Date post: | 27-Mar-2015 |
Category: |
Documents |
Upload: | isabel-gordon |
View: | 213 times |
Download: | 0 times |
Copyright © 2006 Educational Testing Service
Listening. Learning. Leading.
Using Differential Item Functioning to Investigate
the Impact of Accommodations on the Scores of
Students with Disabilities on English-Language Arts
Assessments
Mary Pitoniak, Linda Cook, Frederic Cline, and Cara Cahalan-Laitusis
Educational Testing ServiceNCME Presentation
April 10, 2006
Copyright © 2006 Educational Testing Service 2
P 2
Purpose and Overview of the Study
• The purpose of this study was to examine differential item functioning on the English-Language Arts assessment described by Linda
• DIF analyses are statistical procedures that are used to identify items that function differently for different subgroups of examinees
• DIF “exists when examinees of equal ability differ, on average, according to their group membership in their responses to a particular item” (Standards)
Copyright © 2006 Educational Testing Service 3
P 3
Purpose and Overview of the Study (continued)
• Issues investigated:
– Do 2 different DIF detection methods yield the same results?
– Are the results interpretable in terms of a priori or a posteriori evaluation of item content?
– Of particular interest:When the read-aloud modification is used, do the items function differentially for students?
Copyright © 2006 Educational Testing Service 4
P 4
Purpose and Overview of the Study (continued)
• Features of study:
– 2 DIF detection methods– Large enough sample sizes (not always the
case)– Looked at 3 different criteria (total score,
Reading score, Writing score); we decided to go with total score for several reasons
– Used purification step, as recommended by literature
Copyright © 2006 Educational Testing Service 5
P 5
Comparisons Made in the Study
Comparison Number Reference Group Focal Group
1.3 Without disabilities LD no accommodations
1.4 “ LD IEP/504 accommodations
1.5 “ LD read-aloud modification
(& IEP/504 accommodations)
3.1 LD no accommodations
LD IEP/504 accommodations
3.2 “ LD read-aloud modification
(& IEP/504 accommodations)
Copyright © 2006 Educational Testing Service 6
P 6
DIF Methods Used
• Mantel-Haenszel
• Logistic Regression
• For both methods, we used ETS classification system:
– Category A contains items with negligible DIF;– Category B contains items with slight to
moderate values of DIF;– Category C contains items with moderate to
large values of DIF.
Advantages Disadvantages
Mantel-Haenszel
1. Computationally simple
2. Most powerful DIF detection method when DIF is constant (or uniform) and group mean abilities are equal
3. Low Type 1 error rates when the compared groups have equal mean abilities
4. A relatively small sample size is needed (e.g., 200-250 per group) for reasonable power
5. Effect size measure is relatively more sensitive to actual DIF conditions
1. Power is negatively affected by inadequate sample size and unequal group mean abilities
Logistic Regression
1. Relatively powerful in detecting uniform DIF
2. Superior power in detecting nonuniform DIF
3. Method is very flexible and allows different models to be specified and it can handle multiple ability estimates
1. Computationally intensive
2. The effect size measure (ΔR2) and set guidelines are not as sensitive to the actual magnitude of DIF
Comparison of Mantel-Haenszel vs. Logistic Regression
Example of Uniform DIF
0%
20%
40%
60%
80%
100%
0 16 28 40 52 64
Total Score
Pe
rce
nt
Co
rre
ct
Example of Non-Uniform DIF
0%
20%
40%
60%
80%
100%
0 16 28 40 52 64
Total Score
Pe
rce
nt
Co
rre
ct
Copyright © 2006 Educational Testing Service 10
P 10
Results
• Within this presentation, I will present results only for Reading items (and not Writing), both for time reasons and because we were most interested in the effects of the accommodations on performance on the Reading items
Copyright © 2006 Educational Testing Service 11
P 11
Results (continued)
• Overall
– No items flagged as “C”
– Each method flagged 9 items as “B” (out of 42 items X 5 comparisons, or 210 possible flags)
– However, those 9 items were not the same items—in all, 12 different items were flagged by at least one of the methods
– There were inconsistencies between methods
Copyright © 2006 Educational Testing Service 12
P 12
Number of Items Flagged by Each Method
Method Reference
Group Focal Group
Mantel-Haenszel
Logistic Regression
Non-LD LD no acc
1 6
“ LD IEP/504
1 3
“ LD read-aloud
6 0
LD no acc
LD IEP/504
0 0
“ LD read-aloud
1 0
Total 9 9
Agreement Between Flags for Methods by Comparison Type
Reference Group
Focal Group Agreement
Discrepancy: Uniform vs.
no flag
Discrepancy: Non-uniform vs. uniform
Non-LD LD no acc
36 1 5
“ LD IEP/504
39 2 1
“ LD read-aloud
36 6 0
LD no acc
LD IEP/504
42 0 0
“ LD read-aloud
41 1 0
Mantel-Haenszel Logistic
Regression No flag Favors non-LD
Favors LD no accomm
Total
No flag 36 36
Favors non-LD
0
Favors LD no accomm
1 1
Non-uniform 1
1 1 2
Non-uniform 2
3 3
Total 41 1 0 42
Non-LD vs.
LD No Accommodation
Mantel-Haenszel Logistic
Regression No flag Favors non-LD
Favors LD IEP/504
Total
No flag 39 39
Favors non-LD
0
Favors LD IEP/504
2 2
Non-uniform 1
1 1
Non-uniform 2
0
Total 41 1 0 42
Non-LD vs.
LD IEP/504 Accommodation
Mantel-Haenszel Logistic
Regression No flag Favors non-LD
Favors LD Read-Aloud
Total
No flag 36 1 5 42
Favors non-LD
Favors LD Read-Aloud
Non-uniform 1
Non-uniform 2
Total 36 1 5 42
Non-LD vs.
LD Read-Aloud Modification
Mantel-Haenszel Logistic
Regression No flag Favors
LD Non-Acc Favors LD
IEP/504
Total
No flag 42 42
Favors LD Non-Acc
Favors LD IEP/504
Non-uniform 1
Non-uniform 2
Total 42
LD Non-Accommodated vs.
LD IEP/504 Accommodation
Mantel-Haenszel Logistic
Regression No flag Favors
LD-Non Acc. Favors LD
Read-Aloud
Total
No flag 40 42
Favors LD-Non Acc.
1
Favors LD Read-Aloud
1
Non-uniform 1
Non-uniform 2
Total 42 42
LD Non-Accommodated vs.
LD Read-Aloud Modification
Example of Discrepancies in Flags
Item Flags
M-H—Uniform LR—No flag
The items flaggedby MH (but not LR) as favoring students with read-aloud modification did show differences such as these graphically for LR
0%
20%
40%
60%
80%
100%
0 16 28 40 52 64
Total Score
Pe
rce
nt
Co
rre
ct
Example of Discrepancies in Flags
0%
20%
40%
60%
80%
100%
0 16 28 40 52 64
Total Score
Pe
rce
nt
Co
rre
ct
Item Flags
M-H—Uniform
LR—No flag
Copyright © 2006 Educational Testing Service 21
P 21
A Priori Theories About Read-Aloud Modification Results
• 5 items were easier for students who received the read-aloud modification than for non-LD students.
• A priori theories were not that accurate!– Item A: harder (refer back)
– Item B: easier (short item; intonation/body language)
– Item C: easier (intonation/body language)
– Item D: harder (char. of options)
– Item E: harder (length of options)
Copyright © 2006 Educational Testing Service 22
P 22
A Posteriori Interpretation About Read-Aloud Modification Results
• The reasons why these 5 items were easier with read-aloud accommodation were not obvious to test developers
Copyright © 2006 Educational Testing Service 23
P 23
What Do the Results Say About the 3 Questions Posed
• Do 2 different DIF detection methods yield the same results?
– Neither flagged an item as “C.”
– There were discrepancies in “B” flags, however.
– Some discrepancies are explainable in terms of advantages/disadvantages of methods as listed earlier.
Copyright © 2006 Educational Testing Service 24
P 24
• Are the results interpretable in terms of a priori or a posteriori evaluation of item content? – Not consistently
• Of particular interest:When the read-aloud modification is used, do the items function differentially for students?– Yes, some items were easier when
read-aloud, which supports this state’s decision to view read-aloud as a modification
3 Questions (continued)
Copyright © 2006 Educational Testing Service 25
P 25
• ELL and ELL/LD groups to be compared
• Grade 8 ELA to be evaluated
• DIF analysis paradigm to be utilized
Next Steps