1. BRIDGING THE GAP
BETWEEN ITEM
RESPONSE THEORY AND
STRUCTURAL EQUATION
MODELLING
PhD thesis Eveline Gebhardt
2. Background
Two related traditions: IRT and SEM
Measuring constructs (IRT)
Relationships between constructs (SEM)
Formally equivalent
Some remaining differences:
IRTmore flexibility measurement models
And in data structure and guessing parameter
SEM more flexible structural models
SEM quicker
7. PhD thesis
Describing similarities and differences
between the measurement models of IRT and
SEM
Finding matching models
Describing estimation of structural models of
SEM (2SLS)
Adding structural features to current IRT
models
Testing and applying of additional features
8. Expanding structural features
Multi-step method
1. Run original IRT model in ACER ConQuest
2. Construct SSCP matrix of the structural part
from CQ output and permute this matrix
3. Apply two-stage least squares (2SLS)
estimation method
10. 1. Running original IRT model in ACER
ConQuest
Latent variables are dependent
Observed variable are independent
11. 2. Using OLS equations - 1
To obtain the full SSCP of the structural model
LL LO
SSCP
OL OO
From ACER ConQuest’s output:
conditional ˆc
covariance matrix of latent variablesO(
L|
)
ˆ
regression c
B coefficients between latent and
observed variables ( )
12. 2. Using OLS equations - 2
Imagine a general regression model
L = OΒ + U
With the OLS equations
1
Β=LO OO
1
UU LL LO OO LO
We can construct each part of the SSCP using
the ConQuest parameters from previous slide
13. 2. Using OLS equations - 3
ˆ
L O Bc O O
OL ˆ
O O Bc
L'L = U ¢ + L'O (O'O ) (L ¢ )¢
-1
U O
¢
U ˆ ˆ
= U ¢ + Bc (Bc (O ¢ ))
O
U ˆ O ˆ
= U ¢ + Bc (O ¢ )B c¢
ˆ ˆ O ˆ
= Σ cL O (N - K ) + Bc (O ¢ )B c¢
( )
é ù
UU ˆ O ˆ ˆ
éL¢ L¢ ù ê ¢ + Bc (O¢ )Bc¢ Bc (O¢ )ú
L O O
SSCP = ê ú= ê ú
ê ¢ O¢ ú ê
ë L
O Oû
ê
ë O ˆ
(O¢ )Bc¢ O¢ ú
O úû
14. 3. Permuting the original SSCP
Latent and observed variables are redefined
as endogenous (Y) and exogenous (X)
variables
Endogenous variables are being explained by the
model (l1, l2, o5)
Exogenous variables only explain other variables
(o1, o2, o3, o4)
And *then reordered using the following
Y'Y Y'X
SSCP
structure
X'Y X'X
15. 4. Applying 2SLS - 1
Two-stage least squares is the most common
estimation method for path models
m is for the current equation
A is for variables included in m
A 1 A A A A 1
Ym X X X X Ym Ym X m γm
A
Ym X X X X Ym
A A
X m Ym A A
Xm Xm βm
A
A
X m Ym
1
A 1 A A A A 1
γ A
m
Y X XX
m XY m Y X
m m Ym X X X X Ym
β A
m
A A
X m Ym A A
Xm Xm A
X m Ym
16. 4. Applying 2SLS - 2
Additional estimates
Standard errors
1
A 1 A A A
2
Y X XX
m XY
m Y X
m m
S s
A A A A
X m Ym Xm Xm
Explained variance
ee
R2 1 2
2 1
Ym N Ym
17. Simulation study
Part 1 - estimated structural parameters were
evaluated by comparing them with the true
structural parameters and their standard error
Part 2 - the same structural parameters were
compared with parameters as estimated by the
2SLS procedure in SPSS and a path model
procedure in Mplus.
Part 3 - results from the full model, including the
measurement model, were compared between the
multi-step approach and a structural equation
model using Mplus
18. Simulation study – part 1
Generate student responses using known
(true) parameters of a hypothetical model (100
replications)
Running original IRT model in ACER
ConQuest
Applying multi-step method
Compare average estimates with true
parameters
19. General model
Number of indicators and values of parameters can be manipulated for
each analysis
20. ,
Set of three equations
Equation 1
,
Equation 2
Equation 3
21. Three simulations
k m-k
Number of
indicators SIM1 10 30
per latent SIM2 30 10
variable
SIM3 50 50
R-squared of Equation 1 Equation 2 Equation 3
each SIM1 .750 .360 .190
equation SIM2 .098 .020 .004
SIM3 .098 .020 .004
25. Standard errors SIM 1
True SE (SD parameter) Mean SE
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
26. Standard errors SIM 2
True SE (SD parameter) Mean SE
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
27. Standard errors SIM 3
True SE (SD parameter) Mean SE
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
28. Conclusions
Regression coefficients are accurately
estimated
R-squared estimates are accurately estimated
Standard errors are equal to standard errors
from 2SLS in SPSS, but different from the SD
of the 100 estimates of each parameters
Measurement error should be included in the
standard errors when measurement model is
included in the full model
Editor's Notes
Developed models to analyse social dataBoth concerned with measuring constructs that cannot be observed directlyEducational science vs econometrics & psychologyAnd with analysing relationships between constructsFormally equivalent if putting constraints on each of the modelsIRT flexibility in data structure and guessing parameterTwo parts in a SEM or modern IRT model.Guessing parameter
Measuring constructs that cannot be directly observed
Two sets of responses:Observed variablesIndicators of the latent variables
Flip between 2 and 3 to show measurement error is missing