detecting dif/dsf with pcmtrees detecting di erential item...
TRANSCRIPT
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Detecting Differential Item and
Differential Step Functioning with
Partial Credit Trees
Basil Abou El-Komboz, Achim Zeileis and Carolin Strobl
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Outline
Testing for DIF in the Rasch model
Standard model tests
Model-based recursive partitioning
Extending the model-based recursive partitioning approach
to the Partial Credit Model (PCM)
Differential item and step functioning in the PCM
(Un)ordered threshold parameters in the PCM
Visualization in Partial Credit trees
Example: Verbal Aggression data
Summary
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Differential Item Functioning (DIF)
is present when one or more items of a test
I are easier or harder to solve for certain subjects
I even though they have the same latent trait
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Standard model tests
I tests for k given groups
graphical test, Andersen’s Likelihood-Ratio Test,
Wald Tests
+ straightforward interpretation
− only detect DIF in specified groups
I latent-class approach
Rost’s “Mixed” (mixture) Rasch model
+ identifies previously unknown groups with DIF
− groups are not directly interpretable
⇒ 2nd step: describe groups with covariates
(e.g., Cohen and Bolt, 2005)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Standard model tests
●
●●●
●
●
●
●●
●
●
●
−3 −2 −1 0 1 2 3
−3
−2
−1
01
23
Geschlecht = Mann
Ges
chle
cht =
Fra
u 1
234
5
6
7
89
10
11
12
●
●●
●
●
●
●
●
●
●
●
●
(Mair, Hatzinger, and Maier, 2010, package eRm)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Standard model tests
I tests for k given groups
graphical test, Andersen’s Likelihood-Ratio Test,
Wald Tests
+ straightforward interpretation
− only detect DIF in specified groups
I latent-class approach
Rost’s “Mixed” (mixture) Rasch model
+ identifies previously unknown groups with DIF
− groups are not directly interpretable
⇒ 2nd step: describe groups with covariates
(e.g., Cohen and Bolt, 2005)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
New: Model-based recursive partitioning
+ identifies previously unknown groups with DIF
+ straightforward interpretationgender
p = 0.006
1
male female
agep < 0.001
2
≤ 34 > 34
Node 3 (n = 35)
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●●
●
1 20
−2.68
4.66Node 4 (n = 74)
●
● ● ●
●
● ●
●
●
●
●● ●
●●
●
●
●
●
●
1 20
−2.68
4.66Node 5 (n = 91)
● ●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●●
1 20
−2.68
4.66
function raschtree in package psychotree
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Approach used in psychotree takes care of...
I selecting splitting variables ⇔ parameter instability tests
scor
e co
ntrib
utio
ns
25 30 35 40 45
−0.
40
0.2
0.4
age
I selecting optimal cutpoints
I other multiple testing issues
I between variables in each split
I over successive splits
(Zeileis and Hornik, 2007; Zeileis, Hothorn, and Hornik, 2008;
Strobl, Malley, and Tutz, 2009; Strobl, Kopf, and Zeileis, 2010a,b)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Extending the model-based partitioning approach
Rasch trees genderp = 0.006
1
male female
agep < 0.001
2
≤ 34 > 34
Node 3 (n = 35)
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●●
●
1 20
−2.68
4.66Node 4 (n = 74)
●
● ● ●
●
● ●
●
●
●
●● ●
●●
●
●
●
●
●
1 20
−2.68
4.66Node 5 (n = 91)
● ●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●●
1 20
−2.68
4.66
Partial Credit trees
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Extending the model-based partitioning approach
Rasch trees genderp = 0.006
1
male female
agep < 0.001
2
≤ 34 > 34
Node 3 (n = 35)
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●●
●
1 20
−2.68
4.66Node 4 (n = 74)
●
● ● ●
●
● ●
●
●
●
●● ●
●●
●
●
●
●
●
1 20
−2.68
4.66Node 5 (n = 91)
● ●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●●
1 20
−2.68
4.66
Partial Credit trees
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Extending the model-based partitioning approach
Rasch model
I scores are 0 or 1
I each item has one location parameter = difficulty
I DIF means item is more/less difficult for certain group
Partial Credit model
I scores are between 0 and mj
I different parametrizations: e.g. mj thresholds
I DIF means entire item is more/less difficult
I DSF means some steps are more/less difficult
(may cancel out so there is no overall DIF)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
(Un)ordered threshold parameters in the PCM
−5 0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
P
(uij=
c|θ i
,δj1,..
.,δj3)
δj1 δj2 δj3
0 1 2 3
P(uij = c |θi , δj1, . . . , δjmj) =
e∑c
k=0(θi−δjk)
∑mj
l=0 e∑l
k=0(θi−δjk)
with∑0
k=0 (θi − δjk) = 0
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
(Un)ordered threshold parameters in the PCM
−5 0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
P(u
ij=c|
θ i,δ
j1,..
.,δj3)
δj1δj2 δj3
0 1
2 3
P(uij = c |θi , δj1, . . . , δjmj) =
e∑c
k=0(θi−δjk)
∑mj
l=0 e∑l
k=0(θi−δjk)
with∑0
k=0 (θi − δjk) = 0
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Visualization in Partial Credit trees
Want Curse
0
0.2
0.4
0.6
0.8
1
δ1δ1 δ2δ2
Do Curse
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Want Scold
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Do Scold
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Want Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1 δ2δ2
Do Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
−2 −1 0 1 2
−2 −1 0 1 2
Category Characteristic Curves
Pro
babi
lity
Latent Trait
Cat. 0 Cat. 1 Cat. 2
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Visualization in Partial Credit trees
Want Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1 δ2δ2
Do Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
−2 −1 0 1 2 3
−2 −1 0 1 2 3
Category Characteristic Curves
Prob
abilit
y
Latent Trait
Cat. 0 Cat. 1 Cat. 2
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Visualization in Partial Credit trees
Want Curse
0
0.2
0.4
0.6
0.8
1
δ1δ1 δ2δ2
Do Curse
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Want Scold
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Do Scold
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
Want Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1 δ2δ2
Do Shout
0
0.2
0.4
0.6
0.8
1
δ1δ1δ2δ2
−2 −1 0 1 2
−2 −1 0 1 2
Category Characteristic Curves
Pro
babi
lity
Latent Trait
Cat. 0 Cat. 1 Cat. 2
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Visualization in Partial Credit trees
Wan
t Cur
se
0
0.2
0.4
0.6
0.81
δ 1δ 1δ 2δ 2
Do
Cur
se
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
Wan
t Sco
ld
0
0.2
0.4
0.6
0.81
δ 1δ 1δ2δ 2
Do
Sco
ld
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
Wan
t Sho
ut
0
0.2
0.4
0.6
0.81
δ 1δ 1δ 2δ 2
Do
Sho
ut
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
−2
−1
01
2
−2
−1
01
2
Cat
egor
y C
hara
cter
istic
Cur
ves
Probability
Late
nt T
rait
Cat
. 0C
at. 1
Cat
. 2
Late
nt tr
ait
−2
−1
01
2
−2
−1
01
2
Want Curse Do Curse Want Scold Do Scold Want Shout Do Shout
inspired by “effect plots”
(Fox and Hong, 2009, package effects)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Visualization in Partial Credit trees
Wan
t Cur
se
0
0.2
0.4
0.6
0.81
δ 1δ 1δ 2δ 2
Do
Cur
se
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
Wan
t Sco
ld
0
0.2
0.4
0.6
0.81
δ 1δ 1δ2δ 2
Do
Sco
ld
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
Wan
t Sho
ut
0
0.2
0.4
0.6
0.81
δ 1δ 1δ 2δ 2
Do
Sho
ut
00.2
0.4
0.6
0.8
1
δ 1δ 1δ 2δ 2
−2
−1
01
2
−2
−1
01
2
Cat
egor
y C
hara
cter
istic
Cur
ves
Probability
Late
nt T
rait
Cat
. 0C
at. 1
Cat
. 2
Late
nt tr
ait
−2
−1
01
2
−2
−1
01
2
Want Curse Do Curse Want Scold Do Scold Want Shout Do Shout
inspired by “effect plots”
(Fox and Hong, 2009, package effects)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Example: Verbal Aggression data
> data("VerbalAggression", package = "psychotools")
responses of 316 subjects to frustrating situations
I here: situation 4 (self-to-blame situation)
“The operator disconnects me when I used up my last 10
cents for a call.”
I items: 3 verbally aggressive responses (curse, scold, shout)
× 2 behavioural models (want, do)
I response categories: 0 = no, 1 = perhaps, 2 = yes
I covariates: gender, trait anger (assessed by the Dutch
adaptation of the state-trait anger scale STAS)
De Boeck and Wilson (2004), Smits, De Boeck, and Vansteelandt
(2004), dichotomized version also available in package difR
(Magis, Beland, and Raiche, 2011)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Example: Verbal Aggression data
genderp = 0.027
1
female male
Node 2 (n = 243)
WantCurse
DoCurse
WantScold
DoScold
WantShout
DoShout
−2
−1
0
1
2
Late
nt tr
ait
angerp = 0.198
3
≤ 21 > 21
Node 4 (n = 49)
WantCurse
DoCurse
WantScold
DoScold
WantShout
DoShout
−2
−1
0
1
2
Late
nt tr
ait
Node 5 (n = 24)
WantCurse
DoCurse
WantScold
DoScold
WantShout
DoShout
−2
−1
0
1
2
Late
nt tr
ait
Partial Credit tree (tweaked a little for visualization)
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Summary
model-based recursive partitioning
I can identify groups of subjects with DIF and DSF that
I need not be pre-specified
I are formed by (combinations of) observed covariates
I with optimally selected cutpoints
I available for
I Rasch model
I Partial Credit model
I and more to come
I results are directly interpretable
, but keep in mind:
observed covariates may be proxies for the true causes
e.g.: gender ⇔ socialization, district ⇔ first language
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Summary
model-based recursive partitioning
I can identify groups of subjects with DIF and DSF that
I need not be pre-specified
I are formed by (combinations of) observed covariates
I with optimally selected cutpoints
I available for
I Rasch model
I Partial Credit model
I and more to come
I results are directly interpretable
, but keep in mind:
observed covariates may be proxies for the true causes
e.g.: gender ⇔ socialization, district ⇔ first language
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Summary
model-based recursive partitioning
I can identify groups of subjects with DIF and DSF that
I need not be pre-specified
I are formed by (combinations of) observed covariates
I with optimally selected cutpoints
I available for
I Rasch model
I Partial Credit model
I and more to come
I results are directly interpretable
, but keep in mind:
observed covariates may be proxies for the true causes
e.g.: gender ⇔ socialization, district ⇔ first language
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
Summary
model-based recursive partitioning
I can identify groups of subjects with DIF and DSF that
I need not be pre-specified
I are formed by (combinations of) observed covariates
I with optimally selected cutpoints
I available for
I Rasch model
I Partial Credit model
I and more to come
I results are directly interpretable, but keep in mind:
observed covariates may be proxies for the true causes
e.g.: gender ⇔ socialization, district ⇔ first language
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
References I
Cohen, A. and D. Bolt. “A Mixture Model Analysis of
Differential Item Functioning.” Journal of Educational
Measurement 42 (2005): 133–148.
De Boeck, P. and M. Wilson, editors. Explanatory Item
Response Models: A Generalized Linear and Nonlinear
Approach. New York: Springer, 2004.
Fox, John and Jangman Hong. “Effect Displays in R for
Multinomial and Proportional-Odds Logit Models:
Extensions to the effects Package.” Journal of
Statistical Software 32 (2009): 1–24.
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
References II
Magis, D., S. Beland, and G. Raiche. difR: Collection of
methods to detect dichotomous differential item
functioning (DIF) in psychometrics, 2011. R package
version 4.1.
Mair, Patrick, Reinhold Hatzinger, and Marco Maier. eRm:
Extended Rasch Modeling., 2010. R package version
0.13-0.
Smits, D., P. De Boeck, and K. Vansteelandt. “Inhibition of
Verbally Aggressive Behaviour.” European Journal of
Personality 18 (2004).
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
References III
Strobl, C., J. Kopf, and A. Zeileis. A New Method for
Detecting Differential Item Functioning in the Rasch
Model. Technical Report 92, Department of Statistics,
Ludwig-Maximilians-Universitat Munchen, Germany, 2010.
URL: http://epub.ub.uni-muenchen.de/11915/.
Strobl, C., J. Kopf, and A. Zeileis. “Wissen Frauen weniger
oder nur das Falsche? – Ein statistisches Modell fur
unterschiedliche Aufgaben-Schwierigkeiten in
Teilstichproben.” Allgemeinbildung in Deutschland –
Erkenntnisse aus dem SPIEGEL Studentenpisa-Test. Ed.
S. Trepte and M. Verbeet Wiesbaden: VS Verlag, 2010,
255–272.
Detecting
DIF/DSF with
PCMtrees
Testing for DIF in
the RM
Standard tests
Model-based recursive
partitioning
Extension to the
PCM
DIF/DSF in the PCM
(Un)ordered threshold
parameters
Visualization
Example: Verbal
Aggression data
Summary
References
References IV
Strobl, C., J. Malley, and G. Tutz. “An Introduction to
Recursive Partitioning: Rationale, Application and
Characteristics of Classification and Regression Trees,
Bagging and Random Forests.” Psychological Methods 14
(2009): 323–348.
Zeileis, A., T. Hothorn, and K. Hornik. “Model-Based
Recursive Partitioning.” Journal of Computational and
Graphical Statistics 17 (2008): 492–514.
Zeileis, Achim and Kurt Hornik. “Generalized M-Fluctuation
Tests for Parameter Instability.” Statistica Neerlandica 61
(2007): 488–508.