Polytomous Item Explanatory IRT Models with Random Item Effects: Concepts, Simulations, Applications Jinho Kim University of California, Berkeley Presented on October 2, 2018 at the BEAR Seminar, Berkeley Way West 1217
Polytomous Item Explanatory IRT Models with Random Item Effects:
Concepts, Simulations, Applications
Jinho KimUniversity of California, Berkeley
Presented on October 2, 2018at the BEAR Seminar, Berkeley Way West 1217
2
CONTENTS
1. Research Goals & Backgrounds2. Polytomous Item Response Models (with Fixed Item Effects)
2.1. Partial Credit Model2.2. Item Explanatory Extensions of the Partial Credit Model
3. Polytomous Item Explanatory Models with Random Item Effects3.1. Polytomous Random Item Effects 3.2. Polytomous Item Explanatory Models with Random Item Errors
4. Estimation: Bayesian methods5. Simulation Studies
5.1. Study 1: Model Comparison5.2. Study 2: Parameter Recovery
6. Empirical Study: An application to the verbal aggression data 7. Conclusion and Discussion
3
DESIGN OF VERBAL AGGRESSION ITEMS
• Verbal aggression items were designed using three design factors: (1) the Behavior Mode factor: Want and Do (2) the Behavior Type factor: Curse, Scold, and Shout(3) the Situation Type factor: Other-to-blame and Self-to-blame
• Four frustrating situations were used for the item stem.(a) Other-to-blame: A bus fails to stop for me. (Bus),
I miss a train because the clerk gave me faulty information. (Train)(b) Self-to-blame: The grocery store closes just as I am about to enter. (Store)
The operator disconnects me when I used up my last 10 cents for a call. (Call)
• 24 items were written to fit a 2 × 3 × 2 design with two replications within each cell
3
44
Item Behavior Mode Behavior Type Situation Type
Num. Situation Want Do Curse Scold ShoutOther-to-blame
Self-to-blame
01 Bus 1 0 1 0 0 1 002 Bus 1 0 0 1 0 1 003 Bus 1 0 0 0 1 1 0~ ~
13 Bus 0 1 1 0 0 1 014 Bus 0 1 0 1 0 1 015 Bus 0 1 0 0 1 1 0~ ~
DESIGN OF VERBAL AGGRESSION ITEMS
• The item 01: “A bus fails to stop for me. I would want to curse.”• This item was constructed by Want (Behavior Mode), Curse (Behavior Type),
and Other-to-blame (Situation Type):
5
DESIGN OF VERBAL AGGRESSION ITEMS
• Sample Size:- A total of 316 persons responded to the 24 items without missing values.
• Polytomous Data: - The item responses were three ordered-category responses:
“no” = 0, “perhaps” = 1, and “yes” = 2.
• Item Properties:- The three design factors were considered as categorical item properties (or
features) for item explanatory models.- The categorical item properties were dummy coded:
Want, Shout, and Self-to-blame were a reference for each item property.
5
6
RESEARCH GOALS
• How can we explain or predict the polytomous item difficulties by observed item properties?
• Considering the uncertainty in explanation or prediction, how can we add random item errors to account for residual variation in the polytomous item difficulties?
• How can we develop and estimate polytomous item explanatory models with random item errors?
• In what ways do the proposed models work in simulation and empirical studies?
6
7
BACKGROUNDS• LLTM+e approach as a random item explanatory IRT model
– The Linear Logistic Test Model (LLTM; Fischer, 1973) decomposes the difficulties of specific items into linear combinations of elementary components such as item properties.
– The Linear Logistic Test Model with item error (LLTM+e; Janssen et al., 2004) can consider the uncertainty in explanation and enhance the precision of estimation of the item difficulties by accounting for residual variation.
• LLTM+e approach to polytomous data– Extensions of the LLTM+e approach to polytomous data is more complicated and
demanding than to dichotomous data.– (1) item parameterization with item properties in the polytomous models is
complicated,– (2) adding random item errors to the polytomous models is also complicated, – (3) allowing random effects on both item and person sides makes for crossed
random effects, which is demanding for estimation.
8
Polytomous Item Response Models (with Fixed Item Effects)
1. Partial Credit Model (PCM)
2. Item Explanatory Extensions of the Partial Credit Model: Many-
Facet Rasch Model (MFRM) and Linear Partial Credit Model
(LPCM) approaches
8
9
Partial Credit Model (PCM) with two different item parameterizations
𝑙𝑙𝑙𝑙𝑃𝑃𝑃𝑃 𝑦𝑦𝑝𝑝𝑝𝑝 = 𝑚𝑚 𝜃𝜃𝑝𝑝
𝑃𝑃𝑃𝑃 𝑦𝑦𝑝𝑝𝑝𝑝 = 𝑚𝑚 − 1 𝜃𝜃𝑝𝑝= 𝜃𝜃𝑝𝑝 − 𝛽𝛽𝑝𝑝 + 𝜏𝜏𝑝𝑝𝑖𝑖 , 𝑚𝑚 = 1, … ,𝑀𝑀𝑝𝑝 , 𝜏𝜏𝑝𝑝𝑖 = 0, and∑𝑖𝑖=1
𝑀𝑀𝑖𝑖 𝜏𝜏𝑝𝑝𝑖𝑖 = 0
𝛽𝛽𝑝𝑝 is an item location parameter for item i, interpreted as the overall item difficulty, 𝜏𝜏𝑝𝑝𝑖𝑖 is a step deviation parameter for the m-th step within item i
• PCM with onefold item parameterization (original parameterization)
𝑙𝑙𝑙𝑙𝑃𝑃𝑃𝑃 𝑦𝑦𝑝𝑝𝑝𝑝 = 𝑚𝑚 𝜃𝜃𝑝𝑝
𝑃𝑃𝑃𝑃 𝑦𝑦𝑝𝑝𝑝𝑝 = 𝑚𝑚 − 1 𝜃𝜃𝑝𝑝= 𝜃𝜃𝑝𝑝 − 𝛿𝛿𝑝𝑝𝑖𝑖, 𝑚𝑚 = 1, … ,𝑀𝑀𝑝𝑝,𝛿𝛿𝑝𝑝𝑖 = 0, and 𝜃𝜃𝑝𝑝~𝑁𝑁 0,𝜎𝜎𝜃𝜃2
𝛿𝛿𝑝𝑝𝑖𝑖 is a step difficulty parameter for scoring m rather than 𝑚𝑚 − 1 on item 𝑖𝑖
• PCM with twofold item parameterization (ConQuest parameterization)
9
10
Item Explanatory Extensions of the PCM using two polytomous item explanatory approaches
Polytomous Rasch Model Partial Credit Model
Item Parameterization 𝛿𝛿𝑝𝑝𝑖𝑖 = 𝛽𝛽𝑝𝑝 + 𝜏𝜏𝑝𝑝𝑖𝑖 (two fold) 𝛿𝛿𝑝𝑝𝑖𝑖 (one fold)
Target Item Difficulty Overall Item Difficulty 𝛽𝛽𝑝𝑝 Step Difficulty 𝛿𝛿𝑝𝑝𝑖𝑖
Relevant ItemExplanatory Approach
Many-Facet Rasch Model (MFRM; Linacre, 1989)
with a Decomposition of the Item Location Parameters
Linear Partial Credit Model (LPCM; Fischer and Ponocny, 1994)
with a Decomposition of the Step Difficulty Parameters
Item Explanatory Model with
Fixed Item Effects
𝛿𝛿𝑝𝑝𝑖𝑖′ = �𝑘𝑘=𝑖
𝐾𝐾
𝛾𝛾𝑘𝑘𝑥𝑥𝑝𝑝𝑘𝑘 + 𝜏𝜏𝑝𝑝𝑖𝑖
∑𝑖𝑖=1𝑀𝑀𝑖𝑖 𝜏𝜏𝑝𝑝𝑖𝑖 = 0
𝛿𝛿𝑝𝑝𝑖𝑖′ = �𝑘𝑘=𝑖
𝐾𝐾
𝜔𝜔𝑘𝑘𝑖𝑖𝑥𝑥𝑝𝑝𝑘𝑘
Item Property Effects item property effects 𝛾𝛾𝑘𝑘 step specific item property effects 𝜔𝜔𝑘𝑘𝑖𝑖10
11
Polytomous Item Explanatory Models with Random Item Effects
1. Polytomous Random Item Effects
2. Polytomous Item Explanatory Models with Random Item Errors
11
12
Polytomous Random Item Effects
12
PCM’s Item Parameterization
Random Item Parameter
Polytomous Random Item Effects Distribution Assumption
Two fold𝛿𝛿𝑖𝑖𝑚𝑚 = 𝛽𝛽𝑖𝑖 + 𝜏𝜏𝑖𝑖𝑚𝑚
Item Location Parameter 𝛽𝛽𝑖𝑖
Overall ItemRandom Effects
𝛽𝛽𝑖𝑖 ~ 𝑁𝑁 𝜇𝜇𝛽𝛽,𝜎𝜎𝛽𝛽2
and ∑𝑚𝑚=1𝑀𝑀𝑖𝑖 𝜏𝜏𝑖𝑖𝑚𝑚 = 0
One fold𝛿𝛿𝑖𝑖𝑚𝑚
Step Difficulty Parameter 𝛿𝛿𝑖𝑖𝑚𝑚
Step SpecificRandom Item Effects
𝜹𝜹𝒊𝒊 𝛿𝛿𝑖𝑖𝑖, … ,𝛿𝛿𝑖𝑖𝑚𝑚 ′~ 𝑀𝑀𝑀𝑀𝑁𝑁𝑚𝑚 𝝁𝝁,𝚺𝚺
One fold𝛿𝛿𝑖𝑖𝑚𝑚
Step Difficulty Parameter 𝛿𝛿𝑖𝑖𝑚𝑚
Item-StepRandom Effects 𝛿𝛿𝑖𝑖𝑚𝑚 ~ 𝑁𝑁 𝜇𝜇𝛿𝛿 ,𝜎𝜎𝛿𝛿2
𝝁𝝁𝜹𝜹 = �𝜇𝜇𝛿𝛿1𝜇𝜇𝛿𝛿2⋮
𝜇𝜇𝛿𝛿𝑚𝑚� , 𝚺𝚺 =
⎣⎢⎢⎡ 𝜎𝜎𝛿𝛿1
2 𝜎𝜎𝛿𝛿1𝛿𝛿2 ⋯ 𝜎𝜎𝛿𝛿1𝛿𝛿𝑚𝑚
𝜎𝜎𝛿𝛿2𝛿𝛿1 𝜎𝜎𝛿𝛿22 ⋯ 𝜎𝜎𝛿𝛿2𝛿𝛿𝑚𝑚
⋮ ⋮ ⋱ ⋮𝜎𝜎𝛿𝛿𝑚𝑚𝛿𝛿 1 𝜎𝜎𝛿𝛿𝑚𝑚𝛿𝛿 2 ⋯ 𝜎𝜎𝛿𝛿𝑚𝑚2 ⎦
⎥⎥⎤,
13
Polytomous Item Explanatory Models with Random Item Errors
13
PCM’s Item Parameterization
Target Item Difficulty
Parameter
Relevant Explanatory
Approach
Polytomous Random Item Error Model Specification
Two fold𝛿𝛿𝑖𝑖𝑚𝑚 = 𝛽𝛽𝑖𝑖 + 𝜏𝜏𝑖𝑖𝑚𝑚
Item Location(Overall Item
Difficulty) Parameter 𝛽𝛽𝑖𝑖
Many-Facet Rasch Model
𝛽𝛽𝑝𝑝′ = �𝑘𝑘=𝑖
𝐾𝐾
𝛾𝛾𝑘𝑘𝑥𝑥𝑝𝑝𝑘𝑘
Overall ItemRandom
Error (OIE) 𝜖𝜖𝑖𝑖
𝛿𝛿𝑖𝑖𝑚𝑚 = �𝑘𝑘=0
𝐾𝐾
𝛾𝛾𝑘𝑘𝑥𝑥𝑖𝑖𝑘𝑘 + 𝜖𝜖𝑖𝑖 + 𝜏𝜏𝑖𝑖𝑚𝑚
𝜖𝜖𝑖𝑖~ 𝑁𝑁 0,𝜎𝜎𝜖𝜖2
∑𝑚𝑚=1𝑀𝑀𝑖𝑖 𝜏𝜏𝑖𝑖𝑚𝑚 = 0
One fold𝛿𝛿𝑖𝑖𝑚𝑚
Step Difficulty Parameter 𝛿𝛿𝑖𝑖𝑚𝑚
Linear Partial Credit Model
𝛿𝛿𝑖𝑖𝑚𝑚′ = �𝑘𝑘=0
𝐾𝐾
𝜔𝜔𝑘𝑘𝑚𝑚𝑥𝑥𝑖𝑖𝑘𝑘
Step SpecificRandom Item Error (SSE) 𝜉𝜉𝑖𝑖𝑚𝑚
𝛿𝛿𝑖𝑖𝑚𝑚 = �𝑘𝑘=0
𝐾𝐾
𝜔𝜔𝑘𝑘𝑚𝑚𝑥𝑥𝑖𝑖𝑘𝑘 + 𝜉𝜉𝑖𝑖𝑚𝑚
𝝃𝝃𝒊𝒊 𝜉𝜉𝑖𝑖𝑖, … , 𝜉𝜉𝑖𝑖𝑚𝑚 ′ ~ 𝑀𝑀𝑀𝑀𝑁𝑁𝑚𝑚 𝟎𝟎,𝚺𝚺
Item-StepRandom
Error (ISE) 𝜀𝜀𝑖𝑖𝑚𝑚
𝛿𝛿𝑖𝑖𝑚𝑚 = �𝑘𝑘=0
𝐾𝐾
𝜔𝜔𝑘𝑘𝑚𝑚𝑥𝑥𝑖𝑖𝑘𝑘 + 𝜀𝜀𝑖𝑖𝑚𝑚
𝜀𝜀𝑖𝑖𝑚𝑚~ 𝑁𝑁 0,𝜎𝜎𝜀𝜀2
1414
• Crossed random effects in random item effects IRT models
• Intractable and demanding integration for crossed random effects models due to high-dimensional integrals and prohibitive numerical integration
• Practical difficulty in estimating polytomous random item effects IRT models due to lack of statistical software using an adjacent-categories logit link
• I used the RStan R package implementing a No-U-Turn sampler (NUTS), which is more efficient and robust than Gibbs sampling
Estimation of Crossed Random Effects Models: Bayesian methods
15
Simulation Studies
1. Simulation Design and Data Generation
2. Study 1: Model Comparison
3. Study 2: Parameter Recovery
15
16
• Two simulation studies were conducted:- Model comparison study (one replication) to compare the performance of the
proposed models with the fixed item effects models and the PCM.- Parameter recovery study (30 replications) to examine model parameter recovery
when the simulated datasets were analyzed using a data generating model.- In both simulation studies, the effects of model misspecification on parameter
estimation were also examined.
• Two different explanatory conditions in both simulation studies: - Using the concept of the coefficient of determination (R2), the low and high information conditions (R2 = 0.3 or 0.9) were set for data generation.
• Two polytomous item explanatory approaches in both simulation studies: - Because the MFRM and the LPCM are conceptually and functionally different approaches, the MFRM + OIE and the LPCM + SSE were used for data generation.
16
Simulation Design and Data Generation
17
Data Generation using the MFRM + OIE(Suppose a reading comprehension test situation)
17
Information condition 𝑅𝑅2 = 0.3 (Low) 𝑅𝑅2 = 0.9 (High)
Model parameter Generating valueFixed effects
Intercept 𝛾𝛾0 − 1.00 − 1.00Global comprehension 𝛾𝛾1 0.80 0.80Evaluative reasoning 𝛾𝛾2 0.60 0.60Inferential reasoning 𝛾𝛾3 0.40 0.40Text complexity 𝛾𝛾4 0.20 0.20
Step deviation (1st step) 𝜏𝜏𝑖𝑖1− 0.10 for 𝑖𝑖 ∈ 1, 3, … , 29,− 0.20 for 𝑖𝑖 ∈ 2,4, … , 30
− 0.10 for 𝑖𝑖 ∈ 1, 3, … , 29,− 0.20 for 𝑖𝑖 ∈ 2,4, … , 30
Step deviation (2nd step) 𝜏𝜏𝑖𝑖2 𝜏𝜏𝑖𝑖𝑖 = − 𝜏𝜏𝑖𝑖𝑖 𝜏𝜏𝑖𝑖𝑖 = − 𝜏𝜏𝑖𝑖𝑖Random effects
Overall item error variance 𝜎𝜎𝜖𝜖2 0.852 0.192
Person variance 𝜎𝜎𝜃𝜃2 1.502 1.502
18
Data Generation using the LPCM + SSE
18
Information condition 𝑅𝑅2 = 0.3 (Low) 𝑅𝑅2 = 0.9 (High)Model parameter Generating value
Fixed effectsIntercept 𝜔𝜔01 − 1.50 − 1.50
𝜔𝜔02 − 0.50 − 0.50Global comprehension 𝜔𝜔11 0.50 0.50
𝜔𝜔12 1.10 1.10Evaluative reasoning 𝜔𝜔21 0.40 0.40
𝜔𝜔22 0.80 0.80Inferential reasoning 𝜔𝜔31 0.20 0.20
𝜔𝜔32 0.60 0.60Text complexity 𝜔𝜔41 0.10 0.10
𝜔𝜔42 0.30 0.30Random effects
Item error variance (1st step) 𝜎𝜎𝜉𝜉12 0.512 0.112
Item error variance (2nd step) 𝜎𝜎𝜉𝜉22 1.202 0.262
Correlation between the steps 𝜌𝜌𝜉𝜉12 0.00 0.00Person variance 𝜎𝜎𝜃𝜃2 1.502 1.502
19
1. Model Comparison Study(1) MFRM approach-based simulation
- Goodness-of-fit of the models fitted to the MFRM + OIE generation data
19
The three random item effects models showed a significantly superior goodness-of-fit in both information conditions.
20
1. Model Comparison Study(1) MFRM approach-based simulation
- Graphical agreement of the step difficulties between the models
20
In terms of reconstructing the step difficulties, the three random item effects models significantly performed better than the models without errors in both information conditions.
21
1. Model Comparison Study(1) MFRM approach-based simulation
- Effects of model misspecification via scatter plots with 95% CIs
21
• The MFRM + OIE is a data generating model, and the MFRM is a misspecifiedmodel
• The MFRM is inadequate for making accurate statistical inferences such as estimation and hypothesis testing for the item property effects, particularly in the low information condition.
• The MFRM + OIE is better for making accurate statistical inferences in that its CIs contain the generating values for all the parameters
22
1. Model Comparison Study(2) LPCM approach-based simulation
- Goodness-of-fit of the models fitted to the LPCM + SSE generation data
22
The three random item effects models showed a significantly superior goodness-of-fit in both information conditions.
23
1. Model Comparison Study(2) LPCM approach-based simulation
- Graphical agreement of the step difficulties between the models
23
In terms of reconstructing the step difficulties, the three random item effects models significantly performed better than the models without errors in both information conditions.
24
1. Model Comparison Study(2) LPCM approach-based simulation
- Effects of model misspecification via scatter plots with 95% CIs
24
• The LPCM + SSE is a data generating model, and the LPCM and the LPCM + ISE are misspecified models.
• The LPCM is inadequate for making accurate statistical inferences such as estimation and hypothesis testing
25
2. Parameter Recovery Study(1) MFRM approach-based simulation
- Parameter recover over 30 replications via bias plots with 95% Cis
25
26
2. Parameter Recovery Study(1) MFRM approach-based simulation
- Averaged estimates and empirical standard error (SE) for parameters
26
𝑹𝑹𝟐𝟐 = 0.3 𝑹𝑹𝟐𝟐 = 0.9 MFRM + OIE MFRM MFRM + OIE MFRM
Gen Estimate SE Estimate SE Estimate SE Estimate SE 𝛾𝛾0 -1.0 -1.10 0.41 -0.89 0.06 -0.99 0.12 -0.98 0.08 𝛾𝛾1 0.8 0.93 0.33 0.73 0.03 0.81 0.08 0.79 0.03 𝛾𝛾2 0.6 0.64 0.41 0.51 0.03 0.57 0.10 0.56 0.03 𝛾𝛾3 0.4 0.49 0.41 0.40 0.03 0.38 0.10 0.38 0.03 𝛾𝛾4 0.2 0.20 0.12 0.16 0.01 0.20 0.03 0.20 0.01
• The standard errors for the MFRM + OIE were larger than those for the MFRM in both
information conditions. Much larger in the low information condition.
• The MFRM + OIE is more robust and accurate for parameter estimation and statistical inference, compared to the MFRM
27
2. Parameter Recovery Study(2) LPCM approach-based simulation
- Parameter recover over 30 replications via bias plots with 95% Cis
27
28
2. Parameter Recovery Study(2) LPCM approach-based simulation
- Averaged estimates and empirical standard error (SE) for parameters
28
• The standard errors for the LPCM + SSE and the LPCM + ISE were larger than those for the LPCM in both information conditions. Much larger in the low information condition.
• The LPCM + SSE and the LPCM + ISE are more robust and accurate for parameter estimation and statistical inference, compared to the LPCM.
LPCM + SSE LPCM + ISE LPCM Gen Estimate SE Estimate SE Estimate SE
𝑹𝑹𝟐𝟐 = 0.3 𝜔𝜔01 -1.5 -1.51 0.27 -1.51 0.43 -1.35 0.08 𝜔𝜔02 -0.5 -0.45 0.59 -0.44 0.43 -0.35 0.08 𝜔𝜔11 0.5 0.55 0.21 0.55 0.35 0.50 0.05 𝜔𝜔12 1.1 1.19 0.48 1.18 0.35 0.96 0.05 𝜔𝜔21 0.4 0.42 0.26 0.42 0.43 0.38 0.05 𝜔𝜔22 0.8 0.85 0.57 0.85 0.43 0.70 0.06 𝜔𝜔31 0.2 0.19 0.26 0.20 0.43 0.17 0.06 𝜔𝜔32 0.6 0.61 0.57 0.61 0.43 0.52 0.05 𝜔𝜔41 0.1 0.10 0.08 0.09 0.12 0.09 0.02 𝜔𝜔42 0.3 0.24 0.17 0.24 0.12 0.19 0.02
𝑹𝑹𝟐𝟐 = 0.9 𝜔𝜔01 -1.5 -1.53 0.11 -1.54 0.13 -1.53 0.09 𝜔𝜔02 -0.5 -0.51 0.15 -0.52 0.13 -0.51 0.09 𝜔𝜔11 0.5 0.50 0.06 0.50 0.09 0.49 0.05 𝜔𝜔12 1.1 1.08 0.11 1.08 0.09 1.07 0.05 𝜔𝜔21 0.4 0.39 0.08 0.39 0.11 0.39 0.06 𝜔𝜔22 0.8 0.80 0.14 0.80 0.11 0.79 0.06 𝜔𝜔31 0.2 0.19 0.08 0.19 0.11 0.19 0.06 𝜔𝜔32 0.6 0.61 0.14 0.61 0.11 0.60 0.06 𝜔𝜔41 0.1 0.10 0.02 0.10 0.03 0.10 0.02 𝜔𝜔42 0.3 0.30 0.04 0.30 0.03 0.30 0.02
29
Empirical Study: An application to the verbal aggression data
29
30
Model Comparison: Goodness-of-fit
30
Model k DIC LOOIC WAIC
PCM 49 12312.49 12324.91 12320.29 MFRM 30 12448.42 12456.71 12452.81 MFRM + OIE 31 12312.39 12322.23 12317.86 LPCM 11 12454.69 12464.35 12460.36 LPCM + SSE 14 12299.66 12308.67 12304.47 LPCM + ISE 12 12307.09 12315.88 12311.65
• It was revealed that the order of DIC, LOOIC, and WAIC across the models was
consistent: the LPCM + SSE, the LPCM + ISE, the MFRM + OIE, the PCM, the MFRM, and the LPCM fit better to the verbal aggression data in sequence.
31
Model Comparison: Graphical agreement of the step difficulties
31
• In terms of reconstructing the step difficulties, the three random item effects models significantly performed better than the models without errors in both information conditions.
• Compared to the LPCM in the bottom panels, the MFRM in the top panel had similar dispersion of the points but a few points were farther from the line
32
Item Property Effects in the MFRM-based polytomous item explanatory models
32
Predictor Parameter MFRM MFRM + OIEEstimate SE Estimate SE
Fixed effectsIntercept 𝛾𝛾0 1.58 0.08 1.69 0.18Behavior Mode (Do) 𝛾𝛾𝐷𝐷𝐷𝐷 0.43 0.04 0.49 0.15Situation Type (Other-to-blame) 𝛾𝛾𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑃𝑃 -0.82 0.04 -0.89 0.14Behavior Type (Curse) 𝛾𝛾𝐶𝐶𝐶𝐶𝑃𝑃𝐶𝐶𝑂𝑂 -1.28 0.05 -1.38 0.17Behavior Type (Scold) 𝛾𝛾𝑆𝑆𝑆𝑆𝐷𝐷𝑙𝑙𝑆𝑆 -0.63 0.05 -0.70 0.18Step deviation parametersfor each item
𝜏𝜏11 ~ 𝜏𝜏241Random effectsOverall item error variance 𝜎𝜎𝜖𝜖2 0.322 0.07Person variance 𝜎𝜎𝜃𝜃2 0.952 0.05 0.972 0.05
• How does the Situation Type affect the overall item difficulties?
3333
Predictor Parameter LPCM LPCM + SSE LPCM + ISEEstimate SE Estimate SE Estimate SE
Fixed effectsIntercept 𝜔𝜔01 1.43 0.08 1.45 0.18 1.45 0.17
𝜔𝜔02 1.77 0.12 1.96 0.19 1.90 0.20Behavior Mode (Do) 𝜔𝜔𝐷𝐷𝐷𝐷𝑖 0.54 0.06 0.53 0.14 0.56 0.14
𝜔𝜔𝐷𝐷𝐷𝐷𝑖 0.36 0.07 0.44 0.15 0.42 0.16Situation Type (Other-to-blame) 𝜔𝜔𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑃𝑃𝑖 -0.70 0.06 -0.71 0.15 -0.70 0.15
𝜔𝜔𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑃𝑃𝑖 -0.97 0.07 -1.09 0.16 -1.06 0.16Behavior Type (Curse) 𝜔𝜔𝐶𝐶𝐶𝐶𝑃𝑃𝐶𝐶𝑂𝑂𝑖 -1.68 0.08 -1.71 0.17 -1.71 0.19
𝜔𝜔𝐶𝐶𝐶𝐶𝑃𝑃𝐶𝐶𝑂𝑂𝑖 -0.93 0.10 -1.11 0.18 -1.06 0.20Behavior Type (Scold) 𝜔𝜔𝑆𝑆𝑆𝑆𝐷𝐷𝑙𝑙𝑆𝑆𝑖 -0.80 0.07 -0.83 0.19 -0.84 0.18
𝜔𝜔𝑆𝑆𝑆𝑆𝐷𝐷𝑙𝑙𝑆𝑆𝑖 -0.50 0.10 -0.61 0.19 -0.56 0.20Random effectsCorrelation between the steps 𝜌𝜌𝜉𝜉𝑖𝑖 0.82 0.16Item error variance (1st step) 𝜎𝜎𝜉𝜉1
2 0.332 0.07Item error variance (2nd step) 𝜎𝜎𝜉𝜉2
2 0.312 0.08Item-Step error variance 𝜎𝜎𝜀𝜀2 0.332 0.05Person variance 𝜎𝜎𝜃𝜃2 0.952 0.05 0.972 0.05 0.972 0.05
Item Property Effects in the LPCM-based models• How does the Situation Type affect the step difficulties for each step?
34
CONCLUSION AND DISCUSSION• The results show that
– Random item effect models performed better than fixed item effect models.– By taking the uncertainty in explanation and prediction into account, the random
item effects models were more robust and accurate for parameter estimation and statistical inferences.
– In the MFRM with item errors, we could see how item properties have an effect on the overall item difficulties.
– In the LPCM with item errors, we could see how item properties have an effect on the step difficulties for each step.
• Potential value of the proposed models– Can predict item locations or step difficulties of the newly developed items in a
more scientific and systematic rather than intuitive manner.– Can enhance the prediction of the polytomous item difficulties.– Can be a methodological foundation for advanced measurement models.
35
REFERENCE
Linacre, J. M. (1989). Multi-facet Rasch measurement. Chicago: MESA Press.Masters, G. N. (1982). A Rasch model for partial credit scoring. Psychometrika, 47(2), 149-174.Fischer, G. H. (1973). The linear logistic test model as an instrument in educational research. Acta Psychologi
ca, 37, 359-374.Fischer, G. H., & Ponocny, I. (1994). An extension of the partial credit model with an application to the measu
rement of change. Psychometrika, 59(2), 177-192.Janssen, R., Schepers, J., & Peres, D. (2004). Models with item and item group predictors. In De Boeck, P. &
Wilson, M. (Eds.), Explanatory item response models: A generalized linear and nonlinear approach (pp. 189-212). New York: Springer-Verlag.
Spiegelhalter, D., Thomas, A., & Best, N. (2003). WinBUGS version 1.4 [Computer program]. UK: MRC Biostatistics Unit, Institute of Public Health, Robinson Way, Cambridge CB2 2SR.
Van den Noortgate, W., De Boeck, P., & Meulders, M. (2003). Cross-classification multilevel logistic modelsin psychometrics. Journal of Educational and Behavioral Statistics, 28(4), 369-386.
35
Thank [email protected]