Charters and TPSs - Effects on Student Achievement

Transcription

1 CHARTER SCHOOL OUTCOMES IN CALIFORNIA Ron Zimmer and Richard Buddin The RAND Corporation Prepared for the National Conference on Charter School Research at Vanderbilt Universy on September 29, 2006 Portions of this research were sponsored by the California Legislative Analyst s Office and the Smh Richardson Foundation. The opinions and conclusions expressed in this research are the authors alone and should not be interpreted as representing those of RAND or of the sponsors of this research. This working paper is supported by the National Center on School Choice, which is funded in part by the Department of Education s Instute of Education Sciences (R305A040043). For more information, please vis the National Center on School Choice s webse at

2 INTRODUCTION Background This chapter summarizes a series of studies we have conducted over the last four years that evaluate the effectiveness of California charter schools. California has been an important state in the debate over charter schools. California became the second state to authorize charter schools in More than 210,000 students now attend about 550 charter schools in California the largest charter sector in the nation. Almost 200 new charter schools have opened in California in the past five years. Charter schools are found in most parts of the state from districts in small, rural counties to large, urban districts (Edwards, 2006). What also makes California important to the overall charter debate is the array of charter schools whin the state. In California, some charters are conversions and others are startup charter schools. Conversion schools previously existed as tradional public schools and they typically retain an existing facily as well as faculty and students when they become charter schools. Startup schools, by contrast, are new enties that acquire facilies, faculty, and students at their inception. In addion, the motivation to start charter schools may differ between these types of charter schools wh conversion schools becoming charters to reduce their bureaucracy from the districts and to change specific educational programs while startup schools may be iniated to create a new holistic approach to schools, including curriculum programs, instructional practices, governance structures, and overall mission of the schools. California charter schools also use two distinct instructional approaches most rely exclusively on instruction in tradional classroom settings, but some make extensive use of nonclassroom settings, such as homeschooling, independent study, and distance learning. Charter schools that use tradional classroom instruction are more likely to have similar curricula and operation than a school that provides a significant portion of instruction outside of a classroom.

3 Research Approach This study uses information from California to address four key policy questions about charter schools. Do charter and tradional school operations differ? What types of students do charter schools serve? Do charters improve the performance of charter students? Is charter competion improving the performance of tradional schools? These questions address important issues about the effectiveness of charter schools and whether charters are meeting the objective of charter school proponents or causing the adverse effects anticipated by charter opponents. These analyses relied on student-level longudinal data for charters as well as a survey of California charters and a matched sample of tradional public schools (TPSs). Statewide student-level data. The California Department of Education collected individual-level test scores for about 20 million test records between and The Stanford 9 achievement test was administered in grades 2 through 11 in the spring of each academic year. The accompanying administrative data include information on student ethnicy, English learner status, school lunch eligibily, and student mobily. These data do not include an individual student identifier to track individual student progress from year to year. Nevertheless, the file is used to track gains of grade cohorts while precisely accounting for factors that might influence test scores. District student-level data. We also collected student-level data wh individual identifiers from six districts wh a large share of charter schools (Chula Vista, Fresno, Los Angeles, Napa Valley, San Diego, and West Covina). These data include the same information as the statewide data, but also include individual identifiers allowing for the measurement of the amount of time spent in different schools, student movement from TPSs to charter schools and vice versa, and differences in charter and TPS populations. School survey. School-level information was also collected through a survey of principals in all California charter schools and a demographically matched set of 2

4 tradional schools. 1 The surveys were used to assess operational differences between tradional and charter schools and to examine the oversight of charter schools. In addion, more informal insights into charter school operations were obtained from case studies of nine charter schools in the state. The remainder of the chapter is divided into four sections (one on each question) and a conclusion. Separate statistical methods are used in each section, so the methods will be described in each section of the study. DO CHARTER AND TRADITIONAL SCHOOL OPERATIONS DIFFER? One of the major arguments for charter schools is their abily to be innovative. The premise is that competion to attract students from tradional schools will lead charters to iniate new programs and improve student learning (Kolderie, 2004). Charters are not a specific reform, however, so they may implement a variety of policies and procedures in response to parental concerns about nearby TPSs. In the paper entled Getting Inside the Black Box: Examining how the Operation of Charter Schools Affects Performance, 2 we examine the operational and structural differences between charter schools and TPSs. Table 1 presents operational or structural differences between charter and TPSs. The information is drawn from the survey of school principals in charters and a matched set of TPSs. School features are grouped into five categories: school-organizational features, school-level control, teacher qualy issues, curriculum allocations, and principal background. Because many California charter schools (and some tradional schools) incorporate home schooling as part of their instructional design, we asked principals what portion of their student population is instructed at home. Table 1 shows that home school instruction is much more popular among charter schools. Furthermore, the percentage of students instructed at home increases at higher grade levels about 20 and 26 percent of charter middle and high school students, respectively, receive home school instruction. 1 Specifically, we used propensy scores derived based on school-level demographic characteristics of TPSs and charter schools to create matches. For more information on the approach, see Zimmer et al, (2003). 2 Zimmer and Buddin (2005) 3

5 Table 1 School Characteristics by School Type for Charter and TPS Schools Elementary School Middle School High School Charter TPS Charter TPS Charter TPS School Features Students schooled at home (%) 11.64* * * 6.98 Instructional days per year * * Parental involvement compose 0.27* * * Suspensions in school (%) Suspensions out of school (%) Expulsions from school (%) School focuses on specific student group 0.29* * * 0.12 Disabled Low income Minory Low-English proficiency Low-academic performers * 0.16 Disciplinary risks * 0.11 Gifted Revenue per pupil in $1000 (charters) NA NA NA Facily from district (charters) 0.53 NA 0.42 NA 0.29 NA School-Level Control (4-Point Scale Where 4 = Full Control) Student disciplinary policies 3.48* * * 3.16 Student assessment policies 3.46* * * 3.02 Staff salaries and benefs 2.63* * * 1.45 Other budgetary expenses 3.41* * * 2.82 Curriculum 3.45* * * 2.66 Staff hiring, discipline, and dismissal 3.28* * * 2.95 Teacher Qualy Issues Emphasis on full standard credential * 3.64 College major in teaching field * 3.52 Incentives to offset teacher shortages * 2.13 Professional Development (Annual Participation Quartile) Workshops or conferences * 3.20 Courses for college cred 2.05* * Teacher study groups Mentoring or coaching 2.59* * Curriculum Allocations (5-Point Scale: 1 = 0 Hours/Week,..., 5 = 7 or More Hours/Week) English/language arts 4.41* NA NA Mathematics * 3.60 NA NA Computer skills NA NA Social studies NA NA Sciences NA NA Foreign language 1.76* NA NA Fine or performing arts 2.37* NA NA Required Years of Study for High School Graduation Mathematics NA NA NA NA Computer skills NA NA NA NA 0.74* 0.48 English/language arts NA NA NA NA 4

6 Social studies NA NA NA NA Sciences NA NA NA NA Foreign language NA NA NA NA Fine or performing arts NA NA NA NA Principal s Background (Years) Leadership at current school * 4.75 Administrative experience 8.66* * Teaching experience Notes: An asterisk indicates that the corresponding charter school value varies significantly from the TPS value at the 5 percent significance level. Some schools (especially charters) have grades that span across the classification of elementary, middle, and high schools. These schools are included in the category for each type of school. The curriculum allocations reflect fourth and seventh grade students for elementary and middle, respectively. Another structural feature that may vary between charter and TPSs is the instructional days per year. Many charter school proponents argue that one way to improve learning is through more instructional time. For instance, the educational management organization (EMO) of Edison schools, which includes a large number of charter schools nationwide and in California, provides longer school years (Gill et al., 2005). The data in Table 1 indicates that charter schools in this study have longer instructional days than TPSs at middle and high school levels, but not at the elementary school level. Parental involvement at charter schools is consistently higher at all grade levels in charters than in the matched TPSs. Parental involvement was measured as a compose variable that combines information from several survey ems. Principals were asked to characterize parent participation in their school on a five-point scale ranging from few to most for the following ems: Open hours or back-to-school night Regularly scheduled schoolwide parent-teacher conferences Special subject-area events (e.g., science fair, concert) Parent education workshops or courses Wrten contract between school and parent Parents involved in instruction issues Parents involved in governance Parents involved in budget decisions 5

7 The scaled variable was based on the standardized values for each em averaged across all times. The alpha reliabily for the scaled variable was 0.74 wh average interem correlation of School disciplinary practices differ ltle between charter and the matched TPSs. The survey results show that school suspension (eher in or out of school) or expulsion rates are about the same for both types of schools at the elementary, middle, and high school levels. Charters are much more likely than TPSs to focus on particular student groups. TPSs generally serve a local neighborhood and their composion largely reflects the characteristics of students living in the nearby area. In contrast, charters have more flexibily to draw students from a wider area and concentrate on particular groups of students. The extent of charter specialization ranges from 29 percent for elementary students to 42 percent at the high school level. Only 18 percent of TPS elementary schools focus on a specific student group and this falls to 12 percent for high schools. Charters are more likely than TPSs to focus on student groups, but the particular focus varies ltle between the two groups of schools. The key difference between charters and TPS is at the high school level, where charters are much more likely to focus on low performer or disciplinary risks than are TPSs. We restricted our survey questions of revenue per pupil only to charter schools because we found that many TPS principals were not fully aware of the schools finances as they are generally managed by the district. In addion, we did not ask TPSs whether the district provided a facily to the school or not because all TPSs are provided facilies. However, we did look at both issues across different types of charter schools and found that revenue per pupil increases as grade levels increase (which is consistent wh state funding allocations). In contrast, school districts are less likely to provide facilies to schools wh higher grade arrangements wh 53 percent of elementary schools receiving facilies from a district as compared wh 42 and 29 percent for middle and high schools, respectively. One of the strongest arguments put forth by charter advocates for these schools is the abily to make autonomous decisions to meet the needs of their students and allow 6

8 for greater innovation (Kolderie, 2004; Finn et al., 2000; Nathan 1996). However, in cases when the charter arrangement is formed between a district and a school, the district may restrict certain liberties. Therefore, we also asked both charter and TPS principals about the level of control over certain operational decisions whin their schools. In the survey, principals responded to a four-point scale, where 1 represented no control and 4 represented full control. Table 1 indicates that charter school principals across all levels reported a greater level of control than TPS principals on all the dimensions listed. These data show that charter school principals do indeed have a higher level of autonomy in their schools. One of the ways in which charter schools could use this greater freedom is in the hiring and professional development of teachers. 3 To examine these issues, we asked principals how much emphasis (as measured in a four point scale that ranged from not important to very important) they put on hiring credentialed teachers and teachers who majored in the field in which they teach. We also asked what percentage of teachers (through a four quartile response, wh 1 indicating 0 to 25 percent and 4 indicating a 75 to 100 percent) participated in different professional development activies, including workshops or conferences, courses for college cred, teacher study groups, and mentoring and coaching. In terms of hiring practices, elementary and middle charter schools and TPSs placed a similar emphasis on hiring credentialed teachers as well as teachers who majored in the field in which they will teach. However, at the high school level, charter schools placed less of an emphasis on both. For professional development, charter schools at all levels placed a greater emphasis on mentoring and coaching than tradional schools (the differences are significant for elementary and middle schools, but not for high schools). The rest of the results for professional development are more mixed wh charter elementary and middle school emphasizing college courses more, while charter high schools emphasized workshops or conferences more. 3 Although evidence from the lerature is mixed on whether credentialing matters in relation to student achievement (e.g., Goldhaber and Brewer, 2000, 2001), policians and educators across the country often advocate increasing the proportion of credentialed teachers whin the classroom, especially in low-income schools (e.g., Darling-Hammond et al., 2001). 7

9 Another way in which charter schools could use their autonomy is through the amount of instructional time spent on curriculum subjects. We asked both charter and TPS principals how much time they spent on various subjects (through a five-point scale in which 1 represented zero hours and 5 represented seven or more hours per week?). 4 Both charter and TPSs reported similar number of hours spent on core subjects, wh only a few statistically significant differences. Charter elementary schools reported spending a ltle less time on English/language arts, and charter middle schools reported spending a ltle more time on mathematics. However, charter schools reported spending significantly more time on foreign language and fine or performing arts at the elementary level. The survey results show almost no differences between charters and TPSs in the required years of high school study in particular subjects. Required years do not differ significantly in English/language arts, mathematics, social studies, sciences, foreign language, and fine or performing arts. Charter schools do have a higher requirement in computer skills than do TPSs. We also examined the principal s background in both charter and TPSs. The lerature has noted the importance of principal leadership (Gates et al., 2003) to student achievement. Leadership skills may be even more important for charter schools when the principals are not only the instructional leader but also manage the operation and finances of the school, in many cases from scratch. Table 1 shows that principals at charter schools have significantly less experience at their current school across all levels (the effect is significantly different only for high schools), less general administrative experience for elementary and middle charter schools, and teaching experience similar to their TPS counterparts at all levels. Finally, we should note that in a separate paper entled Charter School Type Matters When Examining Funding and Facilies: Evidence From California, 5 we found 4 Elementary and middle school principals were asked about the time allocations of students in the forth and seventh grades, respectively. High school principals were not asked about curriculum allocations across subjects, but they were asked about the years of study in particular subjects that were required for graduation. 5 Krop and Zimmer (2005). 8

10 that charter schools, especially startup charter schools, tend to participate less in categorical programs, which leads to less revenue and expendures relative to charter schools. Overall, the analysis shows differences in the operation and structure of charter schools and TPSs including operational autonomy, professional development, home schooling, parental involvement, and curriculum emphasis at the elementary school level. Later, in our student achievement analysis, we examine how variations in these operational and structural features affect performance by combining the survey data wh the student achievement data. WHAT TYPES OF STUDENTS DO CHARTER SCHOOLS SERVE? Because charter schools are schools of choice, is important to examine whether they are serving the full range of the student population and whether they are doing so in integrated settings. Charter school crics argue that charter success might be illusory if charter schools are simply recruing the best students from TPSs or if they further stratify an already ethnically or racially 6 stratified system (Cobb and Glass, 1999; Wells et al., 1998). In general, these crics fear that charter schools may not only have negative consequences for the charter students who attend these schools, but if charter schools skim off high achieving students, they may also have adverse social and academic effects for students who remain in TPSs. 7 However, proponents of charter schools argue that charter schools will improve racial integration by letting families choose schools outside of neighborhoods where housing is racially segregated (Finn, 2000; Nathan, 1996). Previous studies have examined the racial/ethnic makeup of charter-schools relative to the average racial/ethnic makeup of their surrounding districts and states 6 For simplicy, we will refer to race/ethnicy as race throughout the rest of the paper. 7 The interaction of students wh diverse backgrounds and abily levels can have posive social and academic effects for students (Frankenberg and Lee, 2003; Zimmer, 2003; Zimmer & Toma, 2000; Summers & Wolfe, 1977; and Henderson et al., 1978). If charter schools affect the distribution of students both by race/ethnicy and abily, then they may have implications not only for charter students, but also for students who remain in tradional public schools. Many fear that charter schools will further racially/ethnically stratify an already deeply stratified system and will skim off the best students from tradional public schools, reducing the peer effects whin the tradional schools. 9

11 (Powell et al.,1997; RPP International, 2000; Miron and Nelson, 2002; Frankenberg and Lee, 2003). While this research has provided some insights, does have some weaknesses. First, has exclusively examined race/ethnicy, ignoring altogether the issue of stratification by abily. 8 Second, has used cross-sectional snapshots of schools enrollments, which does not perm examination of the movement of students between schools. Understanding how charter schools affect the mixing of students requires a dynamic model that uses longudinal data to examine the movements of individual students. One exception is a study by Bifulco and Ladd (2006b) in which they examine migration patterns of students of different race/ethnicy as they choose to go to charter schools in North Carolina. They find that black students are more likely to go to charter schools wh higher concentration of black students than their exing school. These results highlight the importance of using student-level data. Like Bifulco and Ladd, we use student-level data from the six districts in which we can link students over time to examine the racial/ethnic makeup of students most likely to enroll at a charter school and to examine whether charter schools are cream skimming the best students or are actually attracting the lowest-performing students, which is presented in a paper entled "The Effect of Charter Schools on School Peer Composion." 9 The longudinal nature of the data is used to track student movements from school to school. This data feature allows us to examine the characteristics of students that migrate from a TPS to a charter and compare the students characteristics wh the distribution of students at the old and new schools. Race/Ethnicy Sorting Do charter students move to more or less racially/ethnically integrated school? Table 2 examines the peer environment in both schools for charter transfers. Column two of the table shows the overall results for all students, and columns three through six show detailed results by each racial/ethnic group. Students are moving to charter schools wh a 8 Sass (2006) does find that students enrolled in gifted programs are less likely to swch to charter schools, but does not consider more general measures like the level of prior test scores. 9 Booker, Zimmer, and Buddin (2005) 10

12 higher percentage of black students and a lower percentage of Hispanic students than the TPSs that they previously attended. Table 2: Comparing Tradional Public and Charter Peer Environments for Charter Movers by Race/Ethnic Background of Student Total Whe Black Hispanic Asian Number of students 14,210 1,834 5,342 5, Whes at tradional (%) Whes at charter (%) Difference * * 9.2* Blacks at tradional (%) Blacks at charter (%) Difference 13.6* 5.6* 11.7* 10.6* 2.6 Hispanics at tradional (%) Hispanics at charter (%) Difference -11.9* 4.2* -21.0* -12.0* -5.1* Asians at tradional (%) Asians at charter (%) Difference -1.8* -2.6* Herfindahl at tradional Herfindahl at charter Difference * 0.071* * * * Difference is significant at 5% level The effects of charters on racial/ethnic diversy are also examined using a Herfindahl index for each school campus. The index is defined as the sum of the squared shares of each racial/ethnic group and is scaled from 0 to 1, where 1 indicates total concentration of enrollment in one racial/ethnic group. 10 The higher the index, therefore, the less diverse the school. Students are transferring to charters slightly more diverse than the tradional schools they are leaving. The results in Table 2 show some interesting movement patterns for students in different race/ethnic groups. The typical black transfer moves from a TPSs that is 39 percent black to a charter that is 51 percent black. The Herfindahl measure also shows 10 A number of recent studies have used the Herfindahl index to measure competion from private schools or among TPSs, including Zanzig (1997), Borland and Howsen (1992, 1993, 1996), Hoxby (1994). 11

13 that black students are transferring to charters that are less diverse than the tradional schools that they left. In contrast wh black students, whe, Hispanic, and Asian students in California tend to go to charter schools that have a lower percent of students of the same race/ethnicy and that are more diverse than their TPSs. Overall, appears that black students are choosing charter schools that are less diverse and more like them. This result is consistent wh patterns of charter school transfers in North Carolina (Bilfulco and Ladd, 2006b). However, whe students are shifting to schools that have a lower percentage of whes than their TPSs. Hispanic students move to charter schools wh a lower percentage of Hispanics than the TPSs they leave. Abily Sorting An analysis of abily sorting for charter school movers is shown in Table 3. Individual test scores of charter students in the year before their move are shown along wh the means of student peers in each student s tradional school. A comparison of the test scores of movers and those of peers left behind shows whether charter schools are creaming the best tradional school students or providing options for students who have not succeeded in TPSs. Table 3: Comparing Average Math and Reading Scores of Charter Movers to Other Students at the TPSs That They Leave Total Number of students 13,863 Math score of movers before move Math score of tradional peers Difference wh tradional peers * Reading score of movers before move Reading score of tradional peers Difference wh tradional peers * * Difference is significant at 5% level The results provide ltle evidence of abily sorting into charter schools. Students moving to charter schools have slightly lower test scores than their tradional school 12

14 peers (0.09 and 0.04 standard deviations in math and reading, respectively), which leads us to conclude that California charters are not creaming the best students from tradional schools, but they are also not providing an outlet for the low-achieving students. DO CHARTERS IMPROVE THE PERFORMANCE OF CHARTER STUDENTS? A recent consensus panel on charter schools stressed the importance of analysis that relies on student-level data and a value-added approach to track student progress over time (Betts and Hill, 2006). The panel stressed the biases inherent in school-level comparisons between charter and tradional schools. A key weakness of a school-level analysis is the high degree of aggregation, which masks changes over time in the school s population of students, and variation of performance across different subjects and grades. In essence, school-level data may not pick up the nuances of school characteristics and can only provide an incomplete picture of why outcomes vary across schools. Similarly, cross-sectional comparisons of individual student scores in charter and tradional schools are likely to be misleading, because unmeasured student characteristics (perhaps parental involvement or motivation) may differ across school types and distort the true contribution of charter schools to student learning. The past four years have seen several key studies that use longudinal data to isolate charter effects. Hanushek, Kain, and Rivkin (2002) found negative effects of Texas charters on student achievement during the inial years of charter operation, but no significant difference in achievement between charter and tradional schools after about three years of operation. Buddin and Zimmer (2003) found that student achievement in California charters was keeping pace wh tradional schools. We also found substantial differences in performance across charter school types wh startups out performing conversions, and charters wh nonclassroom-based instruction doing poorly. Booker et al. (2004a) adjusted for charter-specific mobily effects in Texas and found that charter students do eventually outperform their tradional school counterparts. Bilfulco and Ladd (2006a) found that charter students in North Carolina made smaller achievement gains over time than those students would have made in tradional schools. Sass (2006) found that charter students inially had lower performance than tradional public school 13

15 students, but charter schools produced similar gains to tradional schools as the schools matured. 11 Our analysis has relied upon a two pronged approach to assess charter performance in California. The first approach relies on statewide student-level data to examine the patterns in student achievement in charter and tradional schools and especially to isolate the different effects of alternative charter types (startup and conversion, classroom- and nonclassroom-based instruction). A subset of the statewide data is also used to examine how school operations affect achievement in charters and a matched sample of tradional schools. The statewide data provides the broadest picture of how charters are doing across the state and does include controls for student background and demographics at each school and over time. A limation of the statewide data, however, is that is does not include a student-level identifier that would allow the tracking of student progress from year to year California is planning a statewide student identifier, but does not currently have one. The second approach relies on the district student-level data that does include student-level identifiers. Longudinal tracking of student progress minimizes the problem of selection bias by examining the academic gains made by individual students over time, factoring out students baseline achievement levels. Moreover, they perm whin-student comparisons of achievement gains, examining changes in the achievement trajectories of individual students who move from tradional public schools to charter schools, or vice versa. 11 Hoxby and Rockoff (2004) also examined four charter schools in Chicago, which provided some evidence that charter students outperform non-charter students. Their analysis capalized on the fact that these schools are oversubscribed and used a lottery mechanism to adm students. Presumably the lottery winners and losers are similar in every way except admission into these schools. Tracking performance of both sets of students then creates an unbiased perspective of performance. However, Hoxby and Rockoff s study has a major drawback in that may have limed implications for those schools that do not have wa lists. In fact, you would expect schools wh wa lists to be the best schools and would be surprising if they had the same results as other charter schools. 14

16 Statewide Student-Level Data Charter Type and Academic Achievement The statistical model of student achievement in the paper entled A Closer Look at Charter School Student Achievement 12 is based on a multilevel approach where random effects are estimated for each school and each grade cohort whin each school. The school-level random effect allows for unobserved heterogeney across schools i.e., the learning environment may vary systematically at different schools in ways that have a common achievement effect on the students that attend those schools. The grade-cohort effect is designed to capture the possibily that some groups of students may have a persistent achievement score effect. For example, if third grade students at a particular school score high in math in one year (relative to past cohorts wh similar student characteristics), then the next year s fourth grade students at the same school are also likely to excel in math because of unmeasured attributes of students in the grade cohort. In addion to the random components, the model also adjusts for the individual characteristics of each student taking the test. The key student characteristics available are limed English proficiency, race/ethnicy (whe non-hispanic, black, Asian, and Hispanic), gender, and eligibily for the free or reduced-price school lunch program (a common measure of socioeconomic status). The California test data also include an indication of the parental education for each student. Finally, the data indicate whether students are in their first year at their current school. Students (and their parents) may have difficulty in making the transion to friends and teachers at a new school, so their test scores in the new school may be lower than those of similar students who remain at the same school (Hanushek, Kain, & Rivkin, 2002). The formal model describes the relationship between student-level test scores, student and school characteristics, a random school effect, and a random grade-cohort effect whin each school. The model includes dummy variables to describe any possible time trend in test scores over the five years from 1998 through The dependent 12 Buddin and Zimmer (2005a) 15

17 variable is the student percentile test score, T ijkt, observed for student i in school j and grade cohort k at time t. 13 The formal model is T ijkt = x β + w jt γ + θ j + ψ j ( k ) + ε ijkt (1) where x and w jt are vectors of measured time-varying student and school characteristics, respectively; β and γ are parameter vectors for student and school effects; and school and grade-cohort time heterogeney are represented by θ j and ψ j(k), respectively. The last component of the model is a random error term, ε ijkt, which is orthogonal to all other effects in the model. The x vector includes elements for school type. In particular, we estimated three types of charter school models: 1) A basic model that simply compares test scores between charter and TPSs. 2) A model that separates the performance of conversion and startup charter schools. 3) A more general model that examines how classroom- and nonclassroom-based instruction affects achievement in conversions and startups. 14 Later in this section, we will expand the set of school variables to include survey information on school operations in an analysis of the charters and matched TPSs. The regression results in Table 4 show that the charter effects for elementary students differ somewhat across alternative types of charter schools. 15 Model 1 results indicate that charter students are performing the same in reading and 1.44 percentile points lower in math than are comparable students in tradional schools. Model 2 shows that conversion charter schools have a small, posive effect on reading and a small, negative effect in math compared to tradional public schools. In contrast, startup 13 Assessments of standardized achievement tests are often subject to score inflation (Hamilton, 2003). Researchers find that scores in a state or school tend to rise over time whout any commensurate increase in general learning or proficiency. This is often attributed to teaching to the test. 14 We should note that not all students that attend a nonclassroom-based charter school receive nonclassroom-based instruction because many of these schools are actually hybrid schools that include classroom and nonclassroom instruction. In addion, our data do not distinguish whether a particular student uses any form of nonclassroom instruction, but we do know which charter schools offer these programs. Our survey suggests that about 63 percent of students in a school wh nonclassroom-based instruction receive at least some instruction away from the formal school se. 15 Buddin and Zimmer (2005) include more specifications of this model including a comparison of charter effects for new and established charter schools. Tables 4 and 5 summarize regressions that also control for student demographics. The full specifications are also reported in Buddin and Zimmer(2005). 16

18 students score about 5 to 7 percentage points lower than do similar students in tradional schools in reading and math, respectively. The results from Model 3 show the importance of separating charter schools based on whether they provide classroom- or nonclassroom-based instruction. Students in nonclassroom-based conversion and startup schools have much lower test score than do comparable students in tradional public schools. Classroom-based charters are keeping pace wh tradional public schools the test scores in classroom-based conversions are one percentile point higher in reading half a point lower in math, while startup scores do not differ significantly from those in tradional schools. Table 4 Stanford 9 Test Regressions for Elementary Students by Charter School Type Reading Math Variable Coefficient Standard Error Coefficient Standard Error Model 1: Comparison of Charter and Tradional Schools Charter * R-Square Model 2: Comparison of Conversion and Startup Schools Charter Type Conversion Charter * * Startup Charter * * R-Square Model 3: Comparison of Classroom- and Nonclassroom-Based Charter Schools Charter Type: Conversion and Classroom * * Conversion and Nonclassroom * * Startup and Classroom Startup and Nonclassroom * * R-Square Notes: Standard errors are in parentheses. Entries wh asterisks are associated wh coefficients that are significantly different from zero at the α=0.05 confidence level. The r-square statistic shows the reduction in the variance components for this model relative to an uncondional mean model that only adjusts for the two random effects (Bryk and Raudenbush, 2002). Sample size is 9,114,624 student records. The secondary school results in Table 5 show that charter students overall are scoring 1.46 and 2.26 percentile points lower than similar students in tradional schools (see Model 1). The results in Model 2, however, show that startup school students have 17

19 lower reading and math scores than do comparable students in TPSs. The conversion coefficients are insignificant for reading and slightly negative for math. Finally, the Model 3 results in Table 5 show that nonclassroom-based schools are pulling down the average test scores for both conversion and startup schools. Nonclassroom-based charter schools are performing poorly compared wh tradional public schools. Classroom-based conversion schools have mixed results the reading score is no different than in tradional public schools, but the math score is slightly lower. The test scores for classroom-based startups are higher than those of tradional public schools, after adjusting for the mix of students attending these schools. Table 5 Stanford 9 Test Regressions for Secondary Students by Charter School Type Reading Standard Math Standard Error Variable Coefficient Error Coefficient Model 1: Comparison of Charter and Tradional Schools Charter * * R-Square Model 2: Comparison of Conversion and Startup Charter Schools Charter Type: Conversion Charter * Startup Charter * * R-Square Model 3: Comparison of Classroom-Based and Nonclassroom-Based Charter Schools Charter Type: Conversion and Classroom * Conversion and Nonclassroom * Startup and Classroom * * Startup and Nonclassroom * * R-Square Notes: Standard errors are in parentheses. Entries wh asterisks are associated wh coefficients that are significantly different from zero at the α=0.05 confidence level. The r-square statistic shows the reduction in the variance components for this model relative to an uncondional mean model that only adjusts for the two random effects (Bryk and Raudenbush, 2002). Sample size is 12,647,295 student records. The evidence shows important differences in test score performance across different types of charter schools, but the underlying reasons for these differences are not clear whout greater analysis of the operation of these schools. In addion, the schools may differ from one another in the types of students that they attract. For instance, 18

20 TPSs. 16 The school characteristics are drawn from the survey data described previously, nonclassroom-based students may be different in unique ways from students in tradional public schools that are not captured by the demographic factors in the analysis. More explicly, if nonclassroom-based students have been pulled out of tradional public schools because of problems in tradional settings, then tradional students who do not have these problems do not make a good comparison group. Wh longudinally-linked student-level data, an analysis would be better able to control for these unobservable differences. Nevertheless, the differences in performance among charter schools in our analysis are compelling and underscore the importance of considering these differences when interpreting charter school results. In the next subsection, we will use survey data to explore whether we can identify factors that lead to variations in performance and later we use longudinally-linked student-level data to control for student unobservable differences. Charter Operations and Student Achievement In the previous section, we identified that achievement of charter schools can vary by charter type. Here, we summarize results from a paper entled "Getting Inside the Black Box: Examining how the Operation of Charter Schools Affects Performance" that used the survey data and the statewide achievement data to see if there are certain factors that lead to higher or lower student achievement in charter schools. The analysis includes 257 charters in the state that responded to the survey along wh a matched sample of 184 which included five categorizations of the operational differences: school features, school-level control, teacher qualy issues, curriculum allocations, and principal background. The multilevel model in Equation 1 is run separately for several specifications, so model parameters are not inherently constrained across dissimilar groups of students or schools. 16 Our response rate for the surveys was 73 percent for charter schools and 75 percent for TPSs. 19

21 Reading and mathematics scores. Separate models are estimated for reading and mathematics test scores because student and school characteristics may have different effects on the learning environment in each area. Tradional and charter schools. Models are estimated for tradional and charters schools separately to assess how differences in school operations among each type of school affects achievement. In addion, a pooled regression shows whether differences in operations between tradional and charter schools affect student achievement. We also run the analysis separately for elementary, middle, and high schools, but because of space limations and because our sample size is largest for elementary schools, we display the results only for elementary schools. 17 Table 6 shows the reading and mathematics score regressions for tradional and charter schools as well as for both types of schools pooled together. The main focus of this analysis is on the effects of school operations variables on student achievement, and the regression controls for student characteristics are designed largely to isolate the effects of the operations measures. As expected, limed English proficiency, minory status, low socioeconomic status (eligibily for free or reduced-price school lunch or low parental education levels) are inversely related to reading and mathematics test scores in elementary, middle, and high schools. Other things being equal, girls generally score higher than boys especially on the reading test. New students at a school consistently score lower on both tests than do students who are continuing in the same school. The overall trend in test scores is upward an indication of overall improvement in students or schools or perhaps evidence of schools teaching to the test. Student characteristics have similar qualative effects on test scores in both tradional public and charter schools. The results indicate that home schooling is inversely related to reading and math test scores, which is consistent wh our results for nonclassroom-based schools. While charters have more home-schooled students, the magnude of this effect is similar in tradional public and charter schools. There is no clear evidence on whether the lower 17 Zimmer and Buddin (2005b) show the full results for middle and high school students. 20

22 test scores for home school students reflect poorer learning opportunies or whether these students differ in some unmeasured way from students who receive classroom instruction. The results show that the number of instructional days has no effect on achievement in reading or math. However, parental involvement has a substantial posive effect on student test scores across both charters and TPSs, which is consistent wh the lerature (Jeynes 2003, 2005; Fan and Chen 2001). The results show that a one standard deviation change in school-level parental involvement is linked wh a 4- or 5- percentile increase in student achievement in both reading and math. Differences in school disciplinary practices and a focus on particular student groups have no significant effect on student achievement. Disabled students score significantly lower in reading when both charters and TPSs are pooled, but the separate specifications for charters and TPSs do not show a significant effect for schools wh a focus on disabled students. The school resource variables for charter schools do not have a significant effect on student achievement. Achievement scores do not vary wh eher revenue per pupil or whether the school received their facily at ltle or no cost from the district. This counters some of the charter advocates who have argued that charter schools have not achieved all they could because of poor resources and lack of facilies. Controlling for differences in operational measures and students, the overall scores of charter school students are about 1 percentile point lower in math than those for TPS students. In reading, the differences in scores by school type are not statistically significant. The results also indicate that greater school-level autonomy in charter schools has no significant effect on student achievement scores. Table 1 showed that charter school principals had more control over discipline, assessment, salaries, expenses, curriculum, and staff management than did TPS principals. After controlling for other student and school characteristics, the regression results in Table 6 show that school autonomy does not translate into better academic success in the classroom. Greater autonomy may have other benefs for teachers, parents, and students, but the evidence does not show that autonomy translates into test score improvement. 21

23 A school s emphasis on teacher qualy issues is largely unrelated to test scores. An emphasis on teachers having a full teaching credential is posively related to reading and mathematics scores for combined charter and TPS specifications, but this factor is not significant in explaining achievement for eher group separately. Scores do not vary significantly based on whether the school focuses on hiring teachers wh a major in their teaching field 18 or wh whether the school offers incentives to offset teacher shortages. Professional development opportunies in both tradional public and charter schools are largely unrelated to the achievement scores of students in those schools. Different curriculum allocations at elementary schools have ltle effect on reading and math scores. More time on language arts is associated wh better reading and math scores when charters and TPSs are combined, but the effect is insignificant for the separate school groups. Time commments to math, computer skills, social sciences, sciences, and fine or performing arts do not significantly affect reading or math scores in eher tradional public or charter schools. Foreign language instruction may be crowding out other learning, however. Schools wh more foreign language instruction have consistently lower achievement scores in both reading and mathematics. 19 Principal experience has ltle impact on student achievement. The results show that years of leadership at the school, administrative experience, and teaching experience have almost no significant effect on student achievement at the school. However, the principal experience measures may not capture the leadership skills, enthusiasm, and creativy that may contribute to school success. Table 6: Multilevel Regression Results for Charter and Tradional Public Elementary Schools Reading Mathematics Charter TPS Both Charter TPS Both Student Characteristics Limed English Proficiency * * * -7.22* -7.48* -7.41* (0.21) (0.15) (0.13) (0.22) (0.16) (0.13) Black * -9.22* -9.83* * * * (0.29) (0.19) (0.16) (0.29) (0.20) (0.17) Asian 2.72* 2.11* 2.33* 7.85* 7.04* 7.32* 18 Teaching in field is probably more relevant for middle and high school teachers. In these higher grades, schools may try to recru teachers wh math and science majors for instruction in these subjects. At the elementary level, teachers generally have a multi-subject credential, teach a variety of subjects in their classrooms, and have a college minor in elementary school education. 19 The model specification includes a control for whether individual students are LEP, so the foreign language training here represents addional training outside of English. 22

24 (0.40) (0.25) (0.21) (0.41) (0.26) (0.22) Hispanic -5.15* -3.74* -4.26* -4.04* -3.42* -3.66* (0.22) (0.15) (0.12) (0.22) (0.16) (0.13) Non-high school graduate -2.67* -2.45* -2.50* -2.21* -2.06* -2.13* (0.26) (0.20) (0.16) (0.27) (0.21) (0.16) Some college 4.19* 4.88* 4.60* 4.11* 4.47* 4.32* (0.24) (0.17) (0.14) (0.24) (0.18) (0.14) College graduate 8.81* 7.50* 8.01* 8.60* 7.23* 7.75* (0.25) (0.18) (0.15) (0.26) (0.19) (0.15) Some graduate school 14.00* 11.67* 12.58* 13.31* 11.12* 11.97* (0.30) (0.22) (0.18) (0.31) (0.23) (0.19) Free/reduced school lunch -5.06* -4.69* -4.83* -4.37* -4.88* -4.66* (0.21) (0.17) (0.13) (0.22) (0.18) (0.14) Female 2.71* 3.03* 2.91* * 0.14 (0.13) (0.09) (0.08) (0.13) (0.10) (0.08) New school this year -2.54* -3.28* -2.98* -2.86* -3.93* -3.49* (0.17) (0.13) (0.10) (0.17) (0.13) (0.11) Year * 2.90* 2.96* 3.49* 2.77* 2.94* (0.32) (0.17) (0.15) (0.32) (0.17) (0.15) Year * 3.91* 4.39* 5.29* 4.28* 4.42* (0.41) (0.25) (0.21) (0.42) (0.26) (0.22) Year * 3.82* 4.35* 5.07* 3.65* 3.92* (0.42) (0.26) (0.22) (0.43) (0.28) (0.23) Year * 2.80* 3.22* 3.21* 2.45* 2.45* (0.43) (0.28) (0.23) (0.44) (0.29) (0.24) School Features Students schooled at home (%) -0.07* * -0.10* -0.11* -0.11* (0.02) (0.03) (0.02) (0.02) (0.03) (0.02) Instructional days per year (0.08) (0.16) (0.05) (0.09) (0.17) (0.06) Parental involvement compose 4.97* 3.87* 4.05* 5.52* 4.79* 4.57* (1.22) (1.25) (0.78) (1.35) (1.40) (0.88) Suspensions in school (%) (0.18) (0.09) (0.07) (0.19) (0.10) (0.08) Suspensions out of school (%) (0.24) (0.09) (0.08) (0.26) (0.10) (0.09) Expulsions from school (%) * (1.69) (2.96) (1.29) (1.87) (3.30) (1.45) School focuses on specific students Disabled * (3.79) (2.78) (2.19) (4.21) (3.12) (2.48) Low income (3.38) (3.35) (2.13) (3.75) (3.76) (2.41) Minory (3.20) (3.59) (2.20) (3.56) (4.03) (2.49) Low-English proficiency (3.52) (3.63) (2.28) (3.90) (4.08) (2.58) Low-academic performers (4.22) (5.45) (2.84) (4.68) (6.12) (3.22) Disciplinary risks (4.29) (5.02) (2.81) (4.76) (5.64) (3.18) Gifted (2.81) (2.41) (1.72) (3.13) (2.70) (1.94) Facily from district (charters) (1.52) (1.68) Revenue per pupil in $1000 (charters)

25 (0.33) (0.37) Charter School * (0.30) (0.31) School-Level Control (4-Point Scale Where 4 = Full Control) Student disciplinary policies * 0.67 (1.12) (1.13) (0.77) (1.25) (1.27) (0.87) Student assessment policies (1.31) (0.87) (0.68) (1.46) (0.98) (0.77) Staff salaries and benefs * (0.72) (0.74) (0.45) (0.80) (0.83) (0.51) Other budgetary expenses (1.17) (1.00) (0.70) (1.30) (1.12) (0.79) Curriculum (1.24) (1.13) (0.73) (1.38) (1.27) (0.83) Staff hiring, discipline, and dismissal (1.04) (1.00) (0.68) (1.15) (1.12) (0.77) Teacher Qualy Issues Emphasis on full standard credential * * (1.10) (1.15) (0.72) (1.22) (1.29) (0.82) College major in teaching field (0.82) (0.75) (0.50) (0.91) (0.84) (0.57) Incentives to offset teacher shortages * (0.96) (0.87) (0.59) (1.07) (0.97) (0.67) Professional Development (Annual Participation Quartile) Workshops or conferences * (0.88) (0.79) (0.54) (0.97) (0.89) (0.60) Courses for college cred * (0.82) (0.74) (0.51) (0.90) (0.83) (0.57) Teacher study groups (0.62) (0.58) (0.39) (0.69) (0.65) (0.45) Mentoring or coaching (0.60) (0.65) (0.42) (0.67) (0.73) (0.47) Curriculum Allocations (5-Point Scale: 1 = 0 Hours/Week,..., 5 = 7 or More Hours/Week) English/language arts * * (1.29) (1.21) (0.86) (1.44) (1.36) (0.97) Mathematics (1.65) (1.43) (1.01) (1.83) (1.61) (1.14) Computer skills (0.92) (1.25) (0.66) (1.02) (1.40) (0.74) Social studies * (1.35) (1.34) (0.92) (1.50) (1.50) (1.04) Sciences (1.27) (1.34) (0.85) (1.41) (1.50) (0.96) Foreign language -2.29* -1.89* -1.81* -2.38* -1.59* -1.75* (0.68) (0.72) (0.52) (0.76) (0.81) (0.59) Fine or performing arts * (0.97) (1.05) (0.65) (1.08) (1.18) (0.74) Principal s Background (Years) Leadership at current school (0.23) (0.15) (0.12) (0.26) (0.17) (0.14) Administrative experience * (0.10) (0.10) (0.07) (0.11) (0.12) (0.07) Teaching experience (0.08) (0.09) (0.05) (0.09) (0.10) (0.06) Constant (18.05) (27.52) (11.03) (19.97) (30.94) (12.44) 24

26 Random Effects (St. Dev.) School 6.14* 5.54* 5.77* 6.90* 6.22* 6.55* (0.53) (0.48) (0.34) (0.57) (0.54) (0.37) Grade cohort whin school 4.76* 3.85* 4.19* 4.95* 4.49* 4.79* (0.19) (0.14) (0.12) (0.18) (0.15) (0.12) Residual 20.01* 18.68* 19.22* 20.49* 19.54* 19.93* (0.05) (0.03) (0.03) (0.05) (0.03) (0.03) R-squared Number of Observations Notes: An asterisk indicates that the corresponding regression coefficient is significantly different from zero at the 5 percent confidence interval. Standard errors are reported in parentheses. The r- squared statistic shows the reduction in the variance components for this model relative to an uncondional mean model that only adjusts for the two random effects (Bryk and Raudenbush, 2002). So, overall, while we did find some differences in the operation of structure of charter schools relative to TPSs in Table 1, we did not find that these differences led to much of a difference in the student achievement levels of schools. District Student-Level Data Statistical Model The statistical model in Equation 1 is modified to capalize on the linkage of individual student records from year to year. Equation (2) describes a basic approach for estimating a charter school effect: T = µ + γc + x β + υ (2) i where i and t index individual students and years, respectively; T is test score; µ is an unobserved student-specific factor that does not vary over time; γ is an unobserved parameter reflecting the possible effect of charter school attendance on T; C is an indicator variable that equals one if the school is a charter school and zero otherwise, x is a 1 K vector of K observable factors affecting s, β is a K 1 vector of unobserved parameters, and ν is a random error term. The model includes observed family background characteristics like race/ethnicy and other demographics that are likely to affect student achievement. Two common approaches to estimating a school-level effect over time are a random-effect or a fixed-effect model. The most appropriate approach depends upon the correlation between µ and the observed factors (C and x). A random-effects model 25

27 assumes that unobserved permanent factors affecting student achievement (µ) are uncorrelated wh observed factors (C and x). This type of model would seem appropriate if the vector of student characteristics contains a relatively complete set of observed factors affecting student achievement, including measures of previous academic progress prior to the year of the test. Alternatively, the fixed-effect model uses the longudinal nature of the data to difference out the µ for observations on the same individual. In our analysis, we estimated both a random effect and fixed model and compared the parameter estimates from each model wh a Hausman test, which examines the correlation between the error term and the regressors. The Hausman test showed that the estimated student-specific error term was significantly correlated wh student background variables in the model. Therefore, the parameter estimates from the randomeffects model are inconsistent due to an omted variable bias. This violation of the random-effects assumptions suggested that a fixed-effect model is the more appropriate approach for estimating the charter effect. To estimate the charter effect through a fixed-effects approach, we average the variables for the i th individual student and subtract this result from equation (2), so the transformed fixed-effects equation is s s = γ(c C ) + (x - x )β + (υ υ ) (3) i i where the bar above each variable is the corresponding variable mean. The student fixed effect combines all student-level factors that are invariant over time and affect student achievement, so the results do not include separate parameter estimates for student factors like ethnicy than are invariant over time. 20 The available student background variables do not vary over time, so the x vector consists of a test year variable to detect any trend in scores. 21 i i 20 In our analysis, we test for serial correlation in the residuals in equation (3). First differencing is a preferred estimation method if there is strong posive serial correlation in panel data (Wooldridge, 2002). In this case, our test of serial correlation was weak, so the parameters from the fixed-effect model reported below are similar to those from the first-differenced model. 21 As mentioned previously, assessments of standardized achievement tests are often subject to score inflation (Hamilton, 2003). 26

28 A more general random-growth model is also used to control for the heterogeney of students attending different tradional and charter schools (Heckman and Hotz,1979; Papke, 1994; Wooldridge, 2002). The random-growth specification generalizes the fixed-effects model to allow for individual students to differ not only wh respect to a constant factor (µ) but also wh the rate of test score growth over time. The basis for the random-growth model is equation (4): T = µ + τ t + γc + x β + υ (4) i i where τ is an individual-specific growth rate. Equation 4 is a more general version of Equation 2 that allows for individual-specific differences in both the test score intercept and slope. The model is now first-differenced to obtain equation (5): i T = τ + γ C + x β + υ (5) The differencing eliminates the µ, and τ becomes the intercept of the differenced equation. 22 Equation 5 is estimated by fixed effects to eliminate τ i and to obtain estimates of γ and β. The potential advantage of the random-growth model over the fixed-effect model depends on two factors. First, the random-growth model is more appropriate if the test score trend varies across individual students (remember that the fixed-effects model also controls for an overall test score trend). Second, the random-growth approach is preferred if the student-specific trend is correlated wh charter school enrollment or other exogenous variables in the model. For example, if individual students wh strong posive trends in achievement were prone to transfer to charters, then the fixed-effects approach might suggest a posive charter effect on achievement irrespective of whether those schools did anything to enhance the learning of the students that transferred from tradional to charter schools. The estimated parameters γ and β will vary ltle between these two model if these two factors do not hold. 22 The growth term simplifies because τ i t-τ i (t-1)= τ i. In this specification, the year-to-year change in the trend term is perfectly collinear wh the constant and cannot be estimated. The model does include a parameter for the change in the quadratic trend term. 27

29 A limation of the random-growth model is that requires at least three successive years of test score data to isolate a test score trend, and this requirement implicly excludes many students from the analysis. For example, from the cohorts of students enrolled in elementary school in , only second and third graders are observed for three consecutive elementary school years by the end of the observation period in Similarly students that enter second grade in or are not observed for three consecutive years and are not included in the estimates of the random-growth model in Equation 5. As an extension to these models, we estimate the charter interaction effects by race/ethnicy (black, Hispanic, and other race/ethnicy) and by whether the student is classified as LEP. For the fixed-effects analysis, equation 3 is modified to add interaction terms between the charter school dummy variable and a vector of student characteristics as specified in equation 6: T T i = γ(c C ) + (C i R i C R ) δ + (x - x )β + (υ υ ) (6) i i i i where the charter dummy variable ( C ) for student i in time period t is interacted wh a vector of demographic characteristics (R i ) of student i, including indicators for black, Hispanic, and limed English proficiency (LEP), and δ is a 3 1 vector of unobserved effects. In this formulation of the model, γ represents the effect of charter schools on test scores of students who are not black, Hispanic, or classified as LEP. The coefficients on the interaction terms (δ) indicate whether charter effects are higher or lower for black, Hispanic, or LEP students than for similar students in TPSs. The random-growth model was also modified to add interaction terms that show whether there is a charter effect difference by race/ethnicy or LEP status. s = τ i + γ C + (C R ) δ + x β + υ (7) i As in the earlier random-effects specification, this differenced model is estimated by fixed effects to eliminate the individual student effect and derive consistent estimates of γ, δ, and β. 28

30 The charter effect in both the fixed-effects and random-growth models is identified from the behavior of students that eher swch from a charter to a TPS or from a TPS to a charter. The differencing approach in both models isolates the student achievement gains that are associated wh these school transions, while holding constant permanent time-invariant factors (observed and unobserved) that affect achievement. The approaches compare the achievement gains of students in charters wh those of the same student in a TPS and visa versa. The charter effect is the differential effect on student learning of eher transferring to a charter from a TPS or from a TPS to a charter. The identification of a charter effect from swchers has several potential limations. First, the approach does not compare the growth in student achievement for students who attend only a tradional or charter school over the five year enrollment period Table 3 shows that these students comprise the bulk of enrollments in both districts. The Hausman test comparing random- and fixed-effects models showed that omted variable bias was substantial, however, so the parameter estimates from the random-effects model are inconsistent. Second, student or family factors may change over time in ways that coincide wh the transion to a charter school. For example, parents might provide more homework assistance if they perceive that a TPS is low qualy and reduce their input after transferring their child to a charter. If so, the estimated charter effect would understate the true effect of charters. 23 Basic Charter Effect Table 7 shows the coefficient estimates for the fixed-effects and random-growth model across the six districts. The fixed-effects results are reported to add robustness to our results as the results are similar in magnude and significance across the two models, which suggest that the estimation of the charter effect is not particularly sensive to the modeling approach. To the extent that the estimates differ, the random-growth model is preferred, since provides a more complete accounting of differences in the underlying growth trajectory of individual students. 23 An addional limation is that these results do not adjust for the possible competive effects of charter schools on TPSs. In the next section, we show that these effects are small in California. 29

31 Table 7 Fixed-Effects and Random-Growth Models of Student Achievement and Charter Status Fixed-Effects Model Random-Growth Model Reading Math Reading Math Elementary Schools Charter * * (0.0882) ( (0.1589) (0.1912) Trend * * * * (0.0080) (0.0095) (0.0135) (0.0163) Constant * * (0.0251) ( Observations Secondary Schools Charter * * * (0.0921) (0.1021) (0.1714) (0.1885) Trend * * * * (0.0070) (0.0078) (0.0138) (0.0152) Constant * * (0.0224) (0.0247) Observations Notes: Standard errors are in parentheses. Entries wh asterisks are associated wh coefficients that are significantly different from zero at the α=0.05 confidence level. As shown in Equation 5, the constant term in the random-growth model represents the trend in student achievement, so there is no separate constant term in this equation. The random-growth estimates in Table 7 show that elementary math students in charter schools have test scores 1.64 percentile points lower than comparable students in TPSs. At the secondary school level, reading scores are 0.34 percentile points higher in charters than in tradional schools. School type differences have no significant effect on elementary school reading or on secondary school math. These results show that achievement in charter schools is largely keeping pace wh tradional schools. The charter effects in the random-growth model are similar to those in the statewide analysis shown in Table 4. The estimates in Table 7 are substantially more complete, however, because they implicly control for unobserved student-level factors that might distort the estimates of the charter effect. 30

32 Charter Effects by Demographic Groups Because many urban leaders, including mayors and school district superintendents, have iniated charter schools as a mechanism to improve learning for disadvantaged students, we also examined the effects of charter schools on urban districts student achievement generally and on different demographic groups, using data from Los Angeles and San Diego. The results are based on the analysis from the paper entled Charter School Performance in Urban Districts. 24 To carry out this analysis, we again used a random-growth model as displayed in equation 2, but added terms that interacted the student racial/ethnic characteristics and LEP status of students wh charter variable. 25 The interaction of charter status wh student type is useful for examining whether charter effects differ for different groups. Black, Hispanic, and LEP students lag behind other students in both of these districts as is the case in most urban districts. If charters reduce student achievement gaps for these groups, then they might pose an important policy option for improving the performance of at risk students. The interaction terms in the regression specification complicate the interpretation of the charter effects for different types of students. Given the complexy of the regression specification, we report summary tables of the charter school effects for different types of students in Table 8 (elementary students) and Table 9 (secondary students). 26 The tables also show the average test score effect overall and by student type in Los Angeles and San Diego. As expected, the averages indicate that Black and Hispanic students lag substantially behind the other race/ethnic group. LEP students, who are predominantly Hispanic in California, also have lower test scores than similar students wh stronger English skills in each race/ethnic group. 24 Zimmer and Buddin (2006) 25 Another commonly used indicator of socioeconomic status for students is participation in the Department of Agriculture s free-and-reduced lunch program. San Diego Unified School District does not include this lunch-status variable in s research data files, so the information was not included in our breakdowns or analysis. 26 Zimmer and Buddin (2005) shows the complete regression specifications for these models. 31

33 The Los Angeles elementary school results in Table 8 show that nearly all types of charter students are doing about the same in charters as they would have done in tradional schools. Black non-lep students are the only group wh a significant charter effect in eher reading or math test scores, and these students do worse in charters than in TPSs. Table 8 Summary of Elementary School Results Reading Test Math Test TPS Score Charter Effect TPS Score Charter Effect Los Angeles Schools Overall Student Type Black & not LEP * * Black & LEP Hispanic & not LEP Hispanic & LEP Other & not LEP Other & LEP San Diego Schools Overall * * Student Type Black & not LEP * * Black & LEP Hispanic & not LEP * Hispanic & LEP * Other & not LEP * * Other & LEP * Notes: An asterisk indicates that the coefficient is statistically significant at the 5% level. The San Diego elementary school results in Table 8 show that the effects of charters are predominately negative wh the magnude of the effects being more negative in math than in reading. Black charter school students have test scores that are 2.88 and 3.79 points lower and statistically significant in reading and math, respectively, compared to comparable Black students in TPSs. Hispanic students have worse math scores than their counterparts in TPSs, but reading scores do no differ significantly across the two types of schools. Interestingly, LEP status for Hispanic students (the largest share of LEP students) does not affect the differential achievement of students between charter and tradional schools LEP and other Hispanic charter students do no better in 32

34 reading than their tradional school counterparts, but both groups of Hispanics do worse in math. The secondary school charter effects for particular student groups are summarized in Table 9. In Los Angeles, the results show charter students performing 1.15 percentile points lower in reading than comparable students in tradional schools, but charter math scores are 1.28 percentile points higher than those for TPSs. The results show that Black students are performing at the same level in both charters and TPSs. Hispanic and other charter students are doing worse in reading than are comparable student in TPSs. Hispanics charter student are scoring higher in math than are similar Hispanic students in TPSs. Table 9 Summary of Secondary School Results Reading Test Math Test TPS Score Charter Effect TPS Score Charter Effect Los Angeles Schools Overall * * Student Type Black & not LEP Black & LEP Hispanic & not LEP * * Hispanic & LEP * * Other & not LEP * Other & LEP * San Diego Schools Overall * * Student Type Black & not LEP * * Black & LEP * * Hispanic & not LEP Hispanic & LEP * Other & not LEP * * Other & LEP * * Notes: An asterisk indicates that the coefficient is statistically significant at the 5% level. The secondary school results for San Diego show posive reading effects (1.49 points) and negative math effects (1.69 points). The posive reading and negative math effects are dominated by the Black and other student effects in both areas. Hispanic students do about the same in charters as in tradional schools. 33

35 In general, the results of the above analysis suggest that charter schools are having, at best, mixed results for students of different racial/ethnic categories and LEP students. While there are some cases in which charter schools do improve the performance of blacks and Hispanics, is clear that they are not consistently creating greater gains than their TPS counterparts. We should note that these results are for two large districts in California, and the effects of charter schools may differ in other states and districts, where the charter laws and operational environment differ from that in these California districts. Other Measures of Student Outcomes Our analysis, like several recent charter studies, has focused on student achievement in reading and math as the key measure of charter and tradional school performance. A more complete analysis would include other measures: 1) school factors like safety and learning environment, 2) student achievement in other dimensions than reading and math, 3) noncognive student measures like self-confidence, motivation, and self-esteem, and 4) high school completion and college transion. These factors are hard to measure objectively, and data availabily is scarce. Charter schools may be outpacing tradional schools in several of these dimensions, but the evidence is not yet available. If charters are making substantial improvements to student safety, the learning environment, or noncognive student factors, however, we would expect many of these gains to translate into improvement in student reading and math achievement. The broad picture from reading and math scores is that charters are keeping pace wh tradional public schools, irrespective of changes in other factors related to school type. IS CHARTER COMPETITION IMPROVING THE PERFORMANCE OF TRADITIONAL SCHOOLS? While much of the existing research on charter schools has focused on student achievement effects for students who choose to attend charter schools, we argue that this focus may be too narrow. Supporters hope that charter schools can exert healthy competive pressure on the existing K 12 educational system by giving families alternatives to TPSs. In fact, given that charter schools will probably never educate a substantial portion of the nation s student population, charter advocates argue that these 34

36 schools may have their greatest impact through systemic effects the competive effects of charter schools could improve the performance of TPSs and enhance the performance of students who do not attend charter schools. The challenge in evaluating possible competive effects is in knowing when district or school personnel will perceive a competive threat. Do charter schools create competive pressure when they are located near a TPS or when they first appear in a district? Do charter schools only create competive pressure when they start recruing students away from a particular school, or do they exert pressure when they capture a certain portion of students whin a marketplace? Addionally, the local environment may influence the competive pressure that charter schools create. For instance, some districts may have well developed, preexisting choice programs, including magnet schools or open enrollment policies. Also, some districts may be experiencing significant growth or already have overcrowded schools, in which case charter schools may act more like a release valve than a source of competive pressure. Researchers investigating this issue have made crical assumptions about the competive process, which could affect conclusions about the competive effects. For example, Hoxby (2001) defines competion as whether there is minimum market penetration of charter schools whin a district and finds substantial posive competive effects in Arizona and Michigan. 27 However, Bettinger (2005), using an instrumental variable strategy, also examines competive effects in Michigan as measured by distance and finds no effects. Using school-level data in North Carolina, Holmes, DeSimone, and Rupp (2003) also used distance as a proxy for competion and found substantial competive effects. In contrast, Bifulco and Ladd (2006a) use student-level data in North Carolina and map out the distances of students exing public schools to enter charter schools. Using this mapping, they analyze the effect charter schools have on TPSs whin concentric distances of charter schools. Their analysis finds no competive effects. Sass (2006) and Booker et al. (2004b) also use student-level data in Florida and Texas, respectively, to examine competive effects. Similar to Bifulco and Ladd, Sass 27 More specifically, she identifies competion has occurring only when charter schools represent more than 6 percent of district s total student population. 35

37 uses concentric circles around public schools and measures whether a charter school is whin these concentric circles and what proportion of total students are enrolled in charter schools. Using these approaches, Sass finds posive, but small, competive effects in Florida. Booker et al. uses two approaches, which find consistent and substantial competive effects. First, as in the Hoxby study, the authors use market penetration measures at the district level. Second, they use a campus-level market penetration measure, which is defined by the percentage of students at a tradional campus that leave the school for a charter school. They find competive effects across both measures. In a paper entled Is Charter School Competion in California Improving the Performance of Tradional Public Schools? we examine the competive effect charter schools have on TPSs using our survey and district student-level data. Because of space limations, we will only summarize the results of the surveys here. 28 Overall, our survey of principals from TPSs felt very ltle competive pressure and they indicated that they had changed very ltle in terms of their operation in response to charter schools being introduced to the market. Below, we highlight our results from the student achievement analysis, which confirms our results from the survey analysis. Data Description The analysis uses several measures to characterize the extent of pressures on a TPS to improve s performance in response to competion from charters and other public schools. Distance to charter or other public school. A nearby charter school may create competion for a public school. The pressure may come from students swching to the charter or from parents demanding higher standards in their current school. Similarly, a nearby TPS or magnet school may pressure a school to improve. Presence of charter and other alternatives whin 2.5 miles. This measure resembles the earlier distance measure, but focuses on the specific area near 28 A more complete description of the findings are in Zimmer and Buddin (2005b). 36

38 a school, which could be thought of as an educational marketplace. Distant charters or magnet schools may have much less effect on competion than the presence of nearby alternatives. Number of charters and other alternatives whin 2.5 miles. More alternatives give parents more opportunies to shop for a school and demand better classroom performance. Share of charter and other students whin 2.5 miles. This measure is a more precise indication of the availabily of classroom seats in the neighborhood of a school. If the share of students in the TPS is small, then faces great pressure to keep pace wh nearby alternative schools. Students lost to other schools whin 2.5 miles. This is a tally of the percentage of students swching to a nearby charter or other school in the previous year. If transfer rates are high, the school may be pressured to improve s performance relative to nearby schools to avoid the loss of revenues and personnel. 29 Table 10 shows the means for these school-level competion measures for all non-charter schools in California?. A key indication of charter importance in California is that 37 percent of elementary TPS students are whin 2.5 miles of an elementary charter school. Similarly, 24 and 9 percent of middle and high school students in a TPS are whin 2.5 miles of a charter middle and high school, respectively. The charter enrollment shares are small, however, ranging from an average of about 3 percent for elementary and middle school students to an average of about 1 percent of high school students. The analysis will focus on whether differences in these competive measures across schools and over time affect student achievement scores in these TPSs. 29 These types of measures have been used by Bifulco and Ladd (2003), Booker et al. (2004b), Bettinger (2005), and Sass (2005) to assess how charter school competion was affecting public school performance in North Carolina, Texas, Michigan, and Florida, respectively. Some of these studies also included competion from schools beyond 2.5 miles. In our analysis, however, the districts are all urban, so the 2.5-mile radius typically included other schools and at least some charter schools. 37

39 Table 10 Means and Standard Deviations by School Type Elementary School Middle School High School Standard Standard Standard Mean Deviation Mean Deviation Mean Deviation Student-Level Variables Reading Percentile Math Percentile Black Hispanic Asian Limed English Proficiency Free/Reduced Lunch School-Level Variables Magnet School Free/Reduced-Price Lunch (%) English Learner (%) First-Year Students (%) School-Level Competion Measures Distance to Nearest Charter Distance to Nearest Startup Distance to Nearest Conversion Distance to Nearest TPS Distance to Nearest Magnet Any Charter whin 2.5 Miles Any Startup whin 2.5 Miles Any Conversion whin 2.5 Miles Any Other TPS whin 2.5 Miles Any Magnet whin 2.5 Miles Number of Charters whin 2.5 Miles Number of Startups whin 2.5 Miles Number of Conversions whin 2.5 Miles Number of Other TPS whin 2.5 Miles Number of Magnets whin 2.5 Miles Share of Charters whin 2.5 Miles Share of Startups whin 2.5 Miles Share of Conversions whin 2.5 Miles Share of Other TPS whin 2.5 Miles Share of Magnets whin 2.5 Miles Lost to Charter in Past Year (%) Lost to Conversion in Past Year (%) Lost to Startups in Past Year (%) Lost to Other TPS in Past Year (%) Lost to Magnets in Past Year (%) Sample Size 1,143, , ,183 Note: The competion measures refer to schools of a similar type. For example, the share of charters whin 2.5 miles for elementary schools is based on the share of charter students in grades K through 5 relative to all students in grades K through 5 whin 2.5 miles of the TPS. 38

40 The last few years have been marked by dramatic growth in charter schools in California. Table 11 describes the pattern of charter school growth in the six districts used in our analysis. The results show that the numbers of charter schools and students have more than doubled for elementary, middle, and high schools for the five-year period from 1998 through Similarly, the share of students enrolled in charter schools has also doubled over this period. This charter school growth means that more students have opportunies to attend charters and that more TPSs face competion from charters in the local school marketplace. Model Table 11: Trends in Charter Schools in Six California School Districts for Student Achievement Analysis All Schools Number of Charters Number of Charter Students Share of Charter Students (%) Elementary Schools Number of Charters Number of Charter Students Share of Charter Students (%) Middle Schools Number of Charters Number of Charter Students Share of Charter Students (%) High Schools Number of Charters Number of Charter Students Share of Charter Students (%) Note: Some charter schools span elementary, middle, and high schools grades, so the sum of the number of schools by grade exceeds the overall number of schools. The model of student achievement in public schools is based on a three-way error component where the three components consist of a student-specific effect, a schoolspecific effect, and a year-specific effect (Abowd, Creecy, and Kramarz, 2002; Abowd, Kramarz, and Margolis, 1999; Andrews, Schank, and Upward, 2004; Andrews, Schank, and Upward, 2005). The dependent variable is the student test score, T ijt, observed for student i in school j at time t. Separate test specifications are estimated in reading and math. The formal model is 39

41 T ijt = x β + w γ + θ + ψ + µ + ε (8) jt i j t ijt where x and w jt are vectors of measured time-varying student and school characteristics, respectively; β and γ are parameter vectors for student and school effects; and unobserved, student, school, and time heterogeney are represented by θ i, ψ j and µ t. The last component of the model is a random error term, ε ijt, that is orthogonal to all other effects in the model. In many circumstances, a random-effects approach is used to estimate models like that in Equation 8. This approach assumes that the student, school, and time heterogeney terms are orthogonal to the observed student- and school-level variables. In this suation, this seems unlikely because the measures of these variables are incomplete. For example, student motivation and parental support are important determinants of schooling outcomes, but these factors are not measured in test score databases, such as those used in this study. Similarly, schools may differ from one another in unmeasured ways that may be correlated wh measured factors in the model. As a result, random effects estimation of Equation 8 is likely to yield inconsistent estimates of the parameters β and γ. In preliminary statistical regressions, we compared coefficient estimates wh random and fixed effects models using a Hausman specification test. The results indicated that unobserved student and school heterogeney was significantly correlated wh observed factors in Equation 1 for separate runs for reading and math in elementary, middle, and high school. These results indicated that a fixed-effect approach was more appropriate for this statistical problem. Fixed-effects methods produce consistent estimates of β and γ in Equation 8. The parameter for time heterogeney was estimated directly wh time dummies because the period of observation consisted of five consecutive years. Student and school heterogeney are more complex. Student test scores are observed over time and, in many cases, across different schools. For a particular student spell at a school, the terms θ i and ψ j are both fixed. As a result, student and school heterogeney can be eliminated from the model by taking spell-specific fixed effects for each student-school combination where: 40

42 T ijt T = x x ) β + ( w w ) γ + ( ε ε ). (9) s ( s jt s ijt s This approach means that parameters corresponding to student and school characteristics that are invariant across spells are not identified. The model does eliminate student and school heterogeney, however, whout restricting factors to be orthogonal to measured student and school attributes. Results As we noted earlier, researchers have used a variety of measures of charter competion, partially because no one really knows when TPSs may actually feel a competive threat. The results look at various measures of competion for elementary, middle, and high schools in reading and math. After controlling for student and school heterogeney and measured school variables, ltle evidence was found that charters are affecting student achievement in other public schools at all. Table 12 summarizes the effects of estimating various versions of Equation 9 wh alternative measures of student competion. Nearly every measure of charter competion has a statistically insignificant effect on student achievement in nearby TPSs. Only one of the measures has a statistically significant effect at the elementary, middle, and high school level. Even then the significant factor differs across school types, and the effect does not persist across reading and math whin the same school type. Two of the three significant effects have the oppose sign predicted by a theory of charter competion. In elementary reading, the regression result suggests that a larger share of charters whin 2.5 miles reduces reading scores at the TPS, but has no effect on math achievement. In high school reading, the distance to nearest charter has a posive effect on reading scores i.e., reading scores are higher if charters are further away from the TPS. The result for middle school math does show that the presence of a charter whin 2.5 miles of the TPS is associated wh a higher math achievement score. Overall, these student achievement results suggest that charter schools are not having an effect on the performance of TPSs, which is consistent wh our survey results from school principals. Table 12 Effects of Charter School Competion on TPS Performance Reading Math 41

43 Standard Error Standard Error Coefficient Coefficient Elementary Schools Distance to Nearest Charter Any Charter whin 2.5 Miles Number of Charters whin 2.5 Miles Share of Charters whin 2.5 Miles * Lost to Charter in Past Year (%) Middle Schools Distance to Nearest Charter Any Charter whin 2.5 Miles * Number of Charters whin 2.5 Miles Share of Charters whin 2.5 Miles Lost to Charter in Past Year (%) High Schools Distance to Nearest Charter * Any Charter whin 2.5 Miles Number of Charters whin 2.5 Miles Share of Charters whin 2.5 Miles Lost to Charter in Past Year (%) Although Table 12 does not suggest a competive effect from charter school generally, the effects charter schools have on TPSs may vary across different types of charter schools, which we examine by looking at startup and conversion schools separately. Again, we found no competive effects from eher type of charter schools (Buddin and Zimmer, 2005). Also, as one further sensivy analysis, we examined whether the level of preexisting competion may affect the competive impacts of charter schools. As we have previously noted, a preexisting competive market could blunt the competive effects of charter schools. Therefore, we also examined the charter competive effect while adding measures of competive effects from other TPSs or magnet schools. Again, the measures of competion from charter schools generally indicated no competive effect. This suggests that competion among the existing public schools cannot explain the lack of competive effects from charter schools. The absence of a competive effect, however, could also be explained by the generally low share of students charter schools represent in any of these districts never more than three percent or by the fact that charter schools are acting as a release valve in these growing districts. It is possible that a broader implementation of charter schools (than that observed in California) would exert pressure on tradional public schools to improve their performance. 42

44 SUMMARY AND CONCLUSIONS The charter movement grew out of a hope that by providing greater autonomy to schools, they would be able to cut through bureaucratic frustrations and offer innovative, efficient, and effective educational programs, provide new options to families, and promote healthy competion for TPSs. Our results from California show that charter schools generally perform on par or slightly below the achievement levels of TPSs, they have not closed the achievement gaps for minories, and have not had the expected competive effects on TPSs. On a more posive note, they have achieved comparable test score results wh fewer public resources than tradional schools and emphasized non-core subjects. The evidence shows that charter schools have not created whe enclaves or skimmed high-qualy students from TPSs in fact, charter schools have proven to be more popular among black and lower-achieving students and may have actually created black enclaves. Finally, we discovered few measures of school operations that predicted high performing schools. In particular, greater school autonomy associated wh charter schools has ltle effect on student achievement. In sum, the results suggest that charter schools are not a silver bullet for school improvement but may offer certain educational alternatives that are attractive to some parents. 43

45 REFERENCES Abowd, J., F. Kramarz, & Margolis, D. (1999). High-Wage Workers and High-Wage Firms, Econometrica, 67, Abowd, J., Creecy, R., & Kramarz, F. (2002). Computing Person and Firm Effects Using Linked Longudinal Employer-Employee Data, Technical Paper , U.S. Census Bureau, available at accessed September 6, Andrews, M., Schank, T., & Upward, R. (2004). Practical Estimation Methods for Employer-Employee Data, IAB Discussion Paper No. 3, available at accessed September 6, 2005., (2005). Practical Fixed Effects Estimation Methods for the Three-Way Error Components Model, available at accessed September 6, Bettinger, E.P., The Effect of Charter Schools on Charter Students and Public Schools, Economics of Education Review 2005, Vol 24(2): Betts, J.R., & Hill, P.T., Charter School Achievement Consensus Panel, Key Issues in Studying Charter Schools and Achievement: A Review and Suggestions for National Guidelines. Seattle, WA: National Charter School Research Project, Center on Reinventing Public Education, Bifulco, R. and Ladd, H. The Impact of Charter Schools on Student Achievement: Evidence from North Carolina. Education Finance Policy, Vol. 1(1), Winter 2006a, pp Bifulco, R. & Ladd, H., Race and charter schools: evidence from North Carolina, Forthcoming, Journal of Policy Analysis and Management., 2006b Booker, K., Gilpatric, S., Gronberg, T. J., & Jansen, D. W. (2004a). Charter School Performance in Texas, College Station, Tex., Private Enterprise Research Center Working Paper, Texas A&M Universy. (2004b). The Effect of Charter Schools on Tradional Public School Students in Texas: Are Children Who Stay Behind Left Behind? working paper, Texas A&M Universy. 44

46 Booker, K., Zimmer, R., & Buddin, R. "The effect of charter schools on school peer composion." RAND Working Paper: WR-306-EDU, 2005, Available at: accessed August 21, Borland, M. V. and R. M. Howsen. (1992). Student Academic Achievement and Degree of Market Concentration in Education. Economics of Education Review 11: (1993). On the Determination of the Crical Level of Market Concentration in Education. Economics of Education Review 12: (1996). Competion, Expendures, and Student Performance in Mathematics: A Comment on Couch et al., Public Choice 87: Bryk, A.S., and S. W. Raudenbush, Hierarchical Linear Models: Applications and Data Analysis Methods, Newbury Park, CA: Sage, Buddin, R. & Zimmer, R., Academic Outcomes, in Charter School Operations and Performance: Evidence from California. RAND Corporation: Santa Monica, CA. (2003). Buddin. R., & Zimmer, R.W., A Closer Look at Charter School Student Achievement. Journal of Policy Analysis and Management, 2005a, Vol. 24(2): Buddin, R. & Zimmer, R. Is Charter School Competion in California Improving the Performance of Tradional Public Schools? RAND Working Paper: WR-297- EDU, 2005b, Available at (Accessed on August 21, 2006). Cobb, C.D. and G.V. Glass. Ethnic Segregation in Arizona Charter Schools. Education Policy Analysis Archives, Vol. 7, No. 1, Darling-Hammond, L., Berry, B., and Thoreson, A. Does Teacher Certification Matter? Evaluating the Evidence, Educational Evaluation and Policy Analysis, Vol. 23, No. 1, 2001, pp Edwards, Brian, Characteristics and Demographics of California Charter Schools and Charter School Students, EdSource, Testimony before the California Senate Select Commtee on the Master Plan for Education, August, Fan, X. & Chen, M. (2001). Parental Involvement and students academic achievement: A meta-analysis. Educational Psychology Review, 13, Finn, C. E., B. V. Manno, and G. Vanourek, Charter Schools in Action. Princeton, NJ: Princeton Universy Press,

47 Frankenberg, E., and C. Lee (2003). Charter schools and race: A lost opportuny for integrated education. Education Policy Analysis Archives, 11(32). Retrieved from Gates, S. M, Ringel, J. S. Santibanez, L. Ross, K. E. and Chung, C. H.. Who Is Leading Our Schools? An Overview of School Administrators and Their Careers. Santa Monica, CA: RAND Corporation, MR-1679, Gill, B., Hamilton, L. Lockwood, JR., Marsh, J., Zimmer, R., Hill, D., and Pribesh, S. Inspiration, Perspiration, and Time: Operations and Achievement in Edison schools. Santa Monica, CA: RAND Corporation, MG-351, Goldhaber, D., andbrewer, D. Does Teacher Certification Matter? High School Teacher Certification Status and Student Achievement, Educational Evaluation and Policy Analysis, Vol. 22, No. 2, 2000, pp Goldhaber, D., and Brewer, D. Evaluating the Evidence on Teacher Certification: A Rejoinder, Educational Evaluation and Policy Analysis, Vol. 23, No. 1, 2001, pp Hamilton, L., Assessment as a Learning Tool, Review of Research in Education, 2003, 27: Hanushek, E. A., Kain, J. F., & Rivkin, S. G. The impact of charter schools on academic achievement. 2002, Available at: < accessed September 2, Heckman, J.J., and V.J. Hotz, Choosing among Alternative Nonexperimental Methods for Estimating the Impact of Social Programs: The Case of Manpower Training, Journal of the American Statistical Association 1989, 84: Henderson, V., P. Miezkowski, and Y. Suavageau (1978). Peer-Group Effects and Educational Production Functions. Journal of Public Economic, 10, Holmes, G. M., DeSimone, J., & Rupp, N. (2003). Does School Choice Increase School Qualy? WP9683, NBER. Hoxby, C. M. (1994). Does Competion Among Public Schools Benef Students and Taxpayers? NBER Working Paper Series, no. 4978, Cambridge, MA: National Bureau of Economic Research. Hoxby, C.M. (2001). How School Choice Affects the Achievement of Public School Students. Paper prepared for Koret Task Force meeting, Hoover Instution, Stanford, CA (September). 46

48 Hoxby, C. M., & Rockoff, J.E. (2004). The Impact of Charter Schools on Student Achievement, Cambridge, Mass.: Harvard Universy, available at accessed November 15, Jeynes, W.H. Parental Involvement and Student Achievement: A Meta-Analysis. Research Digest. Harvard Family Research Project. 2005, Available at: (accessed 4/25/06). Jeynes, W. H. A meta-analysis: The effects of parental involvement on minory children's academic achievement. Education & Urban Society, 35(2), 2003, pp Kolderie, T. Creating the Capacy for Change: How and Why Governors and Legislatures Are Opening a New-Schools Sector in Public Education. Education Week Press, Krop, C & Zimmer, R. Charter School Type Matters When Examining Funding and Facilies: Evidence From California. Education Policy Analysis Archives, Vol. 13 (50), Miron, G. N., and C. Nelson (2002). What s Public About Charter Schools? Thousand Oaks, CA: Corwin Press, Inc. Nathan, J., Charter Schools: Creating Hope and Opportuny foramerican Education, San Francisco, CA: Jossey-Bass, Papke, L.E., Tax Policy and Urban Developments: Evidence from the Indiana Enterprise Zone Program. Journal of Public Economics, 1994, Vol. 54: Powell, J., J. Blackorby, J. Marsh, K. Finnegan [QUERY: Is this different from the K. Finnigan ced above?], and L. Anderson. (1997). Evaluation of Charter School Effectiveness. Menlo Park, CA: SRI International. RPP International. (2000). The State of Charter Schools: National Study of Charter Schools Fourth-Year Report. Washington, D.C.: Office of Educational Research and Improvement, U.S. Department of Education. Sass, T.R. Charter Schools and Student Achievement in Florida. Education Finance and Policy, Vol. 1(1), Winter 2006, pp Summers, A. and B. Wolfe, (1977). Do Schools Make a Difference? American Economic Review, 67,

49 Wells, A.S., L. Artiles. S. Carnochan, C.W. Cooper, C. Grutzik, J.J. Holme, A. Lopez, J. Scott, J. Slayton, and A. Vasudeva. Beyond the rhetoric of charter school reform: A study of ten California school districts. Los Angeles: UCLA, Wooldridge, Jeffrey M., Econometric Analysis of Cross Section and Panel Data, MIT Press: Cambridge, Massachusetts., (2002). Zanzig, Blair R. (1997). Measuring the Impact of Competion in Local Government Education Markets on Cognive Achievement of Students. Economics of Education Review 16: Zimmer, R. W. and E. F. Toma, (2000). Peer Effects in Private and Public Schools across Countries. Journal of Policy Analysis and Management, 19(1): Zimmer, R. W. (2003). A New Twist in the Educational Tracking Debate. Economics of Education Review, 22: Zimmer, R., Buddin, R., Chau, D., Daley, G., Gill, B., Guarino, C., Hamilton, L., Krop, C., McCaffrey, D., Sandler, M., Brewer, D., Charter School Operations and Performance: Evidence from California. RAND Corporation: Santa Monica, CA. (2003). Zimmer, R. & Buddin, R. "Getting Inside the Black Box: Examining how the Operation of Charter Schools Affect Performance." RAND Working Paper: WR-305-EDU, 2005, Available at: accessed August 21, 2006 Zimmer, R. & Buddin, R. Charter School Performance in Urban Districts, Journal of Urban Economics, Vol. 60(2), ,