Teacher Turnover,Teacher Quality,and Student Achievement in DCPS

Abstract

In practice, teacher turnover appears to have negative effects on school quality as measured by student performance. However, some simulations suggest that turnover can instead have large positive effects under a policy regime in which low-performing teachers can be accurately identified and replaced with more effective teachers. This study examines this question by evaluating the effects of teacher turnover on student achievement under IMPACT, the unique performance-assessment and incentive system in the District of Columbia Public Schools (DCPS). Employing a quasi-experimental design based on data from the first years of IMPACT, we find that, on average, DCPS replaced teachers who left with teachers who increased student achievement by 0.08 standard deviation (SD) in math. When we isolate the effects of lower-performing teachers who were induced to leave DCPS for poor performance, we find that student achievement improves by larger and statistically significant amounts (i.e., 0.14 SD in reading and 0.21 SD in math). In contrast, the effect of exits by teachers not sanctioned under IMPACT is typically negative but not statistically significant.

Keywords

teacher quality teacher turnover teacher evaluation

Having an effective teacher can dramatically alter students’ educational and economic outcomes. Yet, we know that there are substantial differences in the quality of public school teachers, and there is increasing evidence that in some urban areas less effective teachers are often concentrated in lower-performing schools serving disadvantaged students. Policymakers and researchers recognize these issues and have sought policies to provide all children with effective teachers. The selective retention of effective teachers has been one of the most-discussed strategies that may contribute to this goal. In theory, districts could dismiss ineffective teachers, hire more effective teachers, and redouble efforts to retain effective teachers in these schools. However, we know relatively little about how such policies would work in practice. In particular, the capacity of districts to identify effective teachers at the hiring stage is limited (Boyd, Lankford, et al., 2008; Rockoff, Jacob, Kane, & Staiger, 2011; Rockoff & Speroni, 2010). Furthermore, research and practice have only recently begun making progress on accurately and reliably assessing teacher effectiveness. Some simulations (Hanushek, 2009; Staiger & Rockoff, 2010) estimate that dismissal of the least effective teachers would dramatically improve student achievement. However, these simulations make assumptions regarding the retention of more effective teachers and the labor supply of new teachers that may be overly optimistic. For example, if teacher-evaluation and dismissal policies cause more effective teachers to feel their jobs are threatened, such policies may have the unintended consequence of actually lowering teacher quality (Rothstein, 2015). The ultimate outcome depends on the details of the policy, the behavioral response of teachers, and the characteristics of the local labor market from which new teachers are hired.

Some school districts have begun to implement rigorous teacher-evaluation policies that could systematically dismiss meaningful numbers of ineffective teachers (Thomsen, 2014). However, we are unaware of any research that documents how the patterns of teacher turnover created by such policies (i.e., the attrition of teachers sanctioned for low performance, other teachers choosing to leave, and the hiring of new teachers) influence student achievement. In this study, we provide such evidence by examining the effects of teacher turnover under IMPACT, a seminal teacher-evaluation and compensation system introduced in the District of Columbia Public Schools (DCPS). Implemented at the beginning of the 2009–2010 school year, IMPACT evaluates all teachers annually based on multiple measures of effectiveness. Teachers rated as “Highly Effective” are offered large financial and nonfinancial rewards; those rated as “Ineffective” or twice consecutively “Minimally Effective” are separated from the district. In a purely descriptive sense, DCPS appears to be successfully employing compositional change to improve the quality of teaching. The average IMPACT scores of entering teachers exceed those of exiting teachers in each of the 3 years for which we have data (Figure 1); average differences vary between a third and a half of a standard deviation of teacher quality. Although these are impressive differences, there are a variety of reasons why the overall averages may misrepresent the ability of DCPS to systematically improve teacher quality in the classrooms where they are most needed.

Figure 1.

Average IMPACT scores of all general education teachers (Groups 1 and 2) by year.

We employ a quasi-experimental event study to examine teacher turnover and its effect on student achievement in DCPS in 2011 through 2013. Specifically, we rely on school-grade-year cells as our fundamental unit of observation and examine, in “difference-in-differences” specifications, how the patterns of teacher mobility influence student test performance in math and English-language arts (ELA). We find that teacher turnover in DCPS had an overall positive effect on student achievement in math (i.e., 0.08 SD), and that the effect of turnover in reading is also positive (i.e., 0.046 SD) but is only significant at the 10% level. However, the overall effect of teacher turnover masks considerable heterogeneity. We find that, when low-performing teachers (i.e., those with “Ineffective” or “Minimally Effective” ratings under IMPACT) leave the classroom, student achievement grows by 21% of a standard deviation in math and 14% of a standard deviation in reading. We also find that the attrition of high-performing teachers (i.e., those rated “Effective” or “Highly Effective”) has a negative but statistically insignificant effect on student performance.

To be clear, this article should not be viewed as an evaluation of IMPACT, or even as an assessment of IMPACT’s differential effect on teacher composition in DCPS.¹ Instead, we believe that this article makes an important contribution by examining the effects of teacher turnover under a unique policy regime. The existing literature finds that teacher turnover negatively influences student achievement (Bryk & Schneider, 2002; Guin, 2004; Ronfeldt, Loeb, & Wyckoff, 2013), perhaps due to the difficulty of replacing experienced teachers who leave and through disruptive effects among the teachers who remain. However, teacher turnover may instead have large positive effects if school districts can accurately and systematically identify low-performing teachers and replace them with more effective teachers. But whether that can be achieved at scale in a real-world setting is an open empirical question. In our study of DCPS schools under IMPACT, we find overall positive effects of teacher turnover. However, these effects are highly heterogeneous. The exits of teachers identified as low-performing on average meaningfully improve student achievement, while in some cases exits of high-performing teachers negatively influence achievement. Critically, the supply of entering teachers appears to be of sufficient quality to sustain a relatively high turnover rate. Nonetheless, retaining more high-performing teachers would provide substantial direct and indirect benefits.

Background

Improving teacher quality in schools with poor, low-performing, and largely non-White students has become an imperative of education policy. A recent body of research has made it clear that the variance in teacher effectiveness is qualitatively large and that more effective teachers can dramatically improve students’ short- and long-run life outcomes (Aaronson, Barrow, & Sander, 2007; Chetty, Friedman, & Rockoff, 2014; Jackson, 2012; Rivkin, Hanushek, & Kain, 2005; Rockoff, 2004). Furthermore, the evidence of stark inequities in access to effective teachers (Goldhaber, Lavery, & Theobald, 2015; Isenberg et al., 2013; Sass, Hannaway, Xu, Figlio, & Feng, 2012) has motivated efforts to improve teacher effectiveness as a means of reducing educational and economic inequality. Policies to improve teacher effectiveness can be conceptualized as either improving the performance of existing teachers or altering the composition of teachers. In this study, we focus on how changes in the composition of the teacher workforce (i.e., due to turnover and hiring) influence student outcomes under a system of performance assessment.²

Composition of the Teaching Workforce

The composition of a district’s teachers improves when their policies retain the most effective teachers, exit poorly performing teachers, and select the most able entering teachers. High-performing teachers leave their schools and districts for a variety of reasons, some personal, but most related to attributes of their jobs (Johnson, Kraft, & Papay, 2012). Several descriptive studies link teacher turnover to negative school environments and poor student outcomes (Bryk & Schneider, 2002; Guin, 2004). And this teacher turnover is likely to further exacerbate poor school performance through several mechanisms. Quasi-experimental evidence from New York City finds that teacher turnover leads to lower student achievement (i.e., −0.08 SD in math and −0.05 SD in reading) and that some of these effects reflect the disruptive nature of turnover on the students of teachers who remain (Ronfeldt et al., 2013). This evidence suggests that, in the absence of policies that effectively improve the composition of teachers, we should expect turnover to result in a decrease in student achievement.

Increasing the retention of effective teachers would appear to be an obvious strategy to improve teaching effectiveness, yet over a third of high-performing teachers report that they received little encouragement from their principals to remain at their current school (The New Teacher Project, 2012). There is only limited evidence that financial incentives make a difference in retaining teachers generally (Clotfelter, Glennie, Ladd, & Vigdor, 2008; Glazerman & Seifullah, 2012) and high-performing teachers specifically (Dee & Wyckoff, 2015). Teacher tenure decisions offer an opportunity to differentially retain the most effective novice teachers. Although states are increasingly employing more rigorous evaluations as part of tenure reviews, nearly all teachers reviewed are granted tenure (National Council on Teacher Quality, 2012). A tenure review process that more meaningfully differentiates teacher effectiveness is associated with substantial improvements in teacher quality and student achievement (Loeb, Miller, & Wyckoff, 2015).

In the absence of real-world evidence on the effects of policies that improve teacher composition, researchers have simulated the effects of such policies employing data-driven assumptions. It is estimated that annually replacing teachers who fall in the bottom 5% to 10% of the value-added distribution would improve student achievement by 50% of a standard deviation (Hanushek, 2009). Staiger and Rockoff (2010) suggest that replacing 80% of first-year teachers with new hires would increase student achievement of all students by 8% of a standard deviation. However, making alternative assumptions regarding the reliability of teacher quality measures and teachers’ behavioral responses to retention policies can lead to different outcomes. Districts whose evaluation system leads to the dismissal of meaningful numbers of teachers may face a limited supply of high-quality teachers. Teachers may find the stress and uncertainty of these working conditions outweigh the benefits, including compensation. As a result, such policies will need to be accompanied by improved working conditions or increased compensation (Rothstein, 2015).

In sum, there is at best limited empirical evidence of the effects of differential retention policies on teacher quality and student achievement. What evidence does directly bear on this issue are simulations dependent on a series of simplifying assumptions about the policies and the behavioral responses of existing teachers and the available labor market. Our study leverages the implementation of IMPACT, the high-stakes teacher-evaluation and compensation system in DCPS, to examine this issue directly in an at-scale setting.

Teacher Evaluation in DCPS

In just the last few years, the design and implementation of teacher evaluation has evolved quickly as many districts look to improve teacher performance, partly under the encouragement of federal policies such as waivers from the No Child Left Behind (NCLB) Act and the Teacher Incentive Fund (TIF). While these policy innovations are still a work in progress, best-practice principles of effective evaluation are beginning to emerge (Donaldson & Papay, 2015). DCPS was an early and influential adopter. DCPS began evaluating teachers under IMPACT, a new performance-assessment and incentive system, during the 2009–2010 school year. The design of IMPACT appears consistent with virtually all of the emerging best-practice principles. First, all teachers are evaluated on a multifaceted measure of teacher performance (e.g., clearly described standards, the use of multiple teacher observations made by different observers, and the use of student outcomes). Second, these evaluations are linked to high-powered incentives that include the potential dismissal of low-performing teachers and very large financial incentives for high-performers. Third, in addition to the feedback associated with the evaluations, teachers are provided with various supports, including instructional coaching to assist in improving their teaching practice. Because IMPACT has been implemented at scale over a sustained period, we have an opportunity to observe behavioral responses to the policy.

Each year, teachers receive a final IMPACT score that determines their IMPACT rating and their associated consequences. IMPACT scores range from 100 to 400 points and include several components, which depend on a teacher’s grade and subject of instruction. The teachers in this analysis all taught in tested grades and subjects and thus have a value-added component, which during the period of our analysis comprised 50% of their IMPACT score. Thirty-five percent of their IMPACT score was determined by rigorously scored classroom observations tied to the district’s Teaching and Learning Framework (TLF). The TLF specifies the criteria by which DCPS defines effective instruction and structures a scoring rubric. The TLF includes multiple domains such as leading well-organized, objective-driven lessons, checking for student understanding, explaining content clearly, and maximizing instructional time.³ A teacher’s TLF score is typically based on five formal observations: three by an administrator (e.g., a principal or assistant principal) and two by a “master educator” (i.e., an expert practitioner who travels across multiple schools to conduct TLF observations independently of administrators). Only the administrator’s first observation is announced in advance.

All teachers are also assessed by their administrators on a rubric that measures their support of school initiatives, efforts to promote high expectations, and partnerships with students’ families and school colleagues: the Commitment to the School Community (CSC) measure. CSC is weighted to represent 10% of the overall IMPACT scores. Teachers also received a score based on their school’s estimated value-added (SVA), which contributes 5%. Finally, principals assess each teacher on their “Core Professionalism” (CP). The rubric for CP rates teachers on the basis of attendance, punctuality, policies and procedures and respect. Teachers are assumed to be professionals, and, therefore, CP scores can only reduce a teacher’s overall IMPACT score.

During the period of our study, IMPACT scores were translated into one of four IMPACT ratings, which dictated consequences as shown in Table 1.⁴ Specifically, “Ineffective” teachers are separated from the district, as are teachers who receive two consecutive “Minimally Effective” ratings. The financial incentives available to high-performing teachers through IMPACTplus include one-time bonuses of up to US$25,000⁵ and permanent increases to base pay of up to US$27,000 per year, the present value of which is worth up to US$185,259 in current dollars.⁶ These design features of IMPACT create sharp incentive contrasts for teachers with scores local to the ME/E threshold (i.e., dismissal threats) and the HE/E threshold (i.e., the possibility of a permanent base-pay increase). An earlier study, employing a regression-discontinuity design, shows that, once these incentives become politically credible, they meaningfully increased the likelihood that teachers rated as ME exited DCPS (Dee & Wyckoff, 2015).

Table 1

IMPACT Ratings, Scores, and Consequence, 2009–2010 Through 2011–2012

IMPACT score	IMPACT rating	Consequence
100–174	Ineffective (I)	Dismissal
175–249	ME	Salary not advanced on salary schedule after first ME rating; dismissal after second consecutive ME rating
250–349	Effective (E)	None
350–400	HE	Bonus; eligible for permanent base-pay increase after second consecutive HE rating

Note. ME = Minimally Effective; HE = Highly Effective.

Teacher Retention in DCPS

During the period of our analysis, the average DCPS teacher attrition was 18% (Figure 2).⁷ A recent study of teacher attrition in 16 urban school districts across seven states finds year-to-year district attrition averages 13% but varies between 8% and 17% (Papay, Bacher-Hicks, Page, & Marinell, 2015). This suggests that the annual attrition in DCPS is comparatively high, which may reflect the intended and unintended effects of IMPACT as well as unique features of the local labor market. Interestingly, the attrition rate among teachers rated as “Effective” or “Highly Effective” (high-performers) was only 13%. Attrition of these higher-performing teachers was 10% in low-poverty schools and 14% in high-poverty schools.⁸ Some attrition of high-performing teachers undoubtedly results from the same forces that cause attrition in many districts (e.g., demanding working conditions or unsupportive leadership). However, in DCPS some high-performing teachers may leave because they find some features of IMPACT demotivating or stressful. If this response is sufficiently large, IMPACT could reduce teacher quality and student achievement. However, IMPACT also leads to the dismissal of ineffective teachers and induces other low-performing teachers to exit voluntarily (Dee & Wyckoff, 2015). The annual attrition rate of low-performing teachers is 46%. This implies that low-performing teachers were more than 3 times as likely to leave as high-performing teachers (Figure 2) and accounted for 36% of all exits from DCPS (Figure 3).

Figure 2.

Proportion of teachers exiting DCPS, by teacher performance and school poverty.

Figure 3.

Proportion of exiting teachers who are high- or low-performing, by school poverty status.

This descriptive summary of retention highlights the challenges confronting DCPS to improve student achievement by improving the composition of its teacher workforce. Losing 13% of the best teachers each year places strong demands on teacher recruitment to prevent a reduction in achievement in those classrooms. However, exiting 46% of low-performing teachers creates substantial opportunity to improve achievement in the classrooms of low-performing teachers. In the remainder of this article, we explore how teacher turnover in DCPS under IMPACT affects student achievement. We examine this question in the aggregate and separately for low- and high-performing teachers. We also consider whether the relationship between teacher turnover and student achievement varies across schools and over time.

Methodology: Conceptual Model and Empirical Strategy

To examine the effects of teacher turnover on student achievement, we employ a panel-based research design that effectively compares how outcomes in school-grade cells changed following the exit of a teacher to the contemporaneous change in school-grade cells where no turnover occurred. We particularly want to understand whether teacher effectiveness and student achievement are higher or lower as a result of exiting teachers. Changes in overall teacher effectiveness depend upon the magnitude of mean difference in effectiveness between entering and exiting teachers and the proportion of teachers who turn over. Changes in student achievement depend on these differences and on the relationship between measured teacher effectiveness and student achievement. Our empirical model attends to these relationships.

To illustrate how our research design utilizes student- and teacher-level data, we begin with the commonly used specification of student-level achievement shown in Equation 1. The achievement of student i in school s, grade g, assigned to teacher j and class c during year t ( $A_{i s g j c t}$ ) is a function of that student’s observables, including prior achievement (X_isgjct) and the attributes of classroom peers, ${\bar{X}}_{s g j c t}$ , the teacher’s value-added (µ_jsgt), a school fixed effect ( $π_{s}$ ), a year fixed effect (τ_t), and an idiosyncratic error term (ϵ_isgjct) that captures random noise that may occur at the individual and higher levels of aggregation (e.g., school, grade, classroom):

A_{i s g j c t} = β_{0} + X_{i s g j c t} β_{1} + {\bar{X}}_{s g j c t} β_{2} + μ_{j s g t} + π_{s} + τ_{t} + ε_{i s g j c t} .

To control for our student-level covariates, while facilitating further aggregation of this specification, we replace the dependent variable in Equation 1 with the student-level residuals ( $A_{i s g j c t}^{*}$ ) obtained by regressing $A_{i s g j c t}$ on $X_{i s g j c t}$ . Aggregating the resulting equation to the teacher level, we have

{\bar{A}}_{s g c j t}^{*} = β_{0} + {\bar{X}}_{s g c j t} β_{2} + μ_{s g c j t} + π_{s} + τ_{t} + {\bar{ε}}_{s g c j t}^{*} .

Consider the case of teacher j in a particular school and grade who was hired in year t to replace teacher j′ that left the school and grade at the end of the prior school year. Equation 3 shows the difference in average achievement of students taught by the entering teacher in the next year compared with that of the exiting teacher in the prior year:

\begin{array}{l} Δ {\bar{A}}_{s g j t}^{*}^{\underline{\underline{def}}} {\bar{A}}_{s g j t}^{*} - {\bar{A}}_{s g j^{'} t - 1}^{*} \\ = ({\bar{X}}_{s g j t} - {\bar{X}}_{s g j^{'} t - 1}) β_{2} + μ_{s g j t} - μ_{s g j^{'} t - 1} + π_{s} - π_{s} + τ_{t} - τ_{t - 1} + {\bar{ε}}_{s g j t}^{*} - {\bar{ε}}_{s g j^{'} t - 1}^{*} \\ = Δ {\bar{X}}_{s g t} β_{2} + Δ {\bar{μ}}_{s g j t} + Δ τ_{t} + Δ {\bar{ε}}_{s g j t}^{*} . \end{array}

That is, Equation 3 models the change in student achievement (i.e., conditional on student traits) as a teacher-level function of the change in classroom peers, the change in teacher quality, and other unobserved time-varying changes. However, conducting the analysis at the teacher level would have several prohibitive limitations. For example, if the attrition of a teacher has negative consequences on the productivity of his or her grade-level colleagues (Ronfeldt et al., 2013), this specification would not capture it. Furthermore, a teacher-level specification may exacerbate the bias due to internal-validity threats (Chetty et al., 2014). For example, in the presence of teacher turnover, some more motivated parents may seek to have their children placed in the classrooms of returning teachers, leaving entering teachers with lower-performing students (i.e., net of observed traits). Such nonrandom sorting of students to teachers is much less likely at the school-by-grade level, because the sorting would need to occur across schools to affect school-grade outcomes. Furthermore, observations at the annual school-grade level capture spillover effects that may exist among members of a school-grade team. Aggregating to the grade level also avoids the need to match each exiting teacher with the teacher replacing her. The student-weighted aggregation of Equation 3 to the school-grade-year level is shown here:

Δ {\bar{A}}_{s g t}^{*} = Δ {\bar{X}}_{s g t} β_{2} + Δ {\bar{μ}}_{s g t} + ω_{t} + Δ {\bar{ε}}_{s g t} .

Our analysis aims to understand how student achievement changes as a function of teacher turnover, rather than as a function of changes in teacher value-added in Equation 4. That is, teacher turnover may change teacher quality (e.g., Δµ_sgt), which in turn changes student achievement. We estimate a reduced-form model of changes in residualized student achievement as a function of teacher turnover at the end of the prior school year. E_sgt–₁ in Equation 5 is the student-weighted share of teachers in school s and grade g in year t − 1 who exit DCPS by the beginning of year t.

Δ {\bar{A}}_{s g t}^{*} = Δ {\bar{X}}_{s g t} β_{2} + γ_{1} E_{s g t - 1} + ω_{t} + Δ ε_{s g t}^{*} .

The identification strategy implied by this research design has a straightforward “difference-in-differences” logic.⁹ That is, we effectively examine the change in student performance in a school-grade cell before and after teacher turnover has occurred. These student-performance changes reflect both the effect of teacher turnover and the effect of other time-varying determinants. A second difference—the contemporaneous performance change in school-grade cells that did not experience turnover—captures the effects of those other time-varying determinants. The difference in these differences then isolates the effect of teacher turnover. Ideally, we would complement this analysis with similar results for the period that preceded the introduction of IMPACT. However, reliable data for teachers linked to students do not exist prior to 2009–2010.

Our quasi-experimental specification unrestrictively controls for several unobserved determinants of student achievement. More specifically, this specification identifies the effect of teacher turnover controlling for time-invariant traits specific to each school-grade cell, time-varying traits shared across all schools and grades, and various student-level traits including prior achievement. However, the internal validity of the inferences based on this basic model still rests on several critical assumptions that we engage directly. First, our design implicitly assumes that students do not sort to (or from) turnover classes by switching schools in a way that biases the results. Second, as currently specified, our approach implicitly assumes that, when filling vacancies due to turnovers, schools do not manipulate teacher transfers within DCPS in a manner that biases turnover results. For example, when an exit occurs within a school, principals do not systematically move the most or least effective teachers from other grades to fill that vacancy. Although there are slight variations across years and subjects, on average 55% of replacement teachers come from outside the DCPS system, 34% transfer within DCPS schools, and 11% transfer across DCPS schools. Our specification also assumes that these teacher transfers have no achievement implications for the “sending” school-grade cell (e.g., due to disruption of teacher peers). Third, our design assumes that there are no important unobserved factors changing at the school- or grade level that influence student achievement and that are also correlated with turnover (e.g., increasingly effective principals).

To address the robustness of our results in the presence of these potential confounds, we modify our basic estimation approach and conduct several robustness tests. First, we add several additional controls to our empirical models to address potential challenges to internal validity. To address concerns that within-school or across-school transfers may influence our results, we also introduce direct controls for these transfers. To assess the relevance of unobserved school trends that are correlated with turnover, we also employ specifications that include school fixed effects. Time-invariant school effects have already been eliminated from our design as a result of first-differencing school-grade observations. Adding a school fixed effect to our first-difference specification implies that we are also controlling flexibly for school-specific changes over time (e.g., school trends in culture and leadership).¹⁰ Second, we also estimate auxiliary regressions to examine whether our treatment explains changes in observed student attributes. To the extent that teacher turnover predicts changes in students traits, it would suggest endogenous switching based on unobserved correlates of these observed traits.

Third, we estimate the effects of teacher turnover on teacher quality directly. If, as our conceptual model suggests, teacher quality is the mechanism through which turnover influences student achievement, we should observe consistent results for the effects of turnover on both teacher quality and student achievement. To provide increased assurance that any student achievement changes associated with teacher turnover reflect its effects on teacher quality, we estimate some specifications where we employ IMPACT scores as the dependent variable.¹¹ As will be seen in Tables 3 to 5, in every instance where turnover is estimated to positively or negatively affect achievement, we observe an effect of turnover on observed teacher quality that is of the same sign and usually of proportionate magnitude. This increases our confidence that the change in teacher quality is the primary factor determining the effect of turnover on student achievement, and not other factors that could plausibly be associated with teacher exits.

We create three treatment variables to examine different types of teacher turnover. As before E_sgt−₁ is the proportion of students in a school-grade-year cell in the prior year whose teacher left the district and is used in specifications in which we examine the overall effects of turnover. In other specifications, $E_{s g t - 1}^{L}$ denotes the proportion of students in each such cell whose teacher exited DCPS and was a low-performing (i.e., Minimally Effective or Ineffective) teacher, and $E_{s g t - 1}^{H}$ denotes the proportion of students taught by a high-performing (Effective or Highly Effective) teacher who left the district at the end of year t − 1. In all specifications, we condition on the prevalence of within-school transfers, $S_{s g t - 1}$ , and transfers across schools in the district, $D_{s g t - 1}$ . These controls allow us to condition on the effects that turnover may have on school-grade cells that “send” teachers elsewhere within the district. The resulting specification takes the following form:

Δ {\bar{A}}_{s g t}^{*} = γ_{1} E_{s g t - 1}^{L} + γ_{2} E_{s g t - 1}^{H} + δ S_{s g t - 1} + θ D_{s g t - 1} + Δ {\bar{X}}_{s g t} β_{2} + ω_{t} + φ_{s} + ε_{s g t}^{*} .

Δ {\bar{TQ}}_{s g t} = γ_{1}^{'} E_{s g t - 1}^{L} + γ_{2}^{'} E_{s g t - 1}^{H} + δ^{'} S_{s g t - 1} + θ^{'} D_{s g t - 1} + ω_{t}^{'} + φ_{s}^{'} + ε_{s g t}^{'} .

Finally, we examine whether the effect of teacher turnover varies by year or school characteristics by interacting each treatment variable with the appropriate year or school-characteristic indicator variable (not shown). For instance, we assess whether the effect of teacher turnover differs between high- and low-poverty schools.

DCPS Administrative Data and Sample Construction

Our analysis draws on several sources of student, teacher, and school administrative data from DCPS. Students’ test scores, demographic variables, and teacher assignments come from DC’s Comprehensive Assessment System (DC CAS). These data span the 2009–2010 through 2012–2013 school years and include 56,564 student-years for tested students in Grades 4 through 8 with prior test scores.¹² Teacher administrative data include annual school assignments and annual IMPACT evaluation data. These data also span the 2009–2010 through 2012–2013 school years and include almost 1,900 teacher-years for teachers of students with DC CAS achievement data. Finally, school administrative data identify school type, poverty status, and closure status.

To construct our final analytical sample, we edited the data on the students and teachers in several conventional ways. First, we restricted our sample to general education classrooms, which resulted in dropping 12 special education campuses leaving 103 schools serving students tested in Grades 4 through 8. We then excluded students when they were tested in a grade other than their assigned grade (0.22% of student-year observations) or when they lacked a prior-year score (1.97% of student-year observations). To limit measurement error, we linked teachers to school-grade-year cells if the teacher is assigned to at least 10 tested students in that cell. This restriction eliminated 0.62% of teacher-school-grade-year observations. We also excluded teacher-year observations when those teachers taught in a school that closed at the end of that school year. This restriction also eliminated 0.62% of teacher-school-grade-year observations.

The primary outcome of interest is the year-to-year change in average residualized and standardized student achievement at the school-grade-year level.¹³ To construct this measure, we first standardize students’ scale scores to have mean 0 and unit SD within a subject, grade, and year. Next, we recover the residuals from a student-level linear regression of standardized test scores on lagged standardized test scores and student demographics. Then using the average residuals in each school-grade-year cell, we subtract prior-year outcomes from current-year outcomes. This measure isolates how test performance in each school-grade cell changed across years after controlling for the prior achievement and outcome-relevant baseline traits of the students it served.¹⁴ We similarly constructed first-differenced measures for the IMPACT scores of teachers in each school-grade-year cell. Because we have achievement data and IMPACT scores for 4 years (2009–2010 to 2012–2013), we are able to create 3 years of these differenced outcomes. Aggregating these observations to school-grade-year cells produces the unrestricted sample, whose descriptive statistics are shown in Table 2, columns 1 (math) and 5 (reading).

Table 2

Descriptive Statistics for Math and Reading, Various Samples, 2011–2013

	Math samples				Reading samples
	Unrestricted	Base	Unbalanced	Balanced	Unrestricted	Base	Unbalanced	Balanced
Average student characteristics (N = 56,564 student-year observations)
Students per s-g-y cell	50.6	51.2	51.0	51.1	50.6	51.4	51.7	52.8
Students per s-g-y cell	(44.6)	(45.1)	(44.7)	(44.9)	(44.6)	(45.0)	(45.5)	(46.7)
Proportion male	.51	.51	.51	.51	.51	.51	.51	.51
Proportion male	(.09)	(.09)	(.09)	(.09)	(.09)	(.09)	(.09)	(.09)
Proportion Black	.77	.77	.77	.76	.77	.77	.77	.76
Proportion Black	(.28)	(.29)	(.29)	(.29)	(.28)	(.28)	(.29)	(.29)
Proportion Hispanic	.15	.15	.15	.15	.15	.15	.15	.15
Proportion Hispanic	(.21)	(.21)	(.21)	(.22)	(.21)	(.21)	(.21)	(.21)
Proportion LEP	.08	.08	.08	.08	.08	.08	.08	.08
Proportion LEP	(.12)	(.12)	(.12)	(.12)	(.12)	(.12)	(.12)	(.12)
Proportion SpEd	.18	.18	.18	.17	.18	.18	.18	.17
Proportion SpEd	(.10)	(.09)	(.09)	(.09)	(.10)	(.09)	(.09)	(.09)
Proportion FRPL	.69	.70	.70	.69	.69	.70	.70	.70
Proportion FRPL	(.23)	(.23)	(.23)	(.23)	(.23)	(.23)	(.23)	(.23)
Residualized achievement	−.02	−.02	−.02	−.01	−.03	−.03	−.03	−.03
Residualized achievement	(.23)	(.23)	(.23)	(.23)	(.18)	(.18)	(.18)	(.18)
Average teacher characteristics (N = 1,873 teacher-year observations)
Teachers per s-g-y cell	1.68	1.68	1.68	1.70	1.86	1.86	1.86	1.89
Teachers per s-g-y cell	(.86)	(.87)	(.87)	(.88)	(.93)	(.93)	(.93)	(.95)
Any exit	.21	.21	.20	.19	.19	.20	.19	.19
Any exit	(.36)	(.36)	(.35)	(.34)	(.34)	(.35)	(.34)	(.34)
High-performer exit	.10	.10	.10	.10	.10	.10	.09	.09
High-performer exit	(.26)	(.27)	(.26)	(.25)	(.26)	(.26)	(.25)	(.25)
Low-performer exit	.11	.10	.10	.10	.10	.10	.10	.10
Low-performer exit	(.27)	(.27)	(.26)	(.26)	(.26)	(.26)	(.26)	(.26)
IMPACT score	283.5	283.7	283.7	286.3	284.8	285.2	285.2	286.6
IMPACT score	(51.0)	(51.2)	(51.2)	(50.7)	(48.7)	(48.7)	(48.7)	(48.4)
Teaching experience	9.55	9.63	9.63	9.91	9.30	9.37	9.37	9.56
Teaching experience	(6.92)	(6.89)	(6.89)	(6.92)	(6.77)	(6.76)	(6.76)	(6.72)
Average school characteristics
Number of unique schools	100	97	97	88	100	97	97	90
% High-poverty	80	80.41	80.41	79.55	80	80.41	80.41	81.11
% Elementary	64	64.95	64.95	67.05	64	64.95	64.95	65.56
% Middle	14	13.40	13.40	12.50	14	13.40	13.40	14.44
% Senior high school	1	1.03	1.03	1.14	1	1.03	1.03	1.11
% Education campus	20	20.62	20.62	19.32	20	20.62	20.62	18.89
School-grade-year obs.	838	751	734	663	838	753	733	666

Note. Unrestricted sample includes school-grade-year cells which contain nonmissing data for all variables in our model. The base sample restricts the sample to school-grade-year cells which contain nonmissing outcome data in 2 consecutive years (to form the first differences). The unbalanced sample further restricts to school-grade-year cells which contain both IMPACT scores and student achievement. The balanced sample is limited to school-grade cells which contain all 3 years of first differences. LEP = limited English proficiency; SpEd = special education; FRPL = free/reduced price lunch.

A final set of sample restrictions reflects concerns regarding missing data. First, differenced outcomes can only be created when the school-grade cell contains the outcome of interest in 2 consecutive years. This results in missing observations when schools open or close during the years of our analysis. This restriction produces school-grade-year cells with missing outcome data, which results in a loss of 87 school-grade cells in math (838 observations in the unrestricted sample to 751 in the base sample) and 85 in reading (838–753). Second, some school-grade-year cells are missing IMPACT scores, which results in different estimation samples for changes in IMPACT scores (Equation 6b) versus changes in student achievement (Equation 6a). Because we want to observe the effect of teacher turnover on teacher quality and student achievement in the same school-grade-year cells, we drop cells that are missing differenced IMPACT scores. This results in the loss of 17 school-grade-year cells in the math sample and 20 school-grade-year cells in the reading sample. The remaining sample is unbalanced, in that each school-grade cell is not observed in each year.

Third, we eliminate school-grade cells with fewer than 3 years of differenced outcomes. We are concerned that unbalanced observations introduce structural changes that influence estimates in ways that do not reflect responses to typical teacher exits. For example, school-grade cells may exist in some years but not others because schools close during the time frame of our analysis. In such situations, within-school, time-varying factors which we do not observe may influence student achievement and be correlated with teacher turnover, biasing our estimates. This restriction results in the loss of an additional 71 school-grade-year cells from the math sample and 67 school-grade-year cells from the reading sample, and creates the balanced sample.

Table 2 summarizes average student and teacher characteristics observed in each of the analytic samples. As might be expected, these sample restrictions influence the nature of our sample. Online Appendix Table 1 (available in the online version of the journal) statistically compares the differences in the means of the school-grade-year cells that were deleted in moving from the unrestricted to the balanced samples. As might be expected, relative to cells retained in the balanced sample, the dropped cells have lower residualized achievement and a somewhat higher percentage of Black students, a lower percentage of Hispanic students, and a higher percentage of students attending high-poverty schools. Cells dropped to create the balanced math sample also have lower average IMPACT scores and a higher incidence of teacher turnover. We estimate our basic student achievement model (Equation 6a) with and without these observations to explore how these exclusions affect our estimates. As is evident by comparing the estimates in Online Appendix Tables 2 (math) and 3 (reading; available in the online version of the journal), these estimates are similar. In general, these estimates show that, as we restrict the sample, the estimates of the effect of turnover of low-performing teachers become somewhat more positive, and the effect of exits of high-performing teachers becomes slightly less negative. Results of models that distinguish between high- and low-poverty schools and by year show very similar patterns. These results are available from the authors.

The “treatment” variable in our setting is defined by the proportion of students in a school-grade-year cell experiencing different types of teacher turnovers.¹⁵ Teacher–school assignment rosters and rosters that link teachers to students allow us to identify teacher exits as well as within- and across-school transfers. IMPACT ratings distinguish turnover among high-performing teachers (those rated “Effective” or “Highly Effective”) from turnover among low-performing teachers (those rated “Ineffective” or “Minimally Effective”). Turnover dosages are calculated by employing teacher–student assignment rosters to identify the proportion of students in school-grade-year cells affected by each type of teacher turnover. We identify teachers who move to a new grade cell in the same school or to a new school analogously. These measures control for movement out of the school-grade cell that does not result in an exit from DCPS.

Results

Our conceptual model suggests that the induced turnover of low-performing teachers (i.e., teachers rated by IMPACT as “Ineffective” or “Minimally Effective”) should result in improvements in teaching quality and student achievement, whereas the turnover of high-performing (“Highly Effective” and “Effective”) teachers may well result in a reduction in teacher quality and student achievement depending on the quality of entering teachers. The overall effect, which balances these two types of turnovers, is conceptually ambiguous and depends on the composition of exiting teachers and the quality of entering teachers.

Descriptive Summary

Before turning to our estimates, it may be instructive to examine simple averages of the IMPACT scores of exiting and entering teachers. If our estimates, which control for a variety of potential confounds, are wildly different from these simple means, we would want to understand how our adjustments influence the outcomes. Figure 4 shows the unconditional means of IMPACT scores of all exiting and entering general education teachers (i.e., teachers of all subjects in tested and untested grades) in DCPS.¹⁶ As might be expected, mean IMPACT scores of exiting high-performing teachers exceed those of entering teachers by 12 to 23 IMPACT points (i.e., 25%–45% of a standard deviation of teacher effectiveness [IMPACT scores]) depending on the year. In contrast, exiting low-performing teachers are substantially less effective than the average entering teacher, with differences between 71 and 85 IMPACT points (i.e., 1.4 and 1.7 SD). Across all teachers, entering teachers have IMPACT scores between a third and a half of a standard deviation greater than exiting teachers. A very similar pattern exists when the sample is restricted to teachers who can be matched to students with math achievement scores (Figure 5). Here, the difference between entering and exiting teachers varies by 35% to 55% of a standard deviation of teacher effectiveness depending on the year. The pattern for teachers matched to students with reading scores is identical to somewhat smaller differences between the IMPACT scores of entering and exiting teachers (i.e., 0.25–0.50 SD; results available from authors). Importantly for the purposes of our analysis, the differences in individual value-added (IVA) scores for entering and exiting teachers are very similar to those observed for teacher quality (Figure 6). The patterns for the exits of high-performing and low-performing teachers are identical to those observed for IMPACT scores.

Figure 4.

Average IMPACT scores of all general teachers (IMPACT Group 1 and Group 2) by status of exiting teacher and year.

Figure 5.

Average IMPACT scores of teachers who are matched to students with math achievement scores (IMPACT Group 1) by year.

Figure 6.

Average individual value-added scores of teachers who are matched to students with math achievement scores (IMPACT Group 1) by status of exiting teacher and year.

Comparing the IMPACT scores of entering and exiting teachers suggests that teacher quality is improving as a result of teacher turnover. This is true whether teacher effectiveness is measured by overall IMPACT scores or by value-added. However, when teachers who are judged to be high-performing voluntarily exit, they are replaced on average by somewhat less effective teachers. Contrast that with the exit of teachers who are either forced to leave as a result of IMPACT or whose performance, if not improved, would lead to a forced exit. Turnover in this instance appears to result in a substantial improvement in measured effectiveness. As discussed above, there are a variety of reasons why these simple comparisons may misrepresent the effects of teacher turnover in DCPS. For example, the composition of students may have changed from one year to the next in a way that either favors or disadvantages teachers entering a school-grade cell which experienced teacher turnover. We now turn to the estimation of Equations 6a and 6b, which control for a number of potentially confounding factors.

Quasi-Experimental Estimates

We report our main results (i.e., estimates based on Equations 6a and 6b) in Table 3. The results in the first row identify the estimated effect of overall teacher turnover. Interestingly, these results suggest that the exit of DCPS teachers led to improved teacher quality and student achievement in both math and reading, although the reading estimate is not significant at traditional levels. More specifically, these results imply that if all students in a school-grade cell experienced turnover of the average exiting teacher IMPACT scores would increase by 17.4 points (Table 3, row 1, column 1). This is approximately a third of a standard deviation of teacher effectiveness (SD = 50). The corresponding increase in student achievement is 0.079 SD. In reading, IMPACT scores are estimated to increase by 15.1 IMPACT points, and student achievement is estimated to increase by 0.046 SD but is only significant at the .10 level. Thus, on average, exiting teachers are replaced by teachers who are more effective as measured by IMPACT and who increase student achievement, at least in math.

Table 3

Effect of Teacher Turnover on IMPACT Scores and Math and Reading Student Achievement

	Math				Reading
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS
All exits	17.359*	0.079**			15.066*	0.046^†
All exits	(6.973)	(0.03)			(6.244)	(0.024)
High-performers			−29.720**	−0.055			−17.798*	−0.047
High-performers			(8.486)	(0.039)			(7.697)	(0.034)
Low-performers			63.838**	0.210**			46.129**	0.136**
Low-performers			(8.071)	(0.041)			(7.987)	(0.03)
Student controls		X		X		X		X
Observations	663	663	663	663	666	666	666	666
R ²	.035	.015	.138	.045	.035	.017	.087	.04

Note. Robust standard errors reported in parentheses. All models include year fixed effects and controls for teacher movement within and across schools. Student controls account for the year-to-year, across-cohort change in the percentage of students in a school-grade-year cell who are Black, Hispanic, other non-White race/ethnicity, limited English proficient, special education, or FRPL eligible. DC CAS = District of Columbia Comprehensive Assessment System; FRPL = free/reduced price lunch.

†

p < .10. *p < .05. **p < .01.

In the remaining rows of Table 3, we report the estimates when the effects of teacher turnover are allowed to differ across teacher effectiveness groups (i.e., high- and low-performers). These results indicate that the overall effects of teacher turnover masked considerable heterogeneity across low- and high-performing teachers. Turnover of high-performing teachers results in a decrease in average IMPACT scores of 30 points (i.e., 0.60 SD of teacher effectiveness) in math. This negative effect reflects the difficulty of replacing a high-performing teacher. Our estimates indicate that turnover of a high-performing teacher has a negative but statistically insignificant effect on student achievement (−0.055 SD). Similar, but smaller, results hold for reading.

In contrast, the exit of low-performing teachers substantially increases both teaching quality and student achievement. In math, the exit of low-performing teachers is estimated to improve teaching quality by 64 IMPACT points (1.3 SD) and student achievement by 0.21 SD. The effects on reading are somewhat smaller but still large, 46 IMPACT points and 0.14 SD of student achievement. Over the first 3 years of IMPACT, replacing teachers identified by IMPACT as low-performers leads to substantial improvement in student achievement as, on average, their replacements are meaningfully more effective teachers.

These estimates reflect the effect on student achievement if all teachers in a school-grade cell were of the identified type, for example, low-performing, and exited, and thus would overstate the effect on all the students in that school-grade cell if a low-performing teacher left a school-grade cell and the other teacher(s) in that cell remained. Alternatively, assuming no spillovers from one classroom within a grade to another, these estimates capture the average effect on the students in the exiting teacher’s classroom. A strength of our approach is to capture such spillovers.

Robustness of Results

The consistency of the effects of turnover on teacher quality and student achievement and their robustness to introducing student controls increases our confidence in the internal validity of our estimates. Nonetheless, legitimate concerns may remain that parents or principals may systematically respond to teacher turnover by altering the assignment of students to teachers in ways that threaten internal validity. For example, if turnover predicts changes in student attributes, it may signal strategic behavior by parents or principals that may bias our results. Fortunately, we find nothing of concern when we regress a variety of student characteristics on teacher turnover (Online Appendix Table 4, available in the online version of the journal). Of the 18 estimated coefficients (six student attributes by three types of teachers [all, high-performing, and low-performing]), only one is significant at conventional levels. The exit of all high-performing teachers from a school-grade cell is associated with a 2.4% decrease in limited English proficiency (LEP) students. These results suggest that there is not systematic sorting of students to teachers in response to turnover, and when there is some evidence, the magnitudes are modest. Nonetheless, we include controls for all student variables we can observe.

Another potential threat to the validity of our estimates may be that underlying trends in schools may cause student achievement to increase over time in school-grade cells with turnover but not in school-grade cells without turnover. To address this issue, we estimate first-difference models that introduce school fixed effects and models which include school-by-year fixed effects. The identifying variation for estimates with school fixed effects comes from within-school comparisons of school-grade cells with and without turnover. Adding a school-by-year fixed effect effectively limits our comparisons to grades in the same school and year with and without turnover. Estimates for our base models and those with school and school-by-year fixed effects are shown in Online Appendix Tables 5 (math) and 6 (reading; available in the online version of the journal). Adding school fixed effects to our base model changes the estimates only slightly. The one substantive change is the effect of a typical teacher exit on math student achievement. The coefficient is somewhat smaller (0.058 SD rather than 0.079 SD), the standard error larger (0.038 rather than 0.030), the combination of which results in a statistically insignificant estimate. Adding school-by-year fixed effects has a larger effect on some of the estimates. In math, while still significant and educationally meaningful, the effect of turnover of low-performing teachers on achievement is about half as large as in either of the other two models. In reading, the change is not nearly so dramatic. Adding school-by-year fixed effects substantially reduces the identifying variation in ways that have important implications for the identification of effects and for external validity. For example, 663 school-grade cells contribute to identifying the effects of our preferred specification in math (Table 3). This is reduced to 534 school-grade cells when we include school fixed effects and to only 317 if we include school-by-year fixed effects.

We include additional robustness checks in which we estimate the effects of two “placebo” models. In the first, turnover at the end of 2012–2013 is used to predict changes in student achievement from 2009–2010 to 2010–2011. If turnover is the mechanism that drives our results and not some other attribute of the school-grade cells that experience turnover, then the effects of the placebo estimates should not be similar to the estimates presented in Table 3. We find that they are not. As shown in Online Appendix Table 7 (available in the online version of the journal), none of the estimated coefficients in math or reading are statistically significant.

In the second “placebo” test, we predict student achievement as before—a function of turnover in the same grade cell in the prior year—but also include turnover in the next higher school-grade cell. For example, in observations considering changes in achievement in fourth-grade cells we also include the value of turnover for fifth-grade cells in the same school. If turnover in the next-grade cell predicts achievement in the current grade, we might be concerned that turnover signaled something about the school rather than turnover per se. It is possible that, because teachers may work together across grades, turnover in fifth grade could influence achievement in fourth grade in the following year. Because our analysis is premised on school-grade cells as units of observation, about 40% of our original sample is unavailable when we include turnover in the next school-grade cell as a control.¹⁷

As shown in Online Appendix Table 8 (available in the online version of the journal), next-grade turnover has no effect on the change in current-grade achievement, and the coefficients of current-grade turnover are robust to the inclusion of prior-grade turnover. Column 1 shows the effect of turnover in the current grade on math achievement in the succeeding year. This is the main result from the article for this smaller sample of school-grade cells. Column 2 shows the estimates for both current-grade turnover and next-grade turnover. The estimates of the effects of current-grade turnover remain largely unchanged, and the estimate of next-grade turnover is not significantly different from 0. The results for reading are qualitatively similar although for this reduced sample the main effect is insignificant.

Treatment Heterogeneity

There are several other ways in which the effects of teacher turnover may be heterogeneous. For example, the contexts across low- and high-poverty schools are likely to shape both the prevalence of teacher turnover and its effects on students. Overall, we find that high-poverty schools appear to improve as a result of teacher turnover. We estimate that the overall effect of turnover on student achievement in high-poverty schools is 0.084 in math and 0.052 in reading. Both estimates are statistically distinguishable from 0 (Table 4, row 2). In comparison, the point estimates of the effect of turnover in low-poverty schools are close to 0.

Table 4

Effect of Teacher Turnover on IMPACT Scores and Math and Reading Student Achievement by School Poverty Status

	Math				Reading
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS
All exits
Low-poverty	21.714	0.006			1.738	−0.038
Low-poverty	(19.301)	(0.082)			(8.727)	(0.043)
High-poverty	16.793*	0.084**			16.032*	0.052*
High-poverty	(7.259)	(0.03)			(6.548)	(0.025)
High-performer exits
Low-poverty			23.648	−0.004			1.922	−0.041
Low-poverty			(21.962)	(0.097)			(9.892)	(0.05)
High-poverty			−39.234**	−0.064			−20.437^†	−0.048
High-poverty			(−8.596)	(0.042)			(8.575)	(0.038)
Low-performer exits
Low-poverty			NA	NA			NA	NA
Low-poverty
High-poverty			64.075**	0.209**			46.761**	0.138**
High-poverty			(8.171)	(0.041)			(8.104)	(0.03)
Student controls		X		X		X		X
Observations	663	663	663	663	666	666	666	666

Note. Robust standard errors reported in parentheses. All models include year fixed effects and controls for teacher movement within and across schools. Student controls account for the year-to-year, across-cohort change in the percentage of students in a school-grade-year cell who are Black, Hispanic, other non-White race/ethnicity, limited English proficient, special education, or FRPL eligible. We do not include estimates for low-performer exits in low-poverty schools as these are found in only three school-grade cells. FRPL = free/reduced price lunch; NA = not applicable.

†

p < .10. *p < .05. **p < .01.

DCPS appears to be quite capable of replacing exiting high-performing teachers in low-poverty schools with comparable teachers (Table 4, row 3). However, replacing a high-performing teacher in a high-poverty school is more difficult and is estimated to result in a decrease of 80% of a standard deviation of teacher quality in math and 40% of a standard deviation in reading, though corresponding decreases in student achievement are not significant (Tables 4, row 4). These differences occur even though the average high-performing teacher who exits a low-poverty school is estimated to be 30% of a standard deviation of IMPACT scores more effective than the average high-performing exit from a high-poverty school.

Forty percent of teacher turnover in high-poverty schools is among low-performing teachers (Figure 3). Our estimates indicate that there are consistently large gains from the exit of low-performing teachers in high-poverty schools. In math, teacher quality improves by 1.3 SD and student achievement by 20% of a standard deviation; in reading, these figures are 1 SD of teacher quality and 14% of standard deviation of student achievement. In DCPS, virtually all low-performing teacher turnover is concentrated in high-poverty schools: on average, 1% of students in low-poverty schools experience low-performing teacher turnover.¹⁸

When we examine the effects of DCPS turnovers over time, we observe substantial consistency as well as a few interesting differences. Overall, the effects of DCPS turnover appear to become increasingly positive year to year. However, student achievement is estimated to be unaffected until 2013 when for math (Table 5, columns 1 and 2, first three rows) and reading (Table 5, columns 5 and 6, first three rows) the estimated effect is an improvement of 11% of a standard deviation of student achievement.

Table 5

Effect of Teacher Turnover on IMPACT Scores and Math and Reading Student Achievement, by Year

	Math				Reading
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS	IMPACT score	DC CAS
All exits
2011	13.053	0.092			2.113	−0.039
2011	(11.876)	(0.061)			(9.113)	(0.043)
2012	20.706	0.025			23.396**	0.046
2012	(14.355)	(0.059)			(10.354)	(0.039)
2013	20.017^†	0.112**			20.258^†	0.105*
2013	(11.429)	(0.041)			(10.895)	(0.041)
High-performer exits
2011			−15.404	−0.022			−38.553**	−0.161**
2011			(13.793)	(0.071)			(9.513)	(0.054)
2012			−53.682**	−0.277**			−9.302	−0.042
2012			(15.312)	(0.059)			(12.900)	(0.049)
2013			−21.426	0.057			−9.969	0.008
2013			(13.399)	(0.057)			(11.766)	(0.059)
Low-performer exits
2011			43.824**	0.215*			31.473**	0.050
2011			(14.939)	(0.100)			(10.267)	(0.049)
2012			74.931**	0.243**			54.623**	0.133**
2012			(16.914)	(0.067)			(12.376)	(0.050)
2013			70.166**	0.179**			53.750**	0.208**
2013			(12.678)	(0.051)			(15.822)	(0.047)
Student controls		X		X		X		X
Observations	663	663	663	663	666	666	666	666

Note. Robust standard errors in parentheses. All models include controls for teacher movement within and across schools. Student controls account for the across-cohort change in the percentage of students in a school-grade-year cell who are Black, Hispanic, other non-White race/ethnicity, LEP, special education, or FRPL eligible. LEP = limited English proficiency; FRPL = free/reduced price lunch.

†

p < .10. *p < .05. **p < .01.

For most years, the exit of high-performing teachers does not influence teacher quality or student achievement. However, in one year for math (2012) and reading (2011), the exit of high-performing teachers has a substantial negative effect on teaching quality and student achievement. These estimates are similar across alternative analytic samples that employ the base and unbalanced data. When we examine the individual exiting and entering teachers in the school-grade cells with teacher turnover, we observe the exit of several very effective teachers who are replaced by teachers whose subsequent performance places them among the low-performers.

In contrast, the exit of low-performing teachers yields consistently large improvements in teaching quality and student achievement in math (0.18–0.24 SD of student achievement) and increasing effects over time in reading (0.05 [not significant] to 0.21 SD of student achievement). In almost every year, DCPS has been able to replace low-performing teachers with high-performing teachers who have been able to improve student achievement.

Finally, we examined differences in the effects of turnover between elementary and middle school grades. For math, we find no statistically significant differences in the effects of turnover in elementary and middle school grades for either teacher effectiveness or student achievement. This is true for the overall model and for models that estimate effects for low- and high-performing teachers. For reading, the results are similar with the exception that when a low-performing teacher exits an elementary grade, teacher effectiveness increases substantially more than for a similar exit from a middle school grade (there is not corresponding increase in student achievement). When we divide our sample this way, our sample sizes are reduced, which may limit our ability to discern differences. These results are available from the authors.

Discussion

In general, higher rates of teacher turnover are legitimately thought to negatively influence student outcomes (e.g., Ronfeldt et al., 2013). However, DCPS constitutes a unique and policy-relevant case because, under IMPACT, a substantial fraction of teacher turnover consists of lower-performing teachers who were purposefully compelled or encouraged to leave, thus potentially altering the typical distribution of teacher effectiveness among exiting teachers. We find that the overall effect of teacher turnover in DCPS conservatively had no effect on achievement and, under reasonable assumptions, improved achievement. This average combines the negative, but statistically insignificant, effects of exits of high-performing teachers with the very large improvements in student achievement resulting from the exits of low-performing teachers.

The high stakes associated with IMPACT have been controversial, both within the District of Columbia as well as in broader discussions of education policy. There are elements of both sides of this debate in our estimates. While we are unable to identify high-performing teachers who leave DCPS because of IMPACT, our estimates indicate that replacing high-performing teachers who exit with teachers who perform similarly is difficult. In general, such turnover does not lead to statistically significant reductions in student performance, except in one notable instance (i.e., math teachers in 2011–2012).

Alternatively, IMPACT targets the exit of low-performing teachers. Our estimates show that doing so substantially improves teaching quality and student achievement in high-poverty schools. An improvement of 20% of a standard deviation of student achievement in math is roughly equivalent to 35% to 65% of a year of student learning, depending on grade level (Hill, Bloom, Black, & Lipsey, 2008). Similarly, improvements of 14% of a standard deviation in reading translate to 35% to 55% of a year of learning. More than 90% of the turnover of low-performing teachers occurs in high-poverty schools, where the proportion of exiting teachers who are low-performers is twice as high as in low-poverty schools. An important component of IMPACT’s design is to dismiss teachers rated as “Ineffective” and twice consecutively “Minimally Effective.” As is clear from this analysis, the benefits of that policy primarily redound to high-poverty schools. In comparison with almost any other intervention, these are very large improvements that are situated among some of the neediest students.

We should note that our analysis does not have the causal warrant of an experimental design. Nonetheless, under certain identifying assumptions that we articulate and examine, our quasi-experimental design does identify the change in student achievement caused by teacher turnover. However, we do not claim that IMPACT caused all of the teacher turnover we observe. Although IMPACT certainly caused some teachers to leave DCPS through dismissals, voluntary teacher attrition is likely driven by myriad teacher preferences.¹⁹ While it is possible that teachers may leave DCPS because they are dissatisfied with IMPACT and the human capital strategies in DCPS writ large, we are unable to link the attrition of high-performing teachers to IMPACT.²⁰ Nor do we know whether our turnover results for teachers and students in Grades 4 through 8 in math and reading generalize to turnover for other teachers and students. However, the descriptive summaries in Figures 4 and 5 would suggest they might.

Our empirical results were not inevitable, even for the turnover of low-performing teachers. As Rothstein (2015) makes it clear, there are good reasons to believe that the supply of teachers may be insufficient to maintain teacher quality, especially when teacher quality is difficult to ascertain in advance and challenging to improve in schools where there is substantial turnover. Our estimates suggest that, on average, DCPS is able to recruit replacements for exiting teachers who are at least as effective, and for low-performing teachers, replacements who are substantially more effective. These results are consistent with simple descriptive evidence on the effectiveness of entering and exiting teachers (Figures 1, 4, and 5). This may reflect the compensating differentials available to DCPS teachers in the form of bonuses and increases in base pay, or it may reflect specific aspects of the market for teachers in the District of Columbia. Other school districts may experience different results when implementing a system intended to increase the attrition of low-performing teachers.

The challenge of improving the composition of teachers in DCPS is increasing. First, as the least effective teachers exit, there are fewer such teachers to exit over time, and we would expect the average effectiveness of exiting teachers to continue to increase. Second, in 2012–2013 DCPS adjusted its evaluation system, so that to be rated as “Effective” or better (and thus avoid sanctions) teachers needed IMPACT scores of at least 300 rather than 250 as had been true in 2011–2012. Increasing the threshold for high-performing status will likely lead to the exit of some previously “Effective” teachers who are now classified as “Developing” and may cause some “Effective” and “Highly Effective” teachers to leave as they perceive the system as more stressful. However, DCPS made several other changes to IMPACT in 2012–2013 that may cause the system to be more hospitable, such as reducing the number of teacher observations, increasing access to bonus and base-pay increases, and reducing the weight of value-added for Group 1 teachers.

We expect that both the declining numbers of very low-performing teachers and changes in the IMPACT rating thresholds place strong demands on the system to continue recruiting effective teachers to replace the exit of higher-performing teachers. Figure 1 presents some early evidence of these trends. The teachers exiting at the end of our study window were noticeably more effective than those exiting after IMPACT’s first year (i.e., by about 40% of a teacher-level standard deviation). However, over this same period, the performance of entering teachers also grew appreciably (i.e., 28% of a standard deviation). These trends appear unrelated to the average experience of entering and exiting teachers, which, throughout this period, remained relatively constant at 3.5 and 7 years, respectively. As long as DCPS continues to recruit more able teachers than it loses, compositional change will likely lead to increased student achievement. Whether DCPS can reap further performance benefits from compositional change in its workforce as it increases performance standards appears plausible but remains to be seen. Regardless, our results indicate that, under a robust system of performance assessment, the turnover of teachers can generate meaningful gains in student outcomes, particularly for the most disadvantaged students.

Footnotes

Acknowledgements

We are grateful to the District of Columbia Public Schools (DCPS) for the data employed in this article and to Scott Thompson, Kim Levengood, Alden Wells, and Luke Hostetter of DCPS for addressing our questions regarding the data and IMPACT. We appreciate comments from John Friedman, Steve Glazerman, Hamp Lankford, Luke Miller, and seminar participants at Stanford University, the University of Virginia, Michigan State University, and Vanderbilt, as well as conference attendees at APPAM and AEFP. We also received helpful suggestions from two anonymous reviewers.

Authors’ Note

The views expressed in the article are solely those of the authors and may not reflect those of the District of Columbia Public Schools (DCPS) or the funders. Any errors are attributable to the authors.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: We received financial support for this work from the Institute of Education Sciences Grant R305H140002 and the National Center for the Analysis of Longitudinal Data in Education Research (CALDER). CALDER is supported by IES Grant R305A060018.

Notes

Authors

MELINDA ADNOT, PhD, is a visiting assistant professor of educational policy at Davidson College, North Carolina. Her research interests include educational policy analysis, teacher evaluation, and teacher quality.

THOMAS DEE, PhD, is a professor of education at Stanford University and a research associate in the Programs on Economics of Education, Health Economics and Children at the National Bureau of Economic Research (NBER). His research focuses largely on the use of quantitative methods (e.g., panel data techniques, instrumental variables, and random assignment) to inform contemporary policy debates. Recent examples include econometric evaluations of incentive- and accountability-based reforms and an analysis of recent, stimulus-funded, school-turnaround initiatives.

VERONICA KATZ is a doctoral student in education policy at the University of Virginia. Her research focuses on teacher quality and retention, especially in low-performing schools.

JAMES WYCKOFF is the Curry Memorial Professor of Education and Policy, and director of the EdPolicyWorks, a research center at the University of Virginia. His research focuses on providing evidence for improving teacher quality through a variety of mechanisms including teacher preparation, recruitment, assessment, and retention. Currently, he is working with colleagues to examine the effects of teacher evaluation in the District of Columbia and the ways in which principals improve teacher quality in New York City.

References

Aaronson

Barrow

Sander

(2007). Teachers and student achievement in the Chicago Public High Schools. Journal of Labor Economics, 25, 95–135. doi:10.1086/508733

Boyd

Lankford

Loeb

Rockoff

Wyckoff

(2008). The narrowing gap in New York City teacher qualifications and its implications for student achievement in high-poverty schools. Journal of Policy Analysis and Management, 27, 793–818.

Bryk

Schneider

(2002). Trust in schools: A core resource for improvement. New York, NY: Russell Sage.

Chetty

Friedman

J. N.

Rockoff

J. E.

(2014). Measuring the impacts of teachers II: Teacher value-added and student outcomes in adulthood. American Economic Review, 104, 2633–2679. doi:10.1257/aer.104.9.2633

Clotfelter

C. T.

Glennie

Ladd

H. F.

Vigdor

J. L.

(2008). Would higher salaries keep teachers in high-poverty schools? Evidence from a policy intervention in North Carolina. Journal of Public Economics, 92(5–6), 1352–1370. doi:10.1016/j.jpubeco.2007.07.003

Donaldson

M. L.

Papay

J. P.

(2015). Teacher evaluation for accountability and development. In Ladd

Geortz

(Eds.), Handbook of research in education finance and policy (2nd ed., pp. 174–193). New York, NY: Routledge.

Dee

T. S.

Wyckoff

(2015). Incentives, selection, and teacher performance: Evidence from IMPACT. Journal of Policy Analysis and Management, 34, 267–297. doi:10.1002/pam.21818

Glazerman

Seifullah

(2012). An evaluation of the Chicago Teacher Advancement Program (Chicago TAP) after four years. Princeton, NJ: Mathematica Policy Research.

Goldhaber

Lavery

Theobald

(2015). Uneven playing field? Assessing the teacher quality gap between advantaged and disadvantaged students. Educational Researcher, 44, 293–307.

10.

Guin

(2004). Chronic teacher turnover in urban elementary schools. Education Policy Analysis Archives, 12(42). doi:10.14507/epaa.v12n42.2004

11.

Hanushek

E. A.

(2009). Teacher deselection. In Goldhaber

Hannaway

(Eds.), Creating a new teaching profession (pp. 165–180). Washington, DC: Urban Institute Press.

12.

Hill

C. J.

Bloom

H. S.

Black

A. R.

Lipsey

M. W.

(2008). Empirical benchmarks for interpreting effect sizes in research. Child Development Perspectives, 2, 172–177. doi:10.1111/j.1750-8606.2008.00061.x

13.

Isenberg

Max

Gleason

Potamites

Santillano

Hock

Hansen

(2013, November). Access to effective teaching for disadvantaged students (No. NCEE 2014-4001). Washington, DC: National Center for Education Evaluation, Institute of Education Sciences, U.S. Department of Education. Retrieved from http://ies.ed.gov/ncee/pubs/20144001/pdf/20144001.pdf

14.

Jackson

C. K.

(2012, December). Non-cognitive ability, test scores, and teacher quality: Evidence from 9th grade teachers in North Carolina (Working Paper No. 18624). Cambridge, MA: National Bureau of Economic Research. Retrieved from http://www.nber.org/papers/w18624

15.

Johnson

S. M.

Kraft

M. A.

Papay

J. P.

(2012). How context matters in high-need schools: The effects of teachers’ working conditions on their professional satisfaction and their students’ achievement. Teachers College Record, 114, 100306.

16.

Loeb

Miller

L. C.

Wyckoff

(2015). Performance screens for school improvement: The case of tenure reform in New York City. Educational Researcher, 44, 199–212.

17.

National Council on Teacher Quality. (2012). State of the states 2012: Teacher effectiveness policies–Area 3: Identifying effective teachers. Washington, DC: Author.

18.

Office of the State Superintendent of Education. (2012, February 8). 2011/12 annual enrollment audit overview. Author. Retrieved from http://osse.dc.gov/sites/default/files/dc/sites/osse/publication/attachments/Annual%20Enrollment%20Overview%202011%20to%202012.pdf

19.

Office of the State Superintendent of Education. (2013a). DC CAS math and reading scale scores. Retrieved from http://osse.dc.gov/sites/default/files/dc/sites/osse/publication/attachments/scale_lines3.pdf

20.

Office of the State Superintendent of Education. (2013b). OSSE presentation 2013 DC CAS results (statewide). Retrieved from http://osse.dc.gov/sites/default/files/dc/sites/osse/publication/attachments/OSSE%20Presentation%202013%20DC%20CAS%20Results%20%28Statewide%29-2.pdf

21.

Papay

Bacher-Hicks

Page

Marinell

(2015). The challenge of teacher retention in urban schools: Evidence of variation from a cross-site analysis. Retrieved from https://dx-doi-org.web.bisu.edu.cn/10.2139/ssrn.2607776

22.

Rivkin

S. G.

Hanushek

E. A.

Kain

J. F.

(2005). Teachers, schools, and academic achievement. Econometrica, 73, 417–458. doi:10.1111/j.1468-0262.2005.00584.x

23.

Rockoff

J. E.

(2004). The impact of individual teachers on student achievement: Evidence from panel data. The American Economic Review, 94, 247–252.

24.

Rockoff

J. E.

Jacob

B. A.

Kane

T. J.

Staiger

D. O.

(2011). Can you recognize an effective teacher when you recruit one? Education Finance and Policy, 6, 43–74.

25.

Rockoff

J. E.

Speroni

(2010). Subjective and objective evaluations of teacher effectiveness. The American Economic Review, 100, 261–266.

26.

Ronfeldt

Loeb

Wyckoff

(2013). How teacher turnover harms student achievement. American Educational Research Journal, 50, 4–36. doi:10.3102/0002831212463813

27.

Rothstein

(2015). Teacher quality policy when supply matters. American Economic Review, 105, 100–130. doi:10.1257/aer.20121242

28.

Sass

T. R.

Hannaway

Figlio

D. N.

Feng

(2012). Value added of teachers in high-poverty schools and lower poverty schools. Journal of Urban Economics, 72, 104–122. doi:10.1016/j.jue.2012.04.004

29.

Staiger

D. O.

Rockoff

J. E.

(2010). Searching for effective teachers with imperfect information. Journal of Economic Perspectives, 24, 97–118. doi:10.1257/jep.24.3.97

30.

Thomsen

(2014). Teacher performance plays growing role in employment decisions. Denver, CO: Education Commission of the States.

31.

The New Teacher Project. (2012). The irreplaceables: Understanding the real retention crisis in America’s urban schools. Brooklyn, NY: Author.

32.

Whitehurst

Chingos

Lindquist

(2014). Evaluating teachers with classroom observations: Lessons learned in four districts. Washington, DC: Brookings Institution.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.15 MB