flexMIRT User Manual

R
flexMIRT
: Flexible Multilevel
Multidimensional Item Analysis and Test
Scoring
User’s Manual
Version 2.0
Authored by:
Carrie R. Houts, PhD
Li Cai, PhD
The correct citation for the user’s manual is:
R
Houts, C. R., & Cai, L. (2013). flexMIRT
user’s manual version 2:
Flexible multilevel multidimensional item analysis and test scoring. Chapel
Hill, NC: Vector Psychometric Group.
The correct citation for the software is:
R
Cai, L. (2013). flexMIRT
version 2: Flexible multilevel multidimensional
item analysis and test scoring [Computer software]. Chapel Hill, NC: Vector
Psychometric Group.
c
Copyright 2013
Vector Psychometric Group, LLC
October 2013
Contents
1 Overview
1
2 Basic Calibration and Scoring
2.1 Overview . . . . . . . . . . . . . . . . . . . .
2.2 Single-Group 2PL Model Calibration . . . .
2.3 Single-Group 3PL Model Calibration . . . .
2.4 Single-Group 3PL EAP Scoring . . . . . . .
2.5 Single-Group 3PL ML Scoring . . . . . . . .
2.6 Graded Model Calibration and Scoring . . .
2.7 1PL Model Fitted to Response Pattern Data
2.8 The Nominal Categories Model . . . . . . .
2.8.1 Nominal Model . . . . . . . . . . . .
2.8.2 Generalized Partial Credit Model . .
2.8.3 Rating Scale Model . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
5
12
15
19
20
23
29
29
36
37
3 Advanced Modeling
3.1 Multiple Group Calibration . . . . . . . . . . . . . . . . . . .
3.2 Characterizing Non-Normal Population Distributions . . . . .
3.3 Multiple Group Scoring Only . . . . . . . . . . . . . . . . . .
3.4 DIF Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Multiple Group Multilevel Model . . . . . . . . . . . . . . . .
3.6 Multidimensional Model - 4 Correlated Factors . . . . . . . . .
3.7 The Bifactor Model . . . . . . . . . . . . . . . . . . . . . . . .
3.7.1 Bifactor Model Using the Reparameterized Nominal Categories Model . . . . . . . . . . . . . . . . . . . . . . .
3.7.2 Bifactor Model Using the 2PL . . . . . . . . . . . . . .
40
40
42
43
45
49
53
56
4 Simulation
4.1 A Syntax-Only Example Using Value Statement . . . . . . . .
4.2 The First Example Redone Using Parameter File . . . . . . .
61
61
64
i
56
59
4.3
4.4
4.5
4.6
4.7
4.8
4.9
The Parameter File Layout . . . . . . . . . . . . . . . .
4.3.1 Column Progression for Groups . . . . . . . . .
4.3.2 Column Progression for Items . . . . . . . . . .
A 3PL Model Simulation Example . . . . . . . . . . .
Simulating Data with Non-Normal Theta Distributions
Simulating Bifactor Model Data . . . . . . . . . . . . .
Simulating Group Differences . . . . . . . . . . . . . .
Multilevel Bifactor Model Simulation . . . . . . . . . .
Internal Simulation Function . . . . . . . . . . . . . . .
4.9.1 Accessing the Internal Simulation Function . . .
4.9.2 The Simulation Output File . . . . . . . . . . .
5 Metropolis-Hastings Robbins-Monro Estimation for
Dimensional Models
5.1 Overview of MH-RM Estimation . . . . . . . . . . . .
5.2 MH-RM Specific Syntax Options . . . . . . . . . . . .
5.3 Examples Utilizing MH-RM Estimation . . . . . . . . .
5.3.1 MIRT Model with Correlated Factors . . . . . .
5.3.2 Multi-level MIRT Model . . . . . . . . . . . . .
5.4 Fixed Effects Calibration . . . . . . . . . . . . . . . . .
6 Cognitive Diagnostic Models
6.1 General CDM Modeling Framework . .
6.2 CDM Specific Syntax Options . . . . .
6.3 CDM Modeling Examples . . . . . . .
6.3.1 Introductory CDM Examples .
6.3.2 Replicating Published Examples
7 Details of the Syntax
7.1 The Project Section . . .
7.2 The Options Section . .
7.3 The Groups Section . . .
7.4 The Constraints Section
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 Models
8.1 A Generalized Formulation . . . . . . . . . . . .
8.2 Item Response Models . . . . . . . . . . . . . .
8.2.1 Model with Pseudo-Guessing Parameter
8.2.2 Graded Response Model . . . . . . . . .
ii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
66
67
68
71
72
74
77
80
80
82
High
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
85
85
87
89
89
93
97
.
.
.
.
.
99
100
101
102
103
123
.
.
.
.
138
138
138
151
157
.
.
.
.
162
162
163
163
164
8.2.3
Nominal Categories Model . . . . . . . . . . . . . . . .
Appendices
164
166
A Quick-Start Installation and User Guide
167
A.1 Installing flexMIRT . . . . . . . . . . . . . . . . . . . . . . . . 167
A.2 Using flexMIRT . . . . . . . . . . . . . . . . . . . . . . . . . . 172
B flexMIRT via the Command Line Interface
174
B.1 flexMIRT via the CLI - single file run . . . . . . . . . . . . . . 174
B.2 flexMIRT via the CLI - batch mode . . . . . . . . . . . . . . . 177
C Plotting
180
C.1 Plotting ICCs, TCCs, and Information Functions . . . . . . . 180
D Labeled Output Files
184
E Licensing Information
188
E.1 Commercial use of flexMIRT . . . . . . . . . . . . . . . . . . . 188
E.2 Classroom Discounts for flexMIRT . . . . . . . . . . . . . . . 188
References
189
Index
194
iii
List of Tables
4.1
4.2
4.3
4.4
4.5
Simulation - Graded Model PRM File Content . . . . . .
Key Codes for Columns With Set Options . . . . . . . .
Simulation - 3PL Parameter File with Labeled Columns .
Labeled Parameter File for Bifactor Model Simulation . .
Simulating DIF and Group Mean Differences - Parameter
6.1
6.2
6.3
6.4
6.5
103
119
120
123
6.6
Q-matrix for Basic CDM Demonstrations . . . . . . . . . . . .
Model Fit for Basic CMDs Fit to Simulated Data . . . . . . .
Q-matrix for Testlet-DINA Example . . . . . . . . . . . . . .
Q-matrix for Rupp, Templin, & Henson’s Example 9.2 . . . .
R
Generating latent class mean values and flexMIRT
and MPLUS
estimates for Rupp, Templin, & Henson’s Example 9.2 . . . .
Guessing and Slippage Parameters from flexMIRT and MCMC
7.1
Constraint Commands and Keywords . . . . . . . . . . . . . .
157
D.1
D.2
D.3
D.4
Labeled -sco file - EAP scores, 1 Factor . . . . . . . . . . .
Labeled -sco file - MAP scores, 1 Factor . . . . . . . . . .
Labeled -sco file - EAP scores, 2 Factors . . . . . . . . . .
Labeled -sco file - EAP scores, 2 Factors with User-supplied
Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Labeled -sco file - MAP scores, 2 Factors . . . . . . . . . .
Labeled -prm file - GRM with 1 Factor, 1 Group . . . . . .
Labeled -prm file - 2PL with 6 Factors, 1 Group . . . . . .
Labeled -prm File - 3PL with 1 Factor, 3 groups . . . . . .
184
184
185
D.5
D.6
D.7
D.8
iv
. . .
. . .
. . .
. . .
Values
. .
. .
. .
ID
. .
. .
. .
. .
. .
65
66
70
73
76
128
137
185
185
186
186
187
List of Examples
2.1
2.2
2.3
2.4
2.5
2PL Calibration Syntax . . . . . . . . . . . . . . . . . . . . .
Single Group 3PL Calibration Syntax . . . . . . . . . . . . . .
3PL EAP Scoring example . . . . . . . . . . . . . . . . . . . .
3PL ML Scoring example . . . . . . . . . . . . . . . . . . . . .
Graded Model Combined Calibration and Scoring Syntax with
Recoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 1PL Calibration and Scoring with Response Pattern Data . .
2.7 Nominal Model Calibration . . . . . . . . . . . . . . . . . . .
2.8 Nominal Model Calibration 2 . . . . . . . . . . . . . . . . . .
2.9 Generalized Partial Credit Model . . . . . . . . . . . . . . . .
2.10 Rating Scale Model with Trend Contrasts . . . . . . . . . . .
2.11 Rating Scale Model with Triangle Contrasts . . . . . . . . . .
3.1 2-Group 3PL Using Normal Priors with EAP Scoring . . . . .
3.2 2-Group 3PL with Empirical Histogram (Excerpts) . . . . . .
3.3 2-Group 3PL MAP Scoring from Parameter Estimates . . . .
3.4 2-Group 3PL Summed Score to EAP Conversion Table . . . .
3.5 DIF: Test All Items (Excerpts) . . . . . . . . . . . . . . . . .
3.6 DIF: Test Candidate Items (Excerpts) . . . . . . . . . . . . .
3.7 Multiple Group Multilevel Model (Excerpts) . . . . . . . . . .
3.8 Four Factor GRM Model with Correlated Factors . . . . . . .
3.9 Bifactor Structure - Nominal Model . . . . . . . . . . . . . . .
3.10 Bifactor Structure - 2PL . . . . . . . . . . . . . . . . . . . . .
4.1 Simulation - Graded Model Syntax Only . . . . . . . . . . . .
4.2 Simulation - Graded Model Using PRM File . . . . . . . . . .
4.3 Simulation - 3PL Model Normal Population . . . . . . . . . .
4.4 Simulation - Empirical Histogram Prior . . . . . . . . . . . . .
4.5 Simulation - Bifactor Structure From Existing Parameter File
4.6 Simulation - DIF and Group Mean Differences . . . . . . . . .
4.7 Simulation - Multilevel Model with Bifactor Structure . . . . .
v
5
12
16
19
20
24
29
32
36
37
38
40
42
44
44
45
48
50
53
57
59
62
64
69
71
72
75
77
5.1
5.2
5.3
5.4
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
C.1
<Options> - Engine Mode Settings to Call MH-RM Estimation
and MH-RM Specific Options . . . . . . . . . . . . . . . . . .
MHRM: 4 Factor MIRT Model . . . . . . . . . . . . . . . . .
MHRM: Two-level MIRT Model . . . . . . . . . . . . . . . . .
Fixed Effects Calibration . . . . . . . . . . . . . . . . . . . . .
<Options>, <Groups>, and <Constraints>- CDM Specific Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
C-RUM Model Fit To Simulated Data . . . . . . . . . . . . .
DINA Fit to Simulated Data . . . . . . . . . . . . . . . . . . .
DINO Fit to Simulated Data . . . . . . . . . . . . . . . . . . .
Testlet Structure with DINA model . . . . . . . . . . . . . . .
Rupp, Templin, & Henson - Example 9.2 . . . . . . . . . . . .
Rupp, Templin, & Henson - Example 9.2 without D group . .
de la Torre Subtraction Syntax . . . . . . . . . . . . . . . . .
<Project> section commands . . . . . . . . . . . . . . . . . .
<Options> - Engine Mode and Output Statements . . . . . .
<Options> - Technical Specifications . . . . . . . . . . . . . .
<Options> - GOF and LD Indices Commands . . . . . . . . .
<Options> - DIF Analysis Commands . . . . . . . . . . . . .
<Options> - MH-RM Specific Options . . . . . . . . . . . . .
<Groups> - General Data/Model Descriptors . . . . . . . . .
<Groups> - Specialized Data/Model Descriptors . . . . . . . .
Constraints Across Groups, Specifying a Prior Distribution, and
Freely Estimating Theta . . . . . . . . . . . . . . . . . . . . .
R plotting code . . . . . . . . . . . . . . . . . . . . . . . . . .
vi
87
89
93
97
101
104
111
115
121
123
131
134
138
139
141
145
148
149
151
154
159
181
List of Output
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
2.14
2.15
2.16
2.17
2.18
2.19
3.1
3.2
3.3
3.4
4.1
4.2
4.3
4.4
2PL Calibration Output - Summary and Controls . . . . . . .
2PL Calibration Output - Item and Group Parameters . . . .
2PL Calibration Output - Information Functions . . . . . . . .
2PLM Calibration Output - Goodness of Fit Indices . . . . . .
Single Group 3PL Calibration Output- Item Parameters . . .
Single Group 3PL Calibration Output- Extended GOF Table .
3PL EAP Scoring Output . . . . . . . . . . . . . . . . . . . .
3PL ML Scoring Output . . . . . . . . . . . . . . . . . . . . .
Graded Model Calibration - Slopes and Intercepts . . . . . . .
Graded Model Calibration - Slopes and Thresholds . . . . . .
Graded Model Calibration - Extended GOF Table . . . . . . .
1PL Grouped Data Output - Item Parameters and Factor Loadings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1PL Grouped Data Output - Item Fit Diagnostics . . . . . . .
1PL Grouped Data Output - Summed Score to Scale Score Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1PL Grouped Data Output - JSI matrix . . . . . . . . . . . .
1PL Grouped Data Output - Haberman Residual Table (Excerpt)
Nominal Model Output - Item Parameters . . . . . . . . . . .
Nominal Model 2 Output - Item Parameters . . . . . . . . . .
Nominal Model 2 Output - GOF values . . . . . . . . . . . . .
DIF Analyses Output Table . . . . . . . . . . . . . . . . . . .
Two-Level Model - Means and Variances . . . . . . . . . . . .
Correlated Four Factor Model Output - Loadings, Means and
Variances . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Nominal Model Scoring Function Estimates - Excerpt . . . . .
Simulation Control Output - Graded Model Syntax Only . . .
Simulated Data Set - Graded Model Syntax Only . . . . . . .
Simulated Data Set - Graded Model Using PRM File . . . . .
Simulation Control Output for Multilevel Bifactor Model . . .
vii
7
9
10
11
14
15
17
20
22
22
23
26
27
27
28
28
31
34
35
47
52
55
58
63
64
65
78
4.5
4.6
4.7
5.1
5.2
5.3
5.4
5.5
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
Beginning Columns of Output from 5 Replications . . . . . . .
82
Item Parameters with Parameter Number Labels . . . . . . .
82
ISF output - Switch from Model 1 to Model 2 Values . . . . .
83
Correlated Four Factor Model Output - Processing Pane Output 91
Correlated Four Factor Model Output - Loadings, Means and
Variances . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
Two-Level MIRT Model - Processing Pane Output . . . . . . .
95
Multilevel MIRT Model - Item and Group Parameters . . . . .
96
Fixed Effects Calibration Output - Item Parameters . . . . . .
98
C-RUM Output - Item Parameters and Attribute Profile Posterior Probabilities . . . . . . . . . . . . . . . . . . . . . . . . 107
C-RUM Output –SCO Output . . . . . . . . . . . . . . . . . . 109
DINA output - Item Parameter Estimates . . . . . . . . . . . 113
DINA Summed Score Conversion Tables from -ssc File . . . . 114
DINO Output - Item Parameters . . . . . . . . . . . . . . . . 118
Rupp, Templin, Henson Output - Item Parameters . . . . . . 127
Rupp, Templin, Henson Output - -sco File . . . . . . . . . . . 129
Rupp, Templin, Henson Output - Item Parameters without D
group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
viii
CHAPTER 1
Overview
Vector Psychometric Group, LLC, a North Carolina-based psychometric software and consulting company, is pleased to announce the immediate availR
ability of flexMIRT
Version 2.0, a substantial update to our multilevel and
multiple-group item response theory (IRT) software package for item analysis
R
and test scoring. flexMIRT
fits a variety of unidimensional and multidimensional IRT models to single-level and multilevel data using maximum marginal
likelihood or modal Bayes via Bock-Aitkin EM (with generalized dimension
reduction) or MH-RM estimation algorithms. The generalized dimension reduction EM algorithm, coupled with arbitrary user-defined parameter conR
straints, makes flexMIRT
one of the most flexible IRT software programs
on the market today and the MH-RM algorithm allows users to efficiently
R
estimate high-dimensional models. flexMIRT
produces IRT scale scores using maximum likelihood (ML), maximum a posteriori (MAP), and expected
a posteriori (EAP) estimation. It (optionally) produces summed-score to IRT
scale score (EAP) conversion tables for unidimensional and multidimensional
R
IRT models. As for the item types, flexMIRT
can estimate any arbitrary
mixtures of 3-parameter logistic (3PL) model, logistic graded response model
(which includes 2PL and 1PL as special cases), and the nominal categories
model (including any of its restricted sub-models such as generalized partial
credit model, partial credit model, and rating scale model) for both single-level
and multilevel data, in any number of groups.
R
also has some of the richest psychometric and statistical feaflexMIRT
tures. New to version 2.0, we have implemented features that allows users
to fit Cognitive Diagnostic Models (CDMs). In addition to fully implementing many recently developed multilevel and multidimensional IRT models,
R
flexMIRT
supports numerous methods for estimating item parameter standard errors: Supplemented EM, empirical cross-product approximation, Fisher
(expected) information matrix, Richardson extrapolation, forward difference,
1
and sandwich covariance matrix. A multitude of model fit statistics for dimensionality analysis, item-fit testing, and latent variable normality diagnosis are
R
included in flexMIRT
. Its multiple-group estimation features easily facilitate
studies involving differential item function (DIF) and test linking (including
vertical scaling).
R
is its ability to relax the ubiquitous mulA key innovation in flexMIRT
tivariate normality assumption made in virtually all multidimensional IRT
models. With an extended dimension reduction algorithm, it supports the
non-parametric estimation of latent density shapes using empirical histograms
for both unidimensional and hierarchical (e.g., bifactor and testlet response
theory) item factor models in any number of groups, with support for constraints on group means and variances. Finally, it has a full-featured built-in
Monte Carlo simulation module that can generate data from all models impleR
mented in flexMIRT
.
R
flexMIRT is easy to use. It has an intuitive syntax. It can import
data natively in space-, comma-, or tab- delimited formats. Windows-based
R
flexMIRT
, with a friendly graphical user interface (GUI), is available in both
R
32-bit and 64-bit flavors. flexMIRT
is designed with cross-platform compatibility and large-scale production use from the beginning. The computational
R
engine of flexMIRT
is written using standard C++, which ensures that it
can run on any platform where a C++ compiler exists. A newly-designed
R
memory allocation scheme helps flexMIRT
efficiently handle thousands of
items and millions of respondents, with no imposed upper limit on the size of
the problem.
R
flexMIRT
is fast. For multidimensional and multilevel models that perR
mit dimension reduction, flexMIRT
automatically reduces the amount of
quadrature computation to achieve a dramatic reduction in computational
R
time. As modern CPU architecture trends toward multi-core design, flexMIRT
implements two ways to use multi-threading to further speed up computations
by spreading the load automatically to the multiple available cores or processing units.
This user’s manual is intended both as a guide to those trying to navigate
R
the complex features implemented in flexMIRT
for the first time, and as a
reference source. It is assumed that the user is familiar with the theory and
applications of IRT. Please visit the website http://www.flexMIRT.com for
product updates, technical support, and feedback.
2
CHAPTER 2
Basic Calibration and Scoring
R
syntax. We provide
We begin by discussing the basic structure of flexMIRT
examples that demonstrate some standard IRT analyses, such as calibration
R
and scoring. Syntax files for flexMIRT
may be created in the GUI by selecting “New” or created using a flat text program (e.g., Notepad). If “New" is
R
selected within the flexMIRT
GUI, a command file containing the skeleton
code for a basic analysis opens. Once you have modified this command file, it
R
must be saved prior to being able to submit the analysis to flexMIRT
; the
"Run" button is disabled until the file has been saved to prevent the skeleton
R
code from being over-written. The file name extension for flexMIRT
syntax
files is *.flexmirt. Data files must use space, tab, or comma delimiters between
variables and may not contain a variable name header. Data should be arranged so columns represent items and each observation has its own row in
R
the dataset. Although flexMIRT
has some data manipulation capabilities
(recoding, scoring using a provided answer key, etc), it is expected that the
user has completed necessary pre-analyses (such as dimensionality tests, ensuring that all response options have observations, etc.) prior to submitting
data to the program. Missing data are indicated by -9 by default, although
the missing data code may be modified to any numeric value between -127 and
127. All syntax and data files used in this manual are available in the Support
R
section on the flexMIRT
website (http://www.flexMIRT.com/Support).
2.1. Overview
R
Generally speaking, flexMIRT
syntax files are divided into four required sections, each of which provides instructions for distinct parts of the analysis.
The four sections, which will be discussed in more detail, are
1. <Project> – where you provide a title and a general description of the
analysis for record-keeping.
3
2. <Options> – where the type of analysis and options of the analysis are
specified.
3. <Groups> – where the input data and the IRT models are specified.
4. <Constraints> – where parameter constraints (such as equality or fixed
parameter values) or univariate priors may be specified.
All four section headers must be present in every command file (e.g., even
if no constraints are applied to the parameters, the <Constraints> section
header must still be declared in the syntax). The other requirement is that all
R
statements in the flexMIRT
syntax file must end with a semi-colon. Finally,
the commands in the syntax are not case-sensitive (but file names on certain
operating systems may be).
Within the <Project> section, there are two required commands, Title
= “ ”; and Description = “ ”;. The name of the title and the contents of
the description must be enclosed in double quotation marks.
The <Options> section is where the more technical aspects of the analysis
are specified. It is in this section that analysis details such as convergence
criteria, scoring methods, standard error estimation methods, or requests for
various fit indices may be specified. The only required command is Mode = ;,
where the user is asked to specify what type of analysis will be conducted. The
three options available are Calibration, Scoring, and Simulation. Depending on the Mode of the analysis, there may be additional required commands,
such as the type of IRT scoring to be conducted in a Scoring analysis. These
details will be covered as they emerge in the examples.
As noted previously, <Groups> is the section where information regarding
the groups of item response data is provided. This includes the name(s) of the
file(s) containing the data set(s), the number of items, the number of groups
and participants, and the number of response options for each item, among
other things. Following the <Groups> section header, a group name must be
provided for the first group (even if there is only one group). The group name
is enclosed in percent signs (%) and should be something easy to remember
(e.g., %Group1% or %Grade4% or %Female% or %US%). The group names become
important in specifying constraints in multiple-group analyses. The name is
arbitrary, but spaces are not allowed when specifying group names. After a
group name has been given, the data file name where the item responses are
located is specified, as well as the number of participants in the group, variable
names, the number of possible response categories for each item, and the IRT
model to be estimated (e.g., threePL, Graded(5), etc).
4
The final section, <Constraints>, is reserved for specifying constraints
and prior distributions on the item parameters and latent variable means and
(co)variances. Although the section header is required, there are no mandatory
commands that must be included. Constraints and prior distributions may be
placed on almost any aspect of the model, including means, covariances, and
the various item parameters. Details on how to specify constraints will be
discussed in examples that utilize constraints.
2.2. Single-Group 2PL Model Calibration
The first several examples will use data from two test forms that resemble
those used in North Carolina end-of-grade testing. Data are responses to 12
multiple choice items collected from 2844 test-takers, with responses coded
as correct (1) or incorrect (0). These dichotomous responses may be fitted
with either the 2PL model, if guessing is assumed to be negligible, or with the
3PL model, which explicitly models guessing with an additional parameter.
Example syntax for fitting the 2PL model is presented first.
Example 2.1: 2PL Calibration Syntax
<Project>
Title = "2PL example";
Description = "12 items, 1 Factor, 1 Group 2PL Calibration";
<Options >
Mode = Calibration;
<Groups>
%Group1%
File = "g341-19.dat";
Varnames = v1,v2,v3,v4,v5,v6,v7,v8,v9,v10,v11,v12;
N = 2844;
Ncats(v1-v12) = 2;
Model(v1-v12) = Graded(2);
<Constraints>
Following the previous discussion, section headers (denoted by their enclosure
within < and >) precede the commands. The <Project> header is followed
by a title and description, both of which will be printed in the output. The
<Options> section, by the Mode command, declares this to be a calibration run.
After the <Groups> section header, a name for the first (and only) group is
provided, which is enclosed within %s. Then the name of the data file (File =
5
“ ”;), variable names (Varnames = ;), number of examinees (N = ;), number
of response categories in the data for each item (Ncats( ) = ;), and the IRT
model to be fit to each item (Model( ) = ;) are provided. In this example,
only the file name was provided in the File statement. When the command
file (*.flexmirt) and the data file are in the same folder, the name of the data
R
to locate it. If, however, the command file and
file is sufficient for flexMIRT
the data are in different directories, it is necessary to specify the full path in
the File statement. For example, an alternate statement might be File =
"C:\Users\VPG\Documents\flexMIRT\example 2-1\g341-19.dat";.
When providing names for the variables in the Varnames statement, the
name for every variable was fully written out as v1, v2, ..., v12. For the
Ncats and Model statements, a shortcut was used, which is only available
when variables names end with consecutive integers. As exemplified in the
syntax, v1, v2, ..., v12 can be shortened to v1-v12. This shorthand may
be employed in any statement where variable names are provided.
For the Model statement , the Graded Response Model, with two categories,
is specified as Graded(2). This model is formally equivalent to the 2PL model
R
and as such flexMIRT
does not implement the 2PL model separately. Although not used in this example, it is useful to be aware that multiple Ncat and
Model statements can be specified within one group, which is needed for an
analysis of mixed item types. For example, if items 1 through 6 in the current
data were dichotmous and items 7 through 12 were 5-category polytomous
items, two Ncat statements (i.e., Ncats(v1-v6) = 2; Ncats(v7-v12)= 5;)
and two Model statements (i.e., Model(v1-v6) = Graded(2); Model(v7-v12)
= Graded(5);) would be provided to fit appropriate models to the items.
There are no constraints placed on this model, so after the section header
is declared, the command file syntax ends. At the completion of estimation,
R
flexMIRT
automatically opens the results in the Output viewer. The output is also saved into a text file named with the command file name and
“*-irt.txt” as the extension. The command file used in the first example
is “2PLM_example.flexmirt” and the output is saved to “2PLM_exampleirt.txt”. The output is presented in a straight-forward fashion, so those familiar
with IRT and its parameters and fit values should find it easily understandable.
For completeness, however, we will review the output and highlight relevant
sections. Due to the length of the output file, it will be separated into parts
which are discussed in turn.
6
Output 2.1: 2PL Calibration Output - Summary and Controls
flexMIRT(R) 2.00 (64-bit)
Flexible Multilevel Multidimensional Item Analysis and Test Scoring
(c) 2012-2013 Vector Psychometric Group, LLC., Chapel Hill, NC, USA
2PLM example
12 items 1 Factor, 1 Group 2PLM Calibration
Summary of the Data and Dimensions
Missing data code
-9
Number of Items
12
Number of Cases
2844
# Latent Dimensions
1
Item
1
2
3
4
5
6
7
8
9
10
11
12
Categories
2
2
2
2
2
2
2
2
2
2
2
2
Model
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Graded
Bock-Aitkin EM Algorithm Control Values
Maximum number of cycles:
500
Convergence criterion:
1.00e-004
Maximum number of M-step iterations:
100
Convergence criterion for iterative M-steps:
1.00e-007
Number of rectangular quadrature points:
49
Minimum, Maximum quadrature points:
-6.00,
6.00
Standard error computation algorithm: Cross-product approximation
Miscellaneous Control Values
Z tolerance, max. abs. logit value:
7
50.00
Number of free parameters:
24
Number of cycles completed:
27
Number of processor cores used:
1
Maximum parameter change (P#): 0.000083709 (
Processing times (in seconds)
E-step computations:
0.23
M-step computations:
0.02
Standard error computations:
Goodness-of-fit statistics:
Total:
0.47
13)
0.22
0.00
Output Files
Text results and control parameters: 2PLM_example-irt.txt
Convergence and Numerical Stability
flexMIRT(R) engine status: Normal termination
First-order test: Convergence criteria satisfied
Condition number of information matrix: 46.0221
Second-order test: Solution is a possible local maximum
R
Following the flexMIRT
version and copyright information, the output
begins by printing the title and description provided in the <Project> section
of the syntax file. A broad summary of the data and model is provided on
the next 4 lines, listing the missing value code, number of items, sample size,
and total number of dimensions. In the next section, the number of categories
and the IRT models are listed for each item. The various control values are
listed (e.g., convergence criteria, the maximum number of iterations, number
of quadrature points used, number of free parameters in the model) – if any
control values were changed from defaults, the information printed here serves
as verification.
Next, the program processing time is listed, both in total and broken down
into various stages. Any output files that were generated by the analysis are
named in the next section. The final part of this subsection of the output
R
reports if flexMIRT
terminated normally and if the specified convergence
criteria were met. The next section of the output provides the estimated item
and group parameters.
8
Output 2.2: 2PL Calibration Output - Item and Group Parameters
2PLM example
12 items, 1 Factor, 1 Group 2PL Calibration
2PL Items for Group 1: Group1
Item Label P#
a
s.e.
1
v1
2
1.05
0.07
2
v2
4
1.23
0.11
3
v3
6
0.84
0.06
4
v4
8
1.04
0.07
5
v5 10
0.85
0.06
6
v6 12
1.34
0.10
7
v7 14
1.90
0.17
8
v8 16
0.97
0.08
9
v9 18
1.89
0.15
10
v10 20
1.32
0.10
11
v11 22
0.87
0.06
12
v12 24
1.01
0.07
P#
1
3
5
7
9
11
13
15
17
19
21
23
c
0.30
2.88
-0.05
0.94
0.33
2.09
4.29
1.51
3.35
2.40
-0.26
0.15
s.e.
0.05
0.11
0.04
0.05
0.04
0.09
0.21
0.06
0.16
0.09
0.04
0.05
b
-0.29
-2.35
0.06
-0.90
-0.39
-1.56
-2.26
-1.56
-1.78
-1.82
0.30
-0.15
s.e.
0.05
0.15
0.05
0.06
0.06
0.09
0.12
0.10
0.08
0.10
0.05
0.05
2PLM example
12 items, 1 Factor, 1 Group 2PL Calibration
Group Parameter Estimates:
Group Label P#
mu
s.e.
1 Group1
0.00
----
P#
s2
1.00
s.e.
----
sd
1.00
s.e.
----
Following a reprinting of the title and description, each item is listed in its
own row, with the provided label and assigned parameter numbers, estimated
item parameters, and standard error values going across. The IRT model fit
in this example was the 2PLM, so the reported item parameters are the slope
(labeled a) and intercept (labeled c). In addition, for unidimensional models
the difficulty value (labeled b) is also reported; this value may be found from
the other two parameters as b = c/ − a. Below the item parameter table, the
group parameters are printed. In this example, we had only one group, so the
mean and variance were fixed at 0 and 1, respectively, for model identification.
That these parameters were not estimated is indicated by both the lack of an
assigned parameter number and the dashes in the standard error columns.
9
10
-2.0
0.13
0.36
0.09
0.20
0.12
0.41
0.85
0.22
0.85
0.43
0.08
0.12
-1.6
0.18
0.31
0.11
0.24
0.14
0.45
0.62
0.24
0.87
0.43
0.10
0.16
-1.2
0.22
0.24
0.13
0.26
0.16
0.42
0.38
0.23
0.67
0.37
0.13
0.20
-0.8
0.26
0.17
0.15
0.27
0.17
0.35
0.20
0.21
0.42
0.29
0.15
0.23
-0.4
0.28
0.12
0.17
0.25
0.18
0.26
0.10
0.17
0.23
0.20
0.17
0.25
-0.0
0.27
0.08
0.18
0.22
0.17
0.18
0.05
0.14
0.12
0.13
0.19
0.25
0.4
0.24
0.05
0.17
0.18
0.16
0.11
0.02
0.11
0.06
0.08
0.19
0.24
0.8
0.20
0.03
0.16
0.13
0.14
0.07
0.01
0.08
0.03
0.05
0.18
0.20
1.2
0.16
0.02
0.14
0.10
0.12
0.04
0.01
0.06
0.01
0.03
0.16
0.17
1.6
0.12
0.01
0.12
0.07
0.09
0.03
0.00
0.04
0.01
0.02
0.14
0.13
2.0
0.08
0.01
0.10
0.05
0.07
0.02
0.00
0.03
0.00
0.01
0.11
0.09
2.4
0.06
0.00
0.08
0.03
0.06
0.01
0.00
0.02
0.00
0.01
0.09
0.07
at 15 Values of theta from -2.80 to 2.80 for Group 1: Group1
2.8
0.04
0.00
0.06
0.02
0.04
0.01
0.00
0.01
0.00
0.00
0.07
0.05
Test Information: 3.56 4.38 4.86 4.83 4.41 3.87 3.38 2.97 2.61 2.29 2.01 1.77 1.58 1.42 1.30
Expected s.e.: 0.53 0.48 0.45 0.45 0.48 0.51 0.54 0.58 0.62 0.66 0.71 0.75 0.80 0.84 0.88
Marginal Reliability for Response Pattern Scores: 0.64
Item Information Function Values
Theta:
Item
Label -2.8 -2.4
1
v1 0.07 0.10
2
v2 0.35 0.38
3
v3 0.05 0.07
4
v4 0.12 0.16
5
v5 0.07 0.09
6
v6 0.24 0.33
7
v7 0.70 0.89
8
v8 0.17 0.20
9
v9 0.39 0.64
10
v10 0.29 0.38
11
v11 0.04 0.06
12
v12 0.06 0.09
2PLM example
12 items, 1 Factor, 1 Group 2PL Calibration
Output 2.3: 2PL Calibration Output - Information Functions
The item and test information functions, values for which are presented in
Output 2.3, are the focus of the next section of the output file. The title and
description are printed once again, followed by the label for the information
table. As may be seen, the values of theta used in the table range from -2.8
to 2.8, increasing in steps of 0.4. The information function values for each
item are presented on the appropriate row and the Test Information Function
(TIF) values, which sum the item information values at each theta value (plus
the contribution from the prior), are printed towards the bottom of the table.
Additionally, the expected standard error associated with each theta value
is printed. Finally, the marginal reliability, which is defined as 1 minus the
averaged error variance divided by the prior variance, is reported underneath
the table. After the information and reliability section, the overall goodness
of fit (GOF) indices are presented, which appear in the next output box.
Output 2.4: 2PLM Calibration Output - Goodness of Fit Indices
Statistics based on the loglikelihood of the fitted model
-2loglikelihood:
33408.05
Akaike Information Criterion (AIC):
33456.05
Bayesian Information Criterion (BIC):
33598.92
Full-information fit statistics of the fitted model
The table is too sparse to compute the general
multinomial goodness of fit statistics.
Limited-information fit statistics of the fitted model
The M2 statistics were not requested.
When no option is specified, the GOF reporting will provide only basic
fit indices. These include the values for the −2×Log Likelihood, the Akaike
Information Criterion (AIC), Bayesian Information Criterion (BIC), and when
appropriate, the Pearson X 2 statistic (labeled X2), the likelihood ratio statistic (labeled G2), their associated degrees of freedom, the estimated population discrepancy function value (labeled F0hat), and the RMSEA (Root Mean
Square Error of Approximation; e.g., Browne & Cudeck, 1993). In this examR
ple, flexMIRT
has reported that the underlying contingency table was too
sparse to compute the full-information fit values because the total number of
cross-classifications is 212 = 4096, which is considerably larger than the sample
11
size of 2844.
2.3. Single-Group 3PL Model Calibration
The data in the previous section may also be suitable for the 3PL model, which
explicitly models examinees’ guessing.
Example 2.2: Single Group 3PL Calibration Syntax
<Project>
Title = "3PL example";
Description = "12 items, 1 Factor, 1 Group Calibration,
saving parameter estimates";
<Options>
Mode = Calibration;
GOF = Extended;
savePRM = Yes;
<Groups>
%Group1%
File = "g341-19.dat";
Varnames = v1-v12;
N = 2844;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
<Constraints>
Prior (v1-v12), Guessing : Beta(1.0,4.0);
The syntax for this run differs from the previous one in several ways. In
the <Options> section, the parameter estimates are to be saved to a separate output file (savePRM = Yes;) and additional GOF statistics were requested by specifying GOF = Extended; as an option. In the <Groups> section the fitted model is now the 3PL model (Model keyword is ThreePL). In
the <Constraints> section, a prior distribution is imposed on the Guessing
parameters for all items. In this example the Beta(1.0,4.0) distribution was
used, which corresponds to a prior sample size of 5, with 1.0/(1.0 + 4.0) = 0.2
as its mode. This prior indicates that the prior guessing probability is equal to
0.2, as these multiple-choice items have 5 options. The Normal and LogNormal
distributions are also available when specifying priors. In relation to the lower
asymptote, the Normal prior applies a normal distribution prior on the logit of
the lower asymptote parameter. A prior of Normal(-1.39, 0.5) is a reason12
able prior for the lower asymptote, with a mode around 0.20 in the typical g
metric. Other common normal priors for the logit-g parameter would be N(1.09, 0.5) for items with 4 possible response options, N(-0.69, 0.5) for items
with 3 response options, and N(0.0, 0.5) for True/False type items. The standard deviation of 0.5 is standard across all noted priors and is based primarily
on experience, in that it provides a prior that is diffuse enough to allow estimated guessing parameters to move away from the mean value if necessary
but not so diffuse as to be uninformative.
The output for this analysis is quite similar to that of the previous run,
so only differences will be highlighted. As noted in the syntax discussion,
R
flexMIRT
has been instructed to save the item and group parameter values
into a separate file. This results in an additional file name, with the “*prm.txt" extension, being listed in the output files. The layout of the item
parameter output table (presented in Output 2.5) is essentially the same as
before, with the slopes, thresholds, and difficulty values and their standard
errors reported first. In the last four columns of the table, the estimates for
the lower asymptote (both in its logit form and in terms of the pseudo-guessing
probability) are reported, as well as standard errors.
13
14
3PL Items for Group 1: Group1
Item Label P#
a
s.e.
1
v1
3
1.84
0.30
2
v2
6
1.27
0.12
3
v3
9
1.50
0.26
4
v4 12
1.31
0.19
5
v5 15
1.35
0.22
6
v6 18
1.39
0.15
7
v7 21
1.96
0.18
8
v8 24
1.05
0.13
9
v9 27
1.94
0.18
10
v10 30
1.35
0.13
11
v11 33
1.77
0.32
12
v12 36
1.55
0.24
P#
2
5
8
11
14
17
20
23
26
29
32
35
c
-0.56
2.77
-0.98
0.46
-0.42
1.91
4.27
1.24
3.31
2.28
-1.40
-0.52
s.e.
0.26
0.20
0.29
0.22
0.25
0.18
0.25
0.23
0.19
0.17
0.34
0.23
b
0.31
-2.17
0.65
-0.35
0.31
-1.37
-2.18
-1.19
-1.71
-1.69
0.79
0.33
s.e.
0.10
0.26
0.11
0.21
0.15
0.23
0.17
0.34
0.15
0.22
0.08
0.11
P# logit-g
1
-1.09
4
-2.00
7
-1.19
10
-1.19
13
-1.08
16
-1.93
19
-2.22
22
-1.56
25
-2.59
28
-2.23
31
-1.25
34
-1.37
...
Output Files
Text results and control parameters: g341-19-irt.txt
Text parameter estimate file: g341-19-prm.txt
...
3PL example
12 items, 1 Factor, 1 Group Calibration, saving parameter estimates
Output 2.5: Single Group 3PL Calibration Output- Item Parameters
s.e.
0.22
1.33
0.23
0.45
0.28
1.09
1.46
0.91
1.59
1.36
0.19
0.28
g
0.25
0.12
0.23
0.23
0.25
0.13
0.10
0.17
0.07
0.10
0.22
0.20
s.e.
0.04
0.14
0.04
0.08
0.05
0.12
0.13
0.13
0.10
0.12
0.03
0.05
In addition to the added parameters and output files, the GOF statement was
set to Extended, so additional fit statistics, namely marginal X 2 values and
the local dependence (LD) statistics of Chen and Thissen (1997) are included
in the output. A table containing these values is printed between the group
parameter values and the Information Function values table. This table is
presented below. The values in the second column of this table are the itemspecific marginal X 2 , which may be tested against a chi-square variable with
degrees of freedom equal to the number of categories for that item minus 1. In
this case, all items fit well at the level of univariate margins. The remaining
cells provide pairwise diagnostic information for possible local dependence.
These are standardized X 2 values so values larger than 3 in absolute magnitude
should be examined further.
Output 2.6: Single Group 3PL Calibration Output- Extended GOF Table
Marginal fit
Marginal
Item
Chi2
1
0.0
2
0.0
3
0.0
4
0.0
5
0.0
6
0.0
7
0.0
8
0.0
9
0.0
10
0.0
11
0.0
12
0.0
Marginal
Item
Chi2
11
0.0
12
0.0
(Chi-square) and Standardized LD X2 Statistics for Group 1: Group1
1
2
3
4
5
6
7
8
9
10
-0.7
-0.7
0.9
-0.6
-0.5
0.8
-0.5
-0.6
2.5
-0.6
-0.7
2.0
-0.6
-0.0
-0.7
-0.4
-0.6
0.7
5.0
3.5
-0.6
0.2
-0.1
-0.7
-0.4
-0.7
-0.4
-0.7
-0.1
3.7
5.3
-0.5
-0.6
0.6
-0.4
0.1
-0.3
0.3
-0.3
0.3
-0.4
-0.5
-0.5
-0.4
-0.0
-0.7
0.7
-0.7
-0.0
0.8
-0.3
0.0
4.4
-0.0
0.0
-0.7
0.2
-0.7
-0.7
-0.6
1.4
-0.7
0.4
-0.3
-0.1
11
1.2
2.4. Single-Group 3PL EAP Scoring
There are situations when existing item parameters will be used and only
scoring for a new set of data is desired. This next example will introduce a
command file that does just that, using the saved parameter estimates from a
prior calibration run.
15
Example 2.3: 3PL EAP Scoring example
<Project>
Title = "3PL EAP scoring";
Description = "12 items, 1 Factor, 1 Group Scoring";
<Options>
Mode = Scoring;
ReadPRMFile = "3PL_example-prm.txt";
Score = EAP;
saveSCO=YES;
<Groups>
%Group1%
File = "g341-19.dat";
Varnames = v1-v12;
N =2844;
<Constraints>
For scoring runs the <Project> section stays the same. The first change
encountered is in the <Options> section, in that Mode is now set to Scoring.
Using the command ReadPRMFile, the name of the file that contains the item
parameter estimates is provided. In this case, the file was created by the
previous calibration run. The type of IRT scale scores is selected using the
Score = ; command. In this case, Expected A Posteriori (EAP) scores will
be saved. We may also specify ML for maximum likelihood scoring, MAP for
Maximum A Posteriori scoring, or SSC for summed score to EAP conversion
tables. We have also specified that the scores should be saved into a separate file, which will be labeled with the extension “*-sco.txt”. The command
R
requesting flexMIRT
to output the estimated scores (SaveSco = Yes;) is
redundant for most scoring methods, but it is needed if individual scores are
desired for the SSC method. Without the SaveSco = Yes; command, SSC will
produce only the summed score to scale score conversion table. Notice that
this may be a desirable feature in some situations, when only the conversion
table is needed with actual scoring occurring at a much later stage.
The output of scoring runs differs from calibration runs so we have reproduced the entire scoring control output file, which has the extension“*-ssc.txt”,
in Output 2.7 below.
16
Output 2.7: 3PL EAP Scoring Output
3PL EAP scoring
12 items, 1 Factor, 1 Group Scoring
Summary of the Data and Control Parameters
Missing data code
Number of Items
Number of Cases
Number of Dimensions
Item
1
2
3
4
5
6
7
8
9
10
11
12
Categories
2
2
2
2
2
2
2
2
2
2
2
2
-9
12
2844
1
Model
3PL
3PL
3PL
3PL
3PL
3PL
3PL
3PL
3PL
3PL
3PL
3PL
Scoring Control Values
Response pattern EAPs are computed
Miscellaneous Control Values
Output Files
Text results and control parameters: g341-19-ssc.txt
Text scale score file: g341-19-sco.txt
3PL EAP scoring
12 items, 1 Factor, 1 Group Scoring
3PL Items for Group 1: Group1
Item
a
c
b logit-g
17
g
1
2
3
4
5
6
7
8
9
10
11
12
1.84
1.27
1.50
1.31
1.35
1.39
1.96
1.05
1.94
1.35
1.77
1.55
-0.56
2.77
-0.98
0.46
-0.42
1.91
4.27
1.24
3.31
2.28
-1.40
-0.52
0.31
-2.17
0.65
-0.35
0.31
-1.37
-2.18
-1.19
-1.71
-1.69
0.79
0.33
-1.09
-2.00
-1.19
-1.19
-1.08
-1.93
-2.22
-1.56
-2.59
-2.23
-1.25
-1.37
0.25
0.12
0.23
0.23
0.25
0.13
0.10
0.17
0.07
0.10
0.22
0.20
3PL EAP scoring
12 items, 1 Factor, 1 Group Scoring
Group Parameter Estimates:
Group
Label
1
Group1
mu
0.00
s2
1.00
sd
1.00
As with the calibration output, the file starts with a general data summary
R
and echoing of values that were provided to flexMIRT
. In the scoring mode,
this section of the output is much shorter due to the fact that fewer control
parameters are available for modification. Of note in this section is the line
labeled “Scoring Control Values,” under which the type of IRT scale scores
is printed. Below that, the output files generated are named and tables containing the item and group parameters, as read in from the parameter file are
printed, to allow for the user to verify the parameter values were read in as
expected.
The individual scores were saved into a separate file with the extension
“*-sco.txt.” Opening the individual score file (3PLM_score_example-sco.txt)
reveals that the columns in the score file are formatted as follows: the group
number, the observation number, the scale score, and the standard error associated with the scale score. Optionally, following the group number and the
R
observation number, flexMIRT
will print the case ID number if the caseID
variable is given in the command syntax.
18
2.5. Single-Group 3PL ML Scoring
As noted, maximum likelihood scores are also available. One issue that may
occur when using ML scores is that extreme response patterns on either end of
the continuum (e.g., all correct or all incorrect) are associated with undefined
or unreasonable theta estimates. When Score = ML; the user is required to
specify the desired maximum and minimum scores via the MaxMLscore and
MinMLscore keywords, to obtain practical theta estimates for such cases.
Example 2.4: 3PL ML Scoring example
<Project>
Title = "3PL ML Scoring Example";
Description = "12 items 1 Factor, 1 Group Scoring";
<Options>
Mode = scoring;
ReadPRMFile = "3PL_example-prm.txt";
Score= ML;
MaxMLscore = 3.5;
MinMLscore = -5.5;
saveSCO = YES;
<Groups>
%Group1%
File = "g341-19.dat";
Varnames = v1-v12;
N = 2844;
<Constraints>
This syntax is the same as the previous scoring example, save for the change
from Score = EAP to Score = ML and the additional commands MaxMLscore
and MinMLscore used to “reign in" the extreme ML estimates. We present the
first 10 cases of the associated -sco.txt file, where the scores were saved.
19
Output 2.8: 3PL ML Scoring Output
1
1
1
1
1
1
1
1
1
1
1
2
3
4
5
6
7
8
9
10
2
3
2
3
3
3
3
3
2
3
-0.198234
1.427265
0.113716
1.427265
-0.675141
0.637464
1.382694
-1.205723
3.500000
0.859248
0.573388
0.784667
0.680658
0.784667
0.625639
0.606990
0.792826
0.628273
99.990000
0.658306
Like the previous -sco file, the first column indicates group membership and
R
the second column provides the flexMIRT
assigned observation number. The
third column reports the number of iterations to reach the reported score. The
fourth and fifth columns are the ML theta estimate and the associated SE,
respectively. As can be seen for observation 9, our specified maximum of
3.5 was applied. Because the reported value was assigned as a result of our
MaxMLscore statement, the reported SE is 99.99 to indicate that the score was
assigned, rather than estimated.
2.6. Graded Model Calibration and Scoring
Example syntax for a combined calibration and scoring run is now presented.
R
flexMIRT
code for the graded response model, which may be used with two
or more ordered response options (e.g., Samejima, 1969), is also introduced.
This example employs data from 3000 simulees, using parameter values reported in Table 1 of Edwards’ (2009) paper as generating values. These data
were constructed to mimic responses collected on the Need for Cognition Scale
(Cacioppo, Petty, & Kao, 1984), which is an 18 item scale that elicits responses
on a 5-point Likert-type scale, asking participants to indicate the extent to
which the content of the item prompt describes them.
Example 2.5: Graded Model Combined Calibration and Scoring Syntax with
Recoding
<Project>
Title = "NCS example";
Description = "18 items, 1D, 1 Group Calibration and Scoring";
20
<Options>
Mode = Calibration;
SaveSCO = YES;
Score = MAP;
GOF = Extended;
<Groups>
%Group1%
File = "NCSsim.dat";
Varnames = v1-v18;
N = 3000;
Ncats(v1-v18) = 5;
Code(v1-v18) = (1,2,3,4,5),(0,1,2,3,4);
Model(v1-v18) = Graded(5);
<Constraints>
The only new command encountered in this example is the “Code =”
statement in the <Groups> section. The internal representation of item reR
sponse data in flexMIRT
is zero-based (i.e., response options must start at
zero and go up). The raw item response data file for this example (NCSsim.dat) labels the categories as 1 through 5, so recoding is necessary. The
Code statement does this for variables v1-v18 by specifying the original values in the first pair of parentheses and, following a comma, the desired recoded values in a second set of parentheses. Thus, a command such as
Code(v1-v18)=(1,2,3,4,5),(4,3,2,1,0); accomplishes both reverse coding and makes the recoded item response data zero-based.
Another point of note in this command file is that, even though this is
a calibration run, we have requested MAP scores be saved. This combination
of commands is a short-cut to a combined calibration and scoring run, in
which the calibration is explicitly requested and the SaveSco=Yes; command
implicitly requests a scoring run, using the estimated parameter values from
the calibration.
With respect to the output, the primary change from previous examples
are the additional parameter values in the item parameter table. As may
be seen from the two item parameter estimate output tables below there are
multiple thresholds (or equivalently, intercepts) for each item. For convenience, both of these parameterizations (thresholds and intercepts) are presented in separate tables. However, only the intercept values will be saved
to the -prm file, if it is requested. This behavior can be modified by adding
the statement SlopeThreshold= Yes; to the <Options> section. When the
21
SlopeThreshold keyword is invoked threshold values, rather than intercept
values, will be saved to the -prm file.
Output 2.9: Graded Model Calibration - Slopes and Intercepts
Graded Items for Group 1: Group1
Item Label P#
a
s.e. P#
c 1
1
v1 5 1.58 0.06
1 3.25
2
v2 10 1.85 0.06
6 4.35
3
v3 15 1.79 0.06 11 4.04
4
v4 20 1.73 0.06 16 4.11
5
v5 25 1.44 0.05 21 4.35
6
v6 30 1.19 0.05 26 2.10
7
v7 35 1.24 0.05 31 3.27
8
v8 40 0.72 0.04 36 2.18
9
v9 45 1.08 0.05 41 2.46
10
v10 50 1.37 0.05 46 4.72
11
v11 55 1.62 0.06 51 4.34
12
v12 60 1.44 0.05 56 4.41
13
v13 65 1.12 0.05 61 2.19
14
v14 70 1.37 0.05 66 3.71
15
v15 75 1.57 0.06 71 4.30
16
v16 80 0.80 0.04 76 2.13
17
v17 85 1.17 0.05 81 3.28
18
v18 90 0.61 0.04 86 2.83
s.e.
0.09
0.12
0.11
0.12
0.12
0.06
0.09
0.06
0.07
0.15
0.12
0.13
0.06
0.10
0.12
0.06
0.09
0.08
P#
c 2
2 0.95
7 1.79
12 1.61
17 1.65
22 1.71
27 0.29
32 0.77
37 0.02
42 -0.10
47 2.28
52 1.92
57 1.79
62 0.35
67 1.71
72 1.66
77 0.20
82 0.95
87 1.16
s.e.
0.06
0.07
0.06
0.06
0.06
0.05
0.05
0.04
0.05
0.07
0.07
0.06
0.05
0.06
0.06
0.04
0.05
0.05
P#
3
8
13
18
23
28
33
38
43
48
53
58
63
68
73
78
83
88
c 3
-0.41
0.48
0.29
0.35
0.47
-0.88
-0.04
-0.97
-1.06
0.78
0.72
0.75
-1.09
0.42
0.30
-0.59
0.00
0.21
s.e.
0.05
0.06
0.06
0.06
0.05
0.05
0.05
0.04
0.05
0.05
0.06
0.05
0.05
0.05
0.05
0.04
0.05
0.04
P#
4
9
14
19
24
29
34
39
44
49
54
59
64
69
74
79
84
89
c 4
-3.34
-2.91
-2.40
-2.61
-2.35
-3.30
-2.61
-2.95
-3.06
-2.12
-2.62
-2.03
-3.72
-2.01
-2.47
-2.69
-2.07
-2.13
s.e.
0.09
0.09
0.08
0.08
0.07
0.09
0.07
0.08
0.08
0.07
0.08
0.07
0.10
0.07
0.07
0.07
0.06
0.06
Output 2.10: Graded Model Calibration - Slopes and Thresholds
Graded Items
Item Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
7
v7
8
v8
9
v9
10
v10
11
v11
12
v12
13
v13
14
v14
15
v15
16
v16
17
v17
18
v18
for
P#
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
Group
a
1.58
1.85
1.79
1.73
1.44
1.19
1.24
0.72
1.08
1.37
1.62
1.44
1.12
1.37
1.57
0.80
1.17
0.61
1: Group1
s.e.
b 1
0.06 -2.05
0.06 -2.35
0.06 -2.26
0.06 -2.38
0.05 -3.01
0.05 -1.77
0.05 -2.63
0.04 -3.05
0.05 -2.29
0.05 -3.45
0.06 -2.68
0.05 -3.05
0.05 -1.95
0.05
-2.7
0.06 -2.74
0.04 -2.67
0.05
-2.8
0.04 -4.63
s.e.
0.07
0.08
0.07
0.08
0.12
0.07
0.10
0.17
0.10
0.14
0.09
0.11
0.08
0.10
0.10
0.14
0.11
0.31
b 2
-0.60
-0.97
-0.90
-0.95
-1.19
-0.24
-0.62
-0.03
0.09
-1.67
-1.19
-1.24
-0.31
-1.25
-1.06
-0.25
-0.81
-1.89
s.e.
0.04
0.04
0.04
0.04
0.05
0.04
0.04
0.06
0.04
0.06
0.05
0.05
0.04
0.05
0.04
0.05
0.05
0.13
b 3
0.26
-0.26
-0.16
-0.21
-0.33
0.74
0.03
1.36
0.99
-0.57
-0.45
-0.52
0.97
-0.30
-0.19
0.75
0.00
-0.35
s.e.
0.03
0.03
0.03
0.03
0.04
0.05
0.04
0.09
0.05
0.04
0.03
0.04
0.05
0.04
0.03
0.06
0.04
0.07
b 4
2.11
1.58
1.34
1.51
1.63
2.77
2.10
4.12
2.84
1.56
1.62
1.40
3.31
1.47
1.57
3.38
1.77
3.48
s.e.
0.07
0.05
0.05
0.05
0.06
0.11
0.08
0.23
0.12
0.06
0.06
0.05
0.14
0.06
0.06
0.18
0.07
0.23
As before, the marginal X 2 values and the local dependence (LD) statistic
values of Chen and Thissen (1997) are reported due to the GOF = Extended;
22
statement, but an additional feature is present in the table because of the ordinal nature of the data. The Marginal fit (Chi-square) and Standardized
LD X2 Statistics table is presented in the following Output example. The
standardized LD X 2 value for each item pair is followed by either the letter
“p” or the letter “n”. When the item responses are composed of ordered categories, phi correlations are computed for both the model-implied and observed
bivariate tables. If the model implied correlation is lower than the observed
correlation for a given item pair, a “p” is printed after the calculated X 2 ,
meaning “positive” LD. If the model implied correlation is higher, an “n” is
printed, indicating “negative” LD. Either a “p” or an “n” is printed after every value in the table. The absolute magnitude of the individual X 2 values
should still evaluated against a critical value, such as 3.0, to determine if a
non-ignorable level of LD is present.
Output 2.11: Graded Model Calibration - Extended GOF Table
Marginal fit (Chi-square) and Standardized LD X2 Statistics for Group 1: Group1
Marginal
Item
Chi2
1
0.0
2
0.1
3
0.0
4
0.1
5
0.1
6
0.1
7
0.0
8
0.0
9
0.0
10
0.0
11
0.1
12
0.0
13
0.1
14
0.1
15
0.0
16
0.0
17
0.0
18
0.0
Marginal
Item
Chi2
11
0.1
12
0.0
13
0.1
14
0.1
15
0.0
16
0.0
17
0.0
18
0.0
1
2
3
4
5
6
7
-0.7n
-0.8p
-0.6n
1.0n
-0.2n
-0.4p
-1.2n
0.6n
-0.4p
-0.0p
-0.3p
-1.1p
-0.4p
-1.3n
-0.5p
-0.8n
1.8p
-0.3n
-0.7p
-1.4p
-1.5n
-1.4p
-0.1n
0.6p
-1.3p
-0.9n
0.9n
-0.9n
-1.1n
0.0p
1.2n
-0.8p
0.9p
0.7n
-0.6p
-0.9p
-0.8p
-1.1n
-0.5p
1.0n
1.3p
-1.1p
0.5p
-0.7n
-1.0p
-0.6p
0.6p
-0.6n
-0.8n
-0.3n
0.0n
-0.1n
0.3p
-0.7n
-1.2p
0.6p
0.4n
0.8n
-0.6n
-1.3n
-1.4p
-0.9p
0.7p
-0.2n
-0.3p
1.4n
-0.2p
-0.2p
0.8p
-0.4n
-0.7n
0.4p
-0.4p
-0.8p
2.7n
0.1p
-1.5p
-0.9p
-0.5p
0.7n
-0.1n
-0.7n
0.8p
-1.2p
1.4p
0.7n
1.5p
-1.7p
-0.8p
-1.7n
-1.1p
-0.7n
1.4p
0.7p
-0.9n
0.1p
-0.7n
0.5n
11
12
13
14
15
16
17
0.5p
-0.6p
-1.1p
-1.3n
2.6n
1.8n
-0.5p
-1.0p
-0.9n
1.0n
-1.9p
-1.7p
-1.0p
-1.8p
0.5p
-1.2n
0.1n
-0.7n
-0.4n
-0.5p
-0.9n
0.6n
-1.5n
-0.3p
-0.3n
0.7p
-0.1n
0.7p
8
-1.5p
-1.2p
0.1n
-0.5n
-1.5p
0.3n
-0.4n
-0.2p
-0.1p
-0.4n
9
-0.7n
-0.4p
-0.2n
-0.3p
0.7n
-0.5p
-0.1n
-0.9n
-0.3p
10
0.4p
-1.9p
-1.1n
0.5p
-0.6p
0.4p
-0.6p
-0.2p
2.7. 1PL Model Fitted to Response Pattern Data
Up to this point, analyses have been conducted on raw (individual) data,
in which each line in the data set represents responses from an individual.
R
flexMIRT
is also able to handle data that are grouped by response patterns.
23
Example 2.6: 1PL Calibration and Scoring with Response Pattern Data
<Project>
Title = "LSAT 6";
Description = "5 Items 1PL N=1000 Grouped Data";
<Options>
Mode = Calibration;
GOF = Complete;
JSI = Yes;
HabermanResTbl = Yes;
SX2Tbl = Yes;
FactorLoadings = Yes;
SaveSco = No;
Score = SSC;
<Groups>
%Group1%
File = "lsat6grp.dat";
Varnames = v1-v5,w;
Select = v1-v5; // only the first 5 variables
CaseWeight = w;
N = 30;
Ncats(v1-v5) = 2;
Model(v1-v5) = Graded(2);
<Constraints>
Equal (v1-v5), Slope;
/* Entire
Block of
Comments */
The syntax file above uses the well-known LSAT6 data set, in which the
individuals have been grouped by response patterns. There are 1000 examinees
and 5 dichotomous items. The first use of constraints in the <Constraints>
R
section is also presented, demonstrating how to fit a 1PL model in flexMIRT
.
An examination of the data file (lsat6grp.dat) shows that that the data
for these five items have been grouped by response patterns with the last
variable in the file providing the number of observations per pattern. After
naming the group and file name in the <Groups> section as usual, we use the
Select statement to select only the first 5 variables that will be subject to
24
IRT analysis (i.e., we exclude the weighting variable, which was named “w”
in the Varnames statement). Note that this is also the first time that we have
encountered comments in the syntax file. A comment begins with 2 slashes,
following the standard C/C++ convention. They may appear anywhere in the
syntax, but not within a command statement. In addition, one may also use
a pair of /* */ to create entire blocks of comments that span more than one
lines, as shown at the end of the syntax file.
Returning to the example syntax, we use the CaseWeight command to
identify the variable that contains response pattern frequencies, (i.e., the number of cases per response pattern). The rest of the <Groups> statements, those
defining number of categories, etc., are as before.
Note that the graded model with 2 categories is specified in the Model
statement, which is equivalent to a 2PL model. To obtain a 1PL model, the
slope/discrimination parameters must be constrained equal across items. To
accomplish this, an equality constraint is specified in the <Constraints> section using the Equal statement on the Slope parameters of items v1-v5. The
keywords for other items parameters are Intercept for the location parameter in all models, Guessing for the guessing parameter/lower asymptote in
the 3PL model, and ScoringFn for the scoring function parameters in the
Nominal Categories model.
There are several new commands in the <Options> section. For the goodness -of-fit indices, the full gamut of available GOF indices (detailed in Chapter
7 of the manual) are requested by the keyword Complete. The JSI statement
requests the Jackknife Slope Index (JSI), an LD detection tool, for all pairs
of items. The HabermanResTbl command requests a table that lists the observed and expected frequencies and standardized residuals for each response
pattern, as well as the associated SEs, EAPs and SDs. When SX2Tbl is requested, detailed item fit tables will be printed, in addition to the Orlando
and Thissen (2000) S − X 2 item fit chi-square statistics.
The command
FactorLoadings = Yes; requests the printing of the factor loading values
for each item, which may be converted from the respective slope values (e.g.,
Takane & de Leeuw, 1987; Wirth & Edwards, 2007). Even though the scores
are not saved (SaveSco = No;), the scoring method of SSC is specified. As
noted previously, the SSC method will print the requested summed score to
EAP scale score conversion table in the output, although no individual scores
are calculated/saved. The commands introduced in this example add several
25
new parts to the output file.
Output 2.12: 1PL Grouped Data Output - Item Parameters and Factor
Loadings
LSAT 6
5 Items 1PL N=1000 Grouped Data
2PL Items for Group 1: Group1
Item Label
P#
a
s.e.
1
v1
6
0.76
0.00
2
v2
6
0.76
0.00
3
v3
6
0.76
0.00
4
v4
6
0.76
0.00
5
v5
6
0.76
0.00
P#
1
2
3
4
5
c
2.73
1.00
0.24
1.31
2.10
s.e.
0.01
0.00
0.00
0.00
0.00
b
-3.61
-1.32
-0.32
-1.73
-2.78
s.e.
0.01
0.00
0.00
0.00
0.01
LSAT 6
5 Items 1PL N=1000 Grouped Data
Factor Loadings for Group 1: Group1
Item
Label lambda 1
1
v1
0.41
2
v2
0.41
3
v3
0.41
4
v4
0.41
5
v5
0.41
s.e.
0.00
0.00
0.00
0.00
0.00
The item parameters are printed in their customary order, but note that
the 1PL equal-slope constraint specified in the syntax has been interpreted
correctly, as evidenced by all values in the a column being equal. The factor
loadings are presented below the typical item parameter table, with loadings
appearing in the lambda1 column.
The next new portion of the output are the values and tables associated
with the summed-score based item fit diagnostic (e.g., Orlando & Thissen,
2000), which are printed as a result of the GOF = Complete and SX2Tbl =
Yes; commands, respectively. The S − X 2 value and table for the first item is
presented. When the p-value associated with the S − X 2 value is greater than
the α-level (typically 0.05), the item should be examined for possible misfit.
26
Output 2.13: 1PL Grouped Data Output - Item Fit Diagnostics
Orlando-Thissen-Bjorner Summed-Score Based Item Diagnostic Tables and X2s:
Group
1: Group1
Item 1 S-X2(1) = 48.5, p < 0.0001
Category 0
Category 1
Score Observed
Expected
Observed
Expected
0
1
0.4
1
1.6
1
4
5 1.1
1.5
4
6.9
2
4
1.0
6
9.0
3
4
8 0.6
1.5
4
7.4
4
1
0.1
1
1.9
By specifying Score = SSC; in the command file, we requested a summed
score to EAP scale score conversion table be printed, which is shown below.
With a total of 5 items, there are 6 possible summed scores which are listed in
the first column. Going across a row, after the sum score, the EAP score and
the posterior standard deviation associated with that particular summed score
value are presented. Additionally, the table also includes model-implied and
observed probabilities for the summed scores, appearing in columns labeled
P and O, respectively. Immediately after the table, the marginal reliability of
the scale scores is printed, as well as the value of a chi-square statistic that
provides a test of the underlying normality of the latent variable distribution.
Output 2.14: 1PL Grouped Data Output - Summed Score to Scale Score
Conversion
Summed Score to Scale Score Conversion Table:
Summed
Score
EAP
SD
P
O
0.00 -1.910
0.797 0.0023654 0.0030000
1.00 -1.429
0.800 0.0208898 0.0200000
2.00 -0.941
0.809 0.0885974 0.0850000
3.00 -0.439
0.823 0.2274460 0.2370000
4.00
0.084
0.841 0.3648997 0.3570000
5.00
0.632
0.864 0.2958018 0.2980000
Summed score based latent distribution fit S-D2 = 0.9, p = 0.8166
Marginal reliability of the scaled scores for summed scores = 0.29427
Scrolling further down in the full output file, we encounter sections related
to LD detection. We have previously discussed the Marginal fit and Standardized LD X 2 Statistics table, but the matrix immediately below that in the
27
output is new. The JSI = Yes; statement in the command file produces a
matrix (labeled Edwards-Houts-Cai Jackknife Slope Diagnostic Index) which
provides values for a newly introduced LD detection method (Edwards & Cai,
2011), the calculation of which is described in detail later in the manual.
Output 2.15: 1PL Grouped Data Output - JSI matrix
Edwards-Houts-Cai Jackknife Slope Diagnostic Index for Group 1: Group1
Item
1
2
3
4
5
1
--0.3
0.3
0.3
0.3
2
-0.7
---0.7
-0.7
-0.7
3
2.5
2.5
--2.5
2.5
4
-1.4
-1.4
-1.4
---1.4
5
-1.1
-1.1
-1.1
-1.1
---
For now, it is sufficient to know that a value noticeably larger than others in
the matrix indicates an item pair that should be examined for possible LD.
The last new table is the result of the statement HabermanResTbl=Yes;.
A portion of this table is reproduced below. In this table, observed response
patterns are listed by row. For a given response pattern, the observed and
model-implied frequencies, the standardized residual, EAP score, and standard
deviation associated with the EAP score are printed.
Output 2.16: 1PL Grouped Data Output - Haberman Residual Table (Excerpt)
Response Pattern Observed and Expected Frequencies, Standardized
Residuals, EAPs and SDs
Group 1: Group1
Item:
Frequencies
Standard
1 2 3 4 5 Observed Expected Residual EAP[q |u]
0 0 0 0 0
3.00
2.37
0.41
-1.91
0 0 0 0 1
6.00
5.47
0.23
-1.43
0 0 0 1 0
2.00
2.47
-0.30
-1.43
0 0 0 1 1
11.00
8.25
0.96
-0.94
0 0 1 0 0
1.00
0.85
0.16
-1.43
0 0 1 0 1
1.00
2.84
-1.09
-0.94
...
1 1 0 0 1
56.00
55.17
0.12
-0.44
1 1 0 1 0
21.00
24.96
-0.80
-0.44
1 1 0 1 1
173.00
177.91
-0.41
0.08
28
SD[q |u]
0.80
0.80
0.80
0.81
0.80
0.81
0.82
0.82
0.84
1
1
1
1
1
1
1
1
1
1
1
1
0
0
1
1
0
1
0
1
11.00
61.00
28.00
298.00
8.59
61.23
27.71
295.80
0.83
-0.03
0.06
0.15
-0.44
0.08
0.08
0.63
0.82
0.84
0.84
0.86
2.8. The Nominal Categories Model
R
to
Our last set of basic examples will give demonstrations of using flexMIRT
fit the reparameterized nominal categories model (Thissen, Cai, & Bock, 2010).
As discussed in both Thissen and Steinberg (1988) and Thissen et al. (2010),
models equivalent to the Partial Credit Model (Masters, 1982), and Muraki’s
(1992) Generalized Partial Credit Model, among others, may be obtained as
restricted versions of the full-rank nominal model by the choice of contrast
matrices and constraints.
2.8.1
Nominal Model
The first data set contains responses to items in a quality of life scale. First
reported in Lehman (1988), the data set consists of responses from 586 patients
to 35 items from the classic Quality of Life (QOL) Scale for the Chronically
Mentally Ill. All items have 7 response categories: terrible; unhappy; mostly
dissatisfied; mixed, about equally satisfied and dissatisfied; mostly satisfied;
pleased; and delighted.
Example 2.7: Nominal Model Calibration
<Project>
Title = "QOL Data";
Description = "35 Items - Nominal";
<Options>
Mode = Calibration;
Etol = 1e-4;
Mtol = 1e-5;
Processors = 2;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
N = 586;
Ncats(v1-v35) = 7;
29
Model(v1-v35) = Nominal(7);
<Constraints>
Fix (v1-v35),ScoringFn(1);
In the <Groups> section, we have identified the data file, named the variables,
specified the number of response options, and chosen to fit the Nominal model
with 7 categories to the data. In the <Options> section, the convergence criteria for both the E step and M iterations of the EM algorithm have been
adjusted from the default values, by the Etol and Mtol commands, respectively, so that a tightly converged solution can be obtained.
Because the nominal categories model estimates parameters that differ from
the models previously discussed, excerpts from the item parameter tables will
be presented from the output and briefly discussed.
30
31
3.10
3.32
1.82
1.69
1.10
0.92
1.44
1.35
s 5
3.09
Group 1: Group1
s 2
s 3
s 4
0.79
1.15
1.31
Original (Bock, 1972) Parameters, Nominal
Item
Label
Category:
1
v1
a
c
2
v2
a
c
...
34
v34
a
c
35
v35
a
c
0.89
2.19
0.57
2.37
0.70
1.22
0.46
0.93
0.00
0.00
0.00
0.00
0.54
0.59
0.31
0.66
1: Group1
3
4
0.95
1.09
1.06
1.92
0.43
0.87
0.44
1.86
1.52
2.92
1.13
3.39
5
2.55
2.53
1.54
2.24
1.92
2.87
1.42
3.45
6
3.62
2.24
1.74
2.36
-0.70
-0.77
P# gamma 3
9
-0.54
21
-0.42
6.00
6.00
0.46 405
0.59 417
s.e.
0.50
0.52
3.92
4.18
s 7
6.00
2.93
1.57
2.04
2.58
7
4.96
0.91
2.65
2.02
0.44
0.22
0.16
0.15
P# gamma 4
10
0.27
22
0.05
0.30 406
0.37 418
s.e.
0.28
0.30
1.40 398
2.17 410
-0.98
-0.97
---- 397
---- 409
P# alpha 3
2
0.14
14
0.06
s.e.
0.89
1.71
1: Group1
s.e. P# alpha 2
---1
-1.21
---- 13
-1.01
s 6
4.38
Items for Group
1
2
0.00
0.66
0.00
0.94
0.00
0.31
0.00
0.67
Nominal Model Intercept Contrasts for Items for Group 1: Group1
Item
Label Contrasts P# gamma 1
s.e. P# gamma 2
1
v1
Trend
7
0.15
0.15
8
1.64
2
v2
Trend 19
0.34
0.14 20
0.64
...
34
v34
Trend 403
0.26
0.12 404
1.53
35
v35
Trend 415
0.43
0.15 416
1.12
Nominal Model Scoring Function Values
Item
Label
s 1
1
v1
0.00
...
34
v34
0.00
35
v35
0.00
Nominal Model Slopes and Scoring Function Contrasts for Items for Group
Item
Label P#
a
s.e. Contrasts P# alpha 1
1
v1
6
0.83
0.17
Trend
1.00
2
v2 18
0.44
0.17
Trend
1.00
...
34
v34 402
0.49
0.13
Trend
1.00
35
v35 414
0.34
0.14
Trend
1.00
Output 2.17: Nominal Model Output - Item Parameters
-0.02
0.15
P# gamma 5
11
0.14
23
0.23
0.29 407
0.27 419
s.e.
0.21
0.22
0.07
0.14
P# alpha 4
3
0.29
15
-0.11
0.79 399
1.17 411
s.e.
0.48
0.95
0.04
0.11
P# gamma 6
12
0.09
24
0.26
0.25 408
0.29 420
s.e.
0.20
0.22
0.24
0.20
P# alpha 5
4
0.10
16
0.37
0.82 400
1.11 412
s.e.
0.37
0.70
0.19
0.27
s.e.
0.17
0.20
-0.13
-0.20
P# alpha 6
5
-0.19
17
-0.12
0.75 401
0.93 413
s.e.
0.39
0.74
0.56
0.82
s.e.
0.37
0.61
The first 3 tables in Output 2.17 provide parameter estimates using the
reparameterized version of the model (see Thissen et al., 2010). In the first table, the overall slope value, labeled a, the type of α contrast matrix (Ta) used,
and the estimated contrast values, labeled alpha 1 – alpha 3, are reported.
There are 7 categories so there are 6 scoring function contrasts. The second
table, labeled Nominal Model Scoring Function Values reports the scoring
function values for the categories, which are calculated as Ta α, where α are
the estimated scoring contrasts. The third table of item parameters reports
the type of contrast matrix used for the intercepts (Tc) and the estimated γ
parameters. For convenience, the parameters are also reported in the original
parameterization in the final table of the output presented.
The next example employs a set of data regarding numerical knowledge
from pre-school aged children, most notably analyzed in Thissen and Steinberg (1988) to illustrate the concept of testlets. As described by Thissen and
Steinberg (1988), the data were collected from 4 tasks that involved asking 592
young children to identify the numerals representing the quantities of 3 and
4 and, separately, match the numeral for each number to a correct quantity
of blocks. From these items, called Identify3, Identify4, Match3, and Match4,
pseudo-items were created that summarized a child’s performance on both
numbers. For the new pseudo-items, Identify and Match, the response options
are 0 (child could do the requested action for neither number), 3 (child could
identify/match 3 only), 4 (child could identify/match 4 only), and 34 (child
could identify/match both 3 and 4 successfully). We will again fit the Nominal
Model, but with some additional constraints imposed.
Example 2.8: Nominal Model Calibration 2
<Project>
Title = "Pre-School Numerical Knowledge";
Description = "2 Redefined Items Nominal";
<Options>
Mode = Calibration;
SE = SEM;
Etol = 5e-5;
Mtol = 1e-9;
GOF = Extended;
M2 = Full;
SX2Tbl = Yes;
HabermanResTbl = Yes;
FitNullModel = Yes;
32
SavePRM = Yes;
<Groups>
%Group1%
File = "PreschoolNum.dat";
Varnames = Identify,Match,Identify3,Identify4,Match3,Match4;
Select = Identify,Match;
N = 592;
Ncats(Identify,Match) = 4;
Code(Identify,Match) = (0,3,4,34),(0,1,2,3);
Model(Identify,Match) = Nominal(4);
Ta(Identify,Match) = Identity;
Tc(Identify,Match) = Identity;
<Constraints>
Fix (Identify),ScoringFn(2);
Equal (Match), ScoringFn(2) : Group1, (Match), ScoringFn(3);
Equal Group1, (Match), Intercept(1) : Group1, (Match), Intercept(2);
Within the <Options> section, there are several new commands. The
method of standard error calculation has been changed from the default of
empirical cross-product approximation (Xpd) to Supplemented EM algorithm
(keyword is SEM; see Cai, 2008). We have also requested the full version of the
limited information GOF M2 statistic (Cai & Hansen, 2012; Maydeu-Olivares
& Joe, 2005) and, additionally, specified that the zero-factor Null Model should
be estimated and the associated incremental GOF indices be printed.
In the <Groups> section, we have identified the data file, named the variables, selected only the pseudo-items to retain for analysis, specified the number of response options, recoded the response options into zero-based sequential (but nevertheless nominal) codes, and chosen to fit the Nominal model
with 4 categories to the data. The statements Ta and Tc are specific to the
nominal model and refer to the contrast matrices for the scoring function and
intercept parameters, respectively. The options for the scoring function and
intercept contrasts include a Trend matrix (which uses a Fourier basis, augmented with a linear column), an Identity matrix, or a user-supplied matrix.
In this example, both of the contrast matrices have been set to type Identity
to facilitate equality constraints. Within the <Constraints> section, we encounter a more complicated set of constraints than in the previous example.
Suffice it to say that these constraints are applied to the scoring functions and
intercepts so that the solution matches with Thissen and Steinberg’s (1988)
preferred testlet solution. Estimated item parameters are presented in Output
2.18.
33
34
Original (Bock, 1972) Parameters, Nominal Items for Group 1: Group1
Item
Label
Category:
1
2
3
4
1 Identify
a
0.00
0.00
0.97
2.41
c
0.00
-0.86
-0.80
0.65
2
Match
a
0.00
0.38
0.38
2.45
c
0.00
-0.42
-0.42
0.52
Nominal Model Intercept Contrasts for Items for Group 1: Group1
Item
Label Contrasts P# gamma 1
s.e. P# gamma 2
s.e.
1 Identify Identity
3
-0.86
0.14
4
-0.80
0.43
2
Match Identity
9
-0.42
0.28
9
-0.42
0.28
Nominal Model Scoring Function Values Group 1: Group1
Item
Label
s 1
s 2
s 3
s 4
1 Identify
0.00
0.00
1.21
3.00
2
Match
0.00
0.46
0.46
3.00
P# gamma 3
5
0.65
8
0.52
Nominal Model Slopes and Scoring Function Contrasts for Items for Group 1: Group1
Item
Label P#
a
s.e. Contrasts P# alpha 1
s.e. P# alpha 2
1 Identify
2
0.80
0.61 Identity
1.00
---0.00
2
Match
6
0.82
0.62 Identity
1.00
---7
0.46
Output 2.18: Nominal Model 2 Output - Item Parameters
s.e.
0.20
0.20
s.e.
---0.31
P# alpha 3
1
1.21
7
0.46
s.e.
0.47
0.31
Although several new command statements were introduced in this example, only two result in additions to the output file. Both the M2 and
FitNullModel statements provide additional GOF information that is printed
near the bottom of the output file, with the other overall GOF index values.
The fitted model’s −2×log-likelihood, AIC, and BIC are printed as before,
but additional fit statistics for the zero-factor null model are also included
under the corresponding fitted model section. The estimation of the independence model permits the calculation of the non-normed fit index (NNFI;
aka Tucker-Lewis Index). The usual full-information fit statistics are printed,
but limited-information statistics, based on M2 , for the fitted and zero-factor
model are also given. Taken together, the fit indice values indicate that the
nominal model provides excellent fit to this data.
Output 2.19: Nominal Model 2 Output - GOF values
Statistics based on the loglikelihood of the fitted model
-2loglikelihood:
2766.69
Akaike Information Criterion (AIC):
2784.69
Bayesian Information Criterion (BIC):
2824.14
Statistics based on the loglikelihood of the zero-factor null model
-2loglikelihood:
2885.67
Akaike Information Criterion (AIC):
2895.67
Bayesian Information Criterion (BIC):
2917.58
Full-information fit statistics of the fitted model
Degrees
G2 of freedom Probability
F0hat
8.45
6
0.2064
0.0143
Degrees
X2 of freedom Probability
F0hat
8.99
6
0.1739
0.0152
RMSEA
0.03
RMSEA
0.03
Full-information fit statistics of the zero-factor null model
Degrees
G2 of freedom Probability
F0hat
RMSEA
127.43
10
0.0001
0.2153
0.14
Degrees
X2 of freedom Probability
F0hat
RMSEA
124.58
10
0.0001
0.2104
0.14
With the zero-factor null model, Tucker-Lewis (non-normed) fit index
based on G2 is
0.97
35
With the zero-factor null model, Tucker-Lewis (non-normed) fit index
based on X2 is
0.96
Limited-information fit statistics of the fitted model
Degrees
M2 of freedom Probability
F0hat
RMSEA
0.17
6
0.9999
0.0003
0.00
Note: M2 is based on full marginal tables.
Note: Model-based weight matrix is used.
Limited-information fit statistics of the zero-factor null model
Degrees
M2 of freedom Probability
F0hat
RMSEA
117.87
10
0.0001
0.1991
0.13
Note: M2 is based on full marginal tables.
Note: Model-based weight matrix is used.
With the zero-factor null model, Tucker-Lewis (non-normed) fit index
based on M2 is
1.09
2.8.2
Generalized Partial Credit Model
We return the QOL data again to illustrate Muraki’s (1992) Generalized Partial Credit (GPC) model. The GPC model is a constrained special case of the
nominal model. To set up a GPC model using the reparameterized nominal
model, one fixes all scoring function contrasts. This results in the first scoring
function contrast being fixed to 1.0 and the other contrasts fixed to 0.0. With
such constraints in place, the category scoring function values are equal to
0, 1, . . . , k, which is the necessary structure for fitting the GPC. Because the
R
GPC is such a popular special case of the nominal model in flexMIRT
2.0
we have added GPC() as an option to the Model()= ; statement. This option will automatically fix the scoring functions as needed, freeing the user
from making such technical specifications. The following syntax demonstrates
the use of this new model keyword. As with the Graded model specification,
when one uses the GPC() keyword in the Model statement, the model name is
immediately followed by the number of categories in parentheses.
Example 2.9: Generalized Partial Credit Model
<Project>
Title = "QOL Data";
36
Description = "35 Items - GPC";
<Options>
Mode = Calibration;
Etol = 1e-4;
Mtol = 1e-5;
Processors = 2;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
N = 586;
Ncats(v1-v35) = 7;
Model(v1-v35) = GPC(7);
<Constraints>
2.8.3
Rating Scale Model
Andrich’s (1978) Rating Scale (RS) model may also be obtained as a special case of the reparameterized nominal model. To obtain the RS model in
R
flexMIRT
it is necessary to fix the scoring functions for all items. Additionally, we need to constrain the slopes for all items to be equal, as well as set
each intercept to be equal across the items. This is all accomplished via the
constraints imposed in Example 2.10.
Example 2.10: Rating Scale Model with Trend Contrasts
<Project>
Title = "QOL Data";
Description = "35 Items RS Using Default Trend Contrasts";
<Options>
Mode = Calibration;
SE = SEM;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
37
N = 586;
Ncats(v1-v35) = 7;
Model(v1-v35) = Nominal(7);
<Constraints>
Fix (v1-v35),ScoringFn;
Equal (v1-v35),Slope;
Equal (v1-v35),Intercept(2);
Equal (v1-v35),Intercept(3);
Equal (v1-v35),Intercept(4);
Equal (v1-v35),Intercept(5);
Equal (v1-v35),Intercept(6);
R
The previous RS fitting used the flexMIRT
default trend contrasts. To
obtain results for the RS model similar to those that would be produced by
R
Multilog, the user may provide flexMIRT
with the triangle contrasts utilized
by Multilog. This is done through the Tc command and, rather than using
one of the available keywords, supplying the values that compose the desired
contrast. The triangle contrast is shown in the Tc statement in the following
example syntax.
Example 2.11: Rating Scale Model with Triangle Contrasts
<Project>
Title = "QOL Data";
Description = "35 Items RS Multilog-style Triangle Contrasts";
<Options>
Mode = Calibration;
SE = SEM;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
N = 586;
Ncats(v1-v35) = 7;
Model(v1-v35) = Nominal(7);
Tc(v1-v35) = (
38
0
-1
-2
-3
-4
-5
-6
0
-1
-1
-1
-1
-1
0
0 0 0 0,
0 0 0 0,
-1 0 0 0,
-1 -1 0 0,
-1 -1 -1 0,
-1 -1 -1 -1,
0 0 0 0 );
<Constraints>
Fix (v1-v35),ScoringFn;
Equal (v1-v35),Slope;
Equal (v1-v35),Intercept(2);
Equal (v1-v35),Intercept(3);
Equal (v1-v35),Intercept(4);
Equal (v1-v35),Intercept(5);
Equal (v1-v35),Intercept(6);
R
Thus far, we have covered how to use flexMIRT
to calibrate several of the
most commonly used single-group unidimensional IRT models, and to obtain
IRT scale scores. While these types of models are useful, the greatest strength
R
of flexMIRT
is its ability to readily handle complex models (e.g., multidimensional, bifactor, testlet, etc.) and more complex data structures, such as
multilevel data. In the next chapter, we will provide instruction and examples,
building intuitively from the basic syntax already presented, to demonstrate
R
how flexMIRT
syntax files are constructed to handle more complex models
and data structures.
39
CHAPTER 3
Advanced Modeling
The previous chapter focused on basic unidimensional, single group models to
R
introduce the flexMIRT
syntax, output, and available IRT models. While
these examples are useful in their own right, it is expected that a user working with real-world data will need more advanced modeling options. In this
R
chapter, we will provide examples for using flexMIRT
when multiple groups,
multiple factors, and/or a hierarchical data structure need to be modeled.
3.1. Multiple Group Calibration
The data set used to illustrate 2PL and 3PL examples in the previous chapter
is actually made up of responses from two groups of examinees, 3rd and 4th
graders. Using this already familiar data set, we will demonstrate multiple
group analysis.
Example 3.1: 2-Group 3PL Using Normal Priors with EAP Scoring
<Project>
Title = "G341-19";
Description = "12 Items 3PL 2 Groups
Calibration with Normal Prior, Saving Parameter Estimates,
Followed by Combined Scoring Run for EAP Scores";
<Options>
Mode = Calibration;
Progress = Yes;
SaveSCO = Yes;
SavePRM = Yes;
SaveInf = Yes;
Score = EAP;
FisherInf = 81,4.0;
SE = Xpd;
40
GOF = Extended;
M2 = Full;
FitNullModel = Yes;
<Groups>
%Grade4%
File = "g341-19.grp1.dat";
Varnames = v1-v12;
N = 1314;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
BetaPriors(v1-v12) = 1.5;
%Grade3%
File = "g341-19.grp2.dat";
Varnames = v1-v12;
N = 1530;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
BetaPriors(v1-v12) = 1.5;
<Constraints>
Free Grade3, Mean(1);
Free Grade3, Cov(1,1);
Equal Grade3, (v1-v12), Guessing : Grade4, (v1-v12), Guessing;
Equal Grade3, (v1-v12), Intercept: Grade4, (v1-v12), Intercept;
Equal Grade3, (v1-v12), Slope : Grade4, (v1-v12), Slope;
Prior Grade3, (v1-v12), Guessing : Beta(1.0,4.0);
Two groups (%Grade4% and %Grade3%) are specified in the <Groups> section. An important point to highlight is that the data for each group should
be stored as individual files. To that end, “g341-19.dat” is split into 2 new
files “g341-19.grp1.dat” and “g341-19.grp2.dat”. As may be seen in the N =
statements for each group, the original sample size of 2844 is now divided
between the two groups. Within each group the basic specification statements (VarNames, Ncats, etc) remain the same. There are now two sets of
them, one per group. A new option not specific to multiple group analysis, is
BetaPriors. This statement imposes a prior on the item uniquenesses in the
form of a Beta(α,1.0) distribution with a user-specified α parameter, in this
case α = 1.5 (see Bock, Gibbons, & Muraki, 1988).
41
Within the <Constraints> section, we are allowing the mean and variance
of the second group (Grade3) to be estimated by freeing those parameters via
the first two constraint statements. Additionally, we have specified a model
that assumes measurement invariance by constraining the item parameters to
be equal across the two groups. Finally, we supplied a Beta distribution prior
for the guessing parameters.
There are new commands encountered in the <Options> section as well.
It should be noted that these newly encountered commands are not specific
in any way to the multi-group format and are available for use with single
group analyses as well. For instance, the Progress statement (via Yes/No
R
keywords) determines whether flexMIRT
will print progress updates in the
console window. Using SaveInf= Yes;, we have requested that the Fisher
information function for the items and the total scale be saved to an external
file and have specified via FisherInf = 81,4.0; that we would like information values calculated at 81 points, equally spaced from -4.0 to 4.0. The
method of standard error calculation has been specified as empirical crossproducts approximation, using SE = Xpd (all possible SE calculation methods
and corresponding keywords are covered in Chapter 5: Details of the Syntax).
3.2. Characterizing Non-Normal Population Distributions
This example is substantially the same as the analysis just discussed, with
an additional option included in the calibration portion. As the command
file does not change drastically, only the modified sections are presented here.
The full syntax file (g341-19.calib.EH.flexmirt) is available in the examples
R
provided on the flexMIRT
website.
Example 3.2: 2-Group 3PL with Empirical Histogram (Excerpts)
...
<Options>
...
Etol = 5e-4;
<Groups>
%Grade4%
File = "g341-19.grp1.dat";
Varnames = v1-v12;
N = 1314;
Ncats(v1-v12) = 2;
42
Model(v1-v12) = ThreePL;
BetaPriors(v1-v12) = 1.5;
EmpHist = Yes;
%Grade3%
File = "g341-19.grp2.dat";
Varnames = v1-v12;
N = 1530;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
BetaPriors(v1-v12) = 1.5;
EmpHist = Yes;
...
The primary point of interest here is the EmpHist = Yes; statement found
as the last command provided for each group. With this option invoked,
R
flexMIRT
will estimate the item parameters and the population (theta) distribution simultaneously. The latent variable distribution is estimated as an
empirical histogram (e.g., Mislevy, 1984; Woods, 2007). By “empirical histogram” we mean the normalized accumulated posterior densities for all response patterns at each quadrature node - that is, an empirical characterization
of the shape of the examinee ability distribution in the population. The use of
an empirical histogram is in contrast to a typical calibration run in which the
latent trait is assumed to be normally distributed. This estimation method is
currently only available for single-level unidimensional and single-level bifactor
(or testlet response models) with one general dimension. In the <Options>
section, the convergence criterion for the E step of the EM algorithm has
been increased from the default of 1e-4 to 5e-4 to accommodate the added
uncertainty due to the estimation of the empirical histogram prior.
3.3. Multiple Group Scoring Only
Scoring runs with multiple groups proceed in much the same way as the single
group scoring analyses. The current scoring example makes uses of the item
parameters obtained from the first multiple group example (Example 3.1 which
employed normal priors). The syntax is presented below and, as with the single
group scoring syntax, in the <Options> section Mode is set to Scoring, the type
of scores requested is specified (MAPs in this case) and the file containing the
relevant item parameters (generated by our previous calibration run) is given
using the ReadPRMFile statement. Because of the distinct groups, two sets
of data descriptor statements (e.g., syntax for declaring group label, data file
43
name, variable names, and sample size) are given. Because this example is so
similar to the single group scoring runs, no further discussion is provided. For
explications of any of the included syntax statements, the user should refer to
a previous single-group example (e.g., Example 2.3).
Example 3.3: 2-Group 3PL MAP Scoring from Parameter Estimates
<Project>
Title = "G341-19";
Description = "12 Items 3PL 2 Groups MAP from Parameter Estimates";
<Options>
Mode = Scoring;
ReadPRMFile = "g341-19.calib.normal-prm.txt";
Score = MAP;
<Groups>
%Grade4%
File = "g341-19.grp1.dat";
Varnames = v1-v12;
N = 1314
%Grade3%
File = "g341-19.grp1.dat";
Varnames = v1-v12;
N = 1530
<Constraints>
It some situations, it may be useful to obtain the sum score conversion
table, without actually having to score data. For example, when only the
SSC table is wanted and the sample size is large enough that calculating
scores for every individual would take a non-trivial amount of computing time,
R
flexMIRT
is able to by-pass individual scoring and produce only the necessary table. The syntax to do so is straight-forward and presented below.
Example 3.4: 2-Group 3PL Summed Score to EAP Conversion Table
<Project>
Title = "G341-19";
Description = "12 Items 3PL 2 Groups Summed Score to EAP from
Parameter Estimates Using Estimated Empirical Histogram Prior
Make Table Only";
<Options>
44
Mode = Scoring;
ReadPRMFile = "g341-19.calib.EH-prm.txt";
Score = SSC;
<Groups>
% Grade4 %
Varnames = v1-v12;
% Grade3 %
Varnames = v1-v12;
<Constraints>
Just as a typical scoring run, we have specified the parameter file (now using parameters obtained from the multiple-group calibration which estimated
theta via an empirical histogram). We have set the scoring method to produce
the desired SSC (Summed Score to EAP Conversion) table. In the <Groups>
section, we still provide labels and variable names for both groups, but the
sample size and data file names are omitted because they are not needed for
producing the conversion table without scoring any individuals.
3.4. DIF Analyses
Up to this point, the multiple group examples have assumed the invariance of
item parameters across groups. However, when multiple groups are involved,
it is generally prudent to check that assumption by looking for differential item
R
functioning (DIF) between groups. flexMIRT
provides two general methods
for DIF testing (Langer, 2008): 1) a DIF sweep for all items, and 2) a more
focused examination, where candidate DIF items are tested with designated
anchor items. Both methods utilize the Wald test. We will first cover the more
general DIF sweep, again using the end-of-grade testing data. As the syntax is
quite similar to what was presented earlier, only the relevant changes will be
noted in the excerpt. The full syntax file may be found in the corresponding
R
folder on the flexMIRT
support website.
Example 3.5: DIF: Test All Items (Excerpts)
...
<Options>
...
DIF = TestAll;
DIFcontrasts = (1.0 -1.0);
...
<Constraints>
45
...
Prior Grade3, (v1-v12), Guessing : Normal(-1.1,0.5);
Prior Grade4, (v1-v12), Guessing : Normal(-1.1,0.5);
First, in conjunction with the use of the 3PL model, a logit-normal prior
is applied to the guessing parameters, which was a previously unseen option.
With respect to the DIF analysis, the relevant new statements are contained
in the <Options> section. To enable the general DIF sweep, testing all items,
DIF = TestAll; is specified. Also required when testing for DIF are values
R
with the entries for a DIF contrast matrix, specifying
that supply flexMIRT
how contrasts among groups should be constructed. In this example, we are
conducting a straight-forward comparison of Grade3 against Grade4, so the
statement appears as, DIFcontrasts = (1.0 -1.0); with spaces between the
supplied values for each contrast. When k groups are present, k − 1 contrasts
must be specified, (e.g., for 3 groups, DIFcontrasts = (2.0 -1.0 -1.0, 0.0
1.0 -1.0);, with commas in between). Because the syntax file is free-form,
the contrasts may appear on more than one line. For example,
DIFcontrasts = (
2.0 -1.0 -1.0,
0.0 1.0 -1.0
);
is entirely permissible.
From the DIF analysis, an additional table is presented in the output.
46
47
DIF Statistics for 3PL Items
Item numbers in:
Grp1 Grp2 Total X2 d.f.
p
1
1
1.8
3 0.6081
2
2
8.5
3 0.0370
3
3
0.5
3 0.9177
4
4
17.1
3 0.0007
5
5
1.8
3 0.6053
6
6
0.6
3 0.8922
7
7
0.7
3 0.8622
8
8
8.8
3 0.0317
9
9
3.6
3 0.3093
10
10
1.5
3 0.6779
11
11
1.9
3 0.5977
12
12
0.8
3 0.8486
X2g
0.1
0.2
0.0
0.0
0.1
0.1
0.0
0.0
0.2
0.0
0.9
0.5
Output 3.1: DIF Analyses Output Table
d.f.
1
1
1
1
1
1
1
1
1
1
1
1
p
0.8141
0.6717
0.9096
0.8647
0.7717
0.7909
0.8730
0.9159
0.6730
0.9055
0.3405
0.4906
X2a|g
1.4
4.8
0.4
0.3
1.2
0.4
0.3
7.3
0.4
0.7
0.0
0.0
d.f.
1
1
1
1
1
1
1
1
1
1
1
1
p
0.2371
0.0280
0.5425
0.5910
0.2731
0.5058
0.6061
0.0067
0.5115
0.4001
0.9424
0.9307
X2c|a,g
0.4
3.5
0.1
16.8
0.6
0.1
0.5
1.5
3.0
0.8
1.0
0.3
d.f.
1
1
1
1
1
1
1
1
1
1
1
1
p
0.5399
0.0621
0.7281
0.0001
0.4555
0.7462
0.5004
0.2269
0.0849
0.3726
0.3263
0.5717
For each item that is present in both groups, a line in the table is added
that reports the various Wald tests for that item. As we fit the 3PL to the
items, we are provided with an overall test for DIF, a test of DIF with respect
to the guessing parameter (labeled X2g), a test of the slope, assuming g is
fixed across groups (labeled X2a|g), and a test of the intercept, assuming a
and g are fixed. Each of the reported Wald values also has the associated
degrees of freedom and p-value printed. Woods, Cai, and Wang (2013) found
R
that the Wald-2 procedure, implemented via DIF = TestAll; in flexMIRT
,
tends to result in inflated Type I error rates and suggested that this procedure
may be best used for identifying anchors for a subsequent DIF analysis that
examines only candidate items for DIF, rather than as a final DIF analysis
from which conclusions regarding item functioning are drawn. Additionally,
it is also recommended that when using the TestAll DIF sweep option that
the standard error method be changed from the default. Conclusions reported
in Woods et al. (2013) were based on the use of supplemented EM SEs (SE =
SEM;).
R
flexMIRT
may also be used for more traditional DIF testing of test candidate items against designated anchor items. Only relevant excerpts of the
command file, “g341-19.difcand.flexmirt", are presented. Again, the full file is
available on the support website.
Example 3.6: DIF: Test Candidate Items (Excerpts)
...
<Options>
...
DIF = TestCandidates;
DIFcontrasts = (1.0 -1.0);
<Groups>
%Grade4%
File = "g341-19.grp1.dat";
Varnames = v1-v12;
N = 1314;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
DIFitems = v1,v2;
%Grade3%
File = "g341-19.grp2.dat";
Varnames = v1-v12;
48
N = 1530;
Ncats(v1-v12) = 2;
Model(v1-v12) = ThreePL;
DIFitems = v1,v2;
<Constraints>
Free Grade3, Mean(1);
Free Grade3, Cov(1,1);
Equal Grade3, (v3-v12), Guessing : Grade4, (v3-v12), Guessing;
Equal Grade3, (v3-v12), Intercept: Grade4, (v3-v12), Intercept;
Equal Grade3, (v3-v12), Slope
: Grade4, (v3-v12), Slope;
...
The keyword used with the DIF statement is now TestCandidates, as
opposed to TestAll in the previous example. The DIFcontrasts statement
remains unchanged. To specify candidate DIF items, within the <Groups> section, we simply list the candidate item names using the DIFitems statement,
found as the last entry within each group subsection of the excerpted code.
Here, we are testing Items 1 and 2 for possible DIF. As can be seen in the
<Constraints> section, the item parameters for the anchor items (v3-v12)
are set equal across groups. The DIF table printed in the output is structured
as in the previous example and will not be covered here.
3.5. Multiple Group Multilevel Model
R
Our final multiple group example will cover the flexMIRT
syntax options for
handling nested data. Using the student cognitive outcomes data from the
2000 Program for International Student Assessment (see Adams & Wu, 2002),
31 mathematics items with mixed formats (multiple choice and constructed
response) from students in both the United States and Ireland will be analyzed.
The US data set is comprised of 2115 students, representing 152 different
schools, and the Irish data set has 2125 student observations, coming from 139
different schools. From the general description of the data, it is clear that the
complete set of data for this examples has two groups (US and Irish students)
and within each country, students (level-1 units) are nested within schools
(level-2 units). Because the full command file (PISA00mathL2.flexmirt) is
quite long, only commands relevant to multilevel aspect of the analysis will be
highlighted.
49
Example 3.7: Multiple Group Multilevel Model (Excerpts)
...
<Groups>
%USA%
File = "PISA00Math.US.dat";
Missing = 9;
Varnames = SchID,StdID,
ViewRoomQ1,
BricksQ1,
...
CarpenterQ01,
PipelinesQ1;
Select = ViewRoomQ1,...,PipelinesQ1;
CaseID = StdID;
Cluster = SchID;
Dimensions = 2;
Between = 1;
N = 2115;
...
%Ireland%
File = "PISA00Math.IE.dat";
...
CaseID = StdID;
Cluster = SchID;
Dimensions = 2;
Between = 1;
N = 2125;
...
<Constraints>
...
Free USA,Cov(1,1);
Free Ireland,Mean(1);
Free Ireland,Cov(1,1);
The data files are specified as usual within each group. We have also set a missing data code using Missing = 9;, so all 9s in the data set will be interpreted
as missing values, rather than the default missing value of -9. The variables
are named in the typical fashion; the first two variables are of note because
50
they indicate student ID numbers and school ID numbers, respectively, and
will not be selected for IRT analysis. In general, a multilevel analysis has two
kinds of latent dimensions, Between and Within. The Between dimensions
R
will always precede the Within dimensions when flexMIRT
is assigning dimensions. For this example, there are only two dimensions total so we specify
the total number of dimensions to be used for the model (via Dimensions=2;)
and also denote that at level-2 there is 1 between-school dimension (Between
= 1;). This implies that there is 1 within-school dimension at the student
R
level. In the estimation and output, flexMIRT
will use the first dimension
for between-school variations and the second dimension for within-group variations.
When a 2-level model has been requested, it is required that a variable
defining the higher-level units be specified via the Cluster statement. We
also provide a variable that identifies the first level data (individuals) with
the CaseID statement. The latter can be convenient if the individual IDs are
R
unique so that the scale scores computed by flexMIRT
can be merged into
other data sets for further analysis.
The <Constraints> section contains extensive statements for setting up
the necessary priors and equality constraints both within and across the two
groups so that the desired hierarchical model is estimated. With the betweengroup equality constraints, the item parameters (for both level-1 and level-2
models) are set equal across the two countries, thereby creating anchoring
so that the mean of Ireland schools can be freely estimated. The withingroup equality constraints are set up so that the item slopes for the within
dimension is equal to those for the between dimension. This implies that
a random-intercept only model is fitted in each country, which permits the
decomposition of between-school versus within-school variance.
51
Output 3.2: Two-Level Model - Means and Variances
...
PISA 2000 Math Data: Two-Level Unidimensional
31 Items
US Data: 2115 Students, 152 Schools
IE Data: 2125 Students, 139 Schools
Summary of the Data and Control Parameters
Group
USA
Ireland
Missing data code
9
9
Number of Items
31
31
Number of L-2 Units
152
139
Number of L-1 Units
2115
2125
Number of Dimensions
2
2
Between Dimensions
1
1
Within Dimensions
1
1
...
Group Parameter Estimates:
Group
Label P#
mu 1
s.e. P#
1
USA
0.00
---2
Ireland 79
0.36
0.34
mu 2
0.00
0.00
s.e.
-------
Latent Variable
P# Theta 1
1
0.52
0.00
Variance-Covariance Matrix for Group1: USA
s.e. P# Theta 2
s.e.
0.07
---1.00
----
Latent Variable
P# Theta 1
80
0.11
0.00
Variance-Covariance Matrix for Group2: Ireland
s.e. P# Theta 2
s.e.
0.02
---1.00
----
We can see from the output above that the summary of data shows the multilevel syntax has been correctly interpreted. On average, the overall schoollevel achievement in Ireland is 0.36 standard deviations higher than that of
the US. The more interesting result, however, is in the variance decompositions. In the US, with a fixed within-school variance of 1.0 (for identification),
the between-school variance is estimated as 0.52. This implies an intra-class
correlation of 0.52/(1 + 0.52) = 0.34, showing a significant amount of school
level variations in the US. On the other hand, the intra-class correlation for
Ireland is equal to 0.11/(1 + 0.11) = 0.10, which implies that the school
achievement-levels are much more homogeneous.
52
3.6. Multidimensional Model - 4 Correlated Factors
Although the previous example was multidimensional, estimating two factors,
it was also dealing with nested/hierarchical data. Because the fitting of a
multidimensional model was mixed in with the hierarchical modeling syntax,
we provide a more straight-forward multidimensional example by fitting a
model with four correlated latent factors to a single group. The QOL dataset
was introduced in Chapter 2, but a previously unmentioned aspect of the scale
is it was designed to assess both the overall quality of life as well as the quality
of life in 7 more specific subdomains (family, money, health, leisure, living,
safety, and social), with each subdomain containing 4 to 6 items. Item 1 is a
global quality of life item. We will be fitting a four factor model to a subset
of the items of the previously introduced QOL scale (items from the family,
money, health, and leisure factors) and allowing the latent factors to correlate.
Example 3.8: Four Factor GRM Model with Correlated Factors
<Project>
<Project>
Title = "QOL Data";
Description = "Items 2-21 with Correlated Factors";
<Options>
Mode = Calibration;
Quadrature = 21,5.0;
SavePRM = Yes;
Processor = 4;
FactorLoadings = Yes;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
Select = v2-v21;
N = 586;
Dimensions = 4;
53
Ncats(v2-v21) = 7;
Model(v2-v21) = Graded(7);
<Constraints>
Fix (v2-v21),Slope;
Free
Free
Free
Free
(v2-v5),Slope(1);
(v6-v9),Slope(2);
(v10-v15),Slope(3);
(v16-v21),Slope(4);
Free Cov(2,1);
Free Cov(3,1);
Free Cov(3,2);
Free Cov(4,1);
Free Cov(4,2);
Free Cov(4,3);
Points of note in the syntax, relevant to the fitted multidimensional model are
the Dimensions statement in the <Groups> section and the various constraints
encountered in the <Constraints> section. As described earlier, we are fitting
a model with four factors, hence the Dimensions = 4; statement. Within the
<Constraints> section, we are assigning the items to the appropriate factors.
The first step in assigning items to factors to fix all slope values to 0, via
Fix (v2-v21),Slope;, which allows us, in the next step, to free the individual
items to load on only the appropriate factors. As can be seen in the four Free
statements, we are specifying that Items 2 through 5 load on the first factor,
Items 6 through 9 load on the second factor, Items 10 through 15 load are
assigned to the third factor, and so on. This corresponds to the description
provided earlier, in which Items 2 to 5 belong to the “Family" factor only and
Items 6 to 9 belong to only the “Money" factor, etc. The final constraints,
R
such as Free Cov(2,1);, are used to tell flexMIRT
to freely estimate the
covariance/correlation between the factors. Using the lower triangle of the
variance/covariance matrix of the latent factors to determine positions, the
variance of the first factor is found at Cov(1,1), the variance of the second
factor at Cov(2,2) and the covariance of the two factors is located in Row 2,
Column 1 (i.e., Cov(2,1)) of the matrix and so on.
54
Additionally, in the <Options> section, the Quadrature command is encountered for the first time. This statement allows the user specify the number
of quadrature points to be used, 21 in this case, and the range over which
those points should be spread (from -5.0 to 5.0 here). Reducing the number
of quadrature points from the default value (49 points per dimension) may be
necessary when fitting multidimensional models because the total number of
points used is exponentially related to the number of dimensions. With higher
dimensional models, the total number of quadrature points (and therefore processing time) may become unwieldy if the default value is maintained.
Within the output, the various sections are arranged as with the other examples. In the output file, the factor loadings are printed below the estimated
item parameters, as requested by the FactorLoadings = Yes; statement. As
seen before, the group means of the latent variables are reported followed by
the estimated latent variable variance/covariance matrix. This section of the
output is presented below.
Output 3.3: Correlated Four Factor Model Output - Loadings, Means and
Variances
Factor Loadings for Group 1: Group1
Item
Label lambda 1
1
v2
0.83
2
v3
0.78
3
v4
0.93
4
v5
0.92
5
v6
0.00
6
v7
0.00
7
v8
0.00
8
v9
0.00
9
v10
0.00
10
v11
0.00
11
v12
0.00
12
v13
0.00
13
v14
0.00
14
v15
0.00
15
v16
0.00
16
v17
0.00
17
v18
0.00
18
v19
0.00
19
v20
0.00
20
v21
0.00
s.e. lambda 2
0.03
0.00
0.03
0.00
0.02
0.00
0.02
0.00
---0.87
---0.72
---0.90
---0.88
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
s.e. lambda 3
---0.00
---0.00
---0.00
---0.00
0.03
0.00
0.05
0.00
0.02
0.00
0.03
0.00
---0.66
---0.74
---0.71
---0.72
---0.78
---0.72
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
s.e. lambda 4
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
0.06
0.00
0.05
0.00
0.05
0.00
0.05
0.00
0.04
0.00
0.05
0.00
---0.77
---0.73
---0.76
---0.84
---0.75
---0.49
s.e.
------------------------------------------0.04
0.04
0.04
0.03
0.04
0.07
QOL Data
Items 1-21 Correlated Factors with Correlated Factors
Group Parameter Estimates:
Group
Label
1
Group1
Latent Variable
P# Theta 1
1.00
141
0.41
142
0.55
144
0.55
P#
Variance-Covariance
s.e. P# Theta 2
---0.04
1.00
0.03 143
0.50
0.04 145
0.56
mu 1
0.00
s.e.
----
P#
mu 2
0.00
s.e.
----
P#
mu 3
0.00
Matrix for Group1: Group1
s.e. P# Theta 3
s.e. P# Theta 4
s.e.
---0.04
0.03 146
----
1.00
0.69
---0.03
55
1.00
s.e.
----
P#
mu 4
0.00
s.e.
----
The pattern of the estimated loadings indicates that the items were assigned to the factors as expected, with certain loadings restricted to 0.00 and
having no associated SE. Because there is only a single group, the latent variable means were also restricted to zero and have no associated SE, indicating
they were not estimated. The final section in the output excerpt reports the
correlation matrix of the latent variable - for example, the correlation between
the first two standardized factors is estimated to be 0.41, with a SE of 0.04.
3.7. The Bifactor Model
The item bifactor model is a special multidimensional IRT model, in which
each item is restricted to load on, at most, two factors. The typical set up for
a bifactor model is one factor on which all the items load, termed the general
factor, and several more specific factors, termed group-specific factors, onto
which only subsets of the items can load. A common substantive example of
this would be a reading test made up of several short reading passages that
each have several items probing the students comprehension for a particular
passage. The general factor in this case would be “Reading” and each passage
would have its own specific factor, onto which only the items related to that
passage load.
Note that Bradlow, Wainer, and Wang’s (1999) testlet response theory
model can be understood as a further restricted bifactor model with withinitem proportionality constraints on the general and group-specific slopes (see
also, Glas, Wainer, & Bradlow, 2000). With generalized capability to impose
R
constraints, flexMIRT
can estimate testlet models too. Chief among the
benefits of the item bifactor model is computational efficiency in which fullinformation maximum marginal likelihood estimation of item parameters can
proceed with 2-dimensional integration regardless of how many group-specific
factors are in the model, due to a result first discovered by Gibbons and
R
Hedeker (1992). The bifactor model in flexMIRT
is in fact obtained as a
R
two-tier model (Cai, 2010a). flexMIRT is fully capable of multiple-group
bifactor or two-tier modeling as detailed in Cai, Yang, and Hansen (2011).
3.7.1
Bifactor Model Using the Reparameterized Nominal Categories Model
Here, we use the full 35 item Quality of Life data to illustrate a bifactor
analysis, specifying a general QOL factor and 7 specific factors to capture
the subdomains mentioned earlier. Gibbons et al. (2007) reported a graded
bifactor analysis of this data set using the original 7 categories and we will
56
replicate this structure but employ a full-rank nominal model.
Example 3.9: Bifactor Structure - Nominal Model
<Project>
Title = "QOL Data";
Description = "35 Items Bifactor Nominal Model";
<Options>
Mode = Calibration;
Quadrature = 21,5.0;
Etol = 1e-4;
Mtol = 1e-5;
Processors = 2;
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
N = 586;
Ncats(v1-v35) = 7;
Model(v1-v35) = Nominal(7);
Dimensions = 8;
Primary = 1;
<Constraints>
Fix (v1-v35),ScoringFn(1);
Fix (v1-v35),Slope;
Free (v1-v35),Slope(1);
Free (v2-v5),Slope(2);
Free (v6-v9),Slope(3);
Free (v10-v15),Slope(4);
Free (v16-v21),Slope(5);
Free (v22-v26),Slope(6);
Free (v27-v31),Slope(7);
Free (v32-v35),Slope(8);
With regards to the bifactor structure, in the <Groups> section, a multidimensional model with 8 dimensions is first specified. This is the total number
of domains/factors, both general and specific. The Primary command specifies, out of the total number of dimensions, how many are primary/general
57
dimensions. As noted in the data description, the QOL scale is made of 1
general domain and 7 subdomains.
Once the dimensionality of the model is set up in the <Groups> section,
the details of which items load onto which specific factors are given in the
<Constraints> section. The first constraint serves to fix all item slopes to
zero, facilitating the subsequent freeing of parameters. The next 8 lines set
up the factor structure by 1) freeing all the item slopes on the first dimension,
which will be the general factor, and 2) assigning the items to the appropriate
subdomains, based on the item content. Note that it is absolutely necessary
to ensure the the primary dimension(s) precede the group-specific dimensions
R
when freeing items to load on the factors. flexMIRT
automatically reduces
the dimensionality of integration to 2 and consequently the analysis runs extremely fast.
The Model(v1-v35) = Nominal(7); statement in the <Groups> section
indicates that we wish to fit the Nominal Model (for 7 response categories) to
the items. In the current code, we have opted not to modify either the scoring
functions contrasts (which may accomplished via the Ta = ; statement) or the
intercept contrast matrix (which may accomplished via the Tc = ; statement).
Without changes, the default option of Trend matrices will be used for both
sets of parameters. In the <Constraints> section, the first constraint imposed
fixes the first scoring function contrast for all 35 items (the default value is
1.0), as required for identification (see Thissen et al., 2010). With the default
contrast matrices, the fixed first scoring function contrast of 1.0 ensures that
the first and last scoring function values are fixed to 0 and 6, respectively.
The remaining 5 scoring function values are reparameterized by the contrast
matrices and become estimatable parameters (as shown below).
In the <Options> section, the number of quadrature points is reduced for
time-efficiency, as was done with the other MIRT models, using the Quadrature
statement and the tolerance values are adjusted down to obtain tighter solutions.
Output 3.4: Nominal Model Scoring Function Estimates - Excerpt
Nominal Model Scoring Function Values (s) Group
Item
Label
s 1
s 2
1
v1
0.00
0.76
2
v2
0.00
0.79
3
v3
0.00
1.49
4
v4
0.00
0.77
5
v5
0.00
0.99
6
v6
0.00
0.93
58
1: Group1
s 3
s 4
1.18
1.46
1.53
2.58
1.91
2.08
1.65
2.48
1.71
2.47
2.46
3.13
s 5
3.29
4.18
3.77
4.05
3.93
4.35
s 6
4.61
4.77
4.21
4.76
4.78
5.12
s 7
6.00
6.00
6.00
6.00
6.00
6.00
7
8
9
10
11
12
13
14
15
16
v7
v8
v9
v10
v11
v12
v13
v14
v15
v16
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.79
1.50
1.53
1.41
0.88
0.11
1.45
1.55
1.32
1.52
2.27
2.30
2.54
2.59
1.60
1.06
1.71
1.68
1.09
1.71
2.96
2.92
2.77
2.86
1.89
1.35
1.63
2.59
2.25
2.65
4.25
4.27
4.18
3.93
2.98
2.51
2.71
3.54
3.22
3.96
5.03
4.96
4.85
4.53
4.35
4.15
3.86
4.36
4.16
4.99
6.00
6.00
6.00
6.00
6.00
6.00
6.00
6.00
6.00
6.00
An examination of the output excerpt presented shows that a number of
scoring function values exhibit reversal of categories from the presumed order
(such as in Items 13 and 15) or category boundaries that are very close together
(such as categories 2 and 3 in Items 16). As Preston, Reise, Cai, and Hays
(2011) explained, when the estimated scoring function values deviates from
the presumed order of 0, 1, 2, . . . , 6, the nominal model provides additional
category-specific information that other ordinal polytomous IRT models (e.g.,
partial credit or graded) do not provide. As noted earlier, the items were scored
on a 7-point scale indicating the extent to which the respondent is satisfied
with the statement of the item, where: 0=terrible; 1=unhappy; 2=mostly dissatisfied; 3=mixed, about equally satisfied and dissatisfied; 4=mostly satisfied;
5=pleased; and 6=delighted. There may be some uncertainty as to how people
interpret the order of the middle categories.
3.7.2
Bifactor Model Using the 2PL
Due to the undesirable properties when analyzing the QOL data with all 7
R
categories, we will again demonstrate how flexMIRT
may be used to recode
data. We will fit the same bifactor factor pattern to the QOL data, but
rather than utilizing all 7 categories, we will collapse the data into dichotomous
responses and fit the 2PLM.
Example 3.10: Bifactor Structure - 2PL
<Project>
Title = "QOL Data";
Description = "35 Items Bifactor Showing Recoding";
<Options>
Mode = Calibration;
Quadrature = 21,5.0;
SavePRM = Yes;
Processor = 4;
59
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
N = 586;
Dimensions = 8;
Primary = 1;
Ncats(v1-v35) = 7;
Code(v1-v35) = (0,1,2,3,4,5,6),(0,0,0,1,1,1,1);
Model(v1-v35) = Graded(2);
<Constraints>
Fix (v1-v35),Slope;
Free (v1-v35),Slope(1);
Free (v2-v5),Slope(2);
Free (v6-v9),Slope(3);
Free (v10-v15),Slope(4);
Free (v16-v21),Slope(5);
Free (v22-v26),Slope(6);
Free (v27-v31),Slope(7);
Free (v32-v35),Slope(8);
The Code command is used to collapse the 7 original categories into 2. A
point of note here is that the Ncats value provided refers to the number of
categories in the raw data file, not the number of categories that will be used for
modeling, which is specified with the Model statement. An examination of the
Code statement shows that original responses of 0,1, and 2 will be recoded into
0s and raw responses of 3-6 will become 1s after the recoding by the program.
The bifactor structure remains unchanged from the previous example, as may
be seen in the in the <Constraints> section.
60
CHAPTER 4
Simulation
Simulation studies are common in the IRT literature. Compared to other
commercially available IRT software programs (e.g., Bilog, IRTPro, Multilog,
R
and Parscale), flexMIRT
is unique in that it facilitates simulations by offering
the user the capability of generating data from all implemented IRT models,
including multidimensional, bifactor, and multilevel models. The command
syntax for generating data has the same basic structure as the model-fitting
syntax. In general, it will involve either a relatively small amount of code
in the syntax file, leaving the majority of the model settings and generating
parameter values in the parameter file. Alternatively, one can fully specify the
models and generating parameter values directly in the syntax, without using
the parameter file, or using it only as a supplement. Note that a parameter
file saved from a calibration run is completely compatible in simulations. This
provides a convenient mechanism so that the user may conduct simulations
with generating parameters found in empirical data analysis.
4.1. A Syntax-Only Example Using Value Statement
In this example, we will use the syntax file to specify simulation of a data set
consisting of 1000 simulees’ responses to 4 graded items, each scored with 3 categories. The generating parameter values are specified in the <Constraints>
section. In simulation mode, the Value statement can be used to specify
generating parameters, as is done in the following example.
61
Example 4.1: Simulation - Graded Model Syntax Only
<Project>
Title = "Simulate Data";
Description = "4 Items Graded Syntax Only";
<Options>
Mode = Simulation;
Rndseed = 7474;
<Groups>
%G%
File = "sim1a.dat";
Varnames = v1-v4;
N = 1000;
Model(v1-v4) = Graded(3);
<Constraints>
Value (v1),Slope,0.7;
Value (v2),Slope,1.0;
Value (v3),Slope,1.2;
Value (v4),Slope,1.5;
Value (v1,v2),Intercept(1),2.0;
Value (v1,v2),Intercept(2),1.0;
Value (v3,v4),Intercept(1),0;
Value (v3,v4),Intercept(2),-1.0;
R
One can see in the <Options> section, flexMIRT
is set to run in Simulation mode. When this is case, a random number seed must be supplied via
the Rndseed statement. The group specification remains the same as in calR
ibration runs. In this case, the File statement instructs flexMIRT
to save
the simulated item responses to “sim1a.dat”. The variables are named and the
total number of simulees is set to 1000. The Model statement sets the generating IRT model to the graded response model with 3 categories. Next, in the
<Constraints> section, the generating parameter values are set. For instance,
the first constraint says that the value of the slope parameter for item v1 should
be set to 0.7. The 5th constraint specifies that the first intercept term for item
v1 and v2 should be equal to 2.0. In a single group, the group name can be
omitted, so Value (v1),Slope,0.7; and Value G,(v1),Slope,0.7; would
have the same effect. For multiple group simulations, the group name must
be specified in each Value statement.
62
Output 4.1: Simulation Control Output - Graded Model Syntax Only
...
Summary of the Data and Control Parameters
Missing data code
Number of Items
Number of Cases
Number of Dimensions
Item
1
2
3
4
Categories
3
3
3
3
-9
4
1000
1
Model
Graded
Graded
Graded
Graded
Miscellaneous Control Values
Random number seed:
7474
...
Graded Items for Group 1: G
Item
a
c 1
c 2
1
0.70
2.00
1.00
2
1.00
2.00
1.00
3
1.20
0.00
-1.00
4
1.50
0.00
-1.00
Graded Items for Group 1: G
Item
a
b 1
b 2
1
0.70
-2.86
-1.43
2
1.00
-2.00
-1.00
3
1.20
-0.00
0.83
4
1.50
-0.00
0.67
...
Group Parameter Estimates:
Group
Label
mu
1
G
0.00
s2
1.00
sd
1.00
The first part of the simulation control output echoes the user-supplied
specifications, (e.g. number of items, number of simulees, the number of categories, IRT model types, and the random number seed). The generating item
parameters are then printed. One can see that the syntax statements are correctly interpreted. When everything is set up correctly, a simulated data set is
generated and saved. The first and last three rows of this data set are shown
below. The first 4 columns are the item responses (in 3 categories coded as 0,
63
1, and 2). Column 5 is a case ID column. Column 6 contains the generating
theta value for that simulee.
Output 4.2: Simulated Data Set - Graded Model Syntax Only
0 2
2 2
2 2
...
2 2
2 2
0 0
1 0 0 -0.246011
2 2 1 1.70243
0 2 2 0.71107
1 2 997 -0.0161087
2 1 998 -0.702542
1 0 999 1.09703
As noted previously, simulations may either be conducted by using Value
statements to provide generating parameter values or through the use of a
parameter file. To demonstrate this point, we will now generate data with the
R
parameters values just used, but supply them to flexMIRT
via a parameter
file.
4.2. The First Example Redone Using Parameter File
Example 4.2: Simulation - Graded Model Using PRM File
<Project>
Title = "Simulate Data";
Description = "4 Items Graded Use PRM File";
<Options>
Mode = Simulation;
Rndseed = 7474;
ReadPrmFile = "sim1b-prm.txt";
<Groups>
%G%
File = "sim1b.dat";
Varnames = v1-v4;
N = 1000;
<Constraints>
The ReadPrmFile statement is invoked to read the parameter file. The group
specification only contains the file name, the variable names, and the number
of simulees. Note that the <Constraints> section is empty because the values
of the generating parameters are set in the parameter file. The content of the
64
Table 4.1: Simulation - Graded Model PRM File Content
1
1
1
1
v2
v3
v1
v4
1
1
1
1
1
1
1
1
2
2
2
2
3
3
3
3
2
0
2
0
1 1.0
-1 1.2
1 0.7
-1 1.5
particular parameter file is shown in Table 4.1.
The parameter file has 4 rows, corresponding to the 4 items. Take row 1
as an example. The entry in the first column of 1 specifies that this is a row
containing item parameters. This is a predefined option. The other option is
0, which corresponds to group specifications. The second column specifies the
variable name. Note that the items are not ordered as v1 to v4, but there is
no problem since the items are checked against the list of variable names in
the syntax file and matched automatically. Column 3 indicates that this item
belongs to group 1. Column 4 indicates the number of factors, in this case,
the model is unidimensional. The next two columns, taken together, set the
IRT model to graded (type 2 on column 5) with 3 categories (column 6). Then
the item intercepts (from the intercept corresponding to the lowest category
to the highest category) and the item slope are specified for each item. The
resulting simulated data set is exactly the same as the previous example.
Output 4.3: Simulated Data Set - Graded Model Using PRM File
0 2
2 2
2 2
...
2 2
2 2
0 0
1 0 0 -0.246011
2 2 1 1.70243
0 2 2 0.71107
1 2 997 -0.0161087
2 1 998 -0.702542
1 0 999 1.09703
4.3. The Parameter File Layout
Although the progression of columns of the parameter file must proceed in
a prescribed fashion, the general format of the file is plain ASCII, in which
delimiters may be spaces or tabs. Column headers are not permitted, which
is why the order of values in the columns must follow the rigid structure.
65
For arranging the values in the parameter file, we have found it easiest to
use a spreadsheet program (e.g., Excel) to initially set up the entries and,
when complete, export/paste the values into the flat ASCII text editor. As
discussed earlier, there are set options from which the user must select for
certain columns. Table 4.2 provides a summary of these options.
Table 4.2: Key Codes for Columns With Set Options
Column 1
(Entry Type)
Value
Meaning
0
Group
1
Item
Column 5 - Items
(IRT Model)
Value
Meaning
1
3PL
2
Graded
3
Nominal
Column 5 - Groups
(Prior Type)
Value
Meaning
0
Normal Prior
1
EH Prior
Select Columns
(Nominal Contrast Type)
Value
Meaning
0
Trend
1
Identity
2
User-specified
As indicated in Table 4.2, the first column of the parameter file indicates
whether the information to be supplied across the row refers to an item or a
group. The information that goes into the various subsequent columns differs
somewhat, depending on whether item or group is indicated in column 1, so
the arrangement for these two parts of the parameter file will be discussed
separately.
4.3.1
Column Progression for Groups
R
With a 0 in the first column of the parameter file, flexMIRT
will interpret all
subsequent information on the line as applicable to a group. The 2nd column is
used to supply a group label, which should correspond to the label provided in
the syntax file. The third column indicates to which group this group belongs,
(i.e., 1 if it belongs to the first group, 2 if it belongs to the second group,
etc). Column 4 requires the user to specify the number of latent variables
in the generating model (within the group just specified). The 5th column is
reserved for indicating the type of population distribution (prior) to be applied
to the latent trait. The acceptable values are 0 for normal and 1 for empirical
histogram (see Table 4.2). If a normal prior is chosen, the subsequent columns
will supply the latent factor mean(s) and then (co)variance(s).
The latent factor mean values are entered in order, with the mean of factor
1 listed first, followed by mean 2, mean 3 and so on. For specifying (co)variance
values, the unique elements of the latent factor variance/covariance matrix are
entered in row-major order. To demonstrate this point, suppose the following
R
4 × 4 factor covariance matrix is to be submitted to flexMIRT
:
66
f1
f2
f3
f4
f1
1
2
4
7
f2
2
3
5
8
f3
4
5
6
9
f4
7
8
9
10
The progression of variance/covariance values entered in the parameter file
should be
1 2 3 4 5 6 7 8 9 10
If an empirical histogram prior is chosen, the user must supply additional
information. Column 6 specifies the total number of quadrature points on
which the histogram is defined and Column 7 requires a maximum value for the
quadrature points, which determines the range over which the points are to be
distributed. Similar to specifying Quadrature or FisherInf in the <Options>
section, the quadrature points are spread symmetrically around zero, with the
minimum and maximum value determined by the entry in Column 7. After
the EH specification, the mean(s) and unique element(s) in the covariance
matrix are listed in the subsequent columns, following the same progression of
entries just covered. The histogram itself is supplied next. For instance, if the
entry in Column 6 is 101 and Column 7 is 5, then 101 points will be evenly
spaced over the range of -5 to 5. The user must then supply as the last 101
entries of this particular row, the 101 ordinates of the empirical histogram,
(i.e., the heights of the histogram at the 101 quadrature points). An example
of simulating data with EH priors will be presented later.
4.3.2
Column Progression for Items
The first four columns of a row supplying item level information have the same
generic content as groups. First, a value indicating item information will follow
(i.e., 1) is given in Column 1, then a variable name is assigned in Column 2.
This name must correspond to a variable in the Varnames statement in the
command file. Column 3 assigns the item to a specific group, and Column four
requires the entry of the number of latent factors in the generating model. At
this point (Column 5), the content of columns diverges from what was covered
in the groups subsection. The generating IRT model for the item is set in
Column 5, with the possible models coded as indicated in Table 4.2. For
instance, an item to be generated using Graded model parameters would have
a 2 in the 5th column. The subsequent column supplies the number of possible
response categories for the item.
67
Once an IRT model is chosen, the item parameters must be entered, one
value per column. The progression of expected parameters varies slightly depending on the which IRT model was selected in Column 5. If the 3PL model
was selected, the guessing parameter (in logit metric) is given in Column 7,
followed by the intercept parameter, and then all discrimination/slope parameter values (if there is more than one factor). If the Graded Model is selected
in Column 5, the expected order for parameter values is the ordered intercept
parameters (starting at the lowest in Column 7 and going up) followed by the
discrimination parameter(s).
With respect to the Nominal model, the parameters are entered in the
reparameterized form detailed in Thissen et al. (2010). The type of scoring
contrast is given in Column 7, using one of the code values given in Table 4.2.
If Trend (0) or Identity (1) contrasts are specified in Column 7, the scoring
function contrast values (of which there are the number of categories minus
one) must be supplied starting from Column 8. After those m − 1 values
are given, the slope parameter(s) are entered. Immediately after the slope
parameter, the next column (the column number of which will vary depending
on number of categories and slopes used) is used to indicate the contrast
type for the intercept parameters. Again, if Trend or Identity contrasts are
selected, the columns immediately following the contrast selection contain the
appropriate number of intercept parameter values (again always equal to the
number of categories minus one).
In the rare occasion that user-supplied contrasts are selected for either
the scoring functions or the intercept parameters, it is necessary to supply
the contrast coefficient values as well as the contrasts. In other words, the
built-in contrasts are automatically computed, but any user-defined contrasts
need to be entered into the program. The contrast coefficients are given after
the contrast selection column (e.g., Column 7) but before the actual contrast
values. The contrast coefficients of a user-supplied contrast matrix are expected to be in row major order. That is, for a contrast matrix applied to
nominal items with k categories, the coefficients from that k × (k − 1) matrix
would be entered in the following order: T (1, 1), T (1, 2), T (1, 3), . . . , T (1, k −
1), T (2, 1), . . . , T (2, k − 1), . . . , T (k, k − 1).
4.4. A 3PL Model Simulation Example
In this example, we will simulate item responses to 10 3PL items for a single
group of 500 simulees. The latent trait distribution is standard normal. The
68
syntax file is presented first.
Example 4.3: Simulation - 3PL Model Normal Population
<Project>
Title = "Fit Model";
Description = "10 Items 3PL N=500 Normal Priors";
<Options>
Mode = Simulation;
Progress = No;
ReadPRMFile = "genparams.txt";
RndSeed = 1;
<Groups>
%Group1%
File = "sim1.dat";
Varnames = v1-v10;
N = 500;
<Constraints>
As always, the command file begins with the <Project> section where a title and general description are provided. In the <Options> section we have set
the engine mode to Simulation and, as required when Mode = Simulation;
provided a seed value for the random number generator which will be used
to create the data. Because we are not using Value statements, we have also
provided a file name where item and group parameters may be found.
In the <Groups> section, we label our group and use the File statement to
provide a name for file where the data will be stored once generated. We want
to simulate data for 10 variables so we provide labels for 10 variable names
with the Varnames statement and finally, using N, specify the number of cases
we which to have generated. The variable and group labels provided here must
match those provided in the second column of the parameter file. As may be
seen here, and as mentioned previously, the syntax command file is only used
to set up the very basic structure of the analysis and the specifics are left to
be input through the parameter file, to which we will now turn our attention.
As covered in Section 4.1, the actual parameter file to be submitted to
R
flexMIRT
(Ex1genparms.txt, in this case) may not contain row headers but
they are included in Table 4.3 for illustrative purposes only. In this parameter
file, we have arranged all the item specifications first and the group parameters
are in the bottom row. This is completely arbitrary and the rows of the file
69
may be arranged in any way the user wishes. However, the columns must
follow the arrangement previously detailed. Reading across for the first item,
the 1 in the first column denotes that the values to follow refer to an item. We
then provide a variable name, v1, and assign this item to the first group (which
is the only group in this example). The next column indicates that there is 1
latent factor in the model and the subsequent column (Column 5) says that we
will be providing parameters for the 3PL model (indicated by the 1). The 6th
R
column value tells flexMIRT
that the item will have two response options
(as per the 3PL model). Based on the number of factors, the IRT model, and
the number of response categories, we now supply the appropriate number of
item parameters, in this case 3 parameters (1 for the lower asymptote, 1 for
the intercept, and 1 for the slope). The order of parameters values must be
guessing (in the logit metric), intercept (in the c = −a ∗ b metric, where b is
the difficulty/threshold), and then the slope parameter (a).
Table 4.3: Simulation - 3PL Parameter File with Labeled Columns
Type
1
1
1
1
1
1
1
1
1
1
Type
0
Label
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
Label
Group1
Group #
1
1
1
1
1
1
1
1
1
1
Group #
1
# Factors
1
1
1
1
1
1
1
1
1
1
# Factors
1
Model
1
1
1
1
1
1
1
1
1
1
Prior Type
0
# Cat
2
2
2
2
2
2
2
2
2
2
Mean
0.00
logit g
-1.34
-0.97
-1.64
-2.10
-1.59
-1.72
-1.13
-1.71
-2.79
-1.66
Var
1.00
c
-0.69
-1.30
0.42
0.40
-1.35
-0.56
-0.26
-1.56
0.73
-0.14
a
1.30
1.41
1.05
1.26
2.43
1.67
3.12
2.09
1.17
1.10
The final row of the file is used to describe the group. The first 0 indicates
that the row is describing a group. The second column provides the group a
label, Group1, which should correspond to the label provided in the command
file. We assign the Group1 label to the first group by the 1 in the 3rd column,
and specify that a unidimensional model will be generated, by indicating 1
factor in the 4th column. As noted in Table 4.2, the fifth column of a Group
line indicates the type of prior distribution to be applied (in this case a normal
prior, signified by the 0). After a normal prior is indicated, a group mean of
0.00 and a variance of 1.00 are supplied.
70
4.5. Simulating Data with Non-Normal Theta Distributions
In this example, we will cover data simulated from models with non-normal
population distribution. This is accomplished using the empirical histograms
method (previously used in Calibration Example 3.2). The empirical histogram prior may only be used for unidimensional or bifactor models (testlet
included), as previously noted in the discussion of Example 3.2.
Example 4.4: Simulation - Empirical Histogram Prior
<Project>
Title = "Fit Model";
Description = "10 Items 3PL N=500 Non-Normal Priors";
<Options>
Mode = Simulation;
ReadPRMFile = "EHgenparams.txt";
RndSeed = 1;
<Groups>
%Group1%
File = "sim2.dat";
Varnames = v1-v10;
N = 500;
<Constraints>
The command file for this example is substantively the same as the previous, although we have changed the parameter file name. The added feature
here is that a non-normal distribution will be specified in the generating parameter file (EHgenparams.txt). This change only effects that values entered
into the Group row of the parameter file.
An examination of the generating parameter file will find that the 5th
column of the Group row been changed from the value of 0, seen in Table
4.3, to a value of 1. As noted in Table 4.2, this indicates that an empirical
histogram prior will be used. The 6th column then provides the number of
points to be used in constructing the histogram (121) and the next value
indicates the range over which the points are spread (from -6 to 6 here, as
indicated by the value of 6 in the 7th column of the Group row). Following
the specification of the empirical histogram parameters, the latent variable
mean(s) and covariances are listed. After all means and covariances are listed,
71
the heights of the empirical histogram are provided. For our current example,
we need to provide 121 values, per the entry in Column 6. Figure 4.1 provides
a graphical display of the empirical histogram provided in the parameter file,
to demonstrate the extent of non-normality.
Figure 4.1: Empirical Histogram From EHgenparams.txt
4.6. Simulating Bifactor Model Data
We will be using item parameters obtained from our bifactor analysis in Example 3.10 to simulate items with a bifactor structure from the 2PL model.
As noted previously, when not employing Value statements, only the very basic structure of the analysis is supplied in the simulation syntax file and the
details of the model are left to the parameter file. Table 4.4 presents a labeled
version of the parameter file (QOLbifac-prm.txt) for illustrative purposes.
Example 4.5: Simulation - Bifactor Structure From Existing Parameter File
<Project>
Title = "QOL Data";
Description = "Simulation From PRM File 35 Items Bifactor";
72
<Options>
Mode = Simulation;
ReadPRMFile = "QOLbifac-prm.txt";
Rndseed = 7474;
<Groups>
%Group1%
File = "QOLbifacsim.DAT";
Varnames = v1-v35;
Dimensions = 8;
Primary = 1;
N = 1000;
<Constraints>
Table 4.4: Labeled Parameter File for Bifactor Model Simulation
Type
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
..
.
1
1
Type
0
Label
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
v11
v12
v13
v14
v15
..
.
v34
v35
Label
Group1
Grp#
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
..
.
1
1
Grp#
1
#Fac
8
8
8
8
8
8
8
8
8
8
8
8
8
8
8
..
.
8
8
#Fac
8
Model
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
..
.
2
2
Prior
0
#Cat
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
..
.
2
2
µ1
0.00
c
2.13
3.78
1.69
3.76
2.87
0.86
1.38
0.49
0.30
3.09
2.57
1.88
2.30
5.16
2.09
..
.
2.95
3.22
···
···
a1
2.26
2.31
1.62
3.86
2.77
2.77
1.38
2.60
1.78
1.91
1.42
1.40
1.57
4.02
2.49
..
.
1.67
1.40
µ8
0.00
a2
0.00
2.51
1.42
4.89
3.08
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
..
.
0.00
0.00
σ11
1.00
a3
0.00
0.00
0.00
0.00
0.00
3.20
1.50
2.86
1.97
0.00
0.00
0.00
0.00
0.00
0.00
..
.
0.00
0.00
σ21
0.00
a4
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
1.74
0.71
0.72
0.13
3.80
0.66
..
.
0.00
0.00
σ22
1.00
a5
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
..
.
0.00
0.00
σ31
0.00
a6
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
..
.
0.00
0.00
σ32
0.00
Following the parameter file format, the first column is for entering whether
the values to follow on the row refer to an item or a group, with 1 denoting
items and 0 denoting groups. We will now follow the first line across, discussing
the columns in turn. The next column supplies a variable name for the line
and the following column specifies to which group the item belongs. In this
case, there is only one group, so the entire column consists of 1s. In Column
4, we specify that there are 8 factors in the model and Columns 5 and 6,
R
respectively, tell flexMIRT
that the IRT model is the graded model (type
73
a7
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
..
.
0.00
0.00
σ33
1.00
a8
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
..
.
1.30
0.44
···
···
2) with two possible response categories. The combination of the these three
R
values informs flexMIRT
to expect 9 item parameters, one intercept and 8
slopes. Scanning down the columns of slope parameters, it becomes clear that
the a1 column is for the general factor, with all item loading onto it, and the
next 7 columns represent the group-specific factors, onto which only a subset
of items load.
Turning now to the group specification in the bottom row of the table, the
first column signifies that this line is supplying group information, due the
value of 0. Labeling, assigning the group number and the number of factors
in the model, all similar to what was done in the item rows, come next. In
Column 5, we enter information regarding the type of population distribution
prior. We have chosen Normal priors (via the option 0). Because we have
chosen Normal priors, all that is left to specify are the latent variable means
for each of the eight factors, followed by the unique elements of the covariance
matrix.
4.7. Simulating Group Differences
As seen in the multiple group calibration examples, each group has its own
self-contained subsection within the <Groups> section. While this can lead
to some redundancy in the naming of variables, etc., it does permit greater
R
flexibility. For example, flexMIRT
does not require that the number of items,
the number of categories, the kind of IRT models employed, or the number
of latent dimensions to remain the same across groups. One may specify one
group to have a normal prior distribution and use empirical histograms in the
other group. In particular, one may even set up a multilevel model for one
group and a single level model for the other.
The current example simulates group differences in both population means
and variances as well as item parameters. Each group has 10 items. Group
1 has mean 0, variance 1. Group 2 has the same variance but a mean at 0.2.
Group 3 has a mean of -0.2 and variance of 1.5. The items are for the most
part 3PL items, with the exception of Item 1 in Group 3. The lower asymptote
parameter is set to 0 for that item, making it a 2PL item. Also, the intercept
parameter for Item 9 in Group 2 is chosen such that its threshold value is 0.5
standard deviation higher than the corresponding items in Groups 1 and 3.
74
Example 4.6: Simulation - DIF and Group Mean Differences
<Project>
Title = "Simulate Data";
Description = "3 Groups 10 Items N=1000 (mostly) 3PLs
Group 1 N( 0.0,1.0), Group 2 N( 0.2,1.0), Group 3 N(-0.2,1.5)
Item 9 in Group 2 has 0.5 higher threshold than the corresponding
item in Groups 1 and 3.
Item 1 in Group 3 is actually a 2PL item.";
<Options>
Mode = Simulation;
ReadPRMFile = "genparams.txt";
RndSeed = 7474;
<Groups>
%Group1%
File = "group1.dat";
Varnames = v1-v10;
N = 1000;
%Group2%
File = "group2.dat";
Varnames = v1-v10;
N = 1000;
%Group3%
File = "group3.dat";
Varnames = v1-v10;
N = 1000;
<Constraints>
The <Options> section remains substantially unchanged from previous examples. Within the <Groups> section, we now have labels and specifications
for the three distinct groups for which we wish to generate data. Following
the other examples, the details of the simulation are set up in the parameter
file (genparams.txt). The contents of the parameter file (presented, in part, in
Table 4.5) are now discussed.
As before, the first column still indicates whether the row refers to an item
or a group, and the second column provides a label. The first indication that
this is a multiple group problem appears in Column 3, which is the column
designated for assigning the row (item or group) to a specific group defined in
the syntax. In previous examples, with only one group, all rows displayed a 1
75
Table 4.5: Simulating DIF and Group Mean Differences - Parameter Values
1
1
..
.
v1
v2
..
.
1
1
..
.
1
1
..
.
1
1
..
.
2
2
..
.
-1.33972
-0.96814
..
.
-0.68782
-1.29641
..
.
1.301362
1.409536
..
.
1
1
1
1
..
.
v9
v10
v1
v2
..
.
1
1
2
2
..
.
1
1
1
1
..
.
1
1
1
1
..
.
2
2
2
2
..
.
-2.79306
-1.65629
-1.33972
-0.96814
..
.
0.730972
-0.14263
-0.68782
-1.29641
..
.
1.172468
1.102673
1.301362
1.409536
..
.
1
1
1
1
..
.
v9
v10
v1
v2
..
.
2
2
3
3
..
.
1
1
1
1
..
.
1
1
2
1
..
.
2
2
2
2
..
.
-2.79306
-1.65629
-0.68782
-0.96814
..
.
0.144738
-0.14263
1.301362
-1.29641
..
.
1.172468
1.102673
1
1
0
0
0
v9
v10
Group1
Group2
Group3
3
3
1
2
3
1
1
1
1
1
1
1
0
0
0
2
2
0
0.2
-0.2
-2.79306
-1.65629
1
1
1.5
0.730972
-0.14263
1.172468
1.102673
1.409536
..
.
in this column, where now the first 10 items (4 of which are displayed in the
table) have 1s, the next 10 rows have 2s and so on.
With respect to the differences across the groups, as noted earlier, Item 1 in
group 3 is a 2PL item, so it is specified as a graded model with two response
categories, while item 1 in both Groups 1 and 2 are 3PL items. DIF also
occurs for Item 9. The same generating IRT model is used and the guessing
and discrimination parameter values are identical across groups, but in Group
2, the difficulty parameter value is 0.5 higher (in the b metric) than the value in
Groups 1 and 3. Mean differences among the groups are also generated. The
last 3 rows refer to Groups 1 through 3, respectively. After labeling, specifying
group membership, and declaring the number of factors, normal population
distributions are chosen for all groups (via the 0s in the 5 column). The
sixth column provides the population means and the seventh column provides
population variances.
76
4.8. Multilevel Bifactor Model Simulation
Our final example will illustrate how one may simulate item responses according to a multilevel bifactor model in two groups. We will generate data for 12
dichotomously scored graded items.
Example 4.7: Simulation - Multilevel Model with Bifactor Structure
<Project>
Title = "Two-level Bifactor Model (Two Groups)";
Description = "12 Items, 4 Group Factors,
100 L2 units, 10 respondents within each";
<Options>
Mode = Simulation;
Rndseed = 7471;
ReadPrmFile = "MGsimL2-prm.txt";
<Groups>
%G1%
File = "simL2.g1.dat";
Varnames = v1-v12;
Dimensions = 6;
Between = 1;
N = 1000;
Nlevel2 = 100;
%G2%
File = "simL2.g2.dat";
Varnames = v1-v12;
Dimensions = 6;
Between = 1;
N = 1000;
Nlevel2 = 100;
<Constraints>
Skipping down to the <Groups> section and focusing on the first group’s
subsection, we have provided the data file name, variable labels, and the total
number of dimensions of the generating model with the Dimensions state77
ment. The command Between= 1; indicates that the model has one level-2
(between) latent variable and 5 level-1 (within) latent variables. The Nlevel2
command is required when simulating multilevel data and determines the number of level-2 units (e.g., schools in which students are nested). In this case,
the number of level-2 units is equal to 100. In conjunction with the total
number of level-1 cases specified in the N=1000 line, one can deduce that the
R
will
number of level-1 units within each level-2 unit is equal to 10. flexMIRT
distribute the total sample size as evenly as possible across the level-2 units.
If the sample size value is not a multiple of the number of level-2 units, the
“remainder” observations will be added to the final level-2 unit. Note that
while the specifications for the second group are the same as those in the first
group in this example, it is by no means a requirement. As mentioned earlier,
the number of items, type of IRT model, and the latent dimensionality can be
different across groups.
Output 4.4: Simulation Control Output for Multilevel Bifactor Model
Two-level Bifactor Model (Two Groups)
12 Items, 4 Group Factors, 100 L2 units, 10 respondents within each
Summary of the Data and Control Parameters
Group
G1
G2
Missing data code
-9
-9
Number of Items
12
12
Number of L-2 Units
100
100
Number of L-1 Units
1000
1000
Number of Dimensions
6
6
Between Dimensions
1
1
Within Dimensions
5
5
...
Miscellaneous Control Values
Random number seed:
7471
...
2PL Items for Group 1: G1
Item
a 1
a 2
a 3
a 4
1
2.00
2.00
1.50
0.00
2
1.50
1.50
1.50
0.00
3
1.00
1.00
1.20
0.00
4
2.00
2.00
0.00
1.20
5
1.50
1.50
0.00
1.50
6
1.00
1.00
0.00
1.80
7
1.50
1.50
0.00
0.00
8
1.50
1.50
0.00
0.00
9
1.00
1.00
0.00
0.00
10
2.00
2.00
0.00
0.00
11
2.00
2.00
0.00
0.00
12
1.00
1.00
0.00
0.00
...
Group Latent Variable Means:
a 5
0.00
0.00
0.00
0.00
0.00
0.00
1.20
1.50
1.50
0.00
0.00
0.00
78
a 6
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
1.80
2.00
1.50
c
-1.50
-0.50
0.50
1.50
-1.50
-0.50
0.50
1.50
-1.50
-0.50
0.50
1.50
Group
Label
mu 1
mu 2
mu 3
mu 4
1
G1
0.00
0.00
0.00
0.00
2
G2
0.20
0.00
0.00
0.00
Latent Variable Variance-Covariance Matrix for Group1: G1
Theta 1 Theta 2 Theta 3 Theta 4 Theta 5 Theta 6
0.50
0.00
1.00
0.00
0.00
1.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
0.00
1.00
Latent Variable Variance-Covariance Matrix for Group2: G2
Theta 1 Theta 2 Theta 3 Theta 4 Theta 5 Theta 6
0.25
0.00
1.00
0.00
0.00
1.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
0.00
1.00
mu 5
0.00
0.00
mu 6
0.00
0.00
The simulation control output shows that the input syntax and the parameter values are specified correctly. The item parameters are invariant across
groups so only parameters from Group 1 are shown. The within-item equality
of the slopes implies that the 5 level-1 dimensions (a2 - a6) follow a testlet
response theory model (restricted bifactor) with 4 testlets of 3 items each.
Furthermore, the equality of slopes on the Within general dimension (a2) and
the Between general dimension (a1) implies that a random-intercept model is
set up that provides a decomposition of the Between and Within variances for
the general trait. The mean of Group 1 and 0.20 lower than Group 2, but the
Group 1 Between variance is 0.50 and the Group 2 between variance is 0.25.
79
4.9. Internal Simulation Function
Here we provide an overview for using the internal simulation function (ISF),
which allows users to automate numerous replications of a given simulation and
subsequent calibration. This function uses existing simulation and calibration
syntax files and merely provides a convenient way of obtaining replications; it
is not an alternative method for constructing simulation syntax files.
4.9.1
Accessing the Internal Simulation Function
R
Currently, the ISF is available through the flexMIRT
GUI only; specifically,
the ISF is not available if using the program through the command prompt
or when it is being called by another program, such as R. The ISF is found
R
under the “flexMIRT” menu bar option of the flexMIRT
GUI.
Figure 4.2: Accessing the ISF through the GUI
Once “Simulation” is selected, a pop-up will appear where the user enters
the information for the simulation study to be conducted. As can be seen in
the next image, the user specifies the desired number of replications, the name
and location of the existing simulation syntax file which will be responsible for
the data creation, and the location(s) and name(s) of the calibration syntax
for the model(s) that the users wishes to fit to the simulated data.
Currently, there is no way to save the simulated data but in an upcoming
R
update, flexMIRT
will output the simulated data sets to external files by
default. The simulated datasets will be output to the file name specified
in the simulation syntax file but will have the replication number added to
the output file name. For instance, if we are outputting the data to the file
“SIM.dat" via the File = "SIM.dat"; command in the simulation syntax, the
80
individual replications will be saved to files “’SIM-0.dat," “’SIM-1.dat," etc..
R
However, in the calibration syntax, to direct flexMIRT
to the datafile, we
will still use the statement File = SIM.dat; - without any mention of the
replication number.
When all the syntax files have been identified and a name for the output
file is specified in the “Save Simulation Output to:" blank, the “Run" button
will be enabled.
Figure 4.3: Completed ISF pop-up window
R
Once the simultion has been submitted, flexMIRT
will begin the process
of simulating data and fitting the specified model(s) to the data. The progress
of each replication is printed in the engine viewing pane, as with any oneoff simulation or calibration. When the specified number of replications have
completed, a single output file is created. The format of the ISF output file
is optimized to be read into some other software program (e.g., SAS) and as
R
output. The output
such, appears quite different from the normal flexMIRT
file has no labeled sections or column headers, etc.. Due to this, the layout of
the ISF output file will be described in detail.
81
4.9.2
The Simulation Output File
The first 6 columns in the output for every model summary will contain the
same information. Column 1 has the simulation replication number. Column
2 prints what is called the return code and is an indicator of successful replications. If the return code is 0, the replication completed successfully; if it
takes any value other than 0, a problem was encountered. Column 3 reports
the log likelihood, Column 4 gives the number of estimation cycles that were
completed to fit the model, the 5th column reports the time the calibration
run took, and the 6th column give the inverse of the condition number. The
inverse condition number is equal to the ratio of the smallest to largest eigenvalue of the asymptotic covariance matrix of the item parameters; if it is very
small, it shows instability of the obtained solution.
Output 4.5: Beginning Columns of Output from 5 Replications
1
2
3
4
5
0
0
0
0
0
-4.42522e+003
-4.44206e+003
-4.28512e+003
-4.33888e+003
-4.37110e+003
46
40
41
31
29
1.41000e-001
1.25000e-001
1.25000e-001
1.09000e-001
1.09000e-001
8.61952e-003
1.52923e-002
8.18986e-003
1.63897e-002
1.82420e-002
2.87409e+000
3.00412e+000
2.78522e+000
2.74887e+000
2.58602e+000
1.99261e+000
1.88355e+000
1.87415e+000
1.86261e+000
1.84299e+000
7.62795e-001
7.38170e-001
7.44810e-001
6.77367e-001
8.12819e-001
At the seventh column, reporting of the item and latent variables parameters begins. The parameters are reported in the same order as they are
numbered in a corresponding one-off calibration run output. It is strongly recommended that a single run of the calibration syntax file be conducted prior
to employing the ISF, to both ensure the syntax works properly and to obtain
the ordering of parameters. For example, below is the item parameter portion
R
of the flexMIRT
output for a single run fitting the model found in the file
“Fit1.flexmirt”.
Output 4.6: Item Parameters with Parameter Number Labels
Graded Items for Group 1: G1
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
P#
3
6
9
12
15
18
a
0.81
1.14
1.41
1.25
2.13
0.54
s.e.
0.13
0.15
0.15
0.13
0.28
0.09
P#
1
4
7
10
13
16
c 1
2.59
2.69
2.92
2.88
2.76
2.92
s.e.
0.14
0.16
0.18
0.16
0.24
0.14
P#
2
5
8
11
14
17
c 2
1.84
2.01
0.20
0.13
1.70
0.20
s.e.
0.11
0.13
0.09
0.08
0.18
0.07
The order in which the parameters will be printed in the ISF output cor82
responds to the values list in the columns labeled “P#” in the output. From
this output excerpt, we can see that in the ISF output file, the two intercept
parameters and then the slope parameter for Item 1 will be printed first, as
indicated by the 1, 2, and 3, respectively, in the “’P#” column. These values are followed by the intercept parameters and slope for Item 2, and so on.
Once the item parameters have been reported, the latent variable mean(s) and
covariance(s) are printed. After all point estimates are given, the respective
standard error for each parameter is printed, in the same order as the point
estimates.
The next six columns of the ISF output, following all parameter point
R
estimates and SEs, are fixed. flexMIRT
will report, in order, the G2 , X 2 ,
and M2 overall fit statistics. It should be noted here that the M2 column will
always print, regardless of whether the statistic was requested or not. If the
statistic was not requested, via the GOF and M2 statements in the <Options>
section, placeholder values of 0.0000 will be printed and are not indicative of
the actual value of the M2 statistic, had it been requested.
The overall fit statistics are followed by three members of the powerdivergence family of fit statistics (e.g., Cressie & Read, 1984) for assessing
latent distribution fit, specifically the summed score likelihood based G2 , D2 ,
and X 2 . These indices have been found to accurately detect departures from
normality within the latent distribution while appropriately “ignoring” other
forms of model misspecification such as multidimensionality (Li & Cai, 2012).
Output 4.7: ISF output - Switch from Model 1 to Model 2 Values
...
...
...
...
...
1.29309e-314
9.70097e-315
2.32371e-315
1.04469e-314
1.02812e-314
1.47651e-317
1.47651e-317
1.47651e-317
1.47651e-317
1.47651e-317
|
|
|
|
|
1
2
3
4
5
0
0
0
0
0
-4.52136E+03
-4.40664E+03
-4.44256E+03
-4.45391E+03
-4.58027E+03
31
24
24
21
22
...
...
...
...
...
If the ISF was instructed to fit multiple models to the data, then the values
for next model will follow, on the same line, after the latent distribution fit
statistics of the current model. The new model reporting begins with the
replication number and follows the same general order of reporting detailed
previously. If we fit two models to a given set of simulated data, the output
columns at the point where the reporting for the first model ends and the
second model begins will look something like the values presented in Output
83
4.7 (although we have added the break line for emphasis). Note, however, that
the numbering of parameters may change from model to model. As stated
before, to determine the reporting order of parameters, a single run of any
model being fit is strongly recommended before using the ISF.
84
CHAPTER 5
Metropolis-Hastings Robbins-Monro Estimation for
High Dimensional Models
Although Bock and Aitkin’s (1981) EM algorithm made IRT parameter estimation practical, the method does have shortcomings with the most noticeable
one being its limited ability to generalize to truly high-dimensional models.
This is due primarily to the need to evaluate high-dimensional integrals in the
likelihood function for the item parameters. As the number of dimensions of
a model increases linearly, the number of quadrature points increases exponentially, making EM estimation unwieldy and computationally expensive for
models with more than a handful of latent factors.
A Metropolis-Hastings Robbins-Monro (MH-RM; Cai, 2010b, 2010c) alR
gorithm has been implemented in flexMIRT
to allow for the estimation of
higher dimensional models. We will briefly discuss the algorithm to acquaint
those users less familiar with this recently introduced method, cover the availR
able flexMIRT
options and syntax commands that are specific to this estimation routine, and then provide working examples.
5.1. Overview of MH-RM Estimation
In brief, the MH-RM algorithm is a data augmented Robbins-Monro type (RM;
Robbins & Monro, 1951) stochastic approximation (SA) algorithm driven by
the random imputations produced by a Metropolis-Hastings sampler (MH;
Hastings, 1970; Metropolis, Rosenbluth, Rosenbluth, Teller, & Teller, 1953),
which is a sampling method based on the principles of Markov chain Monte
Carlo (MCMC). The MH-RM algorithm is motivated by Tittering’s (1984)
recursive algorithm for incomplete data estimation, and is a close relative of Gu
and Kong’s (1998) SA algorithm. It can also be conceived of as an extension of
the Stochastic Approximation EM algorithm (SAEM; Celeux & Diebolt, 1991;
Celeux, Chauveau, & Diebolt, 1995; Delyon, Lavielle, & Moulines, 1999) when
85
linear sufficient statistics do not naturally exist in the complete data model.
R
Within the flexMIRT
implementation, the MH-RM iterations use initial values found through an unweighted least squares factor extraction stage
(Stage I) and further improved upon during a supplemented EM-like stage
(Stage II) - both Stage I and Stage II use a fixed number of cycles. Stage
III is home to the MH-RM algorithm and where the primary estimation of
parameters occurs. Cycles in Stage III terminate when the estimates stabilize.
Cycle j + 1 of the MH-RM algorithm for multidimensional IRT consists of
three steps:
• Imputation. In the first steps, conditional on provisional item and latent density parameter estimates β(j ) from the previous cycle, random
samples of the individual latent traits θ(j +1) are imputed using the MH
sampler from a Markov chain having the posterior of the individual latent
traits π (θ|Y, β(j ) ) as the unique invariant distribution.
• Approximation. In the second step, based on the imputed data, the
complete data log-likelihood and its derivatives are evaluated so that
the ascent directions for the item and latent density parameters can be
determined later.
• Robbins-Monro Update. In the third step, RM stochastic approximation
filters are applied when updating the estimates of item and latent density parameters. First, the RM filter will be applied to obtain a recursive
stochastic approximation of the conditional expectation of the complete
data information matrix. Next, we use the RM filter again when updating the new parameter estimates. Cai (2010c) showed that the sequence
of parameters converges with probability 1 to a local maximum of the
likelihood of the parameter estimates, given the observed data.
For more detailed descriptions of the estimation method, see Cai (2010b,
2010c).
86
5.2. MH-RM Specific Syntax Options
Example 5.1: <Options> - Engine Mode Settings to Call MH-RM Estimation
and MH-RM Specific Options
<Options >
Mode = Calibration;
Algorithm = MHRM;
RndSeed = ?;
Imputations = 1;
Thinning = 10;
Burnin = 10;
Stage1 = 200;
Stage2 = 100;
Stage3 = 2000;
Stage4 = 0;
InitGain = 1.0;
Alpha = 1.0;
Epsilon = 1.0;
WindowSize = 3;
ProposalStd = 1.0;
ProposalStd2 = 1.0;
MCsize = 25000;
To call the MH-RM algorithm, the combination of Mode = Calibration;,
Algorithm = MHRM;, and a random number seed via RndSeed = ; is required.
All other listed commands are optional, with the default values shown above.
Imputations controls the number of imputations from the MH step per
RM cycle. A larger value will lead to smoother behavior from the algorithm.
In most cases, a value of 1 should be sufficient for obtaining accurate point
estimations, however this number may be increased for better standard errors.
Thinning refers to the MCMC practice of retaining only every kth draw
from a chain. Thinning is used, in part, to reduce the possible autocorrelation
that may exist between adjacent draws. The Thinning statement sets the
interval for the MH sampler - meaning if Thinning = 10;, every 10th draw
by the sampler will be retained.
The Burnin statement controls the number of draws that are discarded
from the start of the MH sampler. These draws are discarded to avoid using
87
values that were sampled before the chain had fully converged to the target
R
distribution. Burnin = 10; tells flexMIRT
to discard the first 10 values
obtained by the MH sampler.
Stage1 determines the number of Stage I (constant gain) cycles. The
Stage I iterations are used to improve default or user-supplied starting values
for the estimation that occurs in Stage II. Stage2 specifies the number of
Stage II (Stochastic EM, constant gain) cycles. The Stage II iterations are
used to further improve starting values for the MH-RM estimation that occurs
in Stage III. Stage3 sets the maximum number of allowed MH-RM cycles to
be performed.
Stage4 determines the method by which SEs are found. If Stage4 = 0;
then SEs will be approximated recursively (see Cai, 2010b). If a non-zero
value is given then the Louis formula (Louis, 1982) is used directly. If the
Louis formula is to be used, the supplied value determines the number of
iterations of the SE estimation routine; this number will typically need to be
large (e.g., 1000 or more).
InitGain is the gain constant for Stage I and Stage II. If the algorithm
is initially taking steps that are too large this value may be reduced from the
default to force smaller steps.
Alpha and Epsilon are both Stage III decay speed tuning constants. Alpha
is the first tuning constant and is analogous to the InitGain used in Stages
I and II. Epsilon controls the rate at which the sequence of gain constants
converge to zero. Specified Epsilon values must be in the range (0.5, 1], with
a value closer to 1 indicating a faster rate of convergence to zero. WindowSize
allows the user to set the convergence monitor window size. Convergence of
the MH-RM algorithm is monitored by calculating a window of successive differences in parameter estimates, with the iterations terminating only when all
differences in the window are less than 0.0001. The default value of the window size is set at 3 to prevent premature stoppage due to random variation.
ProposalStd and ProposalStd2 control the dispersion of the Metropolis
proposal densities for the first and second level of the specified model, respectively. If a single level model is specified, only the ProposalStd command
will be used. Although default values of 1.0 have been set for both
of these commands, these values need to be adjusted on a case-bycase basis. The values used will depend on the complexity of the model, the
number of items, the type of items (e.g., dichotmous, polytomous), and the
model fit to the items (e.g., graded, 3PLM, Nominal). The user should choose
ProposalStd and ProposalStd2 values so that the long-term average of the
88
acceptance rates (which is printed for each iteration in the progress window)
for level 1 and level 2 (if applicable) are around 0.5 for lower dimensional
models (<= 3 factors) and in the 0.2 - 0.3 range for higher dimensional/more
complex models. Generally speaking, increasing the ProposalStd value will
result in lowered acceptance rates while decreasing the value will result in
higher acceptance rates. Users are directed to Roberts and Rosenthal (2001)
for optimal scaling, choice of dispersion constants, and long-term acceptance
rates of Metropolis samplers.
MCSize is the Monte Carlo size for final log-likelihood, AIC, and BIC approximations.
5.3. Examples Utilizing MH-RM Estimation
All of the calibration examples presented previously in the manual are able
to be estimated using MH-RM estimation. However, requests for additional
things such as LD indices, more complete GOF statistics, etc. will most likely
R
be ignored by flexMIRT
because their behavior, in conjunction with the
MH-RM algorithm, is still under research.
5.3.1
MIRT Model with Correlated Factors
For a first example, we will refit the 4 correlated factors model applied to the
35-item QOL dataset, originally presented in Example 3.8.
Example 5.2: MHRM: 4 Factor MIRT Model
<Project>
Title = "QOL Data";
Description = "Items 2-21 Correlated Factors";
<Options>
Mode = Calibration;
Algorithm=MHRM;
RndSeed = 432;
ProposalStd = 0.5;
SavePRM = Yes;
Processor = 4;
FactorLoadings = Yes;
89
<Groups>
%Group1%
File = "QOL.DAT";
Varnames = v1-v35;
Select = v1-v21;
N = 586;
Dimensions = 4;
Ncats(v1-v21) = 7;
Model(v1-v21) = Graded(7);
<Constraints>
Fix (v1-v21),Slope;
Free
Free
Free
Free
(v1,v2-v5),Slope(1);
(v1,v6-v9),Slope(2);
(v1,v10-v15),Slope(3);
(v1,v16-v21),Slope(4);
Free Cov(2,1);
Free Cov(3,1);
Free Cov(3,2);
Free Cov(4,1);
Free Cov(4,2);
Free Cov(4,3);
R
In the <Options> section, we have specified that flexMIRT
should use the
MH-RM algorithm via Algorithm=MHRM;. In addition, we have provided the
required random number seed for the MH-RM estimation and set ProposalStd
= 0.5;. This value was adjusted from the default of 1.0 because a first run of
the algorithm was observed to have an acceptance rate lower than is desirable.
As noted earlier, lowering the ProposalStd value will generally lead to higher
acceptance rates. Outside of the <Options> section, no syntax was changed
from the example presented in Chapter 3; the same model was fit by both the
90
MH-RM and EM estimation routines.
Output 5.1: Correlated Four Factor Model Output - Processing Pane Output
MH-RM: Stage I
#
1; AR: 0.45,0.00;
#
2; AR: 0.46,0.00;
#
3; AR: 0.39,0.00;
#
4; AR: 0.35,0.00;
#
5; AR: 0.30,0.00;
#
6; AR: 0.27,0.00;
#
7; AR: 0.27,0.00;
#
8; AR: 0.25,0.00;
#
9; AR: 0.24,0.00;
#
10; AR: 0.23,0.00;
#
11; AR: 0.24,0.00;
#
12; AR: 0.23,0.00;
#
13; AR: 0.24,0.00;
...
MH-RM: Stage II
#
1; AR: 0.19,0.00;
#
2; AR: 0.19,0.00;
#
3; AR: 0.18,0.00;
#
4; AR: 0.19,0.00;
#
5; AR: 0.20,0.00;
...
MH-RM: Stage III
#
1; AR: 0.19,0.00;
#
2; AR: 0.19,0.00;
#
3; AR: 0.20,0.00;
...
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
-23418.96
-21948.41
-21098.19
-20510.71
-20335.52
-20145.06
-20118.94
-20052.04
-20065.11
-20092.89
-20071.26
-20086.42
-20087.42
Fn:
Fn:
Fn:
Fn:
Fn:
-20033.45
-19997.80
-20009.39
-20004.81
-20043.30
Fn:
Fn:
Fn:
-20055.31; Gam: 0.50; W: 0; P#:
-20112.81; Gam: 0.33; W: 0; P#:
-20042.63; Gam: 0.31; W: 0; P#:
12; Chg:
0; Chg:
12; Chg:
0.2178
0.0777
0.0943
As noted previously, acceptance rates are printed in the processing pane
for each iteration at all three stages of the MH-RM estimation. Given the
complex structure of the model and the IRT model applied to the items, the
desired goal for our fitted model is to have the reported acceptance rate have an
R
average value between 0.2 and 0.3. Reading across the output, flexMIRT
is
reporting the iteration number, the letters “AR” for acceptance rate, followed
by the level-1 acceptance and then the level-2 acceptance rate. The current
model had only 1 level, so all level-2 values will be 0.00. The final value, labeled
Fn, is an indicator that the program is or is not working towards a converged
solution. While not the log likelihood, it may be interpreted in a similar
fashion, meaning that less negative values indicate a better fit. Examining the
first several rows of the provided iteration history, the acceptance rate starts
somewhat high but quickly drops down to the near desired level by the 10th
iteration of Stage I; during this time, the Fn value also rapidly drops initially
91
and then stabilizes. For all of Stages II and III, both the acceptance rate and
the function value fluctuate, with the acceptance rate staying near the desired
value of 0.20. While the output should be closely examined for issues, this
pattern of values in the processing pane is an early indicator that the MH-RM
algorithm is working towards a good solution.
We present the same section of the output was that given for Example 3.8.
Output 5.2: Correlated Four Factor Model Output - Loadings, Means and
Variances
Factor Loadings for Group 1: Group1
Item
Label lambda 1
1
v2
0.83
2
v3
0.78
3
v4
0.93
4
v5
0.92
5
v6
0.00
6
v7
0.00
7
v8
0.00
8
v9
0.00
9
v10
0.00
10
v11
0.00
11
v12
0.00
12
v13
0.00
13
v14
0.00
14
v15
0.00
15
v16
0.00
16
v17
0.00
17
v18
0.00
18
v19
0.00
19
v20
0.00
20
v21
0.00
s.e. lambda 2
0.03
0.00
0.03
0.00
0.01
0.00
0.01
0.00
---0.87
---0.72
---0.90
---0.88
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
s.e. lambda 3
---0.00
---0.00
---0.00
---0.00
0.02
0.00
0.04
0.00
0.02
0.00
0.02
0.00
---0.66
---0.74
---0.71
---0.72
---0.77
---0.72
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
s.e. lambda 4
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
---0.00
0.05
0.00
0.04
0.00
0.04
0.00
0.04
0.00
0.04
0.00
0.04
0.00
---0.77
---0.73
---0.76
---0.84
---0.75
---0.49
s.e.
------------------------------------------0.03
0.04
0.04
0.03
0.04
0.06
QOL Data
Items 2-21 Correlated Factors
Group Parameter Estimates:
Group
Label
1
Group1
Latent Variable
P# Theta 1
1.00
141
0.40
142
0.54
144
0.54
P#
Variance-Covariance
s.e. P# Theta 2
---0.03
1.00
0.03 143
0.49
0.03 145
0.55
mu 1
0.00
s.e.
----
P#
mu 2
0.00
s.e.
----
P#
mu 3
0.00
Matrix for Group1: Group1
s.e. P# Theta 3
s.e. P# Theta 4
s.e.
---0.03
0.03 146
----
1.00
0.68
---0.02
1.00
s.e.
----
P#
mu 4
0.00
s.e.
----
From the excerpt of the output provided, it may again be observed that
the output from the MH-RM is formatted identically to output from EM estimations. Comparing across the two estimation routine results, the parameter
estimates are nearly identical indicating that the EM and MH-RM estimation
routines are reaching a similar converged solution. In favor of the MH-RM,
however, is the fact that the total time needed to estimate the model and
generate output was around 4 minutes, compared to the 54 minutes used by
the default Bock and Aitkin EM.
92
5.3.2
Multi-level MIRT Model
With the four factor example, we have demonstrated the time benefit that can
R
be obtained through the use of the flexMIRT
MH-RM algorithm when estiR
mating high dimensional models. The MH-RM as implemented in flexMIRT
is also able to fit multi-level models, a feature that, to our knowledge, is not
available in any other commercially available software. We will provide syntax
for and output from fitting such a model. The data used in this example was
simulated to have six dichotomous items with responses from 2000 individualsthese 2000 individuals are nested within 100 higher-order units, 20 per unit.
This type of structure could be seen if 2000 students from 100 different schools
took a brief assessment in which items were graded as right/wrong.
Example 5.3: MHRM: Two-level MIRT Model
<Project>
Title = "Fit Two-level MIRT Model to Simulated Data";
Description = "6 Items, 100 L2 units, 20 respondents within each";
<Options>
Mode = Calibration;
Rndseed = 10;
Algorithm = MHRM;
ProposalStd = 1.0;
ProposalStd2 = 1.0;
Processors = 2;
MCsize = 10000;
<Groups>
%Gr%
File = "simL2.dat";
Varnames = v1-v6,l2id;
Select = v1-v6;
Cluster = l2id;
Dimensions = 4;
Between = 2;
N = 2000;
Ncats(v1-v6) = 2;
Model(v1-v6) = Graded(2);
<Constraints>
Fix (v1-v6),Slope; // fix all slopes to begin with
93
Free
Free
Free
Free
(v1-v3),Slope(1);
(v4-v6),Slope(2);
(v1-v3),Slope(3);
(v4-v6),Slope(4);
//
//
//
//
level-2
level-2
level-1
level-1
factor
factor
factor
factor
1
2
1
2
Equal Gr,(v1-v3),Slope(1):Gr,(v1-v3),Slope(3); // cross-level equality
Equal Gr,(v4-v6),Slope(2):Gr,(v4-v6),Slope(4);
Free
Free
Free
Free
Cov(1,1);
Cov(2,2);
Cov(1,2);
Cov(3,4);
R
As before, we have told flexMIRT
to use the MH-RM algorithm and
provided a random number seed. In this example, we are using both the
ProposalStd and ProposalStd2 keywords because we now have two levels
to our model. The adjustments to the ProposalStd will affect the level-1
acceptance rate and adjustments to the ProposalStd2 value will affect the
R
level-2 acceptance rates. In the <Groups> section we inform flexMIRT
we
will be fitting a model with 4 latent dimensions (Dimensions = 4;), two of
which will be used as higher-level “between” factors (Between = 2;), and
we then indicate a variable that will supply the level-2 group membership
(Cluster = l2id;). Within the <Constraints> section, we create a structure
that has 2 “between” and 2 “within” factors (via the first group of Fix and
Free statements) and assign items to factors as desired. We also constrain
the slopes across the levels to equality using Equal statements and then allow
the variances of the level-2 factors and the covariance of the within factors
and the covariance of the between factors to be freely estimated. It is worth
noting here that there is nothing outside of the <Options> section that is
uniquely related to the requested MH-RM estimation - this specific model can
be estimated using the default Bock-Aitkin EM, albeit with noticeable time
loss. Larger models with more factors would most likely require MH-RM to
complete successfully.
We will cover the processing pane output prior to addressing the results
of the estimation. In particular, we will highlight the differences between the
previous example and the processing pane output for a model with a second
level.
94
Output 5.3: Two-Level MIRT Model - Processing Pane Output
MH-RM: Stage I
#
1; AR: 0.47,0.63;
#
2; AR: 0.48,0.61;
#
3; AR: 0.47,0.59;
#
4; AR: 0.46,0.59;
#
5; AR: 0.45,0.60;
#
6; AR: 0.44,0.57;
#
7; AR: 0.44,0.60;
#
8; AR: 0.43,0.58;
#
9; AR: 0.42,0.61;
#
10; AR: 0.43,0.60;
#
11; AR: 0.43,0.51;
...
MH-RM: Stage II
#
1; AR: 0.41,0.54;
#
2; AR: 0.41,0.59;
#
3; AR: 0.42,0.58;
#
4; AR: 0.42,0.60;
#
5; AR: 0.41,0.61;
...
MH-RM: Stage III
#
1; AR: 0.41,0.56;
#
2; AR: 0.41,0.56;
#
3; AR: 0.41,0.55;
#
4; AR: 0.41,0.54;
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
Fn:
-11738.27
-11666.05
-11556.98
-11447.69
-11366.38
-11236.31
-11164.65
-11131.30
-11147.14
-11092.40
-10987.70
Fn:
Fn:
Fn:
Fn:
Fn:
-10993.18
-10991.12
-11074.56
-11010.83
-11032.11
Fn:
Fn:
Fn:
Fn:
-10954.38;
-10846.63;
-10934.18;
-10915.08;
Gam:
Gam:
Gam:
Gam:
0.50;
0.33;
0.29;
0.25;
W:
W:
W:
W:
0;
0;
0;
0;
P#:
P#:
P#:
P#:
4;
15;
15;
12;
Chg:
Chg:
Chg:
Chg:
0.0451
0.0011
-0.0013
0.0199
While the general structure of the reported values is the same as the previous processing pane, rather than “0.00” being consistently repeated as the
second value after the AR: as in the first example, there is now a value that
changes with each iteration - this is the acceptance rate for the second level.
To be concrete, in Iteration 1 of Stage I, the reported level-1 acceptance rate
is 0.47 and the level-2 acceptance rate is 0.63. As noted before, the desired
range of acceptance rates is between 0.20 and 0.30 for large or complex models, with values around 0.50 being acceptable for smaller, less dimensionally
complex models. Output 5.4 presents the parameters for the estimated model.
It may again be observed that the item and group parameter output from the
MH-RM is formatted identically to output from EM estimations. Additionally,
it appears that the estimation completed successfully, resulting in reasonable
item and group parameter values and SEs.
95
96
P#
7
8
9
a 1
1.59
1.82
0.97
0.00
0.00
0.00
s.e.
0.09
0.11
0.06
----------
Latent Variable
P# Theta 1
13
0.74
15
0.47
0.00
0.00
P#
Variance-Covariance
s.e. P# Theta 2
0.12
0.10 14
0.88
---0.00
---0.00
Group Parameter Estimates:
Group
Label
1
Gr
s.e.
----
P#
10
11
12
P#
mu 2
0.00
a 2
0.00
0.00
0.00
1.79
1.62
0.89
0.14
------16
1.00
0.47
---0.02
Matrix for Group1: Gr
s.e. P# Theta 3
s.e.
mu 1
0.00
Fit Two-level MIRT Model to Simulated Data
6 Items, 100 L2 units, 20 respondents within each
2PL Items for Group 1: Gr
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
Fit Two-level MIRT Model to Simulated Data
6 Items, 100 L2 units, 20 respondents within each
P#
P#
7
8
9
1.00
P# Theta 4
s.e.
----
s.e.
---------0.11
0.10
0.05
----
s.e.
mu 3
0.00
a 3
1.59
1.82
0.97
0.00
0.00
0.00
Output 5.4: Multilevel MIRT Model - Item and Group Parameters
s.e.
----
s.e.
0.09
0.11
0.06
----------
P#
10
11
12
P#
mu 4
0.00
a 4
0.00
0.00
0.00
1.79
1.62
0.89
s.e.
----
s.e.
---------0.11
0.10
0.05
P#
1
2
3
4
5
6
c
-1.04
-0.52
0.68
1.64
-1.82
-0.35
s.e.
0.08
0.07
0.06
0.10
0.10
0.05
5.4. Fixed Effects Calibration
R
New to flexMIRT
is the ability to perform fixed effects (aka fixed theta) calibration, in which the user supplies individual theta values, which are fixed, to
be used in calibrating items. This mode of calibration is potentially useful in
field-testing experimental items. Fixed effects calibration is only available in
R
conjunction with MH-RM estimation. The keyword to induce flexMIRT
to
perform fixed effects calibration is FixedTheta. FixedTheta is a group-specific
command that is used to set the variable(s) that contain(s) the known, individual theta values for that group in the calibration. By implementing fixed
R
effects calibration in flexMIRT
as group-specific, it is possible to construct
a calibration in which one group is based on fixed theta values and the other
group is based on random, estimated thetas.
Example 5.4: Fixed Effects Calibration
<Project>
Title = "LSAT 6 - Fixed Effects Calibration";
Description = "5 Items 1PL N=1000";
<Options>
Mode = Calibration;
Algorithm = MHRM;
Stage1 = 0;
Stage2 = 0;
<Groups>
%Group1%
File = "lsat6.dat";
Varnames = v1-v5,id,map;
Select = v1-v5;
N = 1000;
NCats(v1-v5) = 2;
Model(v1-v5) = Graded(2);
FixedTheta = map;
<Constraints>
Equal (v1-v5),Slope;
97
For any group that will be calibrated with fixed ability values, theta is
read-in as part of the datafile for that group. In this example, the theta values
for %Group1% are in the variable map, the last variable listed in the Varnames =
; statement. As in earlier examples, because we have provided variables that
are not to be calibrated, we then use a Select statement to specify only those
R
with the
items which will be subjected to IRT analysis. To provide flexMIRT
theta values, we use the statement FixedTheta = map; to tell the program
the variable that holds the fixed theta values. Although not seen in this
example, FixedTheta works under multidimensionality - currently, the number
of variables listed in FixedTheta = vars ; must match the number of factors
specified via Dimensions for that group. Of final note in the fixed effects
calibration syntax, the Stage1 and Stage2 keywords in the <Options> section,
which determine the number of constant gain and stochastic EM, constant
gain cycles of the MH-RM estimation method, are both set to 0. This is
R
recommended for fixed effects calibration in flexMIRT
because these cycles
do not contribute much in the fixed effects case and eliminating them improves
the time-efficiency of the estimation.
When FixedTheta is specified and the program can successfully run a
R
problem, flexMIRT
will print in the main output a notification that fixed
effects calibration was performed, prior to reporting the item parameter estimates for the group. Such a notification is presented in the example program
output. Also with regard to the output, please note that for a single-group
fixed effects analysis or a fixed effects analysis that applies to all the groups
in a multiple-group run, the SE of the log-likelihood approximation will be
reported as 0.
Output 5.5: Fixed Effects Calibration Output - Item Parameters
*** Fixed effects calibration in Group 1: Group1
LSAT 6
5 Items 1PL N=1000
2PL Items for Group 1: Group1
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
P#
6
6
6
6
6
a
2.96
2.96
2.96
2.96
2.96
s.e.
0.08
0.08
0.08
0.08
0.08
98
P#
1
2
3
4
5
c
3.59
1.35
0.33
1.76
2.79
s.e.
0.14
0.08
0.08
0.09
0.11
b
-1.21
-0.45
-0.11
-0.59
-0.94
s.e.
0.03
0.01
0.00
0.02
0.03
CHAPTER 6
Cognitive Diagnostic Models
Over the last several decades, a growing field of interest and research has been
the development and estimation of latent variable models that locate respondents in multidimensional space for the purpose of diagnosis; in this context,
“diagnosis” includes the usual meaning, such as diagnosing a psychological disorder, but also encompasses activities such as identifying a student’s mastery
of content areas in an educational setting. Such models have been variously
termed cognitive diagnostic models (CDMs), diagnostic classification models,
cognitive psychometric models, restricted latent class models, and structured
IRT models, among others. The goal of this chapter is not to provide a didactic on CDMs (interested readers are directed to Rupp, Templin, and Henson
(2010) for a thorough discussion of CDMs and related models) but to demonR
strate, for those already familiar with the models, how flexMIRT
may be
used to fit them.
99
6.1. General CDM Modeling Framework
R
Complete coverage of the models flexMIRT
fits to CDMs is provided in the
Models chapter, but we will briefly introduce the models here. Figure 6.1
R
provides a schematic for a generic flexMIRT
CDM with a single higher-order
factor to facilitate the discussion of the mathematical equations.
Figure 6.1: Generic Higher-order CDM
R
The extended CDMs used by flexMIRT
combine a linear predictor with
a link function. The linear predictor may be represented as:
η =α +
K
X
k1 = 1
β k1 x k1 +
1 −1)
K (kX
X
βk1 ,k2 xk1 xk2
(6.1)
k1 = 1 k2 = 1
+ higher-order interaction terms +
p
X
λk ξk
k =1
where α, β, and λ are item parameters, xs are 0/1 attributes, and ξs are
continuous latent variables. The kernel (linear predictor) η can be turned into
item response probabilities through the use of appropriate link functions, such
100
as the logit presented below.
P (yi = 1|x, ξ ) = πi (1|x, ξ ) =
1
1 + exp[−(η )]
R
For the higher-order portion of the CDM, flexMIRT
is modeling the x
QK
probabilities, (i.e., P (x1 , ..., xK |θ1 , . . . , θq ) = k =1 P (xk |θ1 , . . . , θq )) with a
set of (higher-order) latent variables (θ), so that each of the attributes, xs, is
treated as if it is an observed item.
6.2. CDM Specific Syntax Options
Example 6.1: <Options>, <Groups>, and <Constraints>- CDM Specific
Options
<Options >
DMtable = Yes/No;
<Groups >
Attributes = ?;
InteractionEffects = (?,?,...);
Generate = (?,?), (?,?),...;
DM = group;
<Constraints >
Coeff group, (vars), parameter, ?;
Fix group,(vars), MainEffect;
Free group,(vars), Interaction(?, ?,..)
In the <Options> section, the user may control whether or not the (sometimes large) diagnostic classification probability table is printed in the output
via the DMtable command.
Attributes is used to set the number of main effects present in the skill
R
space that is being assessed. This is the keyword that triggers flexMIRT
to
model discrete latent variables, rather than the default continuous latent variables.
R
InteractionEffects instructs flexMIRT
to automatically generate all
possible interaction effects. For instance, with 4 attributes one may specify
101
InteractionEffects = (2,3,4); and, in addition to the 4 main effects, the
program with also generate the 2nd, 3rd, and 4th order interaction terms of
the attributes as dimensions of the model. If only the second order interaction
effects are desired, those would be created by specifying InteractionEffects
= (2);.
Similar to the InteractionEffects keyword, Generate sets up higher order interactions effects. However, rather than creating all possible interaction
effects Generate creates only those effects specified in the statement. For example, Generate = (3,6,7),(4,7); is used to generate the interaction effect
of attributes 3, 6, and 7 and, separately, the interaction effect of attributes 4
and 7. There is no limit to the number of interaction effects that may be
specified in a Generate statement and the order in which interaction effects
are specified does not matter.
The DM command stands for “diagnostic model” and is used to set the
group containing observed data - this is the group to which the higher-order
DM will be applied. The DM command is optional; one could, in principle, fit
a model without the DM (higher-order) latent variables.
In the <Constraints> section, the Coeff command allows user’s to impose proportionality restrictions on the item parameters (i.e., MainEffects and
Interactions). For instance, a DINO model may be parameterized as a logistic model with interactions but the interaction coefficients are -1.0 times the
main effect slopes. With the Coeff keyword, we can change the sign of those
R
specific interaction terms, effectively enabling flexMIRT
to fit more CDMs
than it would be capable of fitting if restricted to only the MainEffect and
Interaction keywords.
The Fix and Free commands noted in the <Constraints> section aren’t
limited to those particular commands or to CDMs, but are provided to show
that rather than using the typical IRT parameter keywords (Slope, etc.),
users may employ the parameter keywords MainEffect and Interaction to
allow syntax specification that is in keeping with the CDM conceptualization/terminology for the models.
6.3. CDM Modeling Examples
R
In this section, we will provide several examples of CDMs fit in flexMIRT
. We
will discuss basic examples that demonstrate how to fit several of the more wellknown CDMs that appear in the literature and then, using datasets that have
R
been fit elsewhere, demonstrate the comparability of flexMIRT
estimates to
those from other programs and estimation methods.
102
6.3.1
Introductory CDM Examples
Here we provide straight-forward demonstrations of several widely discussed
models. We hope that in addition to providing simple, concrete examples for
the chosen models, from these examples users will begin to see the flexibility
R
flexMIRT
has in fitting CDMs, gain insight into the details of our syntax for
CDMs, and be able to extrapolate this to fit the wide variety of alternative
R
CDMs that are not specifically covered but may be fit using flexMIRT
. For
the examples in this section, we will use simulated data for 15 dichotomous
items that were generated to measure four underlying attributes. The Qmatrix, which assigns items to attributes, is presented in Table 6.1. As can
be seen, the majority of items use only 1 attribute, but the final 3 items are
assigned to both attributes 3 and 4. These items could be fit with a variety of
models and we will present examples for three different CDMs.
Table 6.1: Q-matrix for Basic CDM Demonstrations
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Attribute 1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
Attribute 2
0
0
0
1
1
1
0
0
0
0
0
0
0
0
0
Attribute 3
0
0
0
0
0
0
1
1
1
0
0
0
1
1
1
Attribute 4
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
Basic C-RUM Fit to Simulated Data
For the first example, the majority of items will only depend on the main effect
of a single attribute and we will fit the compensatory reparameterized unified
model (C-RUM) to the final three items. The C-RUM model (e.g., Hartz,
103
2002; Choi, Rupp, & Pan, 2013) is a CDM that, when items are informed
by multiple attributes, allows for mastery of one attribute to compensate for
non-mastery on other attributes. The C-RUM estimates for each item an
intercept term and as many slope terms for an item as there are “1” entries
in the Q-matrix. As detailed in Choi et al. (2013), C-RUM is a special case
of the log-linear cognitive diagnosis model (LCDM) in which all interaction
terms are set to 0, meaning only main effects are estimated.
Example 6.2: C-RUM Model Fit To Simulated Data
<Project>
Title = "Calibrate and Score DM Data";
Description = "C-RUM Model";
<Options>
Mode = Calibration;
MaxE = 20000;
MaxM = 5;
Mtol = 0;
Etol = 1e-5;
SE = REM;
SavePrm = Yes;
SaveCov = Yes;
SaveSco = Yes;
Score = EAP;
<Groups>
%G%
File = "DMsim.dat";
Varnames = v1-v15;
N = 3000;
Ncats(v1-v15) = 2;
Model(v1-v15) = Graded(2);
Attributes = 4;
Dimensions = 4;
104
%D%
Varnames = a1-a4;
DM = G;
<Constraints>
Fix G,(v1-v15),MainEffect;
Free
Free
Free
Free
G,(v1-v3),MainEffect(1);
G,(v4-v6),MainEffect(2);
G,(v7-v9, v13-v15),MainEffect(3);
G,(v10-v15),MainEffect(4);
The first points of note in the syntax occur in the <Options> section. The
maximum number of E-steps has been increased from the default, the E tolerance and M tolerance values have been decreased, and the SE calculation
method has been set to REM (Richardson extrapolation method), as experience has found the REM SEs to be preferable to estimates from other methods.
These changes in the <Options> section are made to improve the resulting SE
R
estimates and are recommended for any CDMs fit using flexMIRT
.
The specifications for the CDM begin in the <Groups> section. Although
the naming of the groups is arbitrary, all examples will use the custom of
defining group G as the group containing observed data and group D as the
group containing the higher-order latent dimensions/attributes. As with a
typical IRT model, in group G we read in our data from a named file, provide
R
flexMIRT
with the variable names, and specify the sample size. The items
were simulated as correct/incorrect items, so Ncat= has been set to 2 and we
will fit the 2PLM to all items. For the first statement specific to the CDM,
R
we tell flexMIRT
that 4 latent attributes will be used. Because the C-RUM
model uses only main effect terms, the total number of dimensions is equal to
the number of attributes.
R
In the D group, we inform flexMIRT
that the higher-order portion of
the model we are setting up will be applied to the group with observed data,
group G, using the DM = G; statement. We then name the four attributes via
Varnames = a1- a4;. If no other information is given regarding categories or
R
a model to be fit, flexMIRT
by default will assume there are two categories
for each of the attributes and will fit the attributes with the 2PLM. In the
<Constraints> section, we assign items onto attributes according to the Qmatrix. First, we fix all items to 0 and then free items onto the appropriate
main effect and interaction terms. Items 1 - 12 only have main effects for a
105
single attribute specified in the <Constraints> section. From the Q-matrix,
we know that items 13 - 15 will load on both attributes 3 and 4. These
specifications are incorporated in the statements Free G,(v7-v9, v13-v15),
MainEffect(3); and Free G,(v10-v15), MainEffect(4);, in which items
13 - 15 appear in the list of items freed onto both attribute 3 and attribute 4.
106
107
Label
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
v11
v12
v13
v14
v15
P#
24
25
26
a 1
1.89
1.91
1.91
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
0.14
0.15
0.17
------------------------------------27
28
29
P#
a 2
0.00
0.00
0.00
2.43
2.23
1.88
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
---------0.25
0.17
0.33
---------------------------33
34
35
30
31
32
P#
Pattern
0 0 0 0
0 0 0 1
0 0 1 0
0 0 1 1
0 1 0 0
0 1 0 1
0 1 1 0
0 1 1 1
1 0 0 0
1 0 0 1
1 0 1 0
1 0 1 1
1 1 0 0
1 1 0 1
1 1 1 0
1 1 1 1
Prob
0.15002832
0.16623281
0.03889021
0.19192117
0.00750342
0.03906408
0.00531547
0.11005979
0.00145117
0.01593558
0.00165291
0.07127746
0.00033627
0.01527851
0.00093495
0.18411789
Diagnostic IRT Attributes and Cross-classification Probabilities for Group
Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
1: G
a 3
0.00
0.00
0.00
0.00
0.00
0.00
1.93
2.16
1.69
0.00
0.00
0.00
0.10
0.70
0.55
s.e.
------------------0.14
0.15
0.13
---------0.41
0.28
0.15
36
37
38
39
40
41
P#
4
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.47
1.68
1.92
0.96
1.03
1.85
a
s.e.
---------------------------0.18
0.17
0.20
0.42
0.26
0.20
P#
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Output 6.1: C-RUM Output - Item Parameters and Attribute Profile Posterior Probabilities
c
-1.68
-1.14
-0.63
-0.34
-1.16
1.54
-1.29
-0.85
-0.37
-0.05
0.77
1.14
2.77
1.60
-1.38
s.e.
0.07
0.07
0.06
0.07
0.08
0.07
0.10
0.09
0.08
0.11
0.11
0.12
0.22
0.14
0.13
The first section of the output excerpt presents the item parameters estimates, with freely-estimated versus fixed-at-zero parameters mirroring the Qmatrix specifications. The factors labeled a1-a4 represent main effects/slope
parameters for attributes 1 - 4, respectively, and c is the intercept parameter.
The next section of interest is labeled Diagnostic IRT Attributes and
Cross-classification Probabilities for Group 1: G, which reports the
estimated proportion of the population in each attribute profile. For instance,
it is expected that about 15% of the population have mastered none of the
measured attributes (attribute profile [0 0 0 0]), around 4 % have mastered attributes 2 and 4 (profile [0 1 0 1]), and 18.4% of the population have mastered
all 4 attributes (profile [1 1 1 1]).
We will next cover the structure of the individual scoring file (-sco) that
was requested in the <Options> section. In the -sco output excerpt presented,
we have given only the first 10 cases and have included line breaks (//) and a
separator (|) not in the output file for ease of presentation.
108
109
1
1
0.053631
0.125782
0.164556
0.375016 |
0.225289
0.331604
0.370779
0.484127
//
0.050755
0.023774
0.109961
0.024396
0.033344
0.137477
0.029047
0.049426
0.051856
0.234379 //
0.547785
0.203755
0.044491
0.055075
0.023108
0.040386
0.005129
0.026639
0.002596
0.009569
0.000926
1
-0.863639
0.559221
0.312728
1
2
0.094169
0.197099
0.208524
0.753622 |
0.292064
0.397808
0.406253
0.430901
//
0.085301
0.036159
0.158251
0.035482
0.043046
0.165042
0.021494
0.037324
0.035284
0.185676
//
0.219748
0.436824
0.014035
0.092845
0.009270
0.086583
0.001618
0.044908
0.001060
0.020890
0.000298
2
-0.458755
0.526603
0.277310
1
3
0.028209
0.072130
0.100814
0.199482 |
0.165569
0.258703
0.301083
0.399611
//
0.027413
0.011188
0.066927
0.011742
0.016077
0.090651
0.017207
0.027676
0.029037
0.159689 //
0.720026
0.121128
0.045985
0.025745
0.024803
0.019606
0.004329
0.010169
0.003468
0.005783
0.000973
1
-1.088950
0.574875
0.330481
1
4
0.009269
0.137086
0.038002
0.705494 |
0.095827
0.343938
0.191201
0.455820 //
0.009183
0.003258
0.118294
0.001570
0.007045
0.036558
0.002476
0.028187
0.007495
0.207772 //
0.278809
0.554228
0.003301
0.021836
0.011761
0.109853
0.000381
0.010562
0.000199
0.003930
0.000010
2
-0.674111
0.528978
0.279818
1
5
0.002525
0.007245
0.018326
0.233881 |
0.050183
0.084809
0.134128
0.423297 //
0.002518
0.000148
0.007193
0.000266
0.000372
0.017990
0.001356
0.002651
0.004997
0.179181 //
0.753729
0.219111
0.008923
0.008633
0.002798
0.003822
0000091
0.000367
0.000539
0.001554
0.000028
1
-1.155748
0.576725
0.332611
1
6
0.002825
0.007694
0.376743
0.057087 |
0.053073
0.087379
0.484570
0.232008 //
0.002817
0.000224
0.007635
0.001179
0.002216
0.234808
0.001130
0.001577
0.016181
0.053828 //
0.601189
0.018925
0.334564
0.035049
0.002232
0.000330
0.003394
0.001492
0.000430
0.000134
0.001052
1
7
0.004550
0.083366
0.023717
0.309510 |
0.067300
0.276435
0.152167
0.462292 //
0.004529
0.001691
0.076416
0.000765
0.004134
0.023155
0.002547
0.028993
0.007709
0.213714 //
0.653688
0.243148
0.007739
0.009580
0.027575
0.048194
0.000892
0.004634
0.000468
0.001724
0.000024
1
-1.047601
0.567131
0.321638
1
8
0.020510
0.009435
0.131043
0.184563 |
0.141736
0.096676
0.337448
0.387943 //
0.020089
0.001817
0.009346
0.008108
0.002867
0.113871
0.011562
0.004088
0.029010
0.150500 //
0.731348
0.123032
0.075446
0.042239
0.002715
0.002146
0.000765
0.001798
0.003466
0.005778
0.001571
1
-1.123851
0.575411
0.331098
1
9
0.126701
0.227366
0.460304
0.624168 |
0.332637
0.419131
0.498422
0.484337 //
0.110648
0.052320
0.175671
0.049910
0.064527
0.248424
0.043652
0.064029
0.085670
0.234582 //
0.275306
0.195548
0.076165
0.180042
0.011614
0.038759
0.008781
0.087085
0.001326
0.009335
0.001612
1
-0.403618
0.536756
0.288107
1
10
0.020842
0.009465
0.131177
0.184754 |
0.142855
0.096829
0.337594
0.388098 //
0.020408
0.001846
0.009376
0.008236
0.002886
0.113970
0.011746
0.004110
0.029093
0.150620 //
0.731100
0.122991
0.075420
0.042224
0.002714
0.002145
0.000765
0.001797
0.003522
0.005872
0.001596
1
-1.123361
0.575390
0.331074
Output 6.2: C-RUM Output –SCO Output
0.000259
0.000061
0.007809
0.000007
0.000963
0.033299
0.000009
0.000237
0.000060
0.000039
0.000600
0.007685
0.000554
0.004762
0.000091
0.000207
0.017201
0.000263
0.000507
0.010020
0.000418
0.007549
0.000411
0.001394
0.000010
0.000111
0.003178
0.003819
0.016893
0.007738
0.000067
0.000769
0.000066
0.000012
0.000044
0.000001
0.000005
0.000379
0.000142
0.000442
0.001497
0.072550
0.001474
0.000573
0.000185
0.000045
0.001307
0.008472
0.037478
0.021832
//
//
//
//
//
//
//
//
//
//
The first two columns of the -sco file give the group and observation numR
ber, similar to a standard flexMIRT
IRT score output file. Recall that for
this example, there are 4 attributes and 16 possible attribute profiles. The estimated probability of the respondent having mastered each of the attributes
individually is given, starting after the observation number. For example, for
observation 1, the probability they have mastery of attribute 1 is 0.054, for
attribute 2 it is 0.128 and for attribute 3 it is 0.165 and for attribute 4 it
is 0.375. The following four columns, after the separator, are the estimated
standard errors for those probabilities, respectively.
The entries that have been moved to the second line of each observation
contain the estimated attribute posterior variance/covariance matrix. For observation 1, those values would enter the matrix as:
a1
a2
a3
a4
a1
0.050755
0.023774
0.024396
0.029047
a2
a3
a4
0.109961
0.033344
0.049426
0.137477
0.051856
0.234379
In what we have made the third line for each respondent are the posterior
probabilities for each of the possible attribute profiles. As noted earlier with 4
attributes, there are 16 possible profiles and, therefore, 16 values on the third
line for this example. For observation 1, the posterior probability of being
in attribute profile 1 is 0.548, the probability of possessing attribute profile 2
is 0.204 and so on. The order in which the attribute profile probabilities are
given matches the order of the probability profiles used in the output table
“Diagnostic IRT Attributes and Cross-classification Probabilities.” Looking
back at the output presented for this example, we find that attribute profile
1 corresponds to [0 0 0 0] (meaning no attributes are mastered) and attribute
profile 2 corresponds to [0 0 0 1] (meaning only attribute 4 has been mastered),
etc.
Finally, in what has been moved to the fourth line for each respondent the
most likely attribute profile is listed, followed by the individual’s estimated
theta score for the higher-order latent variable, as well as the the SD and
variance estimates for the higher-order theta estimate. For respondent 1, the
most likely attribute profile is profile 1, corresponding to attribute profile [0 0
0 0] (mastery of no attributes); this attribute can be found to have the highest
posterior probability listed on the line above.
110
Basic DINA Fit to Simulated Data
For the second example, we will fit the deterministic input, noisy-and-gate
(DINA) model to the final three items. Unlike the C-RUM, this model is noncompensatory in nature, not allowing mastery on one attribute to make up
for lack of mastery on other attributes. This is accomplished in the LCDM
framework by constraining main effects to zero and allowing only the most
complex attribute interaction term for an item to take on a non-zero value.
Example 6.3: DINA Fit to Simulated Data
<Project>
Title = "Calibrate and Score DM Data";
Description = "DINA Model";
<Options>
Mode = Calibration;
MaxE = 20000;
MaxM = 5;
Etol = 1e-5;
SE = REM;
SavePrm = Yes;
SaveCov = Yes;
SaveSco = Yes;
Score =SSC;
<Groups>
%G%
File = "DMsim.dat";
Varnames = v1-v15;
N = 3000;
Ncats(v1-v15) = 2;
Model(v1-v15) = Graded(2);
Attributes = 4;
InteractionEffects = (2);
Dimensions = 10; //4 ME and 6 2nd order ints.
%D%
Varnames = a1-a4;
DM = G;
111
<Constraints>
Fix G,(v1-v15),MainEffect;
Free
Free
Free
Free
G,(v1-v3),MainEffect(1);
G,(v4-v6),MainEffect(2);
G,(v7-v9),MainEffect(3);
G,(v10-v12),MainEffect(4);
// item 13-15 fit with DINA model that depends on int of attribs 3,4
Free G,(v13-v15),Interaction(3,4);
All syntax in the <Options> section remains unchanged from the previous
example, with the exception of requesting sum score to EAP conversion tables
(via Score = SSC;). In the G group, we have declared the data file, variable
names, sample size, number of categories and IRT model to be fit to each
item, as in the previous example. We again specify that there are 4 attributes
with main effects, but because we are using the DINA model we will also
need to construct dimensions that will represent the interaction effects. This
is accomplished via InteractionEffects = (2); which is requesting that all
second order interaction terms (i.e., attribute 1 with attribute 2, attribute 1
with attribute 3, etc.) be created for the model. Because we have requested all
second order interactions (of which there are 6 possible with the 4 attributes),
the total number of dimensions for the final model is 10 (4 main effect terms
and 6 second order interactions). Syntax in the D group remains unchanged
from the previous example.
In the <Constraints> section, as with the
previous example, we initially fix all items to zero and then free items onto
the appropriate attributes. Unlike the previous example, which used only main
effects, we are also using the parameter keyword Interaction to free items
13-15 onto the interaction of attributes 3 and 4. As noted when introducing
the model, for items that depend on more than one attribute, only the highest
order interaction terms is allowed to be non-zero, so we do not free items 13-15
onto the main effects of attributes 3 or 4.
Because the majority of the output has the same structure as the previous
example’s, regardless of specific CDM fit to items, we will only briefly discuss
outputted estimates. For this example, we highlight the output the relates to
the item parameters and the sum score to EAP conversion table in the -ssc
file that is printed due to Score = SSC;.
112
113
Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Label
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
v11
v12
v13
v14
v15
2PL Items for Group 1: G
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
7
v7
8
v8
9
v9
10
v10
11
v11
12
v12
13
v13
14
v14
15
v15
P#
P#
24
25
26
a 7
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
a 1
1.87
1.93
1.90
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
----------------------------------------------
s.e.
0.14
0.15
0.17
------------------------------------P#
27
28
29
P#
a 8
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
a 2
0.00
0.00
0.00
2.38
2.23
1.87
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
----------------------------------------------
s.e.
---------0.25
0.18
0.32
---------------------------P#
30
31
32
P#
a 9
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
a 3
0.00
0.00
0.00
0.00
0.00
0.00
2.01
2.16
1.63
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
----------------------------------------------
s.e.
------------------0.15
0.15
0.13
-------------------
Output 6.3: DINA output - Item Parameter Estimates
36
37
38
P#
33
34
35
P#
a 10
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.74
1.33
1.58
a 4
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.53
1.70
1.90
0.00
0.00
0.00
s.e.
------------------------------------0.30
0.22
0.12
s.e.
---------------------------0.17
0.18
0.21
---------P#
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
P#
c
-1.67
-1.14
-0.63
-0.34
-1.17
1.54
-1.55
-1.09
-0.52
0.04
0.83
1.24
3.11
2.01
-0.52
a 5
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
0.07
0.07
0.06
0.07
0.08
0.07
0.13
0.11
0.09
0.09
0.11
0.11
0.17
0.11
0.08
s.e.
----------------------------------------------
P#
a 6
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
----------------------------------------------
The output excerpt presents the estimated parameters for the items, with
estimated versus fixed-at-zero loadings mirroring the Q-matrix specifications.
The factors labeled a1 - a4 represent main effects for attributes 1 - 4 and
a5 - a10 represent the requested 2nd order interactions. Items 1 - 12 have
estimates for main effects on only the attributes specified in the Q-matrix
and item 13-15 have estimates on only the interaction of attributes 3 and
4, which is labeled as a10 in the parameter table. One may notice that dimensions a5 - a9 that have no items assigned to them - these dimensions
correspond to the unused second order interaction effects that were created
via the InteractionEffects keyword. The estimated item values reported
are in the LCDM parameterization, but the classic DINA parameters of guessing and slippage may be obtained by conversion. Guessing is found from the
R
reported flexMIRT
parameters as exp(c)/(1 + exp(c)) and one minus the
slippage term (1 − s) is found as exp(c + a)/(1 + exp(c + a)).
In the -ssc file that was generated by the scoring request, we find tables
that provide conversions from sum scores to the probability of mastery for
each attribute. We present an excerpt of these table below.
Output 6.4: DINA Summed Score Conversion Tables from -ssc File
Summed Score to Attribute Profile Probability Conversion Table:
Summed
Score EAP 1 EAP 2 EAP 3 EAP 4
SD 1
SD 2
SD 3
0.00
0.001
0.001
0.026
0.005
0.030
0.026
0.158
1.00
2.00
0.001
0.001
0.001
0.002
0.029
0.036
0.010
0.021
0.032
0.036
0.035
0.049
0.168
0.186
SD 4
P
0.070 0.0000002
0.097 0.0000093
0.143 0.0001508
Error Covariance Matrix
0.000876
0.000003
0.000700
0.000115
0.000045
0.000017
0.000007
0.024902
-0.000092
0.004948
0.001025
0.000006
0.000147
0.000035
0.001256
0.000085
0.000017
0.028217
-0.000168
0.009475
0.001332
0.000013
0.000214
0.000084
0.002420
0.000173
0.000041
0.034470
-0.000363
0.020508
0.182430
0.027666
0.002884
0.005347
0.163028
0.002254
0.003948
0.009474
0.001326
0.016527
0.091595
0.013076
0.000808
0.001972
0.099016
0.000623
0.001464
0.002789
0.000291
0.007550
0.032539
0.004564
0.000175
0.000572
0.050540
0.000147
0.000493
0.000739
0.000057
0.003578
...
13.00
14.00
15.00
0.760
0.898
0.966
0.795
0.889
0.947
0.990
0.997
0.999
0.983
0.992
0.996
0.427
0.303
0.180
0.404
0.315
0.225
0.097
0.053
0.027
0.129 0.0805989
0.087 0.0356330
0.060 0.0078999
For this example, each summed score is listed with four EAP values, corresponding to the 4 attributes specified in the model, 4 SD values, the ex-
114
pected proportion of respondents who will obtain the listed summed score
(listed in the column labeled p), and the error variance/covariance matrix
associated with that summed score. For example, given a summed score of
0, the probability that the individual has mastered attribute 1 is 0.001, the
probability that they have mastered attribute 3 is 0.026 and so on. For individuals with a summed score of 15, those probabilities increase to 0.966
and 0.999, respectively. Rather than simply providing the most likely atR
tribute profile associated with a summed score, flexMIRT
prints the individual attribute probabilities so researchers have sufficient information to apply
mastery/indifference/non-mastery cut-off values appropriate for the specific
assessment at hand.
Basic DINO Fit to Simulated Data
The final basic example using this simulated dataset will fit the disjunctive,
deterministic noisy-or-gate (DINO) model. Unlike the previous models, the
DINO allows for both non-zero main effect and interaction terms. However,
and again using the LDCM framework, constraints are placed, such that the
main effect slopes are set equal and the interaction term is -1 times the main
effect constrained-to-equality slope. This is done because an item may be
endorsed/answered correctly if any one of the attributes is mastered and there
is assumed to be no added benefit to having mastery of several/all attributes
over mastering only one or a subset of the attributes - the -1 applied to the
interaction term serves to remove such a “benefit” from the model.
Example 6.4: DINO Fit to Simulated Data
<Project>
Title = "Calibrate and Score Simulated DM Data";
Description = "DINO Model";
<Options>
Mode = Calibration;
MaxE = 20000;
Etol = 1e-5;
SE = REM;
SavePrm = Yes;
SaveCov = Yes;
SaveSco = Yes;
Score = EAP;
<Groups>
%G%
File = "DMsim.dat";
Varnames = v1-v15;
115
N = 3000;
Ncats(v1-v15) = 2;
Model(v1-v15) = Graded(2);
Attributes = 4;
Generate = (3,4);
Dimensions = 5; //4 ME and 1 2nd order int.
%D%
Varnames = a1-a4;
DM = G;
<Constraints>
Fix G,(v1-v15),MainEffect;
Free
Free
Free
Free
G,(v1-v3),MainEffect(1);
G,(v4-v6),MainEffect(2);
G,(v7-v9),MainEffect(3);
G,(v10-v12),MainEffect(4);
// items 13-15 fit with DINO model that depend on int of attribs 3,4
Free G,(v13-v15),MainEffect(3);
Free G,(v13-v15),MainEffect(4);
Free G,(v13-v15),Interaction(3,4);
Equal G, (v13), MainEffect(3):G, (v13), MainEffect(4):G, (v13), Interaction(3,4);
Coeff G, (v13), Interaction(3,4), -1;
Equal G, (v14), MainEffect(3):G, (v14), MainEffect(4):G, (v14), Interaction(3,4);
Coeff G, (v14), Interaction(3,4), -1;
Equal G, (v15), MainEffect(3):G, (v15), MainEffect(4):G, (v15), Interaction(3,4);
Coeff G, (v15), Interaction(3,4), -1;
Again, the maximum number of E-step and M-steps have been increased
from the default, the E tolerance and M tolerance values have been decreased,
and the SE calculation method has been set to REM in the <Options> section.
As noted in the output from the previous example, there were several latent
dimensions that had no items loading onto them. Here, rather than using
the general InteractionEffects keyword, we have used the more targeted
Generate to construct only those interaction terms which will be used in the
model. As seen in the Q-matrix and discussed earlier, only the interaction
of attributes 3 and 4 is needed and is specified by the Generate = (3,4);
statement. If additional interactions were needed, they may be added by
placing a comma after the last interaction term and listing the other desired
interactions. For example, if the 3-way interaction of attributes 2, 3 and 4 was
desired in addition to the 3-4 interaction, we could specify that using a single
generate statement of Generate = (3,4), (2,3,4);.
116
As before, we free the first 12 items onto the appropriate single attributes.
Following the comment in the syntax, we free items 13 -15 on both main effect
terms for attributes 3 and 4, as well as the interaction of the two attributes. To
induce the equality specified in the model, we set the main effects and interaction terms equal to each of the items via a statement such as Equal G, (v13),
MainEffect(3):G, (v13), MainEffect(4):G, (v13), Interaction(3,4);
and then apply the -1 to the interaction with a coefficient constraint of the
form Coeff G, (v13), Interaction(3,4), -1;.
117
118
Item Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
7
v7
8
v8
9
v9
10
v10
11
v11
12
v12
13
v13
14
v14
15
v15
P#
24
25
26
a 1
1.89
1.90
1.90
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
0.14
0.15
0.17
------------------------------------27
28
29
P#
a 2
0.00
0.00
0.00
2.47
2.23
1.91
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
s.e.
---------0.26
0.17
0.34
---------------------------36
37
38
30
31
32
P#
Output 6.5: DINO Output - Item Parameters
a 3
0.00
0.00
0.00
0.00
0.00
0.00
1.84
2.17
1.68
0.00
0.00
0.00
1.06
1.51
2.53
s.e.
------------------0.13
0.14
0.13
---------0.29
0.20
0.14
33
34
35
36
37
38
P#
a 4
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.42
1.63
1.91
1.06
1.51
2.53
s.e.
---------------------------0.17
0.17
0.20
0.29
0.20
0.14
36
37
38
P#
a 5
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
-1.06
-1.51
-2.53
s.e.
------------------------------------0.29
0.20
0.14
P#
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
c
-1.68
-1.14
-0.63
-0.33
-1.15
1.55
-1.18
-0.79
-0.32
0.12
0.87
1.25
2.74
1.60
-1.66
s.e.
0.08
0.07
0.06
0.07
0.08
0.07
0.09
0.09
0.08
0.10
0.10
0.11
0.21
0.14
0.13
As with the DINA model, the estimated item values reported are in the
LCDM parameterization but DINO parameters of guessing and slippage may
be obtained by conversion. Again, guessing is found from the reported parameters as exp(c)/(1 + exp(c)) and one minus the slippage term (1 − s) is found
as exp(c + a)/(1 + exp(c + a)). The a term used in the conversion should be
the value reported in one of the main effect columns, not the negative value
associated with the interaction term.
One aspect of the output that we haven’t discussed as of yet is the model
fit. We present the -2 log likelihood, AIC, and BIC values for each of the
three models fit to our simulated data in Table 6.2. As with the output for
more standard IRT models, these values are reported at the bottom of the
R
-irt flexMIRT
output file. As can be seen in Table 6.2, the fit of the models
indicates that using DINO for the final three items was the preferred model
based on all three reported fit indices. It is informative to know that the same
data set was used in all three examples and this data was generated using the
DINO model for the last three items. The advantage for the DINO is sometimes small, but that is most likely a result of the majority of items being fit
by single main effects, the specification of which did not vary regardless of the
model that was used for the items with more complex attribute relationships.
Table 6.2: Model Fit for Basic CMDs Fit to Simulated Data
CRUM
-2loglikelihood:
Akaike Information Criterion (AIC):
Bayesian Information Criterion (BIC):
DINA
DINO
45799.53 45873.59 45798.31
45881.53 45949.59 45874.31
46156.21 46204.17 46128.89
Basic DINA with Testlet Structure
For the final example in this section, we employ a new simulated dataset
that contains responses to 24 dichotomous items that are to be modeled with
4 attributes. In addition to being generated from a CDM, this data was
simulated to incorporate a testlet structure, such as when multiple reading
items use the same passage as a prompt. A graphical depiction of such a
model was given in Figure 6.1, where ys are the observed items, xs are the
discrete attributes, and ξs are the testlet effects.
For the 24 item dataset we will be modeling, items 1 - 6 are members of
testlet 1, items 7 - 12 belong to testlet 2, items 13-18 to testlet 3, and items
119
19-24 are members of the final testlet. The Q-matrix for mapping items onto
attributes supplied in Table 6.3.
Table 6.3: Q-matrix for Testlet-DINA Example
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Attribute 1
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
Attribute 2
1
0
0
1
0
1
1
0
0
1
0
1
1
0
0
1
0
1
1
0
0
1
0
1
Attribute 3
0
1
1
0
0
1
0
1
1
0
0
1
0
1
1
0
0
1
0
1
1
0
0
1
Attribute 4
0
1
0
1
1
0
0
1
0
1
1
0
0
1
0
1
1
0
0
1
0
1
1
0
As can be seen in syntax, the <Options> section is largely unchanged
from the previous examples, although we have reduced the number quadrature
points from the default, due to the high-dimensional nature of this problem. In
the first group, G, the data file, variable names, etc. have been given. As with
previous CDM examples, we specify the number of attributes (Attributes =
R
4;) and also request that flexMIRT
create dimensions for the interaction effects. From the Q-matrix, every possible second order interaction is needed, so
120
we use InteractionEffects = (2); to generate those dimensions automatically. To incorporate the testlet structure into our model, we will be fitting
a bifactor-type model, in which the CDM main effect and interactions are the
primary/general dimensions and the testlet effects are considered the specific
R
dimensions. To indicate this to flexMIRT
we set Primary = 10; which is
the total number of CDM dimensions (4 main effects + 6 interactions) and set
the total number of dimensions for the model equal to 14 (10 CDM dimensions
+ 4 testlet effect dimensions).
In the <Constraints> section, the first group of statements is used to
assign items to the appropriate interaction terms, based on the Q-matrix,
for the DINA model. The second group of constraints, beginning with Free
G,(v1-v6),Slope(11); is used to assign items to their testlet dimension. As
noted earlier, the first 10 latent dimensions are taken up by the CDM, so the
first available dimension for a testlet effect is dimension 11, hence Slope(11) in
the noted Free statement. The specific dimensions are specified as a restricted
bifactor model, where the testlet effect is presumed to be a testlet-specific
random intercept term. This requires the equality of the slopes within factors
and is accomplished with the final group of constraints. With the final Equal
constraint, we are fitting a Rasch-type model to the higher-order dimensions,
constraining the slopes of attribute 1 through attribute 4 in group D to equality.
Example 6.5: Testlet Structure with DINA model
<Project>
Title = "Fit HO-DINA Model with Testlets";
Description = "HO-DINA Model - 4 attributes";
<Options>
Mode = Calibration;
MaxE = 10000;
Etol = 1e-4;
Mtol = 1e-7;
Quadrature = 25,6.0;
Processors = 2;
GOF = Extended;
SE = REM;
SaveDbg = Yes;
SaveSco = Yes;
121
Score = EAP;
<Groups>
%G%
File = "fitDINAt.dat";
Varnames = v1-v24;
Ncats(v1-v24) = 2;
Model(v1-v24) = Graded(2);
Attributes = 4;
InteractionEffects = (2); //6 possible 2nd order ints.
Primary = 10; //number of dimensions to CDM
Dimensions = 14; //total dims, including testlet effects
%D%
Varnames = a1-a4;
DM = G;
<Constraints>
Fix G,(v1-v24),MainEffect;
Free G,(v1,v7,v13,v19),Interaction(1,2);
Free G,(v3,v9,v15,v21),Interaction(1,3);
Free G,(v5,v11,v17,v23),Interaction(1,4);
Free G,(v6,v12,v18,v24),Interaction(2,3);
Free G,(v4,v10,v16,v22),Interaction(2,4);
Free G,(v2,v8,v14,v20),Interaction(3,4);
Free
Free
Free
Free
Equal
Equal
Equal
Equal
G,(v1-v6),Slope(11);
G,(v7-v12),Slope(12);
G,(v13-v18),Slope(13);
G,(v19-v24),Slope(14);
G,(v1-v6),Slope(11);
G,(v7-v12),Slope(12);
G,(v13-v18),Slope(13);
G,(v19-v24),Slope(14);
Equal D,(a1-a4),Slope(1);
122
6.3.2
Replicating Published Examples
Because users are most likely acquainted with fitting CDMs in other programs,
we also provide syntax and output for well-known examples to demonstrate
R
that comparable parameter estimates may be obtained from flexMIRT
and
benefits of either computation-time savings or simplicity of syntax are obtained.
The first example is a replication of Example 9.2 presented in Diagnostic Measurement: Theory, Methods and Applications by Rupp et al. (2010).
The simulated data presented in this example uses 7 dichotomous items that
measure 3 attributes. The Q-matrix given by the authors is duplicated here.
Table 6.4: Q-matrix for Rupp, Templin, & Henson’s Example 9.2
Item
Item
Item
Item
Item
Item
Item
Item
1
2
3
4
5
6
7
Attribute 1
1
0
0
1
1
0
1
Attribute 2
0
1
0
1
0
1
1
Attribute 3
0
0
1
0
1
1
1
R
With the Q-matrix, we are now able to generate our flexMIRT
CDM syntax.
Example 6.6: Rupp, Templin, & Henson - Example 9.2
<Project>
Title = "Rupp Templin Henson Example 9.2";
Description = "Saturated LCDM model - 7 items, 3 attributes";
<Options>
Mode = Calibration;
MaxE = 20000;
Etol = 1e-6;
MaxM = 50;
Mtol = 1e-9;
GOF = Extended;
SE = REM;
SavePrm = Yes;
123
SaveCov = Yes;
SaveSco = Yes;
Score = EAP;
<Groups>
%G%
File = "ch9data.dat";
N = 10000;
Varnames = v1-v7,truec;
Select = v1-v7;
Ncats(v1-v7) = 2;
Model(v1-v7) = Graded(2);
Attributes = 3;
InteractionEffects = (2,3); // generate 2nd- and 3rd-order ints.
Dimensions = 7; // 3 main + 3 2nd-order + 1 3rd-order
%D%
DM = G;
Varnames = a;
Ncats(a) = 8;
Model(a) = Nominal(8);
Tc(a) = Identity;
<Constraints>
Fix G,(v1-v7),MainEffect;
Free G,(v1),MainEffect(1); //
Free G,(v2),MainEffect(2); //
Free G,(v3),MainEffect(3); //
Free G,(v4),MainEffect(1); //
Free G,(v4),MainEffect(2); //
Free G,(v4),Interaction(1,2);
Free G,(v5),MainEffect(1); //
Free G,(v5),MainEffect(3); //
Free G,(v5),Interaction(1,3);
Free G,(v6),MainEffect(2); //
Free G,(v6),MainEffect(3); //
Free G,(v6),Interaction(2,3);
Free G,(v7),MainEffect(1); //
Free G,(v7),MainEffect(2); //
1
2
3
1
2
// 1x2
1
3
// 1x3
2
3
// 2x3
1
1
124
Free
Free
Free
Free
Free
G,(v7),MainEffect(3); // 1
G,(v7),Interaction(1,2); // 1x2
G,(v7),Interaction(1,3); // 1x3
G,(v7),Interaction(2,3); // 2x3
G,(v7),Interaction(1,2,3); // 1x2x3
Fix D,(a),Slope;
Fix D,(a),ScoringFn;
Value D,(a),ScoringFn(1),0.0;
As with the previous CDM examples, the maximum number of E-step and
M-steps have been increased from the default, the E tolerance and M tolerance
values have been decreased, and the SE calculation method has been set to
REM in the <Options> section.
The items were simulated as correct/incorrect items, so Ncat= has been
set to 2 and we will fit the 2PLM to all items. As noted in the introduction,
there are 3 attributes measured by the items, so Attributes = 3;. From the
Q-matrix, we can see that Items 4-6 all measure two of the attributes and
Item 7 measures all 3; to accommodate this aspect of the item assignment,
R
we ask flexMIRT
to generate all second and third order effects for us via
the InteractionEffects command. Finally, we set Dimensions = 7, as the
total number of dimensions of the model is 3 main effects + 3 second order
interactions + 1 third order interaction.
R
In the D group, we inform flexMIRT
that the diagnostic model we are
setting up will be applied to group G, using the DM = G; statement. Within
the D group, we specify that the attribute variable will be called a and that it
has 8 categories, reflecting the 8 possible attribute profiles. Additionally, we
R
tell flexMIRT
that the attribute variable will be fit with the nominal model
and we use an identity matrix for the nominal model intercepts.
In the <Constraints> section, we assign items onto attributes according to
the Q-matrix. First, we fix all items to 0 and then free items onto the appropriate main effect and interaction terms. For example, from the Q-matrix we
see that Item 4 is informed by attributes 1 and 2 and the interaction of those
attributes. In the <Constraints> section we free Item 4 with the statements
Free G,(v4), MainEffect(1);, Free G,(v4), MainEffect(2);, and Free
G,(v4),Interaction(1,2);. The final three constraints are applied to the
higher-level group, group D. The first two statements fix the slope and scoring
function values and the final Value statement is used to fix the first scoring
125
function value to 0 for model identification purposes.
R
With the syntax created, we can now run the CDM in flexMIRT
and
turn to the output.
126
127
27
18
21
P#
15
a 1
1.87
0.00
0.00
2.05
1.79
0.00
2.47
s.e.
0.06
------0.17
0.33
---0.22
P#
24
28
19
16
a 2
0.00
2.04
0.00
1.86
0.00
2.15
2.13
s.e.
---0.21
---0.20
---0.12
0.26
22
25
29
17
P#
Prob
0.24810019
0.07355267
0.09927395
0.09119598
0.08634180
0.09134911
0.09358603
0.21660027
P#
s.e.
----
d 1
0
d 2
----
s.e.
----
30
6
0.00
-1.00
s.e.
0.13
s.e.
----
d 3
----
31
23
P# gamma 4
11
-1.06
P# alpha 3
0.00
8
0.00
-0.14
P#
d 4
----
s.e.
---------0.34
------0.42
s.e.
----
a 4
0.00
0.00
0.00
1.03
0.00
0.00
0.81
7
0.00
-0.97
s 8
7.00
P# gamma 3
10
-1.00
5
0.00
-1.06
s.e.
0.43
s 6
5.00
4
0.00
-1.00
P# gamma 2
9
-0.92
s 5
4.00
Original (Bock, 1972) Parameters, Nominal Items for Group 2: D
Item
Label
Category:
1
2
3
1
a
a
0.00
0.00
0.00
c
-0.00
-1.22
-0.92
Nominal Model Intercept Contrasts for Items for Group 2: D
Item
Label Contrasts P# gamma 1
s.e.
1
a Identity
8
-1.22
0.52
s 4
3.00
s 7
6.00
b
----
P#
20
Nominal Model Scoring Function Values Group 2: D
Item
Label
s 1
s 2
s 3
1
a
0.00
1.00
2.00
s.e.
s.e.
------0.14
---0.36
0.11
0.28
P# alpha 2
0.00
a
0.00
1: G
a 3
0.00
0.00
2.01
0.00
1.72
2.10
2.06
Nominal Model Slopes and Scoring Function Contrasts for Items for Group 2: D
Item
Label P#
a
s.e. Contrasts P# alpha 1
s.e.
1
a
0.00
---Trend
1.00
----
Rupp Templin Henson Example 9.2
LCDM model - 7 items, 3 attributes
GPC Items for Group 2: D
Item
Label
1
a
Rupp Templin Henson Example 9.2
LCDM model - 7 items, 3 attributes
Pattern
0 0 0
0 0 1
0 1 0
0 1 1
1 0 0
1 0 1
1 1 0
1 1 1
Diagnostic IRT Attributes and Cross-classification Probabilities for Group
2PL Items for Group 1: G
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
7
v7
Rupp Templin Henson Example 9.2
LCDM model - 7 items, 3 attributes
s.e.
0.31
s.e.
----
d 5
----
s.e.
------------0.58
---0.44
26
32
P# gamma 5
12
-1.00
s.e.
0.37
s.e.
----
s.e.
----
33
P#
P# alpha 5
0.00
d 6
----
s.e.
---------------0.28
0.44
P# gamma 6
13
-0.97
a 6
0.00
0.00
0.00
0.00
0.00
1.11
0.98
s.e.
----
P#
P# alpha 4
0.00
s.e.
----
a 5
0.00
0.00
0.00
0.00
1.38
0.00
0.75
Output 6.6: Rupp, Templin, Henson Output - Item Parameters
s.e.
0.24
s.e.
----
s.e.
0.05
s.e.
----
s.e.
0.05
0.15
0.10
0.11
0.24
0.05
0.10
P# alpha 7
0.00
s.e.
----
c
-0.91
-1.01
-0.96
-2.44
-2.24
-2.54
-3.63
d 8
----
P#
1
2
3
4
5
6
7
P# alpha 6
0.00
s.e.
----
s.e.
------------------0.84
P# gamma 7
14
-0.14
d 7
----
a 7
0.00
0.00
0.00
0.00
0.00
0.00
-1.30
s.e.
----
The output section labeled 2PL Items for Group 1: G presents the item
parameters for the 7 items, with estimated versus fixed at zero loadings mirroring the Q-matrix specifications, with the factors labeled a1-a3 representing
main effects for attributes 1 - 3 and a4 - a7 representing the 2nd and 3rd
R
order interactions. For those wishing to compare flexMIRT
to MPLUS (the
program used to estimate the book example), this item parameter section corresponds with the “New/Additional Parameters” section of MPLUS output
and the point estimates agree with those given in Figure 9.15 (pg. 218) of
Rupp et al. (2010).
The next section of interest is labeled Diagnostic IRT Attributes and
Cross-classification Probabilities for Group 1: G, which reports the
estimated proportions of respondents in each attribute profile. This section in
R
flexMIRT
corresponds to the MPLUS output section labeled FINAL CLASS
COUNTS AND PROPORTIONS FOR THE LATENT CLASSES... and the reported
values match those given in Figure 9.13 of Rupp et al. (2010).
R
There are some difference between flexMIRT
and MPLUS. For example,
MPLUS fixes the latent variable mean of the last latent class (attribute profile 8
R
in this case) to 0, while in flexMIRT
we have constrained the first latent class
to 0. This results in differences in the latent attribute parameter estimates,
but as we show in the below table these differences are simply a matter of
choice of identification constraints and do not affect the primary estimates of
interests (i.e., estimated profile proportions).
R
Table 6.5: Generating latent class mean values and flexMIRT
and MPLUS
estimates for Rupp, Templin, & Henson’s Example 9.2
c1
c2
c3
µ
exp(µ)
υ
0
1
0.2481
-1.00
0.367879
0.0913
-1.00
0.367879
0.0913
µˆ
exp(µ)
ˆ
υˆ
0
1
0.2481
-1.21583
0.296464
0.0736
-0.91595
0.400136
0.0993
µˆ
exp(µ)
ˆ
υˆ
0.136
1.145682
0.2481
-1.08
0.339596
0.0736
-0.78
0.458406
0.0993
c4
c5
generating
-1.00
-1.00
0.367879
0.367879
0.0913
0.093
flexMIRT
-1.00082
-1.05552
0.367578
0.348011
0.0912
0.0863
MPLUS
-0.865
-0.92
0.421052
0.398519
0.0912
0.0863
c6
c7
c8
sum
-1.00
0.367879
0.0913
-1.00
0.367879
0.0913
0
1
0.2481
4.207277
-0.99914
0.368196
0.0914
-0.97495
0.377211
0.0936
-0.13578
0.87303468
0.2166
4.030631
-0.863
0.421894
0.0914
-0.839
0.432142
0.0936
0
1
0.2166
4.617291
We will next cover the contents of the requested file that contains the
individual scores.
128
129
1
1
0.300930
1
2
0.875910
1
3
0.481882
1
4
0.875910
1
5
0.029570
1
6
0.000006
1
7
0.000007
1
8
0.228668
1
9
0.071059
1
10
0.001455
0.162486
0.019853
0.598993 |
0.368896
0.139494
0.490102 |
0.136084
-0.002661
0.019459
-0.018766
0.517296
0.016153
0.003135
0.083426
0.078495
0.000499
0.000066 2
0.000000
1.000000
1.000000
0.041052
0.047476
0.036249 |
0.198411
0.212654
0.186910 |
0.039367
-0.001706
0.045222
-0.001260
0.035804
0.047016
0.000217
0.040582
0.000228
0.000243
0.000000 1
0.000000
1.000000
1.000000
0.198872
0.190374
0.160731 |
0.399151
0.392596
0.367282 |
0.159322
-0.019122
0.154132
-0.024583
0.147611
0.165898
0.005738
0.172862
0.007272
0.018627
0.000110 1
0.000000
1.000000
1.000000
0.041052
0.047476
0.036249 |
0.198411
0.212654
0.186910 |
0.039367
-0.001706
0.045222
-0.001260
0.035804
0.047016
0.000217
0.040582
0.000228
0.000243
0.000000 1
0.000000
1.000000
1.000000
0.475195
0.091060
0.900778 |
0.499384
0.287695
0.298960 |
0.249385
-0.031539
0.082768
-0.008865
0.415907
0.013637
0.065691
0.053279
0.410184
0.002736
0.008996 2
0.000000
1.000000
1.000000
0.880070
0.979322
0.977730 |
0.324880
0.142303
0.147561 |
0.105547
-0.001801
0.020250
-0.001495
0.000673
0.001170
0.118081
0.000154
0.019844
0.020940
0.839131 8
0.000000
1.000000
1.000000
0.981067
0.979636
0.846212 |
0.136289
0.141243
0.360745 |
0.018575
-0.000279
0.019950
-0.001603
0.000100
0.001301
0.017525
0.001116
0.019142
0.151364
0.809445 8
0.000000
1.000000
1.000000
0.150848
0.679496
0.012669 |
0.357900
0.466670
0.111843 |
0.128093
-0.034142
0.217781
-0.001397
0.009347
0.608329
0.002808
0.082029
0.000460
0.068305
0.000054 3
0.000000
1.000000
1.000000
0.751563
0.081414
0.638255 |
0.432107
0.273469
0.480505 |
0.186716
-0.029184
0.074785
0.013218
0.127968
0.032030
0.017380
0.232326
0.487234
0.026330
0.005674 6
0.000000
1.000000
1.000000
0.143402
0.935046
0.910959 |
0.350483
0.246444
0.284802 |
0.122838
-0.025556
0.060735
-0.031276
0.028628
0.043541
0.782973
0.005167
0.029704
0.038878
0.069654 4
-0.000000
1.000000
1.000000
Output 6.7: Rupp, Templin, Henson Output - -sco File
0.134896 \\
0.034935 \\
0.089377 \\
0.021774 \\
0.130137 \\
0.012509 \\
0.230886 \\
-0.024751
-0.001504
-0.007338
-0.000300
-0.002010
-0.005747
-0.028909
0.081112 \\
0.034935 \\
-0.001504
0.000838
0.240200 \\
-0.008690
As in the previous -sco output, we have given the first 10 cases and have
added a line break and separators. The first two columns of the -sco file give
R
the group and observation number, similar to a standard flexMIRT
IRT score
output file. Recall that for this example, there are 3 attributes and 8 possible
attribute profiles. The probability of the respondent having mastered each
of the attributes individually is given, starting after the observation number.
For example, for observation 1, the probability they have mastery of attribute
1 is 0.16, for attribute 2 it is 0.02 and for attribute 3 it is 0.60. The next
three columns are the estimated SEs for those probabilities, respectively. The
last six entries on the first line of each observation contain the entries for the
estimated attribute variance/covariance matrix.
Finally, on what has been moved to the second line for each respondent (but
is a continuation of the same in the actual -sco file), the posterior probabilities
for each of the attribute profiles (8 possible for this example) are given. For
observation 1, the posterior probability of being in attribute profile 1 is 0.30,
the probability of being in attribute profile 2 is 0.52 and so on. The order in
which the probabilities are given matches the order used in the output table
“Diagnostic IRT Attributes and Cross-classification Probabilities.” Referring
to the output presented for this example, we find that attribute profile 1
corresponds to 0 0 0 (meaning no attributes are mastered) and attribute profile
2 corresponds to 0 0 1 (meaning only attribute 3 has been mastered), etc..
After the probabilities for each of the attribute profiles are presented, the
most likely attribute profile for the respondent is given; for respondent 1 that
is attribute profile 2, for respondents 6 and 7, that is attribute profile 8. The
last three values given in the -sco file are the higher-order latent variable point
estimate followed by the SD and variance for that estimate. In this case, all
individual theta estimates for the higher-order variable are zero and all SD
and variance estimates are 1.00 because the nominal model was fit, which has
zero slope and is a reparameterization of the profile probabilities.
Unstructured Higher-Order Latent Space
R
It is also possible to the fit CDMs in flexMIRT
without using the higherorder latent variable at all. This is accomplished by removing all references to
the D group in setting up the model. We present this syntax below. The only
changes are the removal of the %D% group from the <Groups> section as well
as omitting the parameter constraints that were applied to that group.
130
Example 6.7: Rupp, Templin, & Henson - Example 9.2 without D group
<Project>
Title = "Rupp Templin Henson Example 9.2 redone";
Description = "LCDM model - 7 items, 3 attributes:";
<Options>
Mode = Calibration;
MaxE = 20000;
Etol = 1e-6;
MaxM = 50;
Mtol = 1e-9;
GOF = Extended;
SE = REM;
SavePrm = Yes;
SaveCov = Yes;
SaveSco = Yes;
Score = EAP;
<Groups>
%G%
File = "ch9data.dat";
N = 10000;
Varnames = v1-v7,truec;
Select = v1-v7;
Ncats(v1-v7) = 2;
Model(v1-v7) = Graded(2);
Attributes = 3;
InteractionEffects = (2,3); // generate 2nd- and 3rd-order ints.
Dimensions = 7; // 3 main + 3 2nd-order + 1 3rd-order
<Constraints>
Fix G,(v1-v7),MainEffect;
Free G,(v1),MainEffect(1); //
Free G,(v2),MainEffect(2); //
Free G,(v3),MainEffect(3); //
Free G,(v4),MainEffect(1); //
Free G,(v4),MainEffect(2); //
Free G,(v4),Interaction(1,2);
1
2
3
1
2
// 1x2
131
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
G,(v5),MainEffect(1); // 1
G,(v5),MainEffect(3); // 3
G,(v5),Interaction(1,3); // 1x3
G,(v6),MainEffect(2); // 2
G,(v6),MainEffect(3); // 3
G,(v6),Interaction(2,3); // 2x3
G,(v7),MainEffect(1); // 1
G,(v7),MainEffect(2); // 1
G,(v7),MainEffect(3); // 1
G,(v7),Interaction(1,2); // 1x2
G,(v7),Interaction(1,3); // 1x3
G,(v7),Interaction(2,3); // 2x3
G,(v7),Interaction(1,2,3); // 1x2x3
With respect to the output, presented next, while there are some slight
changes in the estimates, the reported point estimate values for the item parameters remain largely unchanged from the calibration that employed the
higher-order latent variable. Differences in the estimated profile probabilities
are not seen until the 4th or 5th decimal place. If one looked at the -sco file,
there are also only slight changes to the reported values; the most noticeable
difference between the -sco file from the model employing the higher-order
variable and the current example is that final three values discussed in the
previous -sco file (the higher-order latent variable mean, SD, and variance
estimates) are no longer present.
132
133
20
11
14
P#
8
a 1
1.87
0.00
0.00
2.05
1.79
0.00
2.47
s.e.
0.07
------0.18
0.17
---0.57
17
21
12
9
P#
a 2
0.00
2.04
0.00
1.86
0.00
2.15
2.13
s.e.
---0.07
---0.18
---0.19
0.58
15
18
22
10
P#
Pattern
0 0 0
0 0 1
0 1 0
0 1 1
1 0 0
1 0 1
1 1 0
1 1 1
Prob
0.24809775
0.07355411
0.09927471
0.09119567
0.08634185
0.09134873
0.09358578
0.21660141
Diagnostic IRT Attributes and Cross-classification Probabilities for Group
2PL Items for Group 1: G
Item
Label
1
v1
2
v2
3
v3
4
v4
5
v5
6
v6
7
v7
1: G
a 3
0.00
0.00
2.01
0.00
1.72
2.10
2.06
s.e.
------0.07
---0.17
0.20
0.64
23
13
P#
a 4
0.00
0.00
0.00
1.03
0.00
0.00
0.81
s.e.
---------0.31
------0.80
24
16
P#
a 5
0.00
0.00
0.00
0.00
1.38
0.00
0.75
s.e.
------------0.32
---0.85
19
25
P#
a 6
0.00
0.00
0.00
0.00
0.00
1.11
0.98
s.e.
---------------0.37
0.85
Output 6.8: Rupp, Templin, Henson Output - Item Parameters without D group
26
P#
a 7
0.00
0.00
0.00
0.00
0.00
0.00
-1.30
s.e.
------------------1.32
P#
1
2
3
4
5
6
7
c
-0.91
-1.01
-0.96
-2.44
-2.24
-2.54
-3.63
s.e.
0.04
0.04
0.04
0.13
0.11
0.14
0.43
de la Torre Subtraction Example
In a demonstration of a higher order latent trait CDM to non-simulated data,
de la Torre and Douglas (2004) fit a higher-order DINA model to a 20-item
fraction subtraction test given to 2144 examinees. In the noted paper, an 8
attribute model was estimated using MCMC, with final parameter estimates
based on averaging the estimates from 10 parallel chains, each with 20000
iterations of which the first 10000 were discarded as burn-in. We provide
R
syntax and output of a flexMIRT
replication of this analysis.
Example 6.8: de la Torre Subtraction Syntax
<Project>
Title = "Tatsuoka Subtraction Data";
Description = "Restricted Higher-order DINA Model";
<Options>
Mode = Calibration;
MaxE = 20000;
MaxM = 5;
Mtol = 0;
Etol = 1e-5;
GOF = Extended;
SE = REM;
SavePrm = Yes;
SaveCov = Yes;
SaveSco = Yes;
Score = EAP;
Processors = 4;
NewThreadModel = Yes;
<Groups>
%G%
File = "subtraction.csv";
Varnames = v1-v20;
Ncats(v1-v20) = 2;
Model(v1-v20) = Graded(2);
Attributes = 8;
Generate = (3,6,7),(4,7),(2,3,5,7),(2,4,7,8),(1,2,7),(2,5,7,8),
134
(2,5,7),(7,8),(2,4,5,7), (2,7),(1,7),
(2,5,6,7),(1,2,3,5,7),(2,3,5,7);
Dimensions = 22; // 8 main effects + 14 generated higher ord ints
%D%
DM = G;
Varnames = a1-a8;
<Constraints>
Fix G,(v1-v20),MainEffect;
Free G,(v1),Interaction(3,6,7); //3rd-order int of attr 3,6,7
Free G,(v2),Interaction(4,7); //2nd-order int of attr 4,7
Free G,(v3),Interaction(4,7); //2nd-order int of attr 4,7
Free G,(v4),Interaction(2,3,5,7);//4th-order int of attr 2,3,5,7
Free G,(v5),Interaction(2,4,7,8);//4th-order int of attr 2,4,7,8
Free G,(v6), MainEffect(7); //main effect of attr 7
Free G,(v7),Interaction(1,2,7); //3rd-order int of attr 1,2,7
Free G,(v8), MainEffect(7); // main effect of attr 7
Free G,(v9), MainEffect(2); // main effect of attr 2
Free G,(v10),Interaction(2,5,7,8); //4th-ord int of attr 2,5,7,8
Free G,(v11),Interaction(2,5,7); //3rd-order int of attr 2,5,7
Free G,(v12),Interaction(7,8); // 2nd-order int of attr 7,8
Free G,(v13),Interaction(2,4,5,7); //4th-ord int of attr 2,4,5,7
Free G,(v14),Interaction(2,7); //2nd-order int of attr 2,7
Free G,(v15),Interaction(1,7); //2nd-order int of attr 1,7
Free G,(v16),Interaction(2,7); //2nd-order int of attr 2,7
Free G,(v17),Interaction(2,5,7); //3rd-order int of attr 2,5,7
Free G,(v18),Interaction(2,5,6,7); //4th-ord int of attr 2,5,6,7
Free G,(v19),Interaction(1,2,3,5,7);//5th-ord int: att 1,2,3,5,7
Free G,(v20),Interaction(2,3,5,7); //4th-order int: att 2,3,5,7
Equal D,(a1-a8),Slope; // "restricted" higher order model
As with the previous CDM example, the maximum number of E-steps
has been increased from the default, the E tolerance and M tolerance values have been decreased, and the SE calculation method has been set to
REM in the <Options> section. As noted earlier, these changes are made
primarily to improve the resulting SE estimates. We have also requested
the newly implemented multi-core processing model be used for this analysis via NewThreadModel = Yes;. This model is requested due to the high135
dimensional nature of the current problem. Additional information regarding
the new thread model is available in the Syntax chapter.
In the first group section, have used the more targeted Generate to construct only those interaction terms which will be used in the model. The
needed interaction effects are generated in the order they will be used for the
items (e.g., item 1 requires the interaction of attributes 3, 6, and 7) but this
not necessary - as noted earlier, as many interactions as needed may be constructed from a single Generate statement and the order in which interactions
are listed does not matter. Counting the interactions specified in the Generate
statement, we find there are 14 interactions needed to correctly assign items
to correspond to the Q-matrix. Adding the 14 interactions with the 8 main
effects gives us a total of 22 latent dimensions in the model.
In the <Constraints> section, all main effect terms are initially fixed to 0
and then Free statements are added to replicate the Q-matrix for the problem
(which may be found in Table 8 on pg. 347 of de la Torre & Douglas, 2004).
Because a DINA is fit to all items, only the highest-order interaction of attributes for each item is specified and main effect terms are specified for only
those items that depend on a single attribute (i.e., items 6, 8, and 9). The
last statement in the <Constraints> section sets the latent attribute slope
parameters to be equal across attributes, producing a restricted higher-order
model.
As with all examples, the full output file is available on the program’s
Support site. The item parameter table for this example is extremely wide
(22 latent dimension columns, an intercept column, and all associated SEs)
and is not presented here due to size. However, we have summarized the
R
flexMIRT
parameter estimates in Table 6.6. Please note that the slopes
are presented in a single column in this table for efficiency of space; the estimated slope values are not all on the same latent dimension, as can be
R
seen in the full output file. In addition to the flexMIRT
estimates, we also
present the original slippage and guessing parameter values reported by de la
R
Torre and Douglas (2004) in their Table 9 and the flexMIRT
estimates converted to this parameterization. Comparing the point estimates across the two
programs, one notices some differences but overall the values are quite similar. Regressing the original guessing parameter estimates onto the converted
R
flexMIRT
values gives an intercept value close to zero (0.006), a regression
coefficient on 0.97, and an R2 value of 0.99. A similar comparison of the slippage values also results in a low intercept (0.01), a regression coefficient near 1
(0.99), and a high R2 value (0.98). Given the comparability of the parameter
136
R
completed its calibration in
values, it is interesting to note that flexMIRT
20.5 seconds, highlighting the extreme efficiency that may be obtained when
estimating higher-order CDMs via MML.
Table 6.6: Guessing and Slippage Parameters from flexMIRT and MCMC
de la Torre
& Douglas
flexMIRT
Item
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
v11
v12
v13
v14
v15
v16
v17
v18
v19
v20
Intercept (c)
-4.59
-3.42
-4.49
-1.25
-0.84
-4.99
-3.86
-0.20
-1.63
-3.51
-2.66
-1.91
-4.30
-3.20
-3.43
-2.25
-3.00
-1.81
-3.82
-4.14
Slope (a)
6.50
6.90
6.47
3.21
2.35
8.09
5.25
1.65
2.75
4.82
5.24
4.74
5.02
5.79
5.51
4.23
4.83
3.66
4.87
5.68
g
1-s
0.01 0.87
0.03 0.97
0.01 0.88
0.22 0.88
0.30 0.82
0.01 0.96
0.02 0.80
0.45 0.81
0.16 0.75
0.03 0.79
0.07 0.93
0.13 0.94
0.01 0.67
0.04 0.93
0.03 0.89
0.10 0.88
0.05 0.86
0.14 0.86
0.02 0.74
0.02 0.82
g
0.04
0.03
0.00
0.22
0.30
0.03
0.03
0.44
0.18
0.03
0.07
0.13
0.02
0.05
0.04
0.10
0.04
0.13
0.02
0.01
1-s
0.90
0.96
0.88
0.89
0.82
0.96
0.81
0.81
0.75
0.79
0.93
0.96
0.67
0.93
0.90
0.88
0.86
0.85
0.76
0.84
Note: g is found from flexMIRT parameters as exp(c)/(1 +
exp(c)) and 1 − s is found as exp(c + a)/(1 + exp(c + a)).
137
CHAPTER 7
Details of the Syntax
This section provides detailed coverage of all the syntax statements and asR
sociated options that are available in flexMIRT
. Required statements are
listed in bold, optional statements are in normal font. When keywords are
required with statements, the possible options are listed after the statement
and the default setting is underlined. If numeric values are required with a
statement, such as for a convergence criterion, the default value is presented.
If there is no default value, a question mark will denote where values need be
entered.
7.1. The Project Section
Example 7.1: <Project> section commands
Title = “ ”;
Description = “ ”;
Both of these statements must be present in the <Project> section, although the actual content within the quotes may be left blank.
7.2. The Options Section
The <Options> section is where the type of analysis to be conducted is specified and where technical details of the analysis may be modified, including
convergence criteria, scoring methods, and the level of detail desired in the
output. As indicated by the bold, the Mode statement is the only required
command and determines the type of analysis that will be conducted. If Mode
= Scoring; or Mode = Simulation; is selected, some additional technical
commands, covered in the next group of statements become required. The
R
setting of Progress determines if flexMIRT
will print detailed progress in138
formation in the console window, (e.g., iteration number and log-likelihood
value). TechOut determines if controls values such as tolerance values, processing times, names of outputted files, etc. will be printed in the preamble of
the output. NumDec determines the number of decimal places reported for item
parameters in the output file - the default is 2. The other commands allow for
the optional saving of additional output into separate files. SaveCov refers to
the covariance matrix of the parameter estimates; SaveInf, the Fisher information function values of the items and the test; SavePrm, the item and group
parameter estimates; SaveSco, the individual IRT scale scores, and SaveDbg,
additional technical information. Each of the these additional output requests
are saved to files with corresponding extensions; that is, the SaveInf command results in an output file with “*-inf.txt” appended to the command file
name; SaveCov appends “*-cov.txt”, etc.
Example 7.2: <Options> - Engine Mode and Output Statements
<Options >
Mode = Calibration/Scoring/Simulation;
Progress = Yes/No;
TechOut = Yes/No;
NumDec = 2;
SaveCov = Yes/No;
SaveInf = Yes/No;
SavePrm = Yes/No;
SaveSco = Yes/No;
SaveDbg = Yes/No;
DMtable = Yes/No;
NormalMetric3PL = Yes/No;
SlopeThreshold = Yes/No;
Specific to CDMs, the command DMtable controls whether the diagnostic
classification probability table is printed in the *-irt file. The default is to print
the table, but this may be modified by changing the command to DMtable=
No;. Modifying this setting may be helpful if the probability table will be
large.
The final two statements in this section are provided primarily to facilitate the use of parameter estimates obtained from other programs, but users
may find them convenient for other purposes. Both NormalMetric3PL and
SlopeThreshold are used to print, save out, or read-in parameters that are
139
in more traditional IRT metrics than the MIRT-compatible logistic metric
R
flexMIRT
uses; these keywords are only available for use in conjunction with
unidimensional models.
When NormalMetric3PL = Yes;, the program will print normal-metric
parameter estimates in the output and save those values into the -prm file, if
R
will 1) print/save (a/1.702) rather than
requested. Specifically, flexMIRT
the typical logistic metric slope values, 2) instead of the intercept (b ∗ −a)
R
flexMIRT
will print/output the difficulty (b) parameter value, and 3) rather
than the logit-guessing parameter, the program will report/save the customary
g value. Note that even if the normal metric keyword is used during calibration,
any priors that may be specified are applied to the MIRT-compatible metrics
(prior applied to intercept, not difficulty; normal guessing prior still applied to
logit-g distribution, etc). If NormalMetric3PL is set to Yes for a scoring run,
R
then flexMIRT
will expect the all parameter values read-in from the -prm
file for scoring to be in the normal metric.
R
By default, flexMIRT
prints both the intercept and difficulty parameters
in the output for undimensional models, but saves only the intercept values
into a requested -prm file. When SlopeThreshold = Yes;, rather than saving
the intercept values, the program will write the difficulty (b) values to the prm file. If SlopeThreshold is set to Yes in conjunction with a 3PL model,
the psuedo-guessing parameters (gs) will be printed/saved, rather than the
default logit-guessing parameter values. Users should be aware that, unlike
the behavior when NormalMetric3PL is used, the SlopeThreshold keyword
has no effect on the metric of the slope; a values printed to the -prm are in
the logistic metric, regardless of the SlopeThreshold setting.
140
Example 7.3: <Options> - Technical Specifications
<Options >
...
Score = EAP/MAP/ML/SSC;
MaxMLscore = ?;
MinMLscore = ?;
Mstarts = Yes/No;
Rndseed =?;
ReadPRMFile = "*.txt";
SE = Xpd/Mstep/SEM/Sandwich/Fisher/FDM/REM;
SmartSEM = Yes/No;
Perturbation = 0.001;
FisherInf = 1, 0.0;
PriorInf = Yes/No;
logDetInf = Yes/No;
Quadrature = 49, 6.0;
MaxE = 500;
MaxM =100;
Etol = 1e-4;
Mtol = 1e-9;
SEMtol =1e-3;
SStol =1e-4;
Processors = 1;
NewThreadModel = Yes/No;
As detailed previously, when Mode is set to something other than Calibration,
additional statements are required. It was noted earlier that specifying SaveSCO
= Yes during a calibration run triggers a combined scoring run. If a scoring
run is specified, either by explicitly setting Mode to Scoring or implicitly by
requesting a combined scoring run, an IRT scoring method is required. As may
be deduced from the lack of an underlined scoring option, there is no default.
The options available for scoring are Expected A Posteriori (keyword is EAP),
Maximum A Posteriori (keyword is MAP), Maximum Likelihood (keyword is
ML), and Summed Score Conversions to EAP scores (keyword is SSC), which
in addition to individual scores provides a conversion table in the output that
translates summed scores to EAP scale scores.
When using ML scoring, there are two additional keywords that must be
employed. Users are required specify the desired maximum and minimum
141
R
using the MaxMLscore and MinMLscore
scores to be assigned by flexMIRT
keywords, respectively. There are no default values for the ML minimum
and maximum scores. Efforts have been made to make ML scoring as
robust as possible, but it is not recommended for use with multidimensional models. Because ML scoring does not use information
from the population distribution, essential statistical information
about the population distribution (e.g., factor inter-correlations,
means, and variances) is ignored when scoring individual dimensions of MIRT models. Additionally, ML scoring can lead to score
information matrices that are not positive definite, making the SEs
for some estimates undefined.
R
flexMIRT
offers a multiple dispersed starting value option for use with
ML and MAP scoring runs. For unidimensional models, the default theta startR
ing value in MAP/ML scoring is 0. When Mstarts = Yes; flexMIRT
will do
12 additional restarts using a range of starting values from -2.75 to 2.75, in step
sizes of 0.5. The best solution (as determined by the highest log-likelihood)
will be picked from the 13 values. Note that using Mstarts = Yes; will increase the time needed to complete scoring, but provides protection against
local modes in 3PL or nominal model scoring.
When a Simulation run is specified by the Mode command, the random
number generator must be “seeded”. An integer value must be explicitly provided using Rndseed command. There is no default value.
The ReadPRMFile command is used to read parameters values from a text
R
file to flexMIRT
, which may be used as starting values for a calibration run,
as fixed item parameters for a scoring run, and as true generating parameter
values for a simulation run. The proper layout for a parameter file is covered
in detail in Chapter 4: Simulation.
R
flexMIRT
has several options available for computing the standard errors of estimated item/group parameters. The standard error method may be
changed by the user through the SE command. The default method is empirical cross-product approximation (keyword is Xpd). Standard errors may also
be calculated via the supplemented EM algorithm (keyword is SEM; Cai, 2008),
from the Fisher (expected) information matrix (keyword is Fisher), via the
sandwich covariance matrix (keyword is Sandwich), using the forward difference method (keyword is FDM; e.g., Jamshidian & Jennrich, 2000), or from the
Richardson extrapolation method (keyword is REM; e.g., Jamshidian & Jennrich, 2000). Mstep values from the EM algorithm can be requested although,
technically, the M-step standard errors are incorrect. They tend to be too
142
small, but under some circumstances (e.g., huge calibration sample size and
thousands of item parameters), it may be more efficient to bypass the more
elaborate (and correct) standard error computations and focus only on item
parameter point estimates.
As an additional note, three of the available standard error methods have
additional options that may be adjusted by the user. By default the supplemented EM algorithm SEs come from an optimized window of iterations. If
you find that the SEs are failing to converge in that optimized window, the
optional command SmartSEM may be useful. By setting SmartSEM = No; it
R
will allow flexMIRT
to use the full EM iteration history, which may lead to
a better window and converged standard errors.
The Perturbation command controls a value used by both the forward
difference and Richardson extrapolation methods of SE estimation. These
methods require the selection of a perturbation constant that is used in the
denominator of the finite difference approximation; the choice of this perturbation constant may exert substantial influence on the accuracy of the final item
parameter SE estimates (Tian, Cai, Thissen, & Xin, 2012). As noted above,
R
the flexMIRT
default of this perturbation constant is set to 0.001, the value
found to produce the consistently least poor estimates from the two methods
when studied in both simulations and as applied to real data (Tian et al.,
2012). While the forward difference and Richardson extrapolation methods
are available to users, Tian et al. (2012) found them to be less accurate than
the supplemented EM algorithm SEs when applied to standard IRT models,
with no noticeable savings with respect to time when the smart window for
the supplemented EM SEs was employed.
The information function for the individual items and full scale is printed
in the general output (i.e., the “*-irt.txt” file) using theta values from -2.8 to
2.8, changing by steps of 0.4. The SaveInf option addressed earlier allows the
user to output the information function values to a separate “*-inf.txt” file,
R
which is useful for additional computations external to flexMIRT
or when
information functions are plotted. By default, the “*-inf.txt” file has only one
theta value (at 0.0) saved. The FisherInf command allows the number of
points and the symmetric minimum/maximum values of those points to be
modified. For instance, if information values were desired for 81 points with
theta values ranging from -4.0 to 4.0, the command to obtain this output,
in conjunction with SaveInf = Yes;, would be FisherInf = 81, 4.0;, with
the total number of points listed first and the maximum theta value to be
used listed after comma. Using the PriorInf command, the user controls if
143
R
includes the contribution from the prior in the test information
flexMIRT
output - the default setting of Yes includes the prior contribution.
R
Note that for multidimensional models, flexMIRT
will try to print, by default, the trace of the Fisher information matrix calculated at every grid point
as defined by the direct product of the points specified in the FisherInf command. For high dimensional models, this is not advisable. If the logDetInf =
Yes; command is used, the information will be taken as the log determinant
of the Fisher information matrix, rather than the trace.
The Quadrature command, which can modify the number of quadrature
points used in the E step of the EM algorithm, is specified in a similar fashion
as the FisherInf command. The default quadrature rule has 49 rectangular
points over -6 to +6.
The next several statements listed within the technical specifications deal
with convergence criteria values and maximum number of iterations. MaxE
determines the maximum allowed number of E-steps in the EM algorithm,
MaxM the number of allowable iterations in each of the iterative M-steps. Etol
and Mtol set the convergence criteria for the E- and M- steps, respectively.
SEMtol sets the convergence criterion for the Supplemented EM algorithm used
in calculating standard errors, and finally SStol determines the criterion for
declaring two summed scores as equivalent (useful for scoring with weights).
The Processors statement lets the user determine how many processors
(threads) an analysis can utilize. When NewThreadModel = Yes; a new multicore processing model is employed. The new thread model is an updated implementation of multi-core processing that may be more efficient for very highdimensional models with a large number of quadrature points to evaluate. The
level of granularity is smaller than the classical threading model implemented
in flexMIRT; a classical threading model trades more memory for speed, and
this new model does not do that nearly as much, so the parallel speed up may
not be as significant as the classical approach for low dimensional analyses with
large N. User are encouraged to experiment with the threading options. When
the NewThreadModel is active, Parallelization granularity: Fine (may
be more optimal for high dimensionality) will be printed in output with
other control values. Coarse will be printed when the original threading model
is used.
144
Another subset of commands available in the <Options> section deals with
the calculation and reporting of various GOF and local dependence (LD) indices.
Example 7.4: <Options> - GOF and LD Indices Commands
<Options >
...
GOF = Basic/Extended/Complete;
CTX2Tbl = Yes/No;
SX2Tbl = Yes/No;
MinExp = 1.0;
M2 = None/Full/Ordinal/OrdinalSW;
HabermanResTbl = Yes/No;
JSI = Yes/No;
FactorLoadings = Yes/No;
FitNullModel = Yes/No;
StartValFromNull = Yes/No;
...
The GOF statement controls the extent of GOF indices that are reported.
Basic, Extended, and Complete are the three possible levels of GOF reporting. Basic, the default value, prints the −2× log likelihood, the AIC/BIC
values, and when appropriate, the likelihood ratio (G2 ) and Pearson X 2 fullinformation fit statistics. The Extended option results in all the indices included in the Basic reporting, as well as the marginal fit X 2 for each item
and standardized LD X 2 for each item pair (e.g., Chen & Thissen, 1997). The
last keyword, Complete, prints all indices included in the Extended reporting
and also includes, per group, the item-fit index S − X 2 (Orlando & Thissen,
2000). Additionally, under GOF = Complete; mode, an additional statistic
S − D2 is printed below the SSC to EAP conversion table. This statistic is
distributed approximately as a central X 2 variable, with degrees of freedom
equal to [(number of summed scores - 1) - 2], and provides a test of latent
variable distribution assumption.
In both the Extended and Complete GOF modes, the standardized LD
2
X statistics for each item pair may be suppressed by setting the CTX2Tbl
command to No. As noted in the discussion of the LD statistic values of
Chen and Thissen (1997) when applied to ordinal data (i.e., Example 2.5),
phi correlations are examined to determine if difference between the observed
145
and model implied item-correlation is positive or negative. When polytomous
data are involved, there may be times that the contingency table of two items
needs to be collapsed from the edges to the middle so that the cells are not
too sparse. Even if collapsing occurs for the Chen-Thissen LD calculations,
the phi correlations (determining the ”p” or ”n” after each LD value), will be
based on the original (non-collapsed) tables.
The SX2Tbl command is invoked when the user wishes to be able to examine the detailed item fit tables used in the calculation of the S − X 2 statistics.
As described in Orlando and Thissen (2000), when expected frequencies for
responses become small, the accuracy of the X 2 approximation deteriorates.
R
flexMIRT
employs an algorithm to evaluate the expected probabilities prior
to final calculations and collapse the table when needed. The MinExp command allows the user to set the value at which expected counts are considered
too small, (the default value for MinExp is set at 1.0) and when expected counts
lower than the value set by MinExp are encountered, the summed score groups
are collapsed towards the center.
The M2 statement requests that the limited-information fit statistics M2 ,
first introduced by Maydeu-Olivares and Joe (2005) (Full) and expanded
in Maydeu-Olivares, Cai, and Hernandez (2011) (OrdinalSW) and Cai and
Hansen (2012) (Ordinal) be printed. The None option suppresses M2 . It
should be noted that, regardless of M2 request, the M2 values will not be
printed unless GOF is set to something other than Basic.
The command HabermanResTbl is invoked when standardized residuals
for each response pattern, sometimes referred to as Haberman Residuals (see
Haberman, 1979), are desired. In addition to the standardized residuals, the
observed and expected frequencies, as well as the associated EAPs and SD for
each response pattern are also reported in the requested table.
R
There is an additional LD index that flexMIRT
can compute with the
command JSI. The statement JSI = Yes; requests that the Jackknife Slope
Index (JSI; Edwards & Cai, 2011), a recently introduced LD detection technique, be calculated. Based on the observation that locally dependent items
often exhibit inflated slopes, the JSI procedure finds, for each item pair, the
change in the slope parameter of item j when item k is removed from the scale.
Specifically, a single JSI values is obtained by
JSIj (k ) =
aj − aj ( k )
,
se(aj (k ) )
(7.1)
where a is the IRT slope estimate, j indexes the item impacted and k indexes
146
the removed item and se(aj (k ) ) is the SE of the item-removed slope parameter.
The resulting n item by n item matrix, with empty diagonals, is inspected by
the user and item pairs with JSI values substantially larger than the other
values should be noted as possibly exhibiting LD.
The FactorLoadings statement may be invoked when, in addition to the
IRT parameters, normal metric factor loadings should be printed as part of
the general output file. The factor loading values are converted from the IRT
parameters, as described in Wirth and Edwards (2007), among others.
The final two statements of this group are tied to the Null (independence)
model, needed for incremental fit indices. The FitNullModel command, when
R
invoked with the keyword Yes, tells flexMIRT
to fit the Null model and print
the resulting basic GOF measures in the general output file. When the fitting of the Null model is requested, it becomes possible for the Tucker-Lewis
Index to be calculated. Additionally, when the zero-factor Null model is requested, the resulting parameter values from fitting this model may be used
as starting values for the IRT model of interest by supplying the command,
StartValFromNull = Yes;. Obviously, it is necessary that the Null model is
requested, via the FitNullModel command, for the StartValFromNull command to be functional.
147
Example 7.5: <Options> - DIF Analysis Commands
<Options >
...
DIF = None/TestAll/TestCandidates;
DIFcontrasts = (...);
OrthDIFcontrasts = Yes/No;
...
Within the <Options> section, there are three commands related to the
implementation of DIF analyses. The first command is a request for a DIF
analysis, DIF. By default, no DIF analysis is conducted, but this may be
modified by the user to request either a comprehensive DIF sweep (keyword is
TestAll), testing all items, or a focused DIF analysis which looks for DIF only
in the candidate items supplied by the user (keyword is TestCandidates).
If the TestCandidates option is selected, the DIFitems command, found
in the <Groups> section is required. When a DIF analysis is requested by
invoking one of the DIF options, DIFcontrasts becomes a required command.
R
This command is used to supply flexMIRT
with the DIF contrast matrix,
which is used in constructing the contrasts among groups. The total number
of contrasts is one fewer than the total number of groups being analyzed. For
a 2 group DIF analyses, that means there is one allowable contrast, which will
typically be of the form DIFcontrasts = (1 -1);. For analyses with more
than 2 groups, any ANOVA contrasts can be specified. The Helmert contrast
(which is orthogonal) is a reasonable choice, (e.g., for 3-groups:
DIFcontrasts = (
2.0 -1.0 -1.0,
0.0 1.0 -1.0);
When group size is highly unbalanced, OrthDIFcontrasts offers additional
re-weighting of the contrast coefficients such that they become orthogonal with
respect to the group sizes (Langer, 2008). This option is turned off by default.
148
Finally, there are several commands that are specific to or used primarily
by the MH-RM algorithm.
Example 7.6: <Options> - MH-RM Specific Options
<Options >
...
Algorithm = BAEM/MHRM;
Imputations = 1;
Thinning = 10;
Burnin = 10;
Stage1 = 200;
Stage2 = 100;
Stage3 = 2000;
Stage4 = 0;
InitGain = 1.0;
Alpha = 1.0;
Epsilon = 1.0;
WindowSize = 3;
ProposalStd = 1.0;
ProposalStd2 = 1.0;
MCsize = 25000;
...
To call the MH-RM algorithm, the combination of Mode = Calibration;,
Algorithm = MHRM;, and a random number seed via RndSeed = ; is required.
All other listed commands are optional, with the default values shown above.
Imputations controls the number of imputations from the MH step per
RM cycle. A larger value will lead to smoother behavior from the algorithm.
In most cases, a value of 1 should be sufficient for obtaining accurate point
estimations, however this number may be increased for better standard errors.
Thinning refers to the MCMC practice of retaining only every kth draw
from a chain. Thinning is used, in part, to reduce the possible autocorrelation
that may exist between adjacent draws. The Thinning statement sets the
interval for the MH sampler - meaning if Thinning = 10;, every 10th draw
by the sampler will be retained.
The Burnin statement controls the number of draws that are discarded
from the start of the MH sampler. These draws are discarded to avoid using
values that were sampled before the chain had fully converged to the target
149
R
to discard the first 10 values
distribution. Burnin = 10; tells flexMIRT
obtained by the MH sampler.
Stage1 determines the number of Stage I (constant gain) cycles. The
Stage I iterations are used to improve default or user-supplied starting values
for the estimation that occurs in Stage II. Stage2 specifies the number of
Stage II (Stochastic EM, constant gain) cycles. The Stage II iterations are
used to further improve starting values for the MH-RM estimation that occurs
in Stage III. Stage3 sets the maximum number of allowed MH-RM cycles to
be performed.
Stage4 determines the method by which SEs are found. If Stage4 = 0;
then SEs will be approximated recursively (see Cai, 2010b). If a non-zero
value is given then the Louis formula (Louis, 1982) is used directly. If the
Louis formula is to be used, the supplied value determines the number of
iterations of the SE estimation routine; this number will typically need to be
large (e.g., 1000 or more).
InitGain is the gain constant for Stage I and Stage II. If the algorithm
is initially taking steps that are too large this value may be reduced from the
default to force smaller steps.
Alpha and Epsilon are both Stage III decay speed tuning constants. Alpha
is the first tuning constant and is analogous to the InitGain used in Stages
I and II. Epsilon controls the rate at which the sequence of gain constants
converge to zero. Specified Epsilon values must be in the range (0.5, 1], with
a value closer to 1 indicating a faster rate of convergence to zero. WindowSize
allows the user to set the convergence monitor window size. Convergence of
the MH-RM algorithm is monitored by calculating a window of successive differences in parameter estimates, with the iterations terminating only when all
differences in the window are less than 0.0001. The default value of the window size is set at 3 to prevent premature stoppage due to random variation.
ProposalStd and ProposalStd2 control the dispersion of the Metropolis
proposal densities for the first and second level of the specified model, respectively. If a single level model is specified, only the ProposalStd command
will be used. Although default values of 1.0 have been set for both
of these commands, these values need to be adjusted on a case-bycase basis. The values used will depend on the complexity of the model, the
number of items, the type of items (e.g., dichotmous, polytomous), and the
model fit to the items (e.g., graded, 3PLM, Nominal). The user should choose
ProposalStd and ProposalStd2 values so that the long-term average of the
acceptance rates (which is printed for each iteration in the progress window)
150
for level 1 and level 2 (if applicable) are around 0.5 for lower dimensional
models (<= 3 factors) and in the 0.2 - 0.3 range for higher dimensional/more
complex models. Generally speaking, increasing the ProposalStd value will
result in lowered acceptance rates while decreasing the value will result in
higher acceptance rates. Users are directed to Roberts and Rosenthal (2001)
for optimal scaling, choice of dispersion constants, and long-term acceptance
rates of Metropolis samplers.
MCSize is the Monte Carlo size for final log-likelihood, AIC, and BIC approximations.
7.3. The Groups Section
The <Groups> section includes commands for specifying group names, number
of items, models to calibrate, as well as more advanced features, such as commands for setting up the structure of hierarchical or multidimensional models,
as well as empirical histogram priors.
Example 7.7: <Groups> - General Data/Model Descriptors
<Groups >
%GroupName%
File = "*.dat"’;
N = ?;
Varnames = vars;
Select = vars;
Missing = -9;
Ncats(vars) = ?;
Code(vars) = (?), (?);
Key(vars) = (?);
Model(vars) = ThreePL/Graded(?)/Nominal(?)/GPC(?);
Prior to any other group commands, a label must be assigned to a group,
and it will become a token that all subsequent command statements may refer
to. Even if only one group is present it must be given a label. The name
provided to the group is arbitrary and at the discretion of the user; however,
spaces should not be used when naming groups and the chosen name must be
enclosed by percent signs.
The File statement is for specifying the path and file name in which item
response data is held (or to be saved for Simulation runs). The number of
151
cases in the data file is set using the N statement. The File and N commands
are listed as required here, but there is one exception: When the engine is
placed under Scoring mode, and scoring method SSC is selected, the input
file is optional when only the score conversion table is made without scoring
individuals.
Varnames assigns labels to variables by which they will be referred to later
in commands and in the output. Variable names are separated by commas
following the equals sign of the Varnames statement. When variable names
end with consecutive integers (e.g., v1, v2, ..., v100) a short-cut is available
that can be both time and space saving. Rather than listing all the variables,
a dash between the name of the first variable and the last variable (i.e., v1R
v100) will indicate to flexMIRT
that all variables between v1 and 100 are
desired. The optional Select statement allows for a subset of the variables
initially read from the data file to be submitted to IRT analysis. The default
for Select is to use all variables defined in Varnames.
Missing values may be represented by a specific number in data files. The
default missing value is -9, but that may be modified with the Missing command. However, missing values may only be numeric values between -127 and
127. Common missing value symbols, such as a dot or a blank space will not
be interpreted correctly by the program.
The number of valid response categories observed in the data file is specified
with the NCats statements. Within the parentheses of the NCats statement,
the names of the variables affected by the command are listed. For example,
if Item 24, 28, and 32 - 40 all had four response options, this could be entered
R
as NCats(v24,v28,v32-v40) = 4; in flexMIRT
syntax.
R
The internal representation of item response data in flexMIRT
is zerobased (i.e., response options must start at zero and go up). That being said, the
program is capable of recoding the data, when old and new values are specified
with the Code command. If the first 10 items are read from a data file that
had four response options (1, 2, 3, 4), it would be necessary to recode these
to zero-based, prior to fitting the IRT model. The recode statement could be
Code(v1-v10) = (1,2,3,4),(0,1,2,3);. The original values are listed first
and the values to which they should be recoded appear within the second set
of parentheses. If it is necessary to also collapse the responses, so that the
original codes 1 and 2 become a single response option 0 and the original 3
and 4 become a second option 1, that Code command could be Code(v1-v10)
= (1,2,3,4),(0,0,1,1);. Note that recoding does not change the values
of the number of response options in the data file, as defined by the NCats
152
statements.
The Key statement also provides a specific form of recoding that may be
useful when the observed responses from multiple choice items have been read
in. Rather than writing numerous Code statements to convert the observed
responses into correct/incorrect 0/1 data, the Key statement allows the user to
R
to score the items
provide a scoring key for the items and instruct flexMIRT
internally. For instance, if the correct responses to Items 1 through 4 were
2, 0, 1, and 3, respectively, the Key statement to score these items would be
Key(v1-v4) = (2,0,1,3);. Multiple Code and Key statements are permitted, to easily accommodate situations in which items have variable numbers
of response options or, when scoring a large number of items, multiple scoring
keys would result in greater user-readability of the command file by limiting
command line length. It should be noted that both the Code and the Key comR
mands are completely internal to flexMIRT
and the original data file will not
R
be modified. Although flexMIRT can do some basic data management, it
is not built to be a data management tool. It is generally advisable to perform all necessary data cleaning and data management in a general statistical
software package such as SAS or SPSS before exporting a space-, comma-, or
R
tab-delimited file for flexMIRT
analysis.
The Model command follows the same general format as Ncats, but instead
of placing a number after the equals sign, one of the four keywords for the IRT
model must be chosen. For example, to fit the graded model with four categories to the hypothetical items previously mentioned, the Model statement is
Model(v24,v28,v32-v40) = Graded(4);. The question mark in parentheses
following the Graded, Nominal, and GPC keywords in Example Box 7.7 are
place-holders for the number of categories the model should accommodate.
When items are of mixed formats (multiple choice, free response, Likert-type
scale), multiple NCats and Model statements may be needed to correctly describe the data and the chosen IRT models.
We have covered the required commands within the <Groups> section as
well as several optional commands that are useful for selecting or modifying
data. In addition to the syntax statements previously covered, there are other
additional, more specialized commands which are also available.
153
Example 7.8: <Groups> - Specialized Data/Model Descriptors
<Groups >
...
CaseWeight = var;
Dimensions = 1;
Primary = 0;
Between = 0;
Nlevel2 = ?;
Cluster = var;
CaseID = var;
EmpHist = Yes/No;
ItemWeight(vars) = (?);
BetaPriors(vars) = ?;
Attributes = ?;
InteractionEffects = (?,?,...);
Generate = (?,?), (?,?),...;
DM = group;
Ta(vars) = Trend/Identity/(...);
Tc(vars) = Trend/Identity/(...);
DIFitems = vars;
The CaseWeight command is needed when response pattern by frequency
(aka grouped) data is entered. It provides the program with the variable name
for the column that contains the number of observations (or weights) for each
response pattern.
The Dimensions statement is used to specify the total number of dimensions in the model. The default is a single-level unidimensional model (i.e.,
Dimensions = 1;). Changing the number of dimensions is the first step in
setting up multidimensional and multilevel models. The Primary statement is
used to set the number of primary, or general, dimensions/factors that will be
used in a two-tier multidimensional model. In the bifactor or standard testlet
model, Primary=1;.
Several of the syntax statements in the currently discussed box are closely
associated with multilevel models. To initiate the basic structure for a 2-level
model, the Between command is used, which specifies that number of latent
variables at level-2 (aka Between). When multilevel data are simulated, it is
R
necessary to tell flexMIRT
how many level-2 units should be generated. This
is specified by the Nlevel2 command. This command is only necessary when
154
simulating multilevel data, as the program is able to determine the number of
unique level-2 units from the data file, provided that the Cluster command
is employed
Cluster names the variable that contains unique identifying values for the
level-2 units (e.g., schools in which students are nested). CaseID is used to
specify the variable that provides identifiers for the level-1 units.
The EmpHist statement is used to enable the empirical histogram (EH)
characterization of the latent variable distribution, simultaneously estimated
with the item parameters. The program will calibrate item parameters against
the estimated EH prior, which is based on the normalized accumulated posterior densities for all the response patterns at each of the quadrature nodes, and
R
flexMIRT
will save the EH prior in the -prm file. Additionally, for a concurrent scoring run or any future scoring runs that use the saved parameter file
containing the EH values, the program will recognize that the group contains
EH weights and will use them to compute EAPs or SSCs. Empirical histograms may only be used with single-level unidimensional or bifactor/testlet
models. When used with a bifactor-type model, only the general dimension
will be characterized using the EH.
While category weights are typically taken to be the same that the category value (e.g., a response of 3 is assigned a weight of 3), this does not
have to be the case. If alternate category weights for items are desired, the
ItemWeights statement may be used to assign such alternative weights to
items. For dichotomously scored items, a statement of ItemWeights(v1-v3)
= (0.0,0.5); could be used to assign, for variables v1 to v3, a weight of 0 to
all 0 responses and a weight of 0.5 to all 1 responses. For polytomous items,
the same format is holds. Suppose variables v4 to v8 had 5 possible responses
categories, weights alternate to the default values of 0,1,2,3,4 could be assigned
with the command, ItemWeights(v1-v3) = (0.0,0.1,2.0,3.0,3.1);. Due
to this statement, scores of 0 are assigned a category weight of 0, 1s a weight of
0.1, 2s a weight of 2.0, all 3s given a weight of 3.0 and responses of 4 a weight
of 3.1. Multiple ItemWeights statements are permitted, to handle situations
when the number of responses options per item is variable.
155
BetaPriors is used to assign item-specific beta distribution priors on the
uniquenesses; the rational and use of applying priors to the item-uniquenesses
was discussed in Bock et al. (1988) on pp. 269-270. The Beta distribution
R
applied in flexMIRT
has an α parameter set to the user-specified value and
the β parameter fixed at 1. For example, the statement BetaPriors(v1-v5) =
1.6; would apply to the uniquenesses of Items 1 through 5 a prior distribution
of the form Beta(1.6, 1). Multiple BetaPriors statements are permitted.
The next four commands are used exclusively for the fitting of cognitive
diagnostic models. Attributes is used to set the number of main effects
present in the skill space that is being assessed. This is the keyword that
R
triggers flexMIRT
to model discrete latent variables, rather than the default
continuous latent variables.
R
InteractionEffects instructs flexMIRT
to automatically generate all
possible interaction effects. For instance, with 4 attributes one may specify
InteractionEffects = (2,3,4); and, in addition to the 4 main effects, the
program with also generate the 2nd, 3rd, and 4th order interaction terms of
the attributes as dimensions of the model. If only the second order interaction
effects are desired, those would be created by specifying InteractionEffects
= (2);.
Similar to the InteractionEffects keyword, Generate sets up higher order interactions effects. However, rather than creating all possible interaction
effects Generate creates only those effects specified in the statement. For example, Generate = (3,6,7),(4,7); is used to generate the interaction effect
of attributes 3, 6, and 7 and, separately, the interaction effect of attributes 4
and 7. There is no limit to the number of interaction effects that may be
specified in a Generate statement and the order in which interaction effects
are specified does not matter.
The DM command stands for “diagnostic model” and is used to set the
group containing observed data - this is the group to which the higher-order
DM will be applied. The DM command is optional; one could, in principle, fit
a model without the DM (higher-order) latent variables.
The commands Ta and Tc are used in conjunction with the nominal categories model, supplying contrasts for the scoring function and intercepts,
respectively. The default for both Ta and Tc is a trend contrast matrix (keyword is Trend), but an identity matrix (keyword is Identity) which allows for
equality constraints, or a user-supplied matrix (denoted by (...) in the syntax
box) may also be employed. Multiple Ta and Tc statments may be supplied,
allowing for a large degree of flexibility in the parameterization of the nominal
156
model.
The final command DIFitems in the <Groups> section supplies candidate
DIF items. This command is required when DIF= TestCandidates; has been
invoked in the <Options> section. The names of candidate items are listed
across, separated by commas, (e.g., DIFitems = v1-v4,v8;)
7.4. The Constraints Section
Several of the examples have shown that there is no required statement in
the <Constraints> section, and that it may be left blank after the (required)
section header. However as models become more complex, especially if they
include multiple groups, constraints will more than likely become necessary.
Listed in the left-most column of Table 7.1 are the types of constraints
available. Value indicates the starting or fixed value of a parameter. Free
indicates a parameter which is freely estimated, Equal indicates parameters
which will be constrained equal, Fix fixes a parameter at a specific value, Coeff
applies a provided coefficient to a given parameter - enabling proportionality
restrictions, AddConst applies a additive constant to specified parameters, and
Prior provides a prior distribution for that parameter.
Table 7.1: Constraint Commands and Keywords
Constraint Type
Free
Equal
Fix
Value
Coeff
AddConst
Prior
Parameter
Intercept(?)
Slope(?)
Guessing
ScoringFn(?)
Mean(?)
Cov(?,?)
MainEffect(?)
Interaction(?,?,...)
Prior Distribution
Normal(µ, σ)
logNormal(µ, σ)
Beta(α − 1, β − 1)
Presented in the second column of Table 7.1 are the keywords for parameters that may have constraints applied to them. Intercept refers to the location parameter. Slope refers to the slope/discrimination parameter. Guessing
is the lower asymptote parameter in the 3PL model. ScoringFn is the scoring
function contrast parameters of the nominal model. Mean and Cov refer to
the mean and covariance matrix elements of the latent variable distribution.
157
MainEffect and Interaction are available so users may employ syntax specifications that are in keeping with the conceptualization/terminology used in
cognitive diagnostic models.
The question marks in the parentheses following the parameter keywords
permits the referencing of specific parameters. For example, when there are 4
latent variables in the model, Slope(1) refers to the slope on the first factor,
Mean(2) refers to the mean of the second factor, and Cov(3,2) refers to the
covariance of factors 3 and 2. When an item has 5 categories, Intercept(2)
in a graded model or ScoringFn(3) in a nominal model each refers to specific
item parameters.
For the Free, Equal, and Fix constraints, the general format for applying
restrictions to the model is the same. First, the type of constraint to be applied
and the group to which it is being applied are listed, followed by a comma. The
items which are affected by the constraint are then specified in parentheses,
followed by another comma and then the keyword for the parameter that is
affected (e.g., Slope, Guessing, etc). For a unidimensional model in Group1,
the slope parameters for v1 through v5 can be set equal using the following
syntax:
Equal Group1, (v1-v5), Slope;
If there is only one group, the group name can be omitted.
Equal (v1-v5), Slope;
The following syntax fixes the second slope of (v1,v2) to 0 (the default value
unless modified by Value statement).
Fix (v1,v2), Slope(2);
The following syntax fixes the second factor mean of Group2 to 0 (the default
value unless modified by Value statement).
Fix Group2, Mean(2);
158
The following syntax fixes the second factor variance of Group2 to 1 (the
default value unless modified by Value statement).
Fix Group2, Cov(2,2);
The following syntax fixes the covariance between factors 2 and 3 in Group2
to 0 (the default value unless modified by Value statement).
Fix Group2, Cov(3,2);
The following syntax fixes the first scoring function contrast for (v1) to 1 (the
default value unless modified by Value statement).
Fix (v1),ScoringFn(1);
The following syntax fixes the second scoring function contrast for (v1) to 0
(the default value unless modified by Value statement).
Fix (v1),ScoringFn(2);
The following syntax frees the third slope for items v1-v5 in Group2.
Free Group2,(v1-v5),Slope(3);
Cross-group constraints are also allowed. The basic constraint format is still
used, with the addition of the constraint in the second group following a colon.
In one of our multiple group examples, it was desired that the item parameters
for 12 items be the same across the 2 groups. To accomplish this, the Equal
constraints in the excerpted code below were employed.
Example 7.9: Constraints Across Groups, Specifying a Prior Distribution,
and Freely Estimating Theta
...
<Constraints >
Free Grade3, Mean(1);
Free Grade3, Cov(1,1);
Equal Grade3, (v1-v12),
Equal Grade3, (v1-v12),
Equal Grade3, (v1-v12),
Prior Grade3, (v1-v12),
Guessing : Grade4, (v1-v12), Guessing;
Intercept: Grade4, (v1-v12), Intercept;
Slope : Grade4, (v1-v12), Slope;
Guessing : Beta(1.0,4.0);
159
As can be seen, for the three constraints that set the item parameters
across groups, the general constraint format of type group, (items), parameter
is maintained for both groups, with a colon between the listing for the Grade
3 and Grade 4 groups. If more than two groups are to be constrained equal,
additional colons, followed by the additional group information may be added.
Also of note in this syntax excerpt, by default, the latent variable mean vector
and covariance matrix are initially constrained to zero and identity, respectively. Using constraints, the first two lines in the example above specify that
the mean and variance for the latent trait are to be freely estimated in the
Grade 3 group, relative to the fixed values of the Grade 4 group which has a
mean of 0 and variance of 1.
The format of the Value, Coeff, and AddConst constraints is similar to
the previous types of constraints but with an additional piece of information.
After specifying the constraint type, group, items, and parameter, a numeric
value is specified as well which is separated from the parameter keyword by a
comma. For example, a coefficient command could be something like, Coeff
G, (v1-v0), Slope, 1.702; where we are applying a coefficient, in group
G, for variables v1-v10, of 1.702 to the slopes, effectively incorporating the
scaling constant D into the estimation.
The Value statement is a more general purpose constraint command. In
calibration mode, it can be used to specify the values of fixed parameters or
it can be used to supply starting values. In scoring mode, instead of using a
parameter file (which is more convenient), one can in principle use the Value
statement to fix item and group parameters to the calibrated estimates. In
simulation mode, the Value statement can be used to specify generating parameters. For a Value constraint on one or more parameters, if the parameters
are free, the numeric values supplied become starting values. If the parameters
are fixed, Value can be used to change the default fixing values. For example,
for a unidimensional model in a single group, the following syntax results in a
set of constraints that fixes the item slopes to 1.0 and then frees the estimation
of the variance of the latent variable, effectively creating a Rasch-type model.
Fix (v1-v5), Slope;
Value (v1-v5), Slope, 1.0;
Free Cov(1,1);
For applying prior distributions, the format is quite similar to the other
constraints, but with the addition of a distribution type (and its associated
pair of parameter values) after a colon. The syntax still follows the general
160
constraint format of type group, (items), parameter, with the addition of the
requested prior distribution following a colon. In the across groups constraint
example previously used, we specified a beta prior for the guessing parameters.
The prior chosen was a beta distribution with α − 1 = 1.0 and β − 1 = 4.0. As
noted in Table 7.1, the Normal distribution and the logNormal distributions
are also available when assigning priors. For the Normal, the prior mean
and standard deviation are needed. For the logNormal, the prior mean and
standard deviation on the logarithmic scale are needed. For the Normal prior,
if it is used on the guessing parameter, it is in fact imposed as a logit-normal
prior on the logit of the guessing probability.
161
CHAPTER 8
Models
In this chapter, some more technical aspects of the IRT model implemented in
R
flexMIRT
are discussed. We begin with a generalized formulation, focusing
on the linear predictor portion of the model. More specific issues are discussed
next.
8.1. A Generalized Formulation
Consider a linear predictor for a level-1 unit i (e.g., individual) in level-2 unit
j (e.g., school) in a group g (e.g., country):
B
W W
ηijg = AB
g θjg + Ag θijg ,
where ηijg is an ng × 1 vector of linear predictors of the ng items in this group,
AB
g is an ng × p matrix of item slopes on the p level-2 (between) latent dimensions or latent attributes (and interaction terms), and AW
g is an ng × q matrix
of item slopes on the q level-1 (within) latent dimensions or latent attributes
W
(and interaction terms), with θB
jg and θijg being the between and within latent
dimensions or attributes (and attribute interactions). Note that both AB
g and
W
Ag are implicitly assumed to be functions of d unknown structural parameters in ξ, subject to appropriate identification conditions or user specified
constraints.
At the item level, the IRT model is a conditional density fξ (yijkg |ηijkg ) =
W
fξ (yijkg |θB
jg , θijg ), where k = 1, . . . , ng and yijkg is the item response to item
k from level-1 unit i in level-2 unit j in group g. Conditionally on the latent
variable levels the item responses are independent such that
W
fξ (yijg |θB
jg , θijg ) =
ng
Y
W
fξ (yijkg |θB
jg , θijg ).
k =1
W
Let the distribution of θW
ijg be fξ ( θijg ), the level-1 latent variables can be
162
integrated out as
fξ (yijg |θB
jg ) =
Z
W
W
W
fξ (yijg |θB
jg , θijg )fξ ( θijg )dθijg .
Assuming further conditional independence of level-1 units on θB
jg , the marginal
distribution of all item responses in a level-2 unit is
fξ (Yjg ) =
Nj
Z Y
B
B
fξ (yijg |θB
jg )fξ ( θjg )dθjg ,
i=1
where Nj is the number of level-1 units in level-2 unit j, fξ (θB
jg ) is the distribution of the level-2 latent variables, and Yjg = {y1jg , . . . , yNj jg } is the
collection of item response patterns for all ng items from the Nj level-1 units
in level-2 unit j of group g.
Once the item responses are treated as fixed upon observation, the marginal
likelihood of Yjg is defined as
L(ξ|Yjg ) = fξ (Yjg ).
Taking logs and summing over the (assumed) independent level-2 units, as
well as groups, the marginal log-likelihood for the structural parameters is
log L(ξ|Y1 , . . . , Yg ) =
Jg
G X
X
L(ξ|Yjg ),
g =1 j =1
where G is the number of groups, Jg is the number of level-2 units in group g,
and Yg = {Y1g , . . . , YJg g } is the collection of item response patterns for all
level-1 and level-2 units in group g.
8.2. Item Response Models
B θB + aW θW ,
For the kth row in ηijg , the linear predictor is equal to ηijkg = akg
jg
kg ijg
B is the set of p slopes on the between latent variables and aW is the
where akg
kg
set of q slopes on the within latent variables.
8.2.1
Model with Pseudo-Guessing Parameter
This is the multilevel and multidimensional extension of the 3PL model. Let
the correct/endorse (1) response probability be
1 − guesskg
,
Pξ (yijkg = 1|ηijkg ) = guesskg +
1 + exp[−(ckg + ηijkg )]
where guesskg is the item-specific pseudo-guessing probability and ckg is the
item-specific intercept. Consequently, Pξ (yijkg = 0|ηijkg ) = 1.0 − Pξ (yijkg =
1|ηijkg ).
163
8.2.2
Graded Response Model
Suppose item k has K graded categories. Let the cumulative response probabilities be
Pξ (yijkg ≥ 0|ηijkg ) = 1.0,
1
,
1 + exp[−(ckg,1 + ηijkg )]
Pξ (yijkg ≥ 1|ηijkg ) =
..
.
1
Pξ (yijkg ≥ K − 1|ηijkg ) =
Pξ (yijkg
1 + exp[−(ckg,K−1 + ηijkg )]
≥ K|ηijkg ) = 0.0,
,
where the c’s are item intercepts and the boundary cases are defined for consistency. Then the category response probabilities are
Pξ (yijkg = l|ηijkg ) = Pξ (yijkg ≥ l|ηijkg ) − Pξ (yijkg ≥ l + 1|ηijkg ).
8.2.3
Nominal Categories Model
R
The nominal model fit by flexMIRT
is the multilevel extension of the reparameterized nominal model due to Thissen et al. (2010). Suppose item k has
K nominal categories. Category l’s response probability is defined as
exp(skg,l ηijkg + ckg,l )
,
Pξ (yijkg = l|ηijkg ) = PK−1
m=0 exp(skg,m ηijkg + ckg,m )
where skg,l is the scoring function value for category l and ckg,l is the intercept
value for category l. Dropping the kg subscript for a minute, Thissen et al.
(2010) pointed out that for identification, the following restrictions should be
in place: s0 = 0, sK−1 = K − 1, c0 = 0. This can be accomplished by
reparameterizaton. Let


s=

s0
..
.
sK−1





= Ta
1
α
!

, and d = 

c0
..
.
cK−1




= Tc γ.
The vector α is a (K − 2) × 1 vector of scoring function contrasts that defines
the ordering of categories, and γ is a (K − 1) × 1 vector of intercept contrasts.
164
The matrices Ta and Tc are fixed K × (K − 1) matrices of contrast coefficients. By an appropriate choice of T, the identification restrictions will be
automatically satisfied. Thissen et al. (2010) propose the use of the following
linear-Fourier contrast matrix

T=
0
1
2
..
.
t2,2
t3,2
..
.
K −1
0








0
···
0
· · · t2,(K−1)
· · · t3,(K−1)
..
.
···





,



0
where a typical element tl,m for l = 1, 2, . . . , K and m = 1, 2, . . . , K − 1 takes
its value from a Fourier sine-series:
(
tl,m
)
π (l − 1)(m − 1)
= sin
.
K −1
For instance, for K = 4, the contrast matrix is

F=





0
0
0
1 .866
.866
2 .866 −.866
3
0
0



,


and for K = 5, the contrast matrix is

F=








0
0
0
0
1 .707
1 .707 


2
1
0 −1 


3 .707 −1 .707 
4
0
0
0
For the nominal model, one can verify that identification conditions are all
satisfied. For the GPC model, all the α contrasts are set to zero.
165
Appendices
166
APPENDIX A
Quick-Start Installation and User Guide
R
from
This guide is designed to provide an overview for installing flexMIRT
the webpage and initializing the program for use.
A.1. Installing flexMIRT
R
To obtain the flexMIRT
installer, you must first register for an account on
R
the flexMIRT website. The registration page may be found by selecting
R
the My Account link in the flexMIRT
website banner or visited directly at
https://flexmirt.vpgcentral.com/Account/LogOn. From your account page,
R
you will be able change your password, view your current flexMIRT
license
and expiration date information, manage licenses on your various computers
(with a purchased license users are allowed up to 3 separate installs of the
program), as well as make payments for new or renewed licenses.
After selecting Register Here, you will be taken to the registration page
and asked to enter your name, email address, and any affiliation information
you would like to supply.
167
Once you have submitted this information via the Register button, an email
confirming your registration information (including the initial password for
R
your flexMIRT
account) will be sent to the address you provided. The confirmation email will originate from [email protected] - please add this email
address to your safe list and, if you have not received an email after 10 minutes,
check that the email has not been redirected to your spam folder.
With the password obtained from the email, you may now log in to your
R
flexMIRT
account. You may change your password using the Change Password link in the upper right hand corner. Before you are able to download
the installer, you must accept the license agreement, which can be accessed by
clicking the link in the green box.
168
You will be taken to the End User License Agreement. Once you have
reviewed the agreement, there is an “I Agree” check box at the bottom of the
page which you must endorse to continue. After you have accepted the license
R
agreement, you will be able to access the flexMIRT
installer via the newly
available Download Software link.
169
After selecting Download, a pop-up will appear where you will select the
R
R
version of flexMIRT
appropriate for your computer. The flexMIRT
installer
should be saved to a folder where you will be able to locate it later.
Locate the downloaded installer (flexmirtInstaller) and double-click to activate it. NOTE: IT MAY BE NECESSARY FOR SOME USERS, ESPECIALLY THOSE WITHOUT ADMINISTRATIVE PRIVILEGES, TO RIGHTCLICK ON THE INSTALLER AND SELECT “Run as Administrator” FOR
THE PROGRAM TO FUNCTION PROPERLY ONCE IT HAS BEEN INSTALLED.
170
The second step of the installer will ask which folder you wish to install
the program into. If you would like the program installed somewhere other
than the default location, this can be accomplished by selecting the “Change”
button and directing the installer to the desired folder location.
Follow the remaining steps, as directed by the installer, until the program
R
indicates that it has finished installing. Once flexMIRT
have been installed,
a shortcut will appear on the desktop. Double-click this to initiate the proR
gram. On the first start-up flexMIRT
will open a pane where you must
register the software; if you opt to cancel at this point you will receive a
R
warning message and flexMIRT
will completely shut down. You may either
R
supply the flexMIRT account username and password you set up earlier or
may copy the license code from your My Account webpage into the appropriate
box.
171
A.2. Using flexMIRT
With the software registered, you are now able to conduct analyses. The
R
flexMIRT
Support page (https://flexmirt.vpgcentral.com/Support) has the
User’s Manual and numerous example syntax command files with accompanying datasets, to help you become acquainted with the types of analyses that
may be conducted and how command files should be structured. By selecting “New” under “File”’ or using the “New” icon, a command file containing
necessary statements for a basic analysis opens. Once you have modified this
command file, it must be saved prior to being able to submit the analysis
R
to flexMIRT
; the ”Run” button is disabled until the file has been saved to
prevent the example code from being over-written.
To provide a brief tutorial, we will focus on the first example, found in
the folder Example 2-1. Existing files are opened using the Open command
under the File option of the engine viewing pane, which is the pane opened
R
when the program is first initiated. Once the desired flexMIRT
command
file is located and selected, it will be opened in a separate pane, labeled Syntax
Editor: with the file name listed after.
172
The command file is submitted for analyses by clicking the “Run” button
R
on the far right of the icon toolbar. flexMIRT
will print its progress (e.g., EM
iteration number, etc.) in the engine pane and, when complete, will open the
Output Viewer pane, which will contain all results. Additionally, the results
are written to an external text file, which may be opened at a later time with
R
or a text editor such as Notepad.
flexMIRT
173
APPENDIX B
flexMIRT via the Command Line Interface
R
via the Command
This appendix provides a brief overview of using flexMIRT
Prompt, which may be useful to those wishing to batch submit a large number
R
of files or to control flexMIRT
via another program (e.g., R).
B.1. flexMIRT via the CLI - single file run
Begin by accessing the Command Prompt via the Windows Start button→
All Programs→ Accessories→ Command Prompt
174
You will now need to direct the CLI to the folder where the WinFlexMIRT
executable is found. This is done using standard “change directory” comR
mands. For example, if flexMIRT
was located in the folder “C:\flexmirt”,
you would type “cd C:\flexmirt” where “cd” stands for “change directory” and
the remaining text is the path to the folder.
If a correct location has been specified, the working directory will be
changed, which can be seen in the above image; otherwise, an error message
of “The system cannot find the path specified” will be reported.
Once the CLI has been pointed to the folder that contains the WinFlexMIRT executable, all that remains is to supply the command file name
with the appropriate WinFlexMIRT call.
If the command file is found in the same folder as the executable, the call
is of the form:
WinflexMIRT.exe -r mysyntax.flexmirt
where the “-r” tells WinFlexMIRT to run in the CLI and mysyntax.flexmirt
is the command file you wish to run.
175
If the command file is found in a different folder, or has spaces in the file
name, then the full path to the command file must be given in quotation marks.
For example, if we wished to submit the command file “FL_20item.flexmirt”
for analysis and it is located in the folder
“C:\Users\VPG\Documents\FL analyses\Final scale\”, the WinFlexMIRT
call would be:
WinflexMIRT.exe -r “C:\Users\VPG\Documents\FL analyses\Final
scale\FL_20item.flexmirt”
Once a command file has been successfully submitted, the same progress
R
information that is shown in the engine pane of the flexMIRT
graphical user
interface will be printed in the CLI.
When the run has completed successfully, the output files may be found in
the folder in which the command file is located.
176
B.2. flexMIRT via the CLI - batch mode
R
A newer feature in flexMIRT
is the ability to submit multiple syntax files
via a single batch file. The feature is only available in software releases after
R
December 1, 2012. If you are using a version of flexMIRT
1.0 released prior
R
to that date, you will need to download the latest version of flexMIRT
from
the website and reauthenticate your license to have access to batch mode
processing.
The batch file consists of a single column of comma separate values in
which the user provides the locations and file names of the command files
R
to be run. The flexMIRT
syntax files in the batch file must ALWAYS be
referenced with the full path. Below is an example batch file, saved as a CSV
in Excel.
177
After directing the CLI to the folder that contains the WinFlexMIRT executable (see above), the batch mode is called using:
WinflexMIRT.exe -b batch.csv
where the “-b” tells WinFlexMIRT to run in the CLI batch mode and batch.csv
is the name of the batch file. If the batch file is in a different directory than
R
the flexMIRT
executable or contains spaces in the file name, the full path
and file name should be provided and enclosed in quotes, as was shown with
the single-run command line run.
Once the batch file has been submitted, progress will be printed. The
number of files found in the batch file is first reported and then the progress for
R
individual runs starting and finishing will be given. If flexMIRT
is unable to
complete a run successfully, the fact that an error was detected in the run will
also be reported. As can be seen, the file “broken.flexmirt” did not complete
and that fact is reported here quite plainly.
R
Once flexMIRT
has processed all of the syntax files named in the batch
file, it will report that it has finished the batch and print the total and average
processing time per file. It will also give the name and location for a summary
log of the batch run - the log file will be located in the same directory as the
batch file and will be named with the batch file name and a “_log.txt” suffix.
178
The batch log from the example batch is presented here and, as in the
command prompt screen, shows that broken.flexmirt was not completed due
to errors and the run time for each successfully completed syntax file and the
batch as a whole.
The typical output files for the individual runs submitted in batch mode
will be in the same folder as the corresponding *.flexmirt syntax command
file, not with the batch file.
179
APPENDIX C
Plotting
In unidimensional IRT, graphs (e.g., tracelines and information curves) form
R
an integral component of item analysis in practice. flexMIRT
is built to
handle multilevel and multidimensional IRT models. In this context, there
currently exists no universally accepted consensus with regards to graphical
procedures. We do realize, however, that there is a need for publication quality
graphics, and we chose to provide a set of R programs. We implemented the
provided graphing functions in R because it is incredibly powerful, allows
maximal customizability, and is a free program.
R
While flexMIRT
does not have an internal plotting function, supplemenR
tal R code to read-in flexMIRT
parameter files and plot item characteristic curves (ICCs), test characteristic curves (TCCs), item information functions (IIFs), and test information functions (TIFs) is available on the Support
webpage (http://www.flexMIRT.com/Support) in a folder labeled “flexMIRT
plotting examples”. Please note that the provided code is usable with single
factor models only.
C.1. Plotting ICCs, TCCs, and Information Functions
R
The provided R code reads item parameters from a flexMIRT
parameter
output file (with the extension*-prm.txt). For information functions, the code
R
reads the needed values from a flexMIRT
information output file (which has
R
does not generate these files by default, so
the extension *-inf.txt). flexMIRT
it is necessary to explicitly request that the parameters and information values
be saved from the calibration run using the SavePRM = Yes; and the SaveInf
= Yes; commands in the <Options> section, respectively. Additionally, by
default the Information Function output file will only contain the Information
value for a single point at the theta value of 0. To change the number of points
and the range over which they are spaced, one uses the FisherInf = 21,
3.0; command in the <Options> section. This specific command requests
180
that information values at 21 points, spaced evenly between -3.0 and 3.0, be
written to the generated *-inf.txt file; the choice of 21 and 3.0 are used as
placeholders and both values can be modified by the user.
The entirety of the example plotting code (found in the file “DrawICCs_Infs.R”
in the available “flexMIRT plotting examples” folder) is presented below. Because we don’t wish to assume that users are familiar with R, we will briefly
discuss the main points of the R syntax.
Example C.1: R plotting code
##calls the pre-defined plotting functions - REQUIRED AS FIRST STEP
#ALTER THE PATH SO IT DIRECTS TO WHERE flexmirt.R IS SAVED ON YOUR COMPUTER
source("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/flexmirt.R")
##For plotting ICCs and TCCs, use the function call flexmirt.icc("parameter file name")
#PROVIDE THE FULL PATH TO THE -prm FILE TO BE PLOTTED
##1PL example
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/SM3-prm.txt")
##2PL example
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/Preschool.2PL-prm.txt")
##3PL example
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/Preschool.3PL-prm.txt")
##Nominal Model Example
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/PreschoolNum_Nom-prm.txt")
##Graded Example
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/LiberalConservative-prm.txt")
##GPC Example
flexmirt.icc(""C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/LiberalConservative-GPC-prm.txt")
##For plotting IIFs AND TIFs, use the function call flexmirt.inf("inf file name")
#NOTE: THE INF PLOTTING FUNCTION PULLS INFORMATION FROM THE *.flexmirt CALIBRATION FILE
#THAT CREATED THE *.inf FILE.
##IT IS ASSUMED THAT THE *-inf FILE and the *.flexmirt FILE ARE IN THE SAME LOCATION
##information function example
flexmirt.inf("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/LiberalConservative-inf.txt")
##information function example - multiple grp
flexmirt.inf("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/fit2-inf.txt")
As may be seen, there are brief explanations provided in the R code for
those less familiar with the program and its language. Although we have
provided numerous examples, these are primarily to demonstrate that the
provided plotting functions work with a variety of models. There are actually
only two commands that are needed to plot any given set of parameters. First,
R needs to load the pre-defined plotting functions we have provided - these are
found in the file “flexmirt.R” and the file is loaded into R using the “source”
call seen on the first non-comment line (the first line without a # sign at the
beginning).
181
Once the functions have been loaded into R, the next step is to provide
the functions with the location and file name of the values one wishes to plot.
For obtaining ICCs and TCCs, the function call used is “flexmirt.icc()” and
within the parentheses, the user supplies the *-prm file name, with location
path, that contains the parameters that will define the ICCs and TCCs. For
instance, the line
flexmirt.icc("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/Preschool.2PL-prm.txt")
calls the trace line plotting function and tells the function to create plots based
on the parameter values found in the file “Preschool.2PL-prm.txt”.
The graph presented here is representative of the plot output that will be
obtained from the ICC/TCC plotting function.
Figure C.1: GRM ICCproduced by provided R code
Similarly, for obtaining information function plots, the function call of
“flexmirt.inf()" is used and the name and location of an *-inf file is provided
in the parentheses. So, the example syntax line
flexmirt.inf("C:/Users/laptop/Documents/flexMIRT/flexMIRT plotting/fit2-inf.txt")
182
is telling R to use the information plotting function (flexmirt.inf) to plot the
information values found in the file “fit2-inf.txt.” As noted in the comments in
the R code file, the information plotting function pulls information from the
calibration file that created the *-inf.txt file. The program assumes that the
*-inf file and the original calibration syntax file (*.flexmirt) will be in same
folder. If, for some reason, the *.flexmirt file is not in the same folder as the
information file, the plotting function will return an error and is unable to run.
The provided plot gives a representative of the graphics produced by the
IIF/TIF plotting function.
Figure C.2: IIF produced by provided R code
The R programs provided are free software; you can redistribute and/or modify them under the terms of the
GNU General Public License (http://www.gnu.org/copyleft/gpl.html) as published by the Free Software
Foundation; either version 2 of the License, or (at your option) any later version. The programs are
distributed in the hope that they will be useful, but WITHOUT ANY WARRANTY; without even the
implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
General Public License for more details.
183
APPENDIX D
Labeled Output Files
R
creates,
Due to the lack of labels in some of the output files that flexMIRT
we provide this appendix to give labeled examples of several file examples,
specifically -sco and -prm files. For each example, we will provide labels at the
top of each column, which would not normally appear, and a small selection
of lines from the output file.
Table D.1: Labeled -sco file - EAP scores, 1 Factor
Grp#
1
1
1
1
1
flexMIRT Obs #
1
2
3
4
5
θˆ
-0.179
0.957
0.033
0.957
-0.519
ˆ
SE(θ)
0.501
0.609
0.576
0.609
0.515
Table D.2: Labeled -sco file - MAP scores, 1 Factor
Grp#
1
1
1
1
1
flexMIRT
Obs #
1
2
3
4
5
# iterations
to MAP
2
3
2
3
2
θˆ
ˆ
SE(θ)
-0.150
0.939
0.077
0.939
-0.489
0.494
0.564
0.567
0.564
0.518
Note: Scores from maximum likelihood (Score = ML;) will have a similar "# of
iterations to score" column.
184
Table D.3: Labeled -sco file - EAP scores, 2 Factors
Grp#
1
1
1
1
1
flexMIRT
Obs #
1
2
3
4
5
θˆ1
θˆ2
SE(θˆ1 )
SE(θˆ2 )
σˆ 11
σˆ 21
σˆ 22
-0.141
0.192
2.023
1.311
0.667
-0.466
1.407
1.684
0.840
1.182
0.258
0.281
0.527
0.334
0.262
0.355
0.295
0.392
0.290
0.284
0.066705
0.079119
0.277999
0.111558
0.068758
0.001763
0.001470
0.015970
0.003131
0.001159
0.126136
0.087174
0.153786
0.083973
0.080632
Table D.4: Labeled -sco file - EAP scores, 2 Factors with User-supplied ID
Variable
Grp#
1
1
1
1
1
flexMIRT
Obs #
1
2
3
4
5
User ID
θˆ1
θˆ2
SE(θˆ1 )
SE(θˆ2 )
σˆ 11
σˆ 21
σˆ 22
sub1
sub2
sub3
sub4
sub5
-0.141
0.192
2.023
1.311
0.667
-0.466
1.407
1.684
0.840
1.182
0.258
0.281
0.527
0.334
0.262
0.355
0.295
0.392
0.290
0.284
0.066705
0.079119
0.277999
0.111558
0.068758
0.001763
0.001470
0.015970
0.003131
0.001159
0.126136
0.087174
0.153786
0.083973
0.080632
Table D.5: Labeled -sco file - MAP scores, 2 Factors
Grp#
1
1
1
1
1
flexMIRT
Obs #
1
2
3
4
5
# of Iterations
to MAPs
3
4
5
4
4
θˆ1
θˆ2
SE(θˆ1 )
SE(θˆ2 )
σˆ 11
σˆ 21
σˆ 22
-0.146
0.200
1.838
1.261
0.677
-0.405
1.409
1.643
0.831
1.190
0.245
0.263
0.494
0.296
0.247
0.338
0.287
0.368
0.275
0.270
0.059838
0.068942
0.244128
0.087689
0.061086
0.001456
0.001183
0.011428
0.002205
0.000914
0.114251
0.082246
0.135356
0.075827
0.072911
185
186
Label
v1
v2
v3
v4
v5
Label
Group1
Grp #
1
1
1
1
1
Grp#
1
# of Factors
1
1
1
1
1
# of Factors
1
Model
2
2
2
2
2
Prior
0
# of Cats
5
5
5
5
5
µ1
0.00
c1
3.24905e+000
4.35083e+000
4.04014e+000
4.11180e+000
4.34774e+000
σ11
1.00
Type
1
1
1
1
1
Type
0
Label
v1
v2
v3
v4
v5
Label
Group1
Grp #
1
1
1
1
1
Grp#
1
# of Factors
6
6
6
6
6
# of Factors
6
Model
2
2
2
2
2
Prior
0
# of Cats
2
2
2
2
2
µ1
0.00
c
2.12982e+000
3.77704e+000
1.69155e+000
3.76464e+000
2.86759e+000
···
···
Table D.7: Labeled -prm file - 2PL with 6 Factors, 1 Group
Type
1
1
1
1
1
Type
0
Table D.6: Labeled -prm file - GRM with 1 Factor, 1 Group
a1
2.26116e+000
2.30913e+000
1.61782e+000
3.86443e+000
2.76758e+000
µ5
0.00
c2
9.45119e-001
1.79341e+000
1.60919e+000
1.64584e+000
1.71393e+000
a2
0.00000e+000
2.51024e+000
1.41918e+000
4.89228e+000
3.08101e+000
σ11
1.00
c3
-4.10822e-001
4.76297e-001
2.87393e-001
3.54674e-001
4.73887e-001
a3
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
σ21
0.00
c4
-3.33909e+000
-2.91482e+000
-2.39810e+000
-2.61476e+000
-2.35381e+000
a4
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
σ22
1.00
a
1.58287e+000
1.85045e+000
1.78740e+000
1.72958e+000
1.44323e+000
a5
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
σ31
0.00
a6
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
0.00000e+000
···
···
Table D.8: Labeled -prm File - 3PL with 1 Factor, 3 groups
Type
1
1
..
.
Label
v1
v2
..
.
Grp #
1
1
..
.
# of Factors
1
1
..
.
Model
1
1
..
.
# of Cats
2
2
..
.
logit g or c
-1.33972
-0.96814
..
.
c or a
-0.68782
-1.29641
..
.
a
1.301362
1.409536
..
.
1
1
1
1
..
.
v9
v10
v1
v2
..
.
1
1
2
2
..
.
1
1
1
1
..
.
1
1
1
1
..
.
2
2
2
2
..
.
-2.79306
-1.65629
-1.33972
-0.96814
..
.
0.730972
-0.14263
-0.68782
-1.29641
..
.
1.172468
1.102673
1.301362
1.409536
..
.
1
1
1
1
..
.
v9
v10
v1
v2
..
.
2
2
3
3
..
.
1
1
1
1
..
.
1
1
2
1
..
.
2
2
2
2
..
.
-2.79306
-1.65629
-0.68782
-0.96814
..
.
0.144738
-0.14263
1.301362
-1.29641
..
.
1.172468
1.102673
1
1
Type
0
0
0
v9
v10
Label
Group1
Group2
Group3
3
3
Grp#
1
2
3
1
1
# of Factors
1
1
1
1
1
Prior
0
0
0
2
2
µ1
0
0.2
-0.2
-2.79306
-1.65629
σ11
1
1
1.5
0.730972
-0.14263
1.172468
1.102673
Note that for item v1 in group 3, a different model (the 2PL, rather than the
3PL) is used. This is why the item parameter columns “logit g or c” and “c
or a” are labeled as they are; the columns have different meanings, depending
on the model fit to an item.
187
1.409536
..
.
APPENDIX E
Licensing Information
R
R
Purchasing a license for flexMIRT
is easy. If you intend to use flexMIRT
for
academic purposes, simply register at flexmirt.vpgcentral.com/Account/LogOn and,
once logged in, follow the instructions on-screen to download a trial version. At any
R
point in your trial, you can log in to your account and purchase a flexMIRT
license.
Your yearly subscription will begin the moment you purchase the program. Newer
R
versions of flexMIRT
will be updated automatically on any computer with an
internet connection and on which a valid copy of the program has been installed
R
Each single-user license is good for three installs of flexMIRT
so you can easily
R
work with flexMIRT on your academic, home, and laptop computers at no additional cost. If additional installations are required, pricing information may be
obtained by contacting [email protected].
R
Your flexMIRT
license does not automatically renew. You will be asked after
R
your year with flexMIRT
if you would like to renew your subscription. If you desire
a different subscription length, please contact [email protected] for pricing
information.
E.1. Commercial use of flexMIRT
R
This version of flexMIRT
is intended for non-commercial use only (as defined
R
in the flexMIRT End User’s Licensing Agreement). If you are interested in purR
chasing flexMIRT
for commercial use, please contact [email protected].
E.2. Classroom Discounts for flexMIRT
R
For educators who are interested in using flexMIRT
in their classroom, bulk educational discounts are available. Please contact [email protected] for more
information.
188
References
Adams, R., & Wu, M. (2002). PISA 2000 technical report. Paris: Organization
for Economic Cooperation and Development.
Andrich, D. (1978). A rating formulation for ordered response categories.
Psychometrika, 43 , 561–573.
Bock, R. D., & Aitkin, M. (1981). Marginal maximum likelihood estimation
of item parameters: Application of an EM algorithm. Psychometrika,
46 , 443–459.
Bock, R. D., Gibbons, R., & Muraki, E. (1988). Full-information item factor
analysis. Applied Psychological Measurement, 12 , 261–280.
Bradlow, E. T., Wainer, H., & Wang, X. (1999). A Bayesian random effects
model for testlets. Psychometrika, 64 , 153–168.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fit.
In K. Bollen & J. Long (Eds.), Testing structural equation models (pp.
136–162). Newbury Park, CA: Sage.
Cacioppo, J. T., Petty, R. E., & Kao, C. F. (1984). The efficient assessment
of need for cognition. Journal of Personality Assessment, 48 , 306–307.
Cai, L. (2008). SEM of another flavour: Two new applications of the supplemented EM algorithm. British Journal of Mathematical and Statistical
Psychology, 61 , 309–329.
Cai, L. (2010a). A two-tier full-information item factor analysis model with
applications. Psychometrika, 75 , 581–612.
Cai, L. (2010b). High-dimensional exploratory item factor analysis by a
Metropolis-Hastings Robbins-Monro algorithm. Psychometrika, 75 , 33–
57.
Cai, L. (2010c). Metropolis-Hastings Robbins-Monro algorithm for confirmatory item factor analysis. Journal of Educational and Behavioral Statistics, 35 , 307–335.
Cai, L., & Hansen, M. (2012). Limited-information goodness-of-fit testing
of hierarchical item factor models. British Journal of Mathematical and
Statistical Psychology, 66 , 245–276.
189
Cai, L., Yang, J. S., & Hansen, M. (2011). Generalized full-information item
bifactor analysis. Psychological Methods, 16 , 221–248.
Celeux, G., Chauveau, D., & Diebolt, J. (1995). On stochastic versions of the
EM algorithm (Tech. Rep. No. 2514). The French National Institute for
Research in Computer Science and Control.
Celeux, G., & Diebolt, J. (1991). A stochastic approximation type EM algorithm for the mixture problem (Tech. Rep. No. 1383). The French
National Institute for Research in Computer Science and Control.
Chen, W. H., & Thissen, D. (1997). Local dependence indices for item pairs using item response theory. Journal of Educational and Behavioral Statistics, 22 , 265–289.
Choi, H.-J., Rupp, A. A., & Pan, M. (2013). Standardized diagnostic assessment design and analysis: Key ideas from modern measurement theory.
In M. M. C. Mok (Ed.), Self-directed learning oriented assessments in
the Asia-Pacific (pp. 61–85). New York, NY: Springer.
Cressie, N., & Read, T. R. C. (1984). Multinomial goodness-of-fit tests.
Journal of the Royal Statistical Society: Series B, 46 , 440–464.
de la Torre, J., & Douglas, J. A. (2004). Higher-order latent trait models for
cognitive diagnosis. Psychometrika, 69 , 333–353.
Delyon, B., Lavielle, M., & Moulines, E. (1999). Convergence of a stochastic
approximation version of the EM algorithm. The Annals of Statistics,
27 , 94–128.
Edwards, M. C. (2009). An introduction to item response theory using the
Need for Cognition scale. Social and Personality Psychology Compass,
3/4 , 507–529.
Edwards, M. C., & Cai, L. (2011, July). A new procedure for detecting departures from local independence in item response models. Paper presented
at the annual meeting of American Psychological Association, Washington, D.C. Retrieved from http://faculty.psy.ohio-state.edu/
edwards/documents/APA8.2.11.pdf
Gibbons, R. D., Bock, R. D., Hedeker, D., Weiss, D. J., Segawa, E., Bhaumik,
D. K., . . . Grochocinski, V. J. (2007). Full-information item bifactor
analysis of graded response data. Applied Psychological Measurement,
31 , 4–19.
Gibbons, R. D., & Hedeker, D. (1992). Full-information item bifactor analysis.
Psychometrika, 57 , 423–436.
Glas, C. A. W., Wainer, H., & Bradlow, E. T. (2000). Maximum marginal
likelihood and expected a posteriori estimation in testlet-based adaptive
190
testing. In W. J. van der Linden & C. A. W. Glas (Eds.), Computerized adaptive testing: Theory and practice (pp. 271–288). Boston, MA:
Kluwer Academic Publishers.
Gu, M. G., & Kong, F. H. (1998). A stochastic approximation algorithm
with Markov chain Monte-Carlo method for incomplete data estimation
problems. The Proceedings of the National Academy of Sciences, 95 ,
7270–7274.
Haberman, S. J. (1979). The analysis of qualitative data. New York: Academic
Press.
Hartz, S. M. (2002). A Bayesian framework for the unified model for assessing cognitive abilities: Blending theory with practicality. Unpublished
doctoral dissertation, Department of Statistics, University of Illinois at
Urbana-Champaign.
Hastings, W. K. (1970). Monte Carlo simulation methods using Markov chains
and their applications. Biometrika, 57 , 97–109.
Jamshidian, M., & Jennrich, R. I. (2000). Standard errors for EM estimation.
Journal of the Royal Statistical Society: Series B, 62 , 257–270.
Langer, M. M. (2008). A reexamination of Lord’s Wald test for differential item functioning using item response theory and modern error estimation. Unpublished doctoral dissertation, Department of Psychology,
University of North Carolina at Chapel Hill.
Lehman, A. F. (1988). A quality of life interview for the chronically mentally
ill. Evaluation and Program Planning, 11 , 51–62.
Li, Z., & Cai, L. (2012, July). Summed score likelihood based indices for testing
latent variable distribution fit in item response theory. Paper presented at
the annual International Meeting of the Psychometric Society, Lincoln,
NE. Retrieved from http://www.cse.ucla.edu/downloads/files/SD2
-final-4.pdf
Louis, T. A. (1982). Finding the observed information matrix when using
the EM algoritm. Journal of the Royal Statistical Society: Series B, 44 ,
226–233.
Masters, G. N. (1982). A Rasch model for partial credit scoring. Psychometrika, 47 , 149–174.
Maydeu-Olivares, A., Cai, L., & Hernandez, A. (2011). Comparing the fit
of IRT and factor analysis models. Structural Equation Modeling, 18 ,
333–356.
Maydeu-Olivares, A., & Joe, H. (2005). Limited and full information estimation and testing in 2n contingency tables: A unified framework. Journal
191
of the American Statistical Association, 100 , 1009–1020.
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., & Teller,
E. (1953). Equations of state space calculations by fast computing
machines. Journal of Chemical Physics, 21 , 1087–1092.
Mislevy, R. (1984). Estimating latent distributions. Psychometrika, 49 , 359–
381.
Muraki, E. (1992). A generalized partial credit model: Application of an EM
algorithm. Applied Psychological Measurement, 16 , 159–176.
Orlando, M., & Thissen, D. (2000). Likelihood-based item-fit indices for
dichotomous item response theory models. Applied Psychological Measurement, 24 , 50–64.
Preston, K., Reise, S., Cai, L., & Hays, R. (2011). Using the nominal response
model to evaluate response category discrimination in the PROMIS emotional distress item pools. Educational and Psychological Measurement,
71 , 523–550.
Robbins, H., & Monro, S. (1951). A stochastic approximation method. The
Annals of Mathematical Statistics, 22 , 400–407.
Roberts, G. O., & Rosenthal, J. S. (2001). Optimal scaling for various
Metropolis-Hastings algorithms. Statistical Science, 16 , 351–367.
Rupp, A. A., Templin, J., & Henson, R. A. (2010). Diagnostic measurement:
Theory, methods, and applications. New York: The Guilford Press.
Samejima, F. (1969). Estimation of latent ability using a response pattern of
graded scores. Psychometric monograph No. 17 .
Takane, Y., & de Leeuw, J. (1987). On the relationship between item response
theory and factor analysis of discretized variables. Psychometrika, 52 ,
393–408.
Thissen, D., Cai, L., & Bock, R. D. (2010). The nominal categories item
response model. In M. Nering & R. Ostini (Eds.), Handbook of polytomous item response theory models: Developments and applications (pp.
43–75). New York, NY: Taylor & Francis.
Thissen, D., & Steinberg, L. (1988). Data analysis using item response theory.
Psychological Bulletin, 104 , 385–395.
Tian, W., Cai, L., Thissen, D., & Xin, T. (2012). Numerical differentiation methods for computing error covariance matrices in item response theory modeling: An evaluation and a new proposal. Educational and Psychological Measurement. Advance online publication.
doi:10.1177/0013164412465875
Tittering, D. M. (1984). Recursive parameter estimation using incomplete
192
data. Journal of the Royal Statistical Society: Series B, 46 , 257–267.
Wirth, R. J., & Edwards, M. C. (2007). Item factor analysis: Current approaches and future directions. Psychological Methods, 12 , 58–79.
Woods, C. (2007). Empirical histograms in item response theory with ordinal
data. Educational and Psychological Measurement, 67 , 73–87.
Woods, C., Cai, L., & Wang, M. (2013). The Langer-improved Wald test for
DIF testing with multiple groups: Evaluation and comparison to twogroup irt. Educational and Psychological Measurement, 73 , 532–547.
193
Index
AddConst, see <Constraints>, AddConst
Parameter Estimates, 30, 31, 34,
Algorithm, see
59
<Options>, Algorithm
Rating Scale Model, 37–39
Alpha, see
CaseID, see <Groups>, CaseID
<Options>, Alpha
CaseWeight, see
Attributes, see
<Groups>, CaseWeight
<Groups>, Attributes
Cluster, see <Groups>, Cluster
Code, see <Groups>, Code
BetaPriors, see
Coeff, see <Constraints>, Coeff
<Groups>, BetaPriors
Cognitive Diagnostic Models, 99–103
Between, see <Groups>, Between
C-RUM, 104, 105
Between-Groups Factor, see
DINA, 111, 112, 122, 134, 135
Multidimensional IRT,
DINO, 115, 116
Multi-level Model
LCDM, 104, 111, 115, 123, 125,
Burnin, see <Options>,Burnin
132
Q-matrix, 103, 106, 120, 123
Calibration
Cognitive Psychometric Models, see
1PL, 24, 25, 97, 98
Cognitive Diagnostic Models
Parameter Estimates, 26, 98
2PL, 5, 6, 9, 58–60, 94, 96, 105, <Constraints>, 4, 5, 12, 16, 19, 21,
24, 30, 33, 38, 39, 41, 42, 58,
112, 116, 122, 125
157
Parameter Estimates, 8, 9, 96,
AddConst, 157, 160
107, 113
Coeff,
101, 102, 116, 117, 157, 160
3PL, 12, 40, 42, 43, 46, 48, 49
Cov, 41, 42, 49, 50, 54, 90, 94, 158–
Parameter Estimates, 13, 14
160
Graded Response Model, 20, 49,
Equal, 24, 25, 33, 38, 39, 41, 49,
53, 54, 90, 94
94, 97, 116, 117, 121, 122, 135,
Parameter Estimates, 21, 22
136, 157–159
Nominal Categories, 29, 30, 32, 33,
Fix, 30, 33, 38, 39, 54, 57, 60, 90,
57, 125, 156
94, 105, 106, 112, 116, 121,
Generalized Partial Credit Model,
122, 125, 126, 132, 135, 157–
36, 37
194
160
Free, 41, 49, 50, 54, 57, 60, 90, 94,
105, 106, 112, 116, 121, 122,
125, 126, 132, 135, 136, 157,
159, 160
Guessing, 12, 13, 25, 41, 46, 49,
158, 159
Interaction, 101, 102, 112, 116,
122, 125, 126, 132, 135
Intercept, 25, 39, 41, 49, 62, 158,
159
MainEffect, 101, 102, 105, 106,
112, 116, 122, 125, 126, 132,
135
Mean, 41, 42, 49, 50, 94, 158
Prior, 157, 161
Beta, 12, 41, 159, 161
logNormal, 12, 161
Normal, 12, 13, 46, 161
ScoringFn, 25, 30, 33, 38, 39, 57,
125, 126, 158, 159
Slope, 24, 25, 33, 38, 39, 41, 49,
54, 57, 60, 62, 90, 94, 97, 158,
159
Value, 61, 62, 125, 126, 157, 159,
160
Convergence Criteria, 4, 8, 30, 43, 144
CTX2Tbl, see <Options>, CTX2Tbl
Data Format, 3, 21, 153
Multiple Groups, 41
Response Pattern Data, 24, 154
Description, see
<Project>, Description
Diagnostic Classification Models, see
Cognitive Diagnostic Models
DIF, see <Options>, DIF
DIFcontrasts, see
<Options>, DIFcontrasts
DIFitems, see <Groups>, DIFitems
Dimensions, see
<Groups>, Dimensions
DM, see <Groups>, DM
DMtable, see
<Options>, DMtable
EmpHist, see <Groups>, EmpHist
Empirical Histograms, 42, 71, 72, 74,
155
Epsilon, see
<Options>, Epsilon
Equal, see <Constraints>, Equal
Etol, see <Options>, Etol
Factor Loadings, 25, 26, 55, 92
FactorLoadings, see
<Options>, FactorLoadings
File, see <Groups>, File
FisherInf, see
<Options>, FisherInf
FitNullModel, see
<Options>, FitNullModel
Fix, see <Constraints>, Fix
Fixed Effects Calibration, 97, 98
Fixed Theta Calibration, see
Fixed Effects Calibration
FixedTheta, see
<Groups>,FixedTheta
Free, see <Constraints>, Free
Generate, see <Groups>, Generate
GOF, see <Options>, GOF
Goodness of Fit (GOF), 145
Item-Level, 15
Marginal Chi-Square, 15, 22, 23
195
Orlando-Thissen-Bjorner S-X 2 , 25– Model, 5, 6, 153
27
GPC(), 37
Overall, 11, 12, 35, 36, 119
GPC(), 153
Limited Information M2 , 33, 35,
Graded(), 5, 6, 21, 24, 25, 54,
36
60, 62, 90, 94, 97, 105, 112,
Group Label, 4, 5
116, 122, 125, 132, 135, 153
<Groups>, 4, 5, 12, 16, 19, 21, 24, 25,
Nominal(), 30, 33, 38, 39, 57,
30, 33, 37–39, 41, 43–45, 49,
125, 153
54, 57, 60, 90, 94, 97
threePL, 12, 41, 43, 49
Attributes, 101, 105, 112, 116,
N, 5, 6, 12, 16, 19, 21, 24, 30, 33,
121, 122, 125, 132, 135, 156
37–39, 41, 43, 44, 49, 50, 54,
BetaPriors, 41, 43, 154, 156
57, 60, 62, 64, 69, 71, 73, 75,
Between, 50, 51, 77, 94, 154
77, 78, 90, 94, 97, 105, 112,
CaseID, 18, 50, 51, 154, 155
116, 122, 125, 132, 135, 151,
CaseWeight, 24, 25, 154
152
Cluster, 50, 51, 94, 154, 155
Ncats, 5, 6, 12, 21, 24, 30, 33, 37–
Code, 21, 33, 60, 151–153
39, 41, 43, 49, 54, 57, 58, 60,
DIFitems, 49, 148, 154, 157
90, 94, 97, 105, 112, 116, 122,
Dimensions, 37–39, 50, 51, 54, 57,
125, 132, 135, 151, 152
60, 73, 77, 78, 90, 94, 105, 112,
Nlevel2, 77, 78, 154
116, 121, 122, 125, 132, 135,
Primary, 37–39, 57, 58, 60, 73,
154
121, 154
DM, 101, 102, 105, 112, 116, 122,
Select, 24, 25, 33, 50, 94, 97, 151,
125, 135, 156
152
EmpHist, 43, 154, 155
Ta, 58, 154, 156
File, 5, 6, 12, 16, 19, 21, 24, 30,
Identity, 33, 125, 132
33, 37–39, 41, 43, 44, 49, 50,
Tc, 39, 58, 154, 156
54, 57, 60, 62, 64, 69, 71, 73,
Identity, 33
75, 77, 90, 94, 97, 105, 112,
Varnames, 5, 6, 12, 16, 19, 21, 24,
116, 122, 125, 132, 135, 151
25, 30, 33, 37–39, 41, 43–45,
FixedTheta, 97
49, 50, 54, 57, 60, 62, 64, 69,
Generate, 101, 102, 116, 135, 136,
71, 73, 75, 77, 90, 94, 97, 105,
156
112, 116, 122, 125, 132, 135,
InteractionEffects, 101, 112, 114,
151, 152
121, 122, 125, 132
HabermanResTbl, see
ItemWeight, 154, 155
<Options>, HabermanResTbl
Key, 151, 153
Missing, 50, 151, 152
196
Hierarchical Data, see
MinMLscore, see
Multidimensional IRT, Multi<Options>, MinMLscore
level Model
Missing, see <Groups>, Missing
Missing Data, 3, 152
Imputations, see
Mode, see <Options>, Mode
<Options>, Imputations
Model, see <Groups>, Model
Information Function, 10, 11, 42, 143, Models, see also Calibration
144
3PL, 163
InitGain, see
Generalized Formulation, 162
<Options>, InitGain
Graded Response Model, 164
InteractionEffects, see <Groups>,
Nominal Categories, 164
InteractionEffects
Monte Carlo Study, see Simulation
ItemWeights, see
Mstarts, see
<Groups>, ItemWeights
<Options>, Mstarts
Mtol, see <Options>, Mtol
JSI, see <Options>, JSI
Multidimensional IRT
Key, see <Groups>, Key
Bifactor Model, 37–39, 56–60, 119,
122
Local Dependence (LD), 15, 23, 28,
Correlated Factors, 53, 54, 56, 90,
146
94
Jackknife Slope Index (JSI), 25,
Multi-level Model, 49–52, 93–95
28, 147
Quadrature Points, 55, 58
Standardized Chi-square, 15, 22, Multiple Groups
23, 28, 145
Calibration, 40
logDetInf, see
Differential Item Functioning (DIF),
<Options>, logDetInf
45
Anchor Items, 49
M2, see <Options>, M2
Candidate DIF Items, 48
Marginal Reliability, 10, 11, 27
DIF Sweep, 45
MaxE, see <Options>, MaxE
Measurement
Invariance, 42, 79,
MaxM, see <Options>, MaxM
160
MaxMLscore, see
Scoring, 43
<Options>, MaxMLscore
MCsize, see
<Options>, MCsize
Metropolis-Hastings Robbins-Monro
Algorithm, 85, 86, 92, 93
MinExp, see <Options>, MinExp
N, see <Groups>, N
Ncats, see <Groups>, Ncats
Nested Data, see Multidimensional IRT,
Multi-level Model
197
NewThreadModel, see
<Options>, NewThreadModel
Nlevel2, see <Groups>, Nlevel2
Non-Normal Theta Estimation, see
Emprical Histograms
NormalMetric3PL, see
<Options>, NormalMetric3PL
NumDec, see <Options>, NumDec
<Options>, 4, 5, 12, 16, 19, 21, 24, 30,
33, 41, 44, 45, 138
Algorithm, 87, 90, 91, 94, 97, 149
Alpha, 87, 88, 149, 150
Burnin, 87, 88, 149, 150
CTX2Tbl, 145
DIF, 148
TestAll, 46, 148
TestCandidates, 49, 148
DIFcontrasts, 46, 49, 148
DMtable, 101, 139
Epsilon, 87, 88, 149, 150
Etol, 30, 33, 37, 38, 43, 57, 58,
105, 112, 116, 122, 125, 132,
135, 136, 141, 144
FactorLoadings, 24, 25, 54, 55,
90, 92, 145, 147
FisherInf, 41, 42, 141, 143, 144
FitNullModel, 33, 35, 41, 145, 147
GOF, 145
Basic, 11, 145
Complete, 24–26, 145
Extended, 12, 15, 21, 23, 33, 41,
125, 132, 145
HabermanResTbl, 24, 25, 28, 29,
33, 145, 146
Imputations, 87, 149
InitGain, 87, 88, 149, 150
JSI, 25, 28, 146
198
JSI, 145
LDX2, 24
logDetInf, 141, 144
M2, 145, 146
Full, 33, 35, 41
MaxE, 105, 112, 116, 122, 125, 132,
135, 136, 141, 144
MaxM, 105, 112, 116, 122, 125, 132,
135, 136, 141, 144
MaxMLscore, 142
MCsize, 87, 89, 94, 149, 151
MinExp, 145, 146
MinMLscore, 142
Mode, 4, 5, 138, 139, 141
Calibration, 4, 5, 12, 21, 24,
30, 33, 37–39, 41, 54, 57, 60,
87, 90, 94, 97, 105, 112, 116,
122, 125, 132, 135, 149
Scoring, 4, 16, 19, 43–45, 138
Simulation, 4, 62, 64, 69, 71,
73, 75, 77, 138, 142
Mstarts, 142
Mtol, 30, 33, 37, 38, 57, 58, 105,
112, 116, 122, 125, 132, 135,
136, 141, 144
NewThreadModel, 135, 136, 141, 144
NormalMetric3PL, 139, 140
NumDec, 139
OrthDIFcontrasts, 148
Perturbation, 143
PriorInf, 141, 144
Processors, 37, 38, 54, 57, 60, 90,
94, 135, 141, 144
Progress, 41, 42, 69, 138
ProposalStd, 87, 89–91, 94, 149,
151
ProposalStd2, 87, 89, 94, 149, 151
Quadrature, 38, 54, 55, 57, 58, 60,
90, 94, 122, 141, 144
ReadPRMFile, 16, 19, 43–45, 64,
69, 71, 73, 75, 77, 141, 142
Rndseed, 62, 64, 69, 71, 73, 75, 77,
87, 90, 94, 141, 142, 149
SaveCov, 105, 112, 116, 122, 125,
132, 135, 139
SaveDbg, 139
SaveInf, 41, 42, 139, 143
SavePrm, 12, 33, 41, 54, 60, 90,
105, 112, 116, 122, 125, 132,
135, 139
SaveSco, 16, 19, 21, 24, 26, 41,
105, 112, 116, 122, 125, 132,
135, 139
Score, 16, 141, 142
EAP, 16, 41, 105, 112, 116, 122,
125, 132, 135, 142
MAP, 16, 21, 43, 44, 142
ML, 16, 19, 20, 142
SSC, 16, 24, 26, 45, 112, 114,
115, 142
SE, 141, 143
FDM, 143
Fisher, 143
Mstep, 143
REM, 105, 112, 116, 122, 125, 132,
135, 136, 143
Sandwich, 143
SEM, 33, 143
Xpd, 41, 42, 143
SEMtol, 141, 144
SlopeThreshold, 22, 139, 140
SmartSEM, 143
SStol, 141, 144
Stage1, 87, 88, 97, 149, 150
Stage2, 87, 88, 97, 149, 150
Stage3, 87, 88, 149, 150
Stage4, 87, 88, 149, 150
StartValFromNull, 145, 147
SX2Tbl, 24–26, 33, 145, 146
TechOut, 139
Thinning, 87, 149
WindowSize, 87, 88, 149, 150
OrthDIFcontrasts, see
<Options>, OrthDIFcontrasts
Perturbation, see
<Options>, Perturbation
PISA data, 49
Primary, see <Groups>, Primary
Prior, see <Constraints>, Prior
Prior Distribution, 5, 12, 13, 41, 42,
46, 160
PriorInf, see
<Options>, PriorInf
Processors, see
<Options>, Processors
Progress, see <Options>, Progress
<Project>, 3–5, 8, 12, 16, 19, 21, 24,
30, 33, 41, 44, 45, 138
Description, 4, 5, 8, 12, 16, 19,
21, 24, 30, 33, 41, 44, 45, 54,
60, 62, 64, 69, 71, 73, 75, 77,
90, 94, 97, 105, 112, 116, 122,
125, 132, 135, 138
Title, 4, 5, 8, 12, 16, 19, 21, 24,
30, 33, 41, 44, 45, 54, 60, 62,
64, 69, 71, 73, 75, 77, 90, 94,
97, 105, 112, 116, 122, 125,
132, 135, 138
ProposalStd, see
<Options>, ProposalStd
ProposalStd2, see
<Options>,ProposalStd2
199
Value statements, 61, 62, 64
Multidimensional IRT
Bifactor Model, 72–74, 77–79
Multi-level Model, 77–79
ReadPRMFile, see
Multiple Groups, 62, 74–79
<Options>, ReadPRMFile
SlopeThreshold, see
Response Pattern Standardized Resid<Options>, SlopeThreshold
uals, 28, 29
SmartSEM, see <Options>, SmartSEM
Restricted Latent Class Models, see
SStol, see <Options>, SStol
Cognitive Diagnostic Models
Stage1, see <Options>, Stage1
Rndseed, see <Options>, Rndseed
Stage2, see <Options>, Stage2
Stage3, see <Options>, Stage3
SaveCov, see <Options>, SaveCov
Stage4, see <Options>, Stage4
SaveDbg, see <Options>, SaveDbg
Standard Error (SE) Method
SaveInf, see <Options>, SaveInf
EM algorithm, 143
SavePrm, see <Options>, SavePrm
Empirical Cross-Product ApproxSaveSco, see <Options>, SaveSco
imation, 8, 33, 42, 143
Score, see <Options>, Score
Fisher Information Matrix, 143
Scoring, 142
Forward Difference, 143
Expected a posteriori (EAP), 16,
Richardson Extrapolation, 105, 143
109, 129, 130, 142
Sandwich Covariance Matrix, 143
Maximum a posteriori (MAP), 20,
Supplemented EM, 33, 143
21, 43, 44, 142
StartValFromNull, see
Maximum Likelihood (ML), 142
<Options>, StartValFromNull
Sum Score Conversion to EAP (SSC),
SX2Tbl,
see <Options>, SX2Tbl
16, 27, 44, 45, 112, 114, 115,
Quadrature, see
<Options>, Quadrature
Quadrature Points, 8
142, 152
Supplying Item Parameters, 16, 18,
20, 43, 45
SE, see <Options>, SE
Select, see <Groups>, Select
SEMtol, see <Options>, SEMtol
Simulation, 61
2PL, 72–79
3PL, 68–72, 74–76
Graded Response Model, 61–65
Item Parameters, 63, 64, 69, 75,
76, 79
Parameter File, 64–76
Ta, see <Groups>, Ta
Tc, see <Groups>, Tc
TechOut, see <Options>, TechOut
Thinning, see <Options>, Thinning
Title, see <Project>, Title
Value, see <Constraints>, Value
Varnames, see <Groups>, Varnames
Weighted Data, see Data Format, Response Pattern Data
WindowSize, see
<Options>, WindowSize
200
">