sensitivity of teacher value-added estimates to student and peer control variables

12
Sensitivity of Teacher Value- Added Estimates to Student and Peer Control Variables March 2012 Presentation to the Association of Education Finance and Policy Conference Matt Johnson Stephen Lipscomb Brian Gill

Upload: bobby

Post on 05-Jan-2016

30 views

Category:

Documents


1 download

DESCRIPTION

Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables. March 2012 Presentation to the Association of Education Finance and Policy Conference Matt Johnson  Stephen Lipscomb  Brian Gill. VAMs Used Today Differ in Their Specifications. Research Questions. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Sensitivity of Teacher Value-Added Estimates to Student and Peer Control

VariablesMarch 2012

Presentation to the Association of Education Finance and Policy Conference

Matt Johnson Stephen Lipscomb Brian Gill

Page 2: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

VAMs Used Today Differ in Their Specifications

2

Value-Added ModelStudent

CharacteristicsClassroom

Characteristics

Multiple Years of

Prior Scores

Colorado Growth Model No No Yes

DC IMPACT Yes No No

Florida Yes Yes Yes

New York City Yes Yes No

SAS EVAAS No No Yes

Page 3: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

How sensitive are teacher value-added model (VAM) estimates to changes in the model specification?– Student characteristics– Classroom characteristics– Multiple years of prior scores

How sensitive are estimates to loss of students from sample due to missing prior scores?

Research Questions

3

Page 4: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Teacher value-added estimates are not highly sensitive to inclusion of:– Student characteristics (correlation ≥ 0.990)– Multiple years of prior scores (correlation ≥ 0.987)

Estimates are more sensitive to inclusion of classroom characteristics (correlation = 0.915 to .955)

Estimates are not very sensitive to loss of students with missing prior test scores from sample (correlation = 0.992)– Precision increases when two prior scores are used but

fewer teacher VAM estimates are produced

Preview of Main Results

4

Page 5: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Explore sensitivity to several specifications:– Exclude score from two prior years (Yi,t-2)

– Exclude student characteristics (Xi,t)

– Include class average characteristics

Student data from a medium-sized urban district for 2008–2009 to 2010–2011 school years

All models run using the same set of student observations

Instrument using opposite subject prior score to control for measurement error

Baseline Model

5

Page 6: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Student Level Class Level

Free or Reduced-Price Meals • •

Disability • •

Gifted Program Participation • •

Lagged Rate of Attendance • •

Lagged Fraction of Year Suspended • •

Race/Ethnicity •

Gender •

Age/Behind Grade Level •

Average Prior Achievement in Same Subject

Standard Deviation of Lagged Achievement

Number of Students in Classroom •

Student and Class Characteristics

6

Page 7: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Correlation of 6th-Grade Teacher Estimates Relative to Baseline VAM Specification

7

Math (n = 87)

Reading (n = 99)

Exclude Student Characteristics 0.990 0.996

Exclude Prior Score from t-2 0.993 0.987

Exclude Student Characteristics and Prior Score from t-2

0.978 0.970

Add class average variables 0.955 0.915

Baseline: Student Characteristics and Prior Scores from t-1 and t-2

Findings are based on VAM estimates from 2008–2009 to 2010–2011 on the same sample of students.

Page 8: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Exclude Student Characteristics

1st (Lowest) 2nd 3rd 4th

5th

(Highest)

Baseline Model

1st (Lowest) 95 5 0 0 0

2nd 5 90 5 0 0

3rd 0 5 75 20 0

4th 0 0 20 70 10

5th (Highest) 0 0 0 10 90

Percentage of 6th-Grade Reading Teachers in Effectiveness Quintiles, by VAM Specification

8

Findings are based on VAM estimates for 99 reading teachers in grade 6 from 2008–2009 to 2010–2011 for a medium-sized, urban district. Correlation with baseline = 0.996.

Page 9: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Baseline + Class Average Characteristics

1st (Lowest) 2nd 3rd 4th

5th

(Highest)

Baseline Model

1st (Lowest) 80 20 0 0 0

2nd 5 65 30 0 0

3rd 15 10 50 15 10

4th 0 5 10 65 20

5th (Highest) 0 0 10 20 70

Percentage of 6th-Grade Reading Teachers in Effectiveness Quintiles, by VAM Specification

9

Findings are based on VAM estimates for 99 reading teachers in grade 6 from 2008–2009 to 2010–2011 for a medium-sized, urban district. Correlation with baseline = 0.915.

Page 10: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Benefits of including two prior years:– More accurate measure of student ability– Increase in precision of estimates

Costs of using two prior years:– Students with missing prior scores dropped– Some teachers dropped from sample

Relative magnitude of costs/benefits?

One or Two Years of Prior Scores?

10

Page 11: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Estimate two VAMs using one year of prior scores– First VAM includes all students– Second VAM restricts sample to students with nonmissing

second prior year of scores

Correlation between teacher estimates: 0.992

Percentage of students dropped: 6.2

Percentage of teachers dropped: 3.9

Net increase in precision from using two prior years– Increase in average standard error of estimates: 2.3%

when students with missing scores are dropped– Decrease in average standard error of estimates: 7.6%

when second year of prior scores added

One or Two Years of Prior Scores?

11

Page 12: Sensitivity of Teacher Value-Added Estimates to Student and Peer Control Variables

Mathematica® is a registered trademark of Mathematica Policy Research.

Please contact– Matt Johnson

[email protected]

– Stephen Lipscomb• [email protected]

– Brian Gill• [email protected]

For More Information

12