julian r. betts and y. emily tang, university of california, san diego
DESCRIPTION
Julian R. Betts and Y. Emily Tang, University of California, San Diego ( [email protected] , [email protected] ) We are grateful to the Center on Reinventing Public Education, University of Washington, Bothell, for funding this research. - PowerPoint PPT PresentationTRANSCRIPT
THE EFFECT OF CHARTER SCHOOLS ON STUDENT ACHIEVEMENT: A META-ANALYSIS
OF THE LITERATURE
CAMPBELL COLLOQUIUM EDUCATION PANEL, MAY 2012
Julian R. Betts and Y. Emily Tang,University of California, San Diego
([email protected], [email protected])We are grateful to the Center on Reinventing Public
Education, University of Washington, Bothell, for funding this research
2
Introduction and Motivation Selecting Studies to Include Assessment of Alternative Methods of
Evaluating the Impact of Charter Schools
Challenges in Study Collection/Review Process
Description of Methods Used in Review Results Future Research and Policy Implications
OUTLINE
3
SOME BACKGROUND ON US EDUCATION
Persistent concern over the performance of US public schools at the elementary and secondary levels
Elementary Grades K-5 (ages 5-11)
Secondary Middle: Grades 6-8 (ages 11-14) High: Grades 9-12 (ages 14-18)
4
THE US SPENDS A LOT (PER PRIMARY SCHOOL PUPIL) ON EDUCATION, OBTAINS AVERAGE EDUCATIONAL OUTCOMES
Source: Gruber (2010)
5
THE US SPENDS ABOUT AVERAGE (% OF GDP) ON EDUCATION, OBTAINS AVERAGE EDUCATIONAL OUTCOMES
Source: OECD (2011)
6
IN THE US THE SCHOOL THAT A STUDENT ATTENDS IS PRIMARILY DETERMINED BY WHERE HE/SHE LIVES
San Diego Unified School District Elementary School Boundaries 2011-12
7
WHAT IS A CHARTER SCHOOL? Charter schools are a relatively new
alternative to traditional neighborhood public schools ~ 20 years, substantial growth in the 2000s
A succession of U.S. presidents has named charter schools as important agents of school reform
8
APPROXIMATELY 5% OF PUBLIC SCHOOLS ARE CHARTER SCHOOLS, THIS NUMBER IS GROWING
Source: Lake and Gross (2011)
9
WHAT IS A CHARTER SCHOOL? Charter schools are publicly funded,
governed by organization under contract with the state
Charter schools are exempted from parts of the state education code, freeing them to innovate with respect to curriculum, pedagogy and hiring of teachers
10
CHARTER SCHOOLS ARE DIFFERENT FROM EACH OTHER, EXAMPLES FROM SAN DIEGO
Albert Einstein Academy: “independent charter school that would have
a dual instructional focus of German-English immersion within the context of a rigorous academic instructional model”
Charter School of San Diego: initially developed from a state bill
“designed to reduce the dropout rate by recovering students who had been out of school for more than 45 days”
11
SELECTING STUDIES FOR THIS LITERATURE REVIEW
Scope: Include studies of US elementary and secondary charter school performance US public K-12 education is decentralized Most data on student performance are collected at
the level of a US state, or the level of a school district (smaller than a US state)
Outcomes: Include studies that use student performance on math and reading standardized tests as an outcome measure
Methods: Include studies that use credible approaches to address selection bias
12
SELECTION BIAS: MAIN CONCERNS WITH ALTERNATIVE APPROACHES LEADING TO EXCLUSION
Snapshots of average student achievement at one point in time can be misleading as they do not account for self-selection into schools US school attendance based largely
on geographic residence. Students choosing to attend charter
schools are likely different in observable and unobservable ways
13
UNOBSERVED CHARACTERISTICS CORRELATED WITH CHARTER SCHOOL ATTENDANCE
Negative selection (downward bias) Example: An underprivileged, disadvantaged
student without family support is at high risk of dropping out of school. She is advised by her high school counseling staff to transfer to a charter school, and she chooses to transfer.
Problem: Underprivileged, disadvantaged students without family support are not likely to obtain high test scores in any school, traditional or charter.
The estimate of charter school effectiveness based on comparison of charter school student performance and traditional school student performance would be biased downwards.
14
UNOBSERVED CHARACTERISTICS CORRELATED WITH CHARTER SCHOOL ATTENDANCE
Positive selection (upward bias) Example: An active, concerned, involved parent
is dissatisfied with the traditional public school in his/her neighborhood. The parent decides to opt-out of the traditional school and enroll his/her child in a charter school.
Problem: Students with active, concerned, involved parents are likely to obtain high test scores in any school, traditional or charter.
Implication: The estimate of charter school effectiveness based on comparison of charter school student performance and traditional school student performance would be upwardly biased.
15
SELECTING STUDIES FOR THIS LITERATURE REVIEW
National Charter School Research Project issued a White Paper (drafters: Betts and Hill, 2006) arguing that lottery-based studies and student-level longitudinal “value-added” studies were the two most credible approaches
These methods more convincing than other methods.
16
METHODS MATTER
Source: Hill (2006)
17
4 COMMONLY USED METHODS OF ANALYSIS IN THE INCLUDED STUDIES
In the set of studies we include, there are four approaches used
1) Lottery-based studies 2) Fixed-effect studies, that compare a
student’s gains in achievement in years attended a charter to his or her gains in years attended a traditional public school
3) Propensity score matching 4) Other types of matching (e.g. CREDO)
18
LOTTERY-BASED ANALYSIS
Source: Waiting for Superman movie (2010)
19
LOTTERY-BASED ANALYSIS Obvious benefit: expected outcomes
identical for lottery winners and losers if lottery conducted fairly
But several weaknesses to this “gold standard”
External validity Most charter schools not oversubscribed
Mathematica study of charter middle schools: only 130/492 oversubscribed
Could be bias from attrition
20
PROPENSITY SCORE MATCHING Assumes “selection on observables” If students in charter schools have unobserved
variations in ability or motivation, will be biased
Two major studies of KIPP (Knowledge is Power Program) schools have used this approach
CREDO at Stanford has produced string of influential state-level studies. Uses a unique matching process. Not propensity score but has similar issue with “selection on observables”
21
STUDENT FIXED-EFFECTS Benefit: Avoids need to compare one student with another,
instead comparing individual students’ trajectories in charter schools and traditional public schools
But many elementary students never switch between the two types of schools – external validity issue Zimmer et al (2009) compare test-score gains of charter “stayers”
and switchers and do not get clear-cut result. But in some cases “stayers” have higher test-score gains
Suggests downward bias from using this method Zimmer et al (2009) also raise concerns about reversibility –
are the effects of attending a charter dependent on the order in which a student attends the charter and the traditional public school? Find some evidence that this is the case.
Unobserved heterogeneity may change over time. Fixed effects cannot solve
22
INCLUDED STUDIES 40 reports now available, with just under
100 estimates of effects for each of math and English Language Arts (reading)
Lottery-based studies still quite rare: still only 8 papers that use lotteries, covering 90 charter schools
We exclude studies using less rigorous methods, specifically, those that do not use student-level test score gains as outcomes.
23
CHALLENGES IN STUDY COLLECTION/REVIEW PROCESS
Handling large weight (large number of students and large number of schools) studies Solution: Analyze with and without large weight studies
Handling the different methods used in different studies Solution: Investigate whether method of analysis matters
Some reports omit important information, e.g. number of schools in the sample Solution: Email exchange with authors
24
Introduction and Motivation Assessment of Alternative Methods of
Evaluating the Impact of Charter Schools
Selecting Studies to Include Challenges in Study Collection/Review
Process Description of Methods Used in
Review Results Future Research and Policy Implications
25
OUR METHODS OF ANALYSIS Fisher test – Is there evidence that no study finds
negative effects; conversely, evidence of no positive effects?
Formal meta-analysis provides overall estimated effect, its statistical significance and measures of how much true underlying variation there is across studies
Histograms Show variability and the influence of weighting of studies
Vote-counting as a way of assessing variation in results
26
HETEROGENEITY IS AN UNDERLYING THEME
Look for variations in effect by: Subject area tested (math vs. reading) Grade span (E, M, H) Geographic location KIPP vs. non-KIPP Is there a systematic difference in results
based on the method researchers use?
27
METHODS USED IN REVIEW Testing Whether Charter Schools in
Any Study Increase or Decrease Achievement Relative to Traditional Public Schools
Meta-Analysis of Effect Size Histograms and Vote Counting as
Measures of Variation
28
METHOD #1: EVIDENCE OF NO POSITIVE EFFECTS, OR NO NEGATIVE EFFECTS?
Fisher’s combined test
S is distributed with df=2k Null hypothesis: No positive effects Null hypothesis: No negative effects€
S = −2 ln(pi)i=1
k
∑
€
χ2
29
METHOD #1: EVIDENCE OF NO POSITIVE EFFECTS, OR NO NEGATIVE EFFECTS?
We conduct this analysis 12 times: 6 ways of combining grades, and two subjects (math and ELA)
First sign of heterogeneous effects of charter schools: in 9/12 cases there is clear evidence of BOTH negative and positive effects
Three exceptions with evidence of positive effects but no evidence of negative effects: elementary and middle school ELA scores, and middle school math scores
PROBABILITY OF NO POSITIVE EFFECTS IN ANY OF THE STUDIES: ALMOST ZERO
Grade-Span Reading Tests Math TestsElementary <0.001 <0.001
Middle <0.001 <0.001High <0.001 0.001
El’y, Middle, and Combined El’y/Middle
<0.001 <0.001
All <0.001 <0.001
Studies of All Grades or Largest Grade Span(s) If An
All-Grade Study Not Available
<0.001 <0.001
30
PROBABILITY OF NO NEGATIVE EFFECTS IN ANY OF THE STUDIES: ALMOST ZERO IN MOST CASES, AND QUITE HIGH IN 3 CASES
Grade-Span Reading Tests Math TestsElementary 0.987 <0.001
Middle 0.994 0.978High <0.001 0.001
El’y, Middle, and Combined El’y/Middle
<0.001 <0.001
All <0.001 <0.001
Studies of All Grades or Largest Grade Span(s) If An
All-Grade Study Not Available
<0.001 <0.001
31
32
METHODS USED IN REVIEW Testing Whether Charter Schools in Any
Study Increase or Decrease Achievement Relative to Traditional Public Schools
Meta-Analysis of Effect Size Histograms and Vote Counting as
Measures of Variation
33
METHOD #2: FORMAL META-ANALYSIS Assume charter school estimates are
randomly distributed Therefore it is important to estimate both the
mean and the variation Underlying “true” variation across studies is
the extent to which variation cannot be explaining by sampling error (“uncertainty”) in individual estimates
Omitted many studies of individual KIPP schools as they would have disproportionate influence Include KIPP schools in subsidiary analysis
34
THE MEAN EFFECT IS A WEIGHTED AVERAGE
In a random effects meta-analysis, we take a weighted average of the effect sizes across studies. If Yi is the effect size for the ith of k studies, and Wi is the weight for each study, our overall estimated effect size M is :
(1)
1
1
k
i iik
ii
WYM
W
35
WEIGHTS DEPEND ON WITHIN-STUDY VARIANCE AND ESTIMATED ACROSS-STUDY (TRUE) VARIANCE The weight for each study is the inverse of the sum of the
within-study variance (based on the standard error) and an estimate of the true between-study variance, T2:
(2)
T2 based on a method of moments estimate of the variance of true effect sizes.
Note that as T2 becomes large relative to the average within-study variance estimate, then we will tend toward equal weighting across studies; whereas as T2 becomes relatively small, the weights can become highly unequal with heavier weight given to studies with the lowest sampling variance.
2
1
i
iY
WV T
36
AN ESTIMATE OF WHAT % OF THE VARIANCE ACROSS STUDIES IS TRUE
Use the I2 statistic (Higgins et al., 2003) Provides estimate of the percentage of
variation across studies that reflects true underlying variation
37
SAMPLE OF OUR RESULTS ON EFFECT SIZES
* Indicates statistically significant (5% level)
Grade Span Reading Tests Math Tests
E (Elementary) 0.022* (9-7), 77.7%
0.049* (10-8), 94.7%
38
SAMPLE OF OUR RESULTS
* Indicates statistically significant (5% level)
Grade Span Reading Tests Math Tests
E (Elementary) 0.022* (9-7), 77.7%
0.049* (10-8), 94.7%
“On average, attending a charter school is associated with an increase in test scores in reading equal to 0.022 of a standard deviation per year.”
39
SAMPLE OF OUR RESULTS
* Indicates statistically significant (5% level)
Grade Span Reading Tests Math Tests
E (Elementary) 0.022* (9-7), 77.7%
0.049* (10-8), 94.7%
Nine studies covering 7 geographic areas77.7% of the variation across studies
represents true variation in charter school effects, rather than “noise”
40
OVERALL EFFECT SIZE ESTIMATESGrade Span Reading Tests Math Tests
E (Elementary) 0.022* (9-7), 77.7%
0.049* (10-8), 94.7%
M (Middle) 0.011 (9-7), 85.7%
0.055* (10-8), 92.0%
H (High) 0.054 (7-5), 98.3%
-0.015 (8-6), 98.6%
Combined E/M -0.009 (15-12), 93.4%
-0.012 (15-12), 97.9%
E, M, and Combined E/M
0.002 (31-17), 90.3%
0.020* (33-18), 96.8%
All 0.008 (17-14), 98.4%
0.014 (18-15), 97.7%
41
ELEMENTARY/MIDDLE SCHOOL MATH EFFECTS: MEANINGFUL BUT NOT HUGE
Enough to move a student at the 50th percentile to the 52nd percentile after attending a charter for one year
Elementary school reading impact is smaller: enough to boost a student from 50th to about percentile 50.8
42
ELEMENTARY SCHOOL READING EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 77.7%, p = 0.000)
San Diego
Chicago
San Diego
National
California
NYC
Delaware
Study
NYC
Boston
ID
0.02 (0.01, 0.04)
0.04 (-0.01, 0.09)
0.10 (0.03, 0.18)
-0.08 (-0.17, 0.01)
0.01 (0.01, 0.01)
-0.00 (-0.01, 0.00)
0.19 (0.02, 0.35)
0.03 (0.00, 0.07)
0.04 (0.01, 0.07)
0.06 (0.01, 0.10)
ES (95% CI)
100.00
6.80
3.70
2.61
27.01
25.00
0.88
12.45
%
12.83
8.73
Weight
0.02 (0.01, 0.04)
0.04 (-0.01, 0.09)
0.10 (0.03, 0.18)
-0.08 (-0.17, 0.01)
0.01 (0.01, 0.01)
-0.00 (-0.01, 0.00)
0.19 (0.02, 0.35)
0.03 (0.00, 0.07)
0.04 (0.01, 0.07)
0.06 (0.01, 0.10)
ES (95% CI)
100.00
6.80
3.70
2.61
27.01
25.00
0.88
12.45
%
12.83
8.73
Weight
0-.3 -.2 -.1 .1 .2 .3
43
ELEMENTARY SCHOOL MATH EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 94.7%, p = 0.000)
San Diego
San Diego
Chicago
Idaho
ID
Boston
NYC
NYC
National
Delaware
California
Study
0.05 (0.02, 0.08)
0.29 (0.22, 0.37)
-0.19 (-0.30, -0.08)
0.12 (0.04, 0.19)
0.33 (0.03, 0.63)
ES (95% CI)
0.02 (-0.03, 0.07)
0.19 (0.02, 0.36)
0.09 (0.06, 0.12)
-0.00 (-0.00, 0.00)
0.04 (0.01, 0.07)
-0.03 (-0.04, -0.02)
100.00
8.88
5.89
8.51
1.10
Weight
11.44
2.89
14.29
16.50
14.25
16.26
%
0.05 (0.02, 0.08)
0.29 (0.22, 0.37)
-0.19 (-0.30, -0.08)
0.12 (0.04, 0.19)
0.33 (0.03, 0.63)
ES (95% CI)
0.02 (-0.03, 0.07)
0.19 (0.02, 0.36)
0.09 (0.06, 0.12)
-0.00 (-0.00, 0.00)
0.04 (0.01, 0.07)
-0.03 (-0.04, -0.02)
100.00
8.88
5.89
8.51
1.10
Weight
11.44
2.89
14.29
16.50
14.25
16.26
%
0-.3 -.2 -.1 .1 .2 .3 .4
44
MIDDLE SCHOOL READING EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 85.7%, p = 0.000)
Chicago
NYC
ID
San Diego
Delaware
Texas
National
Boston
National
San Diego
Study
0.01 (-0.02, 0.04)
-0.06 (-0.14, 0.01)
0.04 (-0.02, 0.10)
ES (95% CI)
-0.08 (-0.12, -0.04)
0.08 (0.04, 0.12)
0.01 (-0.01, 0.04)
-0.10 (-0.23, 0.03)
0.17 (0.07, 0.27)
0.02 (0.02, 0.02)
0.01 (-0.04, 0.06)
100.00
8.49
10.00
Weight
13.69
13.58
16.06
4.21
5.65
17.37
10.94
%
0.01 (-0.02, 0.04)
-0.06 (-0.14, 0.01)
0.04 (-0.02, 0.10)
ES (95% CI)
-0.08 (-0.12, -0.04)
0.08 (0.04, 0.12)
0.01 (-0.01, 0.04)
-0.10 (-0.23, 0.03)
0.17 (0.07, 0.27)
0.02 (0.02, 0.02)
0.01 (-0.04, 0.06)
100.00
8.49
10.00
Weight
13.69
13.58
16.06
4.21
5.65
17.37
10.94
%
0-.3 -.2 -.1 .1 .2 .3
45
MIDDLE SCHOOL MATH EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 92.0%, p = 0.000)
ID
Boston
Texas
San Diego
Chicago
National
National
Idaho
Delaware
Study
NYC
San Diego
0.05 (0.01, 0.10)
ES (95% CI)
0.54 (0.39, 0.69)
-0.00 (-0.02, 0.02)
0.01 (-0.09, 0.11)
-0.09 (-0.16, -0.02)
-0.08 (-0.20, 0.04)
0.02 (0.02, 0.02)
-0.05 (-0.18, 0.08)
0.09 (0.05, 0.13)
0.24 (0.16, 0.31)
0.06 (0.03, 0.10)
100.00
Weight
4.81
14.17
7.90
10.32
6.31
14.66
5.88
13.10
%
9.70
13.15
0.05 (0.01, 0.10)
ES (95% CI)
0.54 (0.39, 0.69)
-0.00 (-0.02, 0.02)
0.01 (-0.09, 0.11)
-0.09 (-0.16, -0.02)
-0.08 (-0.20, 0.04)
0.02 (0.02, 0.02)
-0.05 (-0.18, 0.08)
0.09 (0.05, 0.13)
0.24 (0.16, 0.31)
0.06 (0.03, 0.10)
100.00
Weight
4.81
14.17
7.90
10.32
6.31
14.66
5.88
13.10
%
9.70
13.15
0-.3 -.2 -.1 .1 .2 .3
46
HIGH SCHOOL READING EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 98.3%, p = 0.000)
National
San Diego
San Diego
Texas
San Diego
Study
ID
Boston
Delaware
0.05 (-0.03, 0.14)
-0.02 (-0.02, -0.02)
0.15 (0.10, 0.20)
0.04 (-0.24, 0.33)
-0.16 (-0.18, -0.14)
0.03 (-0.01, 0.07)
ES (95% CI)
0.16 (0.02, 0.31)
0.21 (0.16, 0.26)
100.00
16.98
15.97
6.09
16.84
16.33
%
Weight
11.60
16.18
0.05 (-0.03, 0.14)
-0.02 (-0.02, -0.02)
0.15 (0.10, 0.20)
0.04 (-0.24, 0.33)
-0.16 (-0.18, -0.14)
0.03 (-0.01, 0.07)
ES (95% CI)
0.16 (0.02, 0.31)
0.21 (0.16, 0.26)
100.00
16.98
15.97
6.09
16.84
16.33
%
Weight
11.60
16.18
0-.3 -.2 -.1 .1 .2 .3
47
HIGH SCHOOL MATH EFFECT SIZES
NOTE: Weights are from random effects analysis
Overall (I-squared = 98.3%, p = 0.000)
National
San Diego
San Diego
Texas
San Diego
Study
ID
Boston
Delaware
0.05 (-0.03, 0.14)
-0.02 (-0.02, -0.02)
0.15 (0.10, 0.20)
0.04 (-0.24, 0.33)
-0.16 (-0.18, -0.14)
0.03 (-0.01, 0.07)
ES (95% CI)
0.16 (0.02, 0.31)
0.21 (0.16, 0.26)
100.00
16.98
15.97
6.09
16.84
16.33
%
Weight
11.60
16.18
0.05 (-0.03, 0.14)
-0.02 (-0.02, -0.02)
0.15 (0.10, 0.20)
0.04 (-0.24, 0.33)
-0.16 (-0.18, -0.14)
0.03 (-0.01, 0.07)
ES (95% CI)
0.16 (0.02, 0.31)
0.21 (0.16, 0.26)
100.00
16.98
15.97
6.09
16.84
16.33
%
Weight
11.60
16.18
0-.3 -.2 -.1 .1 .2 .3
48
READING EFFECT SIZES FOR STUDIES THAT COMBINE ELEMENTARY AND MIDDLE SCHOOLS
NOTE: Weights are from random effects analysis
Overall (I-squared = 93.4%, p = 0.000)
ID
Texas
DC
NYC
Ohio
Chicago
Chicago
Texas
North Carolina
Massachusetts
Georgia
Missouri
Minnesota
Ohio
Arkansas
Arizona
Study
-0.01 (-0.02, 0.00)
ES (95% CI)
-0.08 (-0.10, -0.06)
-0.01 (-0.02, 0.01)
0.09 (0.02, 0.16)
-0.08 (-0.12, -0.04)
0.00 (-0.01, 0.01)
-0.04 (-0.06, -0.02)
0.09 (0.06, 0.12)
-0.09 (-0.12, -0.07)
0.00 (-0.01, 0.02)
0.01 (-0.00, 0.01)
0.03 (0.01, 0.05)
-0.02 (-0.03, -0.01)
-0.00 (-0.01, 0.00)
0.02 (0.00, 0.04)
-0.01 (-0.02, -0.01)
100.00
Weight
6.85
7.10
2.57
4.74
7.64
6.85
5.61
6.00
7.50
7.86
7.05
7.46
7.80
7.00
7.96
%
-0.01 (-0.02, 0.00)
ES (95% CI)
-0.08 (-0.10, -0.06)
-0.01 (-0.02, 0.01)
0.09 (0.02, 0.16)
-0.08 (-0.12, -0.04)
0.00 (-0.01, 0.01)
-0.04 (-0.06, -0.02)
0.09 (0.06, 0.12)
-0.09 (-0.12, -0.07)
0.00 (-0.01, 0.02)
0.01 (-0.00, 0.01)
0.03 (0.01, 0.05)
-0.02 (-0.03, -0.01)
-0.00 (-0.01, 0.00)
0.02 (0.00, 0.04)
-0.01 (-0.02, -0.01)
100.00
Weight
6.85
7.10
2.57
4.74
7.64
6.85
5.61
6.00
7.50
7.86
7.05
7.46
7.80
7.00
7.96
%
0-.3 -.2 -.1 .1 .2 .3
49
MATH EFFECT SIZES FOR STUDIES THAT COMBINE ELEMENTARY AND MIDDLE SCHOOLS
NOTE: Weights are from random effects analysis
Overall (I-squared = 97.9%, p = 0.000)
Texas
Missouri
ID
Ohio
North Carolina
NYC
Minnesota
Chicago
Georgia
Study
Arkansas
DC
Massachusetts
Chicago
Texas
Arizona
Ohio
-0.01 (-0.03, 0.01)
-0.12 (-0.16, -0.08)
0.03 (0.01, 0.04)
ES (95% CI)
-0.06 (-0.07, -0.05)
-0.16 (-0.20, -0.12)
0.12 (0.03, 0.21)
-0.03 (-0.04, -0.02)
0.02 (-0.02, 0.06)
-0.01 (-0.02, -0.00)
0.05 (0.03, 0.07)
0.01 (-0.00, 0.03)
0.06 (0.05, 0.07)
0.02 (0.01, 0.03)
0.08 (0.06, 0.11)
-0.04 (-0.05, -0.04)
-0.18 (-0.26, -0.10)
100.00
6.23
7.24
Weight
7.56
6.11
3.43
7.44
6.23
7.59
%
7.20
7.36
7.48
7.53
7.00
7.60
4.01
-0.01 (-0.03, 0.01)
-0.12 (-0.16, -0.08)
0.03 (0.01, 0.04)
ES (95% CI)
-0.06 (-0.07, -0.05)
-0.16 (-0.20, -0.12)
0.12 (0.03, 0.21)
-0.03 (-0.04, -0.02)
0.02 (-0.02, 0.06)
-0.01 (-0.02, -0.00)
0.05 (0.03, 0.07)
0.01 (-0.00, 0.03)
0.06 (0.05, 0.07)
0.02 (0.01, 0.03)
0.08 (0.06, 0.11)
-0.04 (-0.05, -0.04)
-0.18 (-0.26, -0.10)
100.00
6.23
7.24
Weight
7.56
6.11
3.43
7.44
6.23
7.59
%
7.20
7.36
7.48
7.53
7.00
7.60
4.01
0-.3 -.2 -.1 .1 .2 .3
50
METHODS USED IN REVIEW Testing Whether Charter Schools in Any
Study Increase or Decrease Achievement Relative to Traditional Public Schools
Meta-Analysis of Effect Size Histograms and Vote Counting as
Measures of Variation
51
METHOD #3: HISTOGRAMS Another way of displaying the variation
across studies Tried weighting each study equally and
weighting studies by number of observations Latter approach gives heavy weight to CREDO
studies Our formal meta-analysis is closer to
weighting studies equally than weighting by observation
52
53
54
METHOD #4: VOTE COUNTING Categorize studies by sign of effect and whether
statistically significant Method is problematic because it ignores fact that
many statistically insignificant results all in the same direction may, taken together, suggest a statistically significant result
We use mostly to highlight the heterogeneity Typically find that for most grade spans >50% of
studies show positive effects, but this weakens and sometimes reverses if we weight studies by number of observations
55
RESULTS VARY BY METHOD Lottery results yielded the most positive
results, followed closely by propensity score matching.
These were followed by fixed effects and other matching methods (which are fairly similar with mixed positive and negative results)
But it may not be the method that matters quite so much as the specific schools studied Example: Propensity score results are large but
focus on KIPP schools
56
RESULTS VARY BY METHOD
57
RESULTS VARY BY METHOD
58
REPLICATION OF RESULTS USING DIFFERENT METHODS
There are 3 studies/pairs of studies that replicate lottery results using more traditional “value-added” methods
They generally suggest that value-added models can get close to the lottery results (but in a few cases estimates slightly to meaningfully lower): Boston (Abulkadiroglu et al.) New York (Hoxby et al. and CREDO) San Diego Preuss School (McLure et al., Betts, Tang
and Zau)
59
REPLICATION OF RESULTS USING DIFFERENT METHODS
There are 3 studies/pairs of studies that replicate lottery results using more traditional “value-added” methods
They generally suggest that value-added models can get close to the lottery results (but in a few cases estimates slightly to meaningfully lower): Boston (Abulkadiroglu et al.) New York (Hoxby et al. and CREDO) San Diego Preuss School (McLure et al., Betts, Tang
and Zau)
60
Introduction and Motivation Selecting Studies to Include Assessment of Alternative Methods of
Evaluating the Impact of Charter Schools
Challenges in Study Collection/Review Process
Description of Methods Used in Review Results Future Research and Policy
Implications
61
IMPLICATIONS FOR RESEARCH
Evaluate individual schools Charters are meant to innovate; unlikely
that all charters will have the same impact
Charters should obtain permission from applicants to gather student records
States and chartering authorities should regularly receive lottery data
62
IMPLICATIONS FOR RESEARCH
Focus on successful schools to identify characteristics that may be working E.g. longer day/time, student population targeted,
discipline policies, teacher management
Obtain more details about charter school heterogeneity and study them
Obtain more details about charter school closures
63
IMPLICATIONS FOR RESEARCH
Probably important to examine more than results on math and ELA achievement.
A handful of studies point to positive charter effects on graduation, college attendance and behavior. Expand focus to include outcomes other
than math/reading test scores
64
WHAT WORKS CLEARINGHOUSE (WWC) FOR CHARTER SCHOOL RESULTS
In the long run it would be good to have a non-partisan group that collected and interpreted school-level charter results.
65
IMPLICATIONS FOR POLICY
Status as a charter school vs. traditional public school unlikely to be (on its own) meaningful Promoting charter schools for sake of charter
schools probably not productive path to comprehensive reform
Continue expansion (no particular reason not to) Still only ~5% of traditional public school sector Renew focus on traditional public school reform
Exploit flexibility of charter schools by using them as laboratories to learn what works
66
THANK YOU! Published version available at: http://www.crpe.org/cs/crpe/view/csr_pubs/467 Executive summary at: http://www.crpe.org/cs/crpe/view/csr_pubs/468
67
SUPPLEMENTARY SLIDES
68
ADDING KIPP STUDIES BACK IN HAS BIG EFFECT
Grade Span Reading Tests Math Tests
Including KIPP Schools
M 0.070* (38-33), 88.3%
0.180* (39-34), 96.8%
E, M, and Combined E/M
0.034* (60-43), 90.8%
0.105* (62-44), 98.6%
Results Including Only KIPP Estimates
M 0.096* (29-unknown),
82.7%
0.223* (29-unknown),
93.7%
69
SENSITIVITY TO EXCLUSION OF CREDO STUDIES
CREDO (Stanford) has produced impressive string of mostly state-wide longitudinal student studies. Match each charter student to an average of several similar demographics and test scores
Many charter students are matched based on their test scores AFTER they enter charter schools potential bias
Hoxby (2009) has concerns about measurement error that may bias charter coefficient down CREDO offers partial rebuttal
70
RESULTS SOMEWHAT STRONGER IF OMIT CREDO STUDIES
Grade Span Reading Tests Math Tests
E 0.034* (8-6), 79.5%
0.072* (9-7), 95.2%
M 0.010 (8-7), 87.2%
0.068* (9-8), 92.8%
H 0.072 (6-4), 98.5%
-0.002 (7-5), 97.5%
Combined E/M -0.023 (6-5), 95.5%
-0.041 (6-5), 96.9%
E, M, and Combined E/M
0.008 (22-10), 92.0%
0.038* (24-11), 95.0%
All 0.016 (10-9), 86.6%
0.041* (11-10), 67.7%
71
EFFECTS FOR URBAN DISTRICTS AND SCHOOLS LARGER THAN FOR ALL DISTRICTS
Grade Span Reading Tests Math Tests
E 0.046* (6-4), 61.8%
0.085 (6-4), 92.2%
M 0.009 (5-4), 87.0%
0.139 (5-4), 94.8%
H 0.101* (4-2), 78.2%
0.019 (4-2), 42.7%
Combined E/M -0.003 (4-3), 86.2%
0.021* (4-3), 47.7%
E, M, and Combined E/M
0.016 (15-5), 84.1%
0.077* (15-5), 92.4%
All 0.008 (8-6), 63.2%
0.045* (8-6), 74.8%
72
VARIATIONS BY RACE/ETHNICITY Surprisingly few studies test for variation
by race/ethnicity. CREDO studies an important exception
Patterns not uniform, but overall, charter effects decline in the following order: black > Hispanic > Native American > white Results for whites typically negative, not always
significant. Biggest exception is high school reading, with a positive and significant effect
73
VARIATIONS BY ENGLISH LEARNER, SPECIAL EDUCATION, MEAL ASSISTANCE
Effects often insignificant, perhaps due to smaller sample sizes
But some evidence of positive effects of charter schools on EL and special education students in both math and reading from studies of all grades and studies of middle schools