1 detection of item degradation yongwei yang abdullah ferdous tzu-yun chin university of...
TRANSCRIPT
1
Detection of Item Degradation
Yongwei Yang
Abdullah Ferdous
Tzu-Yun Chin University of Nebraska-Lincoln
In T. L. Hayes (chair), Item degradation: impact, detection, and mitigation, an academic-practitioner collaborative forum conducted at the 22nd annual conference of the Society of Industrial and Organizational Psychology in New York City, NY, April 2007.
2
Item Degradation Item Degradation
Item’s favorable psychometric characteristics deteriorate over time Psychometric characteristics
Content relevance and representativeness Technical characteristics (e.g., “difficulty”/“location”, lack of
bias) Utility (e.g., item-criterion relationship)
Item Degradation vs. Exposure/Compromise Item degradation: observed phenomenon Item exposure/compromise:
Items have become known to test takers prior to administration
Possible reasons for degradation
3
Detection of Item Degradation
Essentially it is about investigating the comparability of item’s psychometric properties over time “temporal stability of the psychometric
characteristics” (Chan, Drasgow, & Sawin, 1999)
Can be evaluated under the framework of: Measurement invariance (MI; Meredith, 1993) Predictive invariance (PI; Millsap, 1995)
Item Degradation as MI or PI
Measurement Invariance (MI)
Same relationship across populations between observed indicators and the latent variables
Degradation noninvariance in such relationships over time Loading, location
4
( | , ) ( | )F x w v F x w= Predictive Invariance (PI)
Same relationship across populations between predictors and criterion
Degradation noninvariance in such relationships over time Indicator-criterion
relationship
( | , ) ( | )F y x v F y x=
Let x be observed indicator that measures latent w and predicts y,
and v be some population indicator
5
Item Degradation Detection Methods
Differential item functioning, item parameter drift
Mean & covariance modeling Assessing invariance in various aspects
pertain to measurement or predictive properties
Statistical process control
Models of change
6
Item Degradation Detection
Differential item functioning, item parameter drift
Mean & covariance modeling Assessing invariance in various aspects
pertain measurement or predictive properties
Statistical process control Cumulative sum (CUSUM) procedure
Models of change
7
CUSUM for Item Degradation Detection
Our approach—Conditional CUSUM Whether item parameters have deviated from target Make use of observed scores The importance of controlling for shifts in traits level over
time “Conditional”—test takers at different time points were matched based on
their total test score
Procedures Initial Item Calibration
Compute target item parameter (e.g., difficulty) using the first n job applicants from the operation sample
Define “time group” Every m applicants from the n+1 applicant to the last person under
investigation Define “trait group” (conditioning variable)
Divide job applicants into groups of reasonable size based on total test scores
Compute and plot CUSUM statistics for each trait group separately
8
Conditional CUSUM—Calculation Two-sided Standardized CUSUM
⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢
⎣
⎡
+−
+
−= +
−+
1
0
20
2
0,0max i
i
i
ii Ck
nn
XXC
σσ0
12 20
0
min 0, ii i
i
i
X XC k C
n n
σ σ− −
−
⎡ ⎤⎢ ⎥
−⎢ ⎥= + +⎢ ⎥⎢ ⎥+⎢ ⎥⎣ ⎦
Initial Status Item VarianceTime Group i Item Variance
Time Group i Item MeanTarget Item Mean
Reference value (k) and Control limit (h)
9
Conditional CUSUM—Data Source A web-based personnel selection assessment
for selecting managers 103 items measuring job-related non-cognitive
attributes CTT-based test construction and scoring Fixed-length, linear test Unproctored
Sample: Job applicants from Oct. 2002 to Sept. 2005 Re-taker excluded Total N = 7,000
10
Conditional CUSUM—Results Among the 103 items
36 flagged for upward shift in item means for at least one trait group
20 flagged for downward shift in item means for at least one trait group
9 flagged for having both upward and downward shifts for different trait groups
38 not flagged for any trait group
A couple examples: it035, it174
Follow-up analysis: Were there differences across item types with respect to the
likelihood of being flagged by conditional CUSUM?
Conditional CUSUM—Follow-up Multinomial logistic
regression DV: condition CUSUM flag;
3 categories; “Not Flagged” as the reference category
IV: ability (6 levels), item type (3 levels, multiple choice (MC) as the reference group
11
78.1% 5.2%16.7 %
86.8 % 5.3% 7.9%
79.9 % 11.9% 8.2%
80.6% 8.4% 11.0%
Ite m T y pe
Forward (n=210 )
Revers e (n =114)
M C ( n=29 4)
Total (n=618 )
Not
Flagge d
Flagg ed for
Downw ard
Shift
Flagg ed for
Upward S hift
Conditional C USUM Fl ag
78.1% 5.2%16.7 %
86.8 % 5.3% 7.9%
79.9 % 11.9% 8.2%
80.6% 8.4% 11.0%
Ite m T y pe
Forward (n=210 )
Revers e (n =114)
M C ( n=29 4)
Total (n=618 )
Not
Flagge d
Flagg ed for
Downw ard
Shift
Flagg ed for
Upward S hift
Conditional C USUM Fl ag
Results GOF statistic indicates appropriate fit of the main effect model
(X2=16.83, df=20, p=.664) The impact of ability levels on the CUSUM flags was not statistically
significant (X2=13.48, df=10, p=.198) The impact of item type on the CUSUM flags was statistically
significant (X2=17.83, df=4, p=.001). MC items were more likely to be flagged by conditional CUSUM for
negative shifts Forward items were more likely to be flagged by conditional
CUSUM for positive shifts
Model of Change Perspective 1:
Understanding patterns of change using examinee characteristics Do the trajectories of item parameter change vary across
different types of examinees? Applicant location, SES, demographics, etc.
Perspective 2: Understanding patterns of change using item characteristics Do the trajectories of item parameter change vary across
different types of items? Item format, complexity, content area, etc.
Formulating these questions in a longitudinal analysis framework
12
Perspective 1 Example
13
Using a 2-level longitudinal model to explore: RQ1: On average, was there a shift in item difficulty? RQ2: Were there variations in the slope of the shift? (If Yes to RQ2) RQ3: Could the variations be explained by job applicants
characteristics (e.g., trait level, region, etc.)? The model:
Analysis with item 174: RQ1: significant positive
slope RQ2: non-significant
variations RQ3: not pursued
0 1
0 00 0
1 10 1
( )ti i i ti ti
i i
i i
Y time e
r
r
π ππ βπ β
= + += += +
Level I:
Level II:
Perspective 2 Example
14
Using a 2-level longitudinal model to explore: RQ1: Across items, on
average was there a change in item difficulty over time?
RQ2: Were there variations in the slope of the change across items?
(If Yes to RQ2) RQ3: Could the variations be explained by item characteristics?
Model B:
Analysis with this data set: RQ3: item type did not
explain a significant portion of the variations in slopes
Perspective 2 Example Model A:
Analysis with this data set: RQ1: average slope
across items was not different from zero
RQ2: significant variations in slopes across items
15
0 1 2
0 00 0
1 10
2 20 2
( ) ( )ti i i t i ti ti
i i
i
i i
Y trait time e
r
r
π π ππ βπ βπ β
= + + += +== +
0 1 2
0 00 0
1 10
2 20 20 2
( ) ( )
( _ )
ti i i t i ti ti
i i
i
i i i
Y trait time e
r
item type r
π π ππ βπ βπ β β
= + + += +== + +
Level I
Level II
Summary and Discussions Two types of methods that serve different purposes:
Statistical process control (e.g., CUSUM): Real-time monitoring of degradation We illustrated conditional CUSUM procedure, but other methods exist
(e.g., an IRT-based moving residual approach by Han & Hambleton [2004])
Explicit modeling of patterns of degradation: Understanding the nature of degradation, exploring potential factors that
impact degradation, assisting the development of prevention and mitigation procedures
We illustrated longitudinal modeling methods, but various methods for studying MI/PI may be applied
These methods can also be used in monitoring and understanding degradation in other parameters (e.g., item variance, discrimination, response time) It might be helpful to monitor/model multiple parameters
simultaneously to (1) “flag” items more accurately and, (2) understand factors behind degradation
16
Summary and Discussions Understanding temporal stability of
measurement properties is essential to: Valid decisions based on test scores Valid inferences in substantive research based on
assessment outcomes Research on Flynn effect (e.g., Wicherts et al., 2004)
Further research is needed, such as What monitoring approaches would better fit personnel
selection assessment programs? What would lead to or impact degradation? How would item-level degradation impact test-level
decisions and inferences? Etc.
17
18
Some Useful References MI & PI Concepts
Mellenbergh (1989) Meredith (1993) Millsap (1995)
Various IPD and Item Exposure Detection Methods Bock, Muraki, & Pfeiffenberger (1988) Chan, Drasgow, & Sawin (1999) DeMars (2004) Donahue & Isham (1998) Han & Hambleton (2004) Kim, Cohen, & Park (1995)
CUSUM and Psychometric Applications: Hawkins & Olwell (1998) Meijer & van Krimpen-Stoop (2003) Montgomery (2005) van Krimpen-Stoop & Meijer (2002) Veerkamp & Glas (2000)
19
Contacts
Yongwei Yang: [email protected] Ferdous:
[email protected] Chin: [email protected]
THANK YOU