plots for item parameters in the GR model. These results were
obtained using the autocall macro %TADPLOT to select the subset of parameters. You can see from the
plots for the selected slope parameters (Outputs 5.4 – 5.6) that the parameters appear to be mixing well or
traversing the parameter space quickly and have reasonably low autocorrelations. This is in contrast to the
diagnostic plots for the selected threshold parameters (Outputs 5.7–5.12). You can see that the chains for
the threshold parameters do not mix as well as the chains for the slope parameters. This results in
somewhat higher autocorrelations and therefore higher dependence between sampled values. But as for
other models, this higher dependence does not affect posterior summary results.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 123
Output 5.4: Selected Diagnostics Plots for Slope Parameters—GR Model (Item 1)
Output 5.5: Selected Diagnostics Plots for Slope Parameters—GR Model (Item 7)
124 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.6: Selected Diagnostics Plots for Slope Parameters—GR Model (Item 10)
Output 5.7: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 1)
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 125
Output 5.8: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 1)
Output 5.9: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 7)
126 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.10: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 7)
Output 5.11: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 10)
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 127
Output 5.12: Selected Diagnostics Plots for Threshold Parameters—GR Model (Item 10)
Comparison of the Posterior Densities for the Three Prior Specifications
Given the differences observed in the prior densities for the threshold parameters (see Output 5.1), you can
examine any subsequent impact on the posterior densities for the parameters. Output 5.13 presents the
posterior densities for the threshold parameters values for Item 1 under the three different prior density
specifications. You can see that very small differences are observed between the methods using lower
bounds (Lower) and the method using sorted values (Sort). However, there are some modest differences for
several of the threshold parameters under the method using lower and upper bounds on the distributions
(Up/Low). In particular, for the thresholds b1, b2, and b3, the densities are shifted slightly higher than those
based on the other two methods. In terms of comparing the first two moments of the distributions, the
moments only differed in the hundredths place. So essentially, the point estimates for the parameters and
parameter standard errors were essentially identical across the three methods.
Thus, estimation of the GR model for the DASH data set was not sensitive to the method used to specify
the prior distributions for the threshold parameters. This is likely due to the relatively large amount of data
in the DASH data set which “overwhelmed” the priors. Of course, since real data is analysed, it is not clear
which density is closer to true values. Therefore more exploration of these different prior specifications
may be needed.
128 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.13: Comparison of Posterior Densities for the Four Threshold Parameters (Item 1)
Computation of Transformations by Post-Processing Posterior Results
Any transformations of parameter estimates may be obtained by post-processing the data set of posterior
results saved using the OUTPOST option on the PROC MCMC command. For example, rather than
compute the threshold parameters (bjk) using the intercept parameter (djk) and slope parameters (ajk) in the
PROC MCMC program (see program 5.2), the OUTPOST data set can be processed. Assuming the
transformation is excluded from Program 5.2, the following DATA step and PROC MEANS command
may be used to compute the values of bjk for each iteration, and compute point estimates for the bjk
parameters from means of the transformed posterior samples for the bjk parameters. The ease in postprocessing of the posterior samples illustrates the flexibility of using PROC MCMC for Bayesian analyses.
data dash_transformed;
set dash_postGR;
array a[10] a1-a10;
array d[4,10] d1_1-d1_10 d2_1-d2_10 d3_1-d3_10 d4_1-d4_10;
array b[4,10] b1_1-b1_10 b2_1-b2_10 b3_1-b3_10 b4_1-b4_10;
do j=1 to 10;
do k=1 to 4;
b[k,j]=d[k,j]/a[j];
end;
end;
run;
proc means data=dash_transformed;
var a1-a10 d1_1-d1_10 d2_1-d2_10 d3_1-d3_10 d4_1-d4_10 b1_1-b1_10
b2_1-b2_10 b3_1-b3_10 b4_1-b4_10 ;
run;
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 129
Specification of the Likelihood Model with Use of the Table Function
Beginning with SAS/STAT 13.2, the likelihood model for items with more than two response categories
can be specified using the TABLE function. The TABLE function supports the specification of a
categorical distribution with an array of probabilities for the discrete values. As an example, the code for
using this function with a 5-category item (Item 1 from the DASH data set) is presented below.
array p_star[4];
array p[5];
do k=1 to 4;
p_star[k]=logistic(a[&j]*theta-d[k,1]);
end;
p[1]=1-p_star[1];
do k=2 to 4;
p[k]=p_star[(k-1)]-p_star[k];
end;
p[5]=p_star[4];
model ix1 ~ table(p);
The TABLE function requires a vector of probabilities. A two-dimensional array of probabilities, indexed
by item and response option categories, cannot be specified. This makes the use of the TABLE function
cumbersome when estimating IRT models because the code must be repeated for each item separately
rather than including the code within a DO LOOP across the items. Thus, you may find it easier to program
the log-likelihood function explicitly as in Program 5.2. However, the macro presented below can be used
to replicate the above code for each of the items in your data set.
array p_star[4];
array p[5];
%macro prob_gr;
%do j=1 %to 10;
do k=1 to 4;
p_star[k]=logistic(a[&j]*theta-d[k,&j]);
end;
p[1]=1-p_star[1];
do k=2 to 4;
p[k]=p_star[(k-1)]-p_star[k];
end;
p[5]=p_star[4];
model ix&j ~ table(p);
%end;
%mend;
%prob_gr;
Estimation of the One-Parameter Graded Response Model
For the GR model, one slope parameter and four threshold parameters are estimated for each item. Under
the assumption that all items have a common slope parameter, the GR model reduces to a one-parameter
GR model with the logistic deviate ‫ݖ‬௝௞ = ‫ܽܦ‬൫Ʌ െ ܾ௝௞ ൯ ‫ݖ ݎ݋‬௝௞ = ‫ܦ‬൫ܽɅ െ ݀௝௞ ൯. Therefore, the oneparameter GR model can be treated as a restricted case of the GR model. These two models can be
compared using Bayesian model comparison methods (see Chapter 7) to determine which model is
preferred in a particular testing application.
The PROC MCMC commands used to estimate the one-parameter GR model are provided in Program 5.6.
You can see that the commands are similar to the commands used to estimate a GR model in Program 5.2,
except that the slope parameter (a) is a single variable reflecting a common slope parameter across items.
Thus, a single variable is specified in a PARMS statement for the parameter, PARMS A 1, in the prior
distribution for the slope parameter, PRIOR A ~ LOGNORMAL(0, VAR=25), and in the response
probability model, P_STAR[J,K]=LOGISTIC(A*THETA-D[K,J]).
No output is provided for this model. But it can be estimated and compared with the GR model to
determine whether slope parameters for each item should be estimated. As it turns out, for the DASH data
set, the point estimate (mean of posterior values) for the slope parameter a is 2.63, which is similar to the
130 Bayesian Analysis of Item Response Theory Models Using SAS
mean value of the slope estimates from the GR model. However, the ܽ௝௞ parameter estimates from the GR
model varied a good deal (see Output 5.3) with a range from 1.82 to 3.51. This would suggest that a GR
model would be more appropriate for the DASH item responses. The model comparison methods discussed
in Chapter 7 would, however, be more useful in determining the appropriateness of the one-parameter GR
model to the DASH data. As for the threshold bjk parameters, most of the values were similar to the values
from the GR model.
Program 5.6: PROC MCMC Code for Estimating the One-Parameter GR Model
proc mcmc data=dash_data outpost=dash_post1PGR seed=23 nbi=5000 nmc=20000
nthreads=8 monitor=(a b) diagnostics=all plots=(trace autocorr);
array
array
array
array
ix[10];
b[4,10] b1_1-b1_10 b2_1-b2_10 b3_1-b3_10 b4_1-b4_10;
d[4,10] d1_1-d1_10 d2_1-d2_10 d3_1-d3_10 d4_1-d4_10;
p_star[10,4]; array p[10,5];
beginnodata;
lprior=0;
do j=1 to 10;
lprior = lprior + lpdfnorm(d[1,j],0,5);
do k=2 to 4;
lprior = lprior + lpdfnorm(d[k,j],0,5,d[k-1,j]);
end;
end;
do j=1 to 10;
do k=1 to 4;
b[k,j] = d[k,j]/a;
end;
end;
endnodata;
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
a 1;
d1_1 -1
d1_2 -1
d1_3 -1
d1_4 -1
d1_5 -1
d1_6 -1
d1_7 -1
d1_8 -1
d1_9 -1
d1_10 -1
d2_1 -0.5 d3_1 0
d2_2 -0.5 d3_2 0
d2_3 -0.5 d3_3 0
d2_4 -0.5 d3_4 0
d2_5 -0.5 d3_5 0
d2_6 -0.5 d3_6 0
d2_7 -0.5 d3_7 0
d2_8 -0.5 d3_8 0
d2_9 -0.5 d3_9 0
d2_10 -0.5 d3_10
d4_1 1;
d4_2 1;
d4_3 1;
d4_4 1;
d4_5 1;
d4_6 1;
d4_7 1;
d4_8 1;
d4_9 1;
0 d4_10 1;
prior a ~lognormal(0, var=25);
prior d: ~general(lprior);
random theta ~ normal(0, var=1) subject=_obs_;
llike=0;
do j=1 to 10;
do k=1 to 4;
p_star[j,k]=logistic(a*theta-d[k,j]);
end;
p[j,1]=1-p_star[j,1];
do k=2 to 4;
p[j,k]=p_star[j,(k-1)]-p_star[j,k];
end;
p[j,5]=p_star[j,4];
llike = llike + log(p[j,ix[j]]);
end;
model general(llike);
run;
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 131
Muraki’s Rating Scale Model
Another restricted version of the GR model is Muraki’s (1990) rating scale graded (RS-GR) model. In this
model, the logistic deviate is ‫ݖ‬௝௞ = ‫ܽܦ‬௝ ൫Ʌ െ (ܾ௝ + ܿ௞ )൯. The threshold parameters (ܾ௝௞ ) in the GR model
are decomposed into two terms: (1) a location parameter (ܾ௝ ) for each item, and (2) one set of category
threshold parameters (ܿ௞ ) that applies to all items. To identify the model, σ ܿ௞ across the k response options
for each item are constrained to be 0. Also note that the thresholds ܿ௞ are ordered similarly to the ܾ௝௞
parameters in the GR model. One advantage of this parameterization is that the ܾ௝ parameters can be used
to directly compare where different items are measuring on the latent θ continuum. Another advantage is
that the number of parameters in the model is reduced significantly as compared to the GR model.
The RS model is a restricted version of the GR model because the category threshold parameters, ܿ௞ , are
assumed to be equal across all items in the model, whereas they are free to vary across items in the GR
model. Thus, the RS-GR model can be compared with the GR model to determine whether one set of
category threshold parameters adequately describes the item responses. This model comparison is
illustrated with the DASH data set in Chapter 7.
Estimating the RS-GR Model
Program 5.7 includes the PROC MCMC code for estimating the RS-GR model with the DASH data. One
slope parameter (ܽ௝ ) and one location parameter (ܾ௝ ) are estimated for each DASH item, and there is one
set of four category threshold parameters (ܿଵ , ܿଶ , ܿଷ , ܿସ ) for the five response categories: 1 (“No Difficulty”)
to 5 (“Unable”). The category thresholds parameters (ܿ௝ ) are declared in a separate PARMS statement ߑ.
Notably, only three thresholds ܿଶ , ܿଷ , and ܿସ are defined in the PARMS statement. The fourth threshold
(ܿଵ ) is obtained using ܿଵ = െ(ܿଶ + ܿଷ + ܿସ ). This computation is embedded in a BEGINNODATA −
ENDNODATA block and is used to satisfy the model identification constraint σ ܿ௞ = 0 ߒ.The lower
value of ܿଷ is specified as ܿଶ , and the lower value of ܿସ is specified as ܿଷ to also satisfy the order constraint
ߓ. The code used to compute the log-likelihood function is the same as for the GR model except for the
response model ߔ.
Note that the same strategy of decomposing the threshold parameter into location (ܾ௝ ) and category
threshold parameters (ܿ௝ଵ , ܿ௝ଶ , ܿ௝ଷ , ܿ௝ସ ) can be done for each item in the GR model. This is illustrated in
Program 5.8, which can be found on the SAS Press authors’ web page.
132 Bayesian Analysis of Item Response Theory Models Using SAS
Program 5.7: PROC MCMC Code for Estimating the RS-GR Model
proc mcmc data=dash_data outpost=dash_postRSGR seed=23 nbi=5000 nmc=20000
nthreads=8 monitor=(a b c) diagnostics=all plots=(trace autocorr);
array ix[10];
array a[10]; array b[10]; array c[4];
array p_star[10,4]; array p[10,5];
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
a1 1
a2 1
a3 1
a4 1
a5 1
a6 1
a7 1
a8 1
a9 1
a10 1
c2 -2
b1 0;
b2 0;
b3 0;
b4 0;
b5 0;
b6 0;
b7 0;
b8 0;
b9 0;
b10 0;
c3 0 c4 6;ߑ
beginnodata;
c1=-(c2+c3+c4);ߒ
endnodata;
prior
prior
prior
prior
prior
a:
b:
c2
c3
c4
~
~
~
~
~
lognormal(0, var=25);
normal(0, var=25);
normal(0, var=25);
normal(0, var=25, lower=c2);ߓ
normal(0, var=25, lower=c3);
random theta ~ normal(0, var=1) subject=_obs_;
llike=0;
do j=1 to 10;
do k=1 to 4;
p_star[j,k]=logistic(a[j]*(theta-(b[j]+c[k])));ߔ
end;
p[j,1]=1-p_star[j,1];
do k=2 to 4;
p[j,k]=p_star[j,(k-1)]-p_star[j,k];
end;
p[j,5]=p_star[j,4];
llike = llike + log(p[j,ix[j]]);
end;
model general(llike);
run;
Output from Estimating the RS-GR Model
Output 5.14 shows the summary statistics for item parameters. Compared to the results for the GR model
(see Output 5.3), the item slope estimates based on two models are similar. For each item, the item location
parameter in the RS model approximates the average of the four thresholds in the GR model. The four
category threshold parameter estimates for all items are ordered as expected. Outputs 5.15 to 5.18 provide
the trace and autocorrelation plots for selected item parameters. You can see that the chains are mixing very
well for the slope parameters with low autocorrelations. The location and category threshold parameters, on
the other hand, exhibit reasonable mixing but a moderate degree of autocorrelation. To reduce
autocorrelations in other models that have been discussed, the models where ‫ݖ‬௝௞ = ‫ܽܦ‬௝ ൫Ʌ െ ܾ௝௞ ൯ were
reparameterized into ‫ݖ‬௝௞ = ‫ܽܦ‬௝ ൫Ʌ െ ݀௝௞ ൯ where ܾ௝௞ = ݀௝௞ /ܽ௝ . However, the RS-GR model is more difficult
to reparameterize into a slope intercept model since the ܿ௞ parameters are not a function of each item.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 133
Output 5.14: Posterior Summary Statistics for the Item Parameters—RS-GR Model
134 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.15: Trace and Autocorrelation Plots for Item Parameters (a1, a2, a3)—RS-GR Model
Output 5.16: Trace and Autocorrelation Plots for Item Parameters (a10, b1, b2)—RS-GR Model
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 135
Output 5.17: Trace and Autocorrelation Plots for Item Parameters (b9, b10, c1)—RS-GR Model
Output 5.18: Trace and Autocorrelation Plots for Item Parameters (c2, c3, c4)—RS-GR Model
136 Bayesian Analysis of Item Response Theory Models Using SAS
The Nominal Response Model
Bock’s (1972) nominal response (NR) model is the most general model for polytomously scored items and
is used to model item responses when response categories are not necessarily ordered along the trait (θ)
scale. In the NR model, the probability of an examinee with trait level θ responding in the ݇ ௧௛ response
option, for item j with ݉௝ response options (or categories), is given by the following:
ܲ௝௞ (Ʌ) =
݁ ௭ೕೖ
௠ೕ
σ௞ୀଵ
݁ ௭ೕೖ
for
݇ = 1, … . , ݉௝ ,
where ‫ݖ‬௝௞ = ܽ௝௞ Ʌ + ܿ௝௞ is a multivariate logit for the kth option in item j, ܽ௝௞ is the option discrimination
parameter for item j and category ݇, and ܿ௝௞ is called the option extremity parameter for item j and category
݇. As discussed in Chapter 1, to identify the model, the constraint σ ܽ௝௞ = σ ܿ௝௞ = 0 across k or the
constraint ܽ௝ଵ = ܿ௝ଵ = 0 for each item is typically specified.
Estimating the NR Model
Program 5.9 includes the PROC MCMC code for estimating the NR model with the DASH data set. For
each DASH item, there are five slope parameters (ܽ௝ଵ , ܽ௝ଶ , ܽ௝ଷ , ܽ௝ସ , ܽ௝ହ ) and five intercept parameters
(ܿ௝ଵ , ܿ௝ଶ , ܿ௝ଷ , ܿ௝ସ , ܿ௝ହ ) corresponding to five response categories. However, for model identification purposes,
the first slope and intercept for each item are constrained to be 0 in a BEGINNODATA−ENDNODATA
block. Two-dimensional arrays are used for the item parameters where the first index corresponds to the
item number and the second index corresponds to the category response number. The multivariate logits
and response probabilities for the 5 options are captured in the Z[5] and P[5] arrays. The item parameters
within each item are treated as a block, and their initial values are set to be 0. The priors for slopes and
intercepts are declared as unbounded normal distributions with mean of 0 and variance of 25 ߑ. To obtain
the likelihood of the data, the multivariate logits for the second to fifth response options (z2 to z5) are
calculated first ޾. Note that the logit for the first option (z1) is 1 since its slope and intercept are both 0. In
the next step, the response probabilities for 5 options are computed ޿.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 137
Program 5.9: PROC MCMC Code for Estimating the NR Model
proc mcmc data=dash_data outpost=dash_postNR seed=23 nbi=5000 nmc=20000
monitor=(a c) diagnostics=all plots=all;
array a[10,5] a1_1-a1_5 a2_1-a2_5 a3_1-a3_5
a6_1-a6_5 a7_1-a7_5 a8_1-a8_5
array c[10,5] c1_1-c1_5 c2_1-c2_5 c3_1-c3_5
c6_1-c6_5 c7_1-c7_5 c8_1-c8_5
array ix[10]; array z[5]; array p[5];
parms a1_2-a1_5 c1_2-c1_5 0;
parms a2_2-a2_5 c2_2-c2_5 0;
parms a3_2-a3_5 c3_2-c3_5 0;
parms a4_2-a4_5 c4_2-c4_5 0;
parms a5_2-a5_5 c5_2-c5_5 0;
parms a6_2-a6_5 c6_2-c6_5 0;
parms a7_2-a7_5 c7_2-c7_5 0;
parms a8_2-a8_5 c8_2-c8_5 0;
parms a9_2-a9_5 c9_2-c9_5 0;
parms a10_2-a10_5 c10_2-c10_5 0;
a4_1-a4_5
a9_1-a9_5
c4_1-c4_5
c9_1-c9_5
a5_1-a5_5
a10_1-a10_5;
c5_1-c5_5
c10_1-c10_5;
beginnodata;
a1_1=0;a2_1=0;a3_1=0;a4_1=0;a5_1=0;a6_1=0;a7_1=0;a8_1=0;a9_1=0;a10_1=0;
c1_1=0;c2_1=0;c3_1=0;c4_1=0;c5_1=0;c6_1=0;c7_1=0;c8_1=0;c9_1=0;c10_1=0;
endnodata;
prior c1_2-c1_5 c2_2-c2_5 c3_2-c3_5 c4_2-c4_5 c5_2-c5_5
c6_2-c6_5 c7_2-c7_5 c8_2-c8_5 c9_2-c9_5 c10_2-c10_5
~ normal(0, var=25);ߑ
prior a1_2-a1_5 a2_2-a2_5 a3_2-a3_5 a4_2-a4_5 a5_2-a5_5
a6_2-a6_5 a7_2-a7_5 a8_2-a8_5 a9_2-a9_5 a10_2-a10_5
~ normal(0, var=25);
random theta ~ normal(0, var=1) subject=_obs_;
llike=0;
do j=1 to 10;
do k=2 to 5;޾
z[k]=exp(a[j,k]*theta + c[j,k]);
end;
do k=2 to 5;޿
p[k]=z[k]/(1 + sum(of z2-z5));
end;
p[1]=1 - sum(of p2-p5); * calc prob for category 1;
llike = llike + log(p[ix[j]]);
end;
model general(llike);
run;
Output from Estimating the NR Model
Output 5.19 shows the summary statistics for selected item parameters. Selected results for Items 1, 7, and
10 are again reported as these items reflect items with varying item properties. The autocall macro for
regenerating results, %POSTSUM, was used, and, as in previous tables, the results from this macro differ
from the default summary statistics produced by PROC MCMC. Note that all item parameter estimates are
available in the section comparing Bayesian and MML results.
As you can see in the table, the ܽ௝௞ parameters across response categories for an item are ordered. Given
that the ܽ௝௞ parameters reflect the slope of the linear regression of the response process (z) on θ for each
response option k, increasing values of the ܽ௝௞ parameters across response categories reflect an increasing
probability of selecting option k+1 over option k. This in turn indicates that the scale of response options is
ordered. Therefore, an ordered scale is indicated for the DASH items as would be expected.
138 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.19: Selected Posterior Summary Statistics for the Item Parameters—NR Model
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 139
Outputs 5.20 to 5.23 provide diagnostic plots for selected item parameters generated using the autocall
macro %TADPLOT. These plots are typical for all the item parameters. As you can see, the chains for
slopes and intercepts reflect chains that were mixing well and relatively low autocorrelations. The
parameter densities reflect approximately normal distributions, although some of these distributions reflect
rather large standard deviations.
Output 5.20: Diagnostic Plots for Selected Item Parameters—NR Model
140 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.21: Diagnostic Plots for Selected Item Parameters—NR Model
Output 5.22: Diagnostic Plots for Selected Item Parameters—NR Model
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 141
Output 5.23: Diagnostic Plots for Selected Item Parameters—NR Model
The parameters can be tranformed to conform to Bock’s constraint, σ ܽ௝௞ = σ ܿ௝௞ = 0 across k, by
centering the samples values. The program code that follows illustrates centering the sampled values that
result from executing Program 5.9, and Output 5.24 presents output from executing the program code. In
the program, centering of the values is obtained at each iteration in the posterior results by computing the
sum of the sampled values for the ܽ௝௞ and ܾ௝௞ paramters for each item. The mean based on this sum (sum /
5) is then used to center the values for ܽ௝௞ and ܾ௝௞ by taking the sampled value minus the mean. PROC
MEANS can then be used to obtain point estimates for the centered values. As you can see in the output
(Output 5.24), the σ ܽ௝௞ = σ ܿ௝௞ = 0 across k for each item.
142 Bayesian Analysis of Item Response Theory Models Using SAS
data dash_centered;
set dash_postNR;
array a[10,5] a1_1-a1_5 a2_1-a2_5 a3_1-a3_5 a4_1-a4_5 a5_1-a5_5
a6_1-a6_5 a7_1-a7_5 a8_1-a8_5 a9_1-a9_5 a10_1-a10_5;
array c[10,5] c1_1-c1_5 c2_1-c2_5 c3_1-c3_5 c4_1-c4_5 c5_1-c5_5
c6_1-c6_5 c7_1-c7_5 c8_1-c8_5 c9_1-c9_5 c10_1-c10_5;
array a_c[10,5] a_c1_1-a_c1_5 a_c2_1-a_c2_5 a_c3_1-a_c3_5
a_c4_1-a_c4_5 a_c5_1-a_c5_5 a_c6_1-a_c6_5 a_c7_1-a_c7_5
a_c8_1-a_c8_5 a_c9_1-a_c9_5 a_c10_1-a_c10_5;
array c_c[10,5] c_c1_1-c_c1_5 c_c2_1-c_c2_5 c_c3_1-c_c3_5
c_c4_1-c_c4_5 c_c5_1-c_c5_5 c_c6_1-c_c6_5 c_c7_1-c_c7_5
c_c8_1-c_c8_5 c_c9_1-c_c9_5 c_c10_1-c_c10_5;
keep a_c1_1--a_c10_5 c_c1_1--c_c10_5;
do j=1 to 10;
m_a=sum(of a[j,2], a[j,3], a[j,4], a[j,5]);
m_c=sum(of c[j,2], c[j,3], c[j,4], c[j,5]);
do k=1 to 5;
if k=1 then do;
a_c[j,k]=-m_a/5;
c_c[j,k]=-m_c/5;
end;
else do;
a_c[j,k]=a[j,k]-m_a/5;
c_c[j,k]=c[j,k]-m_c/5;
end;
end;
end;
run;
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 143
Output 5.24: Centered Item Parameters for Selected Items—NR Model
The Generalized Partial Credit Model
The generalized partial credit (GPC) model (Muraki, 1992) is also widely used to model polytomous item
responses. In this model, for item j with ݉௝ response options (or categories), the probability of an examinee
with trait level θ responding in a particular score category x is given by the following:
ೣ
ܲ௝௫ (Ʌ) =
݁ σೖసబ ௭ೕೖ
௠
೓
ೕ
σ௛ୀଵ
݁ σೖసబ ௭ೕೖ
for
݇ = 0, 1, 2, ‫ݔ‬, … m୨
where ‫ݖ‬௝௞ = ‫ܽܦ‬௝ (Ʌ െ Ɂ௝௞ ), ܽ௝ is the slope parameter, and Ɂ௝௞ is called the step difficulty parameter for
category k. Different from the GR model, the response probabilities are computed directly in one step, and
the step difficult parameters are not necessarily ordered. The GPC model reduces to the partial credit (PC)
144 Bayesian Analysis of Item Response Theory Models Using SAS
model under the assumption that all items have equal slopes. Similar to the RS-GR model, some
parameterizations of the GPC model decompose the parameter Ɂ௝௞ into a location parameter (ܾ௝ ) for each
item and a set of category thresholds parameters (ɒ௝௞ ). Using this parameterization, the logistic deviate is
‫ݖ‬௝௞ = ‫ܽܦ‬௝ (Ʌ െ ܾ௝ + ɒ௝௞ ). To identify and estimate the parameters of the GPC model, the threshold
parameter of the last category for each item is set to equal the negative of the sum of the category threshold
parameters for the previous categories, that is,
௠ೕ ିଵ
ɒ௝௠ೕ = െ ෍ ɒ௝௞
௞ୀଵ
This constrains the set of category thresholds parameters to equal 0 for each item.
Estimating the GPC Model
Program 5.10 includes the PROC MCMC code for estimating the GPC model, where ‫ݖ‬௝௞ = ‫ܽܦ‬௝ (ߠ െ ܾ௝ +
ɒ௝௞ ), using the DASH data. For each DASH item, one slope parameter (ܽ௝ ), one location parameter (ܾ௝ ),
and four category threshold parameters (ɒ௝ଵ , ɒ௝ଶ , ɒ௝ଷ , ɒ௝ସ ) are defined corresponding to four steps (1ĺ2,
2ĺ3, 3ĺ4, 4ĺ5).
In the code, the TAU[4,10] matrix represents the 40 threshold parameters (4 category thresholds for each
item), and the Z[5] array denotes the five logistic deviate z values for each item, ‫ݖ‬௝௞ = ‫ܽܦ‬௝ ൫Ʌ െ Ɂ௝௞ + ߬௝௞ ൯
with D=1.0. SUM_Z[K] is an array reflecting the sum of the logistic deviates for response categories. That
is, the values σ௫௞ୀ଴ ‫ݖ‬௝௞ for ‫ = ݔ‬1, 2, 3, 4, 5. NUMER[K] are an array that reflects the numerator for response
categories in the GPC model, and P[5] is an array of response probabilities for five score categories.
As for other models, the item parameters within each item are treated as a single block. Note that for the
model identification, only the category threshold parameters for the first three steps are declared as
parameters. The last threshold parameter within each item equals to the negative sum of the first three
steps, and is computed in a BEGINNODATA−ENDNODATA block ߑ. The priors for the slope parameters
(ܽ௝ ) are log-normal distributions, and the priors for both location and thresholds parameters are normal
distributions. To calculate the likelihood of the data, the Z values are calculated first ߒfollowed by the
SUM_Z values. Then the numerator ߓand denominator ߔof the GPC model are computed and used to
obtain the response probabilities ߕ.
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 145
Program 5.10: PROC MCMC Code for Estimating the GPC Model
proc mcmc data=dash_data outpost=dash_postGPC seed=23 nbi=5000 nmc=20000
nthreads=8 monitor=(a b tau) diagnostics=all plots=(trace autocorr);
array ix[10]; array a[10]; array b[10];
array tau[4,10] tau1_1-tau1_10 tau2_1-tau2_10 tau3_1-tau3_10
tau4_1-tau4_10;
array z[5] z1-z5;
array sum_z[5] sum_z1-sum_z5;
array numer[5] numer1-numer5;
array p[5] p1-p5;
parms
parms
parms
parms
parms
parms
parms
parms
parms
parms
a1 1
a2 1
a3 1
a4 1
a5 1
a6 1
a7 1
a8 1
a9 1
a10 1
b1 0
b2 0
b3 0
b4 0
b5 0
b6 0
b7 0
b8 0
b9 0
b10 0
tau1_1 0
tau1_2 0
tau1_3 0
tau1_4 0
tau1_5 0
tau1_6 0
tau1_7 0
tau1_8 0
tau1_9 0
tau1_10 0
tau2_1 0
tau2_2 0
tau2_3 0
tau2_4 0
tau2_5 0
tau2_6 0
tau2_7 0
tau2_8 0
tau2_9 0
tau2_10 0
tau3_1 0 ;
tau3_2 0 ;
tau3_3 0 ;
tau3_4 0 ;
tau3_5 0 ;
tau3_6 0 ;
tau3_7 0 ;
tau3_8 0 ;
tau3_9 0 ;
tau3_10 0 ;
beginnodata;ߑ
tau4_1=-(tau1_1+tau2_1+tau3_1);
tau4_2=-(tau1_2+tau2_2+tau3_2);
tau4_3=-(tau1_3+tau2_3+tau3_3);
tau4_4=-(tau1_4+tau2_4+tau3_4);
tau4_5=-(tau1_5+tau2_5+tau3_5);
tau4_6=-(tau1_6+tau2_6+tau3_6);
tau4_7=-(tau1_7+tau2_7+tau3_7);
tau4_8=-(tau1_8+tau2_8+tau3_8);
tau4_9=-(tau1_9+tau2_9+tau3_9);
tau4_10=-(tau1_10+tau2_10+tau3_10);
endnodata;
prior a: ~ lognormal(0, var=25);
prior b: ~ normal(0, var=25);
prior tau1_1-tau1_10 ~ normal(0, var=25);
prior tau2_1-tau2_10 ~ normal(0, var=25);
prior tau3_1-tau3_10 ~ normal(0, var=25);
random theta ~ normal(0, var=1) subject=_obs_;
llike=0;
do j=1 to 10;
z[1]=0;
sum_z[1]=0;
numer[1]=1;
do k=2 to 5;ߒ
z[k]=a[j]*(theta - b[j] + tau[k-1,j]);
sum_z[k] = sum_z[k-1]+z[k];
numer[k] = exp(sum_z[k]);ߓ
end;
denom = sum (of numer1-numer5);ߔ
do k=1 to 5;
p[k]=numer[k]/denom;ߕ
end;
llike=llike+log(p[ix[j]]);
end;
model general(llike);
run;
146 Bayesian Analysis of Item Response Theory Models Using SAS
Output from Estimating the GPC Model
Output 5.25 shows the summary statistics for selected item parameters from the GPC model. As previously
done, selected results were generated with use of the autocall macro %POSTSUM because there are a large
number of item parameters. The slope estimates vary across items, which indicates that the GPC model
may be more appropriate for the DASH item responses than the PC model (a common slope model). The
locations are all greater than 0 except for the last item. The four threshold parameter estimates (means)
within each item are not ordered as in the GR model. Note that the item parameters for all items are
presented in the section comparing Bayesian and MML estimation methods.
Output 5.25: Posterior Summary Statistics for Selected Item Parameters—GPC Model
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 147
Outputs 5.26 to 5.28 illustrate the trace and autocorrelation plots for selected item parameters. The plots
that are presented represent typical plots across the set of items. You can see that the chains for the slope
parameters (ܽ௝ ) and the category thresholds parameters (ɒ௝௞ ) all reflect chains mixing well with low
dependence among sampled values. This is in contrast with the location parameters (ܾ௝ ), which exhibit
reasonably good mixing but also moderately high dependence among sampled values.
Output 5.26: Trace and Autocorrelation Plots for Selected Item Slope Parameters—GPC Model
148 Bayesian Analysis of Item Response Theory Models Using SAS
Output 5.27: Trace and Autocorrelation Plots for Selected Item Location Parameters—GPC Model
Output 5.28: Trace and Autocorrelation Plots for Item Tau Parameters—GPC Model
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 149
Comparison of Results Based on MML Estimation
To compare the Bayesian estimation results with traditional MML estimation results, we estimated
different models with the DASH data using the computer program IRTPRO (Cai, Thissen, & du Toit,
2011). Tables 5.1, 5.2, and 5.3 present the comparison of Bayesian and MML results for the GR, NR, and
GPC models. Notably, the procedure PROC IRT (SAS, 2014, Chapter 53) could be used for some of these
comparisons. But it could not be used universally for all comparisons in this book, so this SAS procedure
was not used.
Across the tables, you can see that the Bayesian and MML estimates are very similar for the different
models. The one exception may be the comparison for the NR model (see Table 5.2), where the slope
parameters are more similar than the intercept parameters between two estimation methods, and not
unexpectedly, the differences for more extreme values tend to be larger.
Table 5.1: Comparison of Bayesian and MML Item Parameter Estimates for the GR Model
Estimation Method
a
b1
b2
b3
b4
Item 1
Bayesian
MML
1.83
1.82
0.50
0.49
1.23
1.20
2.17
2.16
2.97
2.95
Item 2
Bayesian
MML
2.07
2.06
0.61
0.60
1.23
1.22
1.89
1.88
2.65
2.65
Item 3
Bayesian
MML
3.49
3.49
0.14
0.13
0.79
0.78
1.65
1.63
2.14
2.12
Item 4
Bayesian
MML
2.50
2.49
í
í0.41
0.45
0.44
1.29
1.28
2.13
2.12
Item 5
Bayesian
MML
2.79
2.79
í
í0.85
í
í0.20
0.50
0.49
1.08
1.07
Item 6
Bayesian
MML
3.21
3.19
í
í0.10
0.61
0.59
1.36
1.35
1.84
1.82
Item 7
Bayesian
MML
3.33
3.32
í
í0.40
0.34
0.33
1.18
1.17
1.54
1.53
Item 8
Bayesian
MML
3.18
3.18
í
í0.78
í
í0.11
0.60
0.59
1.16
1.15
Item 9
Bayesian
MML
2.63
2.62
0.53
0.51
1.11
1.09
1.71
1.70
2.08
2.06
Item 10
Bayesian
MML
2.32
2.33
í
í1.59
í
í0.74
0.06
0.04
0.69
0.68
150 Bayesian Analysis of Item Response Theory Models Using SAS
Table 5.2: Comparison of Bayesian and MML Item Parameter Estimates for the NR Model
Estimation
a1
a2
a3
a4
a5
c1
c2
c3
c4
c5
Item 1
Bayesian
MML
0
0
1.02
1.00
2.31
2.30
3.15
3.18
4.64
4.68
0
0
í1.30
Ѹ1.32
í2.36
Ѹ2.40
í4.35
Ѹ4.44
í7.30
Ѹ7.46
Item 2
Bayesian
MML
0
0
1.35
1.33
2.56
2.57
3.32
3.36
5.35
5.44
0
0
í1.61
Ѹ1.62
í2.83
Ѹ2.89
í4.26
Ѹ4.35
í8.02
Ѹ8.27
Item 3
Bayesian
MML
0
0
2.70
2.73
5.55
5.74
7.42
7.80
9.16
9.61
0
0
í0.73
Ѹ0.79
í2.77
Ѹ2.99
í
Ѹ6.75
í9.25
Ѹ9.90
Item 4
Bayesian
MML
0
0
1.64
1.62
3.27
3.22
4.59
4.58
7.80
8.00
0
0
0.33
0.29
í0.42
Ѹ0.47
í2.22
Ѹ2.33
í7.50
Ѹ7.97
Item 5
Bayesian
MML
0
0
1.29
1.24
3.33
3.30
4.90
4.85
6.88
6.83
0
0
0.79
0.77
1.23
1.16
0.30
0.31
í1.21
Ѹ1.34
Item 6
Bayesian
MML
0
0
2.52
2.52
4.89
4.93
6.40
6.48
8.01
8.10
0
0
í0.13
Ѹ0.18
í1.50
Ѹ1.62
í3.98
Ѹ4.17
í6.21
Ѹ6.47
Item 7
Bayesian
MML
0
0
2.10
2.09
4.59
4.60
6.26
6.32
9.54
9.63
0
0
0.55
0.51
í0.24
Ѹ0.34
í2.91
Ѹ3.06
í6.46
Ѹ6.76
Item 8
Bayesian
MML
0
0
1.70
1.69
3.87
3.84
5.52
5.47
8.41
8.33
0
0
1.07
1.04
1.31
1.24
0.15
0.04
í2.45
Ѹ2.59
Item 9
Bayesian
MML
0
0
2.23
2.23
3.24
3.26
4.57
4.59
5.73
5.77
0
0
í1.70
Ѹ1.74
í2.86
Ѹ2.94
í5.25
Ѹ5.43
í6.65
Ѹ6.80
Item 10
Bayesian
MML
0
0
0.77
0.77
2.02
2.00
3.63
3.59
5.38
5.33
0
0
1.39
1.38
2.37
2.32
2.13
2.05
1.54
1.43
Table 5.3: Comparison of Bayesian and MML Item Parameter Estimates for the GPC Model
Estimation
a
b
IJ1
IJ2
IJ3
Item 1
Bayesian
1.04
1.64
0.36
0.74
Ѹ0.53
Ѹ0.57
MML
1.09
1.59
0.36
0.70
Ѹ0.50
Ѹ0.57
IJ4
Item 2
Bayesian
1.18
1.55
0.20
0.53
Ѹ0.13
Ѹ0.60
MML
1.24
1.50
0.21
0.50
Ѹ0.13
Ѹ0.58
Item 3
Bayesian
2.46
1.15
0.84
0.40
Ѹ0.58
Ѹ0.66
Ѹ0.65
MML
2.56
1.12
0.82
0.38
Ѹ0.56
Item 4
Bayesian
1.61
0.88
1.08
0.41
Ѹ0.48
Ѹ1.01
MML
1.67
0.86
1.06
0.39
Ѹ0.45
Ѹ0.99
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 151
Estimation
Item 5
Bayesian
a
b
IJ1
IJ2
1.61
0.15
0.77
0.36
Ѹ0.47
Ѹ0.66
MML
1.66
0.14
0.75
0.35
Ѹ0.45
Ѹ0.65
IJ3
IJ4
Item 6
Bayesian
2.09
0.92
0.82
0.32
Ѹ0.56
Ѹ0.59
MML
2.17
0.89
0.80
0.31
Ѹ0.54
Ѹ0.57
Item 7
Bayesian
2.17
0.67
0.94
0.32
Ѹ0.81
Ѹ0.45
Ѹ0.44
MML
2.25
0.65
0.91
0.31
Ѹ0.78
Item 8
Bayesian
1.94
0.24
0.88
0.35
Ѹ0.49
Ѹ0.74
MML
2.01
0.23
0.86
0.34
Ѹ0.48
Ѹ0.72
Item 9
Bayesian
1.54
1.29
0.28
0.30
Ѹ0.53
Ѹ0.06
0.29
0.30
Ѹ0.52
Ѹ0.07
MML
1.60
1.24
Item 10
Bayesian
1.33
Ѹ0.43
1.10
0.28
Ѹ0.66
Ѹ0.73
MML
1.38
Ѹ0.42
1.08
0.27
Ѹ0.64
Ѹ0.71
Graphs of Item Category Response Functions
Differing from the dichotomous IRT models, polytomous items have more than one response curves, one
for each category. However, as illustrated in Chapter 4 to plot ICCs for binary response items, the category
response curves can be drawn for each polytomous item. Program 5.11 provides the commands for drawing
the item category response curves for the DASH items based on the GR model. The plots are shown in
Output 5.29.
In the program, point estimates for the item parameters are first obtained using PROC MEANS ߑ. These
point estimates are saved in a data set and used to generate a data set of response category probabilities for
each item ߒ. In this data set, the rows reflect discrete values of θ from −4 to 4 in increments of .1. Given
this data set, a graphics template can be defined to plot the category curves using a SERIESPLOT
command for each response category and for each item ߓ. A LAYOUT LATTICE command ߔis used to
graph all the category plots in a single graph with common scales for the x-axes and y-axes.
Program 5.11: Graphing Item Category Response Curves for Polytomous IRT models—GR Model
%let nitems=10; /* number of items */
%let ncats=5;
/* number of response categories */
%let nthres=4; /* number of threshold parameters */
/* save point estimates for item parameters*/
proc means data=dash_postout noprint;ߑ
var a1-a&nitems b1_1-b1_&nitems b2_1-b2_&nitems b3_1-b3_&nitems
b4_1-b4_&nitems;
output out=means_gr mean=;
run;
/* create dataset of expected item response probabilities across theta */
data plotdata_gr;ߒ
set means_gr;
array a[&nitems] a1-a&nitems;
array b[&nthres,&nitems] b1_1-b1_&nitems b2_1-b2_&nitems
b3_1-b3_&nitems b4_1-b4_&nitems;
152 Bayesian Analysis of Item Response Theory Models Using SAS
array p_star[&nitems,&nthres];
array p[&nitems,&ncats];
retain group(0);
do theta=-4 to 4 by .1;
group=group+1;
do j=1 to &nitems;
do k=1 to 4;
p_star[j,k]=logistic(a[j]*(theta-b[k,j]));
end;
p[j,1]=1-p_star[j,1];
do k=2 to 4;
p[j,k]=p_star[j,(k-1)]-p_star[j,k];
end;
p[j,5]=p_star[j,4];
end;
output;
end;
run;
/* define a graph template for category icc plots using seriesplots */
proc template;
define statgraph iccplots;
begingraph / designwidth=630px designheight=840px;
entrytitle "Category ICCs for the &nitems items – GR Model";
layout lattice / columns=3 rows=4 rowgutter=10 columngutter=10;ߔ
layout overlay /
xaxisopts=(label="Theta" linearopts=(viewmin=-4 viewmax=4))
yaxisopts=(label="Item 1" linearopts=(viewmin=0 viewmax=1));
seriesplot x=theta y=p1;ߓ
seriesplot x=theta y=p2;
seriesplot x=theta y=p3;
seriesplot x=theta y=p4;
seriesplot x=theta y=p5;
endlayout;
*Repeat for Item2~Item9;
layout overlay /
xaxisopts=(label="Theta" linearopts=(viewmin=-4 viewmax=4))
yaxisopts=(label="Item 10" linearopts=(viewmin=0 viewmax=1));
seriesplot x=theta y=p46;
seriesplot x=theta y=p47;
seriesplot x=theta y=p48;
seriesplot x=theta y=p49;
seriesplot x=theta y=p50;
endlayout;
endlayout;
endgraph;
end;
run;
/* produce the plots using the template */
proc sgrender data=plotdata_gr template=iccplots;
run;
Chapter 5: Bayesian Estimation of Unidimensional IRT Models for Polytomously Scored Items 153
Output 5.29: Category ICC Plots for the 10 DASH Items
Graphs of Test Information Functions
A program to illustrate the graphing of test information functions for the GR model is also included on the
SAS Press authors’ web page for this book (Program 5.12). Test information functions are computed by
summing up the information about θ provided by each item. These functions are useful in the test
development process for exploring the degree of precision of a test or differing sets of items across the θ
scale (see, for example, Hambleton & Swamination 1985).
154 Bayesian Analysis of Item Response Theory Models Using SAS
From Bayesian Analysis of Item Response Theory Models Using SAS®, by Clement A.
Stone and Xiaowen Zhu. Copyright © 2015, SAS Institute Inc., Cary, North Carolina,
USA. ALL RIGHTS RESERVED.
From Bayesian Analysis of Item Response Theory Models
Using SAS®. Full book available for purchase here.
Index
A
aberrant 206
Akaike Information Criterion (AIC) 39–40, 190
ALPHA option 79–80
analysis, of multiple chains 67–69
See also Bayesian Analysis
Andrich, D. 11, 13
See also Rating Scale (RS) model
arbitrary distributions 48–49
ARRAY statement, MCMC procedure 45, 50, 53, 211
autocall macros, for post-processing MCMC procedure
data sets 78–79
B
Bayes factor (BF) 40–41, 190–191
Bayesian Analysis
See also specific topics
about 25
Bayesian approaches to statistical inference 26–28
burn-in 35–38
convergence 35–38
elements of statistical models and inference 25–26
estimating proportions with 28–31
frequentist approaches to statistical inference 26–
28
informative prior distributions 38–39
MCMC method and Bayesian estimation 31–35
model comparisons 39–41
model fit and posterior predictive model checks
41–42
uninformative prior distributions 38–39
Bayesian comparison, of IRT models 189–204
Bayesian comparison model indices
about 190
comparing GR and RS-GR models for
polytomously scored items 196–200
comparing models for dichotomously scored items
193–196
comparing unidimensional IRT models with
bifactor IRT models 200–204
computing model comparison statistics in MCMC
procedure 191–192
Conditional Predictive Ordinate (CPO) 190–191
Deviance Information Criterion (DIC) 190
Bayesian estimates
comparing frequentist estimates with 30–31
of IRT models using MCMC procedure 43–80
to statistical inference 26–28
of unidimensional IRT models for dichotomously
scored items 81–112
of unidimensional IRT models for polytomously
scored items 113–153
Bayesian Information Criterion (BIC) 39–40, 190
Bayesian model-checking, for IRT models 205–247
BEGINNODATA statement 102
Béguin, A.A. 213
BETA distribution 28–29, 47
between-item multidimensionality 14–15
bifactor IRT model
about 155–156
compared with unidimensional IRT models 200–
204
estimating in MCMC procedure 156–158
"label switching" 165
other multidimensional IRT models 163–165
output from MCMC 158–163
BINARY distribution 47
BINOMIAL distribution 47
Bock, R.D. 11, 52, 136
Bradlow. E.T. 13, 166
built-in distributions 46–47
burn-in 35–38
BY option, RANK procedure 226
C
Cai, L. 11
CALL SYMPUTX statement 218
chain thinning 59–60
chains, analysis of multiple 67–69
Chen, W.H. 20, 214
CHISQUARE distribution 47
chi-squared statistic 207
Cho, S. 17
classical test theory (CTT) 1–2
Cohen, A.S. 17
comparing
GR and RS-GR models for polytomously scored
items 196–200
models for dichotomously scored items 193–196
models using PPMC 243–247
results based on MML estimation 105–107, 149–
151
unidimensional IRT models and bifactor IRT
models 200–204
with WinBUGS 76–77
computing
drawing inferences 30
model comparison statistics in MCMC procedure
191–192
posterior inferences 30
Conditional Predictive Ordinate (CPO)
about 40–41, 190–191
results for comparing models for dichotomously
scored items 194–196
results for comparing models for polytomously
scored items 197–200
results for comparing unidimensional and bifactor
IRT models 201–204
258 Index
convergence
about 35–38
Gelman-Rubin test for 67–69
strategies to improve of Monte Carlo Chain 59–66
CORR procedure 79–80, 220–223
Cronbach's Coefficient Alpha 79
D
DASH survey 114
DATA option, MCMC procedure 44
Dbar 193
De Ayala, R.J. 11, 22–23
DeMars, C.E. 156, 163, 172
density plots 47–48
Deviance Information Criterion (DIC)
about 39–40, 190
results for comparing GR and RS-GR modes for
polytomously scored items 196–197
results for comparing models for dichotomously
scored items 193–194
results for comparing unidimensional and bifactor
IRT models 200–201
DGENERAL function 49
DIAGNOSTICS option, MCMC procedure 45, 59
DIC option, MCMC procedure 45
dichotomously scored items
Bayesian estimation of unidimensional IRT models
for 81–112
comparing models for 193–196
unidimensional IRT models for 5
“difference models” 7
Differential Item Functioning (DIF)
mixture IRT models and 16–17, 182–187
multiple group models for detecting 180–182
difficulty parameter 4
diffuse priors
See uninformative prior distributions
dimensionality 19
discrepancy measures
about 212–213
evaluating model fit 215
item-level measures 213–214
pairwise measures 214
person-fit measures 214–215
test-level measures 213
discrimination parameter 4
displaying item response functions 107–112
distributions
arbitrary 48–49
BETA 28–29, 47
BINARY 47
BINOMIAL 47
built-in 46–47
CHISQUARE 47
GAMMA 47
GENERAL 47
informative prior 38–39, 70
MVN 47
MVNAR 47
nontruncated 90
NORMAL 47
T 47
TABLE 47
total score 216
truncated 90
UNIFORM 47
uninformative prior 38–39, 70–71
Dmean 193–194
DO loop 50, 53, 64, 116, 129
drawing inferences, computing 30
E
Effective Sample Size (ESS) 38, 59
Embretson, S.E. 6, 15
ENDNODATA statement 102
Englehard, G. 12
estimating
bifactor IRT model in MCMC procedure 156–158
Generalized Partial Credit (GPC) model 144–148
hierarchical models 173–175
Nominal Response (NR) model 136–143
1-Parameter (1P) graded response model 129–130
proportions with Bayesian Analysis 28–31
Rating Scale (RS) model 131–135
testlet IRT model 166–167
evaluating
IRT model applications 19–22
model fit 215
F
factor model 163–164
flat priors
See uninformative prior distributions
Fox, J.-P. 18–19, 73, 173, 178
FREQ procedure 216, 227, 231
frequentist estimates
comparing Bayesian estimates with 30–31
to statistical inference 26–28
functions
See specific functions
G
GAMMA distribution 47
Gelman, A. 46, 71, 166
Gelman-Rubin test
about 38, 59
for convergence 67–69
GENERAL distribution 47
GENERAL function 48, 120
Generalized Partial Credit (GPC) model
about 11, 113, 143–144
estimating 144–148
Gewecke statistics 38
Index 259
Gibbons, R.D. 163
Gibbs sampler, MCMC method and 35
Glas, C.A.W. 173, 213
global odds ratios 214
graded response (GR) model
about 6–8, 113, 114
comparing posterior densities for prior
specifications 127–128
comparing separate and joint prior specifications
120–121
comparing with RS-GR models for polytomously
scored items 196–200
computing transformation by post-processing
posterior results 128
estimation of using Method 1 118–120
output from estimating 121–127
program template for 114–116
specifying prior distribution for threshold and
intercept parameters 116–118
Graphics Template Language (GTL) 206, 216
graphs
if item category response functions 151–153
of test information functions 153
guessing parameter 4
H
Hambleton, R.K. 20, 95, 226
Han, Kyung T. 96
Heidelberger-Welch test
for 1P model 58
about 38
hierarchical models
about 17–19
multilevel IRT model 172–179
hierarchical priors
2-Parameter (2P) model with 91–94
about 29, 70, 71–72
hierarchical rating (HR) model 14
HISTOGRAM statement 218–219
hyperparameters 29
hyperpriors 29
hypothesis testing, compared with classical model-fit
analyses 215
testlet 166–172
IRT models
See also specific types
about 2–4
Bayesian comparison of 189–204
Bayesian estimation of using MCMC procedure
43–80
Bayesian model-checking for 205–247
evaluation of applications 19–22
form of 20
MCMC procedure for estimation 49–59
for multidimensional response data 14–16
multilevel 17–19
parameter estimation 22–23
random effects 17–19
testing effects in unidimensional 12–14
unidimensional, for dichotomously scored
responses 5
unidimensional, for polytomously scored responses
5–12
visualization of 3–4
IRT procedure 149
IRTPRO computer program 149
item analyses, preliminary 79–80
item category response functions, graphs of 151–153
item characteristic curves (ICCs) 15
item fit plots
of proportion correct statistics 223–226
using Yen's Q1 measure 226–231
item parameters
treatment of as random effects 73–75
uncertainty in 109–111
item response functions, displaying 107–112
item response theory (IRT) 1–4
See also IRT models
item-free measurement 2
item-level fit 206–207
item-level measures 213–214
J
Johnson, M.S. 14
JOINTMODEL option, MCMC statement 51
Junker, B.W. 14
I
K
“ideal rating” 14
IGAMMA distribution 47
implementing Metropolis sampler 32–34
informative prior distributions 38–39, 70
intercept parameter, specifying prior distributions for
116–118
inverse gamma distribution 91
IRT model extensions
about 155
bifactor model 155–165
differential item functioning 180–187
hierarchical models 172–179
Karabatsos, G. 207
KEEP statement 211
Kelderman, H. 17
Kopec, J.A. 184
L
“label switching” 165, 184–185
Lane, S. 8
Lieberman, M. 52
likelihood model, specifying using TABLE function
129
260 Index
Linacre, J.D. 13
local independence 19–20
locally dependent item sets (testlets) 13
LOGISTIC function 54, 85
LOGNORMAL distribution 47
LOOPITEMS macro 227
Lord, F.N. 95, 102, 107
lower asymptote parameter 3–4
LPDFNORM function 119–120
M
Maij-de Meij, A.M. 17
MAKEGTLLIST macro 109
many-facet Rasch measurement model (MFRM) 13
marginal maximum likelihood (MML) estimates 22,
105–107, 149–151
Mariano, L.T. 14
Markov Chain Monte Carlo (MCMC) method 31–32,
35, 43–44
Masters, G.N. 11
MCMC method
Bayesian estimation and 31–35
Gibbs sampler and 35
MCMC procedure
about 31–32, 35–38, 121–127
autocall macros for data sets 78–79
Bayesian estimation of IRT models using 43–80
Bayesian estimation of unidimensional IRT models
for dichotomously scored items 81–112
computing model comparison statistics in 191–192
estimating bifactor IRT model in 156–163
Graded Response (GR) model and 114–116
processing of SAS data set by 51
sampling algorithms 45–46
statements for IRT model estimation 44–49
template for IRT model estimation 49–59
MEANS procedure 54–55, 77–78, 107, 128, 141, 151–
153, 218, 223–224, 227, 232
Meijer, R.R. 207
Metropolis sampler, implementing 32–34
Metropolis-Hastings algorithm 31–32, 35
MISSING option, MCMC procedure 44
mixed effects models
See hierarchical models
mixture IRT models 16–17, 182–187
model fit
about 20–22, 76
evaluating 215
posterior predictive model checks and 41–42
model fit statistics
about 206
item-level fit 206–207
person fit 207
test-level fit 206
model identification 52
model re-parameterizing 64–66
MODEL statement, MCMC procedure 45, 47–48, 50,
51–52, 54, 114–116, 120
model-predicted results, compared with observed
results 206
models
comparing 39–41, 76
item response theory (IRT) 2–4
MONITOR command 64
MONITOR option, MCMC procedure 44, 54–55, 74,
82
Monte Carlo Chain 31–32, 59–66
Monte Carlo standard error (MCSE) 57
multidimensional response data, IRT models for 14–16
multilevel IRT model
about 17–19, 172
description of 172–173
estimating 173–175
output from MCMC procedure 175–179
multilevel models
See hierarchical models
multiple group models, for detecting differential item
functioning 180–182
Muraki, E. 8, 143–144
See also Rating Scale (RS) model
MVN distribution 47
MVNAR distribution 47
N
NBI option, MCMC procedure 44
NMC option, MCMC procedure 44
Nominal Response (NR) model
about 11–12, 113, 136
estimating 136–143
nontruncated distribution 90
normal deviate 85
NORMAL distribution 47
normal ogive function 85
NTHREADS option, MCMC procedure 44, 82
O
observed and predicted item-test score correlations, as
an example of PPMC application 220–223
observed and predicted odds ratio measure, as an
example of PPMC application 231–235
observed and predicted person-fit statistic, as an
example of PPMC application 239–243
observed and predicted Yen's Q3 measure, as an
example of PPMC application 236–239
observed results 111–112, 206
observed test score distributions, as an example PPMC
application 216–220
odds ratio (OR) 214
1-Parameter (1P) model
about 5
Bayesian estimation of unidimensional IRT models
82–84
example output from estimating 54–59
Index 261
graded response, estimating 129–130
Heidelberger-Welch test for 58
template for 52–54
option discrimination parameter 11
option extremity parameter 11
options
See specific options
OUTPOST option, MCMC procedure 44, 121, 128
OUTPOST SAS data set 77–78
output
from estimating GPC model 146–148
from estimating hierarchical models 175–179
from estimating NR model 137–143
from estimating RS-GR model 133–135
from estimating testlet IRT model 167–172
example from estimating 1-Parameter (1P) model
54–59
from MCMC procedure when estimating bifactor
IRT model 158–163
“overfitting” 26
prior distributions
choosing 29
specifying for threshold and intercept parameters
116–118
prior specifications, comparing separate and joint 120–
121
PRIOR statement, MCMC procedure 45, 46–47, 47–
48, 48–49, 51–52, 53, 116, 120
priors, effect of different 72–73
PROBNORM function 54
procedures
See specific procedures
program template, for graded response (GR) model
114–116
proportion correct statistics, item fit plots and 223–226
proportions
comparing Bayesian and frequentist estimates for
30–31
estimating with Bayesian Analysis 28–31
Pseudo-Bayes factor (PsBF) 40–41, 190–191
P
R
pairwise measures 214
parameter blocking 61–63
parameter estimation, in IRT models 22–23
PARMS statement, MCMC procedure 45, 47–48, 50,
51, 53, 61–63, 74, 117, 131–132, 173
Partial Credit (PC) model 8–11, 113
Patz, R.J. 14
Pearson chi-squared statistic 213
person fit 207
person-fit measures 214–215
person-free measurement 2
PLOTS option, MCMC procedure 45
polytomously scored items
about 4
Bayesian estimation of unidimensional IRT models
for 113–153
comparing models for 196–200
unidimensional IRT models for 5–12
posterior densities, comparing for prior specifications
127–128
posterior inferences, computing 30
posterior predictive distribution 208–212
posterior predictive model-checking (PPMC)
about 21–22, 39, 76, 205–206, 207
examples of applications 216–247
method 207–208
model fit and 41–42
posterior predictive distribution 208–212
posterior predictive probability values (PPP values) 42
%POSTSUM macro 121, 137, 146, 183
PREDDIST statement, MCMC procedure 45, 208–209,
211, 216, 220–223, 231
predicted test score distributions, as an example PPMC
application 216–220
preliminary item analyses 79–80
Raftery-Lewis statistic 38
Ramsay, J.O. 2
random effects, treatment of item parameters as 73–75
random effects models 17–19
See also hierarchical models
RANDOM statement, MCMC procedure 45–46, 46–
47, 47–48, 50, 51–52, 54, 74, 77–78, 158,
167, 180–181, 182–184
RANK procedure 226
Rasch, G. 5, 15, 82–84
rater effects 13–14
Rating Scale (RS) model 8, 11, 13, 113, 131–132
Ratner, P.A. 184
Reckase, M.D. 19, 200
Reise, S.P. 6
re-parameterizing models 64–66
RESTRUCTURE macro 227
Rogers, H.J. 226
RS-GR model
comparing with GR models for polytomously
scored items 196–200
output from estimating 133–135
S
Samejima, F. 6, 114
See also Graded Response (GR) model
sampling algorithms 45–46
Santor, D.A. 2
Sawatzky, R. 184
SCATTERPLOT statement 222
SEED option, MCMC procedure 44
SERIESPLOT statement 216, 224
SGRENDER procedure 107
Sijtsma, K. 207
Sinharay, S. 213
262 Index
slope parameter 3
SPARSE option 227
specifying likelihood model using TABLE function
129
speededness 20
standard error of measurement (SEM) 1–2
statements
See specific statements
statistical inference
Bayesian approaches to 26–28
elements of 25–26
frequentist approaches to 26–28
statistical models, elements of 25–26
STATISTICS option, MCMC procedure 45
Steinberg, L. 7, 11
Stone, C.A. 8
Stout, W.F. 19
SUBJECT option 52
SURVEYSELECT procedure 189, 210
Swaminathan, H. 20, 95, 226
T
T distribution 47
TABLE distribution 47
TABLE function, specifying likelihood model using
129
%TADPLOT macro 122, 139
TEMPLATE procedure 107
templates, for 1-Parameter (1P) model 52–54
test information functions, graphs of 153
testing effects, in unidimensional IRT models 12–14
testlet IRT model
about 166
estimating 166–167
output from MCMC procedure 167–172
testlets 13
test-level fit 206
test-level measures 213
THETA parameter 51
THIN option, MCMC procedure 44
Thissen, D. 7, 11, 20, 214
3-Parameter (3P) model
about 5
Bayesian estimation of unidimensional IRT models
94–105
threshold parameter 3, 116–118
total score distribution 216
truncated distribution 90
2-Parameter (2P) model
about 5
Bayesian estimation of unidimensional IRT models
85–91
with hierarchical priors 91–94
Bayesian estimation of for dichotomously scored
items 81–112
Bayesian estimation of for polytomously scored
items 113–153
compared with bifactor IRT models 200–204
for dichotomously scored responses 5
for polytomously scored responses 5–12
testing effects in 12–14
UNIFORM distribution 47
uninformative prior distributions 38–39, 70–71
V
vague priors
See uninformative prior distributions
VARB parameter 72
visualization, of IRT models 3–4
W
Wainer, H. 13, 166
Wang, X. 13
WinBUGS 43–44, 76–77
within-item multidimensionality 14–15
Wright, B.D. 11
Y
Yen, W.M. 20
Yen's Q1 statistic 213, 226–231
Yen's Q3 statistic 214
Z
Zumbo, B.D. 16, 184
U
“underfitting” 26
unidimensional IRT models
From Bayesian Analysis of Item Response Theory Models Using SAS®, by Clement A.
Stone and Xiaowen Zhu. Copyright © 2015, SAS Institute Inc., Cary, North Carolina,
USA. ALL RIGHTS RESERVED.
About This Book
Purpose
Assessment has played and continues to play an integral role in society and in a culture of inquiry.
Accordingly, numerous instruments have been developed over the years to measure many different
characteristics of individuals in the educational, psychological, health, social, and behavioral sciences.
These instruments have been used to measure the status or level of different characteristics in individuals,
as well as to capture differences in these characteristics across individuals.
To support the development of instruments and measure characteristics in individuals, test theories have
been used to describe how inferences, predictions, or estimates of a particular characteristic, trait, or ability
of a person may be made from responses to items. Test theories such as classical test theory (CTT) and
item response theory (IRT) provide models for explaining test performance in relation to variables that are
assumed to influence behavior. They provide methods for selecting items, evaluating tests or scales,
obtaining scores, and quantifying sources of errors in the measurement process. In the early decades of the
21st century, psychometricians have favored IRT, as opposed to CTT, in scale development and assessment
applications.
IRT models consist of a family of mathematical models that predict item performance by using parameters
that characterize both the items in an instrument and the respondents. Although numerous methods for
estimating the parameters of IRT models exist, interest in estimating the parameters using Bayesian
methods has grown tremendously. In part, this growth is due to the appeal of the Bayesian paradigm among
psychometricians and statisticians, as well as to the advantages of these methods with small sample sizes,
more complex or highly parameterized models (such as multidimensional IRT models), and interest in
simultaneous estimation of item and person parameters. In contrast to traditional approaches for estimating
model parameters, a Bayesian paradigm considers model parameters to be random variables and uses Bayes
theorem to obtain distributions for the model parameters.
Recently, routines have become available in the SAS system software to implement general Bayesian
analyses (PROC MCMC). Use of the SAS system for Bayesian analysis of IRT models has several
significant advantages over other available programs: (1) It is commonly used by researchers across
disciplines; (2) it provides a robust programming language that extends the capability of the program—in
particular, the capability for model checking; and (3) it shows increased performance and efficiency
through the use of parallel processing. The purpose of this book is to illustrate Bayesian estimation and
evaluation of a variety of IRT models that are of interest to psychometricians, scale developers, and
practitioners responsible for implementing assessment programs.
Is This Book for You?
This book is designed for psychometricians, scale developers, and individuals interested in applications of
Bayesian methods and model checking of IRT models for the development and analysis of assessment data.
In addition, you may find this book useful if you are interested in applications with small sample sizes,
applications requiring more complex or highly parameterized models (such as multidimensional IRT
models), and simultaneous estimation of item and person parameters.
x Bayesian Analysis of Item Response Theory Models Using SAS
Prerequisites
Although introductory material related to IRT and Bayesian analysis is included in the book, some prior
knowledge about these topics is necessary for you to better understand and implement the procedures in
your own assessment applications. For detail and more technical information, see the references provided.
Scope of This Book
This book is example-driven, with sections presenting code for estimating particular models and results that
illustrate convergence diagnostics and inferences for parameters, as well as results that can be used
specifically by scale developers—for example, plotting item response functions.
Chapters 1 and 2 provide introductory treatments of IRT and Bayesian analysis. These chapters are not
meant to provide detailed information for individuals interested in learning about these topics. Rather, these
chapters provide supporting documentation for the subsequent chapters that describe Bayesian estimation
of different IRT models with the use of SAS PROC MCMC. Chapter 1, on IRT, motivates the use of the
different models and provides the model formulations that will be used in subsequent chapters. Chapter 2,
on Bayesian analysis, provides supporting documentation for terms and concepts integral to the use of
PROC MCMC and the subsequent examples.
Chapter 3 presents an overview of the PROC MCMC syntax commands and a template for the estimation
of IRT models, but focuses on syntax that is most relevant to the estimation of IRT models. As a result,
there is some duplication of this material with other SAS documentation for Bayesian analysis and the use
of PROC MCMC. However, presentation of PROC MCMC in this book provides a self-contained guide to
the use of PROC MCMC and Bayesian analysis of IRT models.
Chapters 4 to 6 provide detailed presentations of the syntax used to estimate IRT models for different types
of assessment applications. In these chapters, the IRT models and PROC MCMC syntax to estimate the
models are discussed. Examples then illustrate the generation of convergence diagnostics, use of different
prior distributions, and summary statistics from the posterior distributions for parameters. In some cases,
item responses from data sets that are commonly analyzed in the literature are used. In other cases, item
responses are simulated so that direct comparisons between estimated parameters and population
parameters can be made. The comparison between estimated parameters and parameters used to obtain
simulated item responses facilitates the evaluation of using PROC MCMC to estimate the models. In
addition, comparisons of model parameter estimates derived from using PROC MCMC with other available
software are also presented.
Specifically, Chapter 4 illustrates estimation of IRT models for dichotomously scored items, and Chapter 5
illustrates estimation of models for polytomously scored items. For both of these chapters, the models that
are discussed assume that a single person parameter determines the performance of respondents
(unidimensional IRT models).
Chapter 6 illustrates extensions of IRT models to more highly parameterized models, including
multidimensional IRT models, models that account for any dependence between sets of items (testlet IRT
models), multilevel IRT models, and applications to differential item functioning (DIF). Multidimensional
IRT models are considered when more than one person parameter is assumed. Testlet IRT models may be
used for testing applications in which individuals respond to a set of items based on a single or common
stimulus (testlet). Multilevel or hierarchical models may be estimated in testing applications in which there
exists a nested structure of item responses within clusters of persons or within clusters of items. Finally,
models for evaluating differential item functioning and the use of mixture IRT models are considered to
examine whether items function differently across subpopulations of examinees (for example, masters
versus nonmasters or males versus females). In the case of mixture IRT models, these models assume that
the population of examinees consists of qualitatively different subpopulations or latent classes that are
unknown. These models provide a mechanism for clustering examinees to better understand any
heterogeneity in item responses.
About this Book xi
Although the focus is often on estimating IRT model parameters, it is equally important to evaluate the fit
of a particular IRT model to the item responses. When a model doesn’t fit the data, the validity of any
inferences for the model parameter estimates is threatened. The SAS system and PROC MCMC affords
advantages over other available software for Bayesian analysis of IRT models. For evaluating model-datafit, in particular, SAS provides a robust programming language and built-in statistical and graphical tools.
The programming language and the tools expand the capability of the program beyond estimating the
model to computing many different types of statistics for comparing competing models and evaluating
model fit.
The topics of model comparison and model checking in a Bayesian analysis are discussed in Chapters 7 and
8 of this book. Chapter 7 considers methods for comparing competing models so that a preferred model can
be identified. Methods based on information criteria indices and the Bayes Factor are considered. Chapter 8
begins with a discussion of item and person fit in the context of IRT models and a description of a popular
model checking tool in Bayesian analysis—posterior predictive model checking (PPMC). Then possible
discrepancy statistics for a PPMC analysis of IRT models are discussed, followed by examples illustrating
the use of SAS procedures and programming commands to perform a PPMC analysis. The use of output in
the form of tables and graphs is also included to illustrate the capabilities of a PPMC analysis in SAS, as
well as the different types of output that can be produced to report a PPMC analysis.
About the Examples
Software Used to Develop the Book’s Content
PROC MCMC statements and examples using PROC MCMC in this book are based on SAS 9.4 and
SAS/STAT 13.2.
Example Programming Code and Data
Examples are provided to estimate IRT models for dichotomously and polytomously scored items, as well
as more complex IRT models, such as multidimensional and hierarchical models.
You can access the example programs and data sets for this book by linking to its author page at
http://support.sas.com/publishing/authors/stone.html. Select the name of the author. Then look for the cover
thumbnail of this book, and select Example Code and Data to display the SAS programs that are included
in this book.
For an alphabetical listing of all books for which example code and data are available, see
http://support.sas.com/bookcode. Select a title to display the book’s example code.
If you cannot attain access to the code through the website, send email to [email protected].
Output and Graphics
All output was derived from programs using PROC MCMC or from other SAS programs. All graphics
were obtained with the use of SAS Graphics Template Language with default SAS Output Delivery System
settings.
xii Bayesian Analysis of Item Response Theory Models Using SAS
Additional Help
Although this book provides examples of many of the IRT analyses used by psychometricians, questions
specific to your aims and issues may arise. To fully support you, SAS Institute and SAS Press offer you the
following help resources:
Ɣ
For questions about topics covered in this book, contact the author through SAS Press:
ƕ
ƕ
Ɣ
Ɣ
Ɣ
Send questions by email to [email protected]. Include the book title in your correspondence.
Submit feedback on the author’s page at http://support.sas.com/author_feedback.
For questions about topics in or beyond the scope of this book, post queries to the relevant SAS
Support Communities at https://communities.sas.com/welcome.
SAS Institute maintains a comprehensive website with up-to-date information. One page that is
particularly useful to both the novice and the seasoned SAS user is its Knowledge Base. Search for
relevant notes in the Samples and SAS Notes section of the Knowledge Base at
http://support.sas.com/resources.
Registered SAS users or their organizations can access SAS Customer Support at
http://support.sas.com. Here you can pose specific questions to SAS Customer Support. Under
Support, click Submit a Problem. You will need to provide an email address to which replies can
be sent, identify your organization, and provide a customer site number or license information.
This information can be found in your SAS logs.
Keep in Touch
We look forward to hearing from you. We invite questions, comments, and concerns. If you want to contact
us about a specific book, please include the book title in your correspondence.
Contact the Author through SAS Press
Ɣ
By email: [email protected]
‡
Via the web: http://support.sas.com/author_feedback
Purchase SAS Books
For a complete list of books available through SAS, visit sas.com/store/books.
Ɣ
Phone: 1-800-727-0025
‡
E-mail: [email protected]
Subscribe to the SAS Training and Book Report
Receive up-to-date information about SAS training, certification, and publications via email by subscribing
to the SAS Training & Book Report monthly eNewsletter. Read the archives and subscribe today at
http://support.sas.com/community/newsletters/training!
Publish with SAS
SAS is recruiting authors! Are you interested in writing a book? Visit http://support.sas.com/saspress for
more information.
About These Authors
Clement A. Stone is a professor in the Research Methodology program
at the University of Pittsburgh School of Education. He is an expert in
psychometrics, including educational and psychological instrument
development and validation, item response theory (IRT) models and
applications, and Bayesian analysis of IRT models. Also an expert in
SAS software, he has used SAS extensively in research utilizing
simulation methods, and he instructs graduate students in the use of
SAS. He applies IRT in the development and validation of educational,
psychological, and behavioral assessments, including a research focus
on educational achievement, critical thinking, psychosocial stress,
communication outcomes, risk factors for addiction, and physical disability. He has
published numerous articles, and he coauthored a chapter on the development and
analysis of performance assessments in Educational Measuremen. In addition to
publishing in and reviewing for numerous prominent journals, he has served on the
editorial boards for the Journal of Educational Measurement, Applied Measurement in
Education, Educational and Psychological Measurement, and the American Educational
Research Journal. Stone holds a Ph.D. in research methods, measurement, and statistics
from the University of Arizona.
Xiaowen Zhu is an associate professor in the Department of Sociology
and a research fellow in the Institute for Empirical Social Science
Research, Xi’an Jiaotong University, China. Formerly, she was a
psychometrician for the Data Recognition Corporation in the United
States, working on large-scale state educational assessments. She has a
strong theoretical background, operational experience, and research
experience in psychometrics and applied statistics. She is especially
adept at Bayesian analysis of IRT models and has published journal
articles on this subject. In addition, she has used SAS throughout her research career for
general statistical analysis and simulation studies. Zhu holds a Ph.D. in research
methodology from the University of Pittsburgh.
Learn more about these authors by visiting their author pages, where you can download
free book excerpts, access example code and data, read the latest reviews, get updates,
and more:
http://support.sas.com/publishing/authors/stone.html
http://support.sas.com/publishing/authors/zhu.html
Gain Greater Insight into Your
SAS Software with SAS Books.
®
Discover all that you need on your journey to knowledge and empowerment.
support.sas.com/bookstore
for additional books and resources.
SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA registration. Other brand and product names
are trademarks of their respective companies. © 2013 SAS Institute Inc. All rights reserved. S107969US.0613
">