Upload
vanminh
View
220
Download
0
Embed Size (px)
Citation preview
EFFORT AND INFORMATION:
EVIDENCE FROM MBA GRADE NON-DISCLOSURE*
Eric Floyd†
Rady School of Management
University of California at San Diego
Sorabh Tomar
University of Chicago, Booth School of Business
Daniel J. Lee
Rice University, Jesse H. Jones Graduate School of Business
This Draft: February, 2017
Abstract:
In this paper, we study the effect of grade non-disclosure (GND) on student and employer behavior
in the context of an MBA program at a highly-ranked business school. GND precludes students
from revealing their GPAs to prospective employers. Our results suggest students under GND
spend 55.4 hours (5.31%) less time preparing for classes during their MBA program while
weighting their enrollment towards more difficult subjects. We also document diminished GPA-
based matching of students with employers following GND. Overall, our study provides empirical
evidence on the relationship between information and effort that underpins a broad class of
principal-agent models.
JEL Codes: D70, D82, J24, M51
* The authors are grateful to Phil Berger, Matt Bernhardt, Kevin Boyd, Eric Budish, Hans Christensen, Amoray Cragun, Alan Crane, Kevin
Crotty, Eugene Fama, Christine Gramhofer, Yael Hochberg, Emir Kamenica, Christy Leak, Christian Leuz, John A. List, Michael Minnis, Kevin
Murphy, Patricia Naranjo,Valeri Nikolaev, Ray Pauliks, Canice Prendergast, Jonathan Rogers, Shiva Sivaramakrishnan Doug Skinner, Paul Steven, Kevin Troy, Sarah Zechman and seminar participants at the University of California Irvine and the University of Chicago Booth School
of Business for helpful comments and suggestions regarding this research. The authors acknowledge financial support from Rice University Jones
Graduate School of Business. † Corresponding Author: [email protected]
1
1) Introduction
In 1994, MBA students at the Wharton School of Business of the University of
Pennsylvania implemented the first policy allowing students to voluntarily withhold grades
from recruiting employers. Since then, grade non-disclosure (GND) policies have become
common among leading business schools in the United States. In this paper, we explore the
implications of GND policies in MBA programs and labor markets. By limiting the impact
of academic transcripts in the recruiting process, GND can promote collaboration among
students by reducing grade-based competition in the classroom. Furthermore, GND can
encourage involvement in extracurricular activities and encourage students to enroll in
more demanding but potentially more valuable and specialized classes. If GND is effective
in changing student behavior, business schools can use such policies to influence the
skillsets of their graduates (for example, graduates may develop stronger soft skills, such as
negotiating ability). Changes to MBA human capital are likely meaningful because MBA
graduates are a significant proportion of top-level decision-makers within organizations.1 2
However, GND can also reduce student effort or alter labor market matching. Both
effects can plausibly pair employers with students that are less desirable than those under
grade disclosure. Guvenen, Kuruscu, Tanaka, and Wiczer (2015) analyze the importance of
skill mismatch between the requirements of an occupation and the skill set of an employee.
They show significant negative impacts of skill mismatch on both wages and the
probability of switching occupations. GND may exacerbate skill mismatch by making it
more difficult for employers to select a student that best fits their requirements. For these
reasons, GND’s desirability has been the subject of considerable debate between students,
1 Roughly 17,970 of 147,000 high ranking company employees (e.g., CEOs, CFOs, COOs, directors, etc.) in the BoardEx database for 2012
held MBAs from the twenty-five top business schools ordered by mean student GMAT score in 2011 (see Table 1). 2 A movement towards soft and social skills would be consistent with the evidence in Deming (2015), which shows the growing importance of social skills in the labor market.
2
faculty, and school policy makers.3 Providing causal evidence on GND’s effects is
imperative to moving this debate forward.
Our approach, which uses a standard difference-in-differences framework, allows us to
plausibly identify GND’s causal effects by comparing two student groups within the same
school that differ with respect to whether or not GND applies to them. The school we
investigate has both full-time, daytime MBA programs and part-time, evening MBA
programs, but only full-time MBA students are subject to GND. Evening MBA students
provide a useful comparison to those in the full-time program because students in both
programs typically take the same classes with the same professors, but in different course
sections. Evening and full-time MBA students work through the same syllabi and are
assessed on the same rubrics and criteria. In addition, professors at the school use a
common grading curve for their full-time and evening student class sections. These
similarities hold both before and after GND’s adoption.
We utilize two novel datasets that come directly from the school’s administrators.
These datasets provide information on classroom level averages from course evaluation
surveys for all students (including self-reported student effort), as well as individual GPAs
and post-graduation employment outcomes for full-time MBA students (evening students
typically have full-time employment during the program and rarely change employers
afterwards). We estimate that students reduce subject effort by 55.4 hours during their
MBA program following the adoption of GND. An estimate of a 5.31% effort reduction
over a two-year program results from utilizing a control sample to estimate the
counterfactual, and is smaller than estimates in prior literature (e.g. Jain (1997) reports a
32% reduction in student effort). We further show a greater effort reduction in core subjects
3 See Gottlieb and Smetters (2011), Financial Times (2013), and The Economist (2005) for some examples.
3
(which all students are required to take) than in non-core subjects. This finding aligns with
core subject grades carrying more signal value and/or provide timelier information than
non-core subject grades.
We then explore several alternative explanations for our findings. Two primary
considerations are whether the effect we observe is attributable to concurrent changes in
course offerings that differentially affect full time student effort, and whether effort-averse
students select into schools with GND. First, we find that the observed treatment effect
occurs sharply around the policy implementation date. This sharp change in behavior does
not correspond with observed patterns in new course offerings, indicating a concurrent,
asymmetric change in curriculum across the two student groups does not explain our
findings. We also find that the sequence of quarter-by-quarter effort-declines across first-
and second-year subjects supports neither a selection-based nor a macroeconomic (e.g.
financial recession) explanation for effort reduction.
We next examine student course selection. We find evidence that suggests students
enroll in more challenging classes following GND. However, weighting enrollments
towards more difficult classes is not the only avenue through which GND can affect
students’ time allocation. The possibility remains that students respond to GND by trading
off class preparation effort for leisure.
If students respond to GND by altering their allocation of time across human-capital
augmenting activities, then changes in labor market matching could result. Using MBA
career preference data from Universum, an employer-branding firm, we find that the
correlation between GPA and job quality declines after the adoption of GND. This suggests
that the information in GPA is not recoverable from other sources of information available
to employers. One implication of this result is that employers could begin placing more
4
weight on other measures of student quality, such as interview performance.4 Importantly,
data on hiring decisions is only available for those subject to GND so this part of our study
does not benchmark its findings against a control group.
Our study contributes to existing literature in several ways. First, we provide direct
evidence on the effects of implementing a grade non-disclosure policy. The results of our
study are informative about the effects of other policies that change the informativeness of
signals that result from grades, such as pass/fail systems. Thus, our study informs
academics and school administrators interested in understanding the effects of policies that
vary the information derived from grades.
Second, we utilize a natural experiment that isolates the disclosure of grades. This
helps us to understand the impact of educational institutions' measurement of student
ability, in addition to their role in augmenting student ability, on student investment
decisions and job market outcomes. The results of our paper inform the debate on the value
of grades in the labor market (e.g. Weiss, 1995). Our findings reinforce the role that grades
play in how students match with employers.
The contribution of our paper is not intended to be limited to MBA labor markets; we
anticipate our paper will contribute to the extensive literature on principal-agent models,
most of which is theoretical. As Prendergast (1999) points out, empirical confirmations of
agency theory are relatively sparse, despite a large literature that shows incentives can have
large effects for motivating agent’s effort. Our setting is consistent with agency theory
because we not only show that agent’s behavior changes in response to changes in
incentives that result from increased uncertainty, but also that the implicit contract between
employers and MBA students in hiring decisions changes in response to increases in
4 A number of firm MBA recruiters that we contacted suggested this possibility.
5
uncertainty. In addition, as Prendergast (1999) points out, non-contractible effort is
unlikely to be observed by researchers, and therefore existing research typically focuses on
output measures such as number of goods produced. Our ability to observe effort is a novel
feature of our data.5
Finally, we contribute to the subset of empirical economics literature that analyzes the
impact of transparency regulation. The recent increase in the use of transparency to solve
market failures has increased the need to better understand the effects of mandatory
information disclosure on the functioning of economic markets (Fung, Graham, and Weil,
2007). Empiricists now explore the implications of mandatory disclosure, notably real
effects (Kanodia and Sapra 2016), in markets such as accounting and finance, healthcare,
and restaurant hygiene (Dranove et al., 2003; Jin and Leslie, 2003; Christensen et al., 2015,
Christensen et al., 2016). Considerable uncertainty remains regarding the impact of
disclosure policies in a diverse range of settings, and whether the results of existing
research generalize given the complexities and intricacies of different markets. By
analyzing a disclosure policy in a relatively unexplored setting - labor markets - we also
add to this growing body of empirical research.
2) Setting and Theoretical Motivation
2.1 Grade Non-Disclosure
Our study takes place at a top-ranked, U.S. business school. Before beginning their
program in 2001, the incoming full-time MBA class voted to implement GND. The policy
asks full-time students to not disclose their grades and GPAs to prospective employers, for
both on- and off-campus recruiting, and for both internship and graduate employment
recruiting. Further, overturning the GND policy at this school is a two-step process. First,
5 Foster and Rosenzweig (1994) is a notable exception.
6
25% of the incoming class must sign a petition to force a vote. Then, if at least 75% of the
incoming full-time MBA students participate in the vote and two thirds of the votes are in
favor of overturning the status-quo policy, the policy is overturned.
Under the policy, students can release their grades and GPAs upon accepting an offer or
graduating. The policy's reach extends to the Dean's list and other academic honors for
internship recruiting, but not for graduate recruiting. The policy is not enforced by the
school but deviations are subject to reputation penalties for both students and recruiters
(discussed further below). The policy does not extend to evening, executive, or weekend
MBA students. As we later argue, this feature of the GND policy allows us to explore its
causal impact.
Ciolli (2006) provides descriptive evidence on the administration and enforcement of
GND policies at business schools across the United States. Table 1 (reproduced from
Gottlieb and Smetters, 2011) lists the top 25 business schools, ordered by average GMAT
score, and whether the school has an active GND policy. He suggests that the main penalty
for non-compliance with GND is reputation-based, much like that from breaking a social
norm. Many websites, forums, and blogs publish lists of reasons for why one should
embark on an MBA - networking and alumni relations feature prominently in these lists-
which suggests a reputation penalty is non-trivial. Deviating from GND may also indicate
to an employer that a candidate is dishonest or disloyal, potentially more so if the recruiter
is an alumnus of a business school that practiced GND when the recruiter was a student.
Furthermore, it is feasible that there are reputation consequences for recruiters as well. If it
were known that a recruiter solicited grades from students, the student representative board
could reduce their promotion of that recruiter to the student body.
7
We have discussed the issue of GND violations with staff who work closely with MBA
students and have detailed knowledge of the recruiting process as well as with faculty.
While some believe that deviations are commonplace in the upper tail of the grade
distribution, the consensus is that the majority of students comply with GND. That
complete unraveling does not occur is consistent with a reputation cost to disclosing grades
(Verrecchia, 1983). That is, the benefit of disclosing (separating from non-disclosers)
exceeds the expected reputation penalties for only a select subset of students with grades
above some threshold. To the extent the policy does unravel, it increases the likelihood of
observing no consequences stemming from GND.
The benefits and costs of GND policies are contentious issues themselves, as evidenced
by different schools taking different stances towards GND (see Table 1), with schools
sometimes alternating their stance towards GND. For example, in 2005, Harvard Business
School switched from actively enforcing GND to making adherence voluntary. Proponents
of GND claim that the policy allows for more academic freedom and collaboration, and
allows students to take on more difficult and specialized subjects without the fear of grade-
based competition. MBA internet discussion forums and blogs (e.g., GMAT Club, Poets &
Quants) echo these sentiments. For example, one participant notes, “Disclosing grades may
encourage competitive attitude at the cost of collaboration. Employers value collaboration
and hence schools want to promote it.” On the other hand, detractors of GND suggest that
it leads to reduced student effort. Another participant notes, “Given that there is not much
at stake, our effort and desire to learn naturally goes down.”
2.2 Theory
Principal-agent models have been used extensively to explain observed economic
phenomena ranging from employee compensation to education. Examples of these models
8
include signaling (e.g. Spence, 1973) and multitask principal-agent models (e.g.
Holmström and Milgrom, 1991). A common feature of principal-agent models is that the
precision of information available to the principal influences the magnitude of actions
taken by economic agents as well as the contract offered by the principal to the agent.
We model GND by reproducing a multitasking model that captures the basic elements
of our setting (Holmstrom and Milgrom 1991). We consider a vector of effort that contains
two elements, e=(eh,es), which represents effort made by students in hard and soft tasks (i.e.
subject effort produces hard skills and extracurricular activities produce soft skills). Effort
produces a benefit represented by the concave function B(eh,es) and the agent’s cost is
represented by the convex function C(eh,es). The agent is compensated based on signals,
x=f(e)+ε, available to the employer. ε is assumed to be normally distributed with
covariance matrix Σ.
Holmstrom and Milgrom (1991) show that under reasonable assumptions the optimal
performance incentive is linear of the form w(x)= αhxh+αsxs+β. Under the assumption that
agents choose an allocation of effort across hard and soft tasks to maximize their certainty
equivalent, and σhs=0, then αh satisfies the following condition:
𝛼ℎ = (𝐵ℎ −𝐵𝑠𝐶ℎ𝑠
𝐶𝑠𝑠)/[(1 + 𝑟𝜎ℎ
2 (𝐶ℎℎ −𝐶ℎ𝑠
2
𝐶𝑠𝑠)]
This simple condition is sufficient to motivate the effects of GND. Grades provide one
useful signal of hard ability. When grades are no longer available, the information
concerning eh becomes less informative, or σ2h increases. As σ2
h increases, the weight
placed on signals of hard tasks by employers decreases. The weight placed on soft tasks
after GND is less clear and depends crucially on the cost function. If hard and soft tasks are
complements (Chs is negative) then increasing the weight on soft skills after GND can
9
incentivize student’s to exert effort in hard skills even when performance on hard tasks is
difficult to measure. Alternatively, if hard and soft tasks are substitutes, then increasing the
weight on soft skills will reduce the agent’s incentive to invest in hard tasks. When GND
limits the ability of employers to reward effort in hard tasks, employers can best incentivize
student’s investment in hard tasks by reducing the weight placed on soft skills.
Ultimately, whether or not agents reduce effort in hard tasks following GND depends
on the extent in which 1) employers value hard tasks and 2) hard and soft tasks are
complementary in the agent’s cost function in such a way that rewarding soft tasks can
preserve effort in hard tasks. Alternatively, employers may be able to minimize the effort
loss in hard tasks by reducing the extent to which they reward soft tasks if the tasks are
substitutes. Finally, as the measurability of hard tasks decreases, students may begin
placing weight on the relatively better measured task, soft skills, if employers increase the
weight on soft skills in the wage function.
3) Empirical Predictions
Our primary empirical prediction is that GND reduces student effort in classes. All
MBA students at the business school we study are required to complete the following core
subjects, typically in their first year: Business Statistics (or Applied Regression Analysis),
Financial Accounting, and Microeconomics. Thus, these subjects' grades allow employers
to rank students across the full spectrum of student types. If grades have informational
value, grades in core subjects are likely the most informative. Therefore, one potential
consequence is that the effort drop following GND is larger for core subjects, relative to
non-core subjects.6
6 A greater effort reduction for core subjects has an alternative explanation. First year internship recruiting at the school we study typically
begins in January, while second year full-time recruiting typically begins the following October. In both cases, second year grades are not
available for evaluation to recruiters. Thus, a larger decrease in effort for core subjects following GND would also be consistent with first year subjects (both core and non-core) being the only ones that produce timely information. Unfortunately, we cannot identify non-core subjects
10
An effort decrease should occur during the first-year courses for the 2001-incoming
cohort (as compared to 2000-incoming). This is because the 2001-incoming cohort voted
in favor of GND before beginning their MBA program. When they did, GND applied not
only to their cohort, but to all subsequent cohorts as well. Accordingly, we expect to see
no drop in first year effort when comparing the second non-disclosing cohort to the first
non-disclosing cohort (2002-incoming vs. 2001-incoming respectively). We expect to see
no change in effort for second year subjects across any of the years since second year
subject grades lack generalizability to all students, and lack timeliness - second year
subjects are more specialized, and the bulk of recruiting occurs in October, before second
year grades are acquired. Table 2 summarizes these predictions across the two years.
A purported benefit of GND is that it allows students to move out of their academic
comfort zones. Under a grade disclosure regime, students may be compelled to take
subjects that they feel are easier or less adventurous, fearing competition based on grades.
Thus, it is possible that students enroll in more difficult or specialized subjects following
GND. Furthermore, students may use the academic freedom afforded to them under GND
in order to pursue activities that are not directly related to coursework, such as club
activities and networking events.
4) Data
We use an administrative dataset from the business school that reports the mean values
from course evaluation surveys completed by students at the end of each academic quarter.
The data cover classes from autumn 1997 to autumn 2013. Each observation corresponds to
an instance of a class taught by a specific instructor in a given quarter-year (instructor-
that predominantly first years take. Doing so would allow us to separate the effects of signal generalizability (i.e., how wide the support of
student types is) and signal timeliness.
11
subject-quarter-year). Henceforth, we refer to each observation as a “classroom”. We
observe course evaluation data for 5,867 classrooms, and do not observe evaluation data for
80 classrooms. Thus, our sample is relatively complete. The main variables we construct
from the database are as follows: HOURS is the classroom average response to the question
“Excluding class sessions, estimate the average number of hours per week spent in
preparation or review”; FULL is a dummy variable that takes the value 1 for full-time
MBA observations, and 0 otherwise.
FULL information comes from subject codes indicating which classrooms were held in
the downtown campus during weekdays. These classrooms are designated for evening
MBA students; though full-time MBA students may also enroll in them. Any enrollment in
evening classrooms by full-time MBA students (and vice versa) will bias our statistical
tests towards the null since it acts to homogenize the treatment and control groups.
5) Student Response to GND
5.1 Counterfactual
To make causal inferences about GND’s effects on student behavior using a difference-
in-differences design, a valid control group is needed. GND, at the school we study, applies
only to full-time MBA students. We can therefore consider evening MBA, weekend MBA,
executive MBA, PhD, and undergraduate students as potential samples to estimate the
counterfactual. To obtain unbiased estimators of GND’s effects, our estimation strategy
must ensure that by classifying our observations as either treatment or control we do not
sort on unobservable characteristics correlated with the dependent variables of interest.
Executive MBA students, PhD students and undergraduate students are inherently
different from full-time MBA students due to differing job market incentives, demographic
12
profiles, and/or course materials. Therefore, we do not consider them suitable control
groups because the likelihood of correlated omitted variables is high.
Evening MBA students and weekend MBA students are more likely to have
characteristics similar to those of full-time MBA students. In Table 3 we present
descriptive statistics on the school’s students. In 2011, the average ages of full-time MBA
students, evening MBA students, and weekend MBA students were 28, 29, and 30
respectively. The average lengths of prior full-time employment were 5 years, 5 years, and
6 years respectively. In addition, evening and weekend MBA students take many of the
same courses as full-time MBA students. These courses are taught by the same instructors,
use the same teaching assistants, follow the same syllabi, provide the same learning
materials, and involve the same assessment criteria. Instructors also use a common grading
curve across their evening and full-time sections.
Evening MBA students and weekend MBA students typically begin their studies with
secured employment. Therefore, at first glance, their job market incentives may seem
markedly different, i.e., evening and weekend students may appear to have fewer
incentives to focus on grades. However, anecdotal evidence suggests these students
typically aim for promotions, partake in recruiting events (including recruiting events
specifically for part-time students), and often have their studies funded by their employers,
conditional on a minimum academic performance level. Thus, the grades of evening and
weekend MBA students still yield informational value in the labor market. We focus on
evening MBA students since they live in the same city as full-time MBA students and are
exposed to the same local shocks, whereas weekend MBA students often fly in from
around the country.
13
A large concern regarding identification in our setting is that there exist changes in
course offerings over time that differentially impact the treatment and control groups. Our
analysis might be distorted if the school we study offers courses after GND that are both
more time intensive and only taught in the full-time program. For example, interactive lab
classes were introduced after GND; we discuss later how our analysis accounts for this
issue.
To motivate that our use of evening MBA students as a control group is a reasonable
empirical strategy, we plot full-time and evening students' average academic effort in all
subjects, core subjects, non-core subjects, a subset of non-core subjects, and a finer subset
of non-core subjects, in Figures 1a-1e, respectively.
First, in Figure 1a there is no sign of a drop in effort for full-time MBA students around
the policy date. In Figure 1b, however, there is a discontinuous comparative drop in full-
time MBA student core subject effort in 2001 that persists into successive years. This is
consistent with GND reducing incentives to exert effort in core subjects. The drop likely
occurs within subject since these subjects are compulsory and students cannot substitute
away from them.
Second, in Figure 1c, there are comparative increases in full-time MBA student non-
core effort in 2003 and 2005 to 2008. This can be driven by students exerting more effort
within non-core subjects or by a change in the course offering.7
To disentangle the above explanations, we examine non-core subjects that were
available to full-time MBA students pre-and post-GND, along with subjects that were
7 While students can substitute towards more difficult subjects, this would not explain the increase unless substitution was so prevalent that new class sections were added or removed for some subjects in response to changed demand. This is because Figures 1a-1e are based on
reported subject effort at the classroom level—each point is a mean of means. Thus, if only a few students move across classrooms, the mean
of means remains unchanged unless there is sufficient heterogeneity in student effort, and a selected sample of students substitute towards difficult subjects. Subject substitution is discussed later in Section 4.3.
14
available to evening MBA students pre-and post-GND. This involves dropping 677 of
5,063 non-core observations. The resulting effort plot, Figure 1d, eliminates changes in
effort driven by changes in the course offering, since the subsample focuses on a fixed
course offering. The increases in full-time non-core effort in 2003 and 2005 to 2008 are
somewhat reduced. Thus, changes in the course offering drive a portion of the effort
increases for full-time MBA students. Discussions with staff involved with course
planning indicated that changes in the course offering around GND were not spurred by
changes in student demand following GND.
We then focus on subjects that were available to both full-time and evening MBA
students, both pre and post-GND. This involves dropping a further 566 observations. The
resulting effort plot, Figure 1e, additionally eliminates subjects available only to full-time
or only to evening MBA students. In this sense, it is the most comparable set of non-core
subjects for full-time and evening MBA students. The increase in effort after GND for full-
time students is even further reduced. Overall, this highlights the potential need to control
for course offering changes.
5.2 Tests of Reduced Student Effort
To test our main prediction, a reduction in student effort following GND, we estimate
the following difference-in-differences model using OLS:
𝐻𝑂𝑈𝑅𝑆𝑛𝑖𝑠𝑔𝑦 = 𝛼 + 𝛽1𝐹𝑈𝐿𝐿𝑔 + 𝛽2𝐺𝑁𝐷𝑦 + 𝛽3𝐹𝑈𝐿𝐿𝑔 × 𝐺𝑁𝐷𝑦 + ∑𝑘𝐹. 𝐸 + 𝜀𝑛𝑖𝑠𝑔𝑦 (1)
g denotes MBA student type (evening or full-time), y denotes year, i denotes instructor,
s denotes subject, and n indexes the instances of isgy combinations (for example, if
Professor X teaches financial accounting to two different classrooms of full-time MBA
students each year, n = 1,2 for those classrooms). FULL is a dummy set to 1 for full-time
MBA classrooms, and GND is a dummy set to 1 when y ≥ 2001 (the year of GND
15
adoption). The coefficient of interest is β3. It is interpreted as the change in HOURS that
full-time MBA students record when moving to GND, incremental to the change in
HOURS that evening students record over the same period. Standard errors are clustered at
the instructor-subject level to account for serial correlation.
Many schools require faculty to teach in only the full-time, or evening, MBA programs.
The school in our analysis, however differs by often scheduling faculty to teach the same
subject in both programs. Faculty typically teach their full-time classes in the afternoon and
immediately travel to teach an evening MBA class (1,821 classrooms are taught under this
setup). This feature lets us exploit instructor-subject-year and instructor-subject-FULL
interaction fixed effects. These fixed effects let us abstract from effort variation caused by
changes in the instructor/subject menu offered to student types across years. They also
negate the effect of instructors and subjects offered to only one student group, or in only
one of the pre- or post-periods. Thus, our design provides estimates within instructor-
subject (the variation we exploit is between treatment and control group, within the same
instructor and subject pair, around the GND implementation date).8
Columns 1 to 4 of Table 4 show the results from estimating variants of Model 1. β3, the
coefficient estimate on FULL×GND, ranges from 0.513 to -0.398 across the different
specifications. The difference between columns 1 and 2, which is explained by the
inclusion of fixed effects, reinforces the need to account for changes in course offerings
around GND implementation. We supplement this finding with the specification in column
3, which requires courses to be available both before and after GND to ensure a
comparable sample. As in column 2, β3 is significantly negative.
8 As an example, full-time students, but not evening students, enroll in certain lab subjects. These subjects are effort intensive, and require
varied effort across years. From discussing these subjects with their instructors, the required effort variation is not due to GND related factors. Thus, including these subjects would contaminate our results.
16
The estimate of β3 in column 2, -0.277, implies a subject effort decrease of 16.6 minutes
per subject per week under GND. Given that the program requires students to complete
twenty 10-week courses, this translates to 55.4 fewer hours over the program. The mean
number of hours spent per week by full-time MBA students on subjects pre-GND is 5.22,
suggesting a 5.31% student effort reduction under GND. This is substantially below the
32% mentioned in Gottlieb and Smetters, who cite Jain (1997). Jain (1997) uses a pre-post
design to measure the effort reduction following GND at The Wharton School. However,
it is imperative to account for effort time trends, as demonstrated in figures 1b, 1d, and 1e.
To test our second prediction, that the effort reduction is greater for core or first year
classes, we re-estimate Model (1) after interacting the independent variables with dummies
indicating core subjects and dummies indicating non-core subjects. Column 5 of Table 4
supports that the impact of GND is greater in economic magnitude for core subjects than
for non-core subjects. This difference, however, is not statistically significant, (p-value of
0.529). A potential explanation for this is the small number of identifying core subject
observations (n=664) compared to the number of identifying non-core subject observations
(n=3,883). While these results are consistent with greater signal generalizability and/or
signal timeliness for core subjects, given the limited sample size, they should be interpreted
cautiously.
When testing our prediction that the effort drop should be isolated to first year students
we focus closely on GND’s implementation date; however, we have no indicator for
whether a classroom predominantly consists of second year MBA students. To circumvent
this, we test the predictions outlined in Table 2 using autumn quarter observations. First
year students are encouraged to complete their core subjects in their first autumn quarter.
Thus, we assume that mostly first year students enroll in core autumn classrooms
17
(Microeconomics, Financial Accounting, Applied Regression Analysis and Business
Statistics), and mostly second year students enroll in non-core autumn classrooms. We use
instructor-subject fixed effects, which have a coarser structure than the fixed effects used in
Table 4, since our focus on autumn classrooms drastically reduces our sample size.
The results are shown in Table 5. Consistent with the predictions outlined in Table 2,
column 1 shows a first year subject effort decrease when moving from the last disclosing
cohort (2000 incoming) to the first non-disclosing cohort (2001 incoming). Column 3 does
not show a first year subject effort decrease when moving from the first non-disclosing
cohort (2001 incoming) to the second non-disclosing cohort (2002 incoming). Consistent
with second year subjects not being timely signals for employers, columns 2 and 4 show no
statistically significant effort decreases for second year subjects.9
5.3 Assessing Identification Assumptions
Parallel Trends:
To ensure that our results are not driven by pre-existing trends in the treatment-control
group effort differential, in Figures 2a and 2b we map out the treatment effect by plotting
the coefficients on year x indicators (Dx) interacted with FULL, the treatment group
indicator, in the following regression:
𝐻𝑂𝑈𝑅𝑆𝑛𝑖𝑠𝑔𝑦 = 𝛽1𝐹𝑈𝐿𝐿𝑔 + ∑ (𝛽2,𝑥𝐷𝑥 + 𝛽3,𝑥 × 𝐹𝑈𝐿𝐿𝑔)
2004
𝑠=1998
+ 𝜀𝑛𝑖𝑠𝑔𝑦 (2)
Figures 2a and 2b provide support for a causal effect of GND on student effort. 2000,
the last pre-GND year, is the benchmark year. There are no indications of a differential
full-time and evening student trend in the years leading up to GND, and we observe a sharp
9 While the magnitude of the decrease in column 2 (second year subjects, moving from the last disclosing cohort to the first non-disclosing
cohort) stands out, this becomes expected when we consider that not all first year students complete their core subjects in their first quarter, i.e., some first year students take subjects that we label as second year subjects.
18
decrease in effort around the policy date. Interestingly, when we limit the sample to core
subjects there appears to be a reversal in the decrease in student effort. One interpretation
of this finding is that evening students begin to comply with GND over time even though it
does not explicitly apply to them.10
We plot the coefficient from 1998 to 2004 for two reasons. First, in this specification,
fixed effects will not allow us to account for changes in course offerings since the treatment
effect is measured year by year. Limiting the time period to a symmetric period around the
GND implementation date makes it more likely that course offerings are balanced
throughout the sample. Second, the school we study completed construction of a new
building in 2004 and full-time MBA students began taking classes there in the 2005
academic year. The increased resources and capabilities of the building may have led to
increased effort from full-time MBA students. It may also have allowed the administration
to offer new, or adjust existing, courses. In either case, we expect there to be a differential
shock to full-time MBA students in 2005 unrelated to GND, and thus, we narrow our
sample window.
To ensure that our inferences are not sensitive to the sample window choice, we re-
estimate Model (1) and limit our sample to 1998-2004. The results, shown in column 4 of
Table 4 are almost identical to those in column 2. As observations fall further in time
beyond 2001, it becomes less likely that the corresponding instructor-subject pairs also
have observations falling in the pre-period. This, however, is a requirement for
observations to contribute to our estimation of the GND treatment effect.
Selection:
10 University staff have informed us that many evening MBA students prefer to present themselves in a similar manner to full-time students, and thus also comply with GND.
19
Students voluntarily adopted GND. It may be that the 2001 incoming class was
unusually effort averse, driving the popular vote in favor of GND, and also the observed
reduction in effort. If GND was the result of an unusually effort averse incoming class,
then one wouldn’t expect the treatment effect to persist into subsequent years. The fact that
we observe a persistent treatment effect means that either GND is causing the reduction in
effort or GND causes subsequent classes to also be effort averse. In other words, low effort
individuals may begin to select into programs with GND.
Table 5 provides evidence against a selection of effort averse students. If the 2001
incoming class was simply more effort averse, we would expect to see a comparable
decline in second year subject effort in 2002. In other words, to explain our results, the
2001 class would have to have been unusually effort averse only towards first year classes.
In summary, our analysis supports the fact that GND causes a reduction in effort and this
reduction appears to operate through a reduction in incentives to exert effort as opposed to
selection.
Financial Recession:
The early 2000s recession affected economic activity during our sample period. If
recessions affect full-time MBA students and evening MBA students differentially, this
may drive our results. For instance, the 2001 incoming class may have been especially
wary of the economic climate, and instead focused on recruiting activities rather than
academics. To address this concern, we note that we begin observing the effort decline in
fall 2000 (which is the start of the 2001 academic year). This is months before March 2001,
when the stock market reached its peak before the dot-com bubble burst.11
Concurrent Change in Course Offerings:
11 The NBER indicates March 2001 to be the start of the recession.
20
Our main estimation relies on subject-instructor fixed effects that remove the average
effect for each subject-instructor pair. Given this structure, however, an alternative
explanation is a shift in effort towards classes not in the pre-period. This explanation
would be concerning if, after GND implementation, there was a large increase in new
courses offered that received a large amount of student effort. For example, consider two
courses: an existing class offered in both the pre and post-periods and a new class offered
only in the post period. Imagine that overall effort remains constant while students
substitute effort towards the new class. Our estimation would suggest a decrease in effort
because only classes offered in both the pre and post periods provide identifying variation
in effort changes. In this scenario, however, there is no causal effect of GND on effort.
We examine this possibility in Figure 3, where we define a new course as one that is
not offered in prior years (with 1998 being the “first” year). In contrast to the scenario
illustrated above, fewer new courses are developed from 1999-2003. Effort spent on these
new courses aligns generally with the effort expenditures throughout our dataset, and are
not high enough to invalidate our results. Furthermore, Figure 3 shows the trend in new
course offerings does not resemble the treatment effects in Figures 2a and 2b; we do not
see a sharp increase in course offerings that occurs only with GND’s implementation.
Taken together, changes in course offerings unlikely drive our main result.
Additionally, using fixed effects could induce a treatment effect if full-time students
substitute effort towards classes that are taught both before and after GND, but only for
full-time students. Interactive lab classes fit this description. Substitution of effort towards
these classes account for the large increase in effort in Figure 1 Panel A. Importantly, any
treatment effect induced by lab classes appears to occur in 2003, which is after the sharp
decrease in effort we observe around GND. Furthermore, the enrollment for lab classes is
21
weighted towards more advanced students, which is inconsistent with the treatment effect
in Table 5 being focused in first year classes.
5.4 Tests of a Shift Towards More Difficult Subjects
When grades are disclosed, students have incentives to engage in GPA-augmenting
activities such as enrolling in easier classes. As such, a potential effect of the GND setting
is a shift towards enrollment in more demanding but potentially more valuable and
specialized classes. The test for sensitivity to subject difficulty requires not only measures
of subject difficulty, but also subject popularity. Our measure of popularity for subject s
among student type g (e.g., full-time MBA) is the proportion (in basis points) of student
type g's non-core enrollments comprising of subject s enrollments in quarter q of year y,
i.e.:
𝑃𝑂𝑃𝑈𝐿𝐴𝑅𝐼𝑇𝑌𝑠𝑔𝑞𝑦 =type g enrollments in non-core subject s in quarter-year, qy
number of type g non-core enrollments in quarter-year, qy (3)
We focus only on non-core enrollments because core subjects are largely compulsory
for MBA students. Further, we exclude summer quarter classrooms, which account for less
than 10% of our sample, because summer course offerings are much more limited.
Subject difficulty, DIFFICULTY, is measured using mean HOURS calculated at the
subject level. Thus, we assume that difficult subjects require more time. We discuss this
assumption further below.
To formally test for a shift towards more difficult subjects following GND, we restrict
the sample to only full-time classrooms and estimate the following model:
𝑃𝑂𝑃𝑈𝐿𝐴𝑅𝐼𝑇𝑌𝑠𝑡𝑞𝑦 = 𝛼 + 𝛽1𝐷𝐼𝐹𝐹𝐼𝐶𝑈𝐿𝑇𝑌𝑠 + 𝛽2𝐺𝑁𝐷𝑞𝑦 + 𝛽3𝐺𝑁𝐷𝑞𝑦 × 𝐷𝐼𝐹𝐹𝐼𝐶𝑈𝐿𝑇𝑌𝑠 + ∑ 𝐹𝐸𝑘
+ 𝜀𝑠𝑡𝑞𝑦 (4)
We use subject area fixed effects to explore the within-subject area margin of
substitution (e.g., a student switching from an easier accounting class to a more difficult
22
accounting class). Subject areas are collected from yearly student handbooks. We expect β3
to be positive if students become less responsive to subject difficulty and the potential for
poorer grades after GND.
The results from estimating Equation (4) are provided in Table 6. Column 1 uses the
sample of full-time classrooms described above. We see a positive shift towards more
difficult courses after the implementation of GND. As a placebo, column 2 uses only
evening classrooms - no reduction in the sensitivity to subject difficulty is observed. Thus
within subject area, difficult subject areas tend to be less popular, though, exposure to GND
appears to induce a shift towards more difficult classrooms. Finally, we re-estimate
Equation (4) using a constant composition sample that requires subjects to be available for
full-time MBA students pre and post-GND, and also available for evening MBA students
pre and post-GND, and obtain similar results.
To sum, the results in this section suggest that under GND students tend to substitute
towards more difficult classes within subject area. An alternative explanation is that
HOURS measures content volume as opposed to difficulty. However, in a highly-ranked
business school, we expect these two concepts are closely related. Further, in the context
of a principal-agent model, GND reduces the costs of a lower GPA from both content
volume and difficulty. Thus both cases lead to the same directional prediction.
6) Labor Market Outcomes
6.1 Predictions and Data
Effort reallocation is consistent with a student response to a decreased sensitivity of
employment outcomes to GPA. Under this explanation, we expect the relation between
GPA and employer desirability to diminish following the adoption of GND. We focus on
employer desirability rather than salary because salaries are empirically sticky, but not
23
employee-employer match outcomes.12 The relation between GPA and employer
desirability might persist, however, if other information that employers use to select
students (e.g., GMAT scores, interview performance, undergraduate GPAs) preserve the
ordering of MBA GPA. In other words, following GND employers might use other signals
or indicators that substitute for GPA.
Our employer rankings are obtained from Universum, who collect career preference
data from students, including MBAs. Universum rankings are based on a two-step process
in which MBA students are first asked which employers they are considering, and then
which of those employers are ideal. The rankings are based on the number of respondents
selecting each employer as ideal. Our Universum data spans 2000 to 2014 - only 2000 falls
in our pre-period. To increase the length of our pre-period, we extrapolate the year 2000
rankings to prior years.13
We use a dataset from the business school we study that contains student GPAs, GMAT
scores, and employment outcomes at both the internship and graduate recruiting stages. In
our data, we observe these academic variables and internship (graduate) recruiting
outcomes only for full-time cohorts starting during the academic years 1999 (1998) to 2007
(2006), inclusive. Unfortunately, this means we cannot construct a counterfactual in these
analyses. Accordingly, we view this analysis as providing evidence consistent with the
results presented earlier in the paper where we are more comfortable drawing causal
inferences.
Finally, in concordance with confidentiality laws, the data are anonymized such that we
cannot link students in the data to their real life identities. Accordingly, our data only
12 In untabulated results, we find the salary and signing bonus differences between high and low GPA students does not change with respect to
GND 13In untabulated results, we use only the 2000 Universum rankings for all years, and find that our conclusions remain unchanged.
24
allow us to link a student's GPA/GMAT to their employer when that employer hired
multiple candidates for a particular position type (internship or graduate) in a given year.
6.2 Empirical Specification
We examine the relation between employer rank and GPA using yearly regressions of
employer rank on GPA, with GMAT included as a control. GPA and GMAT are interacted
with dummy variables denoting each year in our sample to allow a flexible relation
between these variables and employer rank over time;14
𝑅𝐴𝑁𝐾𝑖𝑗𝑦 = ∑ (𝛽1,𝑥𝐷𝑥 × 𝐺𝑃𝐴𝑖 + 𝛽2,𝑥𝐷𝑥 × 𝐺𝑀𝐴𝑇𝑖) + 𝐹𝐸 + 𝜀𝑖𝑗𝑦
2006
𝑥=1999
(5)
Since students can disclose their GMAT scores under GND, we are interested in the
portion of GPA orthogonal to GMAT score. When a student matches with the same
employer twice, i.e., once for an internship and once for a graduate position, we select the
internship match. GPAs are measured at the end of the first fall quarter for internship
recruiting, and at the end of the first summer quarter for graduate position recruiting. We
are thus using the GPAs that students disclosed, or were precluded from disclosing, when
they first engaged in recruiting activity with their matched employer.
Table 7 shows the results from Model (5). The coefficients on GPA become less
negative moving from 2000 to 2001 (the year of GND implementation). A negative
correlation indicates that better GPA students match with better-ranked employers). The
decrease is statistically significant with a p-value of 0.015. This aligns with GND
diminishing the GPA-employer rank relation by concealing useful data from recruiters.
The sharpness of the decrease, while not a substitute for a control group, gives us
confidence that GND is causing the change.
14 The GMAT (Graduate Management Admission Test) is an aptitude test and a key business school student admission criterion.
25
The coefficient estimates on GPA in 2000 are the most statistically significant. The
2000 incoming class could have been exposed to abnormally high GPA-based matching.
The 2001 incoming class may have then responded by implementing GND. This can occur
in both the case where students know their ranking and the case where students do not
know their ranking prior to joining the program (i.e., a ‘Rawlsian veil’ setting).15 It is then
unclear whether GND per se, or mean-reversion, diminishes the GPA-employer rank
relation. A mean-reversion explanation, however, does not explain reduced student effort
unless students anticipate the eventual extent of GPA-based matching and set their effort
levels accordingly. We view this as unlikely.
To examine further whether employers substitute towards other measures of academic
ability following GND, we compute correlations between GMAT and employer rank by
year. GMAT is potentially a measure of at least one dimension of intrinsic student ability.
Table 8 shows that with the exception of 1999 and 2003, none of the years exhibit
significant negative correlations between GMATs and employer rank, even after GND.
Thus, GND does not appear to induce employers to utilize GMAT scores in their hiring
decisions. Our interpretation is that GPAs are incrementally informative to GMAT scores.
7) Conclusion
Using the MBA student setting at a high-ranking business school, we provide empirical
evidence that changes in information precision through disclosure policy affect economic
actions, in particular human capital investment through academic effort. We find that
following the adoption of a grade non-disclosure (GND) policy intended to preclude the
release of student grades to prospective employers, students reduce academic effort by 55.4
15 If sufficiently risk-averse, students unaware of their ranking amongst their peers would prefer a policy that allows them to pool with others in the case that they turn out to be of low rank, despite being unable to distinguish themselves in the case they turn out to be of high rank.
26
hours over the two-year program. Supporting evidence is consistent with students reducing
effort as a response to employers’ declining use of GPA in hiring decisions. We find some
evidence of students weighting their enrollments towards difficult subjects in response to
GND, in support of the arguments of GND proponents.
To provide context to our analysis, we conducted interviews regarding the results of our
study. We contacted 20 employer-recruiting departments to learn about their policies
regarding GPA. Although many did not respond, the consensus among those that did was
that GPA is used primarily as a preliminary screening device. Therefore, the institutional
driver for our results may be that students put effort into achieving high GPAs to avoid
being ruled out for a particular job, not necessarily to out-compete other students that also
meet the screening requirements. When GPAs are not provided, employers tend to focus on
non-GPA elements of resumes and on-campus interactions with students.
It is important to know whether or not GND is welfare improving. Ultimately, this
analysis provides evidence on behavioral responses to GND but is unable to determine the
overall desirability of GND. There are likely many costs and benefits to GND that we do
not explore, but one important reason we cannot conjecture GND’s overall impact is our
inability to observe firm-level changes in productivity that result from employers hiring
new employees with new skillsets.
At first glance, the informativeness principle (Holmstrom 1979) would suggest that the
change in matches between employers and employees reduces productivity because
employers would rationally ignore GND if it was not a useful signal. However, agency
problems within the firm make this conclusion ambiguous. Hiring departments may hire
students with higher GPA not because they are the best fit for the employer, but because
they are the easiest candidates for hiring departments to justify to those they report to.
27
GND may finally give hiring departments the flexibility to hire without relying on GPA,
which could potentially improve firm-level productivity if the reliance on GPA was
causing firm’s to invest in suboptimal matches.
In addition to providing evidence on the behavioral effect of GND specifically, our
study is likely of interest to many audiences within economics that use the principal-agent
model as a workhorse for explaining economic phenomena. Despite the theorized
importance of information precision in markets, there is a paucity of empirical evidence
demonstrating the magnitude of its relation with economic outcomes (e.g. Hong, Hossain,
List, and Tanaka, 2013) because few settings exist in which the requisite counterfactual
actions are observed and information precision varies exogenously (Floyd and List, 2016).
Given the prevalence of principal-agent models in economic theory, confirming the central
implications of these models and establishing a causal relation between information
precision and economic outcomes has far reaching implications. The GND setting affords
us plausible exogeneity and allows us to provide causal evidence concerning the impact of
information precision on economic actions for both the principal and the agent.
28
References
Al- Ubaydli, Omar, et al. "Carrots that look like sticks: Toward an understanding of
multitasking incentive schemes." Southern Economic Journal 81.3 (2015): 538-561.
Christensen, Hans Bonde, et al. " The Real Effects of Mandated Information on Social
Responsibility in Financial Reports: Evidence from Mine-Safety Records." Available at
SSRN 2680296 (2016).
Christensen, Hans Bonde, Eric Floyd, and Mark G. Maffett. "The Effects of Charge-Price
Transparency Regulation on Prices in the Healthcare Industry." Chicago Booth
Research Paper 14-33 (2015).
Ciolli, Anthony. "Grade Non-Disclosure Policies: An Analysis of Restrictions on MBA
Student Speech to Employers." U. Pa. J. Lab. & Emp. L. 9 (2006): 709.
Daley, Brendan, and Brett Green. "Market signaling with grades." Journal of Economic
Theory 151 (2014): 114-145.
Deming, David J. The growing importance of social skills in the labor market. No. w21473.
National Bureau of Economic Research, 2015.
Dranove, David, et al. "Is More Information Better? The Effects of “Report Cards” on Health
Care Providers." The Journal of Political Economy 111.3 (2003): 555-588.
Floyd, Eric, and John A. List. "Using Field Experiments in Accounting and Finance." Journal
of Accounting Research 54.2 (2016): 437-475.
Foster, Andrew D., and Mark R. Rosenzweig. "A test for moral hazard in the labor market:
Contractual arrangements, effort, and health." The Review of Economics and
Statistics (1994): 213-227.
Fung, Archon, Mary Graham, and David Weil. Full disclosure: The perils and promise of
transparency. Cambridge University Press, 2007.
Gottlieb, Daniel, and Kent Smetters. Grade non-disclosure. No. w17465. National Bureau of
Economic Research, 2011.
Guvenen, Fatih, et al. Multidimensional Skill Mismatch. No. w21376. National Bureau of
Economic Research, 2015.
Holmstrom, Bengt, and Paul Milgrom. "Multitask principal-agent analyses: Incentive
contracts, asset ownership, and job design." Journal of Law, Economics, &
Organization 7 (1991): 24-52.
Hong, Fuhai, et al. Testing the theory of multitasking: Evidence from a natural field experiment
in Chinese factories. No. w19660. National Bureau of Economic Research, 2013.
Jain, Anjani. Rethinking grade nondisclosure. An address to faculty and students, University of
Pennsylvania, 1997.
Jin, Ginger Zhe, and Phillip Leslie. "The Effect of Information on Product Quality: Evidence
from Restaurant Hygiene Grade Cards." The Quarterly Journal of Economics (2003):
409-451.
Keser, Claudia, and Marc Willinger. "Principals’ principles when agents’ actions are
hidden." international Journal of industrial Organization 18.1 (2000): 163-185.
Lazear, Edward P. "Performance Pay and Productivity." The American Economic Review 90.0
(2000): 1346-1361.
"News from the schools: Grade concerns at Harvard." The Economist. The Economist
Newspaper, 16 Dec. 2005. Web. 18 Jan. 2017.
Prendergast, Canice. "The provision of incentives in firms." Journal of economic
literature 37.1 (1999): 7-63.
29
Prendergast, Canice. "Uncertainty and incentives." Journal of Labor Economics 20.S2 (2002):
S115-S137.
Singer, Johanna. "The Perks of GND." Financial Times. N.p., n.d. Web. 25 Jan. 2016.
Spence, Michael. "Job market signaling." The quarterly journal of Economics (1973): 355-374.
Verrecchia, Robert E. "Discretionary disclosure." Journal of accounting and economics 5
(1983): 179-194.
Weiss, Andrew. "Human capital vs. signaling explanations of wages." The Journal of
Economic Perspectives 9.4 (1995): 133-154.
30
Tables and Figures
Figure 1a: Hours Spent Preparing for Subjects
Figure 1b: Hours Spent Preparing for Core Subjects
31
Figure 1c: Hours Spent Preparing for Non-core Subjects
Figure 1d: Hours Spent Preparing for Non-core Subjects Offered Pre and Post-GND for Full-
time MBA Students, or Pre and Post-GND for Evening MBA Students
32
Figure 1e: Hours Spent Preparing for Non-core Subjects Offered Pre and Post-GND for Full-
time MBA Students, and Pre and Post-GND for Evening MBA Students
Notes: The data are from a highly-ranked business school’s course evaluation database, aggregated at the classroom level, spanning autumn 1997 to autumn 2013. A “classroom” is an instance of a subject, taught by an instructor, in one academic quarter. HOURS for a class is the average response to “Excluding class sessions, estimate the average number of hours per week spent in preparation or review.”
33
Figure 2a: Yearly Treatment Effects for All Subjects
Figure 2b: Yearly Treatment Effects for Core Subjects
Notes: The figures above plot estimates of β3,x from estimating the following model:
𝐻𝑂𝑈𝑅𝑆𝑛𝑖𝑠𝑔𝑦 = 𝛽1𝐹𝑈𝐿𝐿𝑔 + ∑ (𝛽2,𝑥𝐷𝑥 + 𝛽3,𝑥 × 𝐹𝑈𝐿𝐿𝑔)
2004
𝑠=1998
+ 𝜀𝑛𝑖𝑠𝑔𝑦
The data are from a highly-ranked business school’s course evaluations, aggregated at the classroom level, spanning autumn 1997 to autumn 2013. A “classroom” is an instance of a subject, taught by an instructor, in one academic quarter. g denotes MBA student type (evening or full-time), y denotes year, i denotes instructor, s denotes subject, and n indexes the instances of isgy combinations. HOURS for a given classroom is the average response to “Excluding class sessions, estimate the average number of hours per week spent in preparation or review.” FULL is a dummy denoting Full-time classroom observations. Dx is a dummy denoting classrooms held in academic year x. Confidence intervals are based on robust standard errors clustered at the instructor-subject level.
34
Figure 3: New Courses over Time by MBA Type
Notes: The data in this figure are from the course catalog of a highly-ranked business school. A new class is
defined as one that was not offered in prior years (with 1998 being the “first” year).
35
Table 1: Grade Non-disclosure (GND) in U.S. Business Schools
School GND? GMATa Ranka
Stanford Yes 728 1
Harvard (HBS) Mixedb 724 2
Yale Yes 722 11
Penn (Wharton) Yes 718 4
MIT (Sloan) No 718 3
UC Berkeley (Haas) Partialc 718 8
Dartmouth (Tuck) No 716 7
U Chicago (Booth) Yes 715 6
NYU (Stern) Yes 715 10
Northwestern (Kellogg) No 714 5
Columbia Yes 712 9
UCLA (Anderson) No 710 14
Michigan (Ross) Yes 704 15
UVA (Darden) No 699 13
Duke (Fuqua) No 697 12
Wash U St Louis (Olin) No 695 20
Carnegie Mellon (Tepper) Partialc 694 18
Minnesota (Carlson) No 694 21
U of Florida (Hough) No 694 47
UC Davis No 692 29
USC (Marshall) No 690 22
Cornell (Johnson) No 687 16
UNC (Kenan-Fiagler) No 686 19
Notre Dame (Mendoza) No 685 38
UT Austin (McCombs) No 684 17
Notes: This table describes the grade disclosure schemes at top business schools across the United states. The
table and accompanying footnotes are sourced from Gottlieb and Smetters (2011). aSource: U.S. News and World Report, 2011 bHarvard’s GND policy was officially enforced by the school until 2005, when a new dean terminated the
school’s support. Student support for GND remained high at 87% (Harvard Business School Alumni Bulletin,
2006). Harvard’s current 1-2-3 point grading system, however, effectively pools many students since most
(75%) receive grade 2. In effect, even after abandoning the official nondisclosure policy, the Harvard system is
very similar to a non-disclosure school with honors and a minimum grade requirement. cGND until second round interview
36
Table 2: Predictions for Tests based on Implementation Date
2001 compared to 2000 2002 compared to 2001
First year subjects Effort decreases Effort unchanged
Second year subjects Effort unchanged Effort unchanged
Notes: This table summarizes the predictions for treatment effects around the GND
implementation date. Grade non-disclosure is predicted to reduce effort exerted on first-year
subject coursework for the 2001-incoming MBA cohort (relative to the 2000-incoming cohort).
Table 3: Descriptive Statistics by MBA Type (2011)
Full-Time Evening Weekend
Mean Age 28 29 30
Mean Work Experience 5 years 5 years 6 years
Program Length 21 months 2.5 to 3 years 2.5 to 3 years
Notes: This table reports means across three MBA program types for age and work experience
upon entry as well as typical time to completion. These data are reported directly by the business
school under consideration.
37
Table 4: The Effect of GND on Student Effort
Variables (1) (2) (3)a (4)b (5)
FULL 0.0692 -273.1 0.0793 762.3
(0.118) (540.8) (0.140) (1,167)
GND -0.129 -0.0778
(0.13) (0.146)
FULL×GND 0.513 -0.277*** -0.398*** -0.263***
(0.507) (0.0621) (0.138) (0.0651)
FULL×GND×NONCORE -0.254***
(0.0626)
FULL×GND×CORE -0.524*
(0.274)
Fixed Effects
Ins-Sub-Year X X X
Ins-Sub-Full X X X
R-Squared 0.005 0.844 0.014 0.936 0.844
Observations 5,834 4,547 1,845 1,904 4,547
Notes: This table reports the results from our analysis on the effect of GND on self-reported student effort
using a Difference-in-Differences OLS framework. The data are from a highly-ranked business school’s
course evaluations, aggregated at the classroom level, spanning autumn 1998 to autumn 2013 unless
otherwise noted. A “classroom” is an instance of a subject, taught by an instructor, in one academic quarter.
The dependent variable, HOURS is the classroom-average response to “Excluding class sessions, estimate
the average number of hours per week spent in preparation or review.” FULL and GND are dummy
variables denoting full-time classroom observations and classrooms held during or after academic year
2001, respectively. Estimated interactions of FULL and GND with CORE and NONCORE are not reported
for brevity. A comparable sample requires instructor-subject to be offered at least once pre and post-GND
for full-time MBA students, and pre and post-GND for evening MBA students. Robust standard errors are
reported in parentheses. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels (two-
tailed), respectively. aThis model uses the comparable sample, which requires instructor-subject to be offered at least once pre
and post-GND for full-time MBA students, and pre and post-GND for evening MBA students. bThis model uses the academic window spanning 1998-2004.
38
Table 5: Tests of Reduced Effort Investment around GND Implementation
Variables (1) (2) (3) (4)
FULL -0.195 0.191 -1.318*** -0.165
(0.230) (0.181) (0.315) (0.213)
D2001 0.648 -0.073
(0.426) (0.173)
FULL×D2001 -1.024*** -0.253
(0.339) (0.221)
D2002 -0.935** 0.315
(0.320) (0.227)
FULL×D2002 0.437 0.115
(0.409) (0.274)
Subject Level First Year Second Year First Year Second Year
Comparison 2000 vs. 2001 2000 vs. 2001 2001 vs. 2002 2001 vs. 2002
R-Squared 0.912 0.919 0.898 0.904
Observations 58 129 56 132
Notes: This table reports the results from our analysis on the incoming class of 2001’s effect on self-
reported effort. The data are from a highly-ranked business school’s course evaluations, aggregated at the
classroom level, spanning autumn 1998 to autumn 2013. A “classroom” is an instance of a subject, taught
by an instructor, in one academic quarter. The dependent variable, HOURS is the classroom average
response to “Excluding class sessions, estimate the average number of hours per week spent in preparation
or review.” FULL is a dummy denoting Full-time classrooms. Dx is a dummy denoting classrooms held in
academic year x. All estimations include Instructor-Subject fixed effects. Robust standard errors are
reported in parentheses. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels (two-
tailed), respectively.
39
Table 6: Tests of Reduced Sensitivity to Subject Difficulty
Variables (1) (2)
GND -22.15 18.81
(16.90) (24.45)
DIFFICULTY -14.59*** -8.057*
(2.80) (4.715)
GND×DIFFICULTY 4.329*** -2.037
(1.873) (4.759)
Sample Full-time Evening
R-Squared 0.137 0.113
Observations 1,864 1,632
Notes: This table reports the results from our analysis on the effect of GND on the popularity of difficult
classes using a Difference-in-Differences OLS framework. The data are from a highly-ranked business
school’s course evaluations, aggregated at the classroom level, spanning autumn 1998 to autumn 2013. A
“classroom” is an instance of a subject, taught by an instructor, in one academic quarter. The dependent
variable, POPULARITY is the the proportion (in basis points) of student type g's non-core enrollments
comprising of subject s enrollments in quarter q of year y. GND is a dummy variable indicating classrooms
held during or after academic year 2001, DIFFICULTY is the classroom average response to “Excluding
class sessions, estimate the average number of hours per week spent in preparation or review.” All
estimations include Subject Area fixed effects. Robust standard errors are reported in parentheses. ***, **,
and * indicate statistical significance at the 1%, 5%, and 10% levels (two-tailed), respectively.
40
Table 7: The Effect of GND on GPA-Based Student-Employer Matching
Variables (RANK)
D1999×GPA -9.681
(-1.489)
D2000×GPA -8.749***
(-4.810)
D2001×GPA 0.217
(0.0694)
D2002×GPA -5.244
(-1.435)
D2003×GPA 5.200*
(1.917)
D2004×GPA -5.121**
(-2.155)
D2005×GPA -2.254
(-0.468)
D2006×GPA 4.628***
(2.832)
R-Squared 0.312
Observations 2,656
Notes: The data come from a highly-ranked business school’s data archives. Each observation corresponds
to a student-employer match. When a student matches with the same employer twice, i.e., once for an
internship and once for a graduate position, we select the internship match. RANK is the employer rank that
comes from Universum, who survey MBA students about the desirability of MBA employers. Universum
data spans 2000 to 2014. Thus, to increase the length of our pre-period, we extrapolate the year 2000
rankings to prior years. GPA is student GPA measured as at the end of the first fall quarter for internship
recruiting, and as at the end of the first summer quarter for graduate position recruiting. GMAT score is the
maximum GMAT score obtained by a student and listed on their MBA program application. Dx is a dummy
denoting students from cohorts commencing their studies in academic year x. All estimations include
Industry-Cohort fixed effects. Robust t-statistics are reported in parentheses. ***, **, and * indicate
statistical significance at the 1%, 5%, and 10% levels (two-tailed), respectively.
41
Table 8: Correlation between Student GMAT Scores and Employer Rank
1999 2000 2001a 2002 2003 2004 2005 2006
Correlation -0.11 -0.01 -0.08 -0.01 -0.12 -0.06 -0.04 0.00
t -1.81 -0.10 -1.23 -0.08 -1.85 -1.08 -0.88 -0.01
Observations 269 274 265 206 255 340 401 376
Notes: The data come from the data archives of a highly-ranked business school. Each observation
corresponds to a student-employer match. When a student matches with the same employer twice, i.e., once
for an internship and once for a graduate position, we select the internship match. GPAs are measured as at
the end of the first fall quarter for internship recruiting, and as at the end of the first summer quarter for
graduate position recruiting. Employer ranks come from Universum, who survey MBA students about the
desirability of MBA employers. Universum data spans 2000 to 2014. Thus, to increase the length of our
pre-period, we extrapolate the year 2000 rankings to prior years. aFirst year that GND was implemented