NBER WORKING PAPER SERIES THE POWER OF CERTAINTY

NBER WORKING PAPER SERIES

THE POWER OF CERTAINTY:

EXPERIMENTAL EVIDENCE ON THE EFFECTIVE DESIGN OF FREE TUITION PROGRAMS

Elizabeth Burland

Susan Dynarski

Katherine Michelmore

Stephanie Owen

Shwetha Raghuraman

Working Paper 29864

http://www.nber.org/papers/w29864

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

March 2022

This project would not have been possible without our collaborators at the University of

Michigan Office of Enrollment Management, particularly Kedra Ishop, Steve Lonn, Paul

Robinson, and Betsy Brown. We are grateful to the Michigan Department of Education (MDE)

and Michigan’s Center for Educational Performance and Information (CEPI) for providing data.

Seminar participants at the University of Michigan, Columbia Teachers College, and the

University of Wisconsin provided helpful comments. Sarah Cohodes, Joshua Goodman, CJ

Libassi, and Matt Notowidigdo generously read initial drafts. The Institute of Education Sciences

of the US Department of Education (through grants R305B1170015), the Smith Richardson

Foundation, and the Andrew Carnegie Fellows Program funded this research. The views

expressed herein are those of the authors and do not necessarily reflect the views of the National

Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been

peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies

official NBER publications.

may be quoted without explicit permission provided that full credit, including © notice, is given

to the source.

The Power of Certainty: Experimental Evidence on the Effective Design of Free Tuition Programs

Elizabeth Burland, Susan Dynarski, Katherine Michelmore, Stephanie Owen, and Shwetha

Raghuraman

NBER Working Paper No. 29864

March 2022

JEL No. I0,I21,I22,I24,I28

ABSTRACT

Proposed "free college" policies vary widely in design. The simplest approach sets tuition to zero

for everyone. More targeted approaches limit free tuition to those who successfully demonstrate

need through an application process. We experimentally test the effects of these two models on the

schooling decisions of low-income students. An unconditional free tuition offer from a large

public university substantially increases application and enrollment rates. A free tuition offer

contingent on proof of need has a much smaller effect on application and none on enrollment. The

results suggest students place a high value on financial certainty when making schooling decisions.

Elizabeth Burland

University of Michigan

Weill Hall

735 S State St

Ann Arbor, MI 48019

[email protected]

Susan Dynarski

Harvard University Graduate

School of Education 13

Appian Way

Cambridge, MA 02138

and NBER

[email protected]

Katherine Michelmore

735 S State St

Ford School of Public Policy

University of Michigan

Ann Arbor, MI 48109

[email protected]

Stephanie Owen

Colby College

5237 Mayflower Hill

Waterville, ME 04901

[email protected]

Shwetha Raghuraman

University of Michigan

238 Lorch Hall

611 Tappan Ave

Ann Arbor, MI 48109-1220

[email protected]

A data appendix is available at http://www.nber.org/data-appendix/w29864

A long line of research examines policies to increase college enrollment, especially among low-income

students.

Recently, the policy debate has focused on a variety of “free college” proposals. These policies

differ in their eligibility and implementation details, with the simplest setting tuition to zero for all students.

More targeted approaches limit free tuition to those who successfully demonstrate need through an application

process.

A compelling literature shows that seemingly minor differences in program design and implementation

can have outsized effects on take-up rates.

We suspect this also holds true for free college programs. We

implement a large-scale ﬁeld experiment to test our hypothesis.

We randomly assign low-income high school students to receive a pre-admission, unconditional offer

of four years of free tuition at the University of Michigan. In another treatment arm, the free tuition offer

is contingent on students demonstrating need through a traditional aid application each year, for up to four

years. A control arm experiences business as usual.

All students in this experiment have family incomes near the poverty line: they are identiﬁed using a

conﬁdential merger with data on individual eligibility for free- and reduced price meals in Michigan’s public

schools. All students who are eligible for subsidized school meals, in expectation, will receive generous

ﬁnancial aid packages covering most if not all of their tuition and fees at the University of Michigan. Though

both experimental arms advertise free tuition, they differ in the level of certainty and timing of when aid

uncertainty is resolved.

Our analysis suggests that students’ application behavior is swayed by personalized encouragement

and the framing of aid as “free tuition.” Students in both treatment arms apply to the University of Michigan

at much higher rates. However, the ﬁnancial certainty of the unconditional, four-year commitment appears

particularly valuable. Among low-income students offered the unconditional, four-year commitment, the rate

of application to the University of Michigan was 63 percent, compared to 44 percent in the arm with the

conditional, one-year offer, and 35 percent in the control group.

Differences in enrollment behavior underscore the importance of ﬁne print and additional veriﬁcation

in aid offers. While the up-front aid commitment increased University of Michigan enrollment by nearly 9

percentage points, the effect of the conditional commitment on enrollment was less than a percentage point

and not statistically signiﬁcant. This difference in part reﬂects the gap in application effects, but we also see

differences in conditional yield rates. We provide suggestive evidence that the income and asset veriﬁcation

process may have resulted in students receiving slightly less than they expected, which likely dulled the

impact of the conditional offer. We conclude that students place a high value on ﬁnancial certainty when

making schooling decisions.

See Dynarski and Scott-Clayton (2013) and Page and Scott-Clayton (2016) for reviews.

See, for example, Johnson and Goldstein (2009); Beshears et al. (2013); Bulman (2015); Pallais (2015); and Goodman (2016).

I BACKGROUND AND RESEARCH DESIGN

Our setting is an ongoing initiative that has dramatically increased application and enrollment rates

among low-income students at a selective public university. Since 2016, the University of Michigan at Ann

Arbor has offered thousands of low-income students an up-front guarantee of four years of free tuition and

fees (the “HAIL Scholarship”). Students randomized to receive this offer are more than twice as likely as

those in a “business as usual” control group to apply to, be admitted by, and enroll at the University of

Michigan (Dynarski et al., 2021).

Our initial analysis of the introduction of this program did not allow us to tease out its mechanisms.

The treatment was a mixed one: a personalized letter of encouragement from the university president, plus an

unconditional offer of four years of free tuition and fees, plus mailings to parents. In the present paper, we

narrow in on which of these mechanisms generates the very large effects on student decisions.

Our experimental test is straightforward. In fall 2019, we sent mailings to high-achieving, low-income,

high school seniors in Michigan about the University of Michigan. We varied the content of these mailings

across three randomly-assigned experimental arms. Photographs of the mailings are in Appendix A.

I. Control

Students are mailed:

(a) recruitment materials typically sent to prospective applicants

(b) general information about ﬁnancial aid

II. HAIL Scholarship

Students and their parents are mailed:

(a) a personalized encouragement to apply from the university president

(b) an unconditional offer of four years of free tuition and fees

III. Go Blue Encouragement

Students and their parents are mailed:

(a) a personalized encouragement to apply from the university president

(b) a conditional offer of up to four years of free tuition and fees subject to veriﬁcation of eligibility

through ﬁnancial aid applications each year

The ﬁrst two arms replicate the design of Dynarski et al. (2021), in which high school seniors in

fall 2015 and 2016 were randomized between control and the HAIL Scholarship. This captured the causal

effect of the mixed treatment: a personalized letter of encouragement from the university president, plus an

unconditional offer of four years of free tuition and fees, plus mailings to parents. As detailed in Dynarski

et al. (2021), this mixed treatment generated very large increases in rates of application (42 percentage points)

and enrollment (15 percentage points).

Further analysis showed that students induced to attend University of Michigan would either have forgone college, attended a

community college, or enrolled in a less-selective four-year college. There is no evidence that the HAIL intervention diverted students

from other highly selective colleges.

While the treatment effects were conclusively large, it remained unclear which dimensions of the

treatment were most effective. This ambiguity is consequential for policy. Consumers of the ﬁndings to

whom we spoke, particularly policy actors, had widely varying takes on the program’s “secret sauce.” Some

believed the personalized letter from the university president was what mattered, or the mailings to parents.

Their interpretation implied that enhanced outreach would generate large changes in student behavior, without

requiring any substantive change in ﬁnancial aid policies.

Others interpreted the framing of aid as “free tuition’ as what mattered most. This interpretation,

too, implied that small tweaks to recruitment materials would reproduce the results. New mailings could, for

example, highlight “free tuition,” with an asterisk and footnote indicating it was free contingent on passing a

needs analysis. Again, this would require no up-front commitment of free tuition to students.

Our own interpretation was that it was the up-front guarantee of four years of free tuition that most

affected student decisions. This conclusion rested, in part, on the null effects of previous experiments that

provided only information about aid (e.g., Bergman, Denning and Manoli 2019 and Bettinger et al. 2012).

Informational interventions in other policy domains have also produced small to zero effects (e.g., Bhargava

and Manoli 2015).

In order to nail down the mechanisms through which the original intervention affected student decisions,

we worked with the university to design a third arm of the experiment.

The new treatment arm highlights

a University of Michigan program, announced in 2017, that targets low- to middle-income students: the Go

Blue Guarantee. This program provides free tuition and fees to Michigan students with family income below

$65,000 and family assets below $50,000.

Communications for the two treatment arms (II and III) were made as similar as possible, but for the

offer of aid. In particular, Arms II and III both highlight the concept of “free tuition,” while Arm I does not.

The conditions of the “free tuition” offer differ between the two treatment arms.

In order to make the two treatment arms as similar as possible, the packets for HAIL and the Go Blue

Encouragement were the same size and were similarly designed with the University of Michigan branding

and bright coloring (Appendix A). Each packet included a letter signed by the president of the university,

praising the student for their achievements and encouraging them to apply. Information was also mailed

to parents and emailed to principals of eligible students. Letters to parents, mailed two weeks after the

student packets, described the program (HAIL or Go Blue) and encouraged them to help their children apply.

Communications with principals, sent around the same time as the student packets, explained the program,

listed eligible students, and asked the principal to transmit the information to school staff who supported

students in their college applications.

It would have been very satisfying to add enough arms to separately identify the effect of every component of the original, mixed

treatment. In practice, we were constrained by the capacity of the university to manage many treatment arms, as well as by the loss of

statistical power from further splitting the sample.

These two eligibility criteria for the Go Blue Guarantee are far simpler than the multivariable formula that is used to determine

traditional aid such as the Pell Grant. From the students’ perspective, however, the process of qualifying for the Go Blue Guarantee is

the same as it is for traditional student aid. Families complete ﬁnancial aid forms, which the university uses to calculate the income

and assets that deﬁne eligibility for the Go Blue Guarantee. Individual students only learn about their eligibility after admission. For

continuing students this process is repeated, with aid potentially changing over time with a student’s ﬁnancial situation.

The communications in the two treatment arms were identical but for their characterization of ﬁnancial

aid. The letters to students from the university president were identical but for a single paragraph. In the HAIL

arm this paragraph read:

We believe you to be an academically excellent student who has worked hard for your achievements. If

you apply to U-M and are admitted for the fall 2020 term, we will reward your hard work with the HAIL

Scholarship, which covers the full cost of your in-state tuition for four years of study at our Ann Arbor

campus. That’s an approximate $66,000 value to you and your family. Additionally, after a review of your

ﬁnancial aid applications, you will likely be eligible for additional aid to cover costs of housing, meals,

textbooks, and other expenses.

For students in the Go Blue Encouragement arm, this paragraph instead read:

We believe you to be an academically excellent student who has worked hard for your achievements. That’s

why we hope you are planning to apply to the University of Michigan. Furthermore, our Go Blue Guarantee

can help you with your college costs, as it covers the full cost of in-state tuition for in-state students who

are admitted to the Ann Arbor campus and whose families earn incomes of $65,000 or less, with $50,000

or less in assets. If your family earns more, you can still Go Blue; we provide tuition support for families

with incomes up to $180,000.

As noted earlier, all the students in the experiment have low incomes and would typically qualify

for free tuition through standard need-based aid programs administered by the university. It is the timing

and certainty of the aid offers that varies across the experimental arms. The HAIL arm receives a four-year

commitment of free tuition in the September of their senior year of high school. The control and Go Blue

Encouragement arms receive a single-year commitment only if admitted, after completing the aid process,

and have to reapply for aid in subsequent years. This is typically in March or April of the senior year of

high school; however, aid commitments can arrive as early as December (for those admitted through an early

decision program) or as late as the summer after high school (for those whose aid applications are hung up by

bureaucratic delays).

II DATA,SAMPLE, AND RANDOMIZATION

We identify students for the intervention using longitudinal, student-level administrative data from the

Michigan Department of Education (MDE) that contain the universe of students attending public high schools

in Michigan (Michigan Department of Education, 2022).

We identify high-achieving students using high school GPA and SAT score, which come from mandatory,

in-school 11th grade testing. GPA is self-reported on the SAT student questionnaire.

For this intervention,

qualifying SAT scores start at 1100 and qualifying GPAs at a B. Students with higher test scores faced a lower

GPA threshold (and vice versa). The Ofﬁce of Enrollment Management at the University of Michigan set the

GPA and score cutoffs; they are similar to the criteria the school uses when gleaning prospective recruits from

national data on SAT takers.

Students in the sample had an average GPA of an A and SAT of 1270.

We identify low-income students using data on qualiﬁcation for the federal subsidized school-meals

The state of Michigan stopped the statewide collection of high school transcripts several years ago; hence our use of self-reported

GPA. For students with grades and scores in the target range in earlier cohorts, self-reported GPA on the SAT questionnaire was closely

aligned to ofﬁcial GPA on transcripts.

Grades and scores alone do not determine admission; like most highly selective colleges, the University of Michigan uses a holistic

admissions process that also considers factors such as family background and extracurricular activities.

program. Nationwide, students with family income below 130 percent of the federal poverty line qualify for

free lunches, while those with incomes up to 185 percent of the poverty line can get subsidized lunches. In

2020, for a family of four, these thresholds were $34,060 for free lunch and $48,470 for subsidized lunch.

Of the approximately 100,000 juniors in Michigan’s public high schools in the 2018-19 school year,

1,796 students from 477 schools meet both the income and academic criteria for the experimental sample.

Four-ﬁfths of our sample qualiﬁes for a free lunch and the remainder for a reduced-price lunch.

A Randomization

All students in a school who meet the income and academic criteria are assigned the same treatment

status. That is, we randomly assign entire high schools to the treatment arms. We do this because we

hypothesize treatment spillovers within schools, which would attenuate estimated effects toward zero.

We stratify the sample by region (Southeast vs. rest of Michigan) and urbanicity (city vs. suburb,

town, or rural)

and randomize within each of the resulting four strata. We chose these strata based on the

ﬁnding in our earlier experiment that students in rural and more remote areas are most affected by HAIL

(Dynarski et al., 2021). The probability of assignment to each arm is one third. We rerandomized to achieve

balance within region on school characteristics (see Appendix Table 1).

The randomization resulted in 610 control students in 159 control schools, 595 HAIL students in 159

HAIL schools, and 591 Go Blue Encouragement students in 159 Go Blue Encouragement schools. Sample

characteristics are shown in Table 1, by experimental arm. A third of the schools are in the Southeast region

of the state, near Ann Arbor, Lansing, and Detroit. Another 15 percent of schools are in the largely rural

Upper Peninsula. The remaining schools are scattered across the Lower Peninsula, with many in the Grand

Rapids area. Over half the schools are rural, about a third are suburban, and the remainder urban.

Using non-mutually-exclusive race categories, our sample is eighty percent White, seven percent

Black, seven percent Hispanic, eleven percent Asian, three percent American Indian, and less than one

percent Native Hawaiian or Paciﬁc Islander. Eight percent of the sample belongs to more than one of the

race categories.

Balance checks are shown in Appendix Table 1. None of the pairwise comparisons between the

treatment and control groups are statistically signiﬁcant at conventional levels; we also found no statistically

signiﬁcant differences between the combined treatment groups and the control group. This is substantiated by

a joint F-test for each pair of treatment arms, which reveal that, together, these characteristics do not predict

treatment status.

Due to a coding error, we implemented the randomization using out-of-date versions of the urbanicity variables in our dataset.

Using up-to-date values of these variables, our treatment arms are not balanced on the proportion of city schools. All of our main results

are robust to controlling for the up-to-date urbanicity values.

III EMPIRICAL STRATEGY

We evaluate the effect of the HAIL and Go Blue Encouragement offers on application, admission, and

enrollment at the University of Michigan, as described in our pre-analysis plan.

We use internal data on these

outcomes from the university (University of Michigan Ofﬁce of Financial Aid 2022; University of Michigan

Ofﬁce of Enrollment Management 2022), as well as enrollment data from the state of Michigan (Michigan

Department of Education, 2022). We estimate the following by ordinary least squares (OLS):

(1) Y

= b

+ b

HAIL

+ b

GBEncouragement

+ S

+ u

where Y

is an outcome of interest at school j. We collapse the individual student data to the school level

(the level of randomization) and conduct analysis on these means. HAIL

and GBEncouragement

indicate

assignment to the HAIL or Go Blue Encouragement treatment group, respectively. S

is a vector of strata

dummies.

and b

are the parameters of interest and measure the causal effect of being randomized into the

HAIL or Go Blue Encouragement treatment arm, respectively, relative to the control arm, i.e., the estimated

effect of the intent to treat (ITT). These parameters represent the treatment effect on the outcomes of interest,

with schools weighted equally. We additionally test whether the HAIL and Go Blue Encouragement treatment

effects are signiﬁcantly different from each other.

In Section V, we present instrumental-variable (IV) estimates where we use assignment to each

treatment as an instrument for application to the University of Michigan. We estimate the causal effect of

application on admission and enrollment for students induced by either intervention to apply.

In these

analyses, we estimate the following systems of equations using two-stage least squares (2SLS):

(2a) Application

= a

+ a

HAIL

+ S

+ e

HAIL

(2b) Y

= b

+ b

HAIL

[

Application

+ S

+ u

HAIL

(3a) Application

= a

+ a

GBEncouragement

+ S

+ e

(3b) Y

= b

+ b

[

Application

+ S

+ u

where

[

Application

is the school-level application rate predicted by equation 2a or 3a. In this set-up, b

HAIL

and b

represent the local average treatment effects (LATE): the causal effects of application on enrollment

(or admissions), for students induced by the HAIL or Go Blue Encouragement treatment, relative to the

control treatment, to apply.

This study is registered at the randomized trial registry of the American Economic Association under RCT ID AEARCTR-0001831,

with 10.1257/rct.1831-3.0 (Dynarski et al., 2019).

Only the ﬁrst stage and the reduced form analysis were included in the pre-analysis plan.

We run 2SLS separately for the control and HAIL arms (equation 2) and the control and Go Blue

Encouragement arms (equation 3), with equation 2 estimating the LATE for control-HAIL compliers, and

equation 3 estimating the LATE for control-Go Blue Encouragement compliers. Each equation is estimated

for admission and enrollment outcomes. Consistent with our ITT analyses, we report school-level results.

IV RESULTS

Panel A of Table 2 shows the main effects of the HAIL and Go Blue Encouragement offers on

application, admission, and enrollment at the University of Michigan; Appendix Figure 1 shows the effects

visually. We ﬁnd that the HAIL scholarship increased application by 28 percentage points, admission by

10 percentage points, and enrollment by 9 percentage points relative to the control group. The Go Blue

Encouragement treatment increased application by 8 percentage points, admission by 3 percentage points,

and enrollment by 1 percentage point, relative to the control group; for the Go Blue Encouragement treatment,

only the effect on application is statistically signiﬁcant at conventional levels.

The HAIL treatment effect is larger than the Go Blue Encouragement treatment effect on both the

application and enrollment margins (p-values on differences are 0.000 and 0.022, respectively), and marginally

signiﬁcant on the admissions margin (p-value on difference is 0.053). Substantively, the treatment effect on

application was 20 percentage points higher for HAIL than Go Blue Encouragement and the admission and

enrollment effects were each 8 percentage points higher than Go Blue Encouragement.

Overall, both HAIL and Go Blue Encouragement signiﬁcantly increased application rates among

low-income, high-achieving students. However, effect sizes are about three times larger for HAIL and we

ﬁnd no signiﬁcant effects of Go Blue Encouragement on enrollment. Since everyone in the control group was

theoretically exposed to the statewide Go Blue policy, this implies that the personalized mailers did little to

increase enrollment among low-income, high-achieving students beyond the statewide Go Blue policy.

The results so far are based on data from University of Michigan on application, admission, and

enrollment. We use data from the National Student Clearinghouse to examine college enrollments nationwide.

(We do not have data on nationwide application or admissions.) The second panel of Table 2 presents these

results.

For HAIL, the point estimates are broadly similar to those for the two cohorts examined in Dynarski

et al. (2021), though they are substantially less precise due to the smaller sample. The results indicate that

HAIL did not “poach” students from other schools as selective as UM, nor did it increase enrollments at such

schools. Students induced into UM by HAIL would not have attended college at all, or attended less selective

colleges, in the absence of the intervention.

None of the point estimates for Go Blue Encouragement are substantively or statistically signiﬁcant.

Go Blue Encouragement had no impact on enrollment at UM, or anywhere else.

The control mean for UM attendance is higher than for the earlier cohorts, while for the HAIL arm average attendance is about

the same. This mechanically produces a smaller treatment effect (9 vs 15 percentage points) than in the ﬁrst two cohorts. This could be

explained by many factors, including variation over time in program materials (the HAIL packet shrank), changes between cohorts in

the deﬁnition of the experimental sample, secular time effects, the introduction of the Go Blue Guarantee, or growing knowledge of the

HAIL Scholarship among low-income students in the state.

V MECHANISMS

Our ITT estimates suggest that the up-front guarantee of free tuition has a profound impact on student

behavior. In this section, we dig further into the data to try to understand why HAIL had such a large effect

compared to the Go Blue Encouragement.

A The Burden of Aid Forms

HAIL guarantees free tuition. It also waives aid forms. Perhaps students respond so strongly to

HAIL, in part, because they really, really despise aid forms. If HAIL increased applications because it waived

paperwork requirements that marginal enrollees found burdensome, we would expect that UM enrollees in the

HAIL arm would be less likely to ﬁll out the FAFSA than those in the Go Blue Encouragement and control

arms.

We ﬁnd that 98-99 percent of enrolled students complete the FAFSA, with no signiﬁcant differences

across the three arms. Nor is the timing of aid applications consistent with students in the HAIL arm avoiding

the aid form. If anything, HAIL students are quicker to submit their FAFSA applications than control and Go

Blue Encouragement students (see Appendix Figures 2 and 3).

The University of Michigan requires a second ﬁnancial aid form, the CSS Proﬁle, which is administered

by the College Board and asks for additional information about family ﬁnances. The patterns for the Proﬁle

submission diverge from those of FAFSA (Appendix Figure 4). Students tend to submit the Proﬁle later

than they submit the FAFSA: very few students from any of the treatment groups submit the Proﬁle before

December 20, by which time many students had submitted their FAFSA. Conditional on enrollment, HAIL

students are slightly less likely than control or Go Blue Encouragement students to ﬁle the Proﬁle at all (92

percent of enrolled HAIL students complete the Proﬁle compared to 96-97 percent of control and Go Blue

Encouragement students) but this difference is small and is not statistically signiﬁcant.

B Yield Rates and Aid Offers

Both HAIL and Go Blue Encouragement increased applications to the University of Michigan, but

only HAIL increased enrollment. In part, Go Blue Encouragement’s lower effect on enrollment is mechanically

a product of its lower effect on application.

We next investigate whether it also reﬂects a lower admission

or enrollment rate, conditional on application. We cannot use intent to treat to causally estimate these

rates, which condition on (the post-treatment outcomes of) application and admission. We therefore turn

to instrumental variables analysis to examine enrollment and admissions decisions among those induced to

apply by the treatments.

In Table 3 we show IV estimates of the causal effect of application on admission and enrollment

Students who complete the FAFSA but not the Proﬁle tend to have higher family income and assets and are likely not eligible

for additional aid. The small differences between FAFSA and Proﬁle application behavior likely reﬂects the fact that the University of

Michigan is the only public university in the state (and one of a handful nationwide) that requires the CSS Proﬁle. Thus, the CSS Proﬁle

constitutes an additional marginal cost of applying to UM.

If Go Blue Encouragement students enrolled at the same rate as HAIL students, conditional on application, Go Blue Encouragement

would have increased enrollment by 2.5 percentage points instead of 1 percentage point, both below the minimum effect size that we

could have detected.

(estimates are weighted by school; we show student-level results in Appendix Table 2). In this analysis,

random assignment to each treatment arm is used as an instrument for application to the University of

Michigan. We estimate effects separately for HAIL relative to control, and Go Blue Encouragement relative

to control. The ﬁrst stages for these IV equations are the same estimated effects on application as the ITT

from the previous sections. HAIL increased the application rate by 28 percentage points while Go Blue

Encouragement increased it by 8 percentage points. It is important to note that while the F-statistic for the

ﬁrst stage for HAIL is above conventional thresholds at 54, for Go Blue Encouragement it is less than ﬁve.

The Go Blue Encouragement arm therefore does not pass the relevance test. We consider IV results for Go

Blue Encouragement and comparisons across the arms to be suggestive rather than deﬁnitive.

For those offered the HAIL treatment (column 1), the IV estimates show that 34% of those induced

to apply were admitted and 31% enrolled. For Go Blue Encouragement (column 2), 30% of those induced to

apply were admitted and just 10% enrolled. The ratio of enrollment and admission IV estimates is the yield

rate for induced admits: 90% (=0.308/0.342) for HAIL and 33% (=0.100/0.299) for Go Blue Encouragement.

The admission rates for induced applicants from the two arms are quite similar. Applications induced

by the two treatments were presumably viewed as similarly qualiﬁed by university admissions ofﬁcers (we

do not have admissions scores or notes). But enrollment rates and yield rates for the two treatments, while

imprecisely estimated for Go Blue Encouragement, are very different, with the yield rate for HAIL almost

triple that for Go Blue Encouragement. Something caused the behavior of students from these two arms to

diverge after admission.

Differences in aid offers are a plausible explanation. HAIL students could, by construction, only get

pleasant surprises in their aid offers, which could not be less than the promised tuition and fees.

On the

other hand, ﬁnancial aid offers for Go Blue Encouragement students may have fallen short of expectations

created by the headline of “free tuition.” We do not have ﬁnancial aid application information, or offer letters,

for all students who were admitted to the University of Michigan. We can only examine differences in the

ﬁnancial aid packages awarded to enrolled students (Figure 1; see Appendix Table 3 for more detail).

While aid packages for control, HAIL, and Go Blue Encouragement students are broadly similar,

enrolled HAIL students receive around $2,700 more in grants, on average, than Go Blue Encouragement

students. Students in the Go Blue Encouragement arm are less likely than those in the HAIL arm to receive

full-tuition scholarships: 80 percent of students from the Go Blue arm have grants that covered tuition and

fees, compared to 97 percent of the HAIL arm.

It is plausible that the gap is even larger among those who

chose not to enroll.

Our data indicate that many students induced to apply to UM by the free tuition offer did not wind up

qualifying for the Go Blue Guarantee free tuition policy. In particular, many who would have qualiﬁed based

on their incomes did not qualify based on their assets (which, as measured by the Proﬁle, include housing

HAIL students typically received far more grant dollars than originally promised, to pay not only for tuition but for room, board

and other schooling costs.

The four HAIL students who did not get full tuition scholarships have grants that only covered one semester, as they did not enroll

in spring 2021. The same is true for ﬁve Go Blue and control students.

equity in the primary home).

By our calculations from the Survey of Income and Program Participation,

more than a third of families in Michigan with incomes below $65,000 have housing equity of at least

$50,000. The University of Michigan meets students’ full ﬁnancial need, so many students who do not meet

the ofﬁcial criteria for the Go Blue Guarantee still receive substantial ﬁnancial aid packages. Nevertheless,

the disappointment of not qualifying for the free tuition guarantee after learning of its existence may have

contributed to their decision not to enroll.

C Effects of the Statewide Go Blue Guarantee Program

The Go Blue Guarantee was implemented for all Michigan students at once, for winter 2018 enrollment

at the University of Michigan. Our experiment, which took place in 2019, does not let us examine the effect

of the program directly. That is why our experimental arm examining its effect is an encouragement exercise;

potentially, everyone in the control arm could cross over into the Go Blue treatment. In the ITT estimates,

any statewide effect of the Go Blue Guarantee is reﬂected in the behavior of the control group.

Time patterns in application, admission, and enrollment at the University of Michigan for high-achieving

students from the state of Michigan shed some light on whether the statewide rollout of the Go Blue Guarantee

had any effect on student decisions. In Figure 2 we plot these rates separately for low-income and non-low-income

students who have SAT scores of at least 1100.

For low-income students, we clearly see the effects of the initial rollout of the HAIL Scholarship for

the 2016 cohort. We see sharp increases when the experiment started, of 8 percentage points in application, 2.8

percentage points in admission, and 2.7 percentage points in enrollment. HAIL students comprise approximately

a quarter of the low-income population depicted in Figure 2.

The experimental results for these cohorts are

roughly four times the magnitude of the time series jumps, which is consistent with the HAIL treatment-group

students producing all of the increase.

The raw time-series is also consistent with the pattern of results in the present paper. When the Go

Blue Guarantee is implemented for the class of 2018, there is a small increase in application rates but none

in admission or enrollment. These descriptive statistics line up with our experimental results: Go Blue had a

moderate effect on application but none on enrollment, while HAIL had large effects on both application and

enrollment.

We have measures of total income and total assets, as differentially captured by the FAFSA and the Proﬁle, for students in our

experimental sample who enroll at UM. Students can qualify for Go Blue only if assets and income as measured by both the FAFSA and

Proﬁle fall below the stated cutoffs. Almost all students passed the Go Blue income test of $65,000, as expected in a population eligible

for subsidized school meals. Almost all would also have passed the asset test of $50,000 had only the FAFSA been used to calculate

assets. We do not have microdata on detailed asset categories. We do know, however, that through the CSS Proﬁle, UM collects data on

housing equity and retirement savings of aid applicants.

Even among students who meet the criteria for the Go Blue Guarantee, only some will see a “Go Blue Grant” printed in their

ﬁnancial aid package. This contrasts with students in the HAIL arm, who all see a “HAIL Scholarship” printed in their ﬁnancial aid

package covering their tuition and fees. This is because the HAIL Scholarship is applied ﬁrst, before other ﬁnancial aid, whereas the “Go

Blue Grant” is packaged last. For both control and Go Blue Encouragement students, the ﬁnancial aid ofﬁce packages other scholarships,

federal grants, and state aid ﬁrst before adding a “Go Blue Grant” for eligible students (only if they have remaining need). Again, HAIL

students got what they were expecting, and likely more, while Go Blue Encouragement students may have been surprised in what they

were offered—both in amount and in presentation.

Our experimental sample is a subset of the low-income sample because Figure 2 is limited to students with a minimum ACT or

SAT score, while eligibility for HAIL also depends on GPA.

VI DISCUSSION

Our ﬁndings are relevant to the design of tuition and aid policy. The University of Michigan’s Go

Blue Guarantee represents a growing trend among post-secondary institutions (including Harvard, Princeton,

Stanford, and the Universities of Illinois, Virginia, and Wisconsin) in offering free tuition or free cost of

attendance to students from low- and moderate-income families. Additionally, some states have implemented

state-wide free tuition or tuition scholarship programs, including the Susan Thompson Buffett Foundation

scholarship award in Nebraska (Angrist, Autor and Pallais, 2020) and the New York Excelsior Scholarship.

These programs all provide either free tuition or free cost of attendance to students with family income below

a certain threshold. Most recently, in 2020, Democrats campaigned on a promise of free tuition.

At ﬁrst glance, all of these policies appear straightforward. But design details matter, and existing and

proposed free tuition policies vary considerably. Our ﬁndings suggest that the implementation details of these

proposals will matter for their effect on student decisions.

Some free-tuition proposals simply set the “sticker price” of tuition to zero. Like the HAIL Scholarship,

these policies allow students to know that their tuition is free before they make application decisions. By

contrast, other proposals leave sticker prices untouched, but set the net price of tuition to zero for the subset

of students who qualify for sufﬁciently large need-based grants.

Like the Go Blue Guarantee, this delays

the revelation of tuition prices until after students have applied to college and completed aid forms.

Our ﬁndings suggest that a straightforward, zero-tuition program like HAIL would substantially expand

enrollments among low-income students. However, we expect little effect of policies that, like the Go Blue

Guarantee, rely on traditional, need-based aid programs and do not resolve uncertainty about aid until after

application.

A downside of a broad-based commitment of free tuition is that it is expensive, since the subsidy goes

to all students regardless of income. But there are several ways to target free tuition to low-income students

without relying on the traditional system of needs analysis. For example, at community colleges (which

largely enroll students of modest means) a zero-tuition approach would convert what is essentially a policy of

free net tuition into a policy of free sticker-price tuition, providing potential students greater certainty while

requiring little change in per-student spending (of course, total spending would rise if the policy attracted

more students to college).

A free-tuition policy at four-year colleges would, by contrast, require substantial government funding,

since these schools typically rely on the tuition revenue of full-paying, upper-income students. These colleges

could create targeted zero-tuition guarantees like the HAIL Scholarship. Data-sharing partnerships with

state educational agencies would allow individual schools, consortia of schools, or entire state systems to

proactively identify low-income students to offer an early free tuition guarantee. Our ﬁndings suggest these

policies would substantially expand the attendance of low-income students at four-year colleges, where they

are currently under-represented.

The institution- and state-level policies above are all examples of free net-price policies. For a recent legislative proposal, see, for

example https://www.nytimes.com/2021/10/08/us/politics/manchin-democrats-means-testing.html

REFERENCES

Angrist, Joshua, David Autor, and Amanda Pallais. 2020. “Marginal Effects of Merit Aid for Low-Income

Students.” NBER Working Paper No. 27834, National Bureau of Economic Research, Cambridge, MA.

Bergman, Peter, Jeffrey T. Denning, and Dayanand Manoli. 2019. “Is information enough? The

effect of information about education tax beneﬁts on student outcomes.” Journal of Policy Analysis and

Management, 0(0): 1–26.

Beshears, J, James Choi, David Laibson, and B Madrian. 2013. “Simpliﬁcation and saving.” Journal of

Economic Behavior and Organization, 95(1): 130–145.

Bettinger, Eric P, Bridget Terry Long, Philip Oreopoulos, and Lisa Sanbonmatsu. 2012. “The role

of application assistance and information in college decisions: Results from the H&R Block FAFSA

experiment.” The Quarterly Journal of Economics, 127(3): 1205–1242.

Bhargava, Saurabh, and Dayanand Manoli. 2015. “Psychological frictions and the incomplete take-up of

social beneﬁts: Evidence from an IRS ﬁeld experiment.” American Economic Review, 105(11): 3489–3529.

Bulman, George. 2015. “The effect of access to college assessments on enrollment and attainment.”

American Economic Journal: Applied Economics, 7(4): 1–36.

Dynarski, Susan, and Judith Scott-Clayton. 2013. “Financial aid policy: Lessons from research.” NBER

Working Paper No. 18710, National Bureau of Economic Research, Cambridge, MA.

Dynarski, Susan, CJ Libassi, Katherine Michelmore, and Stephanie Owen. 2019. “Increasing Economic

Diversity at a Flagship University: Results from a Large-Scale, Randomized Trial.” AEA RCT registry,

DOI 10.1257/rct.1831.

Dynarski, Susan, C.J. Libassi, Katherine Michelmore, and Stephanie Owen. 2021. “Closing the gap: The

effect of reducing complexity and uncertainty in college pricing on the choices of low-income students.”

American Economic Review, 111(6): 1721–1756.

Goodman, Sarena. 2016. “Learning from the test: Raising selective college enrollment by providing

information.” Review of Economics of Statistics, 98(4): 671–684.

Johnson, Eric, and Daniel Goldstein. 2009. “Do defaults save lives?” Science, 302(1): 1338–1339.

Michigan Department of Education. 2022. “Student Administrative Data (2011-2020).” Michigan

Education Data Center.

Page, Lindsay C, and Judith Scott-Clayton. 2016. “Improving college access in the United States: Barriers

and policy responses.” Economics of Education Review, 51: 4–22.

Pallais, Amanda. 2015. “Small differences that matter: mistakes in applying to college.” Journal of Labor

Economics, 33(2): 493–520.

University of Michigan Ofﬁce of Enrollment Management. 2022. “University of Michigan Application

File (2016-2020).” Michigan Education Data Center.

University of Michigan Ofﬁce of Financial Aid. 2022. “University of Michigan Financial Aid Data

(2016-2020).” Michigan Education Data Center.

Table 1

School-Level Summary Statistics by Treatment Arm

Characteristic Control mean HAIL mean GB Encouragement mean

Southeast school 0.352 0.346 0.352

(0.479) (0.477) (0.479)

School in UP 0.151 0.176 0.151

(0.359) (0.382) (0.359)

City school 0.126 0.126 0.126

(0.333) (0.333) (0.333)

Town/rural school 0.528 0.528 0.516

(0.501) (0.501) (0.501)

Suburban school 0.346 0.346 0.358

(0.477) (0.477) (0.481)

Distance of school from UM (miles) 98.9 104.1 97.5

(86.737) (86.649) (75.648)

UM application rate of school, class of 2015 0.065 0.066 0.060

(0.077) (0.096) (0.090)

Average ACT score of school, class of 2015 19.959 19.915 19.886

(1.851) (2.063) (2.071)

Proportion of HAIL students with A or A+ GPA 0.858 0.866 0.839

(0.240) (0.223) (0.263)

Proportion of HAIL students with A-, B+, or B GPA 0.142 0.131 0.160

(0.240) (0.223) (0.262)

Average SAT of HAIL students 1260.270 1264.402 1261.517

(71.144) (72.771) (61.826)

Proportion female 0.564 0.549 0.570

(0.346) (0.357) (0.336)

Proportion under-represented minority 0.170 0.154 0.185

(0.283) (0.269) (0.294)

Average number of HAIL students 3.8 3.7 3.7

(3.500) (3.189) (3.506)

Number of schools 159 159 159

Number of students 610 595 591

Source: Michigan Department of Education (2022), University of Michigan Ofﬁce of Enrollment Management (2022).

Notes: All analyses conducted at the school level. Standard deviations in parentheses. Due to a coding error, we implemented

randomization based on out-of-date values of the city, town/rural, and suburban variables, which we use for this table as well.

Table 2

Estimated Effect of HAIL Scholarship and Go Blue Encouragement Treatments

on College Choice Outcomes

Treatment effect

Go Blue HAIL vs. GBE

Outcome HAIL Encouragement Effects

Panel A. University of Michigan Application, Admission, and Enrollment (UM administrative data)

Applied to University of Michigan 0.280 0.082 0.198

(0.038) (0.039) (0.038)

[0.354]

Admitted to University of Michigan 0.096 0.025 0.071

(0.036) (0.035) (0.037)

[0.230]

Enrolled at University of Michigan (UM data) 0.086 0.008 0.077

(0.033) (0.032) (0.034)

[0.174]

Panel B. Enrollment Outcomes (National Student Clearinghouse data)

University of Michigan (NSC data) 0.089 0.010 0.080

(0.033) (0.032) (0.034)

[0.169]

Highly competitive or above (other than UM) 0.010 -0.002 0.012

(0.016) (0.015) (0.017)

[0.039]

Four-year 0.039 -0.009 0.048

(0.035) (0.036) (0.036)

[0.724]

Two-year 0.002 0.012 -0.010

(0.021) (0.021) (0.022)

[0.071]

Any 0.041 0.002 0.038

(0.031) (0.033) (0.032)

[0.796]

Number of school-years 477

Number of students 1,796

Source: Michigan Department of Education (2022), University of Michigan Ofﬁce of Enrollment Management (2022).

Notes: All analyses done at the school level. Robust standard errors reported in parentheses. Results from a regression of the outcome

on indicators for each treatment status (HAIL, Go Blue Encouragement), and strata indicators. Control means are in square brackets.

The difference, and standard error of the difference, between the HAIL and Go Blue Encouragement effect coefﬁcients reported in

the right-most column. UM application, admission and enrollment measured in the summer and fall following expected high school

graduation. Admission and enrollment are unconditional on application.

Table 3

School-level Instrumental Variables Estimates of the Effect of UM Application

on UM Admission and Enrollment

Comparison

Outcome HAIL - Control GB Encouragement - Control

Admitted 0.342 0.299

(0.103) (0.330)

Enrolled 0.308 0.100

(0.100) (0.356)

Applied (First Stage) 0.280 0.082

(0.038) (0.039)

First-stage F statistic 53.923 4.508

Strata dummies X X

Number of schools 318 318

Source: Michigan Department of Education (2022), University of Michigan Ofﬁce of Enrollment Management (2022).

Notes: All analyses done at the school level. Robust standard errors reported in parentheses. Results from two-stage least squares

regression, where the ﬁrst stage is a regression of application on treatment and strata dummies. The left column regression sample

includes control and HAIL schools, and the right column regression sample includes control and GB Encouragement schools.

Figure 1

Distribution of Total Grants and Scholarships Awarded to Students, by Treatment Arm

Source: Michigan Department of Education (2022), University of Michigan Ofﬁce of Enrollment Management (2022), University of

Michigan Ofﬁce of Financial Aid (2022).

Notes: Figure plots the distribution of total grant aid by treatment group, among students with aid data reported. Grant aid includes all

institutional and departmental scholarships and grants, federal grants, state grants and scholarships, and private scholarships. The gray

bar represents the range of tuition and fees charged to each student. Tuition and fees range from $15,734.19 to $16,836.19 depending

on the school or college each student is enrolled in at UM. Only the difference between the GBE and HAIL distributions is marginally

signiﬁcant, with an exact p-value from a two-sample Kolmogorov-Smirnov test of 0.051. The four HAIL students who did not get full

tuition scholarships have grants that only covered one semester, as they did not enroll in spring 2021. The same is true for ﬁve Go Blue

and control students.

Figure 2

University of Michigan Application, Admission, and Enrollment Rates

for High-achieving Michigan Public High School Students

(a) Application to University of Michigan (UM) (b) Admission to UM

Source: Michigan Department of Education (2022), University of Michigan Ofﬁce of Enrollment Management (2022).

Notes: Figure plots the number of students who applied (or were admitted/enrolled) to UM over the number of students in each 11th

grade cohort in Michigan public schools. High-achieving students are students who scored at least a 23 on the ACT before 2016, or a

1100 or the SAT in 2016 or later, to correspond with the HAIL academic criteria. The University of Michigan considers the full rollout

of the Go Blue Guarantee to be 2018. There was a partial rollout in 2017, which we label here as the “soft Go Blue” rollout.