Tim R. Sass

Department of Economics

Florida State University

Tallahassee, FL 32306-2180

tsass@fsu.edu.

CHARTER SCHOOLS AND

STUDENT ACHIEVEMENT IN

FLORIDA

Astratto
I utilize longitudinal data covering all public school stu-
dents in Florida to study the performance of charter
schools and their competitive impact on traditional pub-
lic schools. Controlling for student-level fixed effects, IO
find achievement initially is lower in charters. How-
ever, by their fifth year of operation new charter schools
reach a par with the average traditional public school in
math and produce higher reading achievement scores
than their traditional public school counterparts. Among
charters, those targeting at-risk and special education
students demonstrate lower student achievement, while
charter schools managed by for-profit entities peform
no differently on average than charters run by nonprof-
its. Controlling for preexisting traditional public school
quality, competition from charter schools is associated
with modest increases in math scores and unchanged
reading scores in nearby traditional public schools.

C(cid:1) 2006 American Education Finance Association

91

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

F

.

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

1. INTRODUCTION
The efficacy of publicly funded school-choice programs has been fiercely de-
bated in recent years. Proponents claim that school choice will not only provide
a mechanism for students seeking to improve the quality of their own edu-
cation but also engender competition that will lead to improvements in the
quality of education for students who remain in traditional public schools.
Those opposed to publicly funded school choice argue such programs will
skim the best students, drain resources away from public schools, and pro-
mote racial/ethnic segregation.

While much attention has been paid to voucher programs in Milwaukee,
Cleveland, and elsewhere, by far the more common vehicle for school choice is
charter schools. Forty states and the District of Columbia have charter school
laws in place, and approximately 698,000 students attended charters during
the 2003–4 school year.1 Like vouchers, charter schools represent a subsidized
alternative to traditional public schools. Students are not charged any tuition,
and charter schools rely on public funding for their operating budget. While the
specifics vary across states, charter schools typically are not subject to many
of the regulatory constraints governing traditional public schools; charters
generally have considerable freedom in personnel and curriculum decisions.
In this article I utilize a new longitudinal database from Florida to address
three key issues relating to charter schools and student achievement. Primo,
how does the impact of charter schools on student achievement compare with
traditional public schools? Secondo, to the extent that student performance
varies among charter schools, what factors contribute to the difference in
performance? Third, what competitive impact, if any, do charter schools have
on traditional public schools?

To empirically analyze these issues I will focus on student achievement
in traditional public schools and charters in Florida. Due to the size of its
charter sector and the availability of data, Florida is an ideal laboratory for
empirical analysis of charter schools. Charter schools have existed in Florida
since the 1996–97 school year, Quando 5 schools began operation (Vedi la tabella 1).
The number of charter schools has rapidly grown to 258 in the 2003–4 school
year. Florida now ranks third among the states in number of charters oper-
ating, and the 53,000 students enrolled in Florida charter schools in 2003–4
comprise 7.6 percent of the national total. Florida also possesses one of the
most comprehensive systems in the nation for tracking student achievement;
all public school students (whether in traditional or charter schools) must take
annual standardized tests in each of grades 3–10.

1. National data on charter school laws and enrollment are from the Center for Education Reform

(www.edreform.com).

92

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tavolo 1 Charter Schools in Operation by Type, Changes in Operation, and Charter School Student Membership in Florida, 1996/1997–2002/2003

CHARTER SCHOOLS OPERATING

CHANGES IN OPERATION

MEMBERSHIP

School Year

Total

Targeted

Conversion

Run by For-Profit

Openings

Closures

Number of Students

% of all Public School Students

1996–97

1997–98

1998–99

1999–00

2000–1

2001–2

2002–3

5

31

78

113

148

190

232

2

10

26

40

50

58

65

0

1

3

3

4

7

11

0

1

4

9

15

32

48

5

26

49

38

37

53

43

0

2

3

2

11

1

8

400

3,500

10,000

17,200

27,200

39,900

50,700

0.02%

0.15%

0.43%

0.72%

1.12%

1.60%

2.06%

Note: The number of charters operating represents the number of charter schools
in operation during any portion of the school year, including schools that closed
during the school year. The count of charter schools is based on the assignment
of school identification codes by the Florida Department of Education. Così, a sin-
gle charter school may have two branches in distinct geographic locations, or two

charters serving distinct populations may physically reside in the same location.
Charter Schools Opened includes all schools operating that did not exist at the end
of the previous school year. Closures include all schools that ceased operating prior
to the start of the next school year. Student totals are based on attendance during
the October membership survey of public schools.

9
3

io

T
M

R

.

S
UN
S
S

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

I begin by reviewing the extant empirical literature on charter schools in the
next section. In the third section, I discuss methodological issues and present
an empirical model of student achievement. This is followed by a description
of the Florida data and the presentation of the empirical results. A final section
summarizes the findings and their implications for policy.

2. PREVIOUS LITERATURE
Achievement in Charter Schools

Despite the size and importance of the charter school movement, quantita-
tive analysis of the impact of charter schools on student achievement has
been limited. Much of the existing research lacks sufficient controls for stu-
dent characteristics which creates potential selection-bias problems due to the
nonrandom assignment of students between charters and traditional public
schools. Tuttavia, a handful of recent papers account for the impact of student
characteristics on achievement in charter schools by employing longitudinal
data and estimating student-level, fixed-effects models.

Solmon, Paark, and Garcia (2001) analyze scores on the Stanford Achieve-
ment Test (SAT-9) for a panel of Arizona students in grades 3–11 over the 1998-
2000 period.2 Their three-year panel encompasses 40,000 total students, In-
cluding 8,000 students who attended an Arizona charter school for at least
one year. Their models incorporate fixed effects to control for time-invariant
student characteristics but do not include lagged test scores to account for the
cumulative effects of past educational inputs.3 They find the first-year effect
of attending a charter school on achievement is statistically insignificant for
both reading and math. Tuttavia, students who attend a charter school for two
or three years experience achievement gains in both reading and math that
exceed those of traditional public school students. Unfortunately, no measure
of the age of charter schools is included in their analysis. Thus the measured
student tenure effects may in part reflect differences in the maturity of charter
schools, rather than the duration of charter school attendance.

Hanushek, Kain, and Rivkin (2002) analyze individual student achieve-
ment gains for four cohorts of Texas students in grades 4–7 during the

2.

3.

Solmon and Goldschmidt (2004) reanalyze the same data, limiting their analysis to reading scores.
Employing a three-level hierarchial linear model, they find that students attending charter schools
three straight years experience higher achievement growth than students enrolled only in traditional
public schools. Tuttavia, students who start in a charter school and switch to a traditional public
school have significantly higher achievement growth than those who stay in charter schools. Further,
the positive effects of charter attendance diminish with the grade level; achievement growth for
charter-only students exceeds that of traditional-school-only students in elementary school, is equal
by middle school, and is lower in high school.
The inclusion of lagged test scores to control for past educational inputs is discussed in more detail
in the ethodology section below.

94

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

years 1996–2001. Their sample includes over 6,600 students who attended
a charter school during the period and more than 800,000 students in
total. Academic achievement is measured by year-to-year changes in standard-
ized individual scores on the Texas Assessment of Academic Skills (TAAS), UN
criterion-referenced test. In addition to student-level fixed effects, their model
includes controls for both charter school age and student mobility.

Hanushek, Kain, and Rivkin (2002) find that student achievement gains
in both math and reading are lower in first-year charters than in the average
traditional public school. These negative effects diminish rapidly, Tuttavia,
as the charters mature. For students in charters that have existed three years
or more there are no statistically significant differences in reading or math
achievement gains relative to peers attending traditional public schools. These
average effects mask the wide variation in quality among both charters and
traditional public schools. Hanushek, Kain, and Rivkin split their sample into
geographic regions and include school-level fixed effects to measure differ-
ences in school quality. They find that higher-quality charter schools are often
as good or better than traditional public schools, but the bottom quartile of
charters are of much lower quality than the lowest quartile of traditional public
schools in nearly all regions of Texas.

Booker et al. (2004UN) also analyze achievement-score gains in Texas,
though with a larger data set of six cohorts that spans 1995–2002 and cov-
ers 10,000 charter students and 1.4 million students in total.4 In addition
to controls for charter school age and student mobility, they also include
school-level demographics to account for schoolwide peer effects. Similar to
Hanushek, Kain, and Rivkin (2002), they find that new charter schools pro-
duce lower achievement gains in both math and reading than the average
traditional public school and the relative performance of charters improves
over time. Tuttavia, while Hanushek, Kain, and Rivkin find that charters in
operation three or more years are on par with the average traditional public
school, Booker et al. (2004UN) estimate that Texas charters in operation six years
or more surpass the performance of traditional public schools. This effect must
be viewed with caution, Tuttavia, since it is based on the achievement gains
of at most 256 students.5

4. Two of the study’s authors previously conducted a study of Texas charter schools (Gronberg and
Jansen 2001) which covered only three years and fewer than 1,000 charter students. This prior
paper finds students in charter schools do not perform as well on the TAAS exam as do traditional
public school students. Tuttavia, students attending charters primarily serving at-risk students
outperform students in traditional public schools with similar characteristics. These estimates may
be biased, Tuttavia, since Gronberg and Jansen include the lagged test score as an explanatory
variable in their model but ignore the correlation between the lagged score and the error term and
estimate the models using ordinary least squares.
Booker et al. (2004UN) also estimate models that allow for different transition costs of moving from
traditional public schools to charters vis-`a-vis movement between traditional public schools. In this

5.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

95

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Bifulco and Ladd (2004) analyze achievement data for students in North
Carolina over the period 1996–2002. Their data set tracks five cohorts of
students from grades 3–8. Their sample includes 496,000 students in total,
8,700 of which attended a charter school at least one year. Of the 8,700 students
who attended a charter, Di 5,700 are observed in both traditional and charter
schools. They adopt the same methodology as Hanushek, Kain, and Rivkin
(2002) but obtain some contrary results. Like Hanushek, Kain, and Rivkin,
they find that students attending brand-new charters have lower test score
gains in both reading and math than students in the average traditional public
school. Allo stesso modo, they find that the negative charter effects tend to diminish
as charters mature. Tuttavia, unlike Hanushek, Kain, and Rivkin’s results for
Texas, Bifulco and Ladd find that in North Carolina the negative impact of
charter schools on student achievement gains is statistically significant and
quantitatively substantial even for schools in operation for five years.

The recent studies by Hanushek, Kain, and Rivkin (2002), Booker et al.
(2004UN), and Bifulco and Ladd (2004) are laudable for their application of fixed-
effects modeling techniques to large panels of individual student data. How-
ever, their focus on the average effects of charter schools on student achieve-
ment does little to explain why charter schools perform better or worse than
traditional public schools. Likewise, although Hanushek, Kain, and Rivkin doc-
ument large quality variation among charter schools, neither they nor Bifulco
and Ladd nor Booker et al. analyze characteristics of charter schools, other
than age and student mobility, that determine charter school performance.

The Competitive Effects of Charter Schools

Advocates of charter schools claim that charters will not only provide a superior
education to the students who enroll in them but will also foster competition
that will lead to increases in the quality of traditional public schools. A number
of authors, including Bettinger (1999), Eberts and Hollenbeck (2001), Greene
and Forster (2002), and Holmes, DeSimone, and Rupp (2003), have attempted
to test this claim by making cross-sectional school-level comparisons. Their re-
sults are generally quite mixed, ranging from large positive competitive effects
in some instances to small or statistically insignificant competitive impacts in
others. Hoxby (2003) also employs school-level data but compares traditional
school performance before and after the introduction of charter school com-
petition. Defining the competitive threshold as a districtwide 6 percent char-
ter school enrollment share, she finds that in Arizona and Michigan charter

specification the negative first-year charter effect becomes insignificant. Tuttavia, since most first-
year charters are populated by students switching from traditional public schools to charters, IL
insignificant measured impact of first-year charter schools could simply be due to multicollinearity.

96

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

competition boosts traditional public school achievement score levels in both
math and reading in fourth grade as well as increasing seventh-grade math
scores.

Two recent studies, Holmes (2003) and Bifulco and Ladd (2004), exploit
student-level data from North Carolina to estimate the impact of charter school
competition on academic achievement in traditional public schools. Holmes
uses cross-sectional student-level data over multiple years with controls for stu-
dent race/ethnicity and gender. A student’s prior school inputs are measured
by a cubic form of the previous achievement score, while school-level fixed
effects capture any time-invariant attributes of each school. He finds mixed
results for more narrow market definitions; the existence of one or more char-
ter schools within 6 miles is correlated with higher scores in math but not in
reading, whereas the existence of a charter within 12 miles has the opposite
effect—higher reading scores but unchanged math scores. When competition
is measured at the county level, Tuttavia, he finds the existence of a charter is
associated with higher test scores for traditional public schools in both math
and reading.

Bifulco and Ladd (2004) perform an analysis similar to Holmes (2003)
but possess panel data on individual students and can thus fully account
for both time-invariant student and school characteristics via fixed effects. In
contrast to Holmes, Bifulco and Ladd find no significant effect of charter school
competition on traditional public school performance. Neither the existence
of one or more charters within 2.5 miles, 2.5 A 5 miles, O 5 A 10 miles has
any statistically significant effect on test score gains of students in traditional
public schools in North Carolina. Allo stesso modo, variation in the number of charter
schools within 5 miles of a traditional public school does not have a significant
effect on the average traditional public school’s performance.

Booker et al. (2004B) also utilize panel data on individual students to in-
vestigate the impact of charter school competition on student performance
within traditional public schools in Texas. Employing both student and school
fixed effects, they find large and statistically significant positive effects of
charter school competition. UN 1 percent increase in the proportion of public
school students attending charter schools within a district is associated with a
15 percent increase in annual math score gains and an 8 percent increase in
annual reading score gains. Allo stesso modo, UN 1 percent increase in the proportion
of students leaving a school to move to a charter is associated with a 9 per cento
increase in math score gains and a 6 percent increase in reading score gains
for those students who remain in the school. These competitive effects are
even larger when instruments are used to account for possible endogeneity of
charter school penetration. È interessante notare, Booker et al. also find that the im-
pact of charter school competition varies with the initial performance level of

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

97

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

traditional public schools. Schools with the lowest proportion of students pass-
ing the statewide exam experience a large positive impact of charter competi-
zione, whereas schools initially in the highest quintile see a drop in test score
gains with increases in charter penetration.

3. ECONOMETRIC MODEL
Following Todd and Wolpin (2003), consider a general cumulative model of
student achievement:

Ait = At

(cid:1)
Fi(T), Si(T), µi0, εit

(cid:2)

.

(1)

Ait is the achievement level for individual i at the end of their t-th year
of life, Fit is a vector of family/parental inputs supplied during age t, Sit is
a vector of school-supplied inputs during age t, µi0 is a composite variable
representing individual time-invariant characteristics (per esempio., innate ability), E
εit captures any measurement error. Fi(T) and Si(T) represent the entire input
histories of family and school inputs, rispettivamente.

If we assume that the cumulative achievement function, A(·), does not vary
with age6 and is additively separable, then we can rewrite the achievement level
at age t as:

Ait = α1Fit + α2Fit−1 + · · · + αtFi1
2Sit−1 + · · · + β

1Sit + β

+β

tSi1 + γt µi0 + εit

(2)

where α1 and β
family and school inputs, α2, and β
and so on.

1 represent the vectors of weights given to contemporaneous
2, the weights given to last year’s inputs,

Estimation of equation (2) requires data on both current and all prior fam-
ily and school inputs. Tuttavia, administrative records contain only limited
information on family characteristics and no direct measures of parental in-
puts. Therefore I assume that family inputs are constant over time and are
captured by a student-specific fixed component, φi. The marginal effect of
these fixed parental inputs on student achievement may vary over time and is
represented by κt. This of course implies that the level of inputs selected by
families does not vary with the level of school-provided inputs a child receives.
Per esempio, it is assumed that parents do not systematically compensate for

6. This assumption implies that the impact of an input on achievement varies with the time span
between the application of the input and measurement of achievement but is invariant to the age at
which the input was applied. Così, Per esempio, attending a private school in kindergarten has the
same effect on achievement at the end of third grade as attending a private school in second grade
has on fifth-grade achievement.

98

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

low-quality schooling inputs by providing tutors or other resources.7 Given
these assumptions, the achievement-level equation becomes:

Ait = β

1Sit + β

2Sit−1 + · · · + β

tSi1 + κtφi + γtµi0 + εit.

(3)

The need for data covering the entire school input history can be avoided
if one is willing to assume that the marginal impacts of all prior school inputs
decline geometrically with the time between the application of the input and the
measurement of achievement at the same rate (cioè., β
=
λ3β

1, and so on). The achievement equation can then be expressed as:

= λ2β

= λβ

, β

, β

4

2

3

1

1

Ait = β

1Sit + λβ

1Sit−1 + · · · + λt−1β

1Si1 + κtφi + γtµi0 + εit.

(4)

Taking the difference between current achievement and λ times prior achieve-
ment yields:

Ait − λAit =

(cid:1)

β

−

1Sit + λβ
(cid:1)
λ(β

1Sit−1 + · · · + λt−1β

1Si1 + κtφi + γtµi0 + εit

(cid:2)

1Sit−1) + λ(λβ

1Sit−2) + · · · + λ(λt−2β

1Si1)

+ λκt−1φi + λγt−1µi0 + λεit−1] .

(5)

Collecting terms, simplifying, and adding λAit−1 to both sides produces:

Ait = β

1Sit + λAit−1 + (κt − λκt−1)φi + (γt − λγt−1)µi0 + εit − λεit−1.

(6)

Assuming the impact of parental inputs on achievement, κt, and the effect
of the initial individual endowment on achievement, γt, change at constant
rates, Poi (κt − λκt−1) E (γt − λγt−1) can be expressed as constants, κ and γ .
Combining the family/parental inputs with the initial individual endowment
into a single component yields:

Ait = β

1Sit + λAit−1 + νi + ηit.

(7)

where νi = κφi + γ µi0 and ηit = εit − λεit−1. In this so-called value-added spec-
ification, the current achievement level is a function of current school inputs,
lagged achievement, and an individual-specific fixed effect.8

7.

8.

For evidence on the impact of school resources on parental inputs, see Houtenville and Conway
(2003) and Bonesrønning (2004).
It is important to recognize that I am modeling achievement levels, not achievement growth. An
achievement growth model would take the form (cid:8)Ait = πSit + λ(cid:8)Ait + (cid:10)io + ωit.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

.

F

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

99

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Education researchers typically estimate a restricted form of the value-
added specification where the gain in student achievement from one year to
il prossimo (the “gain score”) is a function of contemporaneous inputs and an
individual-specific fixed effect (Vedere, Per esempio, Hanushek, Kain, and Rivkin
2002, Bifulco and Ladd 2004, and Booker et al. 2004UN):

Ait − Ait−1 = (cid:8)Ait = β

1Sit + νi + ηit.

(8)

This of course imposes the restriction that λ = 1 in equation (7).9 As noted
by Boardman and Murnane (1979) and Todd and Wolpin (2003), this implies
that the effect of each input must be independent of when it is applied. In other
parole, school inputs each have an immediate one-time impact on achievement
that does not decay over time. Per esempio, the quality of a child’s kindergarten
must have the same impact on his achievement at the end of age five as it does
on his achievement at age eighteen.10

In order to determine the impact of charter schools on educational
achievement I utilize the value-added model, equation (7). Estimation of
this model by ordinary least squares is problematic, Tuttavia. The lagged
achievement score regressor, Ait−1, will obviously be correlated with the
lagged measurement error, εit−1, and thus ordinary-least-squares (OLS) es-
timates of equation (7) will be biased. In order to obtained unbiased pa-
rameter estimates I employ the dynamic panel data estimator developed by
Arellano and Bond (1991). The Arellano and Bond model uses twice (E
greater) lagged levels of the dependent variable as instruments for the lagged
difference in the dependent variable in order to eliminate the correlation be-
tween the lagged dependent variable and the error term. The student-specific
effect is eliminated by first-differencing the data. Asymptotic standard errors
are utilized that are robust to general heteroskedasticity over individuals and
over time.11

In order to provide a comparison to other studies, I will also estimate the
restricted value-added model specified in equation (8). The obvious disadvan-
tage of this model is the restriction on the decay rate of prior school inputs, 1.
This model does have the advantage that it can be estimated by OLS, Tuttavia,
since there is no lagged dependent variable on the right side of the equation.

9. Alternatively, equation (8) can be interpreted as a special case of an achievement growth model, cioè.,
(cid:8)Ait = πSit + λ(cid:8)Ait + (cid:10)io + ωit, where λ equals zero (achievement growth is independent of past
school inputs).

10. Krueger (1999) makes a similar point and provides empirical evidence that class size reduction has

the greatest impact on achievement the first year a student is in a small class.

11. Standard error estimates adjusted for clustering of errors are not available for the Arellano and

Bond model.

100

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

F

.

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

Robust standard errors are provided, which allow for clustering of errors at the
school level.

Variation in school inputs, Sit, will be measured by student mobility (chang-
ing schools between academic years and within academic years), the general
category of school attended (traditional public or charter) and the type of charter
attended (per esempio., whether or not it targets a specific student population, whether
it is run by a nonprofit organization or a for-profit management firm, E
whether it is a conversion from a traditional public school or a school that be-
gan de novo). Così, any estimated differences between charter and traditional
public schools will represent the combined effects of differences in all other
school inputs (teacher quality, class size, curriculum, eccetera.).

4. DATA AND RESULTS
Data

The Florida Department of Education’s Education Data Warehouse maintains
longitudinal records on all Florida public school students, from preschool
through college, beginning with the 1995/1996 school year. The Data Ware-
house includes not only test scores and student demographic data but also
information on enrollment, attendance, disciplinary actions, and participation
in exceptional student education and limited English proficiency programs.

Although student records are available since the 1995–96 school year,
statewide standardized testing did not begin in Florida until school year 1997–
98. In that year students began taking the “Sunshine State Standards” Florida
Comprehensive Achievement Test (FCAT-SSS). This is a criterion-based exam
designed to test for the skills that students must achieve at each grade level
in order to be promoted and to eventually graduate from high school. IL
FCAT-SSS was administered in selected grades from 1997–98 through 1999–
2000 and has been administered in grades 3–10 since the 2000–1 school year.
Beginning with the 1999–2000 school year a second test, the FCAT Norm-
Referenced Test (FCAT-NRT), has been administered to all third through
tenth graders. The FCAT-NRT is a version of the SAT-9 achievement test used
throughout the country. The FCAT-NRT scores are scaled so that a one-point
increase in the score at one place on the scale represents the same difference
in performance as a one-point increase anywhere else on the scale.12 The
Stanford-9 is a vertically scaled exam, thus scale scores typically increase with

12. See Harcourt Brace Educational Measurement (1997, P. 17). The use of scale scores to evaluate
student achievement is important, since a one-unit change has the same meaning for low- E
high-achieving students. Other types of measures, such as standard deviations from the mean
score, are potentially problematic; it is not clear that a 0.1 standard deviation increase in a test score,
starting one standard deviation from the mean, is the same as a 0.1 standard deviation increase for
someone with an initial score equal to the sample mean.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

101

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Tavolo 2 Characteristics of Students Enrolled in Grades Pre-K–12 by School Type,
2002–3 School Year

Traditional
Public Schools

Charter
Schools

Number of students

2,588,172

52,860

Percent black

Percent Hispanic

Percent female

Percent free/reduced-price lunch

Percent with limited English proficiency

Percent in special education (excluding gifted)

Percent gifted

24.43

20.70

48.38

44.61

10.73

14.47

4.23

28.62

21.07

49.94

36.54

8.76

12.36

2.67

Note: Student totals are based on enrollment throughout the school year. School
type is determined by the longest duration of enrollment. Program participation
is based on the October membership survey of public schools.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

.

F

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

the grade level. I use FCAT-NRT scale scores in all of the subsequent analyses.
In addition to providing four continuous years of data, use of the FCAT-
NRT minimizes potential biases associated with “teaching to the test” since
all school accountability standards and promotion and graduation criteria in
Florida are based on the FCAT-SSS rather than the FCAT-NRT.

Characteristics of Students in Charter Schools

Tavolo 2 provides a breakdown of student characteristics for both traditional
and chartered public schools. The characteristics of students attending charter
schools in Florida are quite similar to those of students in traditional public
schools. Except for a somewhat lower proportion of students from low-income
households (as indicated by free/reduced-price lunch receipt) and gifted stu-
dents and a somewhat higher enrollment of black students, there is little
difference in the measured characteristics of students attending charters ver-
sus traditional public schools in Florida. Così, at least at an aggregate level
there appears to be no strong evidence that charter schools cream the best
students from traditional public schools.

Student Achievement in Charter Schools

Estimation of the value-added and restricted-value-added models with individ-
ual fixed effects requires a minimum of three consecutive years of achievement
score data. Così, my analysis focuses on the sample of Florida public school
students in grades 3–10 who took the FCAT-NRT three consecutive years

102

EDUCATION FINANCE AND POLICY

Tim R. Sass

Tavolo 3 Enrollment Patterns of Students in Sample (Florida Students Who Took the FCAT-NRT
in Three Consecutive Years, 1999–2000 through 2002–3)

Number of Students

Percent

Enrollment Pattern

All Students

Students with All Enrollments Known

Students with No Observed Transitions

Only Tradtional

Only Charter

Students with One or More Observed Transitions

One Transition

Traditional to Charter

Charter to Traditional

Two Transitions

Traditional to Charter to Traditional

Charter to Traditional to Charter

Three Transitions

Traditional to Charter to Traditional to Charter

Charter to Traditional to Charter to Traditional

1,090,242

1,079,028

1,053,765

1,050,131

3,634

25,263

20,746

15,446

5,300

4,430

4,288

142

87

67

20

Students with One or More Unknown Enrollments

11,214

100.00%

98.97%

96.65%

96.32%

0.33%

2.32%

1.90%

1.42%

0.49%

0.41%

0.39%

0.01%

0.01%

0.01%

0.00%

1.03%

during the period 1999–2000 through 2002–3.13 This initial sample includes
Sopra 1 million students, more than 28,000 of which attended a charter school
in one or more of the three sample school years. Since the parameters in
fixed-effects models are identified by within-student intertemporal variation,
transitions between charters and traditional public schools are key to the anal-
ysis. Tavolo 3 provides a breakdown of the enrollment patterns of students and
indicates that over 25,000 students in the sample switched at least once be-
tween traditional public schools and charters—substantially more than in any
previous study of charter school performance.

A potential source of bias in both the value-added and restricted-value-
added achievement models is that the choice of moving to a charter school
may be due to a temporary drop in student performance. Per esempio, UN
student could draw a relatively low-quality teacher one year and perform poorly
relative to past years, which could in turn lead the child’s parents to switch the
student from a traditional public school to a charter school. If the student’s
performance would have rebounded in the next year (even if they had stayed
in a traditional public school), then the measured effect of charter schools will

13. The sample includes students who repeat a grade as well as those making a normal one-grade
progression from year to year. Separate grade-level dummies are included in the models for grade
repeaters.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

F

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

103

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

be biased upward. To investigate the possibility of such mean-reversion bias,
I estimate a probit model of the determinants of switching from a traditional
public school to a charter school.

Computing the deviation in prior-year performance from the past trend
requires four years of data (two years to establish the previous trend, one year
prior to the enrollment decision and a fourth year covering the choice between
traditional public schools and charter schools). Thus my sample is the group
of students in 2002–3 who attended traditional public schools and took the
FCAT-NRT in each of the prior years 1999–2000 through 2001–2. Controls
for student demographics as well as the number of available charter schools
within a 5-mile radius are included in the model. The results, presented in
Tavolo 4, indicate that there is no significant correlation between last year’s
deviation from the prior achievement trend and the current year’s decision
to enroll in a charter school.14 Indeed, the mean-reversion hypothesis would
predict a negative correlation between prior-year deviation from trend and
current-year charter enrollment, yet the point estimates are all positive. Così
estimates of the value-added and restricted-value-added models should not be
subject to bias resulting from reversion to the mean.

Tavolo 5 presents estimates of the average effect of charter schools on stu-
dent achievement in both math and reading. The value-added results indicate
that student achievement in the average charter school is 1.2 scale-score points
lower in math and 0.5 points lower in reading than the average traditional pub-
lic school. The size of these differentials depends on the basis for comparison.
In terms of cross-sectional differences among students, the differentials are
not very large. The estimated scale-score differential in math is equivalent to
2 percent of the standard deviation in all math scores, which equals 49. Like-
wise, the scale-score differential in reading is only 1 percent of the standard
deviation in all reading scores of 47. Tuttavia, if one compares the average-
charter achievement differentials to the average year-to-year score gains, IL
differences are more substantial, equivalent to 8 percent of the average 16-point
year-to-year gain in math and 4 percent of the average 12-point year-to-year gain
in reading.

The effects of student mobility on achievement are captured by three vari-
ables: Number of Schools, Structural Move, and Nonstructural Move. The first
variable, Number of Schools, measures the number of schools a student at-
tended during the current school year, thereby controlling for within-year

14. My measure of deviation from prior achievement is the difference in achievement gains,
(cid:8)At−1 − (cid:8)At−2. This is equivalent to the difference in the achievement level, At−1, and the ex-
pected achievement level based on a linear extrapolation of past performance, A∗
t−1. The predicted
level would be At−2+ (At−2 − At−3) which equals At−2 + (cid:8)At−2. The difference between the actual
achievement level and the predicted level in t−1 is (At−1− (At−2 + (cid:8)At−2) O (cid:8)At−1 − (cid:8)At−2.

104

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

.

F

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

Tavolo 4 Probability of Switching from a Traditional Public School to a Charter School, 2002–
3 (Florida Students Who Were Enrolled in Traditional Public Schools and Took the FCAT-NRT
in Each of the Years 1999–2000 through 2001–2)

Math Achievement Gaint−1–
Math Achievement Gaint−2

2.3 × 10−6
(1.42)

Reading Achievement Gaint−1–
Reading Achievement Gaint−2

Number of Charter Schools

within 5 Milest

Free/Reduced-Price Luncht−1

Limited English Proficiencyt−1

Special Educationt−1

Giftedt−1

Black

Hispanic

Multiracial

Female

2.5 × 10−6
(1.60)

2.3 × 10−8
(0.02)

3.2 × 10−7
(0.21)

1.0 × 10−3∗∗

1.0 × 10−3∗∗

1.0 × 10−3∗∗

(11.35)

(11.36)

(11.13)

1.7 × 10−4
(0.60)

–1.3 × 10−4
(0.23)

1.4 × 10−4
(0.67)

1.9 × 10−4
(0.67)

–2.8 × 10−4
(0.53)

1.6 × 10−4
(0.78)

1.8 × 10−4
(0.64)

–1.9 × 10−4
(0.35)

1.6 × 10−4
(0.77)

–9.0 × 10−4∗
(2.18)

–0.87 × 10−4∗
(2.16)

–9.0 × 10−4∗
(2.17)

1.5 × 10−3∗∗
(5.32)

9.3 × 10−4∗
(2.48)

2.4 × 10−4
(0.35)

–1.1 × 10−4∗∗
(0.83)

1.5 × 10−3∗∗
(5.48)

9.5 × 10−4∗∗
(2.64)

3.6 × 10−4
(0.56)

–1.1 × 10−4
(0.87)

1.5 × 10−3∗∗
(5.30)

9.3 × 10−4∗
(2.50)

2.8 × 10−4
(0.42)

–1.3 × 10−4
(0.99)

Observed probability of

attending a charter school

3.34 × 10−3

3.29 × 10−3

3.33 × 10−3

Number of students

Number of observations

614,271

614,271

630,158

630,158

608,357

608,357

Notes: Absolute values of robust t-ratios adjusted for clustering of errors at the prior-
school level in parentheses. All models include grade dummies and a constant. Re-
ported coefficients are marginal effects. For charter school students the number of
charter schools within a five-mile radius is based on the location of the traditional
public school that a plurality of their prior-year traditional public school classmates
attend in the current year.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

mobility.15 The second and third variables identify two kinds of between-year
school transitions. Following Hanushek, Kain, and Rivkin (2002), strutturale
moves are defined as situations where a student moves from one school to
another and at least 30 percent of his fellow students in the same grade at the
initial school move to the same school. Thus the variable Structural Move cap-
tures the effects of normal transitions from elementary to middle and middle

15. Only schools attended two weeks or more are counted. Per esempio, a student who normally attends
a single school but is temporarily assigned to a juvenile-justice school for a short duration and then
returns to his regular school is counted as attending one school.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

105

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Tavolo 5 Estimates of the Average Effect of Charter Schools on Student Achievement

MATH

READING

Restricted
Value-Added
Model

Value-
Added
Model

Restricted
Value-Added
Model

–0.932∗
(2.56)

–0.638∗∗
(4.88)

–1.933∗∗
(8.01)

–0.870∗∗
(5.57)

–0.022
(0.06)

–0.666∗∗
(5.06)

–1.806∗∗
(9.88)

–0.680∗∗
(5.30)

–0.474∗
(2.32)

–0.638∗∗
(7.26)

–1.128∗∗
(17.81)

–0.555∗∗
(9.72)

0.203∗∗

(104.67)

Value-
Added
Model

–1.202∗∗
(6.47)

–0.742∗∗
(9.64)

–0.967∗∗
(17.03)

–0.633∗∗
(12.37)

0.088∗∗

(36.58)

Chartert

Number of Schoolst

Structural Movet

Nonstructural Movet

Achievement Scoret−1

Number of students

1,065,443

1,065,443

1,068,161

1,068,161

Number of observations

1,704,593

2,768,486

1,724,663

2,791,246

Notes: Absolute values of robust t-ratios appear in parentheses. For the restricted-value-
added model, the reported t-ratios account for clustering of errors at the school level. Tutto
models include time and grade dummies and a constant as appropriate. Reported number
of observations for the value-added model is number of observations after first-differencing
of variables.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

to high school as well as the impact of significant school rezonings. Corre-
spondingly, the Nonstructural Move variable represents students who attend
a school different from the one attended at the end of the preceding school
year but are not joined by at least 30 percent of their former schoolmates.
This encompasses family relocations as well as movements between schools
to attend magnet or other specialized programs. Consistent with the findings
of other researchers, all of the mobility measures have significant negative
effects on student achievement; and structural moves are more harmful to
student learning than nonstructural moves.

It is also interesting to note that the maintained assumption of the
restricted-value-added model, that the effect of past school inputs doesn’t
decay (cioè., λ = 1), is not supported by the data.16 Estimates from the unre-
stricted value-added model indicate a value for λ of 0.09 in math and 0.20 In
reading.17 These estimates suggest that the effects of prior schooling inputs on

16. I also directly estimated the cumulative function specified in equation (3). The effects of lagged
charter school attendance and lagged intrayear and between-year school changes also suggest a
value of lambda less than one.

17. The first-stage estimation of the lagged dependent variable in the value-added model yielded R2
values of 0.12 for math and 0.24 for reading using twice-lagged achievement levels as an instrument
and R2 values of 0.23 for math and 0.35 for reading when both twice-lagged and three-times-lagged

106

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

.

F

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

the current achievement (holding student ability constant) diminish rapidly,
decaying 80–90 percent after one year and 96–99 percent after two.

The results presented in Table 5 mask the important effects of maturation
on charter school performance. As previous studies have demonstrated, one
would expect that performance improves as charter schools age and become
better established. The important issue is how charter schools compare to
traditional public schools in the long run. To address this question I estimate
two models: one with a linear age trend in charter school performance and
another with separate effects for first-year, second-year, third-year, and fourth-
year charters and traditional public schools.18 In the linear-trend model charter
age is defined as the number of prior school years the school has been in
operation, so the age of charters in their first school year of operation is zero.
Estimates of these two models are presented in the upper and lower panels,
rispettivamente, of Table 6.

Estimates of both models (Tavolo 6) indicate that student achievement in
math and in reading are lower, on average, in brand-new charter schools than
in traditional public schools. Value-added model estimates indicate math test
scores are 2.0 A 2.1 points (13 percent of average annual gain) lower in first-
year charters, while point estimates from the restricted-valued-added models
yield somewhat higher differentials of 2.5 A 2.7 points. For reading scores the
value-added linear-trend model indicates a deficit of 1.5 scale points (12 per cento
of average annual gain) for new charters, while the model with an uncon-
strained age profile yields an estimated difference of 1.2 points (10 per cento
of average annual gain). For the restricted-value-added models there are no
statistically significant differences in reading achievement between first-year
charters and mature traditional public schools.

The results presented in Table 6 also show that both math and reading
scores in charter schools generally improve as the school matures. The value-
added model estimates indicate that charters operating five years or more are
on par with the average mature traditional public school in math achievement
and produce reading achievement scores that are 1.1 scale-score points (9 per-
cent of average annual achievement gain) higher than their traditional public

achievement levels are included as instruments. I also estimated value-added models with twice-
lagged levels of student attendance, retention, and disciplinary incidents as additional instruments.
The results differed little from those presented in Table 5, and a Sargan test rejected the validity
of the additional instruments. Allo stesso modo, the R2 values of the first-stage regression changed little
with the additional instruments; they were 0.13/0.24 for math and 0.24/0.36 for reading. In all
cases F-tests reject the null hypothesis that the coefficients on the instruments are zero at better
than a 99.99 percent confidence level.
I make no distinction between charters operating five years and those operating six or more years
since there are fewer than thirty schools that began operation prior to 1998–99 and that were still
in operation during the last year of the sample, 2002–3.

18.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

107

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Tavolo 6 Estimates of the Effect of Charter Schools on Student Achievement, Controlling for School Age

MATH

READING

Charter

Charter % Age of Charter

First-Year Charters

Second-Year Charters

Third-Year Charters

Fourth-Year Charters

Fifth-Year and Older Charters

First-Year

Traditional Public Schools

Second-Year

Traditional Public Schools

Third-Year

Traditional Public Schools

Fourth-Year

Traditional Public Schools

Value-
Added
Model

–2.139∗∗
(7.28)

0.441∗∗
(4.28)

–2.002∗∗
(5.69)
–1.118∗∗
(3.65)
–2.270∗∗
(7.92)
–1.035∗∗
(3.85)
0.258
(0.82)
–0.577∗
(2.14)
–0.720∗∗
(5.50)
0.093
(0.80)
0.413
(3.54)

Restricted
Value-Added
Model

–2.730∗∗
(3.21)

0.810∗∗
(2.62)

–2.510∗
(2.26)
–0.585
(0.80)
–2.876∗∗
(2.92)
–0.894
(1.33)
1.655∗
(2.00)
–0.287
(0.25)
–0.122
(0.24)
0.604
(1.49)
0.648
(1.38)

Value-
Added
Model

–1.504∗∗
(4.60)

0.482∗∗
(4.20)

–1.197∗∗
(3.05)
–0.930∗∗
(2.74)
–1.087∗∗
(3.39)
–0.424
(1.41)
1.103∗∗
(3.13)
–0.189
(0.63)
–0.038
(0.26)
0.106
(0.81)
0.001
(0.00)

Restricted
Value-Added
Model

–1.256
(1.78)

0.552∗
(2.29)

–1.361
(1.44)
–0.140
(0.20)
–0.596
(0.83)
0.173
(0.23)
1.511∗
(2.37)
1.090
(0.96)
0.734
(1.75)
0.596
(1.83)
0.175
(0.40)

Number of students

1,065,443

1,065,443

1,068,161

1,068,161

Number of observations

1,704,593

2,768,486

1,724,663

2,791,246

Notes: Absolute values of robust t-ratios appear in parentheses. For the restricted-value-added
modello, the reported t-ratios account for clustering of errors at the school level. All models include
time and grade dummies and a constant as appropriate. Models also include control variables
reported in Table 5. Reported number of observations for the value-added model is the number of
observations after first-differencing of variables.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

school counterparts. The restricted-value-added model estimates suggest even
greater improvement in charter performance over time; fifth-year and older
charters yield math and reading scores that are 1.7 E 1.5 scale-score points,
rispettivamente, above the average mature traditional public school. In contrast to
charter schools, traditional public schools do not demonstrate any consistent
pattern of maturation effects.19

The results presented in Table 6 may not provide an accurate picture
of charter school maturation effects because they do not control for the age

19. The difference in maturation patterns between traditional public schools and charters is not sur-
prising. Traditional public schools almost always start out with newly constructed facilities, an es-
tablished curriculum, and can draw on administrators and faculty from established public schools.
In contrasto, charters often begin in temporary quarters designed for other purposes, must recruit
faculty from a variety of sources and frequently are attempting to implement a new curriculum.

108

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

F

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

of schools at the time they become charters. The vast majority of charter
schools in Florida are created de novo. Tuttavia, as indicated in Table 1, there
are a growing number of “conversion charters”—traditional public schools
that choose to become charter schools.20 Since these conversion charters
have been in operation for many years, one would expect that the num-
ber of years since becoming a charter school would not significantly affect
their performance. Tavolo 7 presents two sets of estimates that constrain the
effect of age since acquiring charter status to be zero for conversion char-
ters.21 In the top panel, nonconversion charter performance varies with age
in a linear fashion, while in the bottom panel, separate intercepts are in-
cluded for first-year, second-year, third-year, and fourth-year nonconversion
charters.22

When the distinction between conversion and nonconversion charters is
made, a clear maturation pattern for nonconversion charters emerges. Es-
timates from the value-added model indicate new nonconversion charters
produce achievement test scores 2.2 A 2.4 points (14 A 15 percent of annual
achievement gain) lower in math and 2.2 A 2.4 points (18 A 20 percent of
annual achievement gain) lower in reading. These deficits diminish over time,
Tuttavia. In math the gap closes by about a half-point per year, and there is no
statistically significant difference in math achievement scores for fourth-year
and older nonconversion charters and the average traditional public school.
For reading, the improvement in achievement scores with charter age is even
more rapid. Reading test scores are estimated to rise an average of 0.72 points
with each year of charter operation. Charters operating for five years or more
show reading achievement levels that exceed the average traditional public
school by 1.3 scale-score points (10 percent of average annual achievement
gain). Estimates from the restricted-value-added model show similar patterns:
both the intial charter deficit and subsequent improvement from maturation
are greater.

In addition to their age, charter schools vary in numerous dimensions,
including curricular emphasis, the student population they serve, and their

20. In Florida, conversion requires separate affirmative majority votes by both the parents and faculty

21.

of a school and subsequent approval by the local school board.
I also estimated models that allowed for a nonzero maturation effect for conversion charters that
differed from that for nonconversion charters. The maturation effect for conversion charters was
negative and significant in the value-added model for math and insignificantly different from zero
in all other specifications. The other results were nearly identical to the estimates presented in
Tavolo 7. Given the small number of older conversion charters, the negative estimated maturation
effect for math scores could represent some unmeasured attributes of the particular schools that
were early converters.

22. Conversion charters are placed in the same category as nonconversion charters in their fifth or

higher year of operation.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

F

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

109

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Tavolo 7 Estimates of the Effects of Charter Schools on Student Achievement, Controlling for
Charter School Age and Charter School Type

MATH

READING

Value-
Added
Model

–2.215∗∗
(4.63)

0.543∗∗
(3.98)

2.213∗∗
(2.72)

–1.192∗
(2.21)

–0.314
(0.59)

–2.414∗∗
(3.94)

–1.079∗
(2.30)

–1.673∗∗
(4.36)

–0.635
(1.83)

0.360
(0.92)

–0.375
(0.49)

–1.239∗
(2.29)

–0.315
(0.59)

Restricted
Value-Added
Model

–3.401∗∗
(2.67)

1.097∗∗
(2.65)

2.444
(1.22)

–0.292
(0.25)

–0.781
(0.84)

–4.204∗
(2.51)

0.022
(0.02)

–2.323
(1.89)

–0.456
(0.57)

1.997∗
(2.10)

–2.932
(1.63)

–0.477
(0.40)

–1.102
(1.06)

Value-
Added
Model

–2.448∗∗
(4.59)

0.715∗∗
(4.74)

1.233
(1.39)

–0.410
(0.70)

0.156
(0.27)

–2.222∗∗
(3.30)

–1.474∗∗
(2.84)

–0.836
(1.95)

–1.034∗∗
(2.63)

1.276∗∗
(2.94)

–2.487∗∗
(2.98)

–0.499
(0.85)

0.071
(0.12)

Restricted
Value-Added
Model

–2.295∗
(2.32)

0.788∗
(2.50)

1.001
(0.74)

0.084
(0.09)

0.083
(0.08)

–3.306∗∗
(2.81)

–0.380
(0.43)

–0.515
(0.59)

–0.858
(1.23)

1.840∗∗
(2.80)

–3.089∗∗
(2.73)

–0.053
(0.05)

0.035
(0.04)

Charter

Nonconversion Charter
× Age of Charter

Conversion Charter

Targeted Charter

Charter Managed

by For-Profit Firm

First-Year

Nonconversion Charters

Second-Year

Nonconversion Charters

Third-Year

Nonconversion Charters

Fourth-Year

Nonconversion Charters

Fifth-Year and Older Charters

Conversion Charter

Targeted Charter

For-Profit Charter

Number of students

1,061,668

1,061,668

1,064,390

1,064,390

Number of observations

1,696,732

2,756,859

1,716,612

2,779,432

Notes: Absolute values of robust t-ratios appear in parentheses. For the restricted-value-
added model, the reported t-ratios account for clustering of errors at the school level.
Models in the lower panel include dummy variables representing traditional public schools
in their first, second, third, and fourth year of operation. All models include time and grade
dummies and a constant as appropriate. All models also include control variables reported
in Table 5. Reported number of observations for the value-added model is the number of
observations after first-differencing of variables.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

organization. The models presented in Table 7 control for two important at-
tributes of charter schools besides age: whether they target a particular student
population and whether they are operated by a for-profit management com-
pany.

Greene, Forster, and Winters (2003) argue that most comparisons of char-
ter and traditional public schools are biased because a large proportion of

110

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

F

.

.

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

charter schools are targeted to serve educationally disadvantaged populations,
such as students with disabilities, students at risk of dropping out, and low-
income and migrant students.23 Given that I control for student characteristics
with individual-specific fixed effects, this should not be an issue. Tuttavia, if
specialized charters place a greater weight on factors that are not measured by
achievement tests (per esempio., life-management skills, vocational skills, foreign lan-
guages, performing arts), then they may have lower test scores than schools
with a more traditional emphasis on core academic skills. To account for this,
the models presented in Table 7 contain a dummy variable indicating charters
that identify themselves as serving a targeted population.24 In the value-added
models, targeted charters are estimated to have math test scores that are 1.2
scale-score points lower than the scores for charter schools serving a general
population. There are no statistically significant differences in reading test
scores between targeted and nontargeted charters.

The other dimension of charter schools I measure is their organization. In
Florida, as elsewhere, the majority of charter schools are run by local nonprofit
entities. Tuttavia, a growing number of charter schools are being managed
by for-profit firms. If managers of for-profit educational firms are residual
claimants on the net income of schools, they have a clear incentive to operate
schools efficiently. Since operating revenues in charter schools are essentially
constrained to equal the funding level of traditional public schools, for-profit
management companies will have an incentive to minimize the cost of provid-
ing a given level of educational services. In contrasto, operators of not-for-profit
schools will seek to maximize their utility, which typically would include the
welfare of students but might encompass other objectives as well.25 Thus from
a theoretical standpoint it is not clear which type of firm would produce the
greatest contribution to student achievement.

The estimates presented in Table 7 do not show any difference in per-
formance between nonprofit charter schools and charters run by for-profit
management companies. In both the value-added and restricted-value-added

23. Greene, Forster, and Winters (2003) find that when the school-level achievement gains of nontar-
geted charters are compared to those of traditional public schools, charters outperform the average
traditional public school. Tuttavia, their models do not account for the age of charters and exclude
conversion charters. Since most nonconversion charters are young, the relatively high year-to-year
achievement gains may be due to charter maturation rather than an indicator of superior perfor-
mance relative to traditional public schools.

24. The information on targeting by Florida charter schools is based on survey data from Greene,
Forster, and Winters (2003), a separate survey of Florida charters conducted by Robert Crew, E
information from school Web sites.

25. For a more detailed discussion of the incentives in nonprofit versus for-profit firms see Lien (2002)

and references cited therein.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

111

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

models the differences in math and reading achievement test scores between
the two types of organizational forms are statistically insignificant.

In order to measure the effectiveness of charter schools at different grade
levels, I reestimate the value-added and restricted value-added models for
three grade groupings: elementary (grades 4–5), middle (grades 6–8), E
high school (grades 9–10).26 The results are presented in Table 8. For reading,
the negative effect of first-year charters is greatest in the earlier grades, ma il
maturation effect is greater as well. For elementary students, the value-added
model estimates indicate that first-year charters are associated with a 4.1-point
decline in achievement test scores, but this deficit is eliminated by the fourth
year of operation. For middle school students, the first-year-charter deficit is
smaller (2.6 scale-score points), and the annual improvement is smaller as
BENE (0.6 points). At the high school level, there is no statistically significant
difference between reading scores for new charters and the average traditional
public high schools. For math, the pattern is reversed; the differences between
new charters and traditional public schools are greatest at the high school level.
Students attending a brand-new charter high school experience a 3.5-point
deficit in math achievement scores, but after three years of operation the math
achievement score deficit is eliminated. For middle schools the initial deficit
for new charters is only 2.1 scale-score points, but the improvement associated
with charter maturation is also smaller; middle school charters reach a par
with traditional public schools in math after five years of operation. Al
elementary school level, new charters are on par with traditional elementary
schools, and there is no significant improvement in their scores relative to
traditional public schools as the charters mature. È interessante notare, these different
patterns for reading and math achievement also show up in the measured
effects of conversion charters. The effect of conversion charters is significant
only at the high school level for math and at the elementary school level for
reading.

The models presented in Tables 7 E 8 do not differentiate between char-
ters targeting different populations of students. In 2002–3, approximately 22
percent of charter students in my sample attend a school that identified them-
selves as targeting a specific student population. Among students attending
such targeted charters, over half attend a school that is designed to serve
at-risk students. The second most frequent category of targeted charters
are those with programs directed toward serving students with disabilities;
17 percent of students attending targeted charters go to schools that emphasize

26. Given that achievement tests are administered in grades 3–10 and three annual scores are required
to estimate the model, the elementary school group includes fifth-grade students plus fourth graders
who repeated either grade 3 or grade 4.

112

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

Tavolo 8 Value-Added Estimates of the Effects of Charter Schools on Student Achievement by Grade Level

MATH

READING

Elementary Middle
(Grades
(Grades
6–8)
4–5)

High School Elementary Middle
(Grades
(Grades
6–8)
9–10)

(Grades
4–5)

High School
(Grades
9–10)

Value-Added Model

Chartert

–1.459
(1.12)

–2.098∗∗
(3.44)

–3.473∗∗
(3.81)

–4.080∗∗
(2.85)

–2.613∗∗
(3.73)

–0.550
(0.53)

Nonconversion
Chartert
× Age of Chartert

Conversion Chartert

Targeted Chartert

For-Profit Chartert

–0.196
(0.53)

3.832
(1.60)

–6.398∗∗
(3.65)

–2.774
(1.79)

0.416∗
(2.50)

0.406
(0.38)

–0.509
(0.70)

0.439
(0.68)

1.523∗∗
(5.13)

6.208∗∗
(4.38)

–1.979∗
(2.26)

0.057
(0.05)

1.302∗∗
(3.33)

5.993∗
(2.33)

–5.697∗∗
(3.14)

0.034
(0.02)

0.580∗∗
(3.03)

0.661∗
(2.01)

0.620
(0.54)

1.225
(1.50)

0.329
(0.46)

0.239
(0.14)

–1.756
(1.78)

–0.207
(0.18)

Number of students 302,885

615,699

397,188

303,901

617,715

398,387

Number of

observations

310,627

874,490

511,615

311,734

891,467

513,411

Restricted-Value-Added Model

Chartert

–2.641
(0.86)

–2.139
(0.77)

–2.384
(0.40)

–3.216
(1.10)

–2.941
(1.30)

–4.950
(1.51)

Nonconversion
Chartert
× Age of Chartert

Conversion Chartert

Targeted Chartert

For-Profit Chartert

0.420
(0.43)

7.158
(0.65)

–7.709∗
(2.04)

–2.301
(0.79)

0.983
(0.99)

–2.953
(0.91)

1.484
(0.77)

–1.167
(0.54)

1.035
(0.33)

7.610
(0.91)

–5.378
(0.86)

3.486
(0.76)

2.018∗
(2.03)

3.983
(0.60)

–9.765∗∗
(3.53)

–3.299
(1.03)

0.801
(1.15)

–0.825
(0.27)

2.683
(1.48)

–1.772
(0.92)

2.087
(1.55)

–1.767
(0.39)

–2.183
(0.67)

4.939
(1.51)

Number of students 302,885

615,699

397,188

303,901

617,715

398,387

Number of

observations

613,521

1,210,689 646,119

615,664

1,213,506 648,230

Notes: Absolute values of robust t-ratios appear in parentheses. For the restricted-value-added
modello, the reported t-ratios account for clustering of errors at the school level. All models include
time and grade dummies and a constant as appropriate. All models also include control variables
reported in Table 5. Reported number of observations for the value-added model is the number
of observations after first differencing of variables. Since testing begins in grade 3, the grade 4
sample includes only those students who repeated grade 3 or grade 4 and thus have 3 annual test
scores.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

special education. Approximately 25 percent of students attending targeted
charters are at schools with other specializations. Some schools emphasize
performing arts, multicultural education, or primarily serve migrants or gifted
students.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

113

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

Tavolo 9 Value-Added Estimates of the Effects of Charter Schools on Student Achievement, Control-
ling for Charter School Age, Charter School Type, and Target Population

MATH

READING

Value-
Added
Model

–3.206∗∗
(6.42)

0.905∗∗
(6.28)

3.280∗∗
(3.93)

–2.036∗∗
(2.86)

–8.042∗∗
(6.12)

3.829∗∗
(3.95)

0.070
(0.13)

Restricted
Value-Added
Model

–4.716∗∗
(3.29)

1.568∗∗
(3.28)

3.890
(1.81)

–2.120
(1.44)

–7.459∗∗
(3.75)

5.480∗∗
(2.89)

–0.257
(0.28)

Value-
Added
Model

–2.890∗∗
(5.19)

0.882∗∗
(5.52)

1.685
(1.86)

–0.855
(1.06)

–2.409
(1.73)

2.033
(1.90)

0.301
(0.51)

Restricted
Value-Added
Model

–2.656∗
(2.43)

0.926∗∗
(2.61)

1.384
(0.95)

–0.581
(0.41)

–0.589
(0.23)

1.739
(1.08)

0.128
(0.12)

Chartert

Nonconversion Chartert
× Age of Chartert

Conversion Charter

Charter Targeting At-Risk Students

Charter Targeting

Special-Ed. Students

Charter Targeting Other Students

For-Profit Charter

Number of students

1,061,542

1,061,542

1,064,259

1,064,259

Number of observations

1,696,426

2,756,428

1,716,290

2,778,981

Notes: Absolute values of robust t-ratios appear in parentheses. For the restricted-value-added
modello, the reported t-ratios account for clustering of errors at the school level. All models
include time and grade dummies and a constant as appropriate. Models also include control
variables reported in Table 5. Reported number of observations for the value-added model is
the number of observations after first-differencing of variables.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

Estimates of a model that allows the impact of charter schools on student
achievement to vary with the targeted population are presented in Table 9.
Schools targeting at-risk students produce math achievement scores that are
2.0 points lower, on average, than nontargeted charters. The gap is far greater
for charters emphasizing education of students with disabilities. Charters
targeting students with disabilities yield math achievement scores that are
8.0 points (over half the average annual gain) lower than nontargeted charters,
holding student characteristics constant.27 Specialized charters that focus on
programs other than those targeted to at-risk and special education students
possess math scores 3.8 points higher than nontargeted charters. In contrasto,
there are no differences in reading scores among nontargeted and the various

27. The results for charters emphasizing special education are similar when the sample is restricted
to only special education students. The math-score gap for special education students in charters
targeting special education students versus special education students in nontargeted charters is
8.3 points.

114

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

F

.

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

types of targeted charter schools. Although the estimated magnitudes differ,
the restricted-value-added models yield qualitatively similar results for the
effects of charter school specialization.

There are two likely explanations for the observed math achievement differ-
entials between charters targeting special-education and at-risk students and
schools that do not target these groups. Primo, it may be that charters targeting
special-education and at-risk students place greater weight on skills that are
important to these student groups but are not measured by standardized read-
ing and math tests. Per esempio, behavior control, development of social and
oral communication skills, and vocational training may receive greater weight
in these schools. Secondo, there is the potential for negative peer effects in these
schools. Not all students who attend special-education or at-risk charters are
disabled or at risk for dropping out. If disproportionally greater resources are
devoted to students with disabilities and at-risk youth, or special-education and
at-risk students themselves generate negative externalities, the performance
of their peers in the same school may suffer.

The Competitive Effects of Charters on Student Achievement in Traditional

Public Schools

To determine the competitive impact of charters on traditional public schools,
a geographic information systems (GIS) database was constructed covering all
public and private schools in Florida.28 To account for the product dimension
of the education market, enrollment data by grade were collected for all public
and private schools in the state. This information was used to group schools
into three categories: elementary (grades K–5), middle (grades 6–8), and high
school (grades 9–12). Schools could be included in more than one category if
they served students in more than one grade-level grouping. Using the GIS
and enrollment data, the number and enrollment shares of charters, private
schools, and other traditional public schools serving the same grade levels
(elementary, middle, or high school) within 2.5-, 5-, and 10-mile radii of each
traditional public school was determined.29

If charter schools and/or private schools tend to locate where traditional
public schools are performing poorly, then measures of the number and size
of charters and private schools would reflect not only the competitive impact

28. All charter schools, 98 percent of traditional public schools, E 91 percent of private schools were
geocoded based on their street address. The remaining schools were assigned latitude and longitude
values based on the centroid of their five-digit zip code.

29. IL 2 percent of traditional public schools with zip-code-level geocoding were excluded as center
points from the competition analysis but were included in the measures of competing traditional
public schools.

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

.

F

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

115

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

of their existence but also (unmeasured) traditional public school quality. Questo
would tend to bias downward the estimated effects of charter and private
schools on student achievement in traditional public schools. To control for
unmeasured time-invariant traditional school quality, a school fixed effect, θj
(where j indexes schools), can be added to the achievement model.30 Denoting
the vector of school competition measures by Cit, the achievement model
becomes:

Ait = β

1Sit + λAit−1 + ρCit + θj + νi + ηit.

(9)

Direct estimation of equation (9) is problematic, since it requires inclusion
of thousands of dummy variables, one for each traditional public school in the
sample.31 In order to make the problem computationally tractable, I combine
the student and school fixed effects into a single effect, δij = θj + νi, repre-
senting each unique student/school combination or “spell.”32 Employing the
assumption that λ = 1 (the restricted-value-added model) yields:

(cid:8)Ait = β

1Sit + ρCit + δij + ηit.

(10)

Although individual and school effects are not separately identified, both in-
dividual and school heterogeneity can be eliminated by differencing the data
with respect to spell means.33

Tavolo 10 presents estimates of student achievement gains in traditional
public schools as a function of various measures of competition. Each panel
presents estimates of equation (10) using three geographic market definitions
but with different measures of competition added to each. In the first two
panels competition is measured by the existence of charters, private schools,

30. The school fixed effect only accounts for the time-invariant component of preexisting public school
quality. If a temporary reduction in public school quality promotes the entry of new charter schools,
then the measured impact of charter competition could be biased. Tuttavia, given the costs of entry
and exit, it seems unlikely that charters would base their entry decisions on temporary reductions
in public school quality and much more likely that they would enter markets where there are
persistently low-quality traditional public schools.

31. The individual fixed effects can be removed by taking differences from individual means, but that

still leaves the effects of the school dummies to be estimated explicitly.

32. With very limited exceptions, charter schools in Florida must be approved by the local school district.
Thus the spell fixed effects also capture any factors influencing the process of approving new charter
schools.

33. Use of the restricted-value-added model is necessary to allow the use of mean-differencing to
eliminate spell fixed effects. The Arellano and Bond dynamic panel data procedure used to estimate
the unrestricted value-added model relies on first-differencing the data with respect to time. For
a more detailed discussion of the spell-fixed-effects approach see Andrews, Schank, and Upward
(2004). Standard errors for the spell-fixed-effects model are not adjusted for clustering at the school
level, since the school fixed effect should account for any systematic error that is common to all
students attending a particular school.

116

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

F

/

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

F

.

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

Tavolo 10 Restricted-Value-Added Model Estimates of the Effects of School Competition on Student
Achievement Gains in Traditional Public Schools (Student/School Fixed Effects)

MATH

READING

2.5-Mile
Radius

5-Mile
Radius

10-Mile
Radius

2.5-Mile
Radius

5-Mile
Radius

10-Mile
Radius

One or more charters

One or more charters

One or more private

schools

One or more other trad.

public schools

Number of charters

Number of charters

Number of private

schools

Number of other trad.

public schools

Market share of

charters

Market share of

charters

Market share of

private schools

Market share of other
trad. public schools

0.488∗∗
(2.72)

0.485∗∗
(2.71)

–0.180
(0.79)

0.210
(0.85)

0.372∗∗
(2.91)

0.378∗∗
(2.94)

–0.025
(0.57)

0.004
(0.05)

0.075∗∗
(3.19)

0.070∗∗
(2.95)

–0.081∗∗
(3.92)

0.016∗
(2.23)

0.346∗∗
(2.19)

0.346∗∗
(2.19)

0.067
(0.20)

0.053
(0.13)

0.075
(1.11)

0.051
(0.75)

0.065∗∗
(2.62)

–0.078∗
(2.12)

–0.034
(0.94)

–0.047
(1.27)

–0.067∗
(2.25)

–0.016
(1.79)

–0.007
(0.05)

–0.006
(0.04)

0.033
(0.06)

0.052
(0.08)

0.025
(0.65)

0.067
(1.63)

–0.028∗
(1.98)

–0.051∗∗
(2.70)

0.222∗∗
(4.75)

0.213∗∗
(4.30)

–0.016
(0.36)

–0.007
(0.55)

0.241
(1.29)

0.238
(1.27)

–0.137
(0.58)

0.230
(0.89)

0.202
(1.51)

0.216
(1.61)

–0.105∗
(2.25)

0.167∗
(2.13)

0.036
(1.50)

0.039
(1.60)

–0.021
(1.00)

0.017∗
(2.27)

0.215
(1.30)

0.210
(1.27)

0.734∗
(2.10)

0.787
(1.81)

0.028
(0.40)

0.046
(0.65)

–0.052∗
(1.97)

0.065
(1.70)

0.006
(0.16)

0.023
(0.58)

0.041
(1.33)

0.028∗∗
(2.98)

–0.103
(0.65)

–0.099
(0.62)

0.824
(1.43)

0.249
(0.35)

–0.100∗∗
(2.51)

–0.036
(0.84)

–0.072∗∗
(4.92)

0.023
(1.16)

0.035
(0.73)

0.108∗
(2.12)

0.082
(1.78)

0.060∗∗
(4.25)

Number of students

1,707,927 1,707,927 1,707,927

1,724,793 1,724,793 1,724,793

Number of observations 2,703,494 2,703,494 2,703,494

2,723,359 2,723,359 2,723,359

Notes: Absolute values of robust t-ratios appear in parentheses. All models include time and grade
dummies and a constant as appropriate. Models also include control variables reported in Table 5
and fixed effects for each unique student/school combination.

∗ statistical significance at .05 level, ∗∗ significance at the .01 level in a two-tailed test.

and other traditional public schools. Panels 3 E 4 use the number of
alternative schools as the yardstick of competition, and panels 5 E 6 use
enrollment market share to gauge the extent of competition. For each of these
measures, estimates of charter competition alone as well as competition from
charters, private schools, and other traditional public schools are provided.

The presence of one or more charters within a 2.5-mile radius is correlated
con un 0.5 point increase in math achievement score gains. This is equivalent
ad a 3 percent increase in the average yearly math-score gain in traditional
public schools. Consistent with the expectation that more distant schools pro-
vide less competition, the estimated impact of the presence of one or more

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

F

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

117

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

charter schools on traditional public school math achievement gains is found
to diminish with the size of the geographic market definition. The presence of
one or more charters within five miles is associated with a 0.3 point increase
in math-score gains, while the presence of a charter school within 10 miles
has no statistically significant impact on math score gains. The results are
virtually the same when controls for the existence of nearby private schools
and other traditional public schools are taken into account. For reading, IL
point estimates of the presence of a nearby charter school on traditional public
school performance are positive, but not significantly different from zero for
any of the three geographic market definitions.

Each additional charter school located within 2.5 miles of a traditional
public school is associated with a 0.4 higher scale-score gain in math, O
3 percent of the average math score gain. These effects diminish with the
breadth of the geographic market definition and are insignificantly different
from zero using either a 5-mile or 10-mile market definition. Similar results
are obtained when the numbers of competing private schools and traditional
public schools are added to the model. When private, charter, and traditional
public school competition is taken into account, there is no statistically signif-
icant correlation between the number of charter schools and traditional public
school reading achievement test scores.

Model estimates using enrollment shares as the index of competition tell
a similar story. Math achievement in traditional public schools is positively
correlated with charter market share for both the 2.5-mile and 10-mile market
definitions, but not the 5-mile definition. For the 2.5-mile market definition,
each 1 percent increase in charter school enrollment share is associated with a
0.08 increase in math score gains. Thus if charters were successful in gaining
a relatively modest 5 percent market share, the model would predict increases
in traditional public school achievement score gains of 0.4 points in math. For
the broader, 10-mile definition, UN 1 percent increase in charter school market
share is associated with a 0.22 increase in traditional public school math gains.
There is some evidence that charter school market share is positively correlated
with traditional public school achievement scores in reading, though only for
the broadest geographic market definition.

Taken as a whole, the results indicate that the introduction of charter
students is correlated with modest improvements in math achievement and
no reduction in reading achievement in traditional public schools. Whether
the small positive net effect on math scores is due to charter competition
or to positive peer effects is not clear. If charters attract better students and
this worsens the pool of peers who remain in traditional public schools, Poi
the results suggest that quality improvement triggered by charter competition
more than offsets any negative peer effects. Alternatively, if charters attract

118

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

a disproportionate share of disruptive or below-average students, then the
math achievement gains in traditional public schools associated with charter
competition may simply be due to peer effects. In either case, the evidence
suggests that the existence of charter schools does not harm students who
remain in traditional public schools and likely produces some net positive
impacts on mathematics achievement.

5. SUMMARY AND CONCLUSIONS

As the charter school sector proliferates and more resources are devoted to
building and expanding charter schools, it is important to determine what
those dollars are buying. This study begins to provide quantitative evidence
on the effects of charter schools on the achievement of students who attend
charters as well as those who choose to remain in traditional public schools.

Consistent with other recent studies, I find that brand-new charters tend
to have lower student achievement than the average traditional public school.
Tuttavia, of much greater importance is the long-run performance of charter
schools. By their fifth year of operation Florida charter schools are found to
reach a par with traditional public schools in math and to produce reading
achievement scores that exceed those of the average traditional public school
by an amount equal to 10 percent of the average annual achievement gain.

Charter schools are quite diverse; some are similar to traditional public
schools while others seek to serve niches by targeting particular types of stu-
dents (per esempio., special education or at-risk students) or emphasizing particular
programs (per esempio., music, art, and languages). They also vary in their manage-
ment structure, with most run as nonprofit entities but a significant number
operated by for-profit management companies. I find that charter schools
that target special education and at-risk students tend to have lower student
achievement in math than nontargeted charters or the average traditional
public school (holding student characteristics constant). The fact that parents
willingly place their children in these schools (and keep them there) suggests
that special education and at-risk charters may provide other valuable services
beyond the core math and reading instruction tested on standardized exams,
such as behavior management, development of social skills or oral commu-
nication skills. Management structure appears to have no impact on student
achievement in charter schools; charters managed by for-profit firms perform
the same as those operated by nonprofit entities.

Competition from charter schools appears to have a modest net positive
impact on student achievement in Florida’s traditional public schools. Whether
measured by the presence of nearby charter schools, the number of competing
charters, or the enrollment share garnered by charter schools, charter school
competition is associated with higher math and unchanged reading scores

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

/

F

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

.

F

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

119

CHARTER SCHOOLS AND STUDENT ACHIEVEMENT IN FLORIDA

in traditional public schools. The positive effects on math achievement are
neither huge nor trivial; they are equivalent to roughly 3 percent of annual
learning gains.

My findings have several important implications for the evaluation of char-
ter schools. Primo, it is clear that there are significant obstacles associated with
establishing a new charter school and that the performance of charter schools
improves over time. Thus the age of charter schools must be taken into account
when one compares their performance to traditional public schools. In Florida,
charter schools operating five years or more are on par with traditional public
schools in math and surpass the average traditional public school in reading
achievement. Secondo, there is considerable diversity among charter schools.
Charter schools that target specific student populations may have objectives
other than simply maximizing scores on achievement tests in core subjects.
Consequently, simply comparing the achievement of students in targeted char-
ters with students in the average traditional public school may not always be
appropriate. Third, it appears that competition from charter schools has a net
positive impact on the performance of traditional public schools in Florida,
though the size of the effect is modest. The charter sector is still rather small,
Tuttavia, and it is not clear how the magnitude of the competitive effects of
charters will change as charter schools attract a larger proportion of the student
population.

I wish to thank the staff of the Florida Department of Education’s Division of Ac-
countability, Research and Measurement and the K-20 Education Data Warehouse,
particularly Jay Pfeiffer, Jeff Sellers, Kathy Peck, Ruth Jones, Murray Cooper, and Barry
McConnell, for their assistance in obtaining and interpreting the data used in this study.
Tuttavia, the views expressed is this article are solely my own and do not necessarily
reflect the opinions of the Florida Department of Education. I also would like to thank
the Spencer Foundation and the DeVoe Moore Center at Florida State University for
their financial assistance. I am grateful to Bob Bifulco, Scott Gilpatric, Doug Harris,
and Craig Newmark as well as seminar participants at Clemson University, Florida
State University, and the University of Florida for helpful comments. Any remaining
errors are solely my responsibility.

REFERENCES
Andrews, Martyn, Thorsten Schank, and Richard Upward. 2004. Practical estimation
methods for linked employer-employee data. Unpublished paper, University of Manch-
ester.

Arellano, Manuel, and Stephen Bond. 1991. Some tests of specification for panel data:
Monte Carlo evidence and an application to employment equations. Review of Economic
Studi 58:277–97.

Bettinger, Eric P. 1999. The effect of charter schools on charter students and public
schools. Unpublished paper, Case Western Reserve University.

120

EDUCATION FINANCE AND POLICY

l

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

P

:
/
/

D
io
R
e
C
T
.

M

io
T
.

/

F

/

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

l

F
/

/

/

/

/

1
1
9
1
1
6
9
2
9
9
3
e
D
P
2
0
0
6
1
1
9
1
P
D

.

.

F

.

.

.

F

B
sì
G
tu
e
S
T

T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Tim R. Sass

Bifulco, Robert, and Helen F. Ladd. 2004. The impacts of charter schools on stu-
dent achievement: Evidence from North Carolina. Unpublished paper, University of
Connecticut.

Boardman, Anthony E., and Richard J. Murnane. 1979. Using panel data to improve
estimates of the determinants of educational achievement. Sociology of Education
52:113–21.

Bonesrønning, Hans. 2004. The determinants of parental effort in education produc-
zione: Do parents respond to changes in class size? Economics of Education Review 23:1–9.