The impact of college course oerings on the supply of ...theunbrokenwindow.com/Higher Ed/Readings/Impact of College... · 2 Randall Reback/Journal of Econometrics( ) – ARTICLE IN

ARTICLE IN PRESS

Journal of Econometrics ( ) –www.elsevier.com/locate/econbase

The impact of college course o�erings on thesupply of academically talented

public school teachersRandall Reback

Department of Economics, Barnard College, Columbia University, 3009 Broadway,New York, NY 10027, USA

Abstract

This paper identi!es the impact of undergraduate teacher certi!cation programs on the likeli-hood that recent college graduates enter and remain in public school teaching jobs. More selectivepostsecondary institutions are far less likely to o�er teacher certi!cation programs and those thatdo o�er them are less likely to allow students to complete them within their four undergraduateyears. First, to examine the overall relationship between the availability of teacher certi!cationprograms and the likelihood that academically talented students enter teaching careers, I combineBarron’s ratings of college selectivity, detailed data tracking college seniors into the workforce,data on the types of teacher certi!cation programs o�ered by their colleges, and data on states’certi!cation requirements. Next, to isolate the causal e�ect of program o�erings, I treat the se-lection issue related to students sorting into colleges based on pre-existing interest in teachingcareers as an omitted variable problem. Using another data set that surveys high school seniors’career interests and tracks them into college, I estimate an upper bound for the magnitude ofthe bias in the baseline coe.cient estimates. The results suggest that the addition of teachercerti!cation programs that may be completed within four undergraduate years could increaserates of entry into public school teaching by at least 50% among recent graduates of certainselective colleges.c© 2003 Elsevier B.V. All rights reserved.

JEL classi(cation: I20; J45; C20

Keywords: Teacher supply; College selectivity; Certi!cation; Teacher training; Misspeci!ed probit model

E-mail address: [email protected] (Randall Reback).

0304-4076/$ - see front matter c© 2003 Elsevier B.V. All rights reserved.doi:10.1016/j.jeconom.2003.10.009

mailto:[email protected]

2 Randall Reback / Journal of Econometrics ( ) –

ARTICLE IN PRESS

1. Introduction

Given the immediate need for more teachers and the widespread desire to improvestudent achievement, there is increasing concern over how to put high-quality teachersin public school classrooms. What makes a good teacher? Certainly personal qualitiesand useful training experience contribute to a teacher’s e�ectiveness. Just as we wouldnot want to discourage compassionate individuals from teaching, we would not wantto discourage academically talented individuals from teaching. However, it is possiblethat the current system does precisely this. State policies and college 1 course o�eringsput academically talented students at a relative disadvantage. More prestigious collegesare less likely to o�er programs so that students can quickly and easily obtain teachercerti!cation. Thus, these college students have less exposure to pathways to teachingand face higher costs of pursuing certi!cation. They are also far less likely to enterpublic school teaching after graduating from college. Either college course o�eringssimply meet students’ pre-existing demand to pursue teaching careers, or the presenceof undergraduate teacher certi!cation programs has a direct e�ect on individuals’ careerchoices. If the former is true, then di�erential access to teacher certi!cation programshas no e�ect on the supply of teachers. If the latter is true, then barriers to entry maybe just as important as salary comparisons in explaining why more high-achievingindividuals do not enter public school teaching jobs.This paper investigates whether college course o�erings do in fact have a causal e�ect

on the likelihood that graduates of selective colleges enter public school teaching. First,to examine the overall relationship between college course o�erings and the likelihoodthat academically talented students enter teaching careers, I combine Barron’s (1998)ratings of college selectivity, detailed data tracking college seniors into the workforce,data on the types of teaching certi!cation o�ered by their colleges, and data on states’certi!cation requirements. Next, to isolate the causal e�ect of college course o�erings,I address the selection issue related to students sorting into colleges based on theirpre-existing interest in teaching. Treating this selection issue as an omitted variableproblem and using a secondary data set to estimate various parameters, I !nd lowerbound point estimates for the causal e�ect of college course o�erings on entry intopublic school teaching.The results suggest that the addition of undergraduate teacher certi!cation programs

at some elite colleges would lead to an increased presence of academically talentedteachers in public school classrooms. In particular, these programs cause greater ratesof entry for individuals who attend colleges rated in the “Highly Competitive” group.This is Barron’s second highest rating group, consisting of 85 colleges that produceapproximately 11% of U.S. college graduates. The evidence suggests that convertingpost-B.A. or !fth-year teacher certi!cation programs to programs that may also becompleted within one’s four undergraduate years could, on average, increase the rate ofentry into public school teaching by at least 50% at colleges such as Emory University,

1 Throughout this paper, I will use the term “colleges” to describe colleges and universities that o�erBachelor’s degrees.

ARTICLE IN PRESSRandall Reback / Journal of Econometrics ( ) – 3

the University of California (at Berkeley or Davis or Los Angeles or Santa Barbara),and Whitman College.

2. Related literature

Despite frequent pleas to raise teacher salaries and/or to create new teacher induc-tion programs, there is mixed evidence as to what actions would actually improve thequality of public school teachers. While the debate over the e�ects of raising teachersalaries remains wide open, 2 the literature has generally ignored another factor thatmay possibly have an important e�ect on who enters teaching. There may be consid-erable costs associated with entering public school teaching, as individuals may haveto give up a year’s work, or at least part of a year’s work, in order to obtain teachercerti!cation. Furthermore, the costs of entering teaching are often inversely related tothe quality of an individual’s college. The more selective a college, the less likely itis to o�er programs in which students may obtain teaching certi!cations within theirfour undergraduate years.Previous research on entry into the public school teaching profession shows that grad-

uates who earn higher test scores or who attend more selective colleges are less likelyto become teachers. Hanushek and Pace (1995), using the High School and Beyonddata for the high school class of 1980, report that college graduates who completedundergraduate teacher preparation programs had lower test scores than other collegegraduates. Ballou (1996), using the Survey of Recent College Graduates: 1976–1991,!nds that graduates of highly selective colleges are less likely to (1) get certi!ed,(2) apply for teaching jobs once certi!ed, and even slightly less likely to (3) gain em-ployment when applying for teaching jobs. If the third !nding is a true phenomenon,then one would be concerned that increasing the fraction of academically talented grad-uates who pursue teaching careers would not necessarily lead to more of them actuallyteaching in public schools. However, the data used in my study, which allow one tomore e�ectively di�erentiate between the decisions of employers and prospective teach-ers, do not support this !nding. 3 On the other hand, the data are certainly consistent

2 Recent studies (e.g. Figlio, 1997; Loeb and Page, 2000) !nd evidence that “high quality” teachers tendto go to higher paying districts. However, it is unclear whether raising teacher salaries, either generally orcategorically, would entice more people with desirable characteristics to enter the teaching force. In fact,one study theorizes that raising teacher salaries may lead to lower teacher quality (Ballou and Podgursky,1995). Other studies examine the e�ect of teacher salaries on the likelihood of teachers changing schooldistricts or exiting the teaching profession (Hanushek et al., 1999; Imazeki, 2001). These studies !nd thathigher salaries tend to reduce both types of attrition, though the magnitude of the e�ect is not always largeand varies by the type of district.

3 Based on the Baccalaureate & Beyond (B&B) data, graduates from colleges rated in Barron’s top threeselectivity groups who are fully certi!ed to teach are slightly more likely to teach than other fully certi!edgraduates (87.6% versus 86.5%). Fully certi!ed applicants for teaching jobs from these more selectivecolleges are also more likely to receive at least one job o�er (92.7% versus 89.1%). Controlling for thenumber of jobs that they apply to, applicants who graduated from more selective colleges receive morejob o�ers. These !ndings may di�er from Ballou’s (1997) !ndings, because the latter used a proxy for anindividual failing to receive teaching job o�ers.


ARTICLE IN PRESS

with Ballou’s !nding of lower rates of entry into teaching among graduates from rel-atively selective colleges. The purpose of the empirical section below is to determinethe extent to which these trends are due to the causal e�ect of limited undergraduatecerti!cation programs at relatively selective colleges.Since this paper examines how pathways towards teacher certi!cation a�ect whether

individuals from selective colleges become teachers, it is worth reviewing what theschool production function literature reveals about the inNuence that teachers’ under-graduate training and cognitive abilities have on their students’ achievement. The schoolproduction function literature has produced mixed results concerning which inputs aresystematically e�ective in raising student achievement. 4 One problem with estimatingthe returns to various inputs is that there is endogenous sorting of students into schoolswith various traits. There is rarely a valid instrument for the presence of an input, andeven studies that examine net gains in student achievement over time cannot fullycontrol for the relationship between sorting and unobserved variables related to studentacademic growth. Another problem is the strong collinearity of the inputs themselves,causing a high likelihood of serious omitted variable bias. This latter issue is par-ticularly problematic when one hopes to !nd the causal relationship between teachercharacteristics and student achievement. For example, individuals from highly selectivecolleges who pursue teaching careers might tend to have lower alternative wages thantheir college classmates. These individuals may thus have unobserved characteristicsthat are not only negatively correlated with alternative wages, but are also negativelycorrelated with productivity as a teacher. Ehrenberg and Brewer (1994) !nd that theselectivity of a teacher’s undergraduate college is signi!cantly, positively related toteachers’ students’ test score gains, but that this result depends on the other indepen-dent variables used in the regression and on the race of the students. Studies (Ferguson,1991; Ehrenberg and Brewer, 1995; Coleman et al., 1966) generally !nd a positive cor-relation between students’ performance and their teachers’ scores on tests that measurethe teachers’ verbal abilities. As for overall teacher quality, Hanushek et al. (1998) es-timate that variation in teacher quality within Texas public schools accounts for at least7.5% of the variation in student achievement and probably much more. 5 In summary,while the relative importance of teacher characteristics is open to debate, the schoolproduction function literature is consistent with the logical ideas that teacher qualityplays an important role in student achievement and that a teacher’s academic abilitiescontribute to his or her overall quality.

3. Framework: college students’ potential pathways to public school teaching jobs

Before discussing the descriptive and quantitative analyses, it may be helpful todiscuss individuals’ progression towards public school teaching or other careers. All

4 Hanushek (1986, 1997) and Hedges et al. (1994) o�er various interpretations of the school spendingliterature. See Monk and Rice (1999) for an excellent summary of !ndings concerning the productivity ofspeci!c types of inputs.

5 The authors !nd this conservative lower bound by assuming that none of the between-school variationin student test score growth is related to di�erences in teacher quality.


states require public school teachers to have earned a Bachelor’s degree, as well assome type of certi!cation at the elementary or secondary level. 6

This paper focuses on the impact of undergraduate teacher certi!cation programs thatmay be completed within four undergraduate years. For the remainder of the paper, Iwill refer to these programs as UTCPs. If one attends a college without a UTCP, or ifone chooses not to enroll in such a program, then he or she may still obtain certi!cation.This person could attend a post-B.A. certi!cation program at his or her undergraduateinstitution or at another college. These programs are typically 1 year long, and studentsmust cover the full tuition, including paying for the opportunity to serve as a studentteacher in a K-12 classroom. In some states, there are other, “alternative” pathwaystowards certi!cation. Individuals can sometimes obtain emergency certi!cation, whichallows them to take classes towards full certi!cation and simultaneously teach in apublic school district that has traditionally had shortages of quali!ed, certi!ed teachers.Teach for America is a program that speci!cally targets undergraduates at selectivecolleges who are interested in teaching and matches them with school districts thatneed to hire emergency certi!ed teachers. Aside from emergency certi!cation, somecolleges or school districts o�er formal, alternative certi!cation programs that allow fornon-traditional or accelerated training. Teachers with emergency or alternative certi!-cation still face entry costs, because they are required to engage in part-time trainingtowards full certi!cation.One would expect graduates from colleges o�ering UTCPs to be more likely to

ultimately become public school teachers. High school students who are already in-terested in teaching careers may prefer to enroll in these colleges. In addition to thissorting e�ect, individuals may have dynamic preferences, so that the presence of aUTCP could have a causal e�ect on the likelihood that individuals become teachers.Many high school seniors are unsure of their career goals when entering college. Someothers may base their college enrollment decision on factors unrelated to the college’scourse o�erings, such as tuition, prestige, or location. For example, individuals initiallyinterested in teaching may choose to enroll in a college that does not o�er a UTCP,with the intentions of pursuing regular or alternative certi!cation after their four under-graduate years. While individuals attend college, availability of a UTCP could increasethe likelihood that they pursue teaching careers. There are two reasons why a UTCPcould increase the likelihood that an individual becomes interested in teaching or thelikelihood that an individual with a pre-existing interest in teaching remains interested.First, since individuals may complete UTCPs within their four undergraduate years,they can avoid the tuition and opportunity cost associated with pursuing teacher cer-ti!cation after graduation. Second, the mere presence of UTCPs might lead to moreinformation about teaching and thus build greater levels of interest.Many other factors will inNuence the likelihood that an individual becomes a teacher.

Family characteristics, individual characteristics, and one’s pre-existing interest in

6 Elementary certi!cation typically allows one to teach in Kindergarten through sixth grade, while sec-ondary certi!cation allows one to teach in high school. Most states require secondary certi!cation for publicmiddle school teachers, though some use elementary certi!cation or allow certi!cation that is speci!c tomiddle school instruction.


ARTICLE IN PRESS

teaching will inNuence progress towards a teaching career as an individual decideswhich college to attend, which courses to take, whether to pursue teacher certi!ca-tion, and whether to pursue a public school teaching job. These decisions will also beinNuenced by expected teacher salary and working conditions, as compared to one’spotential alternative income. 7 In addition, various state policies will a�ect the relativedesirability of pursuing teaching or of pursuing a certain type of certi!cation. To theextent that certi!cation is non-transferable across states, location preferences could in-Nuence whether an individual pursues certi!cation as an undergraduate. For example,if an undergraduate believes that there is a high probability that he or she will latermove to a state that does not readily accept certi!cation from the state of the under-graduate’s college, then this individual may not want to bother pursuing certi!cationas an undergraduate. In actuality, most states have o.cial reciprocity agreements withother states agreeing to fully recognize teacher certi!cation from the other state. 8

Some factors may also inNuence the marginal e�ect of the presence of a UTCPon the likelihood that one becomes a teacher. The interaction most relevant to thispaper’s analysis is the relationship between the selectivity of an individual’s college andthe marginal e�ect of a UTCP. The expected direction of this relationship is actuallyambiguous. On the one hand, graduates from very selective colleges will generally haverelatively high opportunity costs associated with pursuing certi!cation after collegegraduation. One might thus expect the marginal e�ect of a UTCP to be greater atthese colleges. On the other hand, there may be few students at these colleges who aresu.ciently close to the margin so that the presence of a UTCP actually could a�ecttheir decision. Since their alternative salary is so much greater than their potentialsalary as a teacher, students at an extremely selective college may only decide tobecome teachers when this decision is dominated by non-pecuniary factors. Giventhe ambiguous relationship between college selectivity and the marginal e�ect of thepresence of a UTCP, it will be important to identify which types of colleges areassociated with the largest e�ects.

4. Data

The focus of this paper is on the entry of recent college graduates from selectivecolleges into public school teaching jobs. In order to conduct the main analysis, I userestricted-use data from the National Center for Education Statistics’ Baccalaureate and

7 Since teaching jobs here refer to teaching jobs in public schools, jobs at private schools are included in therange of alternative jobs. However, some of the empirical analysis will also examine what factors inNuenceentry into any type of teaching job. This allows one to determine the extent to which the unavailability ofundergraduate certi!cation programs causes a shift towards private teaching jobs rather than non-teachingjobs.

8 Currently, 40 states and the District of Columbia have o.cial reciprocity agreements with otherstates, with each state typically agreeing to recognize teacher certi!cation from more than 30 other states(NASDTEC, 2000). The majority of the states without o.cial reciprocity agreements grant provisional orfull certi!cation if the outside state’s requirements meet certain standards or if the individual completed ateacher preparation program at an institution certi!ed by the NCATE (National Council for the Accreditationof Teacher Education).


Beyond Longitudinal Study (1993–1997) (B&B) that tracks college Seniors into theworkforce. The B&B data are an excellent source of information on entry into teaching,since the survey was designed partly with this topic in mind. This includes data onwhether individuals receiving Bachelor’s degrees in 1993 later enter various stagesof the “teacher pipeline,” a useful construct to describe progress towards becoming ateacher. The “teacher pipeline” stages include (1) whether the student trains or appliesto teach, (2) whether the student teaches by 1997, and (3) whether someone who hastaught indicates that teaching is in their long-term plans.I combine the B&B data with existing data on state certi!cation policies and with

newly gathered data on college course o�erings. The college course o�ering informationwas collected through email messages, phone calls, analysis of websites, and analysisof information in a published report (NASDTEC, 2000). For each college, at least twoindependent sources were consulted. For both elementary and secondary certi!cation,the college course o�ering information includes: (1) whether the college o�ers thecerti!cation program at all, (2) whether the college o�ers the certi!cation program sothat undergraduates may complete it within their regular timeframe towards the B.A.(roughly 4 years), and (3) whether the college o�ers combined certi!cation/Master’sdegree programs. Question (2) is based on whether some students are able to re-ceive certi!cation within their four undergraduate years, without spending any extrasemesters, summer semesters, or other extra time pursuing this certi!cation. These col-lege course o�ering measures generally pertain to 2000, rather than the time periodof the B&B. Unobserved changes in college course o�erings over this time are thus apotential source of measurement error.The college certi!cation program data are available for all colleges in Barron’s

(1998) top three selectivity rating groups. Analysis of the B&B data suggests thatthese groups individually composed 4.3%, 11.3%, and 26.7% (from most to third-mostselective) of the total fraction of young, recent graduates from nonmilitary colleges. Inorder to isolate recent college graduates who might be interested in teaching careers, Irestrict the sample to students attending nonmilitary colleges who enter college beforethe age of 22 and graduate from college before the age of 30.In addition to these data sources, I use IPEDS (Integrated Post-secondary Educa-

tion Data System) data to further characterize one’s undergraduate institution and useNELS (National Education Longitudinal Survey) data to characterize progress towardsteaching that occurs prior to college enrollment. For all analyses, I use the appropri-ate sample weights given by the NELS or B&B. Using these weights, one may !ndnationally representative estimates for both the total student population and studentsattending various types of colleges.

5. Descriptive analysis

With the framework of Section 3 in mind, this section answers some descriptivequestions about entry into teaching. Though this analysis is not useful for describ-ing causal e�ects or predicting the impact of policy changes, it provides backgroundinformation about who enters public school teaching jobs.


ARTICLE IN PRESS

Table 1

Barron’s college selectivity rating

Most Highly Very Top three groupscompetitive (%) competitive (%) competitive (%) combined (%)

(a) Composition of undergraduates, recent graduates, and recentgraduates who teach by selectivity of their colleges% of all undergraduates 2.5 7.4 17.7 27.6% of all college graduates 4.3 11.3 26.7 42.3% of all college graduates who 3.9 8.2 25.1 37.2

enter teaching within4 years

% of all college graduates 1.9 7.2 25.8 34.9who enter publicschool teaching within4 years

% of all college graduates 1.0 7.3 25.5 33.9who enter publicschool teaching andplan on long-termcareers in K-12 education

(b) Rates within selectivity groups by which graduates reachvarious teaching outcomes% who enter teaching within 14.6 11.5 14.9 14.0

4 years% who enter public school 5.1 7.5 11.3 9.7

teaching within 4 years% who enter public school 1.9 5.1 7.6 6.4

teaching and plan onlong-term careers inK-12 education

Undergraduate percentages are based on IPEDS data for total undergraduate enrollments in all non-militarycolleges during the 1992–1993 school year. Percentages of graduates are based on (sample-weighted)B&B data for individuals who graduated during the 1992–1993 school year, excluding students who at-tended military colleges, entered college after the age of 21, or graduated after the age of 29.

5.1. What fraction of recent college graduates who become teachers come frominstitutions with various college selectivity ratings?

A breakdown of recent college graduates entering teaching is presented inTable 1. Though graduates from more selective colleges are less likely to enter teaching,undergraduates at these schools are also more likely to graduate than undergraduatesat less selective colleges. Thus, the proportion of new, young teachers coming fromhighly selective colleges is similar to the proportion of undergraduates enrolled at thesecolleges. This implies that changes in the fraction of graduates from highly selectiveschools who enter teaching could have a nontrivial impact on the overall distributionof teachers.


Table 2Undergraduate enrollment breakdowns by Barron’s college selectivity rating and the types of teacher certi-!cation programs o�ered

Types of teacher certi!cation Barron’s rating groupprograms o�ered

Most competitive Highly competitive Very competitive(48 colleges) (%) (85 colleges) (%) (245 colleges) (%)

4-year certi!cation program 54 57 81(either elementary or secondary or both)Elementary certi!cation

Any 57 89 914-year 38 53 74¡ 5 year 46 61 75Joint M.A. 42 48 61

Secondary certi!cationAny 84 93 984-year 49 53 72¡ 5 year 61 61 79Joint M.A. 59 63 69

Elementary/secondary combinations4-year elementary, any secondary 57 53 734-year elementary, 4-year secondary 33 50 65Any elementary, any secondary 38 89 91Any elementary, 4-year secondary 41 51 66

Example of interpretation: 89% of undergraduates enrolled at highly competitive colleges are enrolled atcolleges o�ering any elementary teacher certi!cation program. Percentages are based on IPEDS enrollmentdata for all non-military colleges. “¡ 5 year” includes all schools with 4-year programs, as well as othersthat require less than a full !fth year. For example, some programs require nine semesters or require eightregular semesters plus summer classes.

5.2. What fraction of colleges o6er various types of teacher certi(cation programs?

Table 2 displays the breakdown of certi!cation o�erings by the percentage of under-graduates enrolled at schools o�ering the various types of programs within the top threecollege selectivity ratings. 9 One can readily observe that the more selective collegesare less likely to o�er all types of certi!cation programs. The most striking percent-ages are the relatively low fraction of students at Most Competitive institutions whohave access to any elementary certi!cation program and the relatively high fractionof students at Very Competitive institutions who may earn elementary or secondarycerti!cation within their four undergraduate years.

9 The B&B data are representative of these population frequencies. Using the composition of collegegraduates in the B&B data, one estimates similar rates of exposure to certi!cation programs within eachcollege selectivity rating group.


ARTICLE IN PRESS

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

1989 1994 1997 2000

Other (4+)Very Competitive (3)Highly Competitive (2)Most Competitive (1)

Fig. 1. Fraction of individuals in the college class of 1993 revealing a potential long-term interest inteaching: sorted by Barron’s selectivity rating of their college. The “1989” rates are actually based on(sample-weighted) NELS data for the high school class of 1988, while the later data points are based on(sample-weighted) B&B data for the college class of 1993. The “2000” rates are based on a forward-lookingquestion asked in 1997. See Section 5.3 for more details.

5.3. Sorted by the type of certi(cation o6erings at their postsecondary institution,what fraction of students enter various stages of the “teacher pipeline”?

Fig. 1 displays the fraction of students, sorted by college selectivity rating, reachingvarious stages of the teacher pipeline. The stages are intended to measure levels ofinterest in a long-term public school teaching career. The !rst stage, labeled “1989,”uses the NELS data set to measure interest in teaching careers among high schoolseniors. The rates are based on the types of postsecondary institutions actually attendedin 1990 by students in the high school class of 1988, sorted by whether they indicateda future interest in teaching in 1988. I de!ned students as interested in teaching careersif they either chose “School Teacher” as their anticipated choice of occupation at age30 or they chose “Education” as their most likely major in college. 10 An individualgraduating college 4 years after !nishing high school in the B&B sample would havebeen in the high school class of 1989, which is why this point is labeled 1989 eventhough it is based on information from the high school class of 1988 in the NELS.The next three points in the teacher pipeline represented in Fig. 1 are based on

the B&B sample. The 1994 point indicates whether the student, 1 year after receivinga B.A., reports a long-term interest in teaching when asked directly about this. The1997 point indicates whether the student has taught in a public school before or during

10 Not all Education majors entail earning a teacher certi!cation program, though the most common typesof majors focus on elementary education and do entail receiving elementary certi!cation. One certainly neednot major in Education to enter an undergraduate teacher program. Many states require public secondaryschool teachers to major or minor in the !eld that they teach, and some forbid them to major in Education.


1997. The sector (public or private) of employment for some individuals who becameteachers is missing; these individuals are counted as teaching in a public school if theyhave any type of certi!cation (including emergency), a necessary but not a su.cientcondition for teaching in a public school rather than private school. The !nal pointin the pipeline is labeled 2000, though this is actually based on a forward-lookingsurvey question asked in 1997. This point reveals whether individuals who reached theprevious point (teaching in a public school by 1997) indicate that they plan to remainteaching or working in schools beyond 1999. Individuals indicating plans to move tonon-teaching jobs in schools are included, so that movement into jobs such as schoolcounselor or principal is not discounted. As one would suspect, the more selectiveone’s college, the less likely one is to indicate a long-term interest in public schoolteaching at any stage.Fig. 2 displays the fraction of individuals reaching these same stages, but sorted by

whether their college o�ered either type of 4-year undergraduate teacher certi!cationprogram. These relationships will be the main focus of the quantitative analysis below.The qualitative di�erences in these !gures are similar to the e�ects that one observesfrom the probit estimates described in the next section. The second selectivity group,Highly Competitive, has the least amount of initial sorting (di�erence between thetwo percentages in 1989) and the largest di�erences in long-term teaching outcomes(di�erences in 1997 and 2000). There are compelling reasons why the second ratinggroup has less sorting than the !rst (Most Competitive) or third (Very Competitive)rating groups. There is likely less sorting in the second rating group than in the thirdrating group, because students choose colleges based on many other factors and itwould be more di.cult to gain admission to colleges in the second rating group thatboth meet these criteria and also o�er undergraduate teacher training programs. Thereis likely less sorting in the second rating group than in the !rst rating group, becausethe second group contains far more students attending public schools than does the!rst rating group (see Table 3). Students who need or want to attend a public collegewithin their own state may be less able to factor teacher certi!cation o�erings intotheir enrollment decision. The next section will describe the quantitative analysis usedto isolate the direct relationship between certi!cation o�erings and various teachingoutcomes.

6. Quantitative analysis and results

The goal of this analysis is to exploit cross-sectional variation in college teachercerti!cation o�erings in order to estimate the impact of a college adding a UTCP(4-year undergraduate teacher certi!cation program). In particular, one would like toknow how the addition of a UTCP at colleges that currently only o�er post-B.A. or!fth-year certi!cation programs would a�ect the rates by which graduates from thiscollege enter and remain in public school teaching jobs. Probit estimates examiningteacher supply can be used to make these general equilibrium predictions only if twoconditions are met: (1) availability of UTCPs is not endogenous with respect to thesupply and demand for teachers from these rating groups, and (2) the impact of adding


ARTICLE IN PRESS

Most Competitive (1)

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

1989 1994 1997 2000

yesno

Highly Competitive (2)

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

1989 1994 1997 2000

yesno

Very Competitive (3)

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

1989 1994 1997 2000

yesno

Fig. 2. Fraction of individuals in the college class of 1993 revealing a potential long-term interest in teaching:sorted by Barron’s college selectivity rating and whether their college o�ered a 4-year, undergraduate teachercerti!cation program (yes/no). The “1989” rates are actually based on (sample-weighted) NELS data for thehigh school class of 1988, while the later data points are based on (sample-weighted) B&B data for thecollege class of 1993. The “2000” rates are based on a forward-looking question asked in 1997. See Section5.3 for more details.


Table 3Descriptive statistics for independent variables: population means with population standard deviations initalics

Independent variables All top three (1) Most (2) Highly (3) Veryrating groups competitive competitive competitive

College o�ers any undergraduate teacher 0.735 0.459 0.583 0.844certi!cation program? 0.441 0.499 0.493 0.362

Institutional characteristicsPublic college? 0.589 0.18 0.55 0.67

0.492 0.38 0.50 0.47Religiously a.liated? 0.191 0.07 0.17 0.22

0.393 0.25 0.37 0.41Ph.D. o�ered in any subject? 0.704 0.78 0.85 0.63

0.456 0.41 0.36 0.48Undergraduate enrollment 16,268 6,938 18,582 16,782

12,772 4,023 11,826 13,424

Individual characteristicsMinority (non-white) AND male 0.074 0.12 0.09 0.06

0.262 0.32 0.29 0.24Minority AND female 0.080 0.10 0.09 0.07

0.272 0.30 0.28 0.26White AND male 0.420 0.45 0.40 0.42

0.494 0.50 0.49 49Only SAT score available 0.53 0.72 0.56 0.48

0.50 0.45 0.56 0.50SAT verbal score if only took SAT 574 664 578 550

96 79 92 87SAT math score if only took SAT 570 655 581 545

93 73 87 86Only ACT score available 0.16 0.01 0.15 0.19

0.37 0.09 0.35 0.39ACT score if only took ACT 23 28 24 23

4.3 4.5 3.9 4.3Both scores available 0.15 0.09 0.15 0.17

0.36 0.29 0.36 0.37SAT verbal score with ACT also available 562 661 591 541

93 69 86 89SAT math score with ACT also available 571 671 602 549

88 74 82 82ACT score with SAT also available 25 29 27 24

4.9 4.1 4.2 4.8

Parental characteristicsMother: B.A. only 0.30 0.35 0.32 0.28

0.46 0.48 0.46 0.45Father: B.A. only 0.30 0.31 0.31 0.29

0.46 0.46 0.46 0.46Mother: grad. degree 0.18 0.35 0.19 0.15

0.39 0.48 0.39 0.36


ARTICLE IN PRESS

Table 3 (continued)

Independent variables All top three (1) Most (2) Highly (3) Veryrating groups competitive competitive competitive

Father: grad. degree 0.32 0.51 0.34 0.280.47 0.50 0.47 0.45

Either parent: high school 0.06 0.02 0.06 0.06dropout 0.23 0.15 0.23 0.24

Either parent: foreign born 0.17 0.24 0.20 0.140.37 0.43 0.40 0.35

Statistics are based on Baccalaureate and Beyond (B&B) data, using sample probability weights.

a UTCP does not have other general equilibrium e�ects that reduce supply. I willdiscuss the second issue in the next section, and address the !rst issue of potentialendogeneity here.

6.1. Exogeneity of colleges’ teacher certi(cation o6erings

If the availability of UTCPs at colleges is based on the net demand for teachers,then cross-sectional estimates will not measure the causal e�ect of these programs, evenwhen one accounts for the sorting that occurs when individuals select their colleges.For example, states in dire need of teachers might pressure private colleges or mandatepublic colleges to o�er these programs. Another possibility is that states with strongteachers unions inNuence colleges to not o�er these programs, so that the demand forexisting teachers remains high and higher salaries and bene!ts are required.One way around this issue is that a version of each of the models below will

include state !xed e�ects. State dummy variables will control for all of the endogenousvariables a�ecting the net supply of teachers, such as the relative size of the K-12population, teacher salaries relative to others, teachers’ working conditions and bene!ts,and the various state requirements for obtaining teacher certi!cation. Although datadescribing these factors are available, they are so numerous that the number of controlswould approach the number of states in the sample, implying that state !xed e�ectsare appropriate.Even without the inclusion of state !xed e�ects, one need not worry about the

endogeneity of UTCP availability as much as one might suspect. On the surface, theavailability of UTCPs does not appear to be related to the net supply of teachers. 11 InCalifornia, where urban areas traditionally have massive shortages of certi!ed teachers,the public University of California campuses are prohibited from o�ering these UTCPs.No such program exists at the University of California at Los Angeles, while the L.A.

11 Although UTCP o�erings generally seem to be exogenous with respect to the net supply of teachers,these o�erings appear to be endogenous with respect to state policies or labor market conditions that a�ect therelative return of enrollment in di�erent types of certi!cation programs. For example, availability of UTCPs isnegatively correlated with the state’s average salary premium associated with !rst-year public school teacherspossessing a Master’s degree. This reinforces the importance of including geographic dummy variables.


Uni!ed School District has had to create a program by which it hires and internallytrains hundreds of uncerti!ed teachers each year. In private colleges, the availabilityof UTCPs and other certi!cation programs appears to be more closely linked to theperceived prestige of producing graduates from these programs than to the actual marketneed for teachers. An example of the possible importance of occupational prestige is thefact that the most selective colleges are much more likely to o�er secondary certi!cationprograms than elementary certi!cation programs (see Table 2). This may result fromsocietal views that very intelligent people are wasting their talents when teaching inelementary schools. Similarly, many of these elite colleges may want their graduatesto leave college with a wide breath of knowledge and with detailed mastery of anon-Education major. The extra time devoted to train teachers, especially the extra timedevoted to the craft of teaching in elementary schools, may be thus viewed as replacingmore important instructional time. Finally, the lack of UTCPs at an elite college maybe a remnant of the time when this college was an exclusively male institution andteachers in the K-12 education system were predominantly female. Whether a collegewas all-male in 1950 is negatively related to whether it currently o�ers a UTCP. 12

6.2. Empirical methods

The focus of the empirics here will be on the impact of shortening existing cer-ti!cation programs so that a college o�ers either an elementary or secondary 4-yearUTCP. Ideally, one would also be able to estimate the impact of a college adding sometype of teacher certi!cation program when one was not previously o�ered. However,there are simply not enough individuals attending colleges that do not o�er any cer-ti!cation programs (only about 3% of the sample), for this to be identi!ed based oncross-sectional analyses. In order to isolate the e�ect of shortening programs rather thanadding them, these recent graduates from colleges without any certi!cation programsare dropped from the sample. The key independent variable of interest for the baselineregression is whether the college o�ers either type of 4-year program. Though dataare available on the grade level chosen by some teachers, there are too many missingvalues concerning grade level for one to separately measure the e�ects of elementarycerti!cation programs on elementary teaching and secondary programs on secondaryteaching.In order to isolate the impact of the presence of at least one UTCP on entry into

teaching, one must include various controls for the factors described in Section 3. Al-though enrolling in a college with certain characteristics may only be mildly relatedto whether one wants to teach, it is important to include various institutional controlvariables, because they may be correlated with whether the college o�ers a UTCP.Institutional controls include dummy variables for whether the college is public, reli-giously a.liated, o�ers a Ph.D. in any subject, and for the college’s selectivity rating

12 When a dummy variable for whether a college currently o�ers a UTCP is regressed on variables con-cerning various institutional characteristics, the coe.cient on a dummy variable for whether the collegewas all-male in 1950 is negative, though not statistically signi!cant at the 0.10 level. The marginal e�ect,estimated at the mean values, suggests that a college that was exclusively male in 1950 is 10% less likelyto o�er a UTCP.


ARTICLE IN PRESS

group. The other institutional controls are cubic terms for the college’s total under-graduate enrollment. Individual controls include scores on SAT or ACT tests. Thesevariables are important because they are indicators of ability to gain admission to col-leges, potential academic success in college, and earning potential after graduation.Dummy variables that interact race with gender are also included, as these may berelated to one’s interest in teaching. Characteristics of an individual’s parents may alsoinNuence the individual’s interest in teaching and his or her academic success. Thoughparental income is unavailable, measures of parental education and country of birthare included. Estimates of population means and standard deviations, calculated us-ing sample probability weights, for all of these independent variables are displayed inTable 3. To account for the net demand for teachers, including endogenous factorssuch as salaries and requirements, various models add state or regional dummies.Let Pij be a dummy variable concerning whether individual i graduating from college

j becomes a teacher. De!ne X1i as a vector of individual and parental control variables,X2j as a vector of institutional control variables, X ∗

3i as a variable measuring individuali’s interest in a teaching career before enrolling in college, X4j as a dummy variablefor whether the college o�ers an undergraduate teaching program, and Dj as a vectorof regional or state dummy variables for the location of the college. The “true” discretechoice model for individual i graduating from college j would thus be

P∗ij = �0 + X1i�1 + X2j�2 + �3X

∗3i + �4X4j + Dj�5 + �ij; (1)

Pij = 1 if P∗ij ¿= 0; Pij = 0 if P∗

ij ¡ 0: (2)

Since the primary data set begins in individuals’ senior year of college, one may notactually observe X ∗

3i or any discrete version of this variable. The actual probit modelthat I estimate to !nd the baseline relationships is

P∗ij = �o + X1i�1 + X2j�2 + �4X4j + Dj�5 + eij;

Pij = 1 if P∗ij ¿= 0; Pij = 0 if P∗

ij ¡ 0: (3)

Following the work of Yatchew and Griliches (1985), I derive an expression for themagnitude of the bias from omitting X ∗

3i. Assume that X ∗3i and the other independent

variables are related by a linear regression function:

X ∗3i = �0 + X1i�1 + X2j�2 + �4X4j + Dj�5 + uij

with E[uij|X ∗3i] = 0 and Var(uij) = �2u = 1: (4)

Substituting the right-hand side of Eq. (4) into the X ∗3i term in Eq. (1), I re-write

Eq. (1) as

P∗ij = �o + �3�o + X1i(�1 + �3�1) + X2j(�2 + �3�2) + (�4 + �3�4)4X4j

+Dj(�5 + �3�1) + �3uij + �ij: (1′)

If X ∗3i is normally distributed conditional on the other independent variables, then

(1′) and (2) form a valid probit model with the residual having a mean of zero and a


variance of �23�2u + �

2� , where �

2u and �2� are the variances of uij and �ij, respectively.

This suggests that consistent estimators of �4 will converge to

�4 + �3�4√�23�2u + �2�

: (5)

Normally, Eq. (5) would only be useful for determining the direction of the omittedvariable bias. However, by using a secondary data set that surveys high school seniors’career interests and tracks them into college, I can estimate values for the terms inEq. (5). By making some plausible assumptions, I am able to !nd a lower bound pointestimate for �4.Using this secondary data set, the previously mentioned NELS data, I construct a

dummy variable X3i that is equal to one if the individual revealed an interest in ateaching career during the senior year of high school. As described in Section 5.3,I set X3i equal to one if the individual chose “school teacher” as his or her likelyoccupation at age 30 or some sort of education !eld as his or her likely college major;otherwise, X3i equals zero. X3i will equal one if the latent variable, X ∗

3i, exceeds somethreshold. X ∗

3i is simply an indicator variable, so its scale is arbitrary. Given that �2uis assumed to equal one in Eq. (4), 13 I can re-write this equation as a valid probitmodel:

X ∗3i = �0 + X1i�1 + X2j�2 + �4X4j + Dj�5 + uij

X3i = 1 if X ∗3i ¿= c; X3i = 0 if X ∗

3i ¡ c;

where Var(uij) = �2u = 1 and c is some constant: (4′)

I use the NELS data to estimate �4 in the probit model speci!ed in Eq. (4′). 14 I haveassumed that �2u equals one, and �2� is also normalized to one for probit estimation ofEq. (1). Therefore, to solve for �4 in Eq. (5), the only remaining parameter to estimateis �3.Though I cannot estimate �3 directly, I !nd an upper bound value for �3 by using

information revealed from the two data sets. I can estimate an upper bound for �3 equalto E[Pij|X3i = 1] − E[Pij|X3i = 0], the di�erence in the mean entry rates into teachingbetween those who were and were not originally interested in a teaching career whenthey were seniors in high school. This is an upper bound for �3, because the correlationbetween the other independent variables in Eq. (1) and X3i is likely to be in the samedirection as the correlation between these independent variables and Pij. In other words,things that positively inNuence whether one becomes a teacher also positively inNuencewhether one is likely to have had an interest in teaching at an earlier age.

13 In case Eq. (4) is misspeci!ed, one would ideally have additional data so that one could restrict thesample to individuals with X3i=0 and estimate Eq. (1) with X ∗

3i omitted.14 All control variables that I use to estimate this model are identical to the corresponding estimation of

Eq. (3) using the B&B data, except for one. Due to availability, I replace the dummy variable indicatingwhether either of the individual’s parents was born in a foreign country with a dummy variable indicatingwhether any language other than English is spoken at the individual’s home.


ARTICLE IN PRESS

This di�erence in entry rates, E[Pij|X3i =1]− E[Pij|X3i =0], can be estimated usingthe two data sets. First, using my secondary data set (NELS), I can !nd an upper boundfor the mean fraction of those interested in high school who persist in their pursuitof teaching careers, E[Pij|X3i = 1]. Though I do not know the actual career outcomes,interest in teaching in the follow-up year of sophomore year of college serves as agood proxy for actually going on to teach for those who were already interestedduring high school. Let Sij be a dummy variable equal to one if the individual revealsan interest in a teaching career during the sophomore year of college, where interest isde!ned in the same fashion as for X3i. I use E[Sij|X3i =1] as an upper bound estimatefor E[Pij|X3i = 1]. I argue that this measure is an upper bound, because the attritionrate from interest in a teaching career is likely higher than the re-entry rate, i.e., it isprobably the case that more individuals with X3i = 1 and Sij = 1 end up with Pij = 0than those with X3i = 1 and Sij = 0 end up with Pij = 1. Further support for usingE[Sij|X3i=1] as a proxy for E[Pij|X3i=1] is provided by the !nding that this producessimilar estimates concerning entry into teaching as those found by Hanushek and Pace(1996). 15 I can then calculate E[Pij] using my primary (B&B) data set and solve for

E[Pij|X3i = 0] =E[Pij]− E[X3i = 1]E[Pij|X3i = 1]

1− E[X3i = 1]: (6)

It is worth noting that both of my interest in teaching measures, X3i and Sij, captureinterest in teaching in either a public or private school. Though I will use these variablesto estimate the unbiased value of �4 for some models in which the dependent variableequals one only if an individual teaches in a public school, the magnitude of the omittedvariable bias is probably measured more accurately in models where the dependentvariable also equals one if an individual becomes a private school teacher.

6.3. Baseline results ignoring the omitted variable bias

I conduct maximum likelihood estimation of each probit model based on Eq. (3),using the B&B sample probability weights and adjusting standard errors for clusteringat the college level. Inclusion of the sample weights is necessary due to endogenoussample strati!cation. Accounting for these weights, I estimate robust standard errors asderived by Wooldridge (1999).As a baseline, one may estimate Eq. (3) using the full, top three rating group

sample, and de!ning Pi as whether the individual teaches in a public school by 1997.For this model containing all of the top three groups, I add dummy variables forthe Barron’s selectivity rating level of the individual’s college as control variables

15 Hanushek and Pace (1996) estimate that only about 40% of recent college graduates who taught in themid-1980s reported that they were interested in a teaching career when they were high school students.In the NELS data that I use, the fraction of sophomores interested in teaching who were also interestedin teaching when they were high school seniors is 18%, 38%, 39%, and 50% respectively from the mostselective to the least selective rating groups (where the last group includes all schools in the fourth-highestcategory or lower).


Table 4Factors inNuencing entry of college graduates into public school teaching: probit estimates using B&B datafor graduates from colleges in Barron’s top three selectivity rating groups (dependent variable: dummy forwhether individuals receiving B.A. in 1994 taught in a public school by 1997)

Independent variables Coe.cient Standard error

College o�ers any undergraduate teacher 0:254∗ 0.128certi!cation program? (X4j)

Institutional characteristicsCollege is in most competitive rating group(1)? 0.117 0.169College is in highly competitive rating group(2)? −0:120 0.117Public college? 0.212 0.213Religiously a.liated? −0:081 0.134Ph.D. o�ered in any subject? −0:288∗∗ 0.105Undergraduate enrollment 1:2× 10−5 3:2× 10−5

(Undergrad. Enroll)∧2 −1:12× 10−9 1:51× 10−9

(Undergrad. Enroll)∧3 1:73× 10−14 2:09× 10−14

Individual characteristicsMinority (non-white) AND male −0:679∗∗ 0.164Minority AND female −0:031 0.172White AND male −0:582∗∗ 0.088Both SAT and ACT scores available 0.887 0.616Only SAT scores available 0.543 0.460Only ACT scores available −0:172 0.383SAT verbal score (100s) −0:016 0.079SAT math score (100s) −0:104 0.072SAT verbal score (100s) with ACT also available −0:202 0.131SAT math score (100s) with ACT also available −0:024 0.127ACT score 0.014 0.016ACT score with SAT also available 0.016 0.034

Parental characteristicsMother: has B.A. only −0:341∗∗ 0.074Father: has B.A. only 0.084 0.102Mother: has graduate degree −0:168 0.106Father: has graduate degree 0.070 0.103Either parent: high school dropout −0:156 0.133Either parent: Foreign born −0:109 0.127

N = 3061 Pseudo R2 = 0:109

+ signi!cant at 10%; ∗ signi!cant at 5%; ∗∗ signi!cant at 1%. Uses B&B data with sample probabilityweights. Excludes graduates from colleges not o�ering any certi!cation programs, as well as graduates whowere over the age of 21 when they began college or over the age of 29 when they graduated. Independentvariables also include state dummy variables and an intercept term. Standard errors are robust and adjust forclustering at the college level.

and also add state dummy variables. The results for this model are presented inTable 4. As expected, �4 is positive (see !rst row), which means that attending acollege with a four-year undergraduate certi!cation program is positively associated


ARTICLE IN PRESS

with entry into public teaching. Using the estimated average marginal e�ect 16 associ-ated with �4, a student attending a college with a UTCP is 3.4 percentage points morelikely to enter teaching (compared to a mean likelihood of just under 10%).As discussed in Section 3, one would expect the presence of UTCPs to have vary-

ing impacts for students attending colleges in di�erent selectivity groups. Thus, a betterspeci!cation for !nding the impact of UTCPs is to estimate Eq. (3) separately for eachrating group. The estimates of the �4 coe.cients for various rating groups and variousteaching outcomes are displayed in Table 5. In addition to point estimates, Table 5reports the average marginal e�ect of UTCPs for each group. Regardless of the teach-ing outcome used as the dependent variable, the coe.cient estimates in Table 5are only statistically signi!cant when the sample is limited to the second highest ratinggroup. Table 5a shows that having a UTCP is associated with higher rates of entryinto public school teaching at statistically signi!cant levels for individuals who attenda college in the second-highest rating group. Table 5b shows that having a UTCPis associated with an even larger e�ect on entry into either public or private schoolteaching for this group. This suggests that the relationship between certi!cation op-portunities and entry into public school teaching is not simply due to a shifting awayfrom the private sector to the public sector; availability of public school certi!cationopportunities is associated with higher rates of entry into private school teaching. 17

Additional regressions (not reported here) reveal that, for the second-highest collegerating group only, the availability of UTCPs is associated with a statistically signi!cantincrease in the fraction of individuals who teach in public schools and report that theyintend to remain working in schools for the long term.To ensure that �4 is picking up the e�ect of having a shorter program and not the

e�ect of having more programs, I re-estimate the models of Table 5 with the additionof an independent variable to control for whether a college has both elementary andsecondary programs or both types of 4-year programs. The presence of both elementaryand secondary programs is in fact associated with higher entry into teaching, but theinclusion of this variable has little e�ect on the results. 18 The estimated marginal e�ectsand levels of statistical signi!cance corresponding to the �4 point estimates remain verysimilar to those of Table 5. Interestingly, the presence of both elementary and secondaryUTCPs does not have any signi!cant e�ect beyond that of having a UTCP at one level

16 All average marginal e�ect estimates in this paper may be interpreted as estimates of the population aver-age marginal e�ect, because I weight the marginal e�ect estimated for each observation by that observation’sinverse sample probability weight.17 Even if many private schools do not care whether a teacher is certi!ed, there are several reasons why a

UTCP might increase rates of entry into private school teaching: (i) Some individuals planning on teachingin private schools may choose to enroll in a UTCP program, in order to better prepare themselves forteaching or to bolster their resume. This experience may increase the likelihood that they remain interestedin teaching when they graduate. (ii) Some individuals who believe that they would like to teach in a publicschool in the future might begin their careers at private schools. (iii) The presence of UTCP programs mayspark interest in teaching among those who do not enroll in these programs, because there is accessibleinformation about teaching careers (e.g. administrative assistants, peers enrolled in the program) or greateron-campus recruitment of teachers.18 The inclusion of a variable indicating whether a college o�ers a combined certi!cation–Master’s degree

program also has no signi!cant e�ect on the results.


Table 5

A B C

(a) Relationship between colleges o�ering any undergraduate teacher certi!cation program and graduatesbecoming public school teachers (dependent variable: dummy for whether individuals receiving B.A. in1994 taught in a public school by 1997)Most competitive colleges (highest rating group)# of observations 300Estimate −0:173 Not enough positive Not enough positive

outcomes to identify outcomes to identifyStandard error (0.226)Avg. marginal e�ect −0:0143

Highly competitive colleges (second highest rating group)# of observations 801 801 751Estimate 0.309 0.316 0.698Standard error (0.181) (0.177) (0.390)Avg. marginal e�ect 0.0367 0.0361 0.0777

Very competitive colleges (third highest rating group)# of observations 2001 2001 2001Estimate 0.188 0.212 0.049Standard error (0.147) (0.179) (0.181)Avg. marginal e�ect 0.0312 0.0348 0.0082

Geographic controls None Regional dummies State dummies

(b) Relationship between colleges o�ering any undergraduate teacher certi!cation program and graduatesbecoming teachers (dependent variable: dummy for whether individuals receiving B.A. in 1994taught by 1997)Most competitive colleges (highest rating group)# of observations 300 Not enough positive Not enough positive

outcomes to identify outcomes to identifyEstimate 0.093Standard error (0.114)Avg. marginal e�ect 0.0178

Highly competitive colleges (second highest rating group)# of observations 801 801 780Estimate 0.204 0.320 0.993Standard error (0.130) (0.169) (0.225)Avg. marginal e�ect 0.0361 0.0545 0.1592

Very competitive colleges (third highest rating group)# of observations 2001 2001 2001Estimate 0.139 0.194 0.117Standard error (0.126) (0.152) (0.160)

Avg. marginal e�ect 0.0287 0.0390 0.0233


Estimates are point estimates of �4 in Eq. (3) from probit estimation using (sample-weighted) B&B datadivided by college rating group. I applied the same sample restrictions described in the Notes to Table 4.Standard errors are robust and adjust for clustering at the college level.


ARTICLE IN PRESS

and a longer program at the other level. One cannot read very much into this !nding,since there is not much within-state variation between colleges o�ering both types ofUTCP, only one type of UTCP, or neither type of UTCP. However, this !nding looselysuggests that shortening both elementary and secondary programs from post-B.A. or!fth-year programs to 4-year undergraduate programs may not increase rates of entryinto teaching by much more than shortening one type of program to a UTCP.

6.4. Lower bound estimates of causal e6ects

As described in Section 6.2, I can now use these �4 estimates along with otherestimated parameters in Eq. (5) to estimate �4, the true e�ect of availability of UTCPson the likelihood that graduates become teachers. Table 6 displays the estimated valuesfor these parameters, estimates for �4, and the average marginal e�ect implied by these�4 estimates. As discussed in Section 6.2, my estimate of �3 is likely overstating thetrue magnitude of �3 so that these �4 estimates overcorrect for the bias. Since �3 isalways positive, the direction of the bias is the same as the sign of �4, so that thecorrection for this bias is in the opposite direction of the sign of �4. In the usual casewhere �4¿ 0, these adjusted marginal e�ects thus represent lower bound estimates forthe true marginal e�ect. For the other cases where �4¡ 0, the marginal e�ect estimateswithout the correction (displayed in Table 5), may be interpreted as the lower boundestimates of the true e�ect. For each model in Table 6, I display in italics the lowerbound estimates for the average marginal e�ect of UTCPs on the likelihood that onebecomes a teacher. These marginal e�ect estimates are slightly incorrect, because Iam unable to correct for the omitted variable bias of estimates of the other observed,independent variables’ coe.cients.Since only the second-highest rating group has statistically signi!cant coe.cients

in Table 5, the estimates for the marginal e�ects in Table 6 for this group are themost interesting. These estimates remain quantitatively signi!cant for entry into publicschool teaching. Based on the model that includes state !xed e�ects, column C ofTable 6, I estimate that the availability of a UTCP increases the likelihood of enteringpublic school teaching by at least 3.11 percentage points. If one changes the dependentvariable to include private school teaching and conducts similar analyses, then thislower bound estimate increases to 9.53 percentage points. This increase may be due tomore accurate adjustments for the omitted variable bias when private school teachingis included (discussed in the end of Section 6.2), as well as the actual causal impact ofUTCPs on entry into private school teaching (discussed in footnote 17). These estimatesare both quite signi!cant in magnitude when one considers that the mean entry ratesamong graduates from colleges in this rating group that did not o�er a UTCP were4.6% for public schools and 7.5% for public or private schools. Unfortunately, I amunable to calculate the standard errors for these estimates. I can simply observe that thecorresponding estimates ignoring the omitted variable bias were statistically signi!cant,and that estimates using an upper bound value for the size of the omitted variable biasremain quantitatively signi!cant.The marginal e�ect estimates for the highest rating group in Table 6 are not very

informative. Since there are not enough positive outcomes in the data for this group to


Table 6Lower bound estimates for the causal e�ect of colleges o�ering an undergraduate teacher certi!cation programon the likelihood that graduates become public school teachers

A B C

Most competitive colleges (highest rating group)# of NELS observations 204�4 −1:66Standard error of �4 (0.633)�̂3 (upper bound estimate for �3) 0.573

Not enough positive outcomes to identify�̂4 0.750

(Sample-weighted) average value ofdPijd�̂4

0.1150

Lower bound for average marginal e6ect −0.0143

Highly competitive colleges (second highest rating group)# of NELS observations 546 525 512�4 −0:094 0.344 0.873Standard error of �4 (0.259) (0.250) (0.454)�̂3 (upper bound estimate for �3) 0.486 0.486 0.486�̂4 0.389 0.184 0.352


0.0488 0.0192 0.0311

Lower bound for average marginal e6ect 0.0367 0.0192 0.0311

Very competitive colleges (third highest rating group)# of NELS observations 998 968 900�4 0.512 0.607 0.785Standard error of �4 (0.217) (0.229) (0.309)�̂3 (upper bound estimate for �3) 0.612 0.612 0.612�̂4 −0:093 −0:123 −0:423


−0:0130 −0:0163 −0:0541

Lower bound for average marginal e6ect −0.0130 −0.0163 −0.0541


Estimates use (sample-weighted) NELS and B&B data, divided by Barron’s college selectivity ratinggroup. See Section 6.2 for de!nitions of the parameters and Section 6.4 for a description of these results.For each column, I display in italic the (sample-weighted) average marginal e�ect corresponding to therelevant lower bound point estimate for the causal e�ect of an undergraduate teacher certi!cation programon the likelihood that one becomes a teacher.

estimate coe.cients for the state !xed e�ects model, it is worth verifying that the e�ectfor the most academically talented students would not be much larger if state dummyvariables were included. In order to expand the sample size of the most academicallytalented students, I divided students in the top three rating groups who took the SATinto three equally sized groups based on their SAT scores. For students in the topthird of SAT scores (students scoring 1230 and above), the estimated coe.cient onthe UTCP variable in the probit model with state !xed e�ects is positive, but remains


ARTICLE IN PRESS

statistically insigni!cant. This additional result suggests that one would likely !nd apositive, but small, baseline estimate if one could estimate the model with state !xede�ects for the highest rating group.For the third-highest college rating group, almost all of the marginal e�ect estimates

in Table 6 are negative. Furthermore, the estimated marginal e�ect for the model thatincludes state e�ects (column C) is quite large in magnitude: −5:41 percentage points.There is little reason to believe that the presence of UTCPs would have a substantial,negative impact on entry into teaching. This suggests that either my estimate of �4 forthis group is too low, or that the estimates of �4 or �3 in Eq. (5) may signi!cantlyoverstate the true values. One plausible explanation is that UTCPs have little causale�ect on entry into teaching for this group and that the observed control variableseliminate much of the potential omitted variable bias. If the latter phenomenon is alsotrue for the second-highest rating group, then the causal impact of UTCPs on entry intoteaching may be much larger for this group than the lower bound estimates suggest.

6.5. Other interactions

The impact of a UTCP on entry into teaching should vary, because certain types ofindividuals will be more likely to be on the margin concerning whether to teach andbecause the opportunity cost associated with pursuing teacher certi!cation after collegegraduation will also vary. One would thus want to know how each of the within ratinggroup estimates of Tables 5 and 6 changes when the sample is limited to certain typesof individuals or types of colleges. However, dividing the sample by college selectivityrating group and then again along another dimension would not yield enough positiveoutcomes to allow these e�ects to be identi!ed, especially when state !xed e�ectsare included. In order to gain information about the likely qualitative changes to theestimates, I estimate models similar to Table 4 using all students in the top threerating groups, but dividing the sample along various dimensions. The results are notpresented in this paper, but are available in an appendix upon request. One interesting!nding is that the relationship between a college having a UTCP and entry into publicschool teaching is stronger among public colleges. This relationship is also stronger forcolleges in which graduates are likely to reside 19 in states where there are: (1) fewercoursework requirements for certi!cation, (2) greater requirements for student teaching,(3) no emergency certi!cation programs, (4) relatively low starting salaries for teachers,and (5) relatively low salary premiums associated with !rst year teachers possessing aMaster’s degree. These !ndings suggest that UTCPs may be more likely to stimulateentry into teaching when it is relatively easy to balance UTCP coursework withinone’s undergraduate studies and when the !nancial return to pursuing certi!cation aftergraduating from college is relatively low.

19 I calculate migration probabilities for recent college graduates using an outside data source, the 1990Census Public-Use Micro-Sample, in order to avoid potential correlation between observed labor behaviorand migration.


7. Discussion

The previous section’s results suggest that whether a college o�ers a 4-year under-graduate teacher certi!cation program (UTCP) has a signi!cant, causal e�ect on entryinto teaching for individuals attending colleges in the second-highest college selectivityrating category. Regardless of the control variables included, the estimates presented inTables 5 and 6 suggest that UTCPs have a causal, positive impact on entry into teach-ing only for the second-highest rating group. Why does one observe this non-monotonicrelationship between college selectivity and the causal impact of UTCPs? As discussedin Section 3, for the highest rating group, the small e�ect is likely related to the factthat so few of these individuals are on the margin for deciding whether to teach. Forthe third-highest group, this weak relationship might be explained by another !nding:many students at these colleges are unable to complete their Bachelor’s degree orteacher certi!cation within 4 years. This renders shorter certi!cation programs useless.Do the large cross-sectional estimates for colleges in the second-highest rating group

suggest that these colleges should add UTCPs in order to improve the quality of pub-lic school teaching? To answer this, one must consider the other general equilibriume�ects from the addition of UTCPs, as well as the quality of training and prepara-tion o�ered by UTCPs. Since there is currently no source of empirical evidence onhow greater interest in teaching among graduates from highly selective colleges wouldchange teacher labor markets, one can only speculate about these other general equi-librium e�ects. The addition of UTCPs at colleges in the second-highest rating groupwould not have much impact on the national teacher supply curve. Consider the ex-treme case in which every college in the second-highest rating group was forced tohave a UTCP and graduates from these colleges who obtain certi!cation replace po-tential teachers from less selective colleges one-for-one. Based on my cross-sectionalestimates for the state-!xed e�ect models in column C of Tables 5 and 6, the fractionof new, young public school teachers who graduated from colleges in the highest tworating groups (roughly the top 15% of all graduates) would only increase from 9.1%to either 10.4% or 11.2%.Though this change may be a “drop in the bucket” on a national scale, one college

adopting a UTCP may in fact have a signi!cant e�ect on a local or state teachersupply curve if graduates tend to remain in this area. In a competitive labor market,one would expect employee compensation (wages, bene!ts, conditions, etc.) for workersof a given ability to fall as a result of a decrease in the entry cost of this type oflabor. However, the public school teacher labor market is far from competitive. Schooldistricts are usually constrained to o�er all teachers the same salary schedule, andunions often have a large inNuence on labor contracts. Thus, it may be the case that amore attractive pool of potential teachers causes some school districts to raise salariesto compete for the better ones. It may also be the case that some school districts willlower salaries because they have an easier time !nding certi!ed teachers due to anincrease in the total supply. The direction of salary movements, the types of districtsthat experience these salary movements, and the supply elasticities of various types ofpotential teachers will collectively either reinforce or diminish the change in interestin teaching among graduates from selective colleges.


ARTICLE IN PRESS

Even if salaries remain constant due to the rigidity of the labor market, there re-mains the issue of which people interested in teaching will gain employment. One of thebene!ts of expanding interest in teaching among the academically talented may be theexpansion of aggregate teacher supply to help school districts replace the “baby-boom”teachers expected to retire over the next decade. In many regions, the increased interestin teaching might simply o�set the projected rise in the rates of teacher retirement andincreases in the relative size of the school-aged population (Hussar, 1999). Teachershortages are already so severe in some urban districts that they have recently had torecruit teachers from foreign countries. In areas that do not experience a rise in thenet demand for teachers but that do currently allow emergency-certi!ed teachers, theseemergency-certi!ed teachers are likely the ones who would be crowded out in thelong-run, as districts will be forced to hire certi!ed teachers. This could in turn leadindividuals who formerly entered teaching through the emergency-certi!cation route topursue certi!cation prior to teaching, pursue teaching jobs in private schools, or notpursue teaching at all. In other areas, to the extent that public schools “hire the best,”the increased supply of potential teachers who graduate from selective colleges couldcrowd out teachers with less impressive backgrounds. The myriad potential generalequilibrium e�ects 20 prevent one from making any precise claims about how the frac-tion of individuals from selective colleges in particular types of school districts wouldchange.An additional consideration in analyzing the bene!ts of adding UTCPs at certain

colleges is that teacher quality will also be inNuenced by training. While there is ageneral consensus that teacher training can inNuence teacher quality, there is a long-standing debate over the ideal extent and nature of this training. In their review of theschool productivity literature, Monk and Rice (1999) !nd that teacher coursework inboth content area and pedagogical skills are important, but that there are diminishingreturns to training in the content area. They also report that “research shows the rela-tionship between graduate studies and teaching e�ectiveness to be modest (p. 132).”These !ndings both imply that there may be su.cient time within an undergraduate’s4 years to replace some of the higher level subject courses or electives with coursesaimed more speci!cally at future teachers, without compromising the teaching abilitiesof those pursuing certi!cation.This paper’s !ndings also support the idea that, holding teacher salaries and working

conditions constant, relaxing certi!cation requirements could increase the supply ofteachers by lowering entry costs. Of course, there would be many additional generalequilibrium e�ects associated with lowering certi!cation requirements or increasingalternative certi!cation routes. Whether the net e�ect of these policy changes wouldincrease or decrease the overall quality of teachers is ambiguous.

20 Another potential general equilibrium e�ect is changes in college attendance patterns. Potential changesin college attendance decisions could cause cross-sectional estimates to understate the e�ect of adding UTCPson entry into teaching among graduates of selective colleges. In particular, the policy change could increasethe likelihood that students with pre-existing interest in teaching attend selective colleges, since they no longerhave to consider sacri!cing other college characteristics (e.g. prestige, tuition, location) for the availabilityof certi!cation programs.


8. Conclusion

This paper !nds that adding undergraduate teacher certi!cation programs that maybe completed within 4 years would positively impact the supply of public schoolteachers from certain selective colleges. These programs could increase the averagerate of entry of graduates from these colleges into public school teaching by at least50%, provided that greater interest in teaching among these individuals would not leadto perverse general equilibrium e�ects due to changes in local labor markets. Theaddition of these programs would also likely increase the fraction of these individualswho pursue long-term careers in public schools and the fraction who become privateschool teachers. Though academic ability is just one aspect of teacher quality, increasingavailability of undergraduate teacher certi!cation programs may be a way to expand thesupply of teachers and to improve the distribution of academic ability among teacherswithout sacri!cing other characteristics, such as teacher preparation.

Acknowledgements

I would like to thank the National Center for Education Statistics, as well as JulieCullen, for providing access to the restricted-use versions of the National EducationLongitudinal Survey and the Baccalaureate and Beyond. I am grateful for !nancialassistance provided by the Population Studies Center at the University of Michigan andwould like to thank the research assistants who volunteered their time to assist withdata collection, especially Anish Ramachandran. Others providing resourceful assistancewere Susan Cho, Kathleen Gilbert, Lisbeth Goble, Najla Mamou, Anu Shetty, SusieSuh, and David Yu. I would also like to thank Charlie Brown, Julie Cullen, Je�Groen, Jim Hines, Caroline Hoxby, Larry Kenny, Shinichi Sakata, and Je� Wooldridgefor helpful discussions. I thank Larry Marsh and three anonymous referees, as well asparticipants in the May 2002 N.B.E.R. Higher Education Meeting, the University ofMichigan Business School Public Finance Seminar, and the 2002 American EducationFinance Association Conference for their helpful comments. Finally, I thank ClaudiaGoldin for providing data on gender segregation in colleges and thank Lisa Neidert forassisting me with the 1990 Census PUMS data.

References

Baccalaureate and Beyond Longitudinal Study, 1993–1997. Second Follow-up Methodology Report, 1999.National Center for Education Statistics (NCES 99-159), U.S. Department of Education, Washington, DC.

Ballou, D., 1996. Do public schools hire the best? Quarterly Journal of Economics 111, 97–133.Ballou, D., Podgursky, M., 1995. Recruiting smarter teachers. Journal of Human Resources 30, 326–338.Barron, 1998. Barron’s Pro!les of American Colleges, 1998. Barron’s Education Division.Coleman, J.S., et al., 1966. Equality of Educational Opportunity. U.S. O.ce of Education, Washington, DC.Ehrenberg, R.G., Brewer, D.J., 1994. Do school and teacher characteristics matter? Evidence from high

school and beyond. Economics of Education Review 13, 1–17.Ehrenberg, R.G., Brewer, D.J., 1995. Did teachers’ verbal ability and race matter in the 1960s? Coleman

revisited. Economics of Education Review 14, 1–21.


ARTICLE IN PRESS

Ferguson, R.F., 1991. Paying for public education: new evidence on how and why money matters. HarvardJournal of Legislation 28, 465–498.

Figlio, D.N., 1997. Teacher salaries and teacher quality. Economic Letters 55, 267–271.Hanushek, E.A., 1986. The economics of schooling: Production and e.ciency in public schools. Journal of

Economic Literature 24, 1141–1177.Hanushek, E.A., 1997. Assessing the e�ects of school resources on student performance: an update.

Educational Evaluation & Policy Analysis 19, 141–164.Hanushek, E.A., Pace, R.R., 1995. Who chooses to teach (and why)? Economics of Education Review 14,

101–117.Hanushek, E.A., Kain, J.F., Rivkin, S.G., 1998. Teachers, schools, and academic achievement. NBER

Working Paper #6691.Hanushek, E.A., Kain, J.F., Rivkin, S.G., 1999. Do higher salaries buy better teachers? NBER Working

Paper #7082.Hedges, L., Laine, R., Greenwald, R., 1994. Does money matter? A Meta-analysis of studies of the e�ects

of di�erential schooling inputs on student outcomes. Educational Researcher 23, 5–14.Hussar, W.J., 1999. Predicting the Need for Newly Hired Teachers in the United States to 2008-9. National

Center for Education Statistics (NCES 99-26), U.S. Department of Education, Washington, DC.Imazeki, J., 2001. Moving on or moving out?: determinants of job career changes for teachers. Mimeo, San

Diego State University.Loeb, S., Page, M., 2000. Examining the link between teacher wages and student outcomes: the importance

of alternative labor market opportunities and non-pecuniary variation. Review of Economics and Statistics82, 393–408.

Monk, D.H., Rice, J.K., 1999. Modern education productivity research: Emerging implications for !nancingeducation. In: Fowler Jr., W.J. (Ed.), Selected Papers in School Finance: 1997–1999. National Center forEducation Statistics, Washington, DC, pp. 111–139.

NASDTEC, 2000. NASDTEC Manual on the Preparation and Certi!cation of Educational Personnel. NationalAssociation of State Directors of Teacher Education & Certi!cation, Kendall/Hunt, Dubuque, Iowa.

Wooldridge, J.M., 1999. Asymptotic properties of weighted m-estimators for variable probability samples.Econometrica 67, 1385–1406.

Yatchew, A., Griliches, Z., 1985. Speci!cation error in probit models. Review of Economics and Statistics67, 134–139.

Documents

The impact of college course oerings on the supply of ...theunbrokenwindow.com/Higher Ed/Readings/Impact of College... · 2 Randall Reback/Journal of Econometrics( ) – ARTICLE IN