P

A

De

1.

daprculeaanadmsc

Economics of Education Review 36 (2013) 60–72

A

Art

Re

Re

Ac

JEL

I20

J63

Ke

Stu

Sch

Ma

*

1

Ca

Da

Ro

Gli

Co

pri

As

Pro

Stu

gra

02

htt

rincipal turnover and student achievement

shley Miller 1,*

partment of Economics, Mount Holyoke College, 50 College Street, South Hadley, MA 01075-1461, United States

Introduction

School principals play important and varied roles in they to day operations of schools. As instructional leaders,incipals select, monitor and support teachers, designrricula, and manage discipline. As administrativeders, principals set budgets, manage the school facility,d develop relationships with the broader community. Indition to a wide range of responsibilities, principals haveany constituents including students, parents, teachers,hool boards and superintendents (see, for example,

Davis, Darling-Hammond, LaPointe, & Meyerson, 2005;Richardson, Short, & Prickett, 1993). Improving principalquality has become a common focus of school reformefforts: A recent Department of Education report onturning around chronically low-performing schoolsrecommends installing a new principal (Herman et al.,2008) and over the last ten years many states, schooldistricts and non-for-profits have introduced new initia-tives to train and support principals.

Despite the attention currently paid to principals aslevers for school improvement, much remains unknownabout how school leadership affects student learning. Earlywork on how principals affect student achievement hasprimarily focused on the relationship between principalcharacteristics and student test score gains. Both tradi-tional human capital variables such as the education andprior experience of the principal and variables designed tocapture the leadership style of the principal have beenstudied. However, as in the analogous work on teachercharacteristics (Aaronson, Barrow, & Sander, 2007;Hanushek, 1986), the evidence is mixed. Eberts and Stone(1988) find that test score gains are positively correlatedwith the principal’s years of prior teaching experience andadministrative experience, while Brewer (1993) finds no

R T I C L E I N F O

icle history:

ceived 12 December 2011

ceived in revised form 28 March 2013

cepted 28 May 2013

classification:

ywords:

dent achievement

ool leadership

nagerial turnover

A B S T R A C T

Principals have important management roles, including responsibilities for teachers,

curricula and budgets. Schools change principals frequently; about 20% of public school

principals in the United States leave their positions each year. Despite the significance of

principals and the regularity of principal departures, little is known about how turnover

affects schools. Using twelve years of administrative data from North Carolina public

schools, this paper explores the relationship between principal turnover and student

achievement. Principal departures follow a downturn in student performance. Achieve-

ment continues to fall in the two years following the installation of a new principal and

then rises over the next three years. Five years after a new principal is installed, average

academic performance is no different than it was five years before the new principal took

over. Increases in student achievement following a principal transition may reflect mean

reversion rather than a positive effect of principal turnover.

� 2013 Elsevier Ltd. All rights reserved.

Tel.: þ1 412 874 6016.

E-mail address: ashley.ruth.miller@gmail.com.

I appreciate helpful comments from Silvia Barcellos, Leandro

rvalho, Henry Farber, Bo Honore, Stephen Kennedy, Jacob Klerman,

vid Lee, Lars Lefgren, Rob Olsen, Bill Rhodes, Jesse Rothstein and Cecilia

use. Thanks to Debbie Brown of the Cumberland County Schools, Beth

dewell of the Chatham County Schools and Glenda Jones of the Lee

unty Schools for taking time to answer my questions about the

ncipal hiring process in their districts. Financial support from the Abt

sociates Daniel McGillis Development and Dissemination Grant

gram, from the Princeton University Center for Economic Policy

dies, and from the Princeton University Industrial Relations Section is

tefully acknowledged.

Contents lists available at SciVerse ScienceDirect

Economics of Education Review

jo u rn al h om epag e: ww w.els evier .c o m/lo c at e/eco n ed ur ev

72-7757/$ – see front matter � 2013 Elsevier Ltd. All rights reserved.

p://dx.doi.org/10.1016/j.econedurev.2013.05.004

sleatoe

wmTssaMppteappsH

qthwprpnucreInS8ththpetu(BMlo

k

in

th

(1

p

d

e

im

th

tu

H

e

th

te

2

A. Miller / Economics of Education Review 36 (2013) 60–72 61

ignificant effect. And, as Brewer (1993) points out,adership style is likely endogenous since, for example,

student body with an academic focus may cause a school both post large test score gains and to hire principal who

mphasizes academic achievement.More recently, there has been increasing attention on

hether principals are paid for performance and oneasuring the amount of variation in principal quality.

here is evidence that principals are rewarded for strongtudent performance on standardized exams with higheralaries, a lower likelihood of dismissal and principalshipst more desirable schools (Besley & Machin, 2008; Cullen &azzeo, 2008; Hussain, 2007)2, though accountability

olicies are not always effective at rewarding the bestrincipals (Billger, 2007). In line with the findings onacher quality (Nye, Konstantopoulos, & Hedges, 2004)

nd managerial quality (Bertrand & Schoar, 2003), princi-al fixed effects explain a significant portion of variation inrincipal salary (Besley & Machin, 2008) and student testcore gains, particularly for high poverty schools (Branch,anushek, & Rivkin, 2009a3).

Rather than focus on the determinants of principaluality or the amount of variation in principal quality, inis paper I investigate how student performance variesith principal turnover. Principal turnover is a common

henomenon nationwide. Using data from a nationallyepresentative survey, Battle (2010) reports that 21% ofublic school principals left their jobs from one year to theext. Cullen and Mazzeo (2008) reach the same conclusionsing administrative data from Texas. Studies followingohorts of newly hired principals find that about halfemain after four years and 20–40% remain after six (Gatest al., 2006; Papa, 2007; Stoelinga, Hart, & Schalliol, 2008).

the National Center for Education Statistics’ 1999–2000chools and Staffing Survey (SASS), more than half of the524 public school principals surveyed had served for lessan five years in their current position. As shown in Fig. 1,e modal SASS respondent was in her second year as

rincipal, and principals with more than ten years ofxperience leading a school were quite rare. Principalrnover is particularly common at low performing schoolsesley & Machin, 2008; Branch et al., 2009a; Cullen &azzeo, 2008; Fuller, Baker, & Young, 2007), schoolscated in high poverty communities, (Partlow & Ridenour,

2008), and schools with more minority and limited Englishproficiency students (Gates et al., 2006; Papa, 2007).4

Understanding the changes that accompany principalturnover is important not only because it is widespread,but also because any ill effects are disproportionatelyborne by disadvantaged students.

The primary question here is how principal turnoverrelates to student achievement. To date, few large scalestudies address the issue of principal turnover.5 Theexceptions are Weinstein et al. (2009) and Rowan andDenk (1984). Weinstein et al. (2009) focus on principaltransitions in newly formed New York City High Schoolsand consider the relationship between principal turnoverand graduation rates. Years in which schools are led bynew principals are associated with lower graduation rates,and this is particularly true when the school has alreadyexperienced a principal transition. Rowan and Denk (1984)perform a modified GLS analysis of how principaltransitions and school demographics (specifically theproportion of students in the school who received Aid toFamilies with Dependent Children) are related to testscores. These results may be biased by the inclusion of botha school-specific residual component and lagged schoolperformance as a regressor.

Using twelve years of administrative data from NorthCarolina public schools, I follow the method of Jacobson,LaLonde, and Sullivan (1993) to measure student achieve-ment at schools that will undergo a principal transition, areundergoing a principal transition, and have completed aprincipal transition. I take advantage of the panel aspect ofthe data to measure how schools perform relative to their

Fig. 1. Distribution principal experience at current school.

Notes: Data are from the 1999–2000 Schools and Staffing Survey,

administered by the National Center for Education Statistics. The

principals of 8524 public schools responded to the question ‘‘Prior to

this school year, how many years were you employed as the principal of

this school?’’ This is a histogram of their responses.

2 Similar results hold for school superintendents (Ehrenberg, Chay-

owski, & Ehrenberg, 1988; Meier & O’Toole, 2002). For both super-

tendents and principals there is considerable ambiguity about whether

e high scores are actually indicative of high quality. Both Ehrenberg et al.

988) and Hussain (2007) suggest not. Lavy (2008) finds that increasing

rincipal salaries increases the academic performance of students.3 Branch et al. (2009a) note that principals may affect school quality

ifferently at points of their tenure. Their purpose is not to estimate the

ffect of principal turnover, but they recognize that such effects are

pediments to estimating the true variation in principal quality, and

ey make comparisons only among principals with similar job tenure.4 Many other school level factors have been associated with principal

rnover. Large schools experience more turnover than small schools.

igh schools and middle schools experience more turnover than

lementary schools. Urban and rural schools experience more turnover

an suburban schools. Schools with large fractions of uncertified

5 Theoretical frameworks for interpreting principal transitions can be

found in Hargreaves, Moore, Fink, Brayman, and White (2003), Hart

(1991) and Miskel and Cosgrove (1985). Case studies can be found in

Hargreaves et al. (2003), Meyer, Macmillan, and Northfield (2009) and

Weinstein, Jacobowitz, Ely, Landon, and Schwartz (2009). Miskel and

achers also experience excess turnover (Gates et al., 2006; Papa,

007; Young & Fuller, 2009).

Owens (1983) compares schools with and without new principals on

several non-academic measures and finds few differences.

ustracoscfinacdobethlesacstuprrethm

tratraitsenreprAsthonthonThscrotusam

finprm

Ta

De

N

N

N

N

Ed

op

cla

NC

ye

sch

en

A. Miller / Economics of Education Review 36 (2013) 60–7262

ual performance before, during, and after a principalnsition. In line with the previous literature on the

rrelates of principal departure, I find that low performinghools experience more principal turnover. However, Id that this is not driven solely by permanent differencesross schools. In fact, principal departures follow awnturn in school performance, a downturn whichgins some four years before the new principal takese helm. It may be that principals who plan to quit works in the years preceding departure, causing student

hievement to suffer. It is also possible that a drop indent achievement is a cause of some instances of

incipal turnover, either because principals quit insponse to the unpleasantness of working at a schoolat is trending downwards, or because falling test scoresake it more likely that a principal is fired.

The drop in school performance before the principalnsition complicates inference on the causal effect of thensition. Since the typical school is doing badly relative to

usual performance before a new principal starts, it istirely possible that the school would have experienced acovery to its usual performance regardless of whether theincipal was replaced. This phenomenon, known as thehenfelter dip, was first described in Ashenfelter (1978) ine context of estimating the effects of job training programs

wages. De Paola and Scoppa (2008) note its importance ine context of assessing the effects of managerial turnover

performance, albeit in soccer teams rather than schools.is paper is the first to document the Ashenfelter dip in testores before principal transitions. The test score dip is quitebust: Scores drop before a principal transition even in lowrnover schools (with only a single turnover during themple period) and the dip occurs in both elementary andiddle schools.

In line with Branch, Hanushek, and Rivkin (2009b), Id that test scores in the first two years of the newincipal’s tenure are low relative to both school perfor-

performance under the new principal. Performance thenstarts to rebound, returning to pre-transition levels by theend of the fourth year since the principal’s departure.While test scores do eventually increase following princi-pal transitions, this work suggests that the increase in testscores may represent a return to the school’s typicalperformance rather than a positive causal effect ofprincipal turnover.

2. Institutions, data and summary statistics

The main data source is administrative data from publicschools in North Carolina. These data are distributed by theNorth Carolina Education Research Data Center (NCED) and Ihave supplemented the data with additional informationfrom the National Center for Education Statistics’ CommonCore of Data (CCD).6 The NCED data span twelve academicyears, from the 1994–1995 school year through the 2005–2006 school year. During this period, the number of NorthCarolina public schools listed in the CCD rose from 1968 to2348. As shown in Table 1, the sample is restricted to schoolsthat were open (with students enrolled and teachersemployed) for all twelve years.

Given the focus on the correlates of principal turnover,the sample is further restricted to schools in which it ispossible to reliably pin down the timing of principaltransitions. The data contain annual employment fileslisting all principals.7 I match these files across years tomeasure the timing of principal transitions. For some

ble 1

finition of samples.

Number % of universe

(1) (2)

orth Carolina public schools in the 1994–1995 CCD 1968 100.0%

and remain open through 2005–2006 1779 90.4%

and principal transition data 979 49.7%

orth Carolina public primary schools in the 1994–1995 CCD 1170 100.0%

and remain open through 2005–2006 1070 91.5%

and principal transition data 623 53.2%

and test score data (1994–1995 through 2005–2006) 554 47.4%

orth Carolina public middle schools in the 1994–1995 CCD 413 100.0%

and remain open through 2005–2006 364 88.1%

and principal transition data 179 43.3%

and test score data (1996–1997 through 2004–2005) 127 30.8%

otes: Common Core of Data (CCD) information comes from the National Center for Education Statistics’ website, www.nces.ed.gov. North Carolina

ucation Research Data Center (NCED) data is documented online at http://pubpol.duke.edu/centers/child/ep/nceddatacenter/. A school is counted as

en if it enrolls a positive number of students and employs a positive number of teachers (both in the CCD and in the NCED data) in each year. Schools are

ssified as primary, middle, high school or other based on their designation in the 1994–1995 CCD. Principal information and test scores are from the

ED. A school has principal transition data if the school either has a single principal each year or, if a school has two principals, A and B, listed in a given

ar, the school is included if the school has only principal A listed in year N � 1 and only principal B in year N + 1. This restriction eliminates, for example,

ools with two principals in the first or last year of the data, and schools missing a principal at any point. Test scores for primary schools are state-wide

d-of-grade exams for 3rd and 4th graders and test scores for middle schools are end-of-grade exams for 6th, 7th and 8th graders.

6 North Carolina Education Research Data Center data is documented

online at http://pubpol.duke.edu/centers/child/ep/nceddatacenter/. The

Common Core of Data is available online at the National Center for

Education Statistics’ website, http://www.nces.ed.gov/.7 For the purposes of this analysis, principals are the employees who

were paid to be a ‘‘principal/headmaster’’. The employment files include

arate budget codes for other types of administrators (e.g., super-

endents and assistant principals).

ance under the old principal and to subsequent schoolsep

int

sssToinAth(p

ss1Mletwpdvtro

rmB

F

N

w

1

tr

N

fi

w

re

fo

y

si

a

fi1

a

tr

(2

A. Miller / Economics of Education Review 36 (2013) 60–72 63

chools, there is a year when no principal is listed. Thesechools are excluded from the sample. Similarly, there arechools that employ multiple principals in a single year.hese schools are excluded from the sample as well, withne exception: If a school had two principals, A and B, listed

a given year t, it is included in the sample if only principal was listed in year t � 1 and only principal B in year t + 1.8 Inis case, the school is listed as having a new principalrincipal B) in year t. The final sample contains 979 schools.Principal turnover is common in the NCED data. Fig. 2

hows the distribution of job tenure for principals whotarted work at a North Carolina public school during the995–1996, 1996–1997 and 1997–1998 school years.ore than half of these principals left their schools inss than five years.9 As summarized in Table 2, during theelve school years analyzed here schools underwent 1.96

rincipal transitions (and so were led by roughly threeifferent principals) on average. There was substantialariation across schools in the number of principalansitions; some schools had a single principal whilethers had seven principals.10

North Carolina state statutes provide some guidelinesegarding the hiring and firing of principals. Principals

ust be certified as school administrators by the Stateoard of Education and evaluated on an annual basis.

Principals hired by a school district must be appointed bythe district’s board of education based on the recommen-dation of the district’s superintendent. Conversations withhuman resources officials in several North Carolina schooldistricts suggest that typically vacancies are postedthrough a state-wide system. Initial screening of candi-dates is done by an an interview team consisting of districtadministrators as well as teachers, staff and parents at theschool with the opening. The interview team thenrecommends finalists to the superintendent.

When a principal starts working in a school district forthe first time, the contract governing employment condi-tions must be for between two and four years, andsubsequent contracts must last four years. Since contractcan be renegotiated each year, the primary effect of the fouryear rule is that no single contract lasts for more than fouryears. As shown in Fig. 2, there are principal departures atevery year of tenure, and there is no sharp increase indepartures after four years of tenure. At the end of threeyears, principals are eligible for career administrator status.Administrators under contract with career status may not bedismissed without cause (such as inadequate performance,neglect of duty or immorality). Conversations with humanresources officials suggest that outright dismissals are veryrare and non-renewal of contracts more common.

School performance is measured using student scoreson annual state-wide exams. Currently North Carolinapublic school students in 3rd through 8th grade takereading comprehension and math exams at the end of eachschool year, however only 3rd and 4th grade test scores areavailable for all 12 years of the sample period.11 For thisreason, much of the analysis is restricted to primaryschools.

School-year level measures of academic achievementare constructed from normalized student test scores.Specifically, students who are missing test scores areexcluded and scores for the remaining students are thennormalized (separately by grade, subject and year) to bemean zero and standard deviation one. I then compute theaverage 3rd grade and 4th grade reading and math scoresof all the test takers at a given school, and average the fouraverages to produce a single school level score.12 Middleschool test scores, formed in the same way (but with 6th,7th and 8th grade scores in place of 3rd and 4th gradescores) are available from the 1996–1997 school year

ig. 2. Distribution of principal tenure.

otes: This is a histogram of principal tenure for principals who started

orking in North Carolina public schools during the 1995–1996, 1996–

997 and the 1997–1998 school years. All 979 schools with principal

ansition data are included. For schools where principal A works in years

� 1 and N and principal B works in years N and N + 1, principal A’s

nal year is counted here as N � 1, the last year in which principal A

as the sole principal. The rightmost bar includes all principals who

mained on the job for nine or more years.

Table 2

Average number of principal transitions across samples.

Sample Mean SD Min Max N

All schools 1.96 1.08 0 6 979

Primary school sample 1.89 1.05 0 6 554

Middle school sample 2.19 1.12 0 6 127

Notes: This table summarizes the number of times sample schools

changed principals over the course of the 12 school years between 1994–

1995 and 2005–2006, by school type.

8 This restriction eliminates, for example, schools with two principals

r two years in a row and schools with two principals in the first or last

ears of the data. All schools with three or more principals employed in a

ngle year are excluded.9 As noted above, some schools have two principals listed in a single

cademic year. In these cases, it is not possible to determine whether the

rst principal left in October or May.0 In line with the results of Cullen and Mazzeo (2008), Gates et al. (2006)

nd Young and Fuller (2009), primary schools experienced fewer principal

11 More information about these exams is available online at http://

www.dpi.state.nc.us/accountability/testing/eog/.12

ansitions (1.89 on average) than middle schools (2.19) or high schools

.05), with both differences statistically significant at the 10% level.

Taking the first principal component of the four scores yields similar

results.

thsa

scacapsc3rin3rstuth

thsiznoluthwadsaantesm

3.

3.1

pr

Ta

Su

A

T

%

T

%

P

T

%

T

%

3

A

A

M

T

%

T

%

6

A

A

No

sch

lun

sli

the

ye

of

gra

sch

A. Miller / Economics of Education Review 36 (2013) 60–7264

rough the 2004–2005 school year, three-quarters of themple years.Not all students took the state exams, and the potential

ores of excluded students might give a different picture ofademic achievement in some schools. This does notpear to be a significant problem in North Carolina publichools. For example, in the 2000–2001 school year, averaged grade enrollment in sample primary schools, as reported

the CCD, was 83.33 students, while the average number ofd grade test takers with valid scores in the NCED was 79.95dents. While the number of students enrolled is greater

an the number of test takers, the difference is small.Table 3 shows summary statistics for sample schools in

e 2000–2001 school year, including measures of schoole (number of students and teachers), student socioeco-mic status (percentage of the students eligible for free

nch), and teacher retention (the percentage of teachers ate school in the 2000–2001 school year who return toork at the same school in the 2001–2002 school year) indition to academic achievement. Relative to the fullmple, the primary school sample (the focus of the mainalysis) is similar in terms of free lunch eligibility andacher retention but the sample primary schools arealler, with fewer teachers and students.

Results

. Academic performance around principal transitions

I begin by describing the academic performance of

existing literature on principal transitions, student testscores are substantially lower at schools with newprincipals. Column (1) of Table 4 shows the difference inperformance between school-years with a new principaland school-years without a new principal. While there is asignificant gap in test scores, about half of the poorperformance can be accounted for by school fixed effects,as shown in column (2). Principal turnover is morecommon at low performing schools.

This is further illustrated in Table 5: Relative toschools with zero or one principal transitions, schoolswith two principal transitions have lower test scores onaverage. Schools with three or more principal transitionshave still lower scores. Schools with more principalturnover also have a higher fraction of students eligible

ble 3

mmary statistics for the 2000–2001 school year.

Mean SD Min Max N

ll schools

otal number of students (CCD) 618 335 6 2649 979

students free lunch eligible (CCD) 36% 20 0% 95% 929

otal number of teachers (NCED) 39 21 6 154 979

teachers who return (NCED) 79% 11 16% 100% 979

rimary school sample

otal number of students (CCD) 505 187 105 1078 554

students free lunch eligible (CCD) 39% 19 2% 92% 526

otal number of teachers (NCED) 31 10 6 61 554

teachers who return (NCED) 79% 12 16% 100% 554

rd grade test takers (NCED) 80 35 10 294 554

verage 3rd grade math score (NCED) �0.01 0.37 �1.10 1.06 554

verage of average 3rd/4th math/reading scores (NCED) �0.01 0.33 �0.88 0.92 554

iddle school sample

otal number of students (CCD) 674 249 56 1258 127

students free lunch eligible (CCD) 36% 16 7% 86% 123

otal number of teachers (NCED) 44 14 10 76 127

teachers who return (NCED) 76% 10 44% 94% 127

th grade test takers (NCED) 213 88 9 414 127

verage 6th grade math score (NCED) �0.05 0.33 �1.50 0.71 127

verage of average 6th/7th/8th math/reading scores (NCED) �0.03 0.29 �1.46 0.66 127

tes: Means are unweighted averages across schools. The ‘‘all schools’’ sample (979 schools) includes the primary school sample (554 schools), the middle

ool sample (127 schools) and other sample schools. ‘‘Total number of students (CCD)’’ is total enrollment in the 2000–2001 school year. ‘‘% students free

ch eligible (CCD)’’ is the ratio of the number of students eligible for free lunch to the total enrollment in the 2000–2001 school year. Sample sizes are

ghtly reduced for free lunch eligibility because this data is not available for all schools in the 2000–2001 school year. ‘‘Total number of teachers (NCED)’’ is

number of teachers in the 2000–2001 school year. ‘‘% teachers who return (NCED)’’ is the percentage of teachers at the school in the 2000–2001 school

ar who return to teach in the same school in the 2001–2002 school year. ‘‘3rd grade test takers (NCED)’’ and ‘‘6th grade test takers (NCED)’’ are the number

3rd and 6th grade students with valid test scores. School-year test scores are formed by first standardizing individual scores (separately by subject and

de) across test takers to be mean zero and standard deviation one, and then averaging the standardized scores across students at a school to form the

ool subject-grade score. Since the normalization is done at the student level and the school level averages are unweighted, mean scores need not be zero.

Table 4

Cross sectional estimates of the effect of principal transitions.

(1) (2)

New principal �0.0409 �0.0182

(0.0093)*** (0.0054)***

Constant �0.0050 �0.0086

(0.0117) (0.0021)***

Adjusted R2 0.0021 0.7582

Observations 6648 6648

School fixed effects No Yes

Notes: Each column reports coefficients from a separate OLS regression.

The dependent variable is the average elementary school test score in a

given year. N = 6648 = 12 years � 554 schools. Standard errors, clustered

the school level, are shown in parenthesis.

** p < 0.001.

imary school with new principals. In line with theat

*

fotsfrtsle

msbCsinresn

T

S

2

T

O

fr

fu

y

te

A. Miller / Economics of Education Review 36 (2013) 60–72 65

r free lunch and a lower fraction of the teachers returno the same school to teach in the next academic year, ashown in Table 6. Across schools, those which undergoequent leadership changes are more disadvantaged (in

erms of student achievement, student socioeconomictatus, and teacher retention) than those with stableadership.

To assess the relationship between student perfor-ance and principal turnover within schools over time, I

tart with a flexible description of the relationshipetween principal turnover and student test scores.onsider the determinants of school test scores, yitk, forchool i in year t. Let wi represent the effect of the timevariant characteristics of school i on performance and gt

epresent factors that affect all schools in year t. (Yearffects are useful here because not all North Carolina publicchools are included in the sample but test scores areormalized over all test takers, so the average test scores

within the sample may differ across years.) The primaryconcern here is documenting differences in school out-comes around principal transitions. To allow for thispossibility, let dk represent the net effect of having aprincipal with k years of experience.

I focus on primary school test scores which are availablefor the entire sample period and use the strategy adoptedby Jacobson et al. (1993) to measure earnings losses fordisplaced workers. Specifically, I estimate

yitk ¼ wi þ gt þX

m � k � �m

Ditkdk þ eitk (1)

where {Ditk|m � k � � m} are indicator variables forprincipal tenure. For positive k, Dstk is an indicator thatprincipal tenure is k years; Ditk is one if school i got anew principal k years before year t. For negative k, Ditk isone in year t if school i will change principals �k yearslater.

able 5

chool-level academic achievement by number of principal transitions.

Academic achievement (NCED) Mean SD Min Max N

Schools with zero principal transitions 0.02 0.21 �0.45 0.61 38

Schools with one principal transition 0.03 0.28 �0.78 0.78 173

Schools with two principal transitions �0.02 0.26 �0.65 0.67 199

Schools with three or more principal transitions �0.05 0.31 �0.71 0.82 144

All primary schools �0.01 0.28 �0.78 0.82 554

Notes: Academic achievement is the school-level average of the school-year average test scores over the full sample period (1994–1995 through 2005–

006).

able 6

ther school characteristics by number of principal transitions.

Mean SD Min Max N

Student measures

Total number of students (CCD)

Schools with zero principal transitions 651 357 190 2038 65

Schools with one principal transition 601 324 10 2216 282

Schools with two principal transitions 628 325 79 2266 354

Schools with three or more principal transitions 630 320 60 2061 278

All schools 622 325 10 2266 979

% students free lunch eligible (CCD)

Schools with zero principal transitions 34% 16 5% 69% 61

Schools with one principal transition 37% 18 5% 83% 264

Schools with two principal transitions 36% 18 4% 89% 331

Schools with three or more principal transitions 37% 19 4% 85% 267

All schools with free lunch data in every year 37% 18 4% 89% 923

Teacher measures

Total number of teachers (NCED)

Schools with zero principal transitions 39 21 13 120 65

Schools with one principal transition 38 19 8 130 282

Schools with two principal transitions 40 20 10 143 354

Schools with three or more principal transitions 39 19 10 135 278

All schools 39 20 8 143 979

% teachers who return (NCED)

Schools with zero principal transitions 81% 7 56% 92% 65

Schools with one principal transition 80% 6 57% 94% 282

Schools with two principal transitions 79% 7 50% 93% 354

Schools with three or more principal transitions 77% 7 51% 91% 278

All schools 79% 7 50% 94% 979

Notes: ‘‘Total number of students (CCD)’’ is the school-level average of the school-year total enrollment as over the full 12 year sample period. ‘‘% students

ee lunch eligible (CCD)’’ is the school-level average of the school-year ratio of the number of students eligible for free lunch to the total enrollment over the

ll 12 year sample period. This data is not available for all schools in all years. ‘‘Total number of teachers (NCED)’’ is the school-level average of the school-

ear number of teachers over the full 12 year sample period. ‘‘% teachers who return (NCED)’’ is the school-level average of the school-year percentage of

achers who return to teach in the same school in the following school year over 11 years (all but the last year of the full sample period).

mFoyecaanprscmneprassuarpespadprmor

esunpetioburevahatwtraeasc20ona

satrarescef

13

eff

see

at

ex

tur

ma14

pri

to

est

ch15

we

ha16

tra

tra

me

sch

sli

A. Miller / Economics of Education Review 36 (2013) 60–7266

The values of the dk summarize how school perfor-ance differs before, during and after principal transitions.r negative k, dk measures student achievement in thears leading up to a principal transition. For positive k, dk

ptures both the same underlying trends in performanced any causal effects of principal turnover. A newincipal is less familiar with the students, teachers andhool community and this loss of institutional knowledgeay make the causal effect of short principal tenuregative. On the other hand, a new principal may approachoblems with fresh insight and teachers may work harder

they adapt to management policies of the new principal,ggesting a positive causal effect.13 The choice of m isbitrary but should be large enough to capture changes inrformance around a principal transition.14 In the mainecification, dk are estimated for 4 � k � �4, with anditional dummy for five or more years since the newincipal assumed leadership. All principal tenure esti-ates are reported relative to the omitted category of five

more years before the principal transition.While in Jacobson et al. (1993), worker earnings were

timated around a single incident of job loss, most schoolsdergo multiple principal transitions during the sampleriod. Three strategies are used to address this complica-n. First, in the main specification, all schools are includedt scores for schools with multiple transitions are

plicated and included multiple times. School-year obser-tions are included in the regression just once if the schools zero or one principal transition, twice if the school haso principal transitions, etc. A school with two principalnsitions is treated as though it were two different schools

ch with a single principal transition. If, for example, ahool gets a new principal in 1998 and in 2000, the year00 test score for that school enters the regression twice,ce with a two years post dummy turned on and once withyear of the transition dummy turned on.15 Second, themple is restricted to schools with at most one principalnsition. The sample size is substantially smaller but these

sults are simpler to interpret. Third, data from eachhool-year enters the regression once but multiple tenurefects may additively affect a single score.16 All three

strategies yield similar conclusions about school perfor-mance around principal transitions.

In each specification, test scores fall in the last fouryears of the outgoing principal and continue to fall in thefirst two years of the new principal. Scores then begin torise. By the time the new principal has been in place for fiveor more years, scores rise back to baseline levels. Resultsfor the main specification are plotted in Fig. 3. By the yearbefore the new principal starts working at the school,scores average 0.014 lower than the baseline score at fiveor more years before the transition. To put this number incontext, the average within-school standard deviation ofthe test score measure over the course of the twelve yearsample period is 0.142, so this represents a drop of about10% of the typical variation in scores. The school level testscore is an average of student level test scores, so one wayfor this drop to occur is if every student at the school scored0.014 lower on an exam with mean zero and standarddeviation one. Test scores continue to drop, on average,when the new principal starts working at the school,falling, on average to 0.021 below the school’s baseline atthe end of the new principal’s first year, and 0.025 belowthe school’s baseline at the end of the new principal’ssecond year. (All three of these estimates for the dk arestatistically significantly different from zero at the 5%level.) Starting at the end of the third year of the newprincipal, the trend reverses and scores begin to rise. In themain specification, at the end of the fourth year since the

Fig. 3. Average of elementary school test scores as the principal changes.

Notes: This is a plot of the coefficients from a regression of the average of

average 3rd and 4th grade math and reading scores on a set of indicator

variables for years before or since a principal transition. Five or more

years before the principal change is the omitted category. The regression

includes school and year fixed effects. The 95% confidence interval is

formed from standard errors clustered at the school level. The regression

includes 554 primary schools and 12 years of test scores data (1994–1995

through 2005–2006). An observation is a school-year-transition: School-

year observations are included in the regression just once if the school has

zero or one principal transition, twice if the school has two principal

transitions, etc. N = 13,008 = 12 years � 1084 where the number of

school transitions, 1084 = 211 schools with zero or one

transitions + 2 � 199 schools with two transitions + 3 � 108 schools

with three transitions + 4 � 30 schools with four transitions + 5 � 5

schools with five transitions + 6 � 1 school with six transitions . If, for

example, a school gets a new principal in 1998 and in 2000, the year 2000

test score for that school will enter the regression twice, once with a two

years post dummy turned on and once with a year of the transition

dummy turned on.

A natural extension is to allow for the possibility that the sign of the

ect of turnover varies with the performance of the school, since it

ms likely that the loss of institutional knowledge is a bigger problem

schools that are performing well prior to the principal transition. I

plore this possibility empirically but find no evidence that the

nover effect varies in a systematic way with prior school perfor-

nce.

The sample period is 12 years long and it is possible to identify new

ncipals starting in the second year of data, so it is feasible to identify up

eleven leads and lags. However, there is limited data available to

imate long leads and lags and varying the span does not significantly

ange the reported coefficients.

Replicated data may be downweighted so that each school has equal

ight in the regression regardless of how many principal transitions it

s undergone. In practice, this does not significantly change the results.

There are many other possible strategies (such as only considering

nsitions that take place in a single academic year, using the first

nsition each school undergoes, or creating interaction terms to

asure whether principal transitions effect schools differently if the

ool has just experienced a principal transition) available, each yielding

ghtly different interpretations of the estimated tenure effects.

nafoth

CFwetrforHlewsthessth

astsbn

T

T

N

a

(2

T

y

e

tr

tr

p

a

A. Miller / Economics of Education Review 36 (2013) 60–72 67

ew principal, scores are only 0.004 below baseline scores,nd five or more years after the beginning of employmentr the new principal, scores are actually higher on averagean the baseline measure.Results for all three specifications are shown in Table 7.

olumn (1) reports the same results shown graphically inig. 3 and discussed above. Column (2) reports the resultshen schools with multiple principal transitions are

xcluded. The achievement pattern around principalansitions is the same, and the estimated coefficientsr the years immediately around principal transitions are

oughly a third larger than in the main specification.owever, the sample size is smaller and the coefficients aress precisely estimated. Column (3) reports the resultshen multiple principal tenure effects additively affect a

ingle score. As in the restricted regression of column (2),e trends exactly match the main specification but the

stimated coefficients are even larger, and performance istatistically significantly below typical performance for thechool from three years before the new principal until the

ird year of the new principal.Rather than allowing scores to vary completely flexibly

round principal transitions, it is also possible to fit testcores to linear trends before and after principal transi-ions. The same result appears in this more parsimoniouspecification, as shown in Table 8. Scores are slightlyelow baseline and falling during the four years before theew principal. During the first four years of the new

principal, scores are significantly lower than usual (by0.034 points relative to baseline scores) but rising, onaverage each year (by 0.006) suggesting about six yearsuntil a full return to baseline scores. Different methods ofincluding schools with multiple principal transitionsagain yield the same conclusions, as shown in columns(2) and (3).

One important caveat is that these regressions includevery few covariates. In particular, there may be sizableprincipal effects other than those due to principal tenure.Some principals may be permanently better at raisingstudent test scores than other principals and there may beconsiderable variation in the quality of principal-schoolmatches. If permanent principal quality is correlated withprincipal tenure, this will bias the estimated coefficients onprincipal tenure. For example, if high quality principalstend to stay in place and low quality principals tend to beshuffled from school to school, as described in Branch et al.(2009a), the effects of long principal tenure are a mix of thecausal effect of long principal tenure and the high averagequality of survivor principals. However, this phenomenonis not driving the estimates here: The results from singletransition schools, where principals have significantlylonger average tenures (shown in column (2) of Tables 7and 8) are very similar to the results for the full sample.

Thus far, the analysis has focused on elementaryschools. North Carolina End Of Grade Tests are given to3rd-8th grade students, and test score data for middle

able 7

est score changes before and after principal transitions.

(1) (2) (3)

Four years before new principal �0.007 �0.011 �0.013

(0.005) (0.016) (0.007)

Three years before new principal �0.010 �0.025 �0.019*

(0.006) (0.022) (0.009)

Two years before new principal �0.016* �0.035 �0.030**

(0.006) (0.024) (0.010)

The year before new principal �0.014* �0.024 �0.030**

(0.007) (0.027) (0.010)

First year of the new principal �0.021*** �0.036 �0.041***

(0.006) (0.029) (0.010)

Second year of the new principal �0.025*** �0.046 �0.046***

(0.007) (0.032) (0.010)

Third year of the new principal �0.017** �0.025 �0.034**

(0.007) (0.035) (0.011)

Fourth year of the new principal �0.004 �0.026 �0.016

(0.006) (0.039) (0.010)

Five or more years of the new principal 0.005 �0.003 �0.003

(0.005) (0.043) (0.014)

Adjusted R2 0.774 0.752 0.761

N 13,008 2532 6648

School and year fixed effects Yes Yes Yes

otes: Each column reports coefficients from a separate OLS regression. The dependent variable is the school-year test score. All regressions include school

nd year fixed effects. Standard errors clustered at the school level are shown in parenthesis. Columns (1) and (3) include all 554 primary schools. Column

) is restricted to the 211 primary schools without multiple principal transitions. There are 12 years of test scores data (1994–1995 through 2005–2006).

he omitted category is five or more years before the principal transition. In column (1), an observation is a primary school school-year-transition: School-

ear observations are included in the regression just once if the school has zero or one principal transition, twice if the school has two principal transitions,

tc. N = 13, 008 = 12 years � 1084 where the number of school-transitions, 1084 = 211 schools with zero or one transitions + 2 � 199 schools with two

ansitions + 3 � 108 schools with three transitions + 4 � 30 schools with four transitions + 5 � 5 schools with five transitions + 6 � 1 school with six

ansitions In column (2), the regression is restricted to primary schools with zero or one transitions, so N = 2532 = 12 years � 211 since there are 211

rimary schools with zero or one transitions. In column (3), each school-year observation is included once and principal tenure effects are included

dditively. N = 6648 = 12 � 554.

* p < 0.05.

** p < 0.01.

*** p < 0.001.

Table 8

Test score changes before and after principal transitions.

(1) (2) (3)

Intercept for the four years before the new principal �0.006 �0.012 0.001

(0.006) (0.018) (0.009)

Trend in the last four years of the old principal �0.003 �0.005 �0.007*

(0.002) (0.007) (0.003)

Intercept for the first four years of the new principal �0.034*** �0.046 �0.045***

(0.007) (0.029) (0.011)

Trend in the first four years of the new principal 0.006** 0.005 0.005

(0.002) (0.007) (0.003)

Intercept for five or more years of the new principal 0.002 �0.002 0.003

(0.004) (0.043) (0.013)

Adjusted R2 0.774 0.752 0.760

N 13,008 2532 6648

School and year fixed effects Yes Yes Yes

Notes: Each column reports coefficients from a separate OLS regression. The dependent variable is the school-year test score. All regressions include school

and year fixed effects. Standard errors clustered at the school level are shown in parenthesis. Columns (1) and (3) include all 554 primary schools. Column

(2) is restricted to the 211 primary schools without multiple principal transitions. There are 12 years of test scores data (1994–1995 through 2005–2006).

The omitted category is the intercept for five or more years before the principal transition. In column (1), an observation is a primary school school-year-

transition: School-year observations are included in the regression just once if the school has zero or one principal transition, twice if the school has two

principal transitions, etc. N = 13, 008 = 12 years � 1084 where the number of school-transitions, 1084 = 211 schools with zero or one transitions + 2 � 199 -

schools with two transitions + 3 � 108 schools with three transitions + 4 � 30 schools with four transitions + 5 � 5 schools with five transitions + 6 � 1 -

school with six transitions In column (2), the regression is restricted to primary schools with zero or one transitions, so N = 2532 = 12 years � 211 since

there are 211 primary schools with zero or one transitions. In column (3), each school-year observation is included once and principal tenure effects are

included additively. N = 6648 = 12 � 554.

* p < 0.05.

** p < 0.01.

*** p < 0.001.

Fig. 4. Average of middle school test scores as the principal changes.

Notes: This is a plot of the coefficients from a regression of the average of

average 6th, 7th and 8th grade math scores at a school in a given year on a

set of indicator variables for years before or since a principal transition. Five

or more years before the principal change is the omitted category. The

regression includes school and year fixed effects. The 95% confidence

interval formed from standard errors clustered at the school level. The

regression includes 127 middle schools and 9 years of test scores data

(1996–1997 through 2004–2005). An observation is a school-year-

transition: School-year observations are included in the regression just

once if the school has zero or one principal transition, twice if the school has

two principal transitions, etc. N = 2529 = 9 years � 281 where the

number of school-transitions, 281 = 37 schools with zero or one

transitions + 2 � 46 schools with two transitions + 3 � 29 schools with

three transitions + 4 � 11 schools with four transitions + 5 � 3 schools with

five transitions + 6 � 1 school with six transitions . If, for example, a school

gets a new principal in 1998 and in 2000, the year 2000 test score for that

school will enter the regression twice, once with a two years post dummy

turned on and once with a year of the transition dummy turned on.

−.5

0.5

11.

5

−4 −3 −2

Old Prin

cipal

New P

rincip

al 2 3 4Afte

r

Years Before/Since Transition

95% CI Coefficient on Dummy

Fig. 5. % Students free lunch eligible as the principal changes.

Notes: This is a plot of the coefficients from a regression of the percentage

of students in a school in a given year who are eligible for free lunch on a

set of indicator variables for years before or since a principal transition.

This measure is generally available for the full 12 year sample period, but

a few school-years are missing. Five or more years before the principal

change is the omitted category. The regression includes school and year

fixed effects. The 95% confidence interval formed from standard errors

clustered at the school level. The regression includes all sample schools.

An observation is a school-year-transition: School-year observations are

included in the regression just once if the school has zero or one principal

transition, twice if the school has two principal transitions, etc. In the

absence of missing data on free lunch eligibility, there would be N =

23,820 = 12 years � 1985 where the number of transitions, 1985 =

347 schools with zero or one transitions + 2 � 354 schools with two

transitions + 3 � 199 schools with three transitions + 4 � 64 schools with

four transitions + 5 � 13 schools with five transitions + 6 � 2 schools with

six transitions. 179 school-year-transitions observations are lost due to

missing data on free lunch eligibility.

A. Miller / Economics of Education Review 36 (2013) 60–7268

sssoefoinasthrdstrtr

3

oadpofitrsthptr

T

%

a

in

le

a

th

th

A. Miller / Economics of Education Review 36 (2013) 60–72 69

chool students is available for three quarters of theample, from the 1996–1997 school year to the 2004–2005chool year. I use the same flexible estimation techniquen the available years of middle school test scores. In thisntirely distinct sample the same pattern appears: Resultsr middle schools from the main specification are shown Fig. 4. As in the elementary school case, test scores fall on

verage in the years preceding the principal transition. Thecores are lowest in the two years immediately following

e departure of the principal, and scores then begin toebound, returning to the pre-transition mean. The mainifference between the elementary school and the middlechool results is that the drop in scores preceding aansition only begins to occur two years before principalansition for middle schools.

.2. Mechanisms

One possible explanation for these findings is thatutside factors drive changes in school level studentchievement, and that if student achievement continues torop over several years, this increases the likelihood of arincipal turnover. If this is correct, the permanence of theutside factors should influence the interpretation of thesendings. For example, if the typical reason for a pre-ansition dip in performance is that a difficult cohort of

tudents enters the school and gradually works its wayrough the system, a return to the usual school

erformance seems likely whether or not a principalansition occurs. In contrast, if the typical reason for a

pre-transition dip in performance is due to a permanentinflux of needy students then there is no reason to expect areturn to the school mean in the absence of some type ofintervention.

To explore these possibilities, I examine how studentsocioeconomic status, as measured by CCD data on thefraction of students who are eligible for free lunch, varieswith principal transitions. As shown in Fig. 5, there is asmall uptick in the fraction of students eligible for freelunch in the year preceding a principal transition and theserates stay somewhat elevated in the first few posttransition years. However, as shown in Table 9, there isno definitive trend. The same pattern does hold for theschools without multiple principal transitions (column (2))and for the primary school sample (column (3)). But theresults are generally not statistically significantly differentfrom zero, and in the subsample of primary schoolswithout multiple principal transitions (column (4)), thefraction of students eligible for free lunch is actually lowerin the second year of the new principal than in thereference period of five or more years before the principaltransition. These findings do not provide evidence thatsubstantial changes in student socioeconomic status drivethe achievement drops that precede principal transitions.

I next examine how teacher retention varies withprincipal turnover. Many of the ways that principals affectstudents are indirect. To change what students learn,principals need to change what goes on inside theclassroom so one important channel through whichprincipals affect students is selecting, monitoring and

able 9

students free lunch eligible before and after principal transitions.

All schools Primary schools

(1) (2) (3) (4)

Four years before new principal �0.033 0.616 �0.060 0.212

(0.176) (0.738) (0.255) (0.904)

Three years before new principal 0.308 1.640 0.682* 1.226

(0.230) (0.954) (0.281) (1.066)

Two years before new principal �0.031 1.257 �0.150 0.375

(0.241) (1.012) (0.337) (1.357)

The year before new principal 0.558* 2.039 0.647* 1.061

(0.238) (1.115) (0.315) (1.403)

First year of the new principal 0.394 1.022 0.527 0.037

(0.239) (1.271) (0.294) (1.543)

Second year of the new principal 0.471 0.877 0.393 �1.332

(0.271) (1.363) (0.349) (1.772)

Third year of the new principal 0.475 1.893 0.539 0.894

(0.284) (1.552) (0.346) (1.872)

Fourth year of the new principal 0.685* 0.552 0.693* �1.817

(0.268) (1.729) (0.341) (2.178)

Five or more years of the new principal �0.094 1.098 �0.250 �1.105

(0.179) (1.929) (0.233) (2.463)

Adjusted R2 0.817 0.767 0.810 0.793

N 23,641 4117 12,925 2514

School and year fixed effects Yes Yes Yes Yes

Notes: Each column reports coefficients from a separate OLS regression. The dependent variable is the percentage of students who are free lunch eligible in

given school in a given year. This measure is available for all 12 sample years (though not in every year for every school.) The omitted category is the

tercept for five or more years before the principal transition. All regressions include school and year fixed effects. Standard errors clustered at the school

vel are shown in parenthesis. Columns (1) and (2) include all school levels while columns (3) and (4) are restricted to primary schools. Columns (2) and (4)

re restricted to schools without multiple principal transitions (e.g., column (4) consists of an analysis of primary schools led by zero or one principals over

e sample period). In columns (1) and (3), an observation is a school-year-transition: School-year observations are included in the regression just once if

e school has zero or one principal transition, twice if the school has two principal transitions, etc.

* p < 0.05.

susigits

thyetra1.3yeusThsigtyesthtracasc(2wevm

Ta

% t

F

T

T

T

F

S

T

F

F

A

N

S

N

ret

(al

an

(4)

pri

are

*

*

17

ass

pri

ov

pa

A. Miller / Economics of Education Review 36 (2013) 60–7270

pporting teachers.17 High teacher turnover can be both an of underlying problems in school performance andelf a cause of poor performance (Ingersoll, 2001).Using the same specification described above, I find that

e percentage of teachers who return to the school everyar is flat in the years leading up to the principalnsition. At the same time that the old principal leaves,% more teachers leave than usual. Following the first

ar of the new principal, 1.6% more teachers leave thanual. The retention rate then gradually returns to normal.e results are shown in Fig. 6. The drop is statisticallynificantly different from zero but fairly small. For a

pical sample school with roughly 40 teachers, thetimates imply the loss of roughly one more teacheran usual during the years surrounding a principalnsition. Nonetheless, the results show that a signifi-

ntly higher than usual fraction of teachers leave theirhools during a principal turnover. As shown in column) of Table 10, if the sample is restricted to schoolsithout multiple transitions, the estimated coefficients areen larger (though less precisely estimated). Since theain academic achievement measures are only available

ble 10

eachers who return before and after principal transitions.

All schools Primary schools

(1) (2) (3) (4)

our years before new principal �0.122 �0.726 0.171 �0.279

(0.276) (0.913) (0.396) (1.240)

hree years before new principal 0.078 �0.880 0.473 0.390

(0.257) (0.955) (0.380) (1.254)

wo years before new principal 0.082 �0.892 0.506 0.083

(0.248) (1.010) (0.360) (1.346)

he year before new principal �1.300*** �2.018 �1.221*** �1.287

(0.248) (1.107) (0.360) (1.505)

irst year of the new principal �1.552*** �2.808* �1.407*** �1.964

(0.268) (1.190) (0.382) (1.578)

econd year of the new principal �0.538* �2.499* �0.536 �1.346

(0.242) (1.263) (0.352) (1.745)

hird year of the new principal �0.326 �0.706 0.160 0.906

(0.263) (1.401) (0.382) (1.881)

ourth year of the new principal 0.151 �1.074 0.573 �0.357

(0.270) (1.535) (0.374) (2.113)

ive or more years of the new principal 0.149 �0.507 0.317 0.885

(0.166) (1.787) (0.242) (2.461)

djusted R2 0.329 0.269 0.328 0.262

21,835 3817 11,924 2321

chool and year fixed effects Yes Yes Yes Yes

otes: Each column reports coefficients from a separate OLS regression. The dependent variable is the percentage of teachers in a given school in a givenyear who

urn to teach the same school in the following school year. This measure of teacher retention is available from the 1994–1995 through 2004–2005 school years

lbut the last year of the sample period). The omitted category is the intercept for five or more years before the principal transition. All regressions include school

d year fixed effects. Standard errors clustered at the school level are shown in parenthesis. Columns (1) and (2) include all school levels while columns (3) and

are restricted to primary schools. Columns (2) and (4) are restricted to schools without multiple principal transitions (e.g., column (4) consists of an analysis of

mary schools led by zero or one principals over the sample period). In columns (1) and (3), an observation is a school-year-transition: School-year observations

included in the regression just once if the school has zero or one principal transition, twice if the school has two principal transitions, etc.

p < 0.05. **p < 0.01.

** p < 0.001.

−2

−1

01

−4 −3 −2

Old Prin

cipal

New P

rincip

al 2 3 4Afte

r

Years Before/Since Transition

95% CI Coefficient on Dummy

Fig. 6. % of teachers who return as the principal changes.

Notes: This is a plot of the coefficients from a regression of teacher

retention in a school in a given year on a set of indicator variables for years

before or since a principal transition. Specifically, the dependent variable

is the percentage of teachers in a given school in a given year who return

to teach the same school in the following school year. This measure of

teacher retention is available for 11 years (from the 1994–1995 through

2004–2005 school years, all but the last year of the sample period). Five or

more years before the principal change is the omitted category. The

regression includes school and year fixed effects. The 95% confidence

interval formed from standard errors clustered at the school level. The

regression includes 979 schools. An observation is a school-year-

transition: School-year observations are included in the regression just

once if the school has zero or one principal transition, twice if the school

has two principal transitions, etc. N = 21,835 = 11 years � 1985

where the number of transitions, 1985 = 347 schools with zero or one

transitions + 2 � 354 schools with two transitions + 3 � 199 schools with

Fuller et al. (2007) find that principal experience is negatively

ociated with teacher turnover, and Stoelinga et al. (2008) find that new

ncipals have more difficulty recruiting teachers, but both focus on

three transitions + 4 � 64 schools with four transitions + 5 � 13 schools

with five transitions + 6 � 2 schools with six transitions.

erall principal experience in education rather than principal tenure at a

rticular school.

inrps

4

ptrSpppmtrs

eisdbb(1yolasnsptiloom

ppteine

T

D

N

sc

s

A. Miller / Economics of Education Review 36 (2013) 60–72 71

primary schools, I also report estimates of teacheretention around principal transitions separately forrimary schools in columns (3) and (4). The results aretable across the samples.

. Discussion

In each sample and specification, schools experienceoor academic performance around the time of principalansitions. However, interpreting this finding is difficult.

chools may acquire a new principal because the existingrincipal is fired or promoted, moves or retires. The drop inerformance before a principal transition establishes thatrincipal transitions do not occur randomly (and eachodel includes school fixed effects, so neither areansitions as good as random conditional on permanent

chool characteristics).These findings imply that caution is necessary in

stimating the effect of principal transitions. The difficulty illustrated in Table 11, which shows difference-in-ifference estimates of the effect of a principal transitionetween the 2000 and 2001 school year. All estimates areased on eight years of data, but the estimates in columns) and (2) are based on data from the 1995–2002 school

ears, while the estimates in columns (3) and (4) are basedn data from the 1999–2006 school years. The autocorre-tion of school test scores and the presence of negative

election, on average, into principal transitions generateegative estimated effects of principal transition from thepecifications with a long pre-period and a short post-eriod and positive estimated effects of principal transi-on from the specification with a short pre-period and ang post-period. Naıve difference-in-difference estimates

f the effect of principal transition yield unstable andisleading results.

While previous studies have examined the experiencerofile for principals, no one has noted the drop in schoolerformance before new principals take office. The fall inst scores before principals take office alters theterpretation of the existing (ordinary least squares)

significant Ashenfelter dip in test scores before principaltransitions.18 Since the typical school is doing badlyrelative to its usual performance before a new principalstarts, it is entirely possible that the school would haveexperienced a recovery to its usual performance regardlessof whether the principal was replaced. It is difficult todisentangle the positive effects of having a new principalfrom what is merely a return to the permanent state of theschool.

About 20% of public school principals in the UnitedStates leave their positions each year and most school areled by principals with fewer than ten years of tenure. It isimportant to consider how changes in school leadershipaffect students. Using data from North Carolina publicschools, this paper documents that student achievementtypically rises following the installation of a new principal.However, principal turnover typically occurs after a periodof falling achievement. Because schools perform poorly inthe years preceding principal transitions, cross-sectionaland difference-in-difference estimates of the effect ofprincipal transitions are misleading. In this paper, allvarieties of principal turnover are analyzed together.Further work might distinguish between between differ-ent types of turnover. The prior experiences of theincoming principal may be of particular interest. Ultimate-ly, identifying exogenous sources of variation principaldepartures is necessary to reliably pin down the causaleffects of principal turnover.

able 11

ifference-in-difference estimates of the effect of 2001 principal transition.

Data from 1995 to 2002 Data from 1999 to 2006

(1) (2) (3) (4)

School with new principal �0.0627 0.0081 �0.0777 �0.0142

(0.0329) (0.0838) (0.0345)* (0.0892)

2001 or later �0.0033 �0.0022 �0.0181 �0.0047

(0.0081) (0.0262) (0.0075)* (0.0240)

School with new principal and 2001 or later �0.0354 �0.0576 0.0139 0.0031

(0.0203) (0.0446) (0.0176) (0.0500)

Constant 0.0076 0.0283 0.0107 0.0263

(0.0124) (0.0352) (0.0133) (0.0416)

R2 0.0086 0.0028 0.0071 0.0004

Observations 4432 424 4432 424

Notes: Each column reports coefficients from a separate OLS regression. In columns (1) and (3), all primary schools are included and

= 4432 = 8 years � 554 schools. In columns (2) and (4), schools with new principals in years other than 2001 are excluded (so the sample consists of

hools without any principal transitions and schools with a single principal transition resulting in a new principal in 2001) and N = 424 = 8 years � 53 -

chools. Standard errors clustered at the school level are shown in parenthesis.

* p < 0.05.

18 In Ashenfelter (1978), Ashenfelter writes ‘‘. . . the earnings of trainees

tend to fall, both absolutely and relative to the comparison group, in the

year prior to training. . .. this result introduces considerable ambiguity

into the empirical analysis for it suggests that some part of the observed

earnings increase following training may merely be a return to a

permanent path of earnings that was temporarily interrupted by one form

of transitory labor market phenomenon or another. To the extent that this

is the case the earnings generating functions of the trainee and

comparison groups may differ considerably in the period just prior to

training and cause considerable ambiguity in untangling the effect of

training from the effect of this transitory phenomenon.’’ The same

mbiguity arises with regard to the observed test score increase following

rincipal turnover.

stimates of the principal tenure profile. There is aa

p

Re

Aa

As

Ba

Be

Be

Bil

Bra

Bra

Bre

Cu

Da

De

Eb

Eh

Fu

Ga

Ha

Ha

A. Miller / Economics of Education Review 36 (2013) 60–7272

ferences

ronson, D., Barrow, L., & Sander, W. (2007). Teachers and student achieve-ment in the Chicago public high schools. Journal of Labor Economics, 25,95–135.

henfelter, O. (1978). Estimating the effect of training programs on earn-ings. The Review of Economics and Statistics, 60, 47–57.

ttle, D. (2010). Principal attrition and mobility: Results from the 2008–09principal follow-up survey (NCES 2010-337). U.S. Department of Education,National Center for Education Statistics. Washington, DC: US GovernmentPrinting Office.

rtrand, M., & Schoar, A. (2003). Managing with style: The effect ofmanagers on firm policies. Quarterly Journal of Economics, 118, 1169–1208.

sley, T., & Machin, S. (2008). Are public sector CEOs different? Leadershipwages and performance in schools.

lger, S. M. (2007). Principals as agents? Investigating accountability in thecompensation and performance of school principals. Industrial & LaborRelations Review, 61, 90–107.

nch, G. F., Hanushek, E. A., & Rivkin, S. G. (2009a). Estimating principaleffectiveness. Technical Report 32 Center for Analysis of LongitudinalData in Education Research.

nch, G. F., Hanushek, E. A., & Rivkin, S. G. (2009b). Principal turnover andeffectiveness. American Economics Association Annual Meeting.

wer, D. J. (1993). Principals and student outcomes: Evidence from U.S.high schools. Economics of Education Review, 12, 281–292.

llen, J. B., & Mazzeo, M. (2008). Implicit performance awards: An empiricalanalysis of the labor market for public school administrators.

vis, S., Darling-Hammond, L., LaPointe, M., & Meyerson, D. (2005). Schoolleadership study: Developing successful principals. Technical Report Stan-ford University School Redesign Network.

Paola, M., & Scoppa, V. (2008). The effects of managerial turnover: Evidencefrom coach dismissals in Italian soccer teams. Technical Report MunichPersonal RePEc Archive.

erts, R. W., & Stone, J. A. (1988). Student achievement in public schools: Doprincipals make a difference? Economics of Education Review, 7, 291–299.

renberg, R. G., Chaykowski, R. P., & Ehrenberg, R. A. (1988). Determinantsof the compensation and mobility of school superintendents. Industrialand Labor Relations Review, 41, 386–401.

ller, E., Baker, B. D., & Young, M. (2007). The relationship between principalcharacteristics. School-Level Teacher Quality and Turnover, and StudentAchievement.

tes, S. M., Ringel, J. S., Santibanez, L., Guarino, C., Ghosh-Dastidar, B., &Brown, A. (2006). Mobility and turnover among school principals.Economics of Education Review, 25, 289–302.

nushek, E. A. (1986). The economics of schooling: Production and effi-ciency in public schools. Journal of Economic Literature, 24, 1141–1177.

rgreaves, A., Moore, S., Fink, D., Brayman, C., & White, R. (2003). Succeedingleaders? A study of principal succession and sustainability Technical ReportOntario Principal’s Council.

Hart, A. W. (1991). Leader succession and socialization: A synthesis. Reviewof Educational Research, 61, 451–474.

Herman, R., Dawson, P., Dee, T., Greene, J., Maynard, R., Redding, S., et al.(2008). Turning around chronically low-performing schools: A practiceguide (NCEE 2008-4020). National Center for Education Evaluation andRegional Assistance Institute of Education Science US Department ofEducation.

Hussain, I. (2007). School choice, information disclosure and sanctions: Evi-dence from an unusually tough school accountability regime. TechnicalReport University College London.

Ingersoll, R. M. (2001). Teacher turnover and teacher shortages: An organi-zational analysis. American Educational Research Journal, 38, 499–534.

Jacobson, L. S., LaLonde, R. J., & Sullivan, D. G. (1993). Earnings losses ofdisplaced workers. The American Economic Review, 83, 685–709.

Lavy, V. (2008). Does raising the principal’s wage improve the school’soutcomes? Quasi-experimental evidence from an unusual policy exper-iment in Israel. Scandinavian Journal of Economics, 110.

Meier, K. J., & O’Toole, L. J., Jr. (2002). Public management and organizationalperformance: The effect of managerial quality. Journal of Policy Analysisand Management, 21, 629–643.

Meyer, M. J., Macmillan, R. B., & Northfield, S. (2009). Principal successionand its impact on teacher morale. International Journal of Leadership inEducation, 12, 171–185.

Miskel, C., & Cosgrove, D. (1985). Leader succession in school settings. Reviewof Educational Research, 55, 87–105.

Miskel, C., & Owens, M. (1983). Principal succession and changes in schoolcoupling and effectiveness.. American Educational Research AssociationAnnual Meeting.

Nye, B., Konstantopoulos, S., & Hedges, L. V. (2004). How large are teachereffects? Educational Evaluation and Policy Analysis, 26, 237–257.

Papa, F., Jr. (2007). Why do principals change schools? A multivariateanalysis of principal retention. Leadership and Policy In Schools, 6,267–290.

Partlow, M. C., & Ridenour, C. S. (2008). Frequency of principal turnover inOhio’s elementary schools. Mid-Western Educational Researcher, 21, 15–23.

Richardson, M. D., Short, P. M., & Prickett, R. L. (1993). School principals andchange. New York: Garland Publishing Inc.

Rowan, B., & Denk, C. E. (1984). Management succession, school socioeco-nomic context, and basic skills achievement. American Educational Re-search Journal, 21, 517–537.

Stoelinga, S. R., Hart, H., & Schalliol, D. (2008). The work of Chicago publicschools’ principals: Leading in a complex context with high stakes. TechnicalReport Consortium on Chicago School Research.

Weinstein, M., Jacobowitz, R., Ely, T., Landon, K., & Schwartz, A. E. (2009). Newschools, new leaders: A study of principal turnover and academic achieve-ment at new high schools in New York city. Education Finance ResearchConsortium.

Young, M. D., & Fuller, E. (2009). Tenure and retention of newly hired principalsin Texas. University Council for Educational Administration.