The Effects of the Kalamazoo Promise on
College Choice
Rodney J. Andrews1
University of Texas at Dallas
[email protected]
Stephen DesJardins
Center for the Study of Higher and Postsecondary Education
University of Michigan
[email protected]
Vimal Ranchhod
Post-Doctoral Research Fellow
Population Studies Center, University of Michigan and
SALDRU, University of Cape Town
[email protected]
September 22, 2009
1 This
research was supported by a grant from the Center for Local, State, and Urban
Policy at the University of Michigan. We thank James Maxey and ACT, Inc. for providing access to the unit record data. We thank the Upjohn Institute for providing details
about the administration of the the Kalamazoo Promise program. The paper has benefitted
greatly from the comments of Sue Dynarski, Stephanie Cellini, and Robert Toutkoushian.
Abstract
On November 10, 2005, then Superintendent of the Kalamazoo Public School
System, Janice Brown announced—to the surprise of Kalamazoo’s residents—the
beginning of the Kalamazoo Promise. Fully funded by a set of anonymous donors,
the Kalamazoo Promise is an urban revitalization program that offers up to four
years of free tuition to any public college or university in the state of Michigan
for graduates of the Kalamazoo Public School system who meet certain eligibility
criteria. Using the subsidy as a source of exogenous variation in the price of college, we use quasi-experimental methods to evaluate the impact of the subsidy on
college choice. We find that the Kalamazoo Promise increases the likelihood that
students from Kalamazoo Public Schools consider public institutions in Michigan.
In addition, we find that the Kalamazoo Promise especially impacts the college
choice set of students from families who earn less than fifty-thousand dollars in
annual income.
Introduction
To the surprise of the residents of Kalamazoo, Michigan, the Kalamazoo Promise
was announced on November 10, 2005. Fully funded by anonymous donors, the
Kalamazoo Promise offers to pay both the tuition and mandatory fees of graduates of Kalamazoo public high schools at any public college or university located
in Michigan. To be eligible for the scholarship program students must graduate
from a Kalamazoo public high school, reside in the school district, and have been
enrolled in the Kalamazoo Public School (KPS) district for four years or more.
Enrollment and residency must be continuous to be eligible for the financial support. Students must gain admission to and enroll in a public State of Michigan
community college, or four-year college or university. They must make regular
progress toward a degree or certification and maintain a 2.0 grade point average at
their postsecondary institution. Students must complete a minimum of 12 credit
hours per semester, and if their cumulative grade point average drops below 2.0,
they lose the funding, but it may be reinstated if students are able to bring their
grade point average up to at least a 2.0.1
However, the Kalamazoo Promise (henceforth, The Promise) is more than just
a college scholarship program. It is a multi-faceted approach, one component of
which is the tuition support for KPS high school graduates. The Promise is intended to spur urban renewal by attracting new residents who are interested in
having their children’s college subsidized; the scholarship program component is
designed to offer the KPS graduates the opportunity to attend college and hopefully return to work and live in Kalamazoo. The Kalamazoo Promise represents
a substantial philanthropic investment in the citizens of Kalamazoo and it is designed to continue indefinitely.
The urban development aspects of the program are interesting. However, examining the potential of The Promise as a catalyst for development is not the
primary focus of this paper. Herein, we focus on the effect of the scholarship on
college access. The unexpected announcement of The Promise, the significant
size of the subsidy, and the short time between the announcement of the program
and its implementation presents an opportunity to identify the causal effects of the
program on student college choice. To do so, we examine the ACT score report
sending behavior of the first cohort of students who are eligible for the scholarship. We follow Card and Krueger (2005) in using score reports as a proxy for a
1 The
source of the eligibility criteria is the official Kalamazoo Promise Website:
https://www.kalamazoopromise.com/
1
college application and do so because data detailing the early enrollment decisions
of students was unavailable early in the life of the program.2
Our research design exploits the variation in price induced by the Kalamazoo
Promise to examine the effects of changes in the price of some colleges on the
college choice set. The decision of where to go to college is surely among the
last of many decisions that lead to college attendance. The Kalamazoo Promise
is designed to bring about urban revitalization via gains in human capital acquisition. By lessening the price, the Kalamazoo Promise will likely increase the
academic preparation of future students as they take more difficult courses in anticipation of going to college. We should also expect to see college attendance
rising over time due to the following factors: students who absent The Promise
wouldn’t have sufficient funds to attend college will be able to afford to college, an
influx of students from families who relocated to take advantage of The Promise,
and changes in both the community’s infrastructure and attitudes towards college
attendance due to the publicity generated by the promise. Testing the hypotheses
that The Promise should produce more and better prepared college attendees are
fundamental research questions, but are not answerable at present as these effects
will emerge gradually and are outside both the time frame and data we consider
in this paper. Still, the question of The Promise’s influence on college choice is an
important one and central to the design of the subsidy because The Promise only
subsidizes public colleges and universities in Michigan.3
The Promise subsidizes public colleges and universities in Michigan under
the premise that locally educated citizens are more likely to contribute to the local economy. In the short period since its inception, The Promise has garnered
the attention of other communities and philanthropists. For example, there are
“Promise-like” programs in other locales such as Pittsburgh, Pennsylvania; El
Dorado, Arkansas and Denver, Colorado.4 Also, the state of Michigan is con2 Using
data from the University of California System and The College Board, Card and
Krueger (2005) regress the log of the number SAT score reports sent to one of eight UC campuses by members of a racial/ethnic group in a year on the log of the number of applications to the
same UC campus by members of the the same ethnic group-year cell. The result of the regression
is the amount of attenuation from score reports to applications. Their findings indicate that score
reports are a good proxy for applications.
3 We include as an appendix table a comparison between the actual number of promise recipients that enroll at various colleges in Michigan and the number of score reports sent to the same
colleges for students in Kalamazoo Promise schools in 2006.
4 See the following URL hosted by the Upjohn Institute for a complete list of communities with
“Kalamazoo Promise” style programs:
http://www.upjohninst.org/promise/communities.html
2
sidering creating “Promise Zones” that would encourage public and philanthropic
college financing initiatives modeled after The Promise. Given the interest in
these types of programs, this paper provides a crucial step toward determining the
effectiveness of such programs to influence prospective college students’ choice
of postsecondary institution.
Background
Economists have exploited many sources of variation to identify the effects of
changes in college price on college access and attendance.5 For example, within
state changes in tuition (Kane, 1994) and the GI Bill (Bound and Turner, 2002)
have been examined for their effects on college attendance. However, the studies most directly related to our situation are the evaluations that examine how
subsidies to students in specific geographic locations affect postsecondary attendance decisions. For example, Dynarski (2000) examines the impact of Georgia’s
HOPE Scholarship on college attendance. Funded by proceeds from state sanctioned gambling, the HOPE scholarship pays the tuition and fees of residents to
Georgia’s public colleges and universities or pays a fixed amount towards the tuition at a private colleges or university in Georgia. To qualify, Georgia residents
must have a 3.0 grade point average in high school and maintain a cumulative
3.0 grade point average in college. Dynarski finds that the HOPE Scholarship
increased college attendance among 18- and 19-year-olds by 7 to 8 percentage
points. Cornwell, Mustard, and Sridhar (2006) find that over the 1988-1997 period the HOPE Scholarship increased college enrollment by 5.8 percent or 2,889
students per year, with enrollment in four-year colleges accounting for most of the
enrollment gains.
Abraham and Clark (2006) examine the effects of the District of Columbia Tuition Assistance Grant (DCTAG) program. Initiated in 1999, the program subsidies District of Columbia residents’ attendance at public colleges and universities
in the United States. The authors find that the program increases the probability
that students apply to eligible institutions and increases college enrollment among
recent D.C. high school graduates. Kane (2007) also analyzed the DCTAG program, and his results indicate that between the years 1998 and 2000 the number
of District residents attending public institutions in Maryland and Virginia more
than doubled. He also examines the enrollment effects on public institutions in
5 Dynarski
(2002) provides a nice summary of some of this literature
3
other states and finds that the number of D.C. residents attending these institutions nearly doubled.
Following these scholars lead, we exploit a sharp change in the price of Michigan’s public colleges and universities for qualified high school graduates of the
KPS District and use this policy change to identify the effects of the scholarship
on college choice. Although our estimation strategy is similar to those employed
in the papers mentioned above, there are important differences. The papers that
analyze Georgia’s HOPE Scholarship rely on other states to provide the counterfactual to Georgia. That is, these other states act as a comparison group of what
educational outcomes in Georgia would have been absent the HOPE Scholarship.
In our paper we use ACT test-takers in public high schools in Michigan that did
not receive The Promise scholarship (“the treatment”) as the counterfactual.6 The
Georgia HOPE Scholarship also requires that recipients both finish high school
with a 3.0 grade point average and maintain a cumulative 3.0 in college. The
Promise imposes a lower academic standard (simply graduate from high school)
and is therefore more likely to affect the choice sets of students of lower socioeconomic status.
The DCTAG requires a minimum of twelve consecutive months of residency
prior to the applicants first time in college and continued residency throughout the
applicant’s college attendance. It is open to high school graduates—both public
and private—and GED recipients. Recipients of the subsidy offered by the DCTAG program are not required to attend high school in the District but they must
be residents, and the program applies to a larger set of postsecondary institutions
relative to The Promise. In contrast, The Promise subsidizes tuition at any of
Michigan’s public colleges and universities for KPS high school graduates who
have been enrolled in the district’s schools and been residents of the district for at
least four consecutive years. The percent of tuition covered depends on the length
of enrollment in KPS schools. Promise recipients who entered the KPS system
in kindergarten will have one hundred percent of their tuition covered. Promise
recipients who entered the KPS system in either the first grade, second grade, or
third grade will have ninety-five percent of their tuition covered. For Promise recipients who entered the KPS system in grades four through nine, the following
formula describes the percentage of tuition that is covered: .95−.05(G−3) where
G is the grade the student entered the KPS system. For example, if a promise eligible student entered in grade nine, then the percentage of tuition covered is sixtyfive percent. Students who enter the KPS system in grades ten through twelve are
6 Additional
details on the alternative definitions of the counterfactual are provided below.
4
not eligible for the subsidy.
The Promise is similar to Georgia’s HOPE Scholarship and the DCTAG in
that it reduces the price of some colleges for qualified residents. However, The
Promise offers the price subsidy to students in public high schools in a relatively
small geographic area, and the academic performance criteria are not stringent.
We believe our evaluation will add to the general knowledge of how students
respond to price subsidies and provide more specific knowledge about the effects
of The Promise on the postsecondary access of recipients. We know of no other
research that examines the effects of the Kalamazoo Promise on student access to
college.
Data
The primary source of data used in this study consists of unit record data from
ACT’s Student Profile Questionnaire for every ACT test taker in the state of
Michigan for the years 1996-2006. The information collected by the Student Profile Questionnaire (SPQ) is provided by the test taker when the prospective college
student registers for the ACT examination. It contains information on educational
plans, interests, and needs; special educational needs, interests, and goals; college extracurricular plans; financial and demographic information; out-of-class
accomplishments; and high school extracurricular activities. Our unit record data
not only contains a wealth of information from the SPQ, we also have each test
taker’s sub-scores on the English, mathematics, science, and reading sections, as
well as their overall or composite test score. We also know the set of (up to six)
institutions that the test-taker designates ACT to send score reports to. We know
the identity of these institutions and the test-takers order of preference—that is,
whether the institution is the test-taker’s first choice, second choice...sixth choice.
The ACT and the SAT are the two major college entrance examinations accepted by colleges in the United States. We use information from the ACT to
represent the preferences of Michigan’s college going population. Given that students can choose which examination to take, there is a possibility that the ACT
doesn’t accurately represent the college preferences of Michigan’s college-going
students. Data from the National Center for Education Statistics’ Digest of Education Statistics shows that for the years 1996–2006 roughly 10 to 11 percent of
Michigan’s high school graduates attempted the SAT.7 Information obtained from
7 See
http://nces.ed.gov/programs/digest/
5
ACT Inc. shows that for the the years 2001–2006 the percentage of Michigan’s
high school graduates that attempted the ACT was between 64 and 71 percent.8
Given that the majority of Michigan’s high school graduates attempt the ACT,
we are confident that the ACT yields an accurate representation of the college
preferences of Michigan’s college-going population.
To add more detail about the high schools students attend, we merge in high
school level data from the National Center for Education Statistics’ Common Core
of Data. The observations are indexed by a unique high school identifier and year,
so we add yearly information about the high school the ACT test-takers attended.
Added to the SPQ and ACT score data is information on the number of full time
equivalent teachers in the student’s high school, the pupil-to-teacher ratio, the
percentage of students that are eligible for free lunch, and the racial composition
of the high school. These covariates act as controls in our regressions, allowing
us to account for both individual and contextual factors that affect the outcome.
For an observation to be included in the full sample, we require the following:
valid ACT scores, information on the set of institutions designated to receive score
reports, family income data from the SPQ, race/ethnicity information, and nonmissing values for the annual high school data from the Common Core of Data.
As the Common Core of Data only has information on public schools, this means
that only students who attended public schools in Michigan are included in the
sample. The full sample includes 430,702 test-takers from 591 public high schools
in Michigan over the 1996 to 2006 time period.
Methodology
We employ difference-in-differences and difference-in-differences-in-differences
research designs to estimate the effects of The Promise on student college choice.
The unexpected announcement of the program, the substantial size of the subsidy,
and that we consider only the first cohort of students eligible for The Promise allow us to identify the causal effects of the price subsidy on student college choice.
The unexpected announcement of the program means that students could not have
taken actions to prepare for college in anticipation of receiving the subsidy. The
size of the subsidy is large enough to ease the cost of attending college—for example, the subsidy could have been as large as 10,573 dollars, the cost of attending
the University of Michigan-Ann Arbor as a lower division student for the 2005–
8 See
http://www.act.org/news/data.html
6
2006 academic year.9 The short time frame that we analyze means that we need
not worry about the confounding effects of either students who migrate into the
Kalamazoo School District to take advantage of The Promise or the effects of actions that the community has taken to complement the Kalamazoo Promise.10 In
the two paragraphs that follow, we discuss how The Promise affects the college
choice set.
Consider a hypothetical student in Kalamazoo, MI who has just discovered
that she is eligible to receive the tuition subsidy offered by The Promise. Further,
suppose that the student is a senior in high school. How might the tuition subsidy
influence her behavior? Given the timing of the announcement of the program
(November 2005), and the fact that our student has already started her senior year
in high school, she has little ability to alter her academic preparation for college.
Nonetheless, the subsidy should make college more affordable than prior to the
initiation of the program and in-state public institutions relatively more attractive
than out-of-state colleges. If our hypothetical student is interested in attending a
four-year institution, then she will likely have to submit a standardized examination score (such as the ACT) to any institutions to which she applies.
After attempting the ACT, she selects the optimal set of colleges given her
preferences and the cost of attendance. Upon receiving the information that she
is eligible to receive the subsidy offered by the Kalamazoo Promise, the in-state
public colleges and universities are more attractive to her than they would have
been absent the price subsidy. This change in the relative attractiveness of in-state
public institutions means that the probability that she and other eligible students
in public high schools in Kalamazoo will apply to Michigan’s public colleges and
universities increases.
We can estimate these changes at the high school level with the following
framework. Let YST represent the percentage of students at high school S in period T who send a score report to Western Michigan University (located in Kalamazoo). S takes on two values, K and N for Kalamazoo Promise high schools
9 http://sitemaker.umich.edu/obpinfo/files/umaa_tuitfee_
history.pdf is the source for the cost information.
10 Enrollment in the Kalamazoo Public school district increased in 2006 after declining for 17
years. The number of students enrolled in Kalamazoo Public Schools was 10,217 in September of 2005. That number had increased to 11,203 by September of 2006. (see http:
//www.upjohninst.org/promise/index.htm) The district also passed a eighty-five
million dollar bond request for infrastructure due to increased enrollment in May of 2006. The
number of volunteers for Big Brothers Big Sisters of Kalamazoo doubled in the years following
the announcement of the Kalamazoo Promise.(see Evergreen and Miron, 2008)
7
and Non-Promise high schools, respectively. T also takes on two values, 0 for
the period before the initiation of The Promise and 1 for the period after. The
difference-in-differences estimate of the effect of the program on the likelihood of
sending ACT test scores to Western Michigan University is:
∆ = (YK1 −YK0 ) − (YN1 −YN0 )
(1)
The difference in the terms YN1 and YN0 , the means of the percentage of students from Non-Promise high schools sending scores to Western Michigan University in the periods before and after the Kalamazoo Promise, estimates the score
sending behavior in the Kalamazoo Promise Schools in the absence of the policy change. The selection of an appropriate control group is important. We take
two approaches to constructing the control group: 1. We use Public High Schools
in Michigan that did not receive the Kalamazoo Promise. 2. We use a “matched
sample” of high schools. The sample selection procedure is discussed later in
this section. Subtracting the difference between the terms above from the difference between YK1 and YK0 —the means of the percentage of students from Promise
schools sending scores to WMU from the periods before and after The Promise,
respectively—removes the effects of factors that affect score report sending to
Western Michigan University that are common to both Promise high schools and
Non-Promise high schools. The difference in the differences provides an estimate
of the “true” impact of The Promise on ACT score sending to Western Michigan
University. This estimator is easily extended to estimating the effects of score
sending behavior to other colleges and universities.
The differences-in-differences framework is particularly well suited to eliminate the influence of factors that are fixed over time, but nonetheless affect college choice. Two examples of such factors are the existence of other scholarship/subsidy programs and the average distance to college. Let’s consider the
Heyl Scholarship, a competitive scholarship program that began in 1971 which
offers scholarships to students from either Loy Norrix High School or Kalamazoo Central High School who study the natural sciences, mathematics, or computer science at Kalamazoo College or study nursing at Western Michigan University.11 The existence of such a program, however, does not contaminate our
estimates of the impact of the Kalamazoo Promise because the effects of the Heyl
scholarship—which was in existence prior to The Promise and does not change
11 The
information on the Heyl Scholarships was obtained from the following website: http:
//www.kzoo.edu/heyl/contact.html and a phone conversation held on September 9,
2009 with Dr. Diane Kiino, the executive director of the Heyl Scholarship Fund.
8
during the span of our data—are eliminated by the first difference, (YK1 − YK0 ),
from Equation (1).
There is empirical evidence that suggests that distance to a particular college is
negatively related to the likelihood of a student selecting that college. (see Leppel,
1993; Hossler, Braxton, and Coppersmith, 1989; DesJardins, Dundar, and Hendel,
1999) Let’s consider the effects of distance on the likelihood of applying to the
Michigan State University. The students in the Kalamazoo Promise Schools are
the same distance from Michigan State both before and after the promise. If the
effects of distance on choice are constant over time, then, as in the case of the
Heyl Scholarship, the first difference, (YK1 − YK0 ), from Equation (1), removes
the effect of distance, and hence, our estimates of the effect of The Promise on the
likelihood of sending a score report Michigan State University are not confounded
by the effects of distance on the likelihood of sending a score report to Michigan
State University. Given our assumptions, only a factor that both influences the
likelihood of Kalamazoo Promise test-takers sending a score report to a particular
institution and becomes active at the same time as The Promise can contaminate
our estimates of the effects of The Promise on choice.
A test-taker’s individual characteristics likely influence the choice of college.
For example, a student who earns a high ACT score is more likely to apply to a
more selective college than a student who does less well on the ACT test. Failing
to account for the influence of individual characteristics on the decision of where
to send score reports can lead to biased estimates of the effects of The Promise on
our measure of college choice.
High school level characteristics also influence choice. Students from high
schools with populations that have low socio-economic status are less likely to
attend college. The inclusion of high school characteristics reduces the bias of
our estimated treatment effect.
To determine the effects of the Kalamazoo Promise on score report sending
while accounting for the influence of both high school characteristics and individual characteristics, we estimate models of the following form:
0
Yist = α + Xist
β1 + Xst0 β2 + γPOST + ηKP + δ KP × POST + εist
(2)
Yist is an indicator variable for test-taker i in high school s in year t that assumes
a value of one if the test-taker designates a particular school (or set of schools) to
receive a score report and takes on a value of zero otherwise. Xist is a vector of
attributes for test-taker i in school s in year t—for example, the test-taker’s ACT
9
scores and an indicator for the test-taker’s family income category. Xst is a vector
of covariates for high school s in year t. Examples of covariates that comprise
Xst are the number of full time equivalent teachers and the percentage of the high
school on free lunch. POST is an indicator variable that assumes a value of one
for the year that The Promise began, the 2005–2006 academic year. KP is an
indicator variable that assumes a value of one if the student is in a high school
that is eligible for the Kalamazoo Promise. This variable does not change over
time; that is, if a student is in one of the two high schools that are eligible for
The Promise in either the years before or the year following the implementation
of The Promise, KP takes on a value of one. POST × KP is an interaction term; it
is operationalized as an indicator variable that assumes a value of one for students
in high schools that are eligible for The Promise in the year it was announced and
zero otherwise.
The coefficient associated with POST , γ, is an estimate of the factors that
affect score-report sending for both Promise and Non-Promise test-takers for the
academic year The Promise was implemented. The coefficient associated with
KP, η, represents the average difference in the likelihood of sending a score report
to a designated institution between test-takers in Kalamazoo Promise high schools
from the years before the subsidy and students in Non-Promise schools from the
years before the subsidy.
We use a number of years of data prior to the implementation of The Promise.
This allows us to account for several things that could confound our estimates of
the effects of the Kalamazoo Promise. First, we are using repeated cross-sections,
so using a number of years of data of prior to the implementation of the Kalamazoo Promise allows us to accurately account for the effects of Xst and Xist —which
vary both cross-sectionally and temporally—on score report sending behavior.
Second, we include both linear and quadratic trends in our specifications. Using
a number of years of data prior to the implementation of the Kalamazoo Promise
allows us to more accurately account for the presence of trends which could bias
the estimates of The Promise. Third, using a number of years of data allows us
to more accurately control for factors that lead to differences in score report sending behavior between Kalamazoo Promise and Non-Promise Schools in the years
prior to the implementation of the Promise, which is represented by η.
We are most interested in the value of δ , the coefficient associated with the
interaction term POST × KP. It represents the difference-in-differences estimate
of the impact of the price subsidy on the likelihood of sending a score report—our
proxy for applying to college—to a particular set of institutions and is the regression adjusted analogue of ∆ in equation (1). The estimate of δ produced by the
10
difference-in-differences research design is an unbiased estimate of the treatment
effect if two conditions hold. First, there must be no changes in other factors affect
test-takers in the Kalamazoo Promise schools that coincide with the implementation of the Kalamazoo Promise. Second, the control group’s outcomes should
approximate what would have happened to the to test-takers in the Kalamazoo
Promise absent the subsidy. We estimate a series of linear probability models
with the standard errors clustered at the high school level. We cluster to deal with
both the inference issues that arise from using aggregate variables to explain the
behavior of micro-units as pointed out in Moulton (1990), and the serial correlation issues discussed in Bertrand, Duflo, and Mullainathan (2004).
Test-takers from families of limited means are more likely to be sensitive to
a subsidy to the cost of college than test-takers from richer families (see Leslie
and Brinkman, 1987; Heller, 1997; Gallet, 2007). The difference in sensitivity to
the price of college means that the effects of The Promise on application behavior
will likely vary by family income. To test this hypothesis, we estimate models of
the following form:
Yist = X 0 β + γPOST + ηKP + ρFI + δ1 POST × KP + δ2 POST × FI
+ δ3 KP × FI + δ4 POST × KP × FI + εist
(3)
Yist is defined above. X is a vector of characteristics that contains all the elements of Xist and Xst . POST , KP, and POST × KP are defined above. FI is an
indicator variable that assumes a value of one if the test-taker’s family income is
less than $ 50,000 per year. In these specifications we collapse the income indicators to the binary variable FI. We do this for two reasons. First, a total family
income of $ 50,000 per year is about the limit for Pell Grant eligibility; therefore,
we view $ 50,000 in income as a good benchmark for families who may find it difficult to pay for college. The second reason deals with estimation problems. There
are ten income categories. Attempting to estimate triple interactions for each of
the income categories means that our estimates would be based on a small number of respondents, increasing the likelihood of obtaining imprecise estimates. By
collapsing the categories, we lose the ability to estimate effects for the impact of
The Promise on all the income classifications, but we can more accurately estimate effects for a meaningful aggregate—that is, test takers from families that
earn less than $ 50,000.
POST × FI is an interaction term; it is operationalized as an indicator variable
that assumes a value of one if a test-taker’s family income is less than $ 50,000 and
11
the observation comes from the 2005–2006 academic year. The coefficient associated with the interaction term, δ2 , is the difference in the likelihood of submitting
a score report to a particular college for a test-taker with a family income that is
less than $ 50,000 in the 2005–2006 academic year. KP × FI assumes a value of
one for students in Kalamazoo Promise high schools with families that earn less
than $ 50,000. The coefficient associated with KP × FI, δ3 is the difference in the
probability of submitting a score report to a college for a test-taker in the eligible
high schools from families who earn less than $ 50,000 per year. The coefficient
of interest is δ4 , the difference-in-difference-in-differences estimate (DDD); it estimates the change in the likelihood of a test-taker in a Kalamazoo Promise high
school from a family with less than $ 50,000 in income sending a score report to
a particular college relative to to test-takers in Kalamazoo Promise high schools
from families who earn more than $ 50,000 in income. (The test-takers who earn
more than $ 50,000 in income are the control group.) Testing the hypothesis of
differences in price sensitivity amounts to inference on δ4 .
In estimating the effects of the Kalamazoo Promise on choice using the above
methodologies, we first use the entire sample of public highs schools in Michigan
that did not receive the Kalamazoo Promise as the “counterfactual” and depend on
the inclusion of covariates in a linear regression to control for differences between
ACT test-takers in Kalamazoo Promise high schools and the other public high
schools in our sample. We implement a combination of matching and regression
adjustment to both reduce the dependence on a linear specification and to select
a more comparable set of control schools. (see Rubin, 1973; Ho, Imai, King,
and Stuart, 2007) The procedure is implemented as follows: 1. We estimate the
propensity score or the likelihood of a public High school in Michigan receiving
the Kalamazoo Promise as a function of the High School’s 2005 characteristics.12
2. We match on the propensity score and select the public high schools in Michigan with values of the propensity score closest to the values obtained by the 2
Kalamazoo Promise High Schools. 3. The 37 control high schools and test-takers
in those schools that survive the matching and sample selection procedure and the
2 Kalamazoo Promise and the test-takers in the Kalamazoo Promise schools are
taken as the sample. We then estimate the parametric models described above on
this “matched sample”.
12 The
propensity score is a logit model with the following covariates entered linearly: the Percentage of a high school that is white, the percentage of a high school that is black, the percentage
of a high school that is Hispanic, the percentage of a high school on free meals, the percentage
of test-takers who come from families that earn less than fifty thousand dollars per year, and the
percentage of test-takers that send to a public college or university in Michigan.
12
In the following section, we discuss the results of our empirical investigation
of The Promise’s effect on choice.
Results
Difference-in-Differences
Full Sample
Table 1 summarizes the characteristics of the test takers in the sample. After the
establishment of the program there is a decline in performance on the ACT within
KPS (Promise) high schools—average scores drop on all four sections of the ACT
examination, with the drop ranging from 12 percent of a standard deviation in the
pre-period for Reading to a 25 percent of a standard deviation in the pre-period for
Science. The English and Math scores for Non-Promise test takers are unchanged
over time; whereas, the Science and Reading scores for Non-Promise test-takers
exhibited very small, but statistically significant, declines. The average GPA for
test-takers in the Promise schools declined by .04 points but rose by .08 points
in the Non-Promise high schools. These data are consistent with a story that the
Kalamazoo Promise incentivized students of lower academic ability to take the
ACT test, consider sending scores to, and applying to more selective colleges.
[Table 1 about here]
The racial/ethnic profile of the test-takers from the high schools that received
the subsidy differ greatly from the state average. The Promise schools have a
smaller fraction of white students and a larger fraction of black prospective college students relative to the averages in the Non-Promise schools. The Promise
and Non-Promise schools are nearly identical with respect to the fraction of testtakers who identify as being Latino. Kalamazoo Promise schools have higher
proportions of students from families that earn less than $ 50,000 per year.
Table 2 summarizes the characteristics of the high schools of the test-takers.
Kalamazoo Promise high schools have more full time equivalent teachers and
lower pupil-to-teacher ratios relative to the average of the Non-Promise Schools.
The Promise high schools have student bodies that have far higher fractions of
black students, Hispanic students, and students who receive free meals.
13
[Table 2 about here]
Table 3 contains the first set of difference-in-differences estimates, with the
column headings indicating the respective institution or combination of institutions where score reports were sent to by students in the sample. For example,
the U of M heading provides the results for changes in score sending behavior to
the University of Michigan-Ann Arbor campus. In this regression, the outcome is
a binary variable that assumes a value of one if a test-taker sends a score report
to the University of Michigan and assumes a value of zero otherwise. The table displays the estimates of the coefficients that are associated with the variables
of interest: POST , KP, and the interaction of these two variables (indicated by
POST × KP).
[Table 3 about here]
The second column of Table 3 contains the estimates for the University of
Michigan-Ann Arbor (denoted U. of M.). The estimated coefficient associated
with POST means that on average test-takers in the period following the announcement of the program are 4.24 percentage points more likely to send a score
report to the U. of M. than before the provision of the Promise scholarship. Also,
test-takers from the high schools that are eligible for the Kalamazoo Promise are
no more likely to send a score report to the U. of M. than their Non-Promise counterparts. The estimate is small (approximately .06 percentage points) and statistically insignificant. The difference-in-differences estimate of the impact of the
Kalamazoo Promise on the likelihood of sending a score report to the U. of M. is
7.84 percentage points and is statistically significant at the .1 percent level. This
estimate is roughly a 25 ( .0784
.32 ) percent gain relative to the mean percentage of
test-takers at Kalamazoo Promise schools who sent score reports to the University
of Michigan-Ann Arbor in the years prior to the announcement of the Kalamazoo
Promise.
Next, we estimate whether there were changes in score sending behavior to
Michigan State University. Test-takers are nearly 5 percentage points more likely
to send a score report to Michigan State University in the Post-Promise period.
Test-takers who attended Kalamazoo Promise high schools before the announcement of the Promise are nearly 3 percentage points less likely to send a score
report to Michigan State University, but this difference is measured imprecisely
and is not statistically significant. The difference-in-differences estimate of the
impact of the Promise on the probability of sending a score report to Michigan
14
State University is 12.5 percentage points or nearly 35 percent relative to the
mean percentage of score reports sent to Michigan State University by students
from Kalamazoo Promise high schools in the years prior to the Promise. The
estimate is statistically significant at the 1 percent level.
The fourth column contains the estimates for the probability of sending a score
report to Western Michigan University (WMU), the four-year research institution
located in Kalamazoo. All test-takers in the Post-Promise period increased their
likelihood of sending a score report WMU by slightly more than 2 percentage
points. This estimate is significant at the .1 percent level. Test-takers from Kalamazoo Promise high schools in the years prior to the program are 34.8 percentage
points more likely to send a score report to WMU. Given the close proximity of
WMU to KPS students, and given that students are more likely to attend colleges
that are close to home, this large estimate is not surprising. The difference-indifferences estimate of the impact of the program on the likelihood of sending
scores to WMU is 12.2 percentage points or 23 percent of the mean percentage
of students from Promise high schools that submitted a score report to WMU in
the years prior to the beginning of The Promise. But there is a caveat. In the
first year of The Promise, WMU initiated a program that offered to pay room and
board for Kalamazoo Promise recipients matriculating to the institution. Thus, the
difference-in-differences estimate is conflated with the effects of Western Michigans room and board subsidy. Given that we expect the room and board subsidy
to positively affect the likelihood that a student applies to WMU, this implies that
our estimate of the treatment effect is biased upwards.
The likelihood of test-takers in the post-period sending score reports to any of
the public universities in Michigan (Pub. U in MI) is both small (approximately
.2 percentage points) and statistically insignificant. Test-takers from Kalamazoo
Promise high schools are 4 percentage points less likely to send a score report to
these institutions (but this estimate is not significant at conventional levels). The
difference-in-differences estimate of the Promise on the likelihood of sending to
any of the public colleges and universities is 6.3 percentage points which is 8
percent of the mean percentage of test-takers in Kalamazoo Promise high schools
who sent score reports to such institutions in the years before The Promise was
announced.
Test-takers are approximately 1.5 percentage points less likely to send a score
report to Wayne State University (WSU), which is located in Detroit, Michigan.
Test-takers in Promise eligible high schools prior to the announcement are 14.2
percentage points less likely to send a score report to Wayne State University. The
difference-in-differences estimate is 6.5 percentage points, which is more than
15
three times the mean percentage of test-takers from Promise high schools who
sent score reports to WSU. This estimate should be viewed with some caution. On
November 17, 2005, just a week following the announcement of the Kalamazoo
Promise, WSU announced a program that offers a 50 percent reduction in the
cost of room and board to Kalamazoo Promise recipients who enroll at WSU.
This means that our estimate of the treatment effect is likely biased upwards,
as the difference-in-differences estimate, in part, picks up the effects of WSU’s
complimentary funding.
The next outcome we examine is the likelihood of sending a score report to
Kalamazoo College. Kalamazoo College is a private liberal arts college located
in Kalamazoo. Kalamazoo College provides a good contrast to the local public
institutions. It is located in close proximity to Kalamazoo Promise recipients, so
it is comparable to local public institutions with respect to distance. However,
Kalamazoo College is private, so recipients of The Promise cannot apply the subsidy towards tuition at Kalamazoo College. We find a small, .1 percentage point,
increase in the likelihood of all test takers sending a score report to Kalamazoo
College following the announcement of The Promise. Test takers from Kalamazoo Promise Schools are roughly 12.7 percentage points more likely to send a
score report to Kalamazoo College in the years before the announcement of The
Promise, which is sensible given Kalamazoo College’s proximity. The estimate
of the impact of The Promise on the likelihood of sending to Kalamazoo College
is -2.2 percentage points. The estimate is consistent with the hypothesis that The
Promise reduces the attractiveness of private institutions; however, the estimate is
not statistically significant.
The next two outcomes are particularly interesting. The eighth column provides estimates for Kalamazoo Valley Community College (KVCC) which is also
located in Kalamazoo. There is virtually no change in the likelihood of all testtakers in the post-period sending a score report to KVCC. However, test-takers
from Promise high schools in the years prior to the program are 16 percentage
points more likely to send a score report to KVCC This result is statistically significant at the .1 percent level. The difference-in-differences estimate is 2.4 percentage points or about 13 percent of the mean percentage of test-takers in Promise
eligible high schools before the Kalamazoo Program. KVCC is open enrollment,
that is, a high school diploma or its equivalent is all that is required for admittance.
That the Kalamazoo Promise program increases the probability of sending a score
report to a community college implies that the Promise affects a set of students
who, perhaps, are not prepared for an academically rigorous institution such as
WMU or prefer to experiment with college at an institution that provides the type
16
of academic courses offered by a community college. That KVCC is open enrollment means that our estimate understates the impact of The Promise on attending
KVCC. Students who attempt the ACT are likely to be more academically prepared than students who do not attempt an college entrance examination. Because
a college entrance examination is not required for attending KVCC, our data is
not representative of the set of students who are likely to be influenced by The
Promise to attend KVCC. Taken together, this means that the actual set of students who are influenced to consider KVCC because of The Promise and a lack
of academic preparedness is greater than our results indicate.
The final column contains estimates for the flagship universities located in
Michigan. The “Flagship” variable assumes a value of one if a test-taker sends
a score report to both the University of Michigan and Michigan State University.
Test-takers in the year after The Promise began are roughly 3.3 percentage points
more likely to submit score reports to both flagship institutions. Test-takers in
Promise schools in the years prior to The Promise are roughly 1.72 percentage
points less likely to to submit score reports to both universities; however, this estimate is statistically insignificant. The difference-in-differences estimate is 11.3
percentage points and is statistically significant at the .1 percent level. The estimate is relatively large at roughly seventy two percent of the Pre-Promise mean.
Thus far, the results provide strong evidence that The Promise affects the
choice set of potential applicants. However, the above estimates do not include
some covariates that may affect student decisions about where to submit a score
report. For instance, ones race and family income may affect an applicants propensity to send scores to particular institutions thus controlling for these factors may
be important. We do so, and the results are presented in Table 4. The results produced when we control for the race and income of the test-taker are very similar
to the results discussed above in Table 3.
[Table 4 about here]
Matched Sample
Table 5 contains the summary statistics for the set of ACT test takers from high
schools in the matched sample for the years prior to the Kalamazoo Promise. The
differences in the ACT test scores between Promise and Non-Promise test-takers
from the matched sample are smaller in magnitude relative to the differences in
test scores between Promise and non-Promise from the pre-period for the full sam17
ple in Table 1. The differences for the ACT English, ACT math, ACT Science, and
ACT Reading sections from the matched sample are .4 points, .4 points, .1 points,
and .1 points, respectively, smaller than the differences between Promise and NonPromise test-takers from the pre-period for the full sample. The difference in
the GPA between Promise and Non-Promise test-takers is slightly higher in the
matched sample, .15 grade points versus .11 grade points for the full sample. The
differences in racial/ethnic composition between Promise and Non-Promise testtakers from the matched sample in the pre-period are smaller than the differences
between the Promise and Non-Promise from the full sample in the pre-period. The
Kalamazoo Promise test-takers are 57 percent white versus 55 percent white for
the Non-Promise test-takers, a difference of 2 percentage points; this difference
for the full sample was 21 percentage points. The percentage of test-takers who
are black is higher for the Non-Promise test-takers from the matched sample, 27
percent black versus 25 percent for the Promise test-takers. The magnitude of this
difference is far smaller than the 15 percentage point difference between Promise
and Non-Promise students from the full sample. The difference between Promise
and Non-Promise test-takers in the percentage of test-takers from families who
earn less than fifty thousand dollars per year is 5 percentage points larger than the
analogous difference for the full sample.
[Table 5 about here]
Table 6 summarizes the high school characteristics for the set of high schools
in the matched sample. The magnitude of the difference between Promise and
Non-Promise high schools in the number of classroom teachers declines dramatically when comparing the matched sample to the full sample in the pre-period; the
difference goes from nearly 36 teachers, 78.0 Promise versus 42.1 Non-Promise,
for the full sample to less than one for the matched sample. The magnitude of the
difference for the Pupil to Teacher ratio increases for the matched sample relative
to the full sample. The pre-period differences between the Promise and NonPromise high schools for Proportion White and Proportion Black are a lot smaller
than the analogous difference for the full sample. The pre-period difference in
the Proportion Hispanic remains the same for both the matched and full sample.
The magnitude of the pre-period difference in the Proportion of the students who
receive free meals increases.
[Table 6 about here]
Overall, the matching and sample selection procedure produces a set of “counterfactual” test-takers and high schools with characteristics that are closer to the
values we observe for the test-takers and high schools that received the Kalamazoo
Promise. In the cases where the samples become more dissimilar, the magnitudes
18
of the differences are not large in absolute value. Table 7 contains the results for
the difference-in-differences analysis on the matched sample. The estimates presented in Table 7 are the matched sample analogue of the estimates presented in
Table 4.
[Table 7 about here]
The second column of Table 7 contains the estimates for the dependent variable that assumes a value of one if a test-taker chooses to submit a score report to
the University of Michigan-Ann Arbor. The difference-in-differences estimate of
the impact of the Kalamazoo Promise on the likelihood of sending a score report
to the U. of M. is 11.4 percentage points and is statistically significant at the .1
percent level. This estimate is more than 41 percent larger than the estimate obtained using the full sample and is equivalent to approximately a 36 percent ( .114
.32 )
increase relative to the Pre-Promise mean.
The following column contains the estimates of the impact of the Kalamazoo
Promise on the likelihood of sending a score report to Michigan State University.
The estimated treatment effect is slightly smaller than the estimate from the full
sample; We estimate an 11.3 percentage point increase for the matched sample
versus a 12.5 percentage increase for the full sample. The estimated treatment
effect for the matched sample is approximately 32 percent of the Pre-Promise
mean.
The fourth column of Table 7 contains the estimates for Western Michigan
University. The estimated treatment effect on the likelihood of sending a score
report to the Western Michigan University is 10.8 percentage points, which is 1.4
percentage points smaller than the analogous estimate for the full sample,‘and is
equivalent to nearly a 21 percent increase relative to the Pre-Promise mean. The
caveat the we offered for the estimate of the treatment effect for the estimated
treatment using the full sample also applies here. Western Michigan University
initiated a program that offered to pay room and board for Kalamazoo Promise recipients matriculating to the institution, so our estimate of the impact of the Kalamazoo Promise on score report sending is conflated with the effects of WMU’s
targeted recruiting and is biased upwards.
The fifth column contains the estimates of the impact of the Kalamazoo Promise
on the likelihood of sending to a Public University in Michigan for the matched
sample. The estimate takes on a value of 6.34 percentage points, which is extremely close to the full sample estimate of 6.52 percentage points. Relative to
the Pre-Promise mean rate of sending to public universities in Michigan, this treatment effect is equivalent to approximately a 8.5 percent increase.
19
In the sixth column, we present estimates for the effects of The Promise on
the likelihood of sending a score report to Wayne State University. We estimate a
treatment that the Kalamazoo Promise increases the likelihood of sending a score
report to Wayne State University by 6.31 percentage points. This estimate is more
than three times the Pre-Promise Mean percentage of students who submitted a
score report to Wayne State University prior to The Promise. Again, this estimate
is likely biased as Wayne State University implemented a program that reduces
the cost of room and board for Kalamazoo Promise Recipients.
The seventh column contains the estimates for Kalamazoo College. We estimate that the Kalamazoo Promise reduced the likelihood of sending a score report
to Kalamazoo Promise by nearly 2 percentage points, which is nearly a 15 percent
reduction relative to the Pre-Promise mean. This estimate, however, is statistically
insignificant.
The eighth column contains the estimates for Kalamazoo Valley Community
College. The Promise increases the likelihood of sending a score report to Kalamazoo Valley Community college 1.9 percentage points. This estimated effect
is nearly a 12 percent change relative to the Pre-Promise mean. This estimate
understates the true impact of The Promise on KVCC because KVCC is open enrollment and our data is not representative of the set of students who are likely to
consider community college as a postsecondary option.
The final column of Table 7 contains the estimates of the likelihood of sending a score report to both the University of Michigan-Ann Arbor and Michigan
State University. We estimate that The Promise increases the likelihood of sending a score report to both of Michigan’s flagship institutions by 12.5 percentage
points. This effect is more than 78 percent of The Pre-Promise mean and is highly
significant.
With the exception of the estimates of the impact of The Kalamazoo Promise
on the likelihood of sending a score report to the University of Michigan, the estimated treatment effects are similar in sign, magnitude, and statistical significance
to the difference-in-differences estimates contained in Table 3. The evidence from
the results obtained using both the full sample and the matched sample indicate
that The Kalamazoo Promise has large and significant effects on the college choice
set.
20
Difference-in-Difference-in-Differences
Full Sample
In this section, we discuss the results from estimating the model detailed in (3).
For the sake of brevity, we focus the discussion on δ4 , the coefficient associated
with the triple interaction term POST × KP × FI. The DDD estimate of the impact of the Kalamazoo Promise on the likelihood of a test-taker from a family
that earns less than $ 50,000 per year submitting a score report to the University
of Michigan-Ann Arbor is 3.4 percentage points and significant at the .1 percent
level. This is a 13.8 percent gain relative to the mean percentage of test-takers
who sent score reports to the U of M and meet the following criteria: a) from
families that earn less than $ 50,000 per year, b) attend one of the Promise eligible
high schools, c) submitted a score report in the years prior to the announcement
of the Promise.13
[Table 8 about here]
The DDD estimate of the effects of The Promise on the likelihood of submitting a score report to Michigan State University is 7.8 percentage points and is
statistically significant at the .1 percent level. This is nearly a twenty-five percent
gain relative to the Pre-Promise of score reports sent to Michigan State University.
With respect to Western Michigan University, the DDD estimate the impact of the
Promise is 2.9 percentage points. This is a 5.6 percent gain relative to the PrePromise mean. We should be careful, however, in attributing this effect solely to
the Kalamazoo Promise. Recall that Western Michigan University implemented
a program that offers to pay room and board for Kalamazoo Promise recipients
who are admitted. This implies that the DDD estimate is likely biased upwards as
Kalamazoo Promise recipients are eligible for the funds and the provision of the
additional support by WMU increases the likelihood of applying to the institution.
The DDD estimate of the effects of The Promise on the likelihood of submitting a score report to a Public University in Michigan is approximately 3.2
percentage points and is statistically significant at the five percent level. This is
roughly a 4.1 percent gain relative to the to the Pre-Promise mean.
Next, we consider the DDD estimate for Wayne State University. The DDD
estimate for the impact of the Kalamazoo Promise on sending to WSU is 5.4 per13 These
criteria define the population that is used to compute the Pre-Promise mean for this
sub-section.
21
centage points. The estimate is statistically insignificant but large relative to the
Pre-Promise mean; the estimate is 1.8 times as large as the Pre-Promise mean.
Recall that WSU also instituted a program that offered to pay room and board
for Promise recipients, so this estimate is also probably biased upward for reasons
similar to the bias that likely contaminates the estimate for Western Michigan University. The estimate of the effects of The Promise on sending to the local private
four-year institution, Kalamazoo College, for students from families that earn less
than $ 50,000 is small, ≈-.5 percentage points, and statistically insignificant.
Results for Kalamazoo Valley Community College are interesting. First, testtakers who earn less than fifty thousand dollars per year in Kalamazoo Promise
schools are 8.3 percentage points more likely to submit a score report to KVCC,
ceteris paribus. After the Promise is implemented, however, test-takers in Kalamazoo Promise Schools who earn less than $ 50,000 per year are 10.3 percentage
points less likely to submit a score report to KVCC. This is a 49 percent decline
relative to the Pre-Promise mean. This estimate should be viewed with some caution. The set of ACT test-takers represents set of students who are likely better
prepared academically. In particular, we are likely estimating the effect of relaxing the budget constraint for a set of relatively well-prepared students who are
of lower socioeconomic status and are choosing to exclude the KVCC from the
choice set at a rate higher than they would have absent. However, due to the limitations of the data, we are unable to observe students of lower socioeconomic
status who do not attempt the ACT, who are likely less prepared academically,
and thus more likely to use The Promise to attend a community college.
The DDD estimate for the impact of The Promise on sending score reports
to both the University of Michigan-Ann Arbor and Michigan State University is
small, ≈ -.003 percentage points and statistically insignificant.
Overall, the full sample DDD estimates using the provide evidence that testtakers from families that earn less than $ 50,000 respond differently to treatment
than their counterparts who are from families with incomes above fifty thousand
dollars per year.
Matched Sample
[Table 9 about here]
Table 9 contains the difference-in-differences-in-differences estimates for the
matched sample. The second column contains the results for the University of
Michigan. The estimated treatment effect is .824 percentage points and is statis22
tically insignificant. The magnitude of the estimated treatment effect represents a
more than 76 percent decrease relative to the estimate obtained using the full sample. This effect is roughly a 3.3 percent increase relative to the pre-promise for
test-takers from Kalamazoo Promise schools for the years prior to The Promise
who are from families that earn less than $ 50,000 per year.
The estimated treatment effect for the impact of The Promise on the likelihood
of a test-taker from a family that earns less than $ 50,000 per year sending a score
report to Michigan State University is 9.90 percentage points and is statistically
significant. This effect is larger than estimate obtained using the full sample and
is equivalent to a nearly 31 percent increase relative to to the pre-promise for testtakers from Kalamazoo Promise schools for the years prior to The Promise who
are from families that earn less than $ 50,000 per year.
The fourth column of Table 9 contains the estimates for Western Michigan
University using the matched sample. The treatment effect is .0224 percentage
points which is an approximate 24 percent decrease in magnitude relative to the
full sample. The estimate is statistically insignificant. Because of the presence of
a contemporaneous and complimentary program instituted by Western Michigan,
this estimate is likely upwards biased.
Column five contains the difference-in-differences-differences estimates for
Public Colleges and Universities in Michigan. The DDD estimate is .0155 percentage points and is statistically significant. Relative to the full sample results,
this estimate declines by about 50 percent in magnitude and loses statistical significance.
At .0596 percentage points, the DDD estimate for the impact of sending a
score report to Wayne State University is larger than the estimate obtained using
the full sample, but like the full sample estimate, is not statistically significant.
The estimate is nearly two hundred percent of the pre-promise mean for Kalamazoo Promise schools for the years prior to The Promise who are from families that
earn less than fifty-thousand dollars per year. The estimate is also likely biased
as Wayne State University implemented a contemporaneous program that reduced
the costs of room and board for Kalamazoo Promise recipients.
The DDD estimate for effects of the Kalamazoo Promise on sending a score
report to Kalamazoo College is small, ≈ .005 percentage points, and statistically
insignificant. The estimate takes on the opposite sign of the estimate obtained
using the full sample, but both estimates are substantively small and imprecisely
estimated.
Test-takers in Kalamazoo Promise Schools who earn less than $ 50,000 per
year are 11.7 percentage points less likely to submit a score report to KVCC after
23
The Promise is implemented. This is a 56 percent decline relative to the PrePromise mean. Like its full-sample analogue, this estimate should be viewed
cautiously for the same reasons we discuss in the previous section.
There is variation between the estimates DDD estimates obtained from using the full sample and the DDD estimates we obtain from using the matched
sample. This is not surprising because we are using two different samples. The
estimates obtained using the full sample can viewed as an “Average Treatment Effect” for public high schools in Michigan. The treatment effect we estimate using
the matched sample can be viewed as an estimate of “Treatment on the Treated”,
as our estimating sample consists of the high schools that receive the treatment
and public high schools that have similar values of the estimated likelihood of
receiving treatment. The DDD results suggest that test-takers who receive The
Promise and come from families that earn less than $ 50,000 per year increase the
probability that they send a score report to Michigan State University and decrease
the likelihood of sending a score report to KVCC.
Conclusion
We provide evidence that test-takers respond to price incentives. Our results
demonstrate that the Kalamazoo Promise incentivizes students to include Michigans public colleges and universities in their college choice sets after the implementation of the program. Moreover, we show that the effects are fairly large
for Michigans most selective public universities, the University of Michigan-Ann
Arbor and Michigan State University, which also are the institutions whose graduates are likely to be the most mobile. We find that the Promise also generates
interest in two of Kalamazoos public colleges, Western Michigan University and
Kalamazoo Valley Community College. Given the urban development focus of
the Kalamazoo Promise program, the findings that show that the recipients of the
Kalamazoo Promise are more likely to consider local institutions is welcome news
as Groen and White (2004) show that in-state students who are educated by public
institutions are more likely to remain in-state. Still, if the the job opportunities are
not available in the city of Kalamazoo, then citizens, whether educated locally or
not, will go elsewhere.
The DDD estimates provide evidence that ACT test takers from families that
earn less than $ 50,000 that receive The Promise respond differently than their
more affluent counterparts. We find little evidence that The Promise affects the
likelihood of such students considering a local private school, Kalamazoo Col24
lege. Promise recipients that earn less than $ 50,000 are more likely to send score
reports to Michigan State University. However, Promise recipients that earn less
than $ 50,000 are less likely to send score reports to Kalamazoo Valley Community College, a cheaper option. Taken together, these estimates suggest that The
Promise allows test-takers who are financially constrained to consider institutions
that are higher priced and more selective.
The goal of The Promise is to revitalize the city of Kalamazoo by securing the
benefits of an educated citizenry—for example, high-tech industries may be more
likely to locate in Kalamazoo if there is an appropriately skilled labor force that
is readily available. By subsidizing the cost of public post-secondary institutions
in Michigan, the Promise aims to attract families who are interested in securing
these benefits for their children and increasing the likelihood that educated citizens
remain in the area. This realization relies on families and students being interested
in considering local institutions. This paper demonstrates that the Promise, at a
minimum, has stimulated interest in Michigan’s local institutions.
We use exogenous variation in the price of college to identify the effects of
price on the college choice set. We add to the literature by demonstrating that a
localized program that is wholly funded by philanthropy is capable of changing
the college choice set. As these programs diffuse across the rest of Michigan and
across the country, objective evaluation of the effects of such programs is needed
to ensure that the benefits of such large gifts are maximized.
References
Abraham, K. G. and M. A. Clark (2006). Financial aid and students college decisions: Evidence from the district of columbia tuition assistance grant program.
Journal of Human Resources 41(3), 578–210.
Bertrand, M., E. Duflo, and S. Mullainathan (2004). How much should we trust
difference in differences estimates? Quarterly Journal of Economics 119(1),
249 – 275.
Bound, J. and S. Turner (2002). Going to war and going to college: Did world
war ii and the g.i. bill increase educational attainment for returning veterans?
Journal of Labor Economics 20(4), 784–815.
Card, D. and A. Krueger (2005). Would the elimination of affirmative action
25
affect highly qualified minority applicants? evidence from california and texas.
Industrial and Labor Relations Review 56, 416 – 434.
Cornwell, C., D. B. Mustard, and D. J. Sridhar (2006). The enrollment effects of
merit-based financial aid: Evidence from georgia’s hope program. Journal of
Labor Economics 24(4), 761–786.
DesJardins, S. L., H. Dundar, and D. D. Hendel (1999). Modeling the college
application decision process in a land-grant university. Economics of Education
Review 18(1), 117 – 132.
Dynarski, S. (2000, September). Hope for whom? financial aid for the middle
class and its impact on college attendance. National Tax Journal 53(3), 629–
662.
Dynarski, S. (2002). The behavioral and distributional implications of aid for
college. American Economic Review 92(2), 279–285.
Evergreen, S. and G. Miron (2008, February). Community group response. Evaluation of the Kalamazoo Promise Working Paper #2, Western Michigan University.
Gallet, C. (2007). A comparative analysis of the demand for higher education:
Results from a meta-analysis of elasticities. Economics Bulletin 9(7), 1–14.
Groen, J. A. and M. J. White (2004, August). In-state versus out-of-state students:
the divergence of interest between public universities and state governments.
Journal of Public Economics 88(9-10), 1793–1814.
Heller, D. (1997). Student price response in higher education: An update to leslie
and brinkman. Journal of Higher Education 68(6), 624–659.
Ho, D. E., K. Imai, G. King, and E. A. Stuart (2007). Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal
Inference. Political Analysis 15(3), 199–236.
Hossler, D., J. Braxton, and G. Coppersmith (1989). Understanding Student College Choice, Volume 5, pp. 231–288. New York: Agathon Press.
Kane, T. J. (1994, October). College entry by blacks since 1970: The role of college costs, family background, and the returns to education. Journal of Political
Economy 102(5), 878–911.
26
Kane, T. J. (2007). Evaluating the impact of the d.c. tuition assistance grant program. Journal of Human Resources 42(3), 555–582.
Leppel, K. (1993). Logit estimation of a gravity model of the college enrollment
decision. Research in Higher Education 34, 387–398.
Leslie, L. L. and P. T. Brinkman (1987). Student price response in higher education. Journal of Higher Education 58(1), 181–204.
Moulton, B. R. (1990). An illustration of a pitfall in estimating the effects of aggregate variables on micro units. The Review of Economics and Statistics 72(2),
334 – 338.
Rubin, D. B. (1973). Matching to remove bias in observational studies. Biometrics 29(1), 159–183.
27
28
20.1
(5.9)
20.8
(5.3)
20.9
(4.8)
21.6
(5.9)
3.31
(.60)
.55
(.50)
.26
(.44)
.02
(.13)
.51
(.50)
Post
−1.1
(.4)
−1.1
(0.4)
−1.2
(.3)
−0.8
(.4)
−.04
(.04)
−.02
(.03)
.01
(.03)
.00
(.01)
.03
(.03)
Post−Pre
Post
20.5 20.5
(5.4) (5.4)
21.3 21.3
(5.0) (4.9)
21.7 21.6
(4.5) (4.4)
21.7 21.6
(5.9) (5.7)
3.24 3.32
(.59) (.58)
.78
.76
(.41) (.43)
.10
.12
(.30) (.32)
.01
.02
(.12) (.13)
.44
.41
(.50) (.49)
Pre
0.0
(.02)
0.0
(0.0)
−0.1
(.02)
−0.1
(.03)
.08
(.00)
−.02
(.00)
.02
(.00)
.00
(.00)
−.03
(.00)
Post−Pre
NonPromise
−1.1
(.38)
−1.1
(.35)
−1.1
(.31)
−0.6
(.40)
−0.12
(.04)
.00
(.03)
−.01
(.02)
.00
(.00)
.06
(.03)
Differencea
Notes: The sample consists of ACT Test Takers with from public high schools in the state of Michigan for the years 1996–2006 with no missing values for any of
the following variables: high school demographics, the ACT scores, High School GPA, the set of school selected to receive score reports, race/ethnicity data, and
family income data. The parentheses in column 2, column 3, column 5, and column 6 contain standard deviations. Proportion White, Proportion Black, and
Proportion Latino refer to the proportion of ACT test takers who report being a member of that particular racial/ethnic group; they do not refer to the
demographics of the high school that the test takers attended.The parentheses in column 4, column 7 and column 8 contain standard errors. There are some
discrepancies in the differences due to rounding.
a The columns labeled Difference contains the values for the difference in the Post-Pre differences for ACT test takers in Kalamazoo Promise and Non-Kalamazoo
Promise Schools.
21.1
(5.7)
ACT Math
21.9
(5.5)
ACT Science
22.1
(4.8)
ACT Reading
22.4
(6.3)
GPA
3.35
(.57)
Proportion White
.57
(.49)
Proportion Black
.25
(.43)
Proportion Latino
.02
(.13)
Proportion w\ Family Income < $50k .49
(.50)
ACT English
Pre
Promise
Table 1: Summary of Test Taker Characteristics
29
70.0
(4.7)
18.8
(1.6)
.44
(.02)
.45
(.05)
.07
(.03)
.45
(.06)
Post
Pre
42.1
(25.2)
20.0
(25.6)
.84
(.25)
.10
(.24)
.03
(.05)
.20
(.18)
Post−Pre
−8.0
(3.3)
1.5
(1.4)
−.07
(.03)
.05
(.03)
.01
(.01)
.10
(.05)
42.3
(26.9)
19.2
(3.8)
.82
(.26)
.12
(.25)
.03
(.06)
.25
(.19)
Post
0.2
(1.1)
−.8
(1.1)
−.03
(.01)
.02
(.01)
.00
(.00)
.05
(.01)
Post−Pre
NonPromise
−8.2
(18.9)
2.2
(18.1)
−.04
(.19)
.03
(.18)
.01
(.04)
.04
(.14)
Differencea
Notes: This table summarizes the high school characteristics for the set of ACT test takers from public high schools in Michigan for the years
1996–2006 who have no missing values for any of the following variables: high school demographics, the ACT scores, High School GPA, the set
of school selected to receive score reports, race/ethnicity data, and family income data. The sample of high schools is drawn from the National
Center for Education Statistics’s Common Core of Data. As such, the Data spans the set of Public High Schools in Michigan. There are some
discrepancies due to rounding.
a The columns labeled Difference contains the values for the difference in the Post-Pre differences for ACT test takers in Kalamazoo Promise and
Non-Kalamazoo Promise Schools.
78.0
(4.4)
Pupil/Teacher Ratio
17.3
(1.1)
Proportion White
.51
(.04)
Proportion Black
.40
(.05)
Proportion Hispanic
.06
(.02)
Proportion on Free Meals .35
(.05)
Classroom Teachers
Pre
Promise
Table 2: Summary of High School Characteristics
30
.32
0.139
430702
Pre-Promise Mean
R2
N
.52
0.020
430702
0.122∗∗∗
(0.00869)
.77
0.006
430702
0.0634∗∗∗
(0.00772)
.02
0.054
430702
0.0654∗∗∗
(0.0190)
-0.142∗∗∗
(0.0132)
-0.0149∗∗∗
(0.00318)
WSU
.14
0.014
430702
-0.0224
(0.03524)
0.12667∗∗∗
(0.00561)
0.00199∗
(0.00100)
KC
.16
0.018
430702
0.0238∗∗∗
(0.00535)
0.161∗∗∗
(0.00540)
0.000457
(0.00151)
KVCC
.16
0.083
430702
0.113∗∗∗
(0.00604)
-0.0172
(0.0150)
0.0330∗∗∗
(0.00292)
Flagships
Notes: Standard errors in parentheses: ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 The models are linear probability specifications with standard errors that are
clustered at the high school level. The sample includes all ACT test takers from public high schools in Michigan with valid information on the following
variables: high school demographics, the ACT scores, High School GPA, the set of school selected to receive score reports, race/ethnicity data, and family
income data. The above models include as controls the scores on the four components of the ACT, the test taker’s high school grade point average, a linear
trend, a quadratic trend, the percentage of the high school that is eligible for free lunch, the percentage of the high school that is black, the percentage of the
high school that is white, the percentage of the high school that is Hispanic, the pupil-to-teacher ratio, and the number of full time equivalent teachers. The
dependent variables are binary variables that assume a value of one if the test taker includes the institution in the choice set. The following abbreviations are
used: U of M = University of Michigan-Ann Arbor, MSU = Michigan State University, WMU = Western Michigan University, Pub. U. in MI = Public
Universities in Michigan, WSU = Wayne State University, KC = Kalamazoo College, KVCC = Kalamazoo Valley Community College, Flagships = University
of Michigan-Ann Arbor and Michigan State University. The variable Post assumes a value of one if the test-taker is from the 2005-2006 academic year. The
variable Kalamazoo Promise assumes a value of one if a test taker is from one of the high schools that is eligible for the Kalamazoo Promise Program. The
Pre-Promise Mean is the portion of test takers from Kalamazoo Promise High Schools in the years before the program who submitted a score report to the
institution(s) listed in the column headings.
.36
0.046
430702
0.125∗∗∗
(0.0236)
0.0784∗∗∗
(0.00938)
Post×Kalamazoo Promise
-0.0402
(0.0244)
0.348∗∗∗
(0.0219)
-0.0293
(0.0198)
0.000599
(0.00692)
Kalamazoo Promise
Pub. U. in MI
0.00176
(0.00372)
WMU
0.0496∗∗∗ 0.0204∗∗∗
(0.00423) (0.00389)
0.0424∗∗∗
(0.00358)
MSU
Post
U of M
Table 3: Difference-in-Differences Estimates of The Effects of The Kalamazoo Promise Program on College Choice
31
.32
0.145
430702
Pre-Promise Mean
R2
N
.52
0.026
430702
0.120∗∗∗
(0.0103)
.77
0.008
430702
0.0625∗∗∗
(0.00997)
.02
0.073
430702
0.0640∗∗∗
(0.0187)
-0.137∗∗∗
(0.0125)
-0.0139∗∗∗
(0.00310)
WSU
.14
0.015
430702
-0.02261
(0.0349)
0.1262∗∗∗
(0.00551)
0.00194
(0.00100)
KC
.16
0.019
430702
0.0235∗∗∗
(0.00502)
0.160∗∗∗
(0.00541)
0.000501
(0.00151)
KVCC
.16
0.088
430702
0.115∗∗∗
(0.00689)
-0.0144
(0.0148)
0.0326∗∗∗
(0.00288)
Flagships
Notes: Standard errors in parentheses: ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 The models are linear probability specifications with standard errors that are
clustered at the high school level. The sample includes all ACT test takers from public high schools in Michigan with valid information on the following
variables: high school demographics, the ACT scores, High School GPA, the set of school selected to receive score reports, race/ethnicity data, and family
income data. The models include as controls the scores on the four components of the ACT, the test taker’s high school grade point average, indicators for
family income, race/ethnicity indicators, a linear trend, a quadratic trend, the percentage of the high school that is eligible for free lunch, the percentage of the
high school that is black, the percentage of the high school that is white, the percentage of the high school that is Hispanic, the pupil-to-teacher ratio, and the
number of full time equivalent teachers. The dependent variables are binary variables that assume a value of one if the test taker includes the institution in the
choice set. The following abbreviations are used: U of M = University of Michigan-Ann Arbor, MSU = Michigan State University, WMU = Western Michigan
University, Pub. U. in MI = Public Universities in Michigan, WSU = Wayne State University, KC = Kalamazoo College, KVCC = Kalamazoo Valley
Community College, Flagships = University of Michigan-Ann Arbor and Michigan State University. The variable Post assumes a value of one if the test-taker
is from the 2005-2006 academic year. The variable Kalamazoo Promise assumes a value of one if a test taker is from one of the high schools that is eligible for
the Kalamazoo Promise Program. The Pre-Promise Mean is the portion of test takers from Kalamazoo Promise High Schools in the years before the program
who submitted a score report to the institution(s) listed in the column headings.
.36
0.050
430702
0.127∗∗∗
(0.0223)
0.0808∗∗∗
(0.0101)
Post×Kalamazoo Promise
-0.0437
(0.0231)
0.336∗∗∗
(0.0219)
-0.0299
(0.0200)
0.00699
(0.00609)
Kalamazoo Promise
Pub. U. in MI
0.00209
(0.00370)
WMU
0.0487∗∗∗ 0.0195∗∗∗
(0.00420) (0.00385)
0.0423∗∗∗
(0.00353)
MI St.
Post
U of M
Table 4: Difference-in-Differences Estimates of The Effects of The Kalamazoo Promise Program on College Choice:
Race/Ethnicity Indicators and Family Income Indicators Included
Table 5: Summary of Test Taker Characteristics Matched Sample: Prior
to Kalamazoo Promise
ACT English
ACT Math
ACT Science
ACT Reading
GPA
Proportion White
Proportion Black
Proportion Latino
Proportion w Family Income¡$50k
Promise
Non-Promise
Difference
21.1
(.04)
21.8
(.15)
22.1
(.04)
22.4
(.03)
3.35
(.01)
.57
(.05)
.25
(.04)
.02
(.003)
.49
(.03)
20.8
(.64)
21.6
(.69)
21.7
(.56)
21.8
(.60)
3.20
(.04)
.55
(.06)
.27
(.07)
.01
(.003)
.39
(.05)
.3
(.64)
.2
(.71)
.4
(.56)
.6
(.60)
.15
(.04)
.02
(.08)
-.02
(.08)
.00
(.004)
.10
(.06)
Notes: The sample includes all ACT test takers from both the Kalamazoo Promise High Schools and from the 37
public high schools in Michigan with propensity score estimates closest to the the estimated values for the
Kalamazoo Promise High Schools from the years prior to the implementation of the Kalamazoo Promise who have
valid values for the following variables: high school demographics, the ACT scores, High School GPA, the set of
colleges selected to receive score reports, race/ethnicity data, and family income data. The data span the years
1996–2005.
32
Table 6: Summary of High School Characteristics after
Matched Sample: Prior to Kalamazoo Promise
Classroom Teachers
Pupil/Teacher Ratio
Proportion White
Proportion Black
Proportion Hispanic
Proportion on Free Meals
Promise
Non-Promise
Difference
78.2
(.7)
17.3
(.3)
.51
(.03)
.41
(.04)
.06
(.01)
.35
(.04)
78.7
(4.7)
21.4
(1.6)
.61
(.06)
.30
(.07)
.03
(.01)
.18
(.03)
-0.5
(4.7)
-4.2
(1.6)
.10
(.07)
.11
(.08)
.02
(.01)
.17
(.05)
Notes: The above table summarizes the high school characteristics of all ACT test takers from both
the Kalamazoo Promise High Schools and from the 37 public high schools in Michigan with
propensity score estimates closest to the the estimated values for the Kalamazoo Promise High
Schools who have valid values for the following variables: high school demographics, the ACT
scores, High School GPA, the set of colleges selected to receive score reports, race/ethnicity data, and
family income data. The data span the years 1996–2005
33
34
.32
0.158
39795
Pre-Promise Mean
R2
N
.52
0.179
39795
0.108∗∗∗
(0.0109)
.77
0.018
39795
0.0652∗∗
(0.0185)
.02
0.113
39795
0.0631∗∗∗
(0.0174)
-0.0892∗∗∗
(0.0128)
-0.0157
(0.00867)
Wayne St.
.14
0.108
39795
-0.0209
(0.0345)
0.105∗∗∗
(0.00993)
0.0145
(0.0125)
KC
.16
0.139
39795
0.0190∗∗
(0.00623)
0.179∗∗∗
(0.0107)
-0.0101
(0.00678)
KVCC
.16
0.076
39795
0.125∗∗∗
(0.0129)
-0.0111
(0.0146)
-0.00652
(0.0195)
Flagships
Notes: Standard errors in parentheses: ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 The models are linear probability specifications with standard errors that are
clustered at the high school level. The sample includes all ACT test takers from both the Kalamazoo Promise High Schools and from the 37 public high
schools in Michigan with propensity score estimates closest to the the estimated values for the Kalamazoo Promsie High Schools who have valid values for
the following variables: high school demographics, the ACT scores, High School GPA, the set of colleges selected to receive score reports, race/ethnicity
data, and family income data. The above models include as controls the scores on the four components of the ACT, the test taker’s high school grade point
average, a linear trend, a quadratic trend, the percentage of the high school that is eligible for free lunch, the percentage of the high school that is black, the
percentage of the high school that is white, the percentage of the high school that is Hispanic, the pupil-to-teacher ratio, and the number of full time equivalent
teachers. The dependent variables are binary variables that assume a value of one if the test taker includes the institution in the choice set. The following
abbreviations are used: U of M = University of Michigan-Ann Arbor, MSU = Michigan State University, WMU = Western Michigan University, Pub. U. in
MI = Public Universities in Michigan, WSU = Wayne State University, KC = Kalamazoo College, KVCC = Kalamazoo Valley Community College,
Flagships = University of Michigan-Ann Arbor and Michigan State University. The variable Post assumes a value of one if the test-taker is from the
2005-2006 academic year. The variable Kalamazoo Promise assumes a value of one if a test taker is from one of the high schools that is eligible for the
Kalamazoo Promise Program. The Pre-Promise Mean is the portion of test takers from Kalamazoo Promise High Schools in the years before the program
who submitted a score report to the institution(s) listed in the column headings.
.36
0.030
39795
0.113∗∗∗
(0.0200)
0.114∗∗∗
(0.0137)
Post×Kalamazoo Promise
0.0209
(0.0258)
0.366∗∗∗
(0.0253)
0.000437
(0.0226)
0.00427
(0.00766)
Kalamazoo Promise
-0.0143
(0.0124)
-0.0240∗
(0.00949)
0.0364∗
(0.0149)
0.000687
(0.0207)
Pub. U. in MI
W. MI
MI St.
Post
U. of M.
Table 7: Difference-in-Differences Estimates of The Effects of The Kalamazoo Promise Program on College Choice: Matched
Sample
35
0.0431∗∗∗
(0.00465)
-0.0325
(0.0292)
-0.0434∗∗∗
(0.00225)
0.0853∗∗∗
(0.00765)
0.0141∗∗
(0.00472)
0.00636
(0.0217)
0.0778∗
(0.0383)
0.0364∗∗∗
(0.00395)
0.0194∗
(0.00945)
-0.00942∗∗∗
(0.00166)
0.0624∗∗∗
(0.00721)
0.0147∗∗∗
(0.00406)
-0.0259∗
(0.0118)
0.0344∗∗∗
(0.00414)
.25
0.144
430702
Post
Kalamazoo Promise
FI<$50k
Post×Kalamazoo Promise
Post×FI<$50k
Kalamazoo Promise×FI<$50k
Post×Kal Prom×FI<$50k
Pre-Promise Mean for FI<$50k
R2
N
.53
0.014
430702
0.0293∗∗∗
(0.00614)
0.0643∗∗∗
(0.00427)
0.0165∗∗∗
(0.00462)
0.102∗∗∗
(0.0127)
-0.0410∗∗∗
(0.00223)
0.306∗∗∗
(0.0247)
0.0129∗∗
(0.00427)
WMU
.77
0.006
430702
0.0317∗
(0.0138)
0.0117
(0.00638)
0.00799
(0.00419)
0.0450∗∗
(0.0152)
0.00245
(0.00203)
-0.0490∗
(0.0221)
-0.00141
(0.00400)
Pub. U. in MI
.03
0.072
430702
0.0543
(0.0364)
-0.0141
(0.0104)
0.00985∗∗
(0.00362)
0.0354∗∗∗
(0.00432)
0.0135∗∗∗
(0.00230)
-0.131∗∗∗
(0.0134)
-0.0183∗∗∗
(0.00335)
WSU
.12
0.040
430702
-0.0046
(0.0052)
-0.0387∗∗∗
(0.00415)
0.0016
(0.0010)
-0.0193
(0.0378)
-0.00316∗∗∗
(0.00048)
0.14505∗∗∗
(0.00312)
0.00128
(0.00118)
KC
.21
0.019
430702
-0.104∗∗∗
(0.0131)
0.0827∗∗∗
(0.00192)
-0.00284∗
(0.00130)
0.0753∗∗∗
(0.00161)
0.00159∗
(0.000725)
0.120∗∗∗
(0.00792)
0.00165
(0.00178)
KVCC
.12
0.087
430702
-0.00281
(0.0233)
-0.0191
(0.0108)
0.00443
(0.00311)
0.1160∗∗∗
(0.0169)
-0.0170∗∗∗
(0.00135)
-0.00501
(0.0187)
0.0310∗∗∗
(0.00320)
Flagships
Notes: Standard errors in parentheses: ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 The models are linear probability specifications with standard errors that are
clustered at the high school level. The sample includes all ACT test takers from public high schools in Michigan with valid information on the following
variables: high school demographics, the ACT scores, High School GPA, the set of school selected to receive score reports, race/ethnicity data, and family
income data. The models include as controls the scores on the four components of the ACT, the test taker’s high school grade point average, race/ethnicity
indicators, a linear trend, a quadratic trend, the percentage of the high school that is eligible for free lunch, the percentage of the high school that is black, the
percentage of the high school that is white, the percentage of the high school that is Hispanic, the pupil-to-teacher ratio, and the number of full time equivalent
teachers. The dependent variables are binary variables that assume a value of one if the test taker includes the institution in the choice set. The following
abbreviations are used: U of M = University of Michigan-Ann Arbor, MSU = Michigan State University, WMU = Western Michigan University, Pub. U. in MI =
Public Universities in Michigan, WSU = Wayne State University, KC = Kalamazoo College, KVCC = Kalamazoo Valley Community College, Flagships =
University of Michigan-Ann Arbor and Michigan State University. The variable Post assumes a value of one if the test-taker is from the 2005-2006 academic
year. The variable Kalamazoo Promise assumes a value of one if a test taker is from one of the high schools that is eligible for the Kalamazoo Promise Program.
The variable FI<$50k is an indicator variable that assumes a value of one if family income is reported as being less than $ 50,000 annually. The Pre-Promise
Mean for FI<$50k is the proportion of test takers from Kalamazoo Promise High Schools in the years before the program who both reported that their families
earned less than $ 50,000 annually and submitted a score report to the institution(s) listed in the column headings.
.32
0.048
430702
MSU
U of M
Table 8: Difference in Differences in Differences Estimates of The Effects of The Kalamazoo Promise Program on College
Choice
36
.25
0.158
39795
Pre-Promise Mean for FI<$50k
R2
N
.32
0.029
39795
0.0990∗
(0.0388)
.53
0.177
39795
0.0224
(0.0163)
.77
0.011
39795
0.0155
(0.0288)
-0.0139
(0.0103)
0.0227
(0.0236)
0.0535∗
(0.0245)
.03
0.113
39795
0.0596
(0.0413)
-0.0343∗
(0.0136)
0.00512
(0.0200)
0.0328∗∗∗
(0.00765)
0.0310∗∗
(0.00987)
-0.0701∗∗∗
(0.0137)
-0.0178
(0.0101)
Wayne St.
.12
0.107
39795
0.00469
(0.00686)
-0.0211∗∗∗
(0.00568)
0.00151
(0.00263)
-0.0225
(0.0315)
0.0155∗
(0.00599)
0.115∗∗∗
(0.0129)
0.0138
(0.0129)
KC
.21
0.145
39795
-0.117∗∗∗
(0.0211)
0.0936∗∗∗
(0.00240)
0.00393
(0.00236)
0.0737∗∗∗
(0.00583)
-0.0195∗
(0.00760)
0.132∗∗∗
(0.0109)
-0.00998
(0.00613)
KVCC
.12
0.076
39795
-0.00173
(0.0256)
-0.0365∗
(0.0136)
0.00428
(0.0108)
0.126∗∗∗
(0.0211)
-0.00995
(0.00670)
0.00976
(0.0207)
-0.00808
(0.0197)
Flagships
Notes: Standard errors in parentheses: ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 The models are linear probability specifications with standard errors that are
clustered at the high school level. The sample includes all ACT test takers from both the Kalamazoo Promise High Schools and from the 37 public high
schools in Michigan with propensity score estimates closest to the the estimated values for the Kalamazoo Promise High Schools who have valid values for
the following variables: high school demographics, the ACT scores, High School GPA, the set of colleges selected to receive score reports, race/ethnicity
data, and family income data. The above models include as controls the scores on the four components of the ACT, the test taker’s high school grade point
average, a linear trend, a quadratic trend, the percentage of the high school that is eligible for free lunch, the percentage of the high school that is black, the
percentage of the high school that is white, the percentage of the high school that is Hispanic, the pupil-to-teacher ratio, and the number of full time
equivalent teachers. The dependent variables are binary variables that assume a value of one if the test taker includes the institution in the choice set. The
following abbreviations are used: U of M = University of Michigan-Ann Arbor, MSU = Michigan State University, WMU = Western Michigan University,
Pub. U. in MI = Public Universities in Michigan, WSU = Wayne State University, KC = Kalamazoo College, KVCC = Kalamazoo Valley Community
College, Flagships = University of Michigan-Ann Arbor and Michigan State University. The variable Post assumes a value of one if the test-taker is from
the 2005-2006 academic year. The variable Kalamazoo Promise assumes a value of one if a test taker is from one of the high schools that is eligible for the
Kalamazoo Promise Program. The variable FI<$50k is an indicator variable that assumes a value of one if family income is reported as being less than $
50,000 annually. The Pre-Promise Mean for FI<$50k is the proportion of test takers from Kalamazoo Promise High Schools in the years before the
program who both reported that their families earned less than $ 50,000 annually and submitted a score report to the institution(s) listed in the column
headings.
0.00824
(0.0120)
0.0492∗∗∗
(0.00810)
-0.0529∗
(0.0201)
-0.0279
(0.0178)
Post×FI<$50k
Kalamazoo Promise×FI<$50k
0.0174
(0.0124)
-0.0136
(0.0253)
0.0465∗∗∗
(0.0119)
Post×FI<$50k
0.0909∗∗∗
(0.0144)
0.0638∗∗∗
(0.0167)
0.107∗∗∗
(0.0126)
Post×Kalamazoo Promise
0.0200∗
(0.00777)
-0.0335∗∗∗
(0.00682)
-0.0176
(0.0117)
-0.00517
(0.00697)
FI<$50k
0.0382
(0.0241)
0.347∗∗∗
(0.0256)
0.0316
(0.0294)
0.0220
(0.0131)
Kalamazoo Promise
-0.0224
(0.0168)
-0.0295∗∗
(0.00909)
0.0424∗
(0.0191)
-0.0192
(0.0209)
Pub. U. in MI
W. MI
MI St.
Post
U. of M.
Table 9: Difference-in-Differences-in-Differences Estimates of The Effects of The Kalamazoo Promise Program on College
Choice: Matched Sample
Appendix Table: Enrollment and Score Report Comparison
for Kalamazoo Promise Schools in 2006
College
University of Michigan
Michigan State University
Western Michigan University
Wayne State University
Kalamazoo Valley Community College
Number Enrolled
Number of Score Reports
17
37
102
7
131
87
111
146
24
47
Notes: Enrollee information obtained from Promise at a Glance in 2006. In total, 318 students are using Promise scholarships to go to college and 23 are going to Non-Promise colleges. Some discrepancies
should be expected as the score reports approximate the choice set and the enrollment numbers reflect
the final choice from the choice set. The Kalamazoo Valley Community College discrepancy arise
because of the ACT data not being reflective of the entire set of students who attend community college.
37