University of Southern California Dissertations and Theses

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

(USC Thesis Other)

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content TWO ESSAYS IN ECONOMIC POLICY:
THE INFLUENCE OF PERCEIVED COMPARATIVE NEED
ON FINANCIAL SUBSIDY REQUESTS
&
UNINTENDED CONSEQUENCES OF THE FOSTA-SESTA LEGISLATION
by
Jacob Joseph Schneider
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2021
Copyright 2021 Jacob Joseph Schneider
TABLE OF CONTENTS
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .iv
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
ABSTRACTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
ESSAY 1: The In
uence of Perceived Comparative Need on Financial Subsidy
Requests: Experimental Evidence from University Travel Programs
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Study Context, Data, & Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4 Summary Stats, Biased Beliefs & Predictors of Aid Requests . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5 The Eect of Information on Requested Financial Assistance . . . . . . . . . . . . . . . . . . . . . . . . . .30
6 Heterogeneous Response to Information on Comparative Need . . . . . . . . . . . . . . . . . . . . . . . . 34
7 Distributional Shifts In Requested Aid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .38
Tables & Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
ESSAY 2: Unintended Consequences of the FOSTA-SESTA Legislation
9 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
10 Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .61
11 Removal of Craigslist Personals, Rape, & Sexual Assault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
12 The Role of \Me Too" and Backpage.com . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .66
13 Impact on Prostitution Arrests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
14 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .73
Tables & Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
APPENDIX A: Supplementary Materials for In
uence of Perceived Comparative Need
A.1 Supplementary Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A.2 Predictors of Aid Requests for Experimental Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
A.3 Estimation of Additional Treatment Eect Specications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
A.4 Full Sample Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
APPENDIX B: Supplementary Materials for Consequences of FOSTA-SESTA
B.1 Supplementary Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
B.2 Robustness: Controlling for MSA Demographic Eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population . . . . . . . . . . . . . . 123
B.4 Robustness: Balanced Panel Results on Prostitution Arrests . . . . . . . . . . . . . . . . . . . . . . . . 128
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
ii
ACKNOWLEDGEMENTS
I am deeply grateful to Paulina Oliva, Vittorio Bassi, Geert Ridder, Tatiana Melguizo, Je Weaver,
and Manisha Shah for their input and advising on these projects. I'd like to thank Charles Mansour,
Je Peak, Annie Weaver, Hallie Dowling-Huppert, Charlotte Brown, Darcy Phillips, Stine Odegard,
J Swanger, Sarah Rielley, Kyle Anderson, Allison Loecke, Ellen Lentine, and Amanda Torrence for
their assistance in the collection of data for the RCT. Additionally, I'd like to express a deep
gratitude to Donald and Susan Schneider for all their support during my academic preparations
and for editorial revisions of these papers. Finally, I'd like to thank Carlos Cambero for his support
throughout the preparation and completion of this work. It would not have been nalized as quickly,
as enjoyably, or as well without him. All errors are my own.
I dedicate this manuscript to SUSAN and DONALD SCHNEIDER
iii
LIST OF TABLES
Essay 1: The In
uence of Perceived Comparative Need
1 Baseline Summary Statistics and Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2 Predictors of Applying for Aid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3 Predictors of Amount of Aid Requested (as % of Student Fee) . . . . . . . . . . . . . . . . . . . . . . . . 44
4 Marginal Treatment Eect along Bias in Prior on Comparative Need . . . . . . . . . . . . . . . . . .46
5 Dierential Treatment Eect: Over/Under Estimators of Need (Binary) . . . . . . . . . . . . . . .47
6 Dierential Treatment Eect: Over/Under Estimators of Need (Terciles of Bias) . . . . . . 47
7 Overall Average Treatment Eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .49
8 Heterogeneous Eects on Amount of Aid Requested (Separate Regressions) . . . . . . . . . . . 50
9 Heterogeneous Eects on Amount of Aid Requested (Single Regression) . . . . . . . . . . . . . . . 51
10 Treatment Eects on Applying for Aid: EFC x Overestimates . . . . . . . . . . . . . . . . . . . . . . . . .53
11 Treatment Eects on Amount of Aid Requested: EFC x Overestimates . . . . . . . . . . . . . . . .53
12 Treatment Eects: EFC and Personal Savings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
13 Distributional Eects on Applying for Aid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
14 Distributional Eects on Amount of Aid Requested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Essay 2: Unintended Consequences of FOSTA-SESTA Legislation
15 DID Estimate: Craigslist Personals on Rape / Sexual Assault Post 2018 . . . . . . . . . . . . . . 76
16 Event Study Estimates: Craigslist Personals on Rape / Sexual Assault . . . . . . . . . . . . . . . . 77
17 MSA Characteristics in 2017 by Craigslist Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
18 DID Estimate: All Treatments on Rape / Sexual Assault Post 2018 . . . . . . . . . . . . . . . . . . . 79
19 Event Study Estimates: All Treatments on Rape / Sexual Assault . . . . . . . . . . . . . . . . . . . . 80
20 DID Estimate: Impact on Prostitution Arrests Post 2018 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
21 Event Study Estimates: Impact on Prostitution Arrests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
iv
LIST OF FIGURES
Essay 1: The In
uence of Perceived Comparative Need
1 Histogram of Bias in Prior Perception of Need . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2 Plot of Actual Decile vs. Perceived Percentile of Need . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .42
3 Inverse Demand for Program Participation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45
4 Histogram of Amount of Aid Requested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45
5 Treatment Eects on Applying for Aid: Terciles of Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6 Treatment Eects on Amount of Aid Requested: Terciles of Bias . . . . . . . . . . . . . . . . . . . . . . 48
7 Treatment Eects on Applying for Aid: EFC x Overestimates . . . . . . . . . . . . . . . . . . . . . . . . .52
8 Treatment Eects on Amount of Aid Requested: EFC x Overestimates . . . . . . . . . . . . . . . .52
9 Treatment Eects: Terciles of EFC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
10 Treatment Eects: Terciles of Savings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .54
11 Treatment Eects on Applying for Aid: EFC x Savings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
12 Treatment Eects on Amount of Aid Requested: EFC x Savings . . . . . . . . . . . . . . . . . . . . . . 56
Essay 2: Unintended Consequences of FOSTA-SESTA Legislation
13 US Trends in Rate of Rape and Sexual Assault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
14 Trend in Rate of Reported Rape and Sexual Assault by Craigslist Treatment . . . . . . . . . .76
15 Event Study Coecients for Rape/Sexual Assault: Craigslist Treatment . . . . . . . . . . . . . . 77
16 Event Study Coes. for Rape/Sexual Assault (Cond. on All Treatments): Craigslist . . 81
17 Event Study Coes. for Rape/Sexual Assault (Cond. on All Treatments): Backpage . . 81
18 Event Study Coes. for Rape/Sexual Assault (Cond. on All Treatments): Me Too . . . . 81
19 Distribution of DID Coecients from Randomization Inference . . . . . . . . . . . . . . . . . . . . . . . 82
20 Trend in Trend in Juvenile Prostitution Arrests by Craigslist Treatment . . . . . . . . . . . . . . 84
21 Event Study Coecients for Juvenile Prostitution Arrests: Craigslist Treatment . . . . . . .85
v
ESSAY 1:
THE INFLUENCE OF PERCEIVED COMPARATIVE NEED
ON FINANCIAL SUBSIDY REQUESTS:
EXPERIMENTAL EVIDENCE FROM UNIVERSITY TRAVEL PROGRAMS
Abstract
I conduct an RCT, in collaboration with 13 U.S. universities, to study the impact of student
mis-perceptions, regarding comparative economic need, on the propensity to apply for nancial
aid for co-curricular educational travel programs and on reported willingness and ability to pay
for one's own participation. Students are randomly assigned to receive accurate information on
university distributions of Expected Family Contribution (EFC), a standard measure of nancial
need in higher education. Individuals who learn that they've overestimated their comparative need
(measured by EFC), reduce demand for nancial assistance relative to peers who underestimate or
have low ex-ante bias, resulting in a decrease in subsidy requests among lower need (high EFC)
students. However, high need (low EFC) students, and especially those with low personal savings,
increase demand for subsidies, driving a 2.8 percentage point net increase in requested assistance in
response to the information. While the treatment eect among students who initially overestimate
their comparative need suggests the salience of distributional fairness concerns among potential
beneciaries of aid, a net increase in requested assistance implies this type of information provision
may have limited use as a policy tool. Institutions must weigh potentially modest distributional
gains against resource eciency in their process of allocating subsidy resources.
vi
ESSAY 2:
UNINTENDED CONSEQUENCES OF THE FOSTA-SESTA LEGISLATION
Abstract
This paper presents the rst causal evidence on the impact of FOSTA-SESTA, a major U.S. policy
aimed at reducing online sex tracking and especially the tracking of minors. I document the
unintended policy consequence of at least a 6% increase in the rate of reported rape and sexual
assault { owing to the elimination of Craigslist Personals { in U.S. Metropolitan Statistical Areas
with high Craigslist usage prior to the policy change. This eect implies a substitution toward rape
by men who would have otherwise utilized Craigslist Personals to match with either commercial
or non-commercial sexual partners. In ruling out potential confounding eects on reported rape
driven by the \Me Too" movement, I oer the rst causal evidence that \Me Too" brought about
a decrease in rape and sexual assault in the U.S. by 2019. Finally, I document an increase in
female juvenile prostitution arrests, post FOSTA, in MSAs with high usage of Craigslist. This
evidence is suggestive of a movement from online to street solicitation by minors who utilized
Craigslist Personals for either coerced or un-coerced sex solicitation prior to the policy change.
Given previous evidence on higher risk of physical harm among women engaged in street vs. online
solicitation, this evidence may suggest limited policy eectiveness in improving the welfare of the
population it aimed to protect, however, further data would be required to speak to the net policy
impact on sexual tracking.
vii
ESSAY 1:
THE INFLUENCE OF PERCEIVED COMPARATIVE NEED
ON FINANCIAL SUBSIDY REQUESTS:
EXPERIMENTAL EVIDENCE FROM UNIVERSITY TRAVEL PROGRAMS
1 Introduction
Beliefs play a profound role in personal economic choice, but also in the determination of societal
outcomes as they shape political engagement and voting behavior and, in turn, policies such as
taxation, re-distribution, and social safety nets. However, personal beliefs are often incorrect leading
individuals to optimize dierently than they would if possessing full and correct information on the
state of the world. Building from Piketty (1995), which rst modeled dierences in redistributive
tax policies as arising from diering beliefs about social mobility, a growing literature has studied
the role of subjective perception, and also mis- perceptions, on demand for redistribution. We
have learned that beliefs about the origins of wealth and poverty and on the fairness of social
competition are important determinants of preferences for redistribution (Alesina and Angeletos,
2005; B enabou and Tirole, 2006; Alesina et al., 2012). But beliefs about others also matter. E.g., the
extent to which others free-ride shapes willingness to contribute toward public goods (Fischbacher
and Gachter, 2010). And in both Europe and the U.S., citizens hold systematic negative biases
toward immigrants { that they are less educated, less employed and more reliant on government
support than is the case { views which in
uence expressed preferences for redistibutive policies
(Alesina et al., 2018).
This study contributes to a subset of the literature on preference for redistribution which has
focused on role of beliefs regarding one's own comparative economic well-being. Perceptions of com-
parative well-being { e.g., one's perceived percentile in the income distribution { matter both for
1
subjective happiness (Perez-Truglia, 2020) and for the extent to which one supports redistribution
policies and gives charitably to the poor (Cruces et al., 2013; Nair, 2018). Moreover, individ-
uals systematically mistake their comparative economic position. Cruces et al. (2013) advances
understanding that such biases arise in part from naive inference on the shape of the income dis-
tribution deriving from the non-representative reference group with which individuals most often
associate. As a result, individuals tend to believe themselves to be more average than they are.
In some advanced economy settings, studies have shown that individuals believe themselves to be
systematically poorer than they actually are (Karadja et al., 2017; Nair, 2018). And these studies
demonstrate that this result has real implications for distributional outcomes.
This paper studies the in
uence of this type of bias in a novel context, broadly relevant to
public policy. It seeks to address the question of how mis- beliefs about one's comparative eco-
nomic well-being distort requests for transfers { to support private consumption { in settings where
subsidy resources are available to individuals with need. I design and conduct a eld experiment
in the context of U.S. higher education. Utilizing the measure of Expected Family Contribution
(EFC) { a nationally standard metric of economic need in higher education { I experimentally vary
the provision of accurate information on relative need ranking to students who are applying to
educational travel programs. Since these programs require a participation fee ($365 on average),
programs oer nancial assistance to equalize educational opportunity. Students, thus, have the
ability to apply for nancial support, subsequent to their program application. By randomizing
the provision of information on the distribution of Expected Family Contribution, we are able to
study how individuals who initially overestimate or underestimate their relative need, compared to
university peers, shift in demand for scholarship assistance.
The experimental design contributes to understanding regarding the impact of biased beliefs
on demand for redistribution in a novel setting: subsidy seeking behavior. It is, at the same time,
motivated by a problem inherent { from the institutional perspective { in the allocation of subsidy
resources. Though executed in an educational context, the economic problem taken up in this
study is broadly relevant to policy setting where government or a private institution would like
to encourage consumption { e.g., of a vaccine or mosquito net { for individuals whose willingness
or ability to pay falls below the cost of provision of a good or service. Regardless of the origin
of subsidy funding { whether a re-distributive taxation policy, institutionally allocated budget, or
2
private donation { an ecient utilization of resources is desired. In the most ideal scenario this
would entail the honest elicitation of potential beneciaries' willingness to pay (WTP) for the good
and the subsidization of the dierence between that amount and the cost (or price).
1
In commercial
settings, a non- incentive compatible elicitation of WTP would be expected to signicantly under
represent actual WTP due to the economic incentive to purchase at the lowest cost and the social
acceptability, in the commercial space, of \negotiating a good deal." In our setting, and perhaps
other policy settings where agents understand they are beneciaries of redistribution intended for
individuals with need, the incentive to under-report WTP, or request a subsidy when it is not truly
needed, may be partially oset by a preference for distributional fairness or honesty. E.g., among
study participants only 34% of control participants apply for nancial assistance even though the
eort required to do so is negligible and there is no criteria which prevents applying (nor does
nancial need have any bearing on program acceptance). Control participants who do apply for
aid report being able to cover 46% of cost on average. And 3.3% of aid applicants (in the control
group) inform the program that, though they would like to be considered for assistance, they would
be able to cover the full cost.
A non-incentive compatible elicitation of WTP from individuals requesting a subsidy in such
setting { even if understood as a lower bound on actual WTP { can only improve the eciency of
resource allocation, by signaling new (though potentially biased) information.
2
To the extent that
individuals perceive as salient and have a preference for distributional fairness, honesty, resource
eciency (of the govenrment or insitution), etc., the signal should become more representative of
actual WTP. In this study we are not directly concerned, nor do we collect data on how programs
actually allocate their aid resources among applicants. However, given that students must choose
1
There is a natural relationship to the problem of price discrimination in eld of Industrial Organization, however,
applied in a context where willingness to pay falls below cost, the optimization problem becomes one of minimizing
loss rather than maximizing producer surplus.
2
The alternative would be to provide a xed amount to all who request support or a tiered schedule of assistance
targeted on the observable characteristics of potential beneciaries. However, even a tiered subsidy schedule, which
is perfectly calibrated to ensure the participation of a desired percentage of individuals, conditional on observable
characteristics, will end up granting an economic surplus to many beneciaries since it cannot account for individual
heterogeneity in actual WTP. Thus, even the best subsidy schedule can be improved upon in terms of resource
eciency { at least marginally { by asking individuals who request assistance to share their willingness and ability
to pay for their own consumption and by requesting that they contribute the maximum of that or their scheduled
amount (or some amount in between). Note that, while incentive compatible elicitation of WTP is useful for
studying demand, it is not implementable as a policy strategy at the individual level because it still inevitably
restricts consumption of the good or service increasingly by \need" of subsidy. I.e., those with the lowest willingness
to pay{often the poorest{will disproportionately be excluded from consumption by standard methods such as Becker-
DeGroot-Marshak (Becker et al. (1964)).
3
whether to apply for aid and how much to indicate to the program that they can cover out of
pocket, we take an interest in how ex-ante bias in beliefs regarding comparative need impact these
equilibrium choices. In our setting, pilot data collection indicated a similar pattern of bias as
that observed by Cruces et al. (2013): students perceive themselves as more average than they are.
That is, high need students tend to underestimate their comparative need while lower need students
overestimate on average. This pattern implies that, ex-ante, the net impact of providing accurate
information on percentile of need is theoretically unclear and a matter for empirical investigation.
If lower need individuals, who are more likely to overestimate their comparative need, tend to
reduce their requests for assistance in response to information, while higher need individuals request
similar levels of assistance, programs can improve their position in terms of eciency. If those who
overestimate need request less support, while those who underestimate request more support, but
in a way that more or less balances net demand for subsidy assistance, there may be no impact on
eciency. However, overall, requested assistance has shifted more toward individuals with higher
need. Finally, if those who underestimate their relative need request more aid in response to
information and this increase is oset little by those who overestimate, there may be a net increase
in requested aid. It will, however, be directed on average toward higher need students.
The main experimental ndings indicate that individuals who underestimate their position of
need { i.e., who discover they are comparatively worse o than they thought { as well as those
who demonstrate little or no bias in ex-ante beliefs, request a larger subsidy on average when
provided accurate information on the distribution of need among peers. This eect appears driven
largely by high need (i.e., low EFC) students with low personal savings who increase their ask. At
the same time, individuals who initially overestimate their comparative need ranking request less
aid than peers in response to the information intervention. This results in an overall decrease in
requested nancial assistance among lower need (high EFC) students. However, the increase in
aid requests, observed among students with high relative need, dominates. Overall, we estimate a
2.8 percentage point increase in nancial aid requested. These results suggest that, while there is
evidence of pro-social response among students who initially overestimate their comparative need,
as well as a distributional shift in assistance requests from lower to higher need students, providing
accurate information to individuals on comparative economic well-being may come with a cost.
Distributional gains must, therefore, be weighed against potential eciency loss.
4
This work contributes primarily to two bodies of literature. The rst is research that has focused
on subjective beliefs and demand for redistribution. In this area, the present study is most closely
related to work by Cruces et al. (2013), Karadja et al. (2017), and Nair (2018), each of which
study how the updating of beliefs about relative income position impact support of redistribution.
Cruces et al. (2013) and Karadja et al. (2017) both nd that individuals respond to information
on the income distribution in a way that is more or less self serving, i.e., individuals who discover
they are relatively worse o increase support of redistributive policy (Cruces et al.) or individuals
who discover they are relatively better-o decrease support of redistributive policy (Karadja et
al.). Nair (2018), however, studying the outcome of charitable giving, documents that individuals
respond in a pro-social manner. Those who learn that they are more well-o than believed increase
donations to global charity.
This paper provides the rst evidence on the impact of biased beliefs, regarding relative eco-
nomic well-being, on the behavior of individuals seeking subsidies for private consumption from an
established redistribution program. Additionally, this is the rst evidence which studies the impact
of this type of information bias on behavioral outcomes deriving from an entirely natural eld set-
ting. I.e., rather than polling individuals on preference for redistribution (or asking them to make
a donation out of their study compensation), we are able to observe actual nancial assistance
requesting behavior subsequent to the information treatment. In this novel setting, we provide
evidence both that information is used in a way that is self supporting and in a way that suggests
pro-social concern. Individuals who receive conrmation that they are at a high percentile of need
request more nancial assistance on average, while those who discover they've overestimated their
comparative need request comparatively less. However, increased requests by high need students
dominates overall.
Second, this paper adds to a broad literature which studies the trade-os between nancial
payos and pro-social concern. Experimental studies in the laboratory have documented the will-
ingness of agents to sacrice some personal monetary benet in favor of distributional concerns
and have tested the explanatory power of various psychological mechanisms at play such as in-
equality aversion, eciency, and fairness and reciprocity considerations (Forsythe et al. (1994),
Fehr and Schmidt (1999), Bolton and Ockenfels (2000), Engelmann and Strobel (2004), Frey and
Meier (2004)). A key insight deriving from this research is that heterogeneity of individuals in the
5
intensity of preference for fairness { i.e., how much personal benet an agent would be willing to
sacrice to achieve a \fair" outcome { and also in beliefs of what constitutes a \fair" distributional
outcome, are both important for understanding the equilibrium choice of individuals when fairness
concerns are at play (Cappelen et al. (2007), Klor and Shayo (2010), Balafoutas et al. (2012)).
Complementary studies of pro-social behavior in the eld have focused extensively on charitable
giving, observing that, while \genuine altruism" does exist { evidenced by anonymous philanthropic
giving { anonymous giving represents at most 1% of donations (Glazer and Konrad (1996)). Lev-
els of donations are strongly aected by the extent to which giving is visible to others (Buraschi
and Cornelli (2013), Nicola and Macis (2008), Lacetera and Macis (2010), Tonin and Vlassopoulos
(2013)). B enabou and Tirole (2006) develop a theoretical model of intrinsic and extrinsic incentives
of \pro-social" behavior and B enabou and Tirole (2010) provide a review and summary of empirical
ndings related to social image concerns and altruistic behavior.
This paper contributes to the pro-social literature by studying the impact of information on
economic behavior in a new and natural eld setting where distributional fairness considerations are
salient. Agents choose whether to request nancial support from a re-distributive program for which
they are eligible, but where resources are nite. Individual choice to honestly disclose ability to pay
cannot be enforced, however this disclosure has implications for the overall eciency of program
resources and potentially also for the payos of peers. In this way, we are particularly interested in
the response of individuals who initially overestimate their position of need. A reduction in nancial
assistance requested by this group in response to accurate information suggests that individuals
have a preference for distributional fairness which in
uences their propensity to apply for assistance
and/or how much they indicate to the program that they are willing and able to contribute out
of pocket. We nd that students who initially overestimate need do request relatively less aid
when exposed to information compared to peers who underestimate or have low bias. This result
appears robust to controlling for other characteristics that dier across bias and may explain the
dierential eect. We also explore heterogeneity in treatment response which suggests that dierent
preferences about fairness in the distribution of nancial assistance matters for how students who
overestimate need respond to information.
The structure of the paper is laid out as follows: Section 2 provides information on the study
context, experimental design, and data collection process. Section 3 describes the empirical methods
6
used and specications estimated to identify treatment eects. Section 4 presents participant
descriptive statistics and balance as well as evidence on biased perceptions at baseline and the
predictors of applying for aid and of willingness to pay for program participation. Section 5 discusses
the primary experimental results: the eect of biased beliefs and the overall average treatment eect
of information provision. Section 6 explores heterogeneity of treatment eects along six measures
pre-specied in the study pre-analysis plan. Section 7 presents evidence on distributional shifts in
assistance requests along the dimensions of Expected Family Contribution and individual savings.
Finally, Section 8 discusses the policy implications of results and concludes.
2 Study Context, Data, & Experimental Design
I study the in
uence of beliefs of comparative economic well-being on subsidy requesting behavior
in the context of educational travel programs in U.S. universities. In many ways this setting
provides an ideal context for this type of study, as institutions of higher education frequently
allocate nancial assistance for co-curricular and extra-curricular programming, desiring to equalize
access to university opportunities for students facing greater economic constraint or hardship. This
section provides context on the the structure and nancial aid application process for the travel
programs as well as details on the data collection for this study and the experimental design. Note
that the elements of the experimental design and primary hypotheses to be tested were laid out in
a pre-analysis plan which was registered prior to data collection.
2.1 Universities and Travel Programs
Data was collected in collaboration with 13 U.S. universities. Participating universities were iden-
tied due to the structure of their travel program and nancial assistance availability conducive to
study participation. By sta at an initial set of 3 universities, where preliminary data was collected
and a pilot intervention was run, 50 other universities were identied having similar program struc-
tures. Each of these was contacted to assess participation interest and whether the administrative
structure of the program would t with the study design. Of the 53 universities, 32 were elimi-
nated after an initial conversation due to lack of interest, having too small a program, or having a
scholarship award process that would not t with the study design (e.g., not having nancial assis-
7
tance available to participants or a generalized process for applying for aid). Of the remaining 21
programs, 7 began initial conversations about participation, but either did not receive managerial
approval or dropped out due to stang changes or time constraints prior to data collection. The
remaining 14 universities participated in baseline data collection and the information intervention.
However, one university did not follow through with delivering outcome data from scholarship ap-
plications following the baseline survey, therefore, they are excluded from the study sample. The
geographic distribution of participating universities is depicted in Appendix Figure A.1.
Though the travel programs vary somewhat in content and structure across universities, they
should be understood as educational opportunities during school recesses{fall, winter, spring or
summer break{where students travel together to a destination, accompanied by a university sta
and/or faculty member, to learn about dierent themes related to their course of study or general
interest, e.g., local social or environmental issues, economic development, sustainability, con
ict and
peace, etc. The trip oerings in our sample range from 3-17 days with the median trip length being
7 days and a mean trip length of 7.3 days. Destinations may be domestic to the U.S. (91.3%) or
international (8.7%). Appendix Figure A.2 depicts photos of student program participants during
their experience.
Important for understanding the study design is the structure and timeline of student applica-
tions for program participation and nancial assistance. All collaborating programs have an online
application or registration process to select and enroll student participants. Additionally, programs
have nancial assistance available which can be requested via an online scholarship application. In
most cases, students desiring to request aid will apply for program participation, followed within
a day or two by a scholarship application, such that a nancial assistance determination can be
provided before the student must commit to participate with a down payment. Several programs
also allow for rolling scholarship applications up until the time of the trip itself, when student fees
must (typically) be paid in full. While the timing of aid applications vary a bit across programs,
essential for the study design, a scholarship application must always be led after completion of
the online program application or registration. Since the baseline survey is conducted jointly with
program applications, this ensures that all students in our sample respond to the baseline survey
and are exposed to the information intervention before a nancial aid application is initiated.
8
2.2 Study Sample
The sampling frame consists of all students from collaborating universities who applied for travel
program participation during the course of the study. Students were asked within their online
program application (or registration) if they would be willing to take part in the study. They were
provided an informed consent indicating that study enrollment was voluntary and would involve
responding to a survey, which would be joined to certain data from their program application and
their nancial aid application (in the case that they applied for nancial assistance). Students were
informed that study participation and any information provided in the survey would have no bearing
on program acceptance or on nancial assistance awarded and that information collected was for
academic research and would not be viewed by university sta. While there was no compensation
oered for responding to the baseline survey, students were informed that they could enter a rae
to win a $100 Amazon gift card.
Of 2,879 total program applications/registrations across the 13 universities during the study
period, we obtain a sample of 1,029 students who opt to participate (35.7% of all applications).
Though we have no individual data on students who declined study enrollment, we do possess the
aggregate count of scholarship applicants. This allows us to evaluate selection into the study along
the margin of student propensity to apply for aid. During the study period there were a total of 829
scholarship applications (28.8% of program applications). In the study control group we observe
180 scholarship applications out of 528 total control participants (34.1%). Conducting a simple one-
sample test of proportion we can reject at the 1% condence level that the control group proportion
of aid applicants could constitute a random draw from the population of total program applicants
suggesting either (a) students who opt to participate in the study are slightly more likely to apply
for aid than the average program applicant, (b) study participation itself increases the propensity to
le an aid application, or (c) both. That study participation increased aid applications, even for the
control group, is feasible given that the baseline survey may raise the salience of available nancial
assistance to students previously unaware; however, it is also quite possible that study respondents
dier from the average applicant given that students with the intention of applying for aid may feel
more drawn to participate due to the study's focus on program nancial assistance. That said, the
study control group and total population proportions of aid applicants do not diverge dramatically,
9
which provides some condence in terms of external validity of experimental results extrapolating
to the total population of university program applicants. If, e.g., a majority of study applicants
were applying for aid when only 29% of total participants do, this would raise greater concern
about how the total population of students may respond dierently to this type of information
intervention. That said, study participants may also dier on other margins which matter and are
not observable, such as prosociality.
2.3 Data Sources and Collection
The data used in this study derives from one of four sources: (1) Some participant information,
such as the program being applied for and demographics such as gender and student gpa come from
student program applications. (2) The majority of baseline characteristics of student participants
were collected via an online study survey which was completed at the time of application to the
university travel program. (3) The baseline survey was then matched to scholarship applications for
all study participants who ended up applying for travel program nancial assistance. (4) Finally,
university and program specic administrative data was utilized to generate certain controls and
measures necessary for the information intervention and analysis of results. The progression of the
student application process and baseline data collection is depicted in Appendix Figure A.3.
The baseline survey was collected on a rolling basis throughout the 2019-2020 academic year
(Aug 2019 - May 2020), coinciding with the timing of travel program application processes. A
pre-analysis plan which was registered prior to the start of data collection indicated the initial
intention to continue incorporating new participants during the 2020-2021 academic year, however,
due to Covid-19, all participating programs ceased their operations and applications making further
data collection impossible. For this reason, some results are under-powered from what was initially
desired, however, results are still informative to the main research questions.
Every student who applied to one of the participating programs over the period of data collec-
tion was asked during the course of their online program application whether they would be willing
to participate in the study. 35.7% of student applicants opted to do so. Those that did responded
to the baseline questionnaire prior to concluding and submitting their program application. The
baseline collected all individual level control variables as well as the measures used for analysis of
10
heterogeneous treatment eects (especially, ex-ante beliefs about one's position of relative nancial
need compared to peers). In addition to collecting participant information, the experimental pro-
vision of the information treatment also occurred within the baseline survey as described in greater
detail in the following section.
The two primary outcomes of interest derive from the second data source: program scholarship
applications. First, whether a particular student has record of a scholarship application serves as a
measure of the extensive margin of making a nancial assistance request. Second, for those students
who do request nancial aid, the scholarship application captures a measure of their reported
willingness to pay for program participation. Greater detail on the elicitation of this metric is
provided in Section 2.4. The timing of scholarship applications varied to some degree based on
university specic deadlines, however, in all cases it occurred after submitting an application to
participate { the time at which the baseline survey and information treatment was conducted {
and typically within a couple of days. Most programs provide students with information on their
scholarship award prior to requiring a down payment to conrm program participation, however, 2
programs also allowed for ongoing submission of a scholarship application up until the time of the
travel program itself in the case that a student's nancial situation changed.
The nal source of data is university and program administrative records. Most importantly,
from Oces of Financial Aid we obtain university distributions of Expected Family Contribution.
This data is used to provide treatment participants accurate information on their percentile of
nancial need according to Free Application for Federal Student Assistance. Detailed information
on EFC is provided in the next section. From travel program administrative records we also obtain
additional trip specic controls such as cost, trip length, whether involving international or domestic
travel, etc.
2.4 Experimental Design and Key Metrics
This RCT was designed to be executed with minimal disruption to the already existing structure
of the university travel program application and nancial assistance processes. The desire was
to be able to study the role of beliefs about relative nancial need in an entirely natural setting
with real nancial implications deriving from agents' behavioral choice to request a subsidy and to
11
report a particular willingness to pay within their aid application. There are several key elements
of the experimental design which are described in detail in this section, including: the Expected
Family Contribution measure, the elicitation of student priors about their comparative nancial
need, the elicitation of reported willingness to pay for program participation, and the information
intervention.
Expected Family Contribution: Expected Family Contribution (EFC) is a U.S. nationally stan-
dardized measure of a family's nancial well-being with respect to the cost of college tuition and
is calculated from information provided in the Free Application for Federal Student Assistance
(FAFSA). The literature studying perceptions of relative economic well-being most often utilizes
the income distribution as metric of comparison, however, there are several theoretical and prac-
tical reasons why EFC is most appropriate in this context. First, EFC is designed precisely as
a summary index capturing several dimensions of economic welfare. The key components used
to calculate EFC are identied by the U.S. oce of Federal Student Aid (2021) as: \taxed and
untaxed income, assets, and benets (such as unemployment or Social Security)...[also]...family size
and the number of members who will attend college or career school during the year."
3
While Ex-
pected Family Contribution may not capture every dimension of family nancial welfare, it remains
the standard summary measure used by both government and university oces of nancial aid.
Second, EFC has very simple inverse relationship to the concept of need for nancial assistance.
It is most natural to frame the elicitation of prior beliefs, as well as the information intervention,
in terms of one's percentile ranking of need for nancial assistance. We do so by simply inverting
the EFC percentile ranking to obtain the percentile ranking of need. One's percentile of need, as
used in this study, is then consistent with ranking given by the federal government's measure of
Federal Need{utilized by university oces of Financial Aid to allocate tuition aid{dened as Cost
of Attendance (COA) minus Expected Family Contribution.
4
Third, though individual students
may have limited information on parental income or assets, a student's EFC is easily accessible to
them. When FAFSA is led it must be registered to the student's social security number. Even if
3
Greater detail on the calculation of EFC can be found at: https://studentaid.gov/complete-aid-process/how-
calculated.
4
In reality Cost of Attendance may vary somewhat across students within the same university due to slight variation
in tuition and fees between schools (e.g., Arts and Sciences vs. Buisness) and whether a student lives on or o
campus. For our purposes, we simplify by abstracting away from this and considering percentile of \need" as 100 -
percentile of EFC.
12
a parent les FAFSA on behalf of their child, the student's FAFSA account, and specically their
EFC, can be retrieved online using the student SSN and birth date. Finally, and importantly, uni-
versities possess the EFC of all attending students who have led FAFSA, therefore, the university
specic distributions of this measure can be used to identify where a student lies within their own
university percentile ranking of nancial need according to EFC.
Prior Beliefs on Relative Need for Financial Assistance: A key precursor to the in-
formation intervention was collecting participant prior beliefs about their percentile of need for
nancial assistance in their university. This allows estimation of dierential treatment eects for
students who initially overestimate and underestimate their comparative need as captured by the
EFC measure.
In this portion of the baseline survey, students were rst asked to estimate their Expected Family
Contribution. Though EFC is relatively easy to lookup, we wanted to ensure obtaining a measure
of how students perceived their family nancial situation and to be able to dene a treatment eect
for over- and under- estimation in relation to EFC percentile, even if a study participant chose not
to take the time to nd their actual Expected Family Contribution. To make this elicitation simple
and without requiring students to look up any information it was framed as follows:
Expected Family Contribution (EFC) is a standardized measure from FAFSA which
roughly indicates the amount that a family should be able to contribute out of pocket
(apart from loans, grants, scholarships, etc.) toward college tuition, fees, and room and
board for a particular child in an academic year, without incurring signicant nancial
hardship. Where would you estimate your family falls in terms of EFC?
Students were asked to respond using a sliding scale from $0 to $200,000 and to indicate the
maximum if they believed their EFC would fall above.
5
Once an estimate of EFC was elicited, students were asked to provide their belief about their
percentile of need for nancial assistance compared to their university peers by responding to the
following on a discreet integer scale from 0 to 100:
Considering university funding for student programs (like the one to which you are
applying), please indicate on the scale where you believe you fall in the distribution of
5
Among participating universities, less than 1% of public and less than 5% of private school students who le FAFSA
have and EFC above $200,000. The xed scale form $0 to $200,000 was utilized to provide reference to a reasonable
range of values.
13
need of nancial assistance compared to other students in the university. As an example,
22 = \I believe I have greater nancial need than 22% of students in the university."
Finally, (on a new page of the survey) students were asked to look up their actual Expected Family
Contribution and given instructions on how to locate it in their email or on FAFSA.gov. As
anticipated, a number of respondents did not take the time to do so. To ensure that students
did not input incorrect information here, they were specically asked to leave this eld blank if
they had not looked up or couldn't nd their EFC, and, in a subsequent question, to indicate that
their actual EFC was reported or that it couldn't be found. Of the full sample of 1,029 baseline
respondents, 413 students (40%) indicate that they did not look up or could not nd their actual
EFC. These respondents were still included in the information intervention and asked to identify
their actual percentile in the university distribution of EFC based on their EFC estimate. However,
given that they did not supply this key piece of information and that their treatment eect and
ex-ante bias are dened in relation to their estimate rather than actual EFC, the primary results
presented in this paper focus on the sample of 616 students excluding these observations. All
results, reproduced for the full sample of 1,029, can be found in Appendix Section A.4.
The Information Intervention: The experimental treatment was an information intervention
which was implemented at the end of the baseline questionnaire. Participants who were assigned to
the treatment group advanced to a new survey page where they were presented with two graphics.
The rst was a university specic histogram of the EFC measure. An example is shown in Figure
A.4. The graphic was accompanied by the following text:
This image shows the distribution of EFC forfRespondent's Universityg Undergraduates
whose families have led FAFSA. Please take a moment to nd where you fall based on
your EFC (fRespondent's reported EFC Amountg). Remember that increasing EFC
indicates decreasing nancial need for aid.
Second, treated participants were shown a table listing the percentiles of need for aid with the corre-
sponding level of EFC. An example table can be viewed in Figure A.5. The table was accompanied
by the following text:
Here's a breakdown of the percentiles of EFC amongfRespondent's Universityg students
who have led FAFSA this year. Please take a moment to see where you fall in the
percentile ranking based on your actual EFC (fRespondent's reported EFC Amountg)
and indicate your percentile below.
14
Participants were also provided a brief explanation of how to interpret the percentile rankings,
e.g., that being at the 34th percentile implies that the student has greater need for aid than 34%
of students and less need than 66% of students, according to the EFC measure. In a follow-up
question, students were asked to indicate their actual percentile in the displayed distribution to
demonstrate that they had thoroughly viewed the information and updated their prior. Participants
assigned to the control group simply progressed to the conclusion of the questionnaire.
Treatment Assignment: Study participants were randomized into the information intervention
at the individual level. The baseline questionnaire automatically routed respondents to the treat-
ment or control branch with 50-50 probability. There was no stratication in the design. While
stratication could have been benecial for improved statistical power, it was not feasible in this
context since study enrollment was performed on a rolling basis as individuals opted to apply to
their university travel program and to participate in the study. Treatment was blind in the sense
that students were not aware that the information on EFC distribution was given out selectively.
Outcomes: We focus on the treatment eects of information on two related outcomes. The rst is
whether a program applicant completes a nancial assistance application. Second, if a scholarship
application is led, we obtain a student's self report of their maximum willingness/ability to pay
(WTP) for program participation. It is important to note that this reported WTP is not elicited
solely for purposes of the study, but as an input for aid determination by the programs. That is to
say, students will view their report as potentially having bite in terms of the amount of aid that
their program will ultimately award. Along with providing other nancial context they may deem
relevant and any other university specic information elicited, students are asked:
Considering both your own nancial situation and what your parents or guardians are
able to provide in support, please indicate the maximum amount that you can contribute
toward the program cost at this time (i.e., paying above this would be prohibitive for you
to participate).
Note that this measure is purposefully not elicited using an incentive compatible mechanism.
Though asymmetric information on actual individual WTP is precisely what creates the chal-
lenge for optimal allocation of aid resources, use of a common mechanism such as Becker-DeGroot-
Marshak (Becker et al., 1964) is not implementable as a policy tool, as it disproportionately excludes
15
individuals with low willingness to pay. This is what the nancial assistance policy seeks to avoid.
Instead, the idea here is to observe how a student's indication of what they can contribute may
respond to updated beliefs on their percentile of need.
2.5 Potential Mechanisms
While the primary eects of interest deriving from the information intervention are the dierential
response between students who initially over- and under-estimate need and the overall average
eect on requests for nancial assistance, at baseline we collect 6 measures which were pre-specied
to test for heterogeneity of treatment response. This section describes each of these measures.
First, we are interested in testing whether students who are more or less nancially constrained
respond dierently to information. For this we use a baseline measure of self reported personal
savings. Savings provides a dierent dimension of student nancial well-being compared to EFC,
which largely captures family socio-economic status.
The second and third metrics were motivated by a desire to test whether the a potential treat-
ment response of students with ex-ante bias could be associated with \altruistic" pro-social concern,
or possibly social image concern (i.e., desiring to not be perceived as opportunistic). As a mea-
sure of pro-social preference we utilize student reports of whether they are engaged in ongoing
volunteerism (44% of baseline respondents). To capture heterogeneity in social image concern at
baseline we ask study participants to respond to the psychometric Balanced Inventory of Desirable
Responding or BIDR scale. For the sake of brevity in the online survey we utilize the short form
developed by Hart et al. (2015). This tool, developed in psychology and calibrated on U.S. college
students, captures the propensity of individuals to adjust their response to sensitive questions in
settings where they are concerned about the perception of others. It was initially developed to
anticipate bias in survey response, however, it is generally understood to re
ect sensitivity of an
individual to social image concern. The scale consists of a series of questions about behaviors which
are either uncommon but socially lauded or highly common but considered socially undesirable.
E.g., respondents are asked to indicate on a scale from 1 (= I totally disagree) to 8 (= I totally
agree) how true the following statement is: I never cover up my mistakes.
The fourth metric is a scale used capture student perceptions about how well resourced they
16
believe their program to be. The rationale for studying heterogeneous response along this measure
is that over-estimators may be less likely to decrease their aid request in response to information
if they believe their action will have no real impact on the awards granted to other. If, however,
aid resources are perceived as scarce, students with a preference for fairness may adjust more
strongly. To capture heterogeneity in perceptions about program resources, students were asked:
On a scale of 1-10, how would you describe the availability of nancial aid resources to support
the participation of students, for the program to which you are applying? Choices were marked by
increasing perception of resources from 0 (=To participate a student must be able to cover costs out
of pocket) to 5 (=The program has moderate resources to provide aid to students with need) to 10
(=The program is very well resourced to provide as much aid as requested to any student.)
The nal two metrics are related to student perceptions about fairness in the distribution of
scholarship assistance. The rst is elicited to capture heterogeneity in student views about the
type of distribution that is fair. Respondents are shown Figure A.6 and asked to indicate which
distribution scenario they would consider most fair in the case that there is surplus budget to
be given out. We dene the variable Surplus Aid Should Go to Highest Need as an indicator for
whether a student responded (C), that an increasing proportion of funds should go to students with
the highest need (25% of baseline respondents). The nal metric seeks to capture diering student
views on whether surplus aid should be given out to participants at all or returned for some other
university purpose. Respondents are given a scenario and asked to allocate nancial aid among
students requesting scholarships, however, the amount available exceeds what applicants require to
be able to participate. Respondents must choose to allocate the excess aid to participants or indicate
that it should be directed to some other purpose as determined by university administration. We
dene Surplus Aid Should Be Returned to University as an indicator for whether a respondent
chooses the latter option (35% of baseline respondents).
2.6 Denition of Key Metrics
This section formally denes 3 key metrics that are utilized for analysis and required for interpre-
tation of empirical results.
Reported Willingness to Pay as a % of Student Fee: The elicitation of a student's willingness
17
and ability to pay for program participation is described in Section 2.4. Because travel programs
vary in cost both within and across universities, to standardize this measure we divide Reported
WTP by the student fee for the program to which an individual is applying. For students who
don't apply for aid, we impute a value of 1 (i.e., reports being able to 100% of the cost).
Reported WTP (% of Student Fee) =
8
>
>
<
>
>
:
Reported WTP
Student Fee
Student Applies for Aid
1 Student Doesn't Apply for Aid
Amount of Assistance Requested (as a % of Student Fee): Though the actual outcome
measure elicited in the scholarship application is a student's report of their willingness to pay,
for simplied interpretation of results, all treatmenent eect regressions utilize the continuous
outcome Amount of Financial Assistance Requested. This is dened simply as the student fee for
program participation minus a student's expressed willingness to pay to participate. The dierence
is eectively the amount a student is communicating in their scholarship application that they
would require to be able to participate. For students who do not apply for aid, this amount is
considered to be 0. Finally, so that this measure is comparable for varying program costs, we
convert it to a percentage of the student student fee.
Aid Requested (% of Student Fee) =
8
>
>
<
>
>
:
(Student FeeReported WTP)
Student Fee
Student Applies for Aid
0 Student Doesn't Apply for Aid
Bias in ex-ante belief on percentile of comparative need: Bias in student priors on com-
parative need are used to dene dierential treatment eects for over- versus under- estimating
ones comparative need and also to study how treatment response evolves marginally along distance
between one's prior and the truth. Bias, as used in this study, is dened as:
Bias =Prior Belief on Percentile of Need Actual Percentile of Need According to EFC
This denition implies that individuals with positive bias overestimate their relative need while
those with negative bias underestimate, however, when bias is utilized as a continuous measure we
take its absolute value to capture one's (positive) distance from the truth. 2 important variables
18
are derived from the bias measure. First, we dene the metric Overestimates Need (overest need)
as 1 if Bias is positive and 0 otherwise. Underestimates Need (underest need) is dened as 1 if
Overestimates Need = 0 and 0 otherwise. To allow for dierential treatment eects for low bias
students we also dene comparable measures, but dividing bias into 3 terciles. To dierentiate from
the binary split between positive and negative bias, these are denoted as underest need

(for the
rst tercile of bias) and overest need

(for the thrid tercile of bias).
3 Empirical Methods
This section presents the empirical methods used for analysis. All treatment eect specications
listed were pre-registered as part of the study pre-analysis plan, except in the case of a couple of
variations where the divergence is noted and explained below.
3.1 Predictors of Requested Financial Assistance
Before focusing on the experimental impact of the information intervention, I provide evidence
related to the baseline predictors of the key outcome variables { Applying for Aid and Amount of
Aid Requested (as % of student fee). This is done by simple, OLS regression of the outcomes on of
relevant covariates collected at baseline. I estimate:
y
i
= +
0
X
i
+
u
+
i
; (1)
wherey
i
is the outcome for studenti, X
i
is a vector of baseline covaraites, the
u
are university xed
eects, and
i
denotes the idiosyncratic error term. The coecient vector re
ects the marginal
relationships of the variables of interest, X
i
, conditional on the included covaraites and university
xed eects. Estimates from this specication are provided for dierent groups of predictors,
including nancial situation, demographics, beliefs and trip specic characteristics. Finally, to chose
the most relevant predictors, all four sets of variables are placed together in a LASSO regression
and those covariates which survive LASSO selection in a rst stage are utilized for the nal OLS
estimation. While these estimates cannot be understood as causal marginal eects, they provide
valuable context of the student characteristics related to requested aid.
In addition to results from Specication 1, I plot the inverse demand curve for program par-
19
ticipation derived from student reports of their WTP when applying for nancial aid. This curve
is generated by inverting the empirical cumulative distribution of Reported Willingness to Pay
(RWTP) in our sample. Specically, observations are ordered by increasing RWTP and, for each
unique value of RWTP in the sample, the share of observations with RWTP greater than the cur-
rent value is calculated. These shares are then plotted against their corresponding value of RWTP,
producing a graph of inverse demand. Since the value of RWTP is imputed as 1 for students will-
ing and able to pay the full student fee { i.e., those who don't apply for nancial assistance { the
demand curve is truncated at 1. This plot is also complemented with a histogram of the Requested
Aid as a percentage of the student fee conditional on having applied for scholarship aid.
3.2 Principal Treatment Eect Specications
Of primary interest is the overall impact of information on the behavioral response of students in
relationship to applying for aid and expressing a certain amount of aid as necessary to make par-
ticipation feasible. However, if students care about their relative position in the EFC distribution,
the nature of the information treatment is characteristically dierent for students who initially
under- versus overestimate their percentile of comparative need. For this reason, we rst estimate
specications which allow us to study the dierential response of students along the magnitude
and the direction of bias in ex-ante beliefs. In doing so, we attempt to derive some causal under-
standing about the eect of bias, per se, i.e., as dierentiated from a treatment response owing
to one's absolute position in the EFC distribution, or to some other student characteristics. This
is of interest in that a causal eect owing to the updating of biased beliefs on one's comparative
position re
ects something about students' level of pro-social concern when requesting aid: peer
comparison matters in how students engage with the nancial assistance policy. It's important to
dierentiate this eect from a shift, e.g., in requested aid among all low EFC students, regardless
of bias, which could be attributed, perhaps, to a signaling eect where students, now aware that
the program possesses this objective measure of need, feel justied to request more at the low end
of the distribution. A dierential response along bias, alternatively, suggests a mechanism driven
by the updating of beliefs{it matters that one's perceived rank was incorrect.
After exploring the role of bias, we estimate the average treatment eect (ATE) { with no
20
interaction with bias { to understand the overall movement in aid requests. This captures the net
eect of information provision. Each of the following specications is estimated for both outcomes:
(1) the propensity to apply for aid and (2) the amount of aid requested as a percentage of the
student fee (dened formally in Section 2.6).
To study the impact of accurate information along the continuous dimension of ex-ante bias
in prior beliefs on need, we estimate regressions of the following form, via OLS, allowing for a
dierential response in both level and marginal treatment eect by the direction of one's prior bias:
y
i
= +
1
T
i
+
2
T
i
overest need
i
+
3
T
i
jbias
i
junderest need
i
+
4
T
i
jbias
i
joverest need
i
+
1
jbias
i
j +
2
overest need
i
+
3
jbias
i
joverest need
i
+
X
cC
f
c
T
i
c
i
+
c
c
i
g +
0
X
i
+
u
+
i
:
(2)
Here, y
i
is the outcome for student i and T
i
is an indicator for treatment assignment (equal to 1
if a student receives the information intervention). overest need
i
and underest need
i
are dummy
variables equal to 1 when a student respectively overestimates or underestimates her percentile of
need andjbias
i
j is the (positive) distance between her ex-ante perceived percentile of need and
her actual percentile of need according to EFC. The
u
are university xed eects, X
i
represents
a vector of baseline control variables included to reduce residual variance and improve estimation
precision, and
i
denotes the idiosyncratic error term. Of primary interest are the coecients
3
and
4
which capture the marginal treatment eect of increasing bias for students who initially
under and overestimate their percentile of need.
In Specication 2 and all regressions where we test for treatment eects dierentiated by direc-
tion of bias, we present results both including and excluding the interaction of treatment assignment
with a set of control covariates (C). Since direction and magnitude of ex-ante bias are not ran-
domly assigned, they may be correlated with other characteristics which respond to the treatment
intervention. In particular, we observe empirically that low EFC (high need) students tend to un-
derestimate relative need while high EFC students overestimate need on average. Moreover, there
may be varying eects in response to the information at dierent quantiles of the EFC distribution,
21
independent of one's bias. The inclusion of treatment interacted with EFC quintile was pre-specied
for this regression, however, we also present estimates including additional control-treatment inter-
actions to check the stability of the marginal treatment eect estimates to additional controls. The
selection of these additional controls is describe below.
To test for the overall average eect of information provision among over- and under- estimators
of need, we estimate the following specication by OLS:
y
i
= +
1
T
i
+
2
T
i
overest need
i
+
1
overest need
i
+
X
cC
f
c
T
i
c
i
+
c
c
i
g +
0
X
i
+
u
+
i
;
(3)
where all regressors are dened as above. Ideally, we could additionally estimate treatment eects
at each decile of bias to capture non-linear treatment eect heterogeneity along bias, however, this
is not feasible given the large sample that would be required. Instead, we estimate the following
specication, splitting the sample into terciles of bias to at least allow a dierential response of
over- and under- estimators with sizeable bias compared to individuals with bias close to zero:
6
y
i
= +
1
T
i
+
2
T
i
underest need

i
+
3
T
i
overest need

i
+
1
underest need

i
+
2
overest need

i
+
X
cC
f
c
T
i
c
i
+
c
c
i
g +
0
X
i
+
u
+
i
:
(4)
Here, rather than being dened by a split at the median (i.e., bias = 0), underest need

i
and
overest need

i
are now indicators for the 1st and 3rd tercile of bias while the reference category
captured by T
i
is the middle tercile: students with low bias.
Finally, we are interested in the overall Average Treatment Eect (ATE) of the information.
6
It should be noted that Specication 4 was not included in the initial pre-analysis plan, however it follows in the
same spirit of Specication 3, only now allowing for a dierential eect for students with low bias. Bias in our
sample is symmetrically distributed around 0, therefore, a tercile split generates evenly proportioned groups where
the middle tercile re
ects low bias. Students in the middle tercile have bias ranging between -9 and +7 percentiles
from the truth.
22
The ATE is identied by OLS estimation of:
y
i
= +T
i
+
0
X
i
+
u
+
i
: (5)
There are three important notes regarding these main specications. First, though the experi-
mental design did not involve stratication by university due to the rolling nature of the applica-
tions, the inclusion of university xed eects is appropriate given the variation which exists across
universities in terms of program structure, student fees, the wealth distribution of students, etc.,
which will impact student reports of willingness to pay. Second, from the full set of possible control
variables collected at baseline, we use the LASSO double selection procedure proposed by Urminsky
et al. (2016) to select the relevant controls which will enter into the treatment eect regressions.
7
Finally, to eliminate researcher discretion in the determination of which covariates will be inter-
acted with treatment and included { along with Treatment x EFC Quintile { as additional controls
for Specications 2, 3, & 4, LASSO is utilized to select the \most relevant" correlates of bias from
the full set of baseline covariates. Then, for each correlate of bias, the two outcomes are regressed
on treatment and the interaction of treatment with the covariate. The correlates whose treatment
interactions demonstrate signicant impact for either outcome are included in the group of controls.
These covariates are Weekly Earnings (during the school year), Personal Savings, Ethnicity, and
being a First Generation student.
3.3 Specications for Treatment Heterogeneity and Distributional Eects
Six measures, collected at baseline, were pre-specied to be tested as potential mechanisms impact-
ing student response to bias. The measures and rationale are described in greater detail in Section
2.5. To test for dierential treatment response to over-estimating one's need along these measures,
7
From a large set of possible controls collected at baseline (X
0
i), separate LASSO regressions of yi on X
0
i and Ti on
X
0
i are used to select the most relevant predictors of yi and Ti in a rst stage. The union of relevant predictors are
then included as Xi for the estimation of each specication by OLS. In Specications 2 and 3, where inclusion of
controls such as overest andbias are required for proper identication of the treatment eect, we force the selection
of these variables in the rst stage.
23
we estimate the following specication, by OLS:
y
i
= +
1
T
i
+
2
T
i
Z
i
+
3
T
i
overest need

i
+
4
T
i
overest need

i
Z
i
+
1
Z
i
+
2
overest need

i
+
3
overest need

i
Z
i
+
X
cC
f
c
T
i
c
i
+
c
c
i
g +
0
X
i
+
u
+
i
:
(6)
Here, Z
i
is the measure of interest and overest need

i
once again re
ects over- estimators in the
third tercile of bias. Since, as we shall see, we observe no dierential treatment eect between under-
estimators and low bias students, to simplify Specication 6 we pool observations belonging to the
rst and second terciles of bias. Results from this specication using a binary split for positive and
negative bias are also reported in Appendix Section A.3. This fully interacted specication allows
us to test for the equivalence of response to the information intervention among over-estimators who
dier in whether they belong to groupZ or not. This test is given by
2
+
4
= 0. Likewise, the test

2
= 0 performs the same test for students in the low bias / under- estimator group. Sine both the
groupings by bias and the division along the variable Z
i
fully partition the sample, the treatment
eect estimates will both be consistent in the large sample approximation and unbiased for a nite
sample (see Nizalova and Murtazashvili, 2016; Athey and Imbens, 2017; Bun and Harrison, 2019).
As supporting results on treatment response, I perform two exercises which were not pre-
specied, however, provide an valuable complement to the overall ATE estimates and dierential
eects for over- estimators. To study how the eect of over-estimating one's relative need diers
across the levels of EFC, I estimate the following specication allowing for the full interaction of
24
EFC tercile with overestimating or underestimating / having low bias:
8
y
i
= +
3
X
j=1

un lb
j
fT
i
EFC T (j)
i
underest low bias

i
g
+
3
X
j=1

ov
j
fT
i
EFC T (j)
i
overest need

i
g +
0
X
i
+
u
+
i
:
(7)
In this regression,EFC T (j)
i
is a dummy variable indicating whether a student belongs to thejth
tercile of Expected Family Contribution and underest low bias

i
is a dummy for the pooled group
of under-estimators of need and low bias students.
9
The nal empirical exercise is aimed at understanding the distributional eects of the informa-
tion treatment. These estimates do not seek to make any causal claim regarding the interaction of
information with student characteristics, but rather to observe the overall distributional movement
resulting from information. To this end, we are interested in how nancial aid applications and
overall amounts requested shifted along two observable dimensions of student nancial situation:
Expected Family Contribution and Personal Savings. By OLS, I estimate
y
i
= +
3
X
j=1
f
j
T
i
EFC T (j)
i
g +
0
X
i
+
u
+
i
: (8)
This specication is also separately estimated using the terciles of personal savings. Finally, to
capture greater nuance over these two dimensions of student nancial need, we estimate a full
interaction of the terciles of Savings and a binary split of EFC at the median:
10
y
i
= +
3
X
j=2
3
X
k=1

j
k
fT
i
EFC B(j)
i
Savings T (k)
i
g +
0
X
i
+
u
+
i
: (9)
8
Note that in this and the following specications, controls for the levels of groupings (e.g., dummies for EFC terciles
x Overestimates uninteracted with treatment) are included, as required, however to simplify notation here, we allow
them to be subsumed into Xi.
9
I also estimate a version of this specication where under estimators are separated from low bias students, however
this reduces power and the results are qualitatively the same so this is reserved for Appendix Section A.3.
10
Note that the full interaction of terciles of both EFC and Savings produces similar qualitative results, however I
present the binary split along EFC which has better power given the sample size. Utilizing terciles of EFC and a
binary split by savings also produces a qualitatively similar result.
25
NowEFC B(1)
i
is a dummy variable indicating a student has below median EFC whileEFC B(2)
i
indicates above median. By including the full set EFC - Savings interactions with treatment,
j
k
is
interpreted as the average treatment eect among students who have low or high EFC and are in
thekth tercile of savings. While the estimated treatment response cannot necessarily be understood
as causal (i.e., owing exclusively to level of EFC and Savings), the estimates of
j
k
are informative
for understanding the overall distributional shift in aid requests over these dimensions.
4 Summary Stats, Biased Beliefs & Predictors of Aid Requests
This section presents baseline summary statistics and balance of the experimental treatment as
well as two important pieces of empirical evidence which help to contextualize the treatment eect
results described in the following sections. These are (1) the distribution of bias in student prior
beliefs regarding their own percentile of relative need at baseline and (2) the baseline covariates
which predict student aid requests.
4.1 Balance and Baseline Summary Statistics
Table 1 reports baseline summary statistics for select variables. Since treatment eect results,
presented in the main text, focus on the subsample of students which excludes those who fail to
report their actual EFC in the baseline survey, this set of observations is re
ected in Table 1.
However, this and all tables of results for the sample including all survey respondents can be found
in Appendix Section A.4. The columns of Table 1 provide means and standard deviations for
(1) the full sample, (2) observations assigned to the control group, and (3) observations assigned
to treatment. Column 4 reports the dierence in means between treatment and control for each
variable and presents the p-value for the test of equality, conditional on university xed eects.
While select variables are displayed for brevity, the F-statistic, reported at the bottom, re
ects
the test of overall signicance when treatment assignment is regressed on the full set of baseline
covariates. A p-value of .43 suggests that treatment is well balanced across observable student
characteristics.
Of particular importance, we observe balance in both actual Expected Family Contribution
and estimated EFC as well as other economic and demographic variables which may in
uence
26
perceptions of need ranking: personal savings, weekly earnings, ethnic minority and rst generation
status. We also observe balance on the measures of the perceived and actual percentiles of a
student's need according to their subjective report and their university distribution of EFC. This
is important as these measures jointly dene ex-ante bias in perceived percentile rank which is
utilized in treatment eect specications.
4.2 Biased Perceptions of Relative Need
In order for new information to induce a behavioral change through the updating of priors, ex-
ante beliefs about relative nancial need must be incorrect { at least for some students. Figure
1 plots the distribution of the baseline measure of bias in student prior beliefs regarding their
percentile of relative need based on Expected Family Contribution. Positive bias re
ects that a
student overestimates their position of relative need while negative bias indicates that a student
underestimates. The distribution of ex-ante bias in our sample is relatively symmetric, but is
skewed slightly toward under- estimation with a mean of -2.47 percentiles (s.d. 1.02).
11
That
the distribution of bias is close to zero centered is not a given. Other studies documenting biased
perceptions of relative income show mixed results depending on context and comparison group. E.g.,
Cruces et al. (2013) nd over and under- estimation of income percentile to be roughly balanced in
Argentina, while Nair (2018) nds that a representative sample of Americans underestimate their
global income position on average by 27 percentiles.
The symmetry of the distribution in Figure 1 masks important heterogeneity which exists in
terms of direction of bias at dierent locations in the distribution of actual percentile of EFC. Figure
2 plots the average perceived percentile of need within each decile of actual need. The shaded 45
degree band indicates the area where average perceived percentile falls within the correct decile.
While this appears to be true for students in the 3rd through 5th deciles of actual need, we observe
that high need students tend to underestimate their comparative rank while low need students tend
to overestimate. This may be in part mechanical { students at a very high (low) percentile have
less room to over- (under-) estimate { however, Cruces et al. (2013) shows that the tendency to
be biased towards the 50th percentile can in part be explained by homophily. Because individuals
tend to associate with others like themselves, one's reference group for estimating rank of economic
11
A two tailed t-test that the population mean bias is equal to zero can be rejected with a p-value of .016.
27
well-being may be skewed, and, in particular, one may perceive him or herself to be more average
than is, in fact, the case.
Regardless of the origin of the pattern observed in Figure 2, it is theoretically suggestive of
the potential for both eciency and/or distributional gains for programs in terms of requested aid
when students are provided accurate information on their percentile rank.
4.3 Predictors of Requested Aid and Demand for Program Participation
Before turning to the treatment eects of information provision on the propensity to apply for
aid and on overall amount of aid requested, it is informative to explore evidence on the baseline
characteristics associated with student aid requests and the shape of the demand curve for program
participation as re
ected by scholarship applicant reports of their willingness to pay. Since willing-
ness to pay is not elicited via an incentive compatible mechanism, estimates of baseline predictors
of reported WTP must be interpreted with the understanding that this measure re
ects a strategic
decision by applicants which balances nancial incentives with other possible considerations such as
fairness, concern for not being perceived as opportunistic, etc. Nonetheless, student socio-economic
situation strongly predicts reported willingness to pay, as would be expected of actual WTP. Tables
2 & 3 present estimates from Specication 1 where the outcomes of Applying for Aid and Amount
of Aid Requested as a percentage of the program fee are regressed on baseline covariates, control-
ling for university xed eects. Recall that Amount of Aid Requested is dened in terms of the
student fee minus expressed willingness to pay. Since the results of Tables 2 & 3 are non-causal
and are informative of the population as a whole, these tables re
ect the full population of baseline
respondents { i.e., including also students who le FAFSA but don't report their actual EFC. The
results are comparable for the subset of students which are the focus of the experimental results
and can be found in Appendix Section A.2.
Columns 1 - 4 re
ect dierent groupings of predictor variables: (1) nancial indicators, (2) de-
mographic information, (3) beliefs expressed at baseline, and (4) program specic characteristics.
In Column 5 all variables are included together in a LASSO regression, then, the predictors selected
as most relevant in the rst stage are used to re-estimate the marginal eects by OLS. Without
placing too much emphasis on the precision of coecients, given the non-causal nature of these es-
28
timates, it is worth making a few observations. First, scholarship requests follow expected patterns
in relationship to the measures of nancial well-being. Higher savings and higher EFC students
request less aid.
12
Students who don't le FAFSA (typically low need), as well as those who do
le, but don't report their EFC in the baseline survey tend to request less aid than the students
who report their actual EFC at baseline. Second, after controlling for both personal and familial
nancial characteristics{savings, weekly earnings, EFC, etc. { status as an ethnic minority or rst
generation student remain strongly predictive of requested aid. Third, student ex-ante perceptions
of their percentile of need are also strongly predictive of whether they apply for aid and how much
is expresses as being required to participate. Finally, the set of six measures which were collected
to study treatment eect heterogeneity are included as potential predictors. Column 5 estimates
suggest that only personal savings and perceptions on program resources and on who merits receiv-
ing surplus aid are directly associated with requested aid. Specically, controlling for university
xed eects and other predictors, students who perceive their program to be better resourced tend
to request more aid, as do students holding the view that surplus aid budget should be directed
largely to the highest need students.
13
Figure 3 plots the inverse demand curve for program participation among students applying for
aid for both the full sample of students and the subsample which excludes those who don't report
their actual EFC. Again, this does not re
ect incentive compatible elicitation of WTP, but rather
the reported WTP of program applicants. Notably, less than 2% of students express being able to
contribute nothing, while roughly 70% indicate being able to cover the full cost { either revealed
by not applying for a scholarship or by expressing an ability to pay the full fee when asked in the
scholarship application. The curve is relatively
at at the high and low end indicating few students
who report an ability to contribute close to (but not exactly) zero or in the range of 90-99% of cost.
Between these extremes we observe almost linear increase in expressed willingness to pay. This is
re
ected also in the nearly uniform distribution of requested aid conditional on applying for aid,
which is plotted in Figure 6. In Figure 6 it becomes apparent that there is a mass of students who,
12
Note, weekly earnings may be insignicant due to the mixed nature of work as a college student. Students who
work more and have higher earnings often do so due to higher economic need, thus weekly earnings may not follow
what would typically be expected in terms of an income eect of aid requests.
13
Interacting the variable Surplus Aid Should go to the Highest Need with Perceived Percentile of Need in this
specication demonstrates that the impact of the former is driven largely by students who hold this view and
perceive themselves to be of high need.
29
despite applying for aid, express that they can cover the full program cost. These are students with
an actual willingness and ability to pay above the requested student fee { and who honestly indicate
such { but who, given the availability of aid resources, consider it worth requesting support. This
type of aid applicant helps explain some of what is observed in the treatment eects results, namely,
aid applications may adjust without observing a large shift on the extensive margin of total support
requested.
5 The Eect of Information on Requested Financial Assistance
This section presents the main ndings of this study: the behavioral response to accurate informa-
tion on the distribution of university EFC among students who initially over vs under- estimated
their position of comparative need and the the overall average treatment eect (ATE) of information
provision on the propensity to apply for aid and on the proportion of aid requested.
I begin by presenting evidence that over and under- estimators respond dierently to the infor-
mation intervention. Table 4 displays results from Specication 2, where the continuous measure
of the magnitude of a student's ex-ante bias in perceived percentile of need is interacted with indi-
cators for both treatment assignment and whether a student's bias is positive (overestimates need)
or negative (underestimates). Including both Underestimates x Bias and Overestimates x Bias in
the specication allows for a direct comparison of these coecients, which should be interpreted
respectively as the marginal change in treatment eect along magnitude of bias for under and over-
estimators of need. Columns (1) - (3) are for the outcome Applied for Financial Aid, (4) - (6) for
the Amount of Aid Requested as a % of the Student Fee. Columns (1) & (4) present estimates from
the unconditional regression, while (2) & (5) condition on treatment interacted with EFC quintile
{ as pre-specied { to address potential confounding of eects due to correlation between bias and
quantile of EFC. Finally, Columns (3) & (6) include additional controls of treatment interacted with
baseline covariates correlated with bias. Since the magnitude of bias is not randomly assigned, this
is performed as an additional check on the stability of our estimates of the marginal eect along
bias. Note that the inclusion of dierent treatment interactions as controls makes the coecients
on Treated not comparable across columns as it represents dierent reference groups, however, the
dierential eect for over-estimators and marginal eects of bias can be compared.
30
Table 4 indicates that there is little to no impact of increasing bias on either outcome for students
who initially underestimate their relative need. For those students who initially overestimate their
comparative need, however, we observe a signicant marginal reduction in requested aid as the
magnitude of bias increases. That is, students who ex-ante consider themselves to be of higher
comparative need than they are, when informed otherwise, decrease the amount of assistance they
request (relative to those with little to no bias) and in increasing proportion to how far their initial
perception was from the truth. This is observed both in a decreased likelihood of applying for aid
and a decrease in the overall amount of aid requested from the program. The estimates of this
eect appear relatively stable and robust to the inclusion of controls of both EFC quintile and other
correlates of bias interacted with treatment.
The lack of a marginal treatment eect for under- estimators does not indicate that students with
negative bias have no response to the information intervention, only that their response does not
appear to depend on how far their prior perception lies from the truth. To consider how aid requests
vary, on average, among over- and under- estimators, we estimate Specications 3 & 4. These results
are presented in Tables 5 & 6. Here we are interested in testing for dierential average treatment
eects among over- and under- estimators of relative need. Once again, estimates are provided
for both outcomes and columns condition on dierent treatment interactions as described above.
Estimates in Table 5 re
ect a binary split between students with positive and negative bias. Once
again, due to conditioning, the coecients for Treated are not comparable across columns, however,
the unconditional regression for the amount of aid requested (Column 4) suggests a signicant
increase of 5 percentage points for students who initially underestimate their comparative need.
Since unconditional, this estimate indicates that, in response to information, among all students
who underestimate need, there is an average increase in the amount of aid communicated as required
for participation. While the dierential eect for over-estimators is not precisely estimated here,
the sign for both outcomes suggests that over- estimators diminish their assistance request relative
to under- estimators as expected.
In part, the dierential eect for over- estimators may be muted here by the fact that the
binary sample division in Table 5 includes many students with bias close to 0. Since over- versus
under- estimating one's comparative need by a few percentiles may eectively make little dierence
in terms of treatment response, we re-estimate these results with the sample divided into terciles
31
of bias, allowing for a separate response among low bias students. For each outcome, Figures
5 and 6 plot the unconditional treatment eects estimated at each tercile of bias. These gures
complement the regression results found in Table 6 and help to visualize how students are responding
to information on comparative need. The thick and thin spikes attached to the point estimates
represent the 90 and 95 % condence intervals. On the extensive margin of applying for aid
(Figure 5), we observe a signicant increase in assistance requests from students who discover that
they've underestimated their comparative need, a null eect for low bias students, and a possibly
negative{though under-powered{estimated eect for over- estimators. Moreover, we can reject the
equivalence of treatment response for under- and over- estimators, as indicated by the p-value of
.048. On the continuous outcome (Figure 6), we see a slightly dierent picture. We observe a
signicant increase in requested aid from both under-estimators and students with low bias. Thus,
while low-bias students see no average shift in propensity to apply for aid, those who do request aid
on average communicate that they require a higher percentage of support after being exposed to
the information intervention. Here, we can statistically reject the equivalence of treatment eects
between over- and under-estimators and also, marginally, between over-estimators and students
with low ex-ante bias.
So, students who do not overestimate their position of comparative need are motivated, on
average, to increase their nancial aid ask after being exposed to information on their university
distribution of aid, while we observe a null eect or slight decrease in requested aid from students
who initially over-estimate. These eects however should not be interpreted as causal estimates of
the in
uence of bias, but re
ect the average movement in these groupings. As noted, there may
be level eects in response to information at various quantiles of Expected Family Contribution or
also due to other factors correlated with bias such as minority status. The unconditional eects
depicted graphically in these gures are presented in Columns (1) & (4) of Table 6. To explore
the robustness of the dierential treatment eects between under-estimators, low bias students,
and over-estimators, Columns (2) & (5) condition on treatment interacted with EFC Quintile.
Columns (3) & (6) then add the additional treatment interactions with baseline correlates of bias.
Including these controls, nothing substantive changes in terms of estimated dierential eects for
over-estimators. If anything, the relative decrease in requested aid due to overestimating need
appears to become stronger when fully controlling for baseline correlates of bias. The estimated
32
eect for under-estimators remains insignicant{though the point estimate increases{conrming
that, on average, under- estimating versus having low bias makes little dierence in terms of how
a student responds to the information provided.
Finally, we turn to the overall Average Treatment Eect (ATE) of information provision. While
the estimates presented in Table 7 are less informative about the nature of student response to
information{as they average over the varying eects for students with positive and negative biases{
they are important for understanding the overall shift in requested aid owing to the intervention.
This is largely relevant from the vantage point of eciency{i.e., was more or less aid requested
overall?{and for evaluating any trade-o between eciency and distributional considerations which
will be discussed in Section 7. We have seen that both under-estimators and low bias students
increase their aid requests, which may partially be oset by students who overestimate, however this
suggests a net increase in requested aid. Since the ATEs reported in Table 7 are not interacted with
bias or other heterogeneous student characteristics, they re
ect the overall causal eect of provision
of accurate information on university distribution of EFC and student ranking of need. Columns
(1) & (3) report the unconditional ATEs while (2) & (4) report conditional ATEs (conditioned on
baseline controls selected by LASSO double selection).
The ATE estimates are statistically insignicant, however the sign of estimates suggests an
overall increase in nancial assistance requested. Table 7 reports the 90% condence intervals for
the estimates, which oer an indication of uncertainty in the overall causal eect as well as an
upper bound on the potential increase in requested aid. With an average student fee of $365 and a
control group mean of 17% of the student fee requested in aid, a back-of-the-envelope calculation
utilizing the point estimate from Column (4) suggests that providing information resulted in an
average increase of $10.11 of requested assistance, while the upper bound of the corresponding
condence interval suggests it could have been as high as $21.90 per student. This amount does not
necessarily translate to an exact cost to the programs of providing this intervention. E.g., students
are not guaranteed to receive the amount they express they need and it remains ultimately up to
the program to determine an aid award deemed reasonable and feasible given budget constraints.
However, if programs take at face value student reports of their willingness to pay and provide
aid accordingly, ATE estimates suggest there would be a net loss of revenue from student fees.
Moreover, the implication of these estimates is that in response to information, students who don't
33
overestimate their comparative need either see a real reduction in willingness to pay to participate
in the program or feel justied expressing to the program that they require more assistance, while
overestimating one's comparative need \undoes" this eect on average.
6 Heterogeneous Response to Information on Comparative Need
Observing that students who overestimate their rank in comparative need respond dierently to
the information provision than those who underestimate or have low bias in their prior belief, we
explore heterogeneity of this response along the 6 dimensions which were pre-specied prior to data
collection. Since we cannot distinguish between treatment eects for under- estimators and low
bias students on the extensive margin, we pool these observations to simplify the analysis and the
interpretation of results. Table 8 reports coecients and statistical tests from 6 separate estimations
of Specication 6, all for the same outcome: Amount of Aid Requested.
14
The coecients interacted
with Overestimates re
ect dierential eects for over-estimators relative to students who either
have low bias or underestimate their need. Each column represents a separate regression with
a dierent heterogeneous variable interaction. The upper panel reports coecient estimates and
standard errors while the lower panel reports the estimated dierential treatment eects along
with p-value for their statistical tests. First (I), we test for the equivalence of treatment response
for students both belonging to the indicated group (Z=1), but diering in whether or not they
overestimate their comparative need. Second (II), and most importantly, we test for equivalent
response between two students, both of whom overestimate their need, but who dier in whether
they belong to the group dened by the heterogeneous variable. All results are conditional on
treatment interacted with EFC quintile and other baseline controls.
The only measure for which we can reject the equivalence of eects among over- estimators
(Test II) is found in Column 1. This measure captures whether, at baseline, a student indicated a
belief that surplus nancial aid budget should be given largely to students with the highest need
(Z=1) versus being distributed among all students (Z=0). We estimate a 20 percentage point drop
in requested aid for over-estimators who express this view relative to over-estimators who do not.
The coecient Treated x Z provides a similar test within the low bias/underestimating group for
14
Analogous results on the outcome of Applied for Aid are presented in Appendix Section A.3. Extensive margin
eects are largely insignicant and less precisely estimated.
34
students with Z=1 versus Z=0. For this group, the only signicant coecient is found in Column
6 where we observe a 13 percentage point increase in requested aid among under-estimators and
low bias students with low savings relative to their counterparts with high savings. This suggests
that the average increase in requested aid observed among students who don't overestimate their
comparative need is driven especially by students with low personal savings.
Since the six measures are potentially correlated with one another, I re-estimate Specication
6, now including all 6 heterogeneous variables interacted with dummies for treatment assignment
and overestimating relative need. The results from this single regression are presented in Table
9. While these measures may still be correlated with other omitted characteristics contributing
to the treatment response of over and under-estimating students, conditioning on the full set of
possible pre-specied mechanisms provides additional evidence of the robustness of dierential
eects observed in Table 8. Now we observe signicant dierential eects among over-estimators
in Columns 1 & 2 and among low bias and underestimating students in Column 6. Taken together,
these results suggest that the overall increase in requested aid is driven by students with low
personal savings who do not substantially overestimate their position of need. Moreover, this
eect is diminished if a student initially overestimates his need, but especially, if he holds the view
that surplus aid should only go to the highest need students or that it should be returned to the
university rather than being distributed. The implication here is that the pro-social response of
over-estimators depends in part on views regarding what is fair in the aid distribution process.
Overall, no dierential impact is estimated for believing that the program is better or more poorly
resourced, or by volunteering { a proxy for pro-social concern { or by one's measure of social
impression management.
As a nal exercise to explore heterogeneity of treatment response by bias, we fully interact
overestimating need with the terciles of Expected Family Contribution. We have observed a dier-
ential eect for students who overestimate their need, however, the average eects by bias plotted
in Figures 5 and 6 mask heterogeneity of response which may exist at dierent quantiles of Ex-
pected Family Contribution. To study how students at dierent EFC levels respond dierently to
bias, Figures 7 and 8 plot estimates of Specication 7. The corresponding point estimates can be
found in Tables 10 and 11. Though these fully interacted estimates, at the current sample size, are
less powered than would be desired, I present them primarily to observe that there does appear
35
to be a dierent pattern of response across the distribution of EFC. In particular, in the low and
middle EFC ranges, we can conrm that nancial assistance requests increase (Figure 11) among
students who don't overestimate their need. Moreover, we can statistically reject the equivalence
of these eects and the null eect observed among high EFC who don't overestimate. At high EFC
there is a marginally signicant drop in requested aid among over-estimators though this cannot
be statistically distinguished from the imprecise zero eects observed for their counterparts in the
low and mid EFC ranges.
These estimates provide evidence on how level of need interacts with student bias in driving
the average positive eect observed among students with low bias or who underestimate their
need and the average negative (though imprecisely estimated) eect observed among students who
overestimate need. Specically, we learn here that the observed increase in requested aid is driven
by a level increase among all low and middle EFC students, and does not occur in the high EFC
range. This suggest that the eect owes to signaling or conrmation of being a high or medium
need student rather than having to do with bias, per se. Due to large standard errors, we cannot
conclude whether over-estimators in the low EFC ranges also responded similarly, though near zero
point estimates would suggest otherwise. We also learn that high EFC students who overestimate
decrease their aid requests on average, though again, we cannot conrm that low and middle EFC
students who overestimate behave dierently, due to imprecision. Thus, we don't learn much here
on the behavior of students who overestimate need and have low or medium EFC, but we can
conrm that students who overestimate and have high EFC are partially osetting the average
increase among lower EFC students.
7 Distributional Shifts In Requested Aid
In this nal empirical section, we step away from concern about identifying the dierential response
of students along their direction of bias. Instead, we ask, what was the overall distributional eect
on requested nancial assistance in response to information? To address this question, I rst
estimate Specication 8 and an equivalent specication where the terciles of EFC are replaced
with the terciles of savings. The coecients of these regressions are plotted in Figures 9 and
10. The estimates re
ect the average eect of the information intervention at each tercile of EFC
36
and Savings. The estimates are not \causal" in the sense of being able to attribute the observed
movement to one's position of EFC or savings, per se. However, to the extent that the sample
is representative of a population where we would like to replicate this type of intervention, the
estimates re
ect the actual shift in requested aid that could be expected. In this sense the estimates
are quite informative of the distributional shifts in demand for nancial assistance which result from
presenting students with accurate information on the distribution of EFC and their position in it.
The estimates depicted in Figures 9 and 10 are reported also in Table 12.
Along Expected Family Contribution, respectively for middle and high EFC students, there is
a 12 percentage point increase and an 8.5 percentage point decrease in the propensity to apply for
aid. Lack of movement in the low EFC group can be explained if these students tend not to be
\marginal" on the extensive margin. E.g., low EFC students already apply for assistance in higher
proportion. The control group mean of Applied for Aid for low EFC students is .51 relative to
.32 and .19, respectively, among middle and high EFC students. Thus, on the extensive margin, a
majority of the increase in aid applications (among middle EFC students) is oset by the decrease
in the high tercile. However, the intensive margin paints a slightly dierent picture. The fact
that less of a decrease is observed in the high EFC tercile in terms of Amount of Aid Requested
suggests that those students who opt not to apply for aid in response to information would have
reported a relatively high willingness to pay in their scholarship application. Moreover, while more
students are not persuaded to submit a scholarship application in the low tercile, those that do
request more aid on average after observing the EFC distribution. We estimate an 7.6 and 5.6
percentage point increase in the amount of aid requested, respectively, for low and middle EFC
students. This is oset little or none by information response at the high end of EFC. The estimates
of average treatment eects by savings terciles suggest there is little to no average movement on the
extensive margin. However, on the intensive margin we observe a 12 percentage point increase in
Aid Requested among low savings students, which is not oset by any substantial decrease among
medium or high savings students.
Finally, to gain a full picture of the movement of aid requested along these two dimension
of student nancial situation, Figures 11 & 12 plot estimates of Specication 9 allowing for the
full interaction of EFC, personal savings, and treatment. While somewhat under-powered on the
extensive margin, we observe a marginally signicant increase in applying for a scholarship among
37
low EFC, high savings. On the intensive margin we estimate a 17.5 percentage point increase
in requested support among the low EFC, low savings students. The other groupings have point
estimates relatively close to zero (though tighter condence bands would be desirable). Thus, while
there may heterogeneous response which is roughly averaged out within the other EFC-Savings
groups, only the low EFC, low savings grouping demonstrates any strong average response to the
information provision. Thus, the overall increase in requested aid is coming from students with
high need in both dimensions of EFC and individual savings, with little adjustment (on average)
elsewhere along the two dimensions of need.
8 Conclusions
To consider the overall impact of this intervention and it usefulness as a policy tool it is helpful to
reiterate the overall picture painted by the empirical results which have been presented. First, we
observe that students who initially overestimate their position of need respond dierently to infor-
mation than those who have low bias in their prior beliefs or who underestimate their need. This is
observed both by dierential average eects of treatment when students are grouped according to
their bias and in an increasing marginal eect along bias for students who overestimate need. These
results are robust to the inclusion of interactions of treatment assignment with baseline correlates
of bias, supporting a causal interpretation: recognizing that one has lower comparative need than
believed causes a reduction in requested aid relative to peers who don't overestimate need.
However, while this dierential \overestimation eect" appears to result in a decrease in re-
quested aid at high levels of EFC (low need), in the middle and low EFC (higher need) ranges it
is re
ected in a null treatment eect for overestimators (though imprecisely estimated), whereas
students with low bias or who underestimate need actually increase their requested aid on average.
Since students with low bias and those who underestimate need respond similarly to the informa-
tion intervention, the increase in requested aid, observed at middle and low EFC, should not be
attributed to an eect owing to the updating of ex-ante mis-perceptions. Were this the case, we
should observe an increasing treatment eect along bias for underestimators. Instead, we observe
an overall increase at middle and low EFC, which appears negated among those students who are
informed that they've overestimated their need.
38
The mechanisms leading to this increase in aid requests are not evident or discernible from the
current research design. Students who discover they were roughly correct or underestimated their
relative position may feel more bold in expressing a lower ability to pay if they were originally quite
uncertain about how they compared to others. Alternatively, observing the EFC distribution could
signal to students that the program possesses this \objective" measure of need which justies a
higher request in the lower half of the EFC distribution. What is evident is that the increase in
assistance requests is driven largely by students with low personal savings, implying it is mediated
in part by students being more nancially constrained.
Finally, there is evidence that the dierential response of over-estimators is correlated with
student perceptions about fairness in the aid distribution process. We observe a larger decrease
in requested aid among students who overestimate their need and hold the view that nancial
assistance should be directed only to the highest need students or that surplus aid budget should
be directed to another university purpose rather than being given out to student participants. This
evidence supports the conclusion that the response of over-estimators re
ects \other regarding
preference," but also that heterogeneity in views regarding fairness is an important determinant of
how individuals weigh pro-social concern against monetary payo in their behavioral choice.
This study was initiated, in part, as an empirical inquiry into the net impact of information
provision on both eciency and distributional equity in a setting where individuals seek subsidy
benets. It is clear that providing students with accurate information on the distribution of EFC
did little to improve potential eciency, and, more likely, resulted in an overall increase in requested
aid. On the Amount of Aid Requested we estimate a 2.8 percentage point increase over the control
mean of 17% of the student fee, however, the 90% condence interval suggests a lower bound of a
1pp decrease in requested aid or an upper bound of a 6pp increase. That said, the evidence suggests
the increase is largely coming from students who have both low EFC and low personal savings, i.e.,
who are objectively of highest need by measure of both familial and individual nancial situation.
Finally, results suggest possibly modest gains in terms of a distributional shift toward lower EFC
students: increased aid requests at the lower end of the EFC distribution are partially oset by a
decrease at high EFC among students who initially overestimate their need ranking.
Overall, it is unclear that the benets of providing accurate information on comparative need
to students before they initiate a nancial assistance request outweigh the potential costs. Though
39
there is evidence of some pro-social adjustment among lower need students who overestimate their
comparative position, the predominant eect is an increase in requested subsidization among stu-
dents with higher need who receive a conrmation of their position. Since an increase in requested
aid most likely re
ects that individuals feel more justied making a larger ask when their position
of high need is conrmed, rather than a real drop in willingness or ability to contribute, the ob-
served treatment eects are possibly best understood in terms of information gained or lost through
a student's signal about their willingness to pay. The policy usefulness of providing this type of
information to individuals will, therefore, depend largely on programmatic objectives and also the
level of sophistication in subsidy allocation rules. Institutions having a strong aim to subsidize the
highest need individuals and not being tightly constrained in terms of resources might consider
the provision of information on comparative need a useful tool. Institutions with more limited
resources, however, may consider the potential reduction of information on actual willingness to
contribute among higher need individuals to be too costly to justify the minor distributional gains
that could be achieved.
40
Table 1: Baseline Summary Statistics and Balance
(Subsample Excluding FAFSA Filers who Don't Lookup Actual EFC at Baseline)
Means and Standard Deviations Dierence
Variable Full Sample Control Treatment Treat vs Cont.
(1) (2) (3) (4)
Total Savings 2,468.581 2,372.410 2,573.228 155.925
(3,706.623) (3,454.523) (3,966.066) f0.601g
Weekly Earnings 105.358 102.190 108.805 4.803
(238.448) (221.699) (255.766) f0.806g
Expected Family Contribution 14,067.063 13,318.533 14,847.615 904.987
(30,368.561) (28,260.033) (32,479.914) f0.751g
Estimated EFC 35,305.676 35,044.754 35,589.594 2,103.238
(51,476.492) (52,190.098) (50,775.664) f0.599g
No FAFSA: Low Need 0.214 0.224 0.203 -0.010
(0.411) (0.418) (0.403) f0.771g
No FAFSA: Ineligible/Didn't Know 0.083 0.078 0.088 0.016
(0.276) (0.268) (0.284) f0.483g
Unsure if Filed FAFSA 0.083 0.090 0.075 -0.015
(0.276) (0.287) (0.263) f0.496g
Yes FAFSA: With Guardians 0.558 0.539 0.580 0.025
(0.497) (0.499) (0.494) f0.531g
Yes FAFSA: Independent 0.062 0.069 0.054 -0.016
(0.241) (0.253) (0.227) f0.428g
Actual Percentile of Need 49.388 49.835 48.902 -1.627
(29.357) (29.004) (29.778) f0.491g
Perceived Percentile of Need 46.917 45.804 48.129 2.215
(30.438) (30.333) (30.557) f0.358g
Program Student Fee 365.114 357.290 373.627 -0.504
(381.670) (373.190) (391.146) f0.980g
Female Student 0.573 0.561 0.586 0.036
(0.495) (0.497) (0.493) f0.208g
Male Student 0.153 0.171 0.132 -0.034
(0.360) (0.377) (0.339) f0.235g
Ethnicity: White 0.529 0.511 0.549 0.018
(0.500) (0.501) (0.498) f0.638g
Ethnicity: Black / African American 0.089 0.097 0.081 -0.018
(0.285) (0.296) (0.274) f0.407g
Ethnicity: Other (Non-White) 0.237 0.246 0.227 0.006
(0.426) (0.431) (0.420) f0.864g
Ethnicity: Hispanic / Latino 0.144 0.146 0.142 -0.005
(0.352) (0.354) (0.350) f0.852g
First Generation College Student 0.323 0.321 0.325 0.003
(0.468) (0.468) (0.469) f0.926g
GPA 3.518 3.485 3.554 0.084*
(0.440) (0.471) (0.403) f0.081g
Test of Overall Signicance F(41,615) 1.02
(including all (26) controls) p-value f 0.43 g
Observations 616 321 295 616
This table presents sample means and standard deviations (in parentheses) of variables at baseline for the full sample
(1) and for the control (2) and treatment (3) groups. Column (4) reports coecients from the regression of the control
variable on a treatment indicator and university xed eects. P-values for the coecients are indicated in brackets
based on heteroskedasticity robust standard errors. Stars indicate a statistically signicant coecient at .1 (*), .05
(**), and .01 (***) levels of signicance. The F-statistic reported is for the overall test of joint signicance of all
controls in a regression of the treatment dummy on all control variables. Subsample excludes FAFSA lers who don't
lookup actual EFC at baseline.
41
Figure 1
Figure 2
42
Table 2: Predictors of Applied for Aid
(1) (2) (3) (4) (5)
Total Savings (Thousands) -0.0162

-0.00645

(0.00340) (0.00332)
Weekly Earnings (Hundreds) 0.00859 0.00673
(0.00577) (0.00562)
EFC (Tens of Thousands) -0.0195

-0.00711

(0.00432) (0.00269)
No FAFSA -0.400

-0.165

(0.0395) (0.0451)
Filed FAFSA as an Independent Adult 0.0558 0.00214
(0.0616) (0.0621)
Filed FAFSA but Doesn't Report EFC -0.199

-0.0769

(0.0323) (0.0342)
Female Student 0.00560
(0.0411)
Ethnicity: Black / African American 0.342

0.271

(0.0567) (0.0523)
Ethnicity: Hispanic / Latino 0.0977

0.0495
(0.0476) (0.0431)
Ethnicity: Other (Non-White) 0.146

0.129

(0.0364) (0.0354)
First Generation College Student 0.164

0.0397
(0.0340) (0.0352)
Mostly Associates with Lower SES Peers 0.130

0.101

(0.0515) (0.0495)
Grade Point Average -0.0862

(0.0465)
BIDR Impression Management Scale 0.00212 0.000573
(0.00168) (0.00158)
Estimated EFC (Tens of Thousands) -0.0121

-0.0102

(0.00268) (0.00277)
Perceived Percentile of Need 0.00542

0.00333

(0.000511) (0.000617)
Perceived Program Resources Scale 0.0154

0.0156

(0.00596) (0.00577)
Surplus Aid Should Go to Highest Need -0.0115 -0.0180
(0.0279) (0.0270)
Surplus Aid Should Be Returned to Univ 0.00241
(0.0266)
Student Cost (Hundreds) 0.00406 0.00838
(0.00694) (0.00530)
International Program 0.219

0.132
(0.106) (0.0842)
Flying to Destination -0.0193
(0.0808)
LASSO Selected Covariates No No No No Yes
Control Mean 0.34 0.34 0.34 0.34 0.34
Observations (N) 1029 1029 1029 1029 1029
Robust Standard Errors are reported in parentheses. Regressions are conditional on University Fixed Eects. The
omitted category for FAFSA groupings is Filed FAFSA with Guardians (and reported actual EFC). The omitted
category for Ethnicity is White. The omitted category for Mostly Associates with Lower SES is Associates with All
SES. Mostly Associates with Middle and Upper SES are both included, but are not displayed (both are insignicant).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
43
Table 3: Predictors of Amount of Aid Requested (as % of Student Fee)
(1) (2) (3) (4) (5)
Total Savings (Thousands) -0.0105

-0.00313

(0.00184) (0.00158)
Weekly Earnings (Hundreds) 0.000686
(0.00435)
EFC (Tens of Thousands) -0.0140

-0.00517

(0.00334) (0.00177)
No FAFSA -0.241

-0.0703

(0.0219) (0.0231)
Filed FAFSA as an Independent Adult 0.0896

0.0451
(0.0450) (0.0430)
Filed FAFSA but Doesn't Report EFC -0.141

-0.0556

(0.0203) (0.0198)
Female Student -0.0124
(0.0254)
Ethnicity: Black / African American 0.192

0.140

(0.0397) (0.0354)
Ethnicity: Hispanic / Latino 0.0938

0.0603

(0.0322) (0.0286)
Ethnicity: Other (Non-White) 0.0807

0.0677

(0.0226) (0.0215)
First Generation College Student 0.125

0.0418

(0.0224) (0.0211)
Mostly Associates with Lower SES Peers 0.0818

0.0495
(0.0376) (0.0345)
Grade Point Average -0.0483
(0.0300)
BIDR Impression Management Scale 0.00207

0.00102
(0.00102) (0.000918)
Estimated EFC (Tens of Thousands) -0.00637

-0.00491

(0.00133) (0.00138)
Perceived Percentile of Need 0.00391

0.00258

(0.000334) (0.000389)
Perceived Program Resources Scale 0.00867

0.00894

(0.00377) (0.00364)
Surplus Aid Should Go to Highest Need 0.0362

0.0325

(0.0175) (0.0169)
Surplus Aid Should Be Returned to Univ 0.00421
(0.0159)
Student Cost (Hundreds) 0.00334 0.00803

(0.00598) (0.00436)
International Program 0.182

0.115

(0.0681) (0.0545)
Flying to Destination -0.0453
(0.0478)
LASSO Selected Covariates No No No No Yes
Control Mean 0.16 0.16 0.16 0.16 0.16
Observations (N) 1029 1029 1029 1029 1029
Robust Standard Errors are reported in parentheses. Regressions are conditional on University Fixed Eects. The
omitted category for FAFSA groupings is Filed FAFSA with Guardians (and reported actual EFC). The omitted
category for Ethnicity is White. The omitted category for Mostly Associates with Lower SES is Associates with All
SES. Mostly Associates with Middle and Upper SES are both included, but are not displayed (both are insignicant).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
44
Figure 3
Figure 4
45
Table 4: Marginal Treatment Eect along Bias in Prior on Comparative Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated -0.0168 0.0703 -0.00765 0.0585 0.0607 0.0284
(0.0792) (0.103) (0.129) (0.0510) (0.0559) (0.0728)
Treated x Overestimates 0.119 0.103 0.0865 0.0527 0.0313 0.0144
(0.111) (0.0989) (0.0989) (0.0746) (0.0627) (0.0627)
Treated x Underestimates x Bias 0.00379 0.00302 0.00368 -0.000267 -0.000601 -0.000167
(0.00291) (0.00293) (0.00285) (0.00169) (0.00168) (0.00168)
Treated x Overestimates x Bias -0.00797

-0.00607

-0.00699

-0.00635

-0.00483

-0.00569

(0.00345) (0.00324) (0.00314) (0.00241) (0.00216) (0.00217)
Controls N Y Y N Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
Tr x Und x Bias = - Tr x Ov x Bias f0.353g f0.477g f0.430g f0.025g f0.045g f0.030g
Control Group Mean 0.36 0.36 0.36 0.17 0.17 0.17
Observations (N) 616 616 616 616 616 616
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects.
Sample excludes students who can't nd or choose not to report their actual EFC. Columns 1 & 4 are otherwise
unconditional, while Columns 2, 3, 5, & 6 control both for covariates determined by LASSO double selection and, as
indicated, for treatment interacted with quintiles of EFC and other baseline correlates of bias in perceived need. The
Treat coecient should, therefore, be interpreted as the CATE for middle EFC, low bias students in (2) and (5) and
for middle EFC, low bias, middle savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
46
Table 5: Dierential Treatment Eect: Over/Under Estimators of Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated 0.0611 0.133 0.0639 0.0517

0.0489 0.0159
(0.0523) (0.0828) (0.112) (0.0308) (0.0420) (0.0629)
Treated x Overestimates -0.0933 -0.0508 -0.0910 -0.0492 -0.0318 -0.0693
(0.0741) (0.0688) (0.0712) (0.0470) (0.0431) (0.0438)
Controls N Y Y N Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
Pvalue: Treat + Tr x Overest f0.544g f0.335g f0.810g f0.945g f0.727g f0.415g
Control Group Mean 0.36 0.36 0.36 0.17 0.17 0.17
Observations (N) 616 616 616 616 616 616
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects.
Sample excludes students who can't nd or choose not to report their actual EFC. Columns 1 & 4 are otherwise
unconditional, while Columns 2, 3, 5, & 6 control both for covariates determined by LASSO double selection and, as
indicated, for treatment interacted with quintiles of EFC and other baseline correlates of bias in perceived need. The
Treat coecient should, therefore, be interpreted as the CATE for middle EFC, low bias students in (2) and (5) and
for middle EFC, low bias, middle savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table 6: Dierential Treatment Eect: Over/Under Estimators of Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated 0.0109 0.109 -0.0115 0.0647

0.0632 -0.00533
(0.0556) (0.0930) (0.112) (0.0388) (0.0543) (0.0635)
Treated x Underestimates 0.0921 0.0656 0.0875 0.00123 0.00561 0.0298
(0.0815) (0.0840) (0.0845) (0.0487) (0.0513) (0.0522)
Treated x Overestimates -0.0779 -0.0772 -0.106 -0.0905 -0.0899

-0.126

(0.0835) (0.0831) (0.0837) (0.0559) (0.0529) (0.0541)
Controls Y Y Y Y Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
P-value: Tr x Under = Tr x Over f0.048g f0.104g f0.028g f0.067g f0.071g f0.003g
Control Group Mean 0.36 0.36 0.36 0.17 0.17 0.17
Observations (N) 616 616 616 616 616 616
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects.
Sample excludes students who can't nd or choose not to report their actual EFC. Columns 1 & 4 are otherwise
unconditional, while Columns 2, 3, 5, & 6 control both for covariates determined by LASSO double selection and, as
indicated, for treatment interacted with quintiles of EFC and other baseline correlates of bias in perceived need. The
Treat coecient should, therefore, be interpreted as the CATE for middle EFC, low bias students in (2) and (5) and
for middle EFC, low bias, middle savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
47
Figure 5: Treatment Eects Associated with Table 6
Figure 6: Treatment Eects Associated with Table 6
48
Table 7: Overall Average Treatment Eects
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4)
Treated 0.0200 0.0168 0.0311 0.0277
(0.0373) (0.0337) (0.0240) (0.0210)
90% Condence Interval (-0.04 , 0.08) (-0.04 , 0.07) (-0.01 , 0.07) (-0.01 , 0.06)
Controls N Y N Y
Control Group Mean 0.36 0.36 0.17 0.17
Observations (N) 616 616 616 616
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects. Sample
excludes students who can't nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
49
Table 8: Heterogeneous Treatment Eects on Amount of Aid Requested (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Separate Regressions with each Variable Z: (1) (2) (3) (4) (5) (6)
Treated 0.0746

0.0781

0.0439 0.0688 0.0512 0.0104
(0.0434) (0.0457) (0.0467) (0.0448) (0.0480) (0.0415)
Treated x Z -0.0331 -0.0312 0.0519 -0.00697 0.0331 0.126

(0.0544) (0.0480) (0.0478) (0.0477) (0.0490) (0.0511)
Treated x Overestimates -0.0502 -0.0711 -0.0877 -0.113

-0.0928 -0.0670
(0.0533) (0.0573) (0.0583) (0.0623) (0.0628) (0.0524)
Treated x Overestimates x Z -0.166 -0.0801 0.0124 0.0289 -0.00746 -0.103
(0.110) (0.0946) (0.0946) (0.0926) (0.0915) (0.0877)
I. Low Bias/Under & Z = Overestimate & Z -0.216

-0.151

-0.075 -0.085 -0.100 -0.170

(Tr x Over + Tr x Over x Z) f0.024g f0.047g f0.314g f0.224g f0.137g f0.016g
II. Overestimates not Z = Overestimate & Z -0.199

-0.111 0.064 0.022 0.026 0.022
(Tr x Z + Tr x Over x Z) f0.036g f0.168g f0.434g f0.782g f0.737g f0.754g
Control Group Mean 0.17 0.17 0.17 0.17 0.17 0.17
Observations Low Bias/Under x NOT Z 309 259 219 236 231 213
Observations Low Bias/Under x Z 107 157 197 180 185 203
Observations Overest x NOT Z 150 140 116 108 89 74
Observations Overest x Z 50 60 84 92 111 126
Observations (Total) 616 616 616 616 616 616
This table presents coecients from 6 regressions (columns) where the listed variable is interacted with treatment
and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors are reported in
parenthases and p-values for statistical tests in brackets. All regressions are conditional on university xed eects,
covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC. Overestimates
& Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not interacted with
Overestimates or Underestimates represent eects for students with low bias. Sample excludes students who couldn't
nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
50
Table 9: Heterogeneous Treatment Eects on Amount of Aid Requested (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
A Single Regression with all Variables Z: (1) (2) (3) (4) (5) (6)
Treated 0.017 0.017 0.017 0.017 0.017 0.017
(0.058) (0.058) (0.058) (0.058) (0.058) (0.058)
Treated x Z -0.030 -0.017 0.041 -0.004 -0.008 0.109

(0.057) (0.048) (0.048) (0.049) (0.050) (0.053)
Treated x Overestimates 0.049 0.049 0.049 0.049 0.049 0.049
(0.093) (0.093) (0.093) (0.093) (0.093) (0.093)
Treated x Overestimates x Z -0.167 -0.151 0.004 0.005 -0.004 -0.125
(0.112) (0.094) (0.093) (0.097) (0.093) (0.092)
I. Low Bias/Under & Z = Overestimate & Z -0.119 -0.103 0.052 0.053 0.044 -0.077
(Tr x Over + Tr x Over x Z) f0.319g f0.349g f0.598g f0.625g f0.682g f0.497g
II. Overestimates not Z = Overestimate & Z -0.198

-0.168

0.045 0.001 -0.012 -0.016
(Tr x Z + Tr x Over x Z) f0.040g f0.036g f0.572g f0.989g f0.880g f0.831g
Control Group Mean 0.17 0.17 0.17 0.17 0.17 0.17
Observations Low Bias/Under x NOT Z 309 259 219 236 231 213
Observations Low Bias/Under x Z 107 157 197 180 185 203
Observations Overest x NOT Z 150 140 116 108 89 74
Observations Overest x Z 50 60 84 92 111 126
Observations (N) 616 616 616 616 616 616
This table presents coecients from a single regression including the interaction of all listed variables (columns)
with treatment and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors
are reported in parenthases and p-values for statistical tests in brackets. The regression is conditional on university
xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC.
Overestimates & Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not
interacted with Overestimates or Underestimates represent eects for students with low bias. Sample excludes
students who couldn't nd or choose not to report their actual EFC.

p < 0:1,

p < 0:05,

p < 0:01,

p< 0:001
51
Figure 7: Treatment Eects Associated with Table 10
Figure 8: Treatment Eects Associated with Table 11
52
Table 10: Treatment Eects on Applying for Aid
Low EFC Medium EFC High EFC
Under/Low Bias 0.081 0.143

-0.100

(0.074) (0.076) (0.058)
Overestimates -0.126 0.092 -0.138
(0.131) (0.106) (0.093)
Observations Under/Low Bias 151 145 120
Observations Overestimates 55 60 85
This table presents estimates of from a single regression including interactions of each tercile of EFC with Overes-
timating need or Underestimating / Having Low Bias and with an indicator for treatment assignment. Coecients
re
ect estimates of the average treatment eect of information provision within each EFC tercile - Bias grouping
pair. Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and
on baseline covariates determined by LASSO double selection. Sample excludes students who couldn't nd or choose
not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table 11: Treatment Eects on Amount Requested (as % of Student Fee)
Low EFC Medium EFC High EFC
Under/Low Bias 0.104

0.078

-0.012
(0.053) (0.038) (0.030)
Overestimates -0.008 0.003 -0.077
(0.100) (0.068) (0.048)
Observations Under/Low Bias 151 145 120
Observations Overestimates 55 60 85
This table presents estimates of from a single regression including interactions of each tercile of EFC with Overes-
timating need or Underestimating / Having Low Bias and with an indicator for treatment assignment. Coecients
re
ect estimates of the average treatment eect of information provision within each EFC tercile - Bias grouping
pair. Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and
on baseline covariates determined by LASSO double selection. Sample excludes students who couldn't nd or choose
not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
53
Figure 9: Treatment Eects Associated with Table 12 Panel A
Figure 10: Treatment Eects Associated with Table 12 Panel B
54
Table 12: Treatment Eect by Tercile of EFC & Savings
Applied for Aid Request (% of Student Fee)
Panel A: Eects by EFC Tercile (1) (2) (3) (4)
Treated x Low EFC 0.0305 0.0280 0.0796

0.0766
(0.0652) (0.0650) (0.0479) (0.0470)
Treated x Mid EFC 0.127

0.122

0.0553 0.0558

(0.0622) (0.0602) (0.0338) (0.0335)
Treated x High EFC -0.0837 -0.0855

-0.0284 -0.0314
(0.0528) (0.0518) (0.0263) (0.0261)
Panel B: Eects by Savings Tercile
Treated x Low Savings 0.0174 0.0279 0.117

0.124

(0.0652) (0.0616) (0.0476) (0.0438)
Treated x Mid Savings -0.0254 -0.0197 -0.0373 -0.0394
(0.0658) (0.0608) (0.0400) (0.0360)
Treated x High Savings 0.0658 0.0521 0.00234 0.00266
(0.0562) (0.0544) (0.0256) (0.0263)
Controls N Y N Y
Control Mean 0.36 0.36 0.17 0.17
Observations (N) 616 616 616 616
This table presents estimates of the average treatment eects of information provision within each tercile of EFC
and each tercile of Personal Savings. Panels A and B are estimated independently (i.e., not conditional on treatment
interactions of the other). Estimates are conditional on baseline covariates determined by LASSO double selection.
Robust standard errors are reported in parenthases Sample excludes students who couldn't nd or choose not to
report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
55
Figure 11: Treatment Eects Associated with Table 13
Figure 12: Treatment Eects Associated with Table 14
56
Table 13: Distributional Eects on Applying for Aid
Low Savings Medium Savings High Savings
Low EFC 0.087 0.045 0.178
(0.077) (0.096) (0.114)
High EFC -0.127 -0.034 0.028
(0.111) (0.081) (0.060)
Observations Low EFC 148 103 58
Observations High EFC 65 108 134
This table presents estimates of from a single regression including interactions of each tercile of EFC and Savings
with one another and with an indicator for treatment assignment according to Specication 9. Coecients re
ect
estimates of the average treatment eect of information provision within each EFC tercile - Savings tercile pair.
Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and on
baseline covariates determined by LASSO double selection. Sample excludes students who couldn't nd or choose
not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table 14: Distributional Eects on Amount Requested (as % of Student Fee)
Low Savings Medium Savings High Savings
Low EFC 0.175

-0.007 0.027
(0.057) (0.064) (0.054)
High EFC 0.003 -0.023 0.002
(0.062) (0.036) (0.027)
Observations Low EFC 148 103 58
Observations High EFC 65 108 134
This table presents estimates of from a single regression including interactions of each tercile of EFC and Savings
with one another and with an indicator for treatment assignment according to Specication 9. Coecients re
ect
estimates of the average treatment eect of information provision within each EFC tercile - Savings tercile pair.
Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and on
baseline covariates determined by LASSO double selection. Sample excludes students who couldn't nd or choose
not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
57
ESSAY 2:
UNINTENDED CONSEQUENCES OF THE FOSTA-SESTA LEGISLATION
9 Introduction
The Allow States and Victims to Fight Online Sex Tracking Act and the Stop Enabling Sex Traf-
ckers Act passed U.S. Congress as the joint FOSTA-SESTA legislation and was signed into law
on April 11, 2018.
15
The FOSTA-SESTA package (henceforth FOSTA) was motivated as a protec-
tion against victims of sex tracking online and especially of minors. The primary contribution of
FOSTA was to clarify that Section 230 of the Communications Decency Act of 1996 did not oer
protections to internet providers and users against enforcement of laws related to sexual exploitation
of children or sex tracking in general. The immediate result of the legislation was the seizure and
shutdown of Backpage.com, which had been under ongoing investigation for \knowing facilitation
of online sex tracking."
16
However, FOSTA also brought about a self-censorship of other websites
concerned with legal repercussions of the ruling for their platforms. Most notably, Craigslist.com
eliminated the Craigslist Personals section of its website on March 22, 2018, in anticipation of the
law.
This paper provides the the rst empirical evidence related to the causal impact of the FOSTA-
SESTA policy change in the United States. The potential eects of the policy are varied along
several dimensions of social welfare.
17
I present evidence on two outcomes. First, I document
15
https://www.congress.gov/bill/115th-congress/house-bill/1865/text
16
The 2017 report from the U.S. Senate hearing on Backpage.com can be found at
https://www.hsgac.senate.gov/download/backpagecoms-knowing-facilitation-of-online-sex-tracking.
17
Despite receiving bipartisan support, the policy met with considerable opposition from U.S. sex workers and also
initially from major web companies concerned about the broader implications of revisions to the Communications
Decency Act on internet freedom. Mass media coverage has focused largely on potential policy impact on the
livelihood and safety of consensual sex workers and on ambiguous ecacy to improve the welfare of adults and
minors who would be tracked, e.g., if the elimination of online platforms used for commercial sex purposes reduces
visibility, but not the exploitation of potential victims.
58
that the elimination of Craigslist Personals resulted in at least a 6% increase in the U.S. rate of
rape and sexual assault among Metropolitan Areas with above median Craigslist usage. As will be
described, this estimate is necessarily a lower bound given the empirical strategy employed. This
is an unintended consequence of FOSTA, which entered neither into the initial policy debate nor
media criticisms after the bill was signed. However, the mechanism suggested by economics liter-
ature is clear. In several settings it has been documented that increasing (decreasing) barriers to
accessing commercial sex results in a substitution toward (away from) rape (Bisschop et al., 2017;
Cunningham and Shah, 2017; Ciacci, 2018; Ciacci and Sviatschi, 2018). While a portion of trans-
actions on Craigslist were related to commercial sex, the observed eect cannot be distinguished
from owing more broadly to increased search barriers for consensual, unpaid sex partners posed by
the elimination of Craigslist Personals, as the platform was used primarily for this purpose.
The second outcome that I study is that of prostitution arrests. I document an increase in juve-
nile prostitution arrests of females, post FOSTA, owing to the elimination of Craigslist Personals.
Ocial arrest records on prostitution are generally understood to re
ect rates of street solicitation
rather than the online activity of sex workers which would coincide with anecdotal reports from
police that street prostitution increased in 2018 (Cunningham and DeAngelo, 2019). To this end,
the observed increase provides suggestive evidence that the commercial sex activity of either co-
erced or un-coerced female minors (or both) shifted toward street solicitation after the shutdown
of the internet platform.
These results are important since they would provide the rst evidence that the market for sex
with minors, which FOSTA sought to block, adjusted, in part, to a new solicitation environment.
It is also important in that research indicates street solicitation to be more dangerous for female
sex workers than online solicitation. For example Cunningham and DeAngelo (2019) document a
drop in the female homicide rate resulting from the initial expansion of Craigslist Erotic Services
(eliminated from Craigslist in 2010). The expansion of the online marketplace protected sex workers
by removing them from the streets and allowing greater screening capabilities of potential partners.
If the observed eect of FOSTA on prostitution arrests is an indication of movement of minors back
onto the streets, it suggests that additional arrests are occurring that would not have occurred in
the absence of FOSTA. However, further evidence would be necessary to draw conclusions as to
the net impact on the welfare of the minors engaged in either coerced or un-coerced commercial
59
sex activity due to the policy.
The evidence on juvenile arrests is suggestive of movement to the streets, as I am not able, at
this time, to distinguish whether the eect is in fact driven partly by a shift in police resources post
FOSTA. For example, the documented eect could also be explained if resources devoted to tracing
online tracking of minors on Craigslist pre-FOSTA were redirected toward street enforcement
once the primary online platforms utilized for commercial sex disappeared. That said, lack of
any observed eect on total prostitution arrests provides support to the conclusion that there was
specically a shift in the behavior of minors. Further investigation would be valuable to disentangle
the potential mechanisms driving the observed increase, as a shift toward street solicitation may
indicate that the policy had limited eectiveness in stopping the type of exploitative activity which
it sought to address and may also have resulted in movement to a solicitation environment with
increased risk.
The primary contribution of this work is in providing the rst empirical evidence on the con-
sequences of the FOSTA policy, evidence which holds implications for government regulation of
both commercial and non-commercial sex markets. Moreover, this is the rst study, utilizing U.S.
nationally representative data, to document the substitution eect between access to consensual
sex markets (whether commercial or not) and the incidence of rape and sexual assault. By doing so
this study contributes most directly to the literature on government regulation and criminalization
of the prostitution market and its eects on societal outcomes (Bisschop et al., 2017; Cunningham
and Shah, 2017; Ciacci, 2018; Cunningham and DeAngelo, 2019; Cameron et al., 2020).
Second, to demonstrate the robustness of the impact of the elimination of Craigslist Personals
on rape and sexual assault, I provide evidence ruling out that the observed increase in the rate
of reported rape was driven by the \Me Too" movement, which saw greater interest and activity
at the end of 2017 and into 2018 and may have encouraged increased reporting of oenses to the
police. In doing so, I provide the rst causal evidence that this social movement brought about a
decrease in rape and sexual assault in the United States beginning in 2019.
To be clear, this paper does not directly address the question of whether sex tracking was
reduced as a result of FOSTA. Though this is a signicant question in its own right, reliable metrics
on human tracking rates, necessary to speak to the ecacy of the policy's primary aim, are not
readily available. Though I explored the use of ocial FBI crime metrics on human tracking
60
incidents and arrests, cases are sparse and do not provide for clean and reliable identication of the
impact of FOSTA using the dierence-in-dierences methodology employed in this paper. Further
research on this topic would be invaluable to be able to assess in a more complete fashion the
benets vs. social cost of the policy change given the current evidence on elevated incidence of
rape.
The paper is laid out as follows. Section 10 provides a description of the data sets used for
analysis. Section 11 presents the main ndings related to the causal impact of the removal of
Craigslist Personals on rates of rape and sexual assault in the US. Section 12 provides evidence
ruling out that the observed eect on rape is driven by the simultaneous impact of the Me Too
movement. Section 13 presents evidence that the elimination of Craigslist Personals resulted in an
increase in prostitution arrests of female minors. Finally, Section 14 concludes with discussion of
the policy implications of these ndings.
10 Data Sets
The main empirical strategy used in this study is a dierence-in-dierences methodology comparing
geographic regions with high and low usage of Craigslist prior to the FOSTA policy change. Four
data sets are combined for the analysis. The rst is google search data from Google Trends utilized
to capture geographic variation in Craigslist usage and also interest in the Me Too movement.
The second and third sources are annual crime rates (aggregated to the Metropolitan Statistical
Area) and monthly arrest rates (at the agency level) originating from the FBI Uniform Crime
Reporting program (UCR). These data sets provide the primary outcomes. The nal source of
data is county level population demographic estimates from the U.S. Census Bureau. These are
used to observe heterogeneity among geographic regions assigned to treatment and control and to
construct demographic \treatment" and \control" groups used to test robustness of the Craigslist
results related to rape and sexual assault.
Since FOSTA simultaneously impacted the entire United States, to plausibly estimate a lower
bound on the causal eect of the policy change I utilize a measure of variation in \exposure" to
the policy. Google Trends provides data on the relative frequency of google search { for a term, a
topic or a website { across geographies in a specied period of time. From Google Trends I obtain
61
a measure of relative search frequency for Craigslist, Inc. (advertising company) during the year
prior to FOSTA (3/15/2017 to 3/15/2018).
18
While data is also available at the state and national
level, I utilize the U.S. metro area data which provides the nest level of geographical heterogeneity
in google search frequency that covers the entire U.S.
19
At the metro level, Google Trends reports
a comparative search score for 210 market areas. The score for a given region, m, and search term
or topic, X, is dened as follows:
Trends Score
m
(X) =
Searchm(X)
Total Searchm
max

Search
j
(X)
Total Search
j
100 j =f1;2;:::;m;:::;210g: (10)
Here Search
m
(X) refers to a count of total searches in region m for topic X, and Total Search
m
refers to a count of total google searches in region m over the same period. Thus, the score is
normalized to the total search frequency in a given area and to the highest relative search frequency
among the geographies being compared such that the region with highest relative search frequency
is reported as 100 and the others are reported as a percentage of the highest. The measure is also
rounded to the nearest integer.
Appendix Figure B.1 plots the Google Trends comparative search score for the 210 market
areas for Criagslist, Inc. (advertising company), as well as for Backpage.com (website) and for
the the Me Too movement (topic) in the year preceding FOSTA.
20
The metro areas are ordered
by increasing search frequency for Craigslist with Salt Lake City, UT showing least interest and
Medford-Klamath Falls, OR the highest. In 2017, Salt Lake City has roughly 14% the level of
search interest in Craigslist as Medford-Klamath Falls.
For the purpose of the dierence-in-dierence estimation, the continuous measure of relative
search frequency is divided at the median value among the observed units. E.g., in the case of crime
data, 218 Metropolitan Statistical Areas make up a balanced panel from 2013-2019. These MSAs
are mapped to their corresponding Trends Metro Area to attribute a relative search score. Then,
the set of MSAs are divided at the median of their search scores to dene treatment and control
18
Results are robust to other time frames such as the 3 months or 6 months prior to the policy change.
19
Some municipal level data is available, but has limitations for relative comparison of U.S. cities across the entirety
of the U.S.
20
In addition to allowing for queries of a specic word or phrase, Google Trends aggregates searches related to known
topics, websites, etc. Thus, Criagslist, Inc. aggregates all searches that Google determines are directed toward
navigating to Craigslist whether or not they include other words.
62
units. Figure B.2 displays the geographic variation in search intensity for Craigslist and marks the
market areas which will be assigned to treatment and control for the primary outcomes related to
crime rates.
21
Outcomes related to crime rates derive from data provided by the FBI Uniform Crime Reporting
program (UCR). UCR provides summary crime data at the Metropolitan Statistical Area (MSA)
annual level. These data contain the annual counts of reported incidents and crime rates in 7
major crime categories: murder, aggravated assault, rape and sexual assault, robbery, burglary,
larceny, and motor vehicle theft. The rate of rape and sexual assault reported in these data serves
as a primary outcome of interest while the other crime categories are used for placebo tests. Since
not all agencies within an MSA may provide data to UCR, an MSA is included only if at least
75% of the MSA population is represented by reporting agencies in a given year. In the case
that the full MSA population is not represented, the FBI provides an extrapolated estimate of
incidents and an estimated crime rate for the full MSA population. These FBI estimates are used
in the analysis that follows, though results are nearly identical when run using the actual reported
numbers (representing a sub-population of the MSA).
The outcome of prostitution arrest rates also derives from the UCR program, but from the
agency-month data on reported arrests. For analysis of eects on prostitution arrests, to maximize
power I utilize an unbalanced panel of 6600 crime reporting agencies from 2012 to 2018 (the last
available data year). Results for a balanced panel of reporting agencies over the same years are
presented in the appendix. The results are similar, however, they become marginally insignicant
due to the large drop in sample size.
The U.S. Census Bureau provides estimates of annual population at the county level bro-
ken down by demographics of gender and race and ethnicity. These data are aggregated to the
Metropolitan Statistical Area to be utilized in the analysis of robustness of results on rape and
sexual assault as will be described in Section 12.
21
In reality the determination of treatment and control is based on division of search scores at the median for the
Metropolitan Statistical Areas which form a balanced panel from 2013 to 2019 in the Uniform Crime Reporting
crime data, thus not all Trends market areas are utilized, however the division in Figure B.2 re
ects search scores
above and below the median MSA score.
63
11 Removal of Craigslist Personals, Rape, & Sexual Assault
This section presents causal evidence that the elimination of the Personals Section of Craigslist.com
in the U.S. on March 22, 2018, in anticipation of the signing of FOSTA into law, resulted in an
increase in the rate of rape and sexual assault in the United States, which persisted through at
least 2019 (the latest data currently available). This eect is identied by estimating a dierence-
in-dierences specication with the outcome of the rate of rape and sexual assault from the FBI
Uniform Crime Reporting data. I utilize the full set of U.S. Metropolitan Statistical Areas (MSAs)
for which annual crime rates are reported for rape and sexual assault for every year from 2013-2019
(a balanced panel of 218 MSAs). I order the MSAs and divide them at the median, based on their
Google Trends search score for Craigslist, Inc. (advertising company) over the year prior to the
implementation of FOSTA-SESTA to obtain the set of high-Craigslist usage MSAs (Treatment) and
low-Craigslist usage MSAs (Control). The geographic division of MSAs into Craigslist treatment
and control is depicted in Appendix Figure B.3.
Figure 14 plots the trends in the rate of rape and sexual assault for both groups over the analysis
period. The gure provides visual evidence of a dierential deviation from trend beginning in 2018.
Moreover, the apparent cleanliness of parallel pre-trends for the groups, prior to 2018, invites the
use of a dierence-in-dierences specication to estimate a lower bound on the causal impact of
FOSTA on the rate of rape and sexual assault in the U.S. The primary specication of interest is:
y
mt
=High Craigslist
m
Post
t
+
m
+
t
+
mt
(11)
where the outcome,y
mt
, is the rate of Rape and Sexual Assault in MSA m and year t,High Craigslist
m
is a dummy equal to 1 if MSA m has an above median Google Trends search score for Craigslist in
the year prior to FOSTA,Post
t
is a dummy equal to 1 for the years 2018 and 2019, and
m
and
t
are MSA and year xed eects.
mt
captures the idiosyncratic error. is the parameter of interest
which captures the dierential deviation from trend in the outcome during 2018 and 2019 for High
(relative to Low) Craigslist usage MSAs.
^
will be an underestimate of the total impact of the policy change for two reasons. First,
Craigslist Personals was removed at the end of the rst quarter of 2018. Inclusion of this untreated
64
quarter in the post-period will dampen any observed treatment eect owing to the policy. Second,
control MSAs, while having relatively lower Craigslist usage, also make use of the platform. Since
all MSAs are, in fact, impacted by the policy change (and should theoretically observe a treatment
eect in the same direction), the control group trend may overestimate the true counterfactual
trend for the treatment group in the absence of the policy change, biasing downward the ATT
estimate.
The estimate of from Equation 11 is presented in Table 15, Column 1. Column 2 presents
results from the the same specication, but where MSA population is introduced as an additional
control. Both regressions are weighted by MSA population and cluster standard errors by Google
Trends Metro Area { the geographical level at which the Google Trends Craigslist search score is
assigned. The point estimate of 2.94 represents roughly a 6% increase in the rate of rape and sexual
assault for high-Craigslist usage MSAs relative to the counterfactual trend predicted by the control
group.
Interpretation of the estimate of from Specication 11 as a causal eect owing to the removal
of Craigslist Personals relies on the assumption that the outcome for the high-Craigslist usage
MSAs would have evolved similarly to the low-Craigslist usage MSAs in the absence of the policy
intervention. There are two primary threats to the validity of this assumption. The rst is that
the low-Craigslist usage MSAs areas generally provide a poor counterfactual in trends to the high-
Craigslist usage MSAs. To address this concern, I test for the comparability of trends prior to 2018.
The second concern is that some other event, deferentially impacting the treatment and control
group as dened (i.e., correlated with regional search interest in Craigslist prior to FOSTA), is
driving the observed eect in 2018 and 2019, despite the seemingly parallel trends in the outcome
prior to 2018. I discuss a particularly salient threat related to this concern in the following section.
To formally test for parallel trends in the rate of rape and sexual assault between high- and
low- Craigslist usage MSAs, prior to 2018, I estimate the following event study specication:
y
mt
=
X
6=2017
f

High Craigslist
m
1( =t)g +
m
+
t
+
mt
; (2013;:::; 2019): (12)
All variables are the same as dened in Specication 11, however now I interact the dummy for
the treatment group with dummies for each year of the panel, omitting 2017 which serves as the
65
reference year.
The point estimates of the

are plotted with their 95% condence intervals in Figure 15. The
wider condence bands for the years 2013 and 2014 likely re
ect the transition, beginning in 2013,
to the new summary rape denition, which will be discussed in more detail in the next section. The
estimates depict a clear discontinuity and signicant increase in 2018 and 2019. Table 16 reports
coecients and standard errors for these estimates as well as p-values for the test of joint signicance
of pre- and post- period coecients. The test cannot reject that the pre-period coecients are
jointly zero, lending credibility to the assumption that, in the absence of a discontinuous event
in 2018, the outcome trends would have evolved in a similar fashion. Moreover, the joint test of
coecients for 2018 and 2019 rejects with 95% condence that there was no divergence from trend
in these years.
Taken together, the estimates presented in Tables 15 and 16 provide strong evidence that
having high versus low usage of Craigslist, as measured by Google Trends search for Craigslist
in the year prior to FOSTA, predicts a jump in the rate of rape and sexual assault in 2018 and
2019, subsequent to the FOSTA-SESTA legislation, which eliminated Craigslist Personals, as well
as Backpage.com and some other online platforms previously utilized for commercial and non-
commercial sex purposes. A remaining threat to being able to attribute this eect precisely to the
removal of Craigslist Personals is the possibility that the high- and low- Craigslist usage MSAs may
have for another reason diverged on the outcome of rape and sexual assault in 2018 and 2019. Of
specic concern here is the strengthening of the Me Too movement late in 2017 and throughout
2018 which may have simultaneously impacted rates of actual and reported rape and sexual assault.
It is to this concern that I turn in the following section.
12 The Role of \Me Too" and Backpage.com
It should be clear from the previous section that the division of MSAs along Craigslist search
interest is a strong predictor of increased rape and sexual assault beginning in 2018, however,
Craigslist search scores are not randomly assigned to MSAs. This implies that Craigslist usage
may be correlated with population characteristics that could respond deferentially to other shocks
{ besides the elimination of Criagslist personals { taking place between 2017 and 2018. For this
66
reason it is important to address the possible in
uence of 2 other events which occurred in the same
time frame and may have aected actual rates of rape and sexual assault or the reporting of rape
and sexual assault to authorities. The rst is the removal of Backpage.com { a direct result of the
FOSTA legislation. The second is the activity of the Me Too movement which gained traction in
late 2017 and into 2018. Figure B.4 plots the time trend of google search interest in the Me Too
movement in the United States. The rst signicant activity took place from May to August 2016
with no apparent impact on aggregate rates of rape and sexual assault { reported or non-reported
(see Figure 13). The second period of activity began in October 2017, sparked by the case of Alyssa
Milano and Harvey Weinstein, followed by ongoing interest thereafter.
While it is most probable that the elimination of Backpage.com would eect an increase in
rape and sexual assault { due to increased search costs for commercial sex transactions { the
theoretical impact of the "Me Too" movement is somewhat more ambiguous. E.g., Me Too could
have resulted in 1) increased perception among victims of what acts qualify as rape/sexual assault
(e.g., more likelihood to report date rape as rape) or 2) an increased willingness to disclose having
been raped or sexually assaulted to others or by reporting to the police, both of which could be
re
ected in an increase in rates of reported rape and sexual assault. Me Too could also result in 3)
a discouragement eect due to increased risk of being reported as a perpetrator and/or increased
social shaming of men on the margin of engaging in such behavior. Mechanisms 1 and 2 could
confound the observed eect attributed to the removal of Craigslist Personals if MSAs with high
Craigslist usage dier from areas with low usage on characteristics that make them more responsive
to the Me Too movement. However, if anything, mechanism 3 would dampen the Craigslist eect.
Table 17 presents summary statistics of the 218 MSAs of the crime rates panel. Column 1
presents means and standard deviations for the full sample while Columns 2 and 3 present the
same for the control and treatment MSAs. Column 4 reports the dierence in means for each
variable and the p-value for the test of equality. The Craigslist treatment group is imbalanced on
the Backpage.com search score. The observed negative correlation between Backpage search and
Craigslist search are to be expected if the platforms are partial substitutes for one another. If the
elimination of Backpage in
uenced an increase in rape and sexual assault, this would potentially
bias downward the estimates of the in
uence of Craigslist due to the observed negative correlation.
Notably, balance is achieved along search interest in the Me Too movement. This provides some
67
evidence that the eect observed for Criagslist is not driven by geographic variation in exposure or
response to Me Too that is correlated with Craigslist usage. While the sample is relatively balanced
on crime rates with the exception of robbery and motor vehicle theft, we observe a signicant
dierence in population along demographic characteristics. Specically, the Craigslist treatment
MSAs are slightly more male and have a lower percentage of Black and African American population
and a higher percentage of other races. This raises the concern that various racial groups may have
cultural dierences resulting in a dierential response to the Me Too movement. In particular the
Me Too movement has received some criticism for too much attention focused on white females of
socioeconomic privilege. To address concerns around the potential simultaneous role of Me Too or
the elimination of Backpage.com, I condition the dierence-in-dierences estimates from Craigslist
using treatment groups dened by high Backpage usage and Me Too interest, similarly measured
by Google Trends. Though Me Too continues to be a topic of cultural interest into 2019, I use the
Google Trends score based on total search over the same period as Craigslist and Backpage (March
15, 2017 - March 15, 2018) to dene the high- and low- intensity groups for the Me Too movement.
This is both for consistency and also to ensure that the treatment measure is not impacted by
FOSTA (e.g., if sexual assault increases post FOSTA, this may increase Google search for the Me
Too movement in areas most impacted). This time frame captures activity both from the initial
spike of media attention in October 2017, as well as the continued search interest up until March
2018. Results presented below are robust to using other time frames to measure search interest in
Me Too such as only October 2017, only November 2017 - March 2018, and all of 2018.
Unfortunately, parallel pre-trends in the rate of rape and sexual assault are not as clean extend-
ing back to 2013 for the division of MSAs into treatment and control based on Backpage.com and
Me Too. Appendix Figure B.5 plots the outcome trends for these treatment and control groups.
The disruption in parallel trends appears driven by the change in the FBI UCR program's summary
denition of rape. Agencies were encouraged beginning in 2013, but not required until 2017, to
utilize a new denition of rape (currently in use). In 2013 9,243 municipal agencies still used the
legacy denition of rape compared to 2,207 in 2014, and 1159 in 2015. By 2017 all agencies were
required to use the new denition to report to the Uniform Crime Report. This is visible in the
trends of treatment and control groups dened by high and low search for Backpage.com and Me
Too. While the Craigslist treatment appears well balanced on the transition timing, the Backpage
68
and Me Too treatment and controls are not in the early years. By 2015 all three treatment and
control groups converge roughly to parallel trends. Thus, to evaluate the potential in
uence of
Backpage.com and the Me Too movement post 2018, and test robustness of the eect of Craigslist
Personals, I restrict the data to a panel from 2015-2019.
Table 18 presents dierence-in-dierences estimates from Specication 11 for each of the three
treatment groups of interest using the balanced panel from 2015-2019. Column 4 reports estimates
from a single regression where all three treatment group indicators, interacted with a dummy for
the years 2018 and 2019, are included. The only coecient which obtains statistical signicance
is for MSAs with high usage of Craigslist prior to FOSTA. Moreover, the estimate for Craigslist
remains strongly signicant when controlling for potential eects of Backpage and Me Too in the
post period.
To verify the reliability of parallel trends for each treatment in the pre- period, and to consider
potential dynamic eects, I estimate Specication 12 for each treatment group independently and,
then, including all treatments together. Table 19 reports these estimates as well as p-values from
the statistical tests for joint signicance of pre- and post- coecients. Figures 16 - 18 plot the
estimates corresponding to each treatment group from the regression where all three treatments
are included together.
Notably, the coecients for high Craigslist usage remain signicant in 2018 and 2019. While
Backpage usage appears to have an eect in 2019, after conditioning on Craigslist and Me Too, the
estimate drops almost by 50% and becomes signicant suggesting it was being driven by correlation
with Craigslist or Me Too or both. Finally, the Me Too treatment group sees a drop in rape and
sexual assault in 2019 which attenuates somewhat, but maintains signicance when controlling for
all treatment eects. In all cases we cannot reject parallel trends prior to 2018.
There are a couple of interesting features to note regarding these results. First, they suggest
that there was a longer term eect of the Me Too movement on a decrease in rape and sexual
assault in the United States. To the extent that Google Search interest appropriately captures
varying geographic salience and engagement with the Me Too movement, the event study estimates
indicate that regions taking more interest in the movement in late 2017 and early 2018 { compared
to less engaged regions { saw a decrease in rates of rape and sexual assault beginning in 2019, though
outcome trends had run cleanly parallel for the 4 years prior. This, in itself, is an important nding.
69
Second, the conditional estimates from Column 4 of Tables 18 and 19 support the conclusion
that the observed increase in rape and sexual assault among MSAs with high usage of Craigslist
cannot be attributed to a dierential response to the Me Too movement rather than to the impact
of FOSTA. To arrive at this conclusion, given this evidence, one would have to believe that the
MSA division by Craigslist usage better proxies dierential in
uence of the Me Too movement on
the reporting of rape and sexual assault than does Google search interest in Me Too itself, and
that, despite Google search for Me Too predicting a drop in rape in 2019, Craigslist usage as proxy
for the in
uence of Me Too observes a positive eect.
As a further check on robustness against the potential in
uence of Me Too acting through ethnic
or cultural dierences geographically, I perform a similar exercise dening treatment groups along
the relevant demographic characteristics. This tests whether the division of MSAs by high male
or high white population, e.g., predicts the increase in 2018 and diminishes the observed eect of
Craigslist. These results are presented in the Appendix Section B.2. The Craigslist result is robust
to controlling for dierential eects by ethnicity and gender composition. However, we do see a
signicant treatment eect predicted by MSAs with a high male-to-female ratio.
22
While this could
theoretically be driven by the role of the Me Too movement, it seems much more plausible that this
would re
ect the in
uence of FOSTA. That is, since the online markets for commercial sex were
eliminated as well as the non-commercial market on Craigslist Personals, regions with higher male
composition saw a greater shock in substitution toward rape and sexual assault.
As a nal piece of evidence that the eect observed among high Craigslist usage areas is driven
by the FOSTA policy, I perform randomization inference over treatment assignment of the 218
MSAs in the 2015-2019 panel. The rationale here is to obtain an exact p-value for where the
coecient estimate deriving from the Craigslist split lies in the distribution of possible dierence-
in-dierences estimates which would arise under the null hypothesis that there are actually no
\treated" MSAs (i.e., any dierential variation in rates of rape and sexual assault, pre- and post
2018, are random). Given the sucient sample size we should expect a similar inference result using
sampling standard errors, however, inference based on exact p-values is actually most appropriate
here given that the source of uncertainty derives not from sampling variation, but from treatment
assignment.
22
It appears this eect could be driven by an upward trend, thus I will test for the inclusion of group time trends.
70
Figure 19 plots the distribution of dierence-in-dierences coecients from Specication 11
which occurs when treatment is randomly assigned among the 218 MSAs of the balanced panel
such that the treatment and control groups are comprised of an equal number of MSAs. The distri-
bution represents 50,000 iterations of random assignment and re-estimation. The (unconditional)
coecient estimates for the Craigslist, Backpage, and Me Too treatments are denoted by vertical
lines. Exact p-values for a two tailed test for each treatment coecient are reported under the
histogram. While Me Too and Backpage do not re
ect unlikely estimates under the null, Craigslist
does. An exact p-value of .0036, indicates that, if the observed dierential trends in rates of rape
and sexual assault occurred randomly and there is nothing special about the Craiglsist treatment
group, there would be a .36% chance we would obtain an estimate as high as what we observe for
Craiglist. Or, alternatively put, the division based on Craigslist usage has better explanatory power
in terms of a dierential increase in rape incidence in 2018-2019 than 99.64% of ways in which we
could equally divide the set of MSAs.
Since the distribution in Figure 19 includes all random assignments of treatment and control
MSAs, we may be concerned about the possible in
uence of assignments for which the treatment and
control group are already trending apart during the pre-period. If anything this would be expected
to produce a higher exact p-value since it could lead to greater positive or negative separation in the
post period for the groups trending apart. Nonetheless, to check robustness I perform an exercise
similar to Abadie et al. (2010). I progressively drop random assignments which fail to pass the test
for parallel trends in the pre-period. Excluding those random treatment assignments which fail
the pre-period F-test at condence levels of .01, .05, .1 and .2, the exact p-value for the Craigslist
treatment remains stable at .0034, .0035, .0035 and .0035, respectively.
Thus, while the endogeneity of the Google Trends measure should give some pause with regard
to other characteristics which may be driving the observed eect, if the Craigslist division is, in
fact, proxying for geographical variation in rape and sexual assault owing to the Me Too movement
or any other event besides the removal of Craigslist Personals, it is doing so surprisingly well.
Before turning to the outcome of prostitution arrests, I should note the observed imbalance
on MSA population indicated in Table 17. This raises the concern that results may be driven by
population outliers or due to an urban vs. rural response to Me Too. I address this by demonstrating
robustness to dropping population outliers above the 95th percentile (see Appendix Section B.3).
71
Once population outliers are dropped, treatment and control MSAs are balanced in population size
and the estimated eects become larger.
13 Impact on Prostitution Arrests
In this section I present evidence on the impact of the FOSTA-SESTA legislation on rates of
prostitution arrests. For this analysis, to maximize power, I use the unbalanced panel of crime
reporting agencies from 2012 to 2018 (the last year for which arrest data is currently available).
Results for a balanced panel of agencies reporting every year from 2012 to 2018 are reported in
Appendix Section B.4. The results for the balanced panel are comparable to the unbalanced panel,
however, they become marginally insignicant given the lower sample size.
Once again I estimate Specication 11 including interactions of both the Craigslist and Backpage
treatment groups with dummies for the post period, however now m indexes the crime reporting
agency and t represents a monthly periodicity. The post period is now considered April-December
2018, capturing the true discontinuity due to the FOSTA policy. Table 20 reports dierence-in-
dierences estimates for the outcome of Total Prostitution Arrests along with rates of prostitution
arrests for females, juveniles, female juveniles and male juveniles. While no eect is observed for the
total or total female arrest rates, we observe an increase in juvenile prostitution arrests in regions
with high Craigslist usage post FOSTA. Columns 4 and 5 indicate that this eect is being driven
specically by female juvenile prostitution arrests.
To conrm the existence of parallel trends during the years pre-FOSTA I estimate the following
event study specication:
y
mty
=
X
6=2017
f

High Craigslist
m
1( =y)g +
m
+
t
+
mt
; (2012;:::; 2018): (13)
Note the dierence from Specication 12, in that, while arrests are observed monthly, the event
study estimates represent yearly averages. This is due to the fact that monthly rates are quite
noisy. Aggregating to the yearly level averages out some of this noise and signicantly reduces the
number of coecient estimates required. Moreover, the post period is comprised of 9 months, thus
yearly periodicity is the time interval over which we want to test for parallel trends. Since the
post period begins in April, I estimate Specication 13 for the full set of observations and then
72
again, dropping all observations belonging to the months of January-March, thus ensuring that no
pre-FOSTA months enter into the estimate for 2018. Both sets of estimates for the outcomes of
juvenile arrests and female juvenile arrests are reported in Table 21. Estimates for the full set of
observations will bias the post period estimate for 2018 downward. Thus, Columns 3 and 4 should
be taken as best estimates. Here we observe that parallel trends holds true and the coecients for
2018 are (marginally) signicant. Figure 20 plots the annual trends in juvenile and female juvenile
arrests for quarters 2-4. The trend in event study coecient estimates are displayed in Figure 21,
demonstrating clean pretends followed by an increase beginning in quarter 2 of 2018.
While the eect on prostitution arrests of minors could possibly be explained by a shift in police
resources toward greater street enforcement post FOSTA, the lack of eect on total prostitution
arrests suggests otherwise. It is unclear why increased street enforcement would only focus on
solicitation by minors and ignore adult solicitation. More likely, the eect owes, therefore, to a
shift toward increased street prostitution among minors who were previously matching with clients
on Craigslist. Note that there is much less concern here about confounding factors correlated with
Craigslist usage as the eect begins specically in April 2018 and there is little concern that the
Me Too movement would have had any in
uence on prostitution arrests.
14 Conclusions
This study has presented evidence that the elimination of Criagslist Personals { a major online
platform utilized for identifying partners for intimacy and sex, as well for commercial sexual activ-
ity { resulted in an increase in the U.S. rate of rape and sexual assault and in juvenile prostitution
arrests. The nding on rape, though an unintended policy consequence, furthers scientic evidence
that increasing barriers to accessing markets for commercial and non-commercial sex causes in-
creased rates of sexual violence. Moreover, the evidence on juvenile prostitution arrests suggests
that the market for sexual activity with minors shifted toward street solicitation post FOSTA.
While the net welfare eect on potentially exploited minors remains unclear, an increase in street
activity may be an indication that the most vulnerable { those minors engaged in prostitution due
to lack of other options or due to coercion { simply adjusted their commercial sex activity to an
environment with increased personal risk.
73
Further evidence on a net impact on tracking activity and on the welfare of minors who would
have been tracked, but still remain under the in
uence of exploitative adults, is necessary for
assessing the overall cost-benet of the FOSTA policy change. None-the-less, the current ndings
suggest policy improvements that may be considered. Specically, growing evidence suggests the
observed eect on rape and sexual assault would have been diminished in an environment where
consensual adult sex work is decriminalized and regulated. Apart from mitigating a substation
toward forced sexual activity, a legal market for adult sex work would reduce liability for a platform
like Craigslist Personals, used widely for non-commercial purposes, but at risk under FOSTA due
to regular postings for sex for pay. Moreover, it would encourage cooperation with law enforcement
in the identication of adds related specically to the exploitation of minors.
74
Figure 13
75
Figure 14
Table 15: DID Estimate: Discontinuity in Rape / Sexual Assault post 2018
Depvar: Rate of Rape / Sexual Assault
(1) (2)
High Craigslist x Post 2.926

2.858

(1.365) (1.303)
MSA Population (in 100,000s) -0.144
(1.103)
N MSAs 218 218
N Years 7 7
N Total 1526 1526
This table presents DID estimates based on a panel of US Metropolitan Statistical Areas from 2013 to 2019 where
treatment is dened as being an MSA with above median usage of Craigslist in 2017 (as proxied by Google search
frequency). The post period is dened as 2018-2019. Robust standard errors, clustered by Google Trends Metro Area,
are reported in parenthases. Regressions are weighted by MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
76
Figure 15
Table 16: Event Study Estimates: Impact of FOSTA/SESTA on Rape / Sexual Assault via
Craigslist
Depvar: Rate of Rape / Sexual Assault
(1) (2)
High Craigslist x 2013 -0.107 0.0310
(2.660) (2.557)
High Craigslist x 2014 -0.700 -0.616
(2.396) (2.256)
High Craigslist x 2015 -0.263 -0.222
(1.157) (1.142)
High Craigslist x 2016 -0.403 -0.378
(0.885) (0.899)
High Craigslist x 2018 2.916

2.900

(1.144) (1.168)
High Craigslist x 2019 2.346

2.343

(1.124) (1.133)
MSA Population (in 100,000s) -0.142
(1.087)
P-val for Joint Test of Pre - Periods 0.980 0.980
P-val for Joint Test of Post- Periods 0.027 0.031
N MSAs 218 218
N Years 7 7
N Total 1526 1526
This table presents Event Study estimates based on a panel of US Metropolitan Statistical Areas from 2013 to 2019
where treatment is dened as being an MSA with above median usage of Craigslist in 2017 (as proxied by Google
search frequency). The post period is considered 2018-2019. Robust standard errors, clustered by Google Trends
Metro Area, are reported in parenthases. Regressions are weighted by MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
77
Table 17: MSA Characteristics in 2017 by Craigslist Treatment
Means and Standard Deviations
Craigslist Craigslist Dierence
Variable Full Sample Control Treatment Treat vs Cont.
(1) (2) (3) (4)
Backpage.com Google Trends Score 32.000 33.509 30.406 -3.103**
(11.602) (10.673) (12.360) f0.049g
Me Too movement Google Trends Score 36.335 36.795 35.849 -0.946
(16.848) (15.833) (17.921) f0.681g
MSA Population (in 100,000s) 6.563 9.018 3.968 -5.050***
(12.764) (17.000) (4.239) f0.003g
Rate of Violent Crime 382.641 389.510 375.172 -14.337
(187.446) (183.275) (192.496) f0.577g
Rate of Murder / Manslaughter 4.757 5.116 4.377 -0.740
(3.676) (3.654) (3.678) f0.138g
Rate of Rape / Sexual Assault 48.205 46.374 50.139 3.765
(20.461) (17.572) (23.053) f0.178g
Rate of Robbery 72.614 79.521 65.315 -14.205**
(52.391) (51.379) (52.700) f0.045g
Rate of Aggravated Assault 256.255 258.499 253.816 -4.682
(141.437) (142.189) (141.269) f0.809g
Rate of Property Crime 2,472.427 2,427.823 2,516.602 88.780
(883.800) (851.203) (916.915) f0.471g
Rate of Burglary 462.571 475.906 448.982 -26.924
(211.128) (233.201) (186.112) f0.353g
Rate of Larceny Theft 1,799.219 1,787.383 1,811.281 23.897
(604.366) (592.100) (619.223) f0.774g
Rate of Motor Vehicle Theft 218.054 187.395 250.159 62.764***
(168.184) (110.283) (208.356) f0.006g
Male : Female Pop Ratio 0.981 0.970 0.992 0.021***
(0.043) (0.038) (0.046) f0.000g
% Pop Male 0.495 0.492 0.498 0.005***
(0.011) (0.010) (0.011) f0.000g
% Pop White Alone 0.823 0.791 0.857 0.067***
(0.108) (0.111) (0.093) f0.000g
% Pop Black/African American 0.110 0.154 0.065 -0.089***
(0.098) (0.111) (0.053) f0.000g
% Pop Hispanic 0.154 0.127 0.183 0.056**
(0.182) (0.138) (0.216) f0.024g
% Pop Two or More Races 0.027 0.024 0.030 0.006**
(0.018) (0.009) (0.024) f0.017g
% Pop White Alone Female 0.415 0.400 0.431 0.031***
(0.056) (0.058) (0.049) f0.000g
Observations 218 112 106 218
This table presents means and standard deviations (in parentheses) of MSA level variables for the full sample of
MSAs forming the balance analysis panel and for the MSAs divided according to the Craigslist Treatment and Control
groups. Column (4) reports the dierence in means as well as p-values for the test of equality of means between
treatment and control. Stars indicate statistical signicance at .1 (*), .05 (**), and .01 (***) levels.
78
Table 18: DID Estimate: Discontinuity in Rape / Sexual Assault post 2018
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4)
High Craigslist x Post 2.807

2.667

(0.898) (0.897)
High Backpage x Post 0.506 0.192
(0.923) (0.984)
High Me Too x Post -1.299 -0.868
(0.993) (1.051)
MSA Population (in 100,000s) -0.209 -0.593 -0.667 -0.374
(0.650) (0.825) (0.772) (0.682)
N MSAs 218 218 218 218
N Years 5 5 5 5
N Total 1090 1090 1090 1090
This table presents DID estimates based on a panel of US Metropolitan Statistical Areas from 2015 to 2019 where
treatment is dened as being an MSA with above median interest in each treatment (as proxied by Google search
frequency). The post period is dened as 2018-2019. Robust standard errors, clustered by Google Trends Metro Area,
are reported in parenthases. Regressions are weighted by MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
79
Table 19: Event Study Estimates: Discontinuity in Rape / Sexual Assault post 2018
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4)
High Craigslist x 2015 -0.210 -0.185
(1.147) (1.204)
High Craigslist x 2016 -0.371 -0.319
(0.884) (0.916)
High Craigslist x 2018 2.895

2.880

(1.158) (1.146)
High Craigslist x 2019 2.342

2.097

(1.129) (1.140)
High Backpage x 2015 1.316 1.153
(1.091) (1.098)
High Backpage x 2016 0.689 0.675
(0.774) (0.862)
High Backpage x 2018 0.356 0.370
(1.000) (1.079)
High Backpage x 2019 2.020

1.259
(1.054) (1.091)
High Me Too x 2015 -0.913 -0.556
(1.144) (1.172)
High Me Too x 2016 -0.237 -0.0308
(0.775) (0.881)
High Me Too x 2018 -0.469 0.0199
(0.991) (1.055)
High Me Too x 2019 -2.926

-2.163

(1.075) (1.127)
P-val for Joint Test of Craigslit Pre 0.893 0.924
P-val for Joint Test of Craigslit Post 0.029 0.035
P-val for Joint Test of Backpage Pre 0.485 0.577
P-val for Joint Test of Backpage Post 0.086 0.471
P-val for Joint Test of Me Too Pre 0.671 0.807
P-val for Joint Test of Me Too Post 0.011 0.086
N MSAs 218 218 218 218
N Years 5 5 5 5
N Total 1090 1090 1090 1090
This table presents Event Study estimates based on a panel of US Metropolitan Statistical Areas from 2015 to
2019 where treatment is dened as being an MSA with above median interest in each treatment (as proxied by
Google search frequency). The post period is considered 2018-2019. Regressions control for MSA population. Robust
standard errors, clustered by Google Trends Metro Area, are reported in parenthases. Regressions are weighted by
MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
80
Figure 16
Figure 17
Figure 18
81
Figure 19
Table 20: DID Estimates: Impact of FOSTA/SESTA on Prostitution Arrests
Total Female Juvenile Female Juv. Male Juv.
(1) (2) (3) (4) (5)
High Craigslist x Post 0.142 0.114 0.00879

0.00680

0.00200
(0.133) (0.0973) (0.00461) (0.00365) (0.00188)
High Backpage x Post -0.0958 -0.0689 0.00153 0.00230 -0.000774
(0.141) (0.107) (0.00427) (0.00357) (0.00158)
Population (100,000s) -1.724

-1.252

-0.0382

-0.0363

-0.00193
(0.376) (0.397) (0.00907) (0.00882) (0.00134)
N Agencies 6600 6600 6600 6600 6600
N Months 84 84 84 84 84
N Total 399468 399468 399468 399468 399468
This table presents DID estimates based on a monthly panel of US Municipal Police Agencies from 2012 to
2018 where treatment is dened as being an agency in a Google Trends Metro Area with above median usage of
Craigslist/Backpage in the year prior to implementation of FOSTA/SESTA legislation. The post period is dened as
2018 Q2-Q4. Robust standard errors, clustered by Google Trends Metro Area, are reported in parenthases. Regres-
sions are weighted by population covered by police agency.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
82
Table 21: Event Study Estimates: Juvenile Prostitution Arrests
Full Panel Quarters 2-4 Only
Juvenile Female Juv. Juvenile Female Juv.
(1) (2) (3) (4)
High Craigslist x 2012 -0.00296 -0.00223 -0.00212 -0.00206
(0.00631) (0.00637) (0.00660) (0.00659)
High Craigslist x 2013 0.00129 0.00354 0.000366 0.00318
(0.00933) (0.00871) (0.0108) (0.0102)
High Craigslist x 2014 -0.00255 -0.000787 -0.00162 0.0000934
(0.00711) (0.00664) (0.00715) (0.00657)
High Craigslist x 2015 -0.00404 -0.00211 -0.00274 -0.00170
(0.00538) (0.00534) (0.00683) (0.00690)
High Craigslist x 2016 0.000577 -0.000344 -0.00110 0.0000178
(0.00587) (0.00515) (0.00559) (0.00506)
High Craigslist x 2018 0.00626 0.00515 0.00744 0.00652

(0.00420) (0.00313) (0.00464) (0.00348)
Population (100,000s) -0.0382

-0.0363

-0.0385

-0.0376

(0.00880) (0.00860) (0.00978) (0.00987)
Conditional on High Backpage x Years Yes Yes Yes Yes
P-val Joint Test of Pre-Periods Craigslist 0.642 0.804 0.988 0.947
P-val Joint Test of Pre-Periods Backpage 0.246 0.108 0.800 0.362
N Agencies 6600 6600 6600 6600
N Months 84 84 77 77
N Total 399468 399468 299601 299601
This table presents Event Study estimates based on a monthly panel of US Municipal Police Agencies from 2012
to 2018 where treatment is dened as being an agency in a Google Trends Metro Area with above median usage
of Craigslist in the year prior to implementation of FOSTA/SESTA legislation. The post period is dened as 2018.
Robust standard errors, clustered byGoogle Trends Metro Area, are reported in parenthases. Regressions are weighted
by population covered by police agency. In columns 3 & 4 data for quarter 1 of each year is dropped to remove dilution
of treatment eect which which begins Q2 of 2018.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
83
Figure 20: Trends in prostitution arrests re
ecting observations from quarters 2-4 of each year.
84
Figure 21
85
Appendix A: In
uence of Perceived Comparative Need
A.1 Supplementary Figures
Figure A.1: Geographic Distribution of University Locations
Figure A.2: Photos from University Travel Programs:
Organic Coee Farming in Morelos, MX (left)
Immigration and Border Issues in Tucscon, AZ - Nogales, MX (center)
History and Cultural Heritage of Navajo Nation (right)
86
A.1 Supplementary Figures
Figure A.3: Study Flow and Data Sources
Figure A.4: Information Intervention Graphic 1 - EFC Distribution
87
A.1 Supplementary Figures
Figure A.5: Information Intervention Graphic 2 - Percentiles of Need
Figure A.6: Graphic to Elicit Views on Fair Distribution of Aid
88
A.2 Predictors of Aid Requests for Experimental Sample
Table A.1: Predictors of Applied for Aid
(1) (2) (3) (4) (5)
Total Savings (Thousands) -0.0128

-0.00356
(0.00432) (0.00450)
Weekly Earnings (Hundreds) 0.00869 0.00625
(0.00857) (0.00811)
EFC (Tens of Thousands) -0.0201

-0.00853

(0.00456) (0.00326)
No FAFSA -0.420

-0.206

(0.0415) (0.0568)
Filed FAFSA as an Independent Adult 0.0158 -0.00282
(0.0790) (0.0781)
Female Student -0.0299
(0.0546)
Ethnicity: Black / African American 0.338

0.252

(0.0714) (0.0628)
Ethnicity: Hispanic / Latino 0.135

0.0882
(0.0642) (0.0573)
Ethnicity: Other (Non-White) 0.184

0.169

(0.0456) (0.0455)
First Generation College Student 0.150

0.00435
(0.0435) (0.0454)
Mostly Associates with Lower SES Peers 0.173

0.127

(0.0661) (0.0642)
Grade Point Average -0.104

(0.0575)
BIDR Impression Management Scale 0.00281 0.00140
(0.00225) (0.00211)
Estimated EFC (Tens of Thousands) -0.0129

-0.00931

(0.00382) (0.00412)
Perceived Percentile of Need 0.00548

0.00312

(0.000641) (0.000834)
Perceived Program Resources Scale 0.0130

0.0125
(0.00786) (0.00785)
Surplus Aid Should Go to Highest Need -0.0546 -0.0523
(0.0365) (0.0362)
Surplus Aid Should Be Returned to Univ 0.0235 0.0154
(0.0345) (0.0340)
Student Cost (Hundreds) 0.00220 0.00836
(0.00934) (0.00690)
International Program 0.298

0.152
(0.137) (0.103)
Flying to Destination 0.0280
(0.114)
LASSO Selected Covariates No No No No Yes
Control Mean 0.36 0.36 0.36 0.36 0.36
Observations (N) 616 616 616 616 616
Robust Standard Errors are reported in parentheses. Regressions are conditional on University Fixed Eects. The
omitted category for FAFSA groupings is Filed FAFSA with Guardians (and reported actual EFC). The omitted
category for Ethnicity is White. The omitted category for Mostly Associates with Lower SES is Associates with All
SES. Mostly Associates with Middle and Upper SES are both included, but are not displayed (both are insignicant).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
89
A.2 Predictors of Aid Requests for Experimental Sample
Table A.2: Predictors of Amount of Aid Requested (as % of Student Fee)
(1) (2) (3) (4) (5)
Total Savings (Thousands) -0.00974

-0.00213
(0.00220) (0.00194)
Weekly Earnings (Hundreds) 0.00231
(0.00622)
EFC (Tens of Thousands) -0.0138

-0.00298
(0.00336) (0.00202)
No FAFSA -0.240

-0.0364
(0.0233) (0.0277)
Filed FAFSA as an Independent Adult 0.104

0.0824
(0.0612) (0.0585)
Female Student -0.0376 -0.0352
(0.0347) (0.0294)
Ethnicity: Black / African American 0.193

0.134

(0.0497) (0.0433)
Ethnicity: Hispanic / Latino 0.113

0.0756

(0.0453) (0.0384)
Ethnicity: Other (Non-White) 0.108

0.0869

(0.0295) (0.0282)
First Generation College Student 0.142

0.0440
(0.0294) (0.0282)
Mostly Associates with Lower SES Peers 0.117

0.0795

(0.0487) (0.0459)
Grade Point Average -0.0460
(0.0378)
BIDR Impression Management Scale 0.00296

0.00143
(0.00145) (0.00129)
Estimated EFC (Tens of Thousands) -0.00721

-0.00656

(0.00191) (0.00209)
Perceived Percentile of Need 0.00427

0.00307

(0.000420) (0.000522)
Perceived Program Resources Scale 0.00836

0.00818

(0.00504) (0.00485)
Surplus Aid Should Go to Highest Need 0.0238 0.0220
(0.0236) (0.0225)
Surplus Aid Should Be Returned to Univ 0.0130 0.0112
(0.0210) (0.0203)
Student Cost (Hundreds) 0.00208 0.0103

(0.00786) (0.00538)
International Program 0.242

0.155

(0.0950) (0.0718)
Flying to Destination -0.0670
(0.0744)
LASSO Selected Covariates No No No No Yes
Control Mean 0.17 0.17 0.17 0.17 0.17
Observations (N) 616 616 616 616 616
Robust Standard Errors are reported in parentheses. Regressions are conditional on University Fixed Eects. The
omitted category for FAFSA groupings is Filed FAFSA with Guardians (and reported actual EFC). The omitted
category for Ethnicity is White. The omitted category for Mostly Associates with Lower SES is Associates with All
SES. Mostly Associates with Middle and Upper SES are both included, but are not displayed (both are insignicant).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
90
A.3 Estimation of Additional Treatment Eect Specications
A.3.1 Full Interaction of Bias Groups with Terciles of EFC
Figure A.7: Treatment Eects Associated with Table A.3
Figure A.8: Treatment Eects Associated with Table A.4
91
A.3.1 Full Interaction of Bias Groups with Terciles of EFC
Table A.3: Treatment Eects on Applying for Aid
Underestimates Low Bias Overestimates
Low EFC 0.159 -0.053 -0.121
(0.098) (0.103) (0.131)
Medium EFC 0.127 0.197

0.093
(0.098) (0.120) (0.107)
High EFC -0.137 -0.088 -0.139
(0.096) (0.071) (0.094)
Observations Low EFC 94 57 55
Observations Med EFC 89 56 60
Observations High EFC 28 92 85
This table presents estimates of from a single regression including interactions of each tercile of EFC and Bias with
one another and with an indicator for treatment assignment. Coecients re
ect estimates of the average treatment
eect of information provision within each EFC tercile - Bias tercile pair. Robust standard errors are reported in
parenthases. Estimates are conditional on university xed eects and on baseline covariates determined by LASSO
double selection. Sample excludes students who couldn't nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table A.4: Treatment Eects on Amount Requested (as % of Student Fee)
Underestimates Low Bias Overestimates
Low EFC 0.119

0.077 -0.004
(0.059) (0.093) (0.100)
Medium EFC 0.047 0.146

0.004
(0.037) (0.076) (0.069)
High EFC -0.007 -0.011 -0.077
(0.027) (0.038) (0.048)
Observations Low EFC 94 57 55
Observations Med EFC 89 56 60
Observations High EFC 28 92 85
This table presents estimates of from a single regression including interactions of each tercile of EFC and Bias with
one another and with an indicator for treatment assignment. Coecients re
ect estimates of the average treatment
eect of information provision within each EFC tercile - Bias tercile pair. Robust standard errors are reported in
parenthases. Estimates are conditional on university xed eects and on baseline covariates determined by LASSO
double selection. Sample excludes students who couldn't nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
92
A.3.2 Heterogeneous Eects On Applying for Aid
Table A.5: Heterogeneous Treatment Eects on Applied for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Separate Regressions with each Variable Z: (1) (2) (3) (4) (5) (6)
Treated 0.168

0.176

0.112 0.159

0.157

0.127
(0.0826) (0.0853) (0.0905) (0.0874) (0.0886) (0.0873)
Treated x Z -0.112 -0.0739 0.0728 -0.0284 -0.0202 0.0557
(0.0934) (0.0847) (0.0836) (0.0839) (0.0868) (0.0909)
Treated x Overestimates -0.107 -0.107 -0.0525 -0.121 -0.134 -0.0584
(0.0900) (0.0919) (0.103) (0.101) (0.111) (0.0989)
Treated x Overestimates x Z -0.0187 -0.0225 -0.0910 0.00859 0.0492 -0.133
(0.166) (0.157) (0.152) (0.153) (0.154) (0.145)
I. Low Bias/Under & Z = Overestimate & Z -0.126 -0.129 -0.143 -0.112 -0.084 -0.192

(Tr x Over + Tr x Over x Z) f0.356g f0.308g f0.190g f0.323g f0.415g f0.068g
II. Overestimates not Z = Overestimate & Z -0.131 -0.096 -0.018 -0.020 0.029 -0.078
(Tr x Z + Tr x Over x Z) f0.332g f0.465g f0.885g f0.876g f0.816g f0.512g
Control Group Mean 0.36 0.36 0.36 0.36 0.36 0.36
Observations Low Bias/Under x NOT Z 309 259 219 236 231 213
Observations Low Bias/Under x Z 107 157 197 180 185 203
Observations Overest x NOT Z 150 140 116 108 89 74
Observations Overest x Z 50 60 84 92 111 126
Observations (Total) 616 616 616 616 616 616
This table presents coecients from 6 regressions (columns) where the listed variable is interacted with treatment
and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors are reported in
parenthases and p-values for statistical tests in brackets. All regressions are conditional on university xed eects,
covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC. Overestimates
& Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not interacted with
Overestimates or Underestimates represent eects for students with low bias. Sample excludes students who couldn't
nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
93
A.3.2 Heterogeneous Eects On Applying for Aid
Table A.6: Heterogeneous Treatment Eects on Applying for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
A Single Regression with all Variables Z: (1) (2) (3) (4) (5) (6)
Treated 0.163 0.163 0.163 0.163 0.163 0.163
(0.112) (0.112) (0.112) (0.112) (0.112) (0.112)
Treated x Z -0.097 -0.051 0.039 -0.002 -0.051 0.052
(0.095) (0.085) (0.085) (0.086) (0.088) (0.096)
Treated x Overestimates 0.040 0.040 0.040 0.040 0.040 0.040
(0.170) (0.170) (0.170) (0.170) (0.170) (0.170)
Treated x Overestimates x Z 0.003 -0.097 -0.104 -0.039 0.056 -0.175
(0.168) (0.157) (0.150) (0.155) (0.156) (0.155)
I. Low Bias/Under & Z = Overestimate & Z 0.044 -0.057 -0.063 0.002 0.097 -0.135
(Tr x Over + Tr x Over x Z) f0.826g f0.772g f0.717g f0.993g f0.594g f0.487g
II. Overestimates not Z = Overestimate & Z -0.094 -0.148 -0.064 -0.041 0.005 -0.123
(Tr x Z + Tr x Over x Z) f0.492g f0.264g f0.601g f0.748g f0.967g f0.332g
Control Group Mean 0.36 0.36 0.36 0.36 0.36 0.36
Observations Low Bias/Under x NOT Z 309 259 219 236 231 213
Observations Low Bias/Under x Z 107 157 197 180 185 203
Observations Overest x NOT Z 150 140 116 108 89 74
Observations Overest x Z 50 60 84 92 111 126
Observations (N) 616 616 616 616 616 616
This table presents coecients from a single regression including the interaction of all listed variables (columns)
with treatment and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors
are reported in parenthases and p-values for statistical tests in brackets. The regression is conditional on university
xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC.
Overestimates & Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not
interacted with Overestimates or Underestimates represent eects for students with low bias. Sample excludes
students who couldn't nd or choose not to report their actual EFC.

p < 0:1,

p < 0:05,

p < 0:01,

p< 0:001
94
A.3.3 Heterogeneous Eects Using Binary Split of Positive/Negative Bias
Table A.7: Heterogeneous Treatment Eects on Reported WTP (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Panel A: Separate Regression Results (1) (2) (3) (4) (5) (6)
Treated 0.0559 0.0726 0.0182 0.0585 0.0189 0.0146
(0.0448) (0.0482) (0.0489) (0.0452) (0.0503) (0.0431)
Treated x Overestimates -0.000271 -0.0509 -0.0115 -0.0636 0.00871 -0.0477
(0.0499) (0.0540) (0.0565) (0.0568) (0.0555) (0.0488)
Treated x Z -0.0353 -0.0630 0.0598 -0.0292 0.0765 0.0799
(0.0621) (0.0561) (0.0558) (0.0563) (0.0573) (0.0599)
Treated x Overestimates x Z -0.105 0.0491 -0.0195 0.0714 -0.0894 -0.0139
(0.0977) (0.0847) (0.0845) (0.0854) (0.0823) (0.0820)
Overestimates not Z = Overestimates & Z f0.060g f0.823g f0.521g f0.498g f0.830g f0.256g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.213g f0.979g f0.626g f0.906g f0.208g f0.369g
(Tr x Overestimates + Tr x Overestimates x Z)
Panel B: Single Regression (all Z Variables) (1) (2) (3) (4) (5) (6)
Treated 0.019 0.019 0.019 0.019 0.019 0.019
(0.064) (0.064) (0.064) (0.064) (0.064) (0.064)
Treated x Overestimates 0.011 0.011 0.011 0.011 0.011 0.011
(0.083) (0.083) (0.083) (0.083) (0.083) (0.083)
Treated x Z -0.052 -0.033 0.032 -0.031 0.040 0.072
(0.065) (0.058) (0.057) (0.057) (0.057) (0.063)
Treated x Overestimates x Z -0.056 -0.015 0.006 0.063 -0.100 -0.018
(0.102) (0.087) (0.086) (0.087) (0.085) (0.086)
Overestimates not Z = Overestimates & Z f0.165g f0.456g f0.558g f0.618g f0.335g f0.385g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.684g f0.969g f0.855g f0.462g f0.373g f0.946g
(Tr x Overestimates + Tr x Overestimates x Z)
Control Group Mean 0.17 0.17 0.17 0.17 0.17 0.17
Observations Underest x NOT Z 239 202 168 177 172 170
Observations Underest x Z 78 115 149 140 145 147
Observations Overest x NOT Z 220 197 167 167 148 117
Observations Overest x Z 79 102 132 132 151 182
Observations (Total) 616 616 616 616 616 616
Panel A presents results from a separate regression including interactions with each heterogeneous variable. Panel B
presents results from a single regression where all heterogeneous variable interactions are included. Robust standard
errors are reported in parenthases and p-values for statistical tests in brackets. All regressions are conditional on
university xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of
EFC. Overestimates is a dummy equal to 1 for students with positive bias, thus terms not interacted with Overesti-
mates represent eects for students who underestimate their relative nancial need. Sample excludes students who
couldn't nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
95
A.3.3 Heterogeneous Eects Using Binary Split of Positive/Negative Bias
Table A.8: Heterogeneous Treatment Eects on Applied for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Panel A: Separate Regression Results (1) (2) (3) (4) (5) (6)
Treated 0.143

0.201

0.105 0.179

0.130 0.117
(0.0867) (0.0889) (0.0959) (0.0920) (0.0930) (0.0907)
Treated x Overestimates -0.0234 -0.147

-0.0253 -0.146 -0.0248 -0.0168
(0.0826) (0.0869) (0.0964) (0.0928) (0.0954) (0.0929)
Treated x Z -0.0900 -0.186

0.0587 -0.124 0.0322 0.0435
(0.112) (0.101) (0.0993) (0.0988) (0.102) (0.107)
Treated x Overestimates x Z -0.0651 0.264

-0.0353 0.215 -0.0648 -0.0852
(0.154) (0.143) (0.141) (0.140) (0.139) (0.138)
Overestimates not Z = Overestimates & Z f0.141g f0.436g f0.812g f0.355g f0.733g f0.665g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.490g f0.300g f0.545g f0.509g f0.376g f0.318g
(Tr x Overestimates + Tr x Overestimates x Z)
Panel B: Single Regression (all Z Variables) (1) (2) (3) (4) (5) (6)
Treated 0.223

0.223

0.223

0.223

0.223

0.223

(0.123) (0.123) (0.123) (0.123) (0.123) (0.123)
Treated x Overestimates -0.103

-0.103

-0.103

-0.103

-0.103

-0.103

(0.155) (0.155) (0.155) (0.155) (0.155) (0.155)
Treated x Z -0.089 -0.161 0.021 -0.113 -0.015 0.055
(0.114) (0.102) (0.100) (0.101) (0.102) (0.114)
Treated x Overestimates x Z -0.003 0.227 -0.028 0.204 -0.049 -0.135
(0.158) (0.145) (0.141) (0.141) (0.141) (0.146)
Overestimates not Z = Overestimates & Z f0.401g f0.517g f0.943g f0.353g f0.509g f0.438g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.558g f0.475g f0.417g f0.554g f0.375g f0.169g
(Tr x Overestimates + Tr x Overestimates x Z)
Control Group Mean 0.36 0.36 0.36 0.36 0.36 0.36
Observations Underest x NOT Z 239 202 168 177 172 170
Observations Underest x Z 78 115 149 140 145 147
Observations Overest x NOT Z 220 197 167 167 148 117
Observations Overest x Z 79 102 132 132 151 182
Observations (Total) 616 616 616 616 616 616
Panel A presents results from a separate regression including interactions with each heterogeneous variable. Panel B
presents results from a single regression where all heterogeneous variable interactions are included. Robust standard
errors are reported in parenthases and p-values for statistical tests in brackets. All regressions are conditional on
university xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of
EFC. Overestimates is a dummy equal to 1 for students with positive bias, thus terms not interacted with Overesti-
mates represent eects for students who underestimate their relative nancial need. Sample excludes students who
couldn't nd or choose not to report their actual EFC.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
96
A.4 Full Sample Results
Table A.9: Baseline Summary Statistics and Balance
Means and Standard Deviations Dierence
Variable Full Sample Control Treatment Treat vs Cont.
(1) (2) (3) (4)
Total Savings 2,474.955 2,312.530 2,646.134 246.406
(3,603.439) (3,410.073) (3,792.484) f0.268g
Weekly Earnings 101.093 102.666 99.435 -5.257
(236.340) (245.642) (226.356) f0.724g
Expected Family Contribution 14,067.063 13,318.533 14,847.615 904.987
(30,368.561) (28,260.033) (32,479.914) f0.751g
Estimated EFC 39,399.895 38,417.582 40,435.145 1,964.942
(51,253.641) (51,009.707) (51,540.168) f0.522g
No FAFSA: Low Need 0.128 0.136 0.120 -0.010
(0.335) (0.343) (0.325) f0.642g
No FAFSA: Ineligible/Didn't Know 0.050 0.047 0.052 0.009
(0.217) (0.213) (0.222) f0.520g
Unsure if Filed FAFSA 0.050 0.055 0.044 -0.010
(0.217) (0.228) (0.205) f0.444g
Yes FAFSA: With Guardians 0.714 0.697 0.733 0.023
(0.452) (0.460) (0.443) f0.414g
Yes FAFSA: Independent 0.058 0.064 0.052 -0.012
(0.234) (0.246) (0.222) f0.430g
Actual Percentile of Need 45.677 46.134 45.196 -1.062
(28.845) (28.781) (28.933) f0.553g
Perceived Percentile of Need 44.587 44.140 45.058 1.100
(28.314) (28.246) (28.406) f0.525g
Program Student Fee 366.934 359.498 374.770 -2.065
(367.115) (367.813) (366.582) f0.888g
Female Student 0.577 0.562 0.593 0.018
(0.494) (0.497) (0.492) f0.398g
Male Student 0.142 0.150 0.134 -0.015
(0.349) (0.357) (0.341) f0.495g
Ethnicity: White 0.568 0.542 0.595 0.025
(0.496) (0.499) (0.491) f0.382g
Ethnicity: Black / African American 0.078 0.089 0.066 -0.021
(0.268) (0.285) (0.248) f0.197g
Ethnicity: Other (Non-White) 0.223 0.237 0.208 -0.006
(0.416) (0.425) (0.406) f0.818g
Ethnicity: Hispanic / Latino 0.132 0.133 0.132 0.002
(0.339) (0.339) (0.339) f0.928g
First Generation College Student 0.289 0.294 0.283 -0.003
(0.453) (0.456) (0.451) f0.926g
GPA 3.528 3.492 3.564 0.073**
(0.419) (0.453) (0.379) f0.044g
Test of Overall Signicance F(42,1028) 1.05
(including all (27) controls) p-value f 0.38 g
Observations 1,029 528 501 1,029
This table presents sample means and standard deviations (in parentheses) of variables at baseline for the full sample
(1) and for the control (2) and treatment (3) groups. Column (4) reports coecients from the regression of the control
variable on a treatment indicator and university xed eects. P-values for the coecients are indicated in brackets
based on heteroskedasticity robust standard errors. Stars indicate a statistically signicant coecient at .1 (*), .05
(**), and .01 (***) levels of signicance. The F-statistic reported is for the overall test of joint signicance of all
controls in a regression of the treatment dummy on all control variables.
97
A.4 Full Sample Results
Table A.10: Marginal Treatment Eect along Bias in Prior on Comparative Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated -0.00461 -0.00699 -0.0179 0.0331 0.000838 -0.00438
(0.0622) (0.0754) (0.0898) (0.0383) (0.0411) (0.0499)
Treated x Overestimates 0.0616 0.0643 0.0374 0.0215 0.0183 -0.000531
(0.0882) (0.0782) (0.0771) (0.0566) (0.0484) (0.0474)
Treated x Underestimates x Bias 0.00165 0.000957 0.00146 -0.000263 -0.000443 -0.0000784
(0.00228) (0.00225) (0.00218) (0.00135) (0.00131) (0.00129)
Treated x Overestimates x Bias -0.00324 -0.00231 -0.00207 -0.00284

-0.00225 -0.00275

(0.00256) (0.00242) (0.00239) (0.00168) (0.00154) (0.00154)
Controls N Y Y N Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
Tr x Und x Bias = - Tr x Ov x Bias f0.641g f0.675g f0.844g f0.148g f0.172g f0.144g
Control Group Mean 0.34 0.34 0.34 0.16 0.16 0.16
Observations (N) 1029 1029 1029 1029 1029 1029
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects. Sample
includes all baseline survey respondents. Columns 1 & 4 are otherwise unconditional, while Columns 2, 3, 5, & 6
control both for covariates determined by LASSO double selection and, as indicated, for treatment interacted with
quintiles of EFC and other baseline correlates of bias in perceived need. The Treat coecient should, therefore, be
interpreted as the CATE for middle EFC, low bias students in (2) and (5) and for middle EFC, low bias, middle
savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
98
A.4 Full Sample Results
Table A.11: Dierential Treatment Eect: Over/Under Estimators of Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated 0.0291 0.0165 0.00245 0.0269 -0.00749 -0.0123
(0.0394) (0.0600) (0.0784) (0.0223) (0.0290) (0.0408)
Treated x Overestimates -0.0329 0.0161 -0.0229 -0.0256 -0.00795 -0.0422
(0.0556) (0.0529) (0.0545) (0.0337) (0.0336) (0.0331)
Controls N Y Y N Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
Pvalue: Treat + Tr x Overest f0.922g f0.614g f0.801g f0.958g f0.672g f0.222g
Control Group Mean 0.34 0.34 0.34 0.16 0.16 0.16
Observations (N) 1029 1029 1029 1029 1029 1029
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects. Sample
includes all baseline survey respondents. Columns 1 & 4 are otherwise unconditional, while Columns 2, 3, 5, & 6
control both for covariates determined by LASSO double selection and, as indicated, for treatment interacted with
quintiles of EFC and other baseline correlates of bias in perceived need. The Treat coecient should, therefore, be
interpreted as the CATE for middle EFC, low bias students in (2) and (5) and for middle EFC, low bias, middle
savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table A.12: Dierential Treatment Eect: Over/Under Estimators of Need
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4) (5) (6)
Treated 0.0122 0.00209 -0.0490 0.0435 0.00510 -0.0352
(0.0418) (0.0696) (0.0840) (0.0280) (0.0384) (0.0462)
Treated x Underestimates 0.0455 0.0341 0.0618 -0.00769 -0.00301 0.0184
(0.0609) (0.0636) (0.0643) (0.0364) (0.0389) (0.0389)
Treated x Overestimates -0.0215 0.0186 0.00114 -0.0740

-0.0493 -0.0710

(0.0616) (0.0629) (0.0622) (0.0397) (0.0395) (0.0390)
Controls Y Y Y Y Y Y
Treatment x EFC Quintile N Y Y N Y Y
Treatment x Correlates of Bias N N Y N N Y
P-value: Tr x Under = Tr x Over f0.291g f0.822g f0.391g f0.067g f0.275g f0.036g
Control Group Mean 0.34 0.34 0.34 0.16 0.16 0.16
Observations (N) 1029 1029 1029 1029 1029 1029
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects. Sample
includes all baseline survey respondents. Columns 1 & 4 are otherwise unconditional, while Columns 2, 3, 5, & 6
control both for covariates determined by LASSO double selection and, as indicated, for treatment interacted with
quintiles of EFC and other baseline correlates of bias in perceived need. The Treat coecient should, therefore, be
interpreted as the CATE for middle EFC, low bias students in (2) and (5) and for middle EFC, low bias, middle
savings, white, non-rst generation, non-working students in (3) and (6).

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
99
A.4 Full Sample Results
Figure A.9: Treatment Eects Associated with Table A.12
Figure A.10: Treatment Eects Associated with Table A.12
100
A.4 Full Sample Results
Table A.13: Overall Average Treatment Eects
Applied for Aid Request (% of Student Fee)
(1) (2) (3) (4)
Treated 0.0147 0.0221 0.0157 0.0166
(0.0279) (0.0253) (0.0170) (0.0153)
90% Condence Interval (-0.03 , 0.06) (-0.02 , 0.06) (-0.01 , 0.04) (-0.01 , 0.04)
Controls N Y N Y
Control Group Mean 0.34 0.34 0.16 0.16
Observations (N) 1029 1029 1029 1029
Robust standard errors are reported in parenthases. All regressions are conditional on university xed eects. Sample
includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
101
A.4 Full Sample Results
Table A.14: Heterogeneous Treatment Eects on Amount of Aid Requested (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Separate Regressions with each Variable Z: (1) (2) (3) (4) (5) (6)
Treated 0.00775 0.0212 -0.00674 -0.00523 0.0205 -0.00828
(0.0297) (0.0310) (0.0322) (0.0304) (0.0322) (0.0294)
Treated x Z -0.0136 -0.0447 0.0235 0.0115 -0.0381 0.0392
(0.0425) (0.0374) (0.0357) (0.0358) (0.0365) (0.0366)
Treated x Overestimates -0.0229 -0.0361 -0.0664 -0.0125 -0.0839

-0.0357
(0.0400) (0.0421) (0.0448) (0.0476) (0.0490) (0.0421)
Treated x Overestimates x Z -0.0978 -0.0521 0.0509 -0.0848 0.0729 -0.0466
(0.0825) (0.0728) (0.0685) (0.0671) (0.0676) (0.0650)
I. Low Bias/Under & Z = Overestimate & Z -0.121 -0.088 -0.016 -0.097

-0.011 -0.082
(Tr x Over + Tr x Over x Z) f0.102g f0.157g f0.776g f0.057g f0.827g f0.120g
II. Overestimates not Z = Overestimate & Z -0.111 -0.097 0.074 -0.073 0.035 -0.007
(Tr x Z + Tr x Over x Z) f0.115g f0.122g f0.202g f0.199g f0.535g f0.891g
Control Group Mean 0.16 0.16 0.16 0.16 0.16 0.16
Observations Low Bias/Under x NOT Z 512 451 368 382 370 344
Observations Low Bias/Under x Z 182 243 326 312 324 350
Observations Overest x NOT Z 249 239 183 180 154 141
Observations Overest x Z 86 96 152 155 181 194
Observations (Total) 1029 1029 1029 1029 1029 1029
This table presents coecients from 6 regressions (columns) where the listed variable is interacted with treatment
and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors are reported in
parenthases and p-values for statistical tests in brackets. All regressions are conditional on university xed eects,
covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC. Overestimates
& Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not interacted with
Overestimates or Underestimates represent eects for students with low bias. Sample includes all baseline survey
respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
102
A.4 Full Sample Results
Table A.15: Heterogeneous Treatment Eects on Amount of Aid Requested (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
A Single Regression with all Variables Z: (1) (2) (3) (4) (5) (6)
Treated 0.007 0.007 0.007 0.007 0.007 0.007
(0.040) (0.040) (0.040) (0.040) (0.040) (0.040)
Treated x Z -0.013 -0.042 0.031 0.015 -0.039 0.038
(0.043) (0.038) (0.036) (0.036) (0.037) (0.038)
Treated x Overestimates 0.028 0.028 0.028 0.028 0.028 0.028
(0.072) (0.072) (0.072) (0.072) (0.072) (0.072)
Treated x Overestimates x Z -0.098 -0.071 0.025 -0.100 0.049 -0.054
(0.083) (0.074) (0.069) (0.067) (0.067) (0.067)
I. Low Bias/Under & Z = Overestimate & Z -0.070 -0.043 0.053 -0.072 0.077 -0.026
(Tr x Over + Tr x Over x Z) f0.472g f0.602g f0.504g f0.360g f0.291g f0.749g
II. Overestimates not Z = Overestimate & Z -0.111 -0.113

0.056 -0.086 0.010 -0.016
(Tr x Z + Tr x Over x Z) f0.113g f0.077g f0.343g f0.135g f0.856g f0.773g
Control Group Mean 0.16 0.16 0.16 0.16 0.16 0.16
Observations Low Bias/Under x NOT Z 512 451 368 382 370 344
Observations Low Bias/Under x Z 182 243 326 312 324 350
Observations Overest x NOT Z 249 239 183 180 154 141
Observations Overest x Z 86 96 152 155 181 194
Observations (N) 1029 1029 1029 1029 1029 1029
This table presents coecients from a single regression including the interaction of all listed variables (columns)
with treatment and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors
are reported in parenthases and p-values for statistical tests in brackets. The regression is conditional on university
xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC.
Overestimates & Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not
interacted with Overestimates or Underestimates represent eects for students with low bias. Sample includes all
baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
103
A.4 Full Sample Results
Table A.16: Heterogeneous Treatment Eects on Applied for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Separate Regressions with each Variable Z: (1) (2) (3) (4) (5) (6)
Treated 0.0322 0.0523 -0.000695 0.0321 0.0848 0.0330
(0.0614) (0.0634) (0.0646) (0.0636) (0.0651) (0.0624)
Treated x Z -0.0280 -0.0775 0.0651 -0.0313 -0.135

-0.0166
(0.0682) (0.0637) (0.0616) (0.0612) (0.0629) (0.0672)
Treated x Overestimates 0.0231 -0.00183 -0.00825 0.0242 -0.0434 0.0296
(0.0678) (0.0696) (0.0789) (0.0779) (0.0831) (0.0755)
Treated x Overestimates x Z -0.0719 0.00335 0.0303 -0.0527 0.117 -0.0731
(0.124) (0.117) (0.111) (0.111) (0.112) (0.110)
I. Low Bias/Under & Z = Overestimate & Z -0.049 0.002 0.022 -0.028 0.073 -0.044
(Tr x Over + Tr x Over x Z) f0.644g f0.988g f0.785g f0.731g f0.352g f0.596g
II. Overestimates not Z = Overestimate & Z -0.100 -0.074 0.095 -0.084 -0.018 -0.090
(Tr x Z + Tr x Over x Z) f0.336g f0.452g f0.298g f0.365g f0.843g f0.317g
Control Group Mean 0.34 0.34 0.34 0.34 0.34 0.34
Observations Low Bias/Under x NOT Z 512 451 368 382 370 344
Observations Low Bias/Under x Z 182 243 326 312 324 350
Observations Overest x NOT Z 249 239 183 180 154 141
Observations Overest x Z 86 96 152 155 181 194
Observations (Total) 1029 1029 1029 1029 1029 1029
This table presents coecients from 6 regressions (columns) where the listed variable is interacted with treatment
and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors are reported in
parenthases and p-values for statistical tests in brackets. All regressions are conditional on university xed eects,
covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC. Overestimates
& Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not interacted with
Overestimates or Underestimates represent eects for students with low bias. Sample includes all baseline survey
respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
104
A.4 Full Sample Results
Table A.17: Heterogeneous Treatment Eects on Applying for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
A Single Regression with all Variables Z: (1) (2) (3) (4) (5) (6)
Treated 0.108 0.108 0.108 0.108 0.108 0.108
(0.083) (0.083) (0.083) (0.083) (0.083) (0.083)
Treated x Z -0.040 -0.076 0.068 -0.026 -0.126

-0.007
(0.070) (0.065) (0.063) (0.062) (0.065) (0.069)
Treated x Overestimates 0.069 0.069 0.069 0.069 0.069 0.069
(0.130) (0.130) (0.130) (0.130) (0.130) (0.130)
Treated x Overestimates x Z -0.044 -0.032 -0.028 -0.095 0.094 -0.078
(0.128) (0.120) (0.112) (0.112) (0.114) (0.113)
I. Low Bias/Under & Z = Overestimate & Z 0.025 0.037 0.040 -0.027 0.162 -0.009
(Tr x Over + Tr x Over x Z) f0.874g f0.796g f0.769g f0.844g f0.213g f0.949g
II. Overestimates not Z = Overestimate & Z -0.084 -0.108 0.040 -0.121 -0.032 -0.085
(Tr x Z + Tr x Over x Z) f0.432g f0.284g f0.665g f0.192g f0.732g f0.356g
Control Group Mean 0.34 0.34 0.34 0.34 0.34 0.34
Observations Low Bias/Under x NOT Z 512 451 368 382 370 344
Observations Low Bias/Under x Z 182 243 326 312 324 350
Observations Overest x NOT Z 249 239 183 180 154 141
Observations Overest x Z 86 96 152 155 181 194
Observations (N) 1029 1029 1029 1029 1029 1029
This table presents coecients from a single regression including the interaction of all listed variables (columns)
with treatment and terciles of ex-ante bias in perceived need according to Specication 6. Robust standard errors
are reported in parenthases and p-values for statistical tests in brackets. The regression is conditional on university
xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of EFC.
Overestimates & Underestimates are dummies equal to 1 for the rst & third quintiles of bias, thus, terms not
interacted with Overestimates or Underestimates represent eects for students with low bias. Sample includes all
baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
105
A.4 Full Sample Results
Table A.18: Heterogeneous Treatment Eects on Reported WTP (as % of Student Fee)
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Panel A: Separate Regression Results (1) (2) (3) (4) (5) (6)
Treated -0.0112 0.0138 -0.0315 -0.00321 -0.00301 -0.00951
(0.0309) (0.0329) (0.0341) (0.0313) (0.0333) (0.0311)
Treated x Overestimates 0.0246 -0.0128 0.00875 -0.00928 -0.00729 -0.00974
(0.0370) (0.0401) (0.0422) (0.0437) (0.0431) (0.0372)
Treated x Z 0.0224 -0.0565 0.0495 -0.0124 -0.0117 0.0194
(0.0491) (0.0428) (0.0414) (0.0416) (0.0422) (0.0432)
Treated x Overestimates x Z -0.122

0.0110 -0.0274 0.00279 0.00156 -0.0105
(0.0734) (0.0649) (0.0617) (0.0617) (0.0614) (0.0598)
Overestimates not Z = Overestimates & Z f0.066g f0.350g f0.626g f0.834g f0.819g f0.834g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.139g f0.973g f0.702g f0.892g f0.906g f0.693g
(Tr x Overestimates + Tr x Overestimates x Z)
Panel B: Single Regression (all Z Variables) (1) (2) (3) (4) (5) (6)
Treated -0.007 -0.007 -0.007 -0.007 -0.007 -0.007
(0.045) (0.045) (0.045) (0.045) (0.045) (0.045)
Treated x Overestimates 0.054 0.054 0.054 0.054 0.054 0.054
(0.061) (0.061) (0.061) (0.061) (0.061) (0.061)
Treated x Z 0.011 -0.052 0.054 -0.009 -0.012 0.016
(0.049) (0.043) (0.042) (0.041) (0.042) (0.044)
Treated x Overestimates x Z -0.096 -0.009 -0.040 -0.003 -0.021 -0.009
(0.074) (0.065) (0.062) (0.061) (0.061) (0.062)
Overestimates not Z = Overestimates & Z f0.130g f0.214g f0.747g f0.782g f0.463g f0.872g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.624g f0.550g f0.831g f0.466g f0.624g f0.538g
(Tr x Overestimates + Tr x Overestimates x Z)
Control Group Mean 0.16 0.16 0.16 0.16 0.16 0.16
Observations Underest x NOT Z 387 338 280 287 273 268
Observations Underest x Z 129 178 236 229 243 248
Observations Overest x NOT Z 374 352 271 275 251 217
Observations Overest x Z 139 161 242 238 262 296
Observations (Total) 1029 1029 1029 1029 1029 1029
Panel A presents results from a separate regression including interactions with each heterogeneous variable. Panel B
presents results from a single regression where all heterogeneous variable interactions are included. Robust standard
errors are reported in parenthases and p-values for statistical tests in brackets. All regressions are conditional on
university xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of
EFC. Overestimates is a dummy equal to 1 for students with positive bias, thus terms not interacted with Overes-
timates represent eects for students who underestimate their relative nancial need. Sample includes all baseline
survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
106
A.4 Full Sample Results
Table A.19: Heterogeneous Treatment Eects on Applied for Aid
Surplus to Return Program High
Highest Surplus Well Volunteers Impression Low
Need to Uni Resourced Management Savings
Panel A: Separate Regression Results (1) (2) (3) (4) (5) (6)
Treated 0.00829 0.0631 -0.0172 0.0710 0.0683 0.0166
(0.0641) (0.0661) (0.0689) (0.0665) (0.0679) (0.0648)
Treated x Overestimates 0.0609 -0.0331 0.0319 -0.0623 -0.00221 0.0677
(0.0625) (0.0647) (0.0724) (0.0710) (0.0723) (0.0697)
Treated x Z 0.0341 -0.135

0.0891 -0.129

-0.112 0.00762
(0.0837) (0.0748) (0.0732) (0.0727) (0.0744) (0.0793)
Treated x Overestimates x Z -0.159 0.146 -0.0360 0.174

0.0474 -0.104
(0.114) (0.106) (0.103) (0.102) (0.103) (0.102)
Overestimates not Z = Overestimates & Z f0.108g f0.894g f0.460g f0.526g f0.370g f0.176g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.311g f0.197g f0.957g f0.144g f0.550g f0.638g
(Tr x Overestimates + Tr x Overestimates x Z)
Panel B: Single Regression (all Z Variables) (1) (2) (3) (4) (5) (6)
Treated 0.134 0.134 0.134 0.134 0.134 0.134
(0.092) (0.092) (0.092) (0.092) (0.092) (0.092)
Treated x Overestimates 0.005 0.005 0.005 0.005 0.005 0.005
(0.116) (0.116) (0.116) (0.116) (0.116) (0.116)
Treated x Z 0.009 -0.146

0.086 -0.131

-0.104 0.014
(0.085) (0.077) (0.074) (0.074) (0.076) (0.080)
Treated x Overestimates x Z -0.101 0.140 -0.052 0.155 0.027 -0.117
(0.118) (0.109) (0.104) (0.103) (0.105) (0.106)
Overestimates not Z = Overestimates & Z f0.265g f0.933g f0.639g f0.732g f0.286g f0.166g
(Tr x Z + Tr x Overestimates x Z)
Underestimates & Z = Overestimates & Z f0.486g f0.271g f0.702g f0.188g f0.802g f0.382g
(Tr x Overestimates + Tr x Overestimates x Z)
Control Group Mean 0.34 0.34 0.34 0.34 0.34 0.34
Observations Underest x NOT Z 387 338 280 287 273 268
Observations Underest x Z 129 178 236 229 243 248
Observations Overest x NOT Z 374 352 271 275 251 217
Observations Overest x Z 139 161 242 238 262 296
Observations (Total) 1029 1029 1029 1029 1029 1029
Panel A presents results from a separate regression including interactions with each heterogeneous variable. Panel B
presents results from a single regression where all heterogeneous variable interactions are included. Robust standard
errors are reported in parenthases and p-values for statistical tests in brackets. All regressions are conditional on
university xed eects, covariates determined by LASSO double selection, and treatment interacted with quintiles of
EFC. Overestimates is a dummy equal to 1 for students with positive bias, thus terms not interacted with Overes-
timates represent eects for students who underestimate their relative nancial need. Sample includes all baseline
survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
107
A.4 Full Sample Results
Figure A.11: Treatment Eects Associated with Table A.20
Figure A.12: Treatment Eects Associated with Table A.21
108
A.4 Full Sample Results
Table A.20: Treatment Eects on Applying for Aid
Low EFC Medium EFC High EFC
Under/Low Bias 0.056 0.040 -0.063
(0.054) (0.054) (0.051)
Overestimates -0.128 0.130 -0.043
(0.111) (0.090) (0.059)
Observations Under/Low Bias 274 253 167
Observations Overestimates 69 90 176
This table presents estimates of from a single regression including interactions of each tercile of EFC with Overes-
timating need or Underestimating / Having Low Bias and with an indicator for treatment assignment. Coecients
re
ect estimates of the average treatment eect of information provision within each EFC tercile - Bias grouping
pair. Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and
on baseline covariates determined by LASSO double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table A.21: Treatment Eects on Amount Requested (as % of Student Fee)
Low EFC Medium EFC High EFC
Under/Low Bias 0.056 0.013 0.014
(0.037) (0.026) (0.026)
Overestimates -0.085 0.056 -0.052

(0.086) (0.053) (0.030)
Observations Under/Low Bias 274 253 167
Observations Overestimates 69 90 176
This table presents estimates of from a single regression including interactions of each tercile of EFC with Overes-
timating need or Underestimating / Having Low Bias and with an indicator for treatment assignment. Coecients
re
ect estimates of the average treatment eect of information provision within each EFC tercile - Bias grouping
pair. Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and
on baseline covariates determined by LASSO double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
109
A.4 Full Sample Results
Figure A.13: Treatment Eects Associated with Table A.22
Figure A.14: Treatment Eects Associated with Table A.23
110
A.4 Full Sample Results
Table A.22: Treatment Eects on Applying for Aid
Underestimates Low Bias Overestimates
Low EFC 0.051 0.057 -0.121
(0.072) (0.077) (0.111)
Medium EFC 0.064 0.017 0.127
(0.069) (0.088) (0.090)
High EFC -0.036 -0.071 -0.044
(0.087) (0.060) (0.060)
Observations Low EFC 164 110 69
Observations Med EFC 154 99 90
Observations High EFC 32 135 176
This table presents estimates of from a single regression including interactions of each tercile of EFC and Bias with
one another and with an indicator for treatment assignment. Coecients re
ect estimates of the average treatment
eect of information provision within each EFC tercile - Bias tercile pair. Robust standard errors are reported in
parenthases. Estimates are conditional on university xed eects and on baseline covariates determined by LASSO
double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table A.23: Treatment Eects on Amount Requested (as % of Student Fee)
Underestimates Low Bias Overestimates
Low EFC 0.046 0.065 -0.078
(0.042) (0.063) (0.086)
Medium EFC 0.008 0.028 0.054
(0.028) (0.050) (0.053)
High EFC 0.042 0.007 -0.053

(0.049) (0.029) (0.030)
Observations Low EFC 164 110 69
Observations Med EFC 154 99 90
Observations High EFC 32 135 176
This table presents estimates of from a single regression including interactions of each tercile of EFC and Bias with
one another and with an indicator for treatment assignment. Coecients re
ect estimates of the average treatment
eect of information provision within each EFC tercile - Bias tercile pair. Robust standard errors are reported in
parenthases. Estimates are conditional on university xed eects and on baseline covariates determined by LASSO
double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
111
A.4 Full Sample Results
Figure A.15: Treatment Eects Associated with Table A.24 Panel A
Figure A.16: Treatment Eects Associated with Table A.24 Panel B
112
A.4 Full Sample Results
Table A.24: Treatment Eect by Tercile of EFC & Savings
Applied for Aid Request (% of Student Fee)
Panel A: Eects by EFC Tercile (1) (2) (3) (4)
Treated x Low EFC 0.0286 0.0386 0.0356 0.0384
(0.0486) (0.0480) (0.0348) (0.0345)
Treated x Mid EFC 0.0671 0.0685 0.0260 0.0243
(0.0470) (0.0461) (0.0238) (0.0235)
Treated x High EFC -0.0490 -0.0326 -0.0182 -0.0121
(0.0400) (0.0388) (0.0198) (0.0199)
Panel B: Eects by Savings Tercile
Treated x Low Savings 0.00797 0.0161 0.0622

0.0591

(0.0499) (0.0463) (0.0343) (0.0317)
Treated x Mid Savings 0.0311 0.0329 0.00113 0.00233
(0.0473) (0.0439) (0.0279) (0.0256)
Treated x High Savings 0.0191 0.0299 -0.0102 -0.00435
(0.0431) (0.0413) (0.0206) (0.0204)
Controls N Y N Y
Control Mean 0.34 0.34 0.16 0.16
Observations (N) 1029 1029 1029 1029
This table presents estimates of the average treatment eects of information provision within each tercile of EFC
and each tercile of Personal Savings. Panels A and B are estimated independently (i.e., not conditional on treatment
interactions of the other). Estimates are conditional on baseline covariates determined by LASSO double selection.
Robust standard errors are reported in parenthases Sample includes all baseline survey respondents.

p < 0:1,

p< 0:05,

p< 0:01,

p< 0:001
113
A.4 Full Sample Results
Figure A.17: Treatment Eects Associated with Table A.25
Figure A.18: Treatment Eects Associated with Table A.26
114
A.4 Full Sample Results
Table A.25: Distributional Eects on Applying for Aid
Low Savings Medium Savings High Savings
Low EFC 0.060 0.037 0.090
(0.061) (0.069) (0.077)
High EFC -0.089 0.043 -0.014
(0.080) (0.058) (0.049)
Observations Low EFC 224 172 119
Observations High EFC 126 179 209
This table presents estimates of from a single regression including interactions of each tercile of EFC and Savings
with one another and with an indicator for treatment assignment according to Specication 9. Coecients re
ect
estimates of the average treatment eect of information provision within each EFC tercile - Savings tercile pair.
Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and on
baseline covariates determined by LASSO double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Table A.26: Distributional Eects on Amount Requested (as % of Student Fee)
Low Savings Medium Savings High Savings
Low EFC 0.103

-0.012 0.003
(0.045) (0.045) (0.037)
High EFC -0.014 0.025 -0.013
(0.043) (0.028) (0.024)
Observations Low EFC 224 172 119
Observations High EFC 126 179 209
This table presents estimates of from a single regression including interactions of each tercile of EFC and Savings
with one another and with an indicator for treatment assignment according to Specication 9. Coecients re
ect
estimates of the average treatment eect of information provision within each EFC tercile - Savings tercile pair.
Robust standard errors are reported in parenthases. Estimates are conditional on university xed eects and on
baseline covariates determined by LASSO double selection. Sample includes all baseline survey respondents.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
115
Appendix B: Consequences of FOSTA/SESTA Legislation
B.1 Supplementary Figures
Figure B.1: Google Trends Scores of geographic relative search frequency across 210 market areas
from 3/15/2017 to 3/15/2018 for Craigslist, Inc. (Advertising Company), Backpage.com (Website)
and Me Too movement (Topic). Markets are ordered by increasing search interest in Craigslist.
116
B.1 Supplementary Figures
Figure B.2: 210 Google Trends Market Areas colored by Craigslist search score 3/15/17 - 3/15/18.
Figure B.3: 218 MSAs in the 2013-2019 balanced panel colored by Craigslist treatment designation.
117
B.1 Supplementary Figures
Figure B.4
118
B.1 Supplementary Figures
Figure B.5
119
B.2 Robustness: Controlling for MSA Demographic Eects
Table B.1: Robustness to Demographics x Post Controls
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4) (5) (6) (7)
High Craigslist x Post 2.282

2.652

2.775

2.507

2.666

2.663

2.552

(0.848) (1.112) (1.091) (0.920) (0.900) (0.867) (0.954)
High Backpage x Post 1.307 0.203 0.118 0.370 0.218 0.199 1.094
(0.910) (1.065) (1.056) (1.019) (0.973) (0.927) (0.935)
High Me Too x Post -1.383 -0.860 -0.923 -0.852 -0.879 -0.879 -1.204
(1.054) (1.056) (1.048) (1.048) (1.071) (1.176) (1.064)
High Male : Female Pop Ratio x Post 2.963

3.447

(0.827) (0.865)
High % White x Post 0.0448 0.645
(1.289) (2.129)
High % White Female x Post -0.319 -0.915
(1.227) (2.090)
High % Black/African American x Post -0.712 0.516
(1.165) (1.079)
High % Hispanic x Post 0.372 -0.256
(1.018) (1.093)
High % of Two or More Races x Post 0.0348 -0.943
(1.126) (1.242)
N MSAs 218 218 218 218 218 218 218
N Years 5 5 5 5 5 5 5
N Total 1090 1090 1090 1090 1090 1090 1090
This table checks robustness of the DID coecients for Craigslist, Backpage and the Me Too movement to condition-
ing on DID estimates for treatment groups dened by population demographics such as the ratio of males to females
and % of population in diernt ethnic/racial groups. Estimates based on a panel of US Metropolitan Statistical
Areas from 2015 to 2019. The Post period is dened as 2018-2019. Regressions control for MSA population. Robust
standard errors, clustered by Google Trends Metro Area, are reported in parenthases. Regressions are weighted by
MSA population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Figure B.6: Plot of estimates corresponding to Table B.2
120
B.2 Robustness: Controlling for MSA Demographic Eects
Figure B.7: Plot of estimates corresponding to Table B.2
121
B.2 Robustness: Controlling for MSA Demographic Eects
Table B.2: Robustness of Event Study Estimates to High M:F Ratio x Post
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4) (5)
High Craigslist x 2015 -0.210 -0.0142
(1.147) (1.200)
High Craigslist x 2016 -0.371 -0.219
(0.884) (0.909)
High Craigslist x 2018 2.895

2.572

(1.158) (1.091)
High Craigslist x 2019 2.342

1.804
(1.129) (1.142)
High Backpage x 2015 1.316 0.598
(1.091) (1.178)
High Backpage x 2016 0.689 0.347
(0.774) (0.939)
High Backpage x 2018 0.356 1.237
(1.000) (1.097)
High Backpage x 2019 2.020

2.043

(1.054) (1.134)
High Me Too x 2015 -0.913 -0.302
(1.144) (1.141)
High Me Too x 2016 -0.237 0.123
(0.775) (0.864)
High Me Too x 2018 -0.469 -0.384
(0.991) (1.118)
High Me Too x 2019 -2.926

-2.526

(1.075) (1.120)
High Male : Female Pop Ratio x 2015 -1.809

-1.448
(1.092) (1.194)
High Male : Female Pop Ratio x 2016 -0.986 -0.845
(0.798) (0.892)
High Male : Female Pop Ratio x 2018 1.901

2.275

(1.073) (1.049)
High Male : Female Pop Ratio x 2019 0.691 2.108

(1.205) (1.058)
P-val for Joint Test of Craigslit Pre 0.893 0.934
P-val for Joint Test of Craigslit Post 0.029 0.056
P-val for Joint Test of Backpage Pre 0.485 0.879
P-val for Joint Test of Backpage Post 0.086 0.194
P-val for Joint Test of Me Too Pre 0.671 0.872
P-val for Joint Test of Me Too Post 0.011 0.054
P-val for Joint Test of High M:F Ratio Pre 0.254 0.471
P-val for Joint Test of High M:F Ratio Post 0.152 0.058
N MSAs 218 218 218 218 218
N Years 5 5 5 5 5
N Total 1090 1090 1090 1090 1090
This table checks robustness of Event Study coecients for Craigslist, Backpage and the Me Too movement to
conditioning on estimates for a treatment group dened by MSA population ratio of males to females. Estimates
are based on a panel of MSAs from 2015 to 2019. The Post period is considered 2018-2019. Regressions control
for MSA population. Robust standard errors, clustered by Google Trends Metro Area, are reported in parenthases.
Regressions are weighted by MSA population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
122
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population
Table B.3: MSA Characteristics in 2017 by Craigslist Treatment (Dropping Population Outliers)
Means and Standard Deviations
Craigslist Craigslist Dierence
Variable Full Sample Control Treatment Treat vs Cont.
(1) (2) (3) (4)
Backpage.com Google Trends Score 31.951 33.151 30.680 -2.471
(11.844) (11.063) (12.550) f0.136g
Me Too movement Google Trends Score 35.956 36.453 35.430 -1.023
(17.124) (16.462) (17.868) f0.670g
MSA Population (in 100,000s) 4.169 4.242 4.091 -0.150
(4.433) (4.546) (4.332) f0.808g
Rate of Violent Crime 378.976 381.449 376.326 -5.123
(189.486) (185.986) (194.088) f0.848g
Rate of Murder / Manslaughter 4.710 5.045 4.354 -0.691
(3.700) (3.640) (3.749) f0.181g
Rate of Rape / Sexual Assault 48.635 47.278 50.074 2.797
(20.798) (18.157) (23.280) f0.339g
Rate of Robbery 69.162 71.841 66.321 -5.520
(49.414) (45.112) (53.683) f0.427g
Rate of Aggravated Assault 255.656 256.793 254.438 -2.355
(144.078) (146.226) (142.481) f0.908g
Rate of Property Crime 2,479.570 2,443.088 2,516.425 73.336
(900.132) (878.613) (924.409) f0.569g
Rate of Burglary 466.352 486.095 446.011 -40.084
(214.806) (237.000) (188.292) f0.185g
Rate of Larceny Theft 1,801.307 1,791.187 1,811.733 20.546
(610.776) (603.819) (620.762) f0.812g
Rate of Motor Vehicle Theft 212.813 175.391 252.107 76.717***
(169.440) (101.347) (212.896) f0.001g
Male : Female Pop Ratio 0.982 0.972 0.992 0.020***
(0.044) (0.039) (0.046) f0.001g
% Pop Male 0.495 0.493 0.498 0.005***
(0.011) (0.010) (0.011) f0.001g
% Pop White Alone 0.829 0.803 0.856 0.053***
(0.106) (0.110) (0.095) f0.000g
% Pop Black/African American 0.108 0.148 0.065 -0.083***
(0.098) (0.113) (0.054) f0.000g
% Pop Hispanic 0.147 0.110 0.187 0.077***
(0.180) (0.119) (0.221) f0.002g
% Pop Two or More Races 0.027 0.024 0.030 0.006**
(0.019) (0.009) (0.025) f0.015g
% Pop White Alone Female 0.418 0.407 0.430 0.024***
(0.055) (0.057) (0.051) f0.002g
Observations 206 106 100 206
This table presents means and standard deviations (in parentheses) of MSA level variables for the full sample of
MSAs forming the balance analysis panel, with outlier MSAs above the 95th percentile of population removed, and
for the MSAs divided according to the Craigslist Treatment and Control groups. Column (4) reports the dierence in
means as well as p-values for the test of equality of means between treatment and control. Stars indicate statistical
signicance at .1 (*), .05 (**), and .01 (***) levels.
123
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population
Figure B.8: Trends excluding MSAs with population above the 95th percentile
Table B.4: DID Estimate: Discontinuity in Rape / Sexual Assault post 2018
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4)
High Craigslist x Post 3.627

3.574

(1.122) (1.074)
High Backpage x Post -0.665 -0.623
(1.197) (1.026)
High Me Too x Post -1.174 -1.252
(1.279) (1.120)
MSA Population (in 100,000s) -1.043 -1.795 -1.555 -1.008
(1.657) (1.814) (1.822) (1.529)
N MSAs 206 206 206 206
N Years 5 5 5 5
N Total 1030 1030 1030 1030
This table presents Event Study estimates based on a panel of US Metropolitan Statistical Areas from 2015 to 2019
where treatment is dened as being an MSA with above median interest in each treatment (as proxied by Google
search frequency). The post period is dened as 2018-2019. Robust standard errors, clustered by Google Trends
Metro Area, are reported in parenthases. Regressions are weighted by MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
124
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population
Table B.5: Event Study Estimates: Discontinuity in Rape / Sexual Assault post 2018
Depvar: Rate of Rape / Sexual Assault
(1) (2) (3) (4)
High Craigslist x 2015 -0.102 0.00434
(1.481) (1.463)
High Craigslist x 2016 0.177 0.211
(1.011) (1.006)
High Craigslist x 2018 4.048

4.017

(1.356) (1.300)
High Craigslist x 2019 3.258

3.274

(1.469) (1.429)
High Backpage x 2015 2.287 2.168
(1.493) (1.437)
High Backpage x 2016 0.950 0.842
(1.006) (1.006)
High Backpage x 2018 -0.458 -0.329
(1.451) (1.274)
High Backpage x 2019 1.287 1.088
(1.531) (1.438)
High Me Too x 2015 -1.043 -0.646
(1.481) (1.376)
High Me Too x 2016 -0.836 -0.673
(0.986) (0.982)
High Me Too x 2018 -0.838 -0.846
(1.436) (1.284)
High Me Too x 2019 -2.760

-2.537

(1.480) (1.385)
MSA Population (in 100,000s) -1.039 -1.794 -1.571 -1.015
(1.660) (1.757) (1.839) (1.468)
P-val for Joint Test of Craigslit Pre 0.953 0.962
P-val for Joint Test of Craigslit Post 0.011 0.008
P-val for Joint Test of Backpage Pre 0.312 0.318
P-val for Joint Test of Backpage Post 0.406 0.583
P-val for Joint Test of Me Too Pre 0.686 0.791
P-val for Joint Test of Me Too Post 0.143 0.180
N MSAs 206 206 206 206
N Years 5 5 5 5
N Total 1030 1030 1030 1030
This table presents Event Study estimates based on a panel of US Metropolitan Statistical Areas from 2015 to 2019
where treatment is dened as being an MSA with above median interest in each treatment (as proxied by Google
search frequency). The post period is considered 2018-2019. Robust standard errors, clustered by Google Trends
Metro Area, are reported in parenthases. Regressions are weighted by MSA Population.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
125
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population
Figure B.9
126
B.3 Robustness: Dropping MSAs Above the 95th Percentile in Population
Figure B.10
127
B.4 Robustness: Balanced Panel Results on Prostitution Arrests
Table B.6: DID Estimates: Impact of FOSTA/SESTA on Prostitution Arrests
Total Female Juvenile Female Juv. Male Juv.
(1) (2) (3) (4) (5)
High Craigslist x Post 0.174 0.0902 0.00876 0.00591 0.00285
(0.146) (0.111) (0.00544) (0.00394) (0.00234)
High Backpage x Post -0.179 -0.113 -0.000970 0.000871 -0.00184
(0.160) (0.126) (0.00496) (0.00390) (0.00195)
Population (100,000s) -1.869

-1.420

-0.0432

-0.0411

-0.00211
(0.418) (0.444) (0.00935) (0.00906) (0.00155)
N Agencies 2840 2840 2840 2840 2840
N Months 84 84 84 84 84
N Total 238560 238560 238560 238560 238560
This table presents DID estimates based on a monthly panel of US Municipal Police Agencies from 2012 to
2018 where treatment is dened as being an agency in a Google Trends Metro Area with above median usage of
Craigslist/Backpage in the year prior to implementation of FOSTA/SESTA legislation. The post period is dened as
2018 Q2-Q4. Robust standard errors, clustered by Google Trends Metro Area, are reported in parenthases. Regres-
sions are weighted by population covered by police agency.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
Figure B.11: Estimates from a balanced panel of municipal police agencies.
128
B.4 Balanced Panel Results on Prostitution Arrests
Table B.7: Event Study Estimates: Juvenile Prostitution Arrests
Full Panel Quarters 2-4 Only
Juvenile Female Juv. Juvenile Female Juv.
(1) (2) (3) (4)
High Craigslist x 2012 0.00175 0.00177 0.00265 0.00167
(0.00666) (0.00670) (0.00692) (0.00699)
High Craigslist x 2013 0.00690 0.00837 0.00673 0.00826
(0.0107) (0.00996) (0.0125) (0.0119)
High Craigslist x 2014 -0.00134 0.000883 -0.000588 0.00210
(0.00767) (0.00759) (0.00788) (0.00765)
High Craigslist x 2015 -0.00329 -0.00154 -0.00278 -0.00164
(0.00597) (0.00597) (0.00781) (0.00807)
High Craigslist x 2016 0.00295 0.000777 -0.00000744 -0.0000645
(0.00702) (0.00622) (0.00683) (0.00620)
High Craigslist x 2018 0.00758 0.00591 0.00933 0.00720

(0.00524) (0.00383) (0.00583) (0.00432)
Population (100,000s) -0.0431

-0.0411

-0.0438

-0.0428

(0.00930) (0.00904) (0.0109) (0.0110)
Conditional on High Backpage x Years Yes Yes Yes Yes
P-val Joint Test of Pre-Periods Craigslist 0.236 0.267 0.665 0.618
P-val Joint Test of Pre-Periods Backpage 0.171 0.131 0.583 0.331
N Agencies 2840 2840 2840 2840
N Months 84 84 77 77
N Total 238560 238560 178920 178920
This table presents Event Study estimates based on a monthly panel of US Municipal Police Agencies from 2012
to 2018 where treatment is dened as being an agency in a Google Trends Metro Area with above median usage
of Craigslist in the year prior to implementation of FOSTA/SESTA legislation. The post period is dened as 2018.
Robust standard errors, clustered byGoogle Trends Metro Area, are reported in parenthases. Regressions are weighted
by population covered by police agency. In columns 3 & 4 data for quarter 1 of each year is dropped to remove dilution
of treatment eect which which begins Q2 of 2018.

p< 0:1,

p< 0:05,

p< 0:01,

p< 0:001
129
References
Abadie, Alberto, Alexis Diamond, and Jens Hainmueller (2010), \Synthetic control methods for
comparative case studies: Estimating the eect of california's tobacco control program." Journal
of the American Statistical Association, 105, 493{505, URL https://doi.org/10.1198/jasa.
2009.ap08746.
Alesina, Alberto and George-Marios Angeletos (2005), \Fairness and redistribution." Ameri-
can Economic Review, 95, 960{980, URL http://www.aeaweb.org/articles?id=10.1257/
0002828054825655.
Alesina, Alberto, Guido Cozzi, and Noemi Mantovan (2012), \The evolution of ideology, fairness
and redistribution*." The Economic Journal, 122, 1244{1261, URL https://onlinelibrary.
wiley.com/doi/abs/10.1111/j.1468-0297.2012.02541.x.
Alesina, Alberto, Armando Miano, and Stefanie Stantcheva (2018), \Immigration and redistribu-
tion." Working Paper 24733, National Bureau of Economic Research, URL http://www.nber.
org/papers/w24733.
Athey, S. and G.W. Imbens (2017), \Chapter 3 - the econometrics of randomized experimentsa."
In Handbook of Field Experiments (Abhijit Vinayak Banerjee and Esther Du
o, eds.), vol-
ume 1 of Handbook of Economic Field Experiments, 73{140, North-Holland, URL https:
//www.sciencedirect.com/science/article/pii/S2214658X16300174.
Balafoutas, Loukas, Rudolf Kerschbamer, and Matthias Sutter (2012), \Distributional preferences
and competitive behavior." Journal of Economic Behavior & Organization, 83, 125{135, URL
https://EconPapers.repec.org/RePEc:eee:jeborg:v:83:y:2012:i:1:p:125-135.
Becker, Gordon M., Morris H. Degroot, and Jacob Marschak (1964), \Measuring utility by a single-
response sequential method." Behavioral Science, 9, 226{232, URL https://onlinelibrary.
wiley.com/doi/abs/10.1002/bs.3830090304.
Bisschop, Paul, Stephen Kastoryano, and Bas van der Klaauw (2017), \Street prostitution zones and
130
crime." American Economic Journal: Economic Policy, 9, 28{63, URL https://www.aeaweb.
org/articles?id=10.1257/pol.20150299.
Bolton, Gary E. and Axel Ockenfels (2000), \Erc: A theory of equity, reciprocity, and competition."
American Economic Review, 90, 166{193, URL https://www.aeaweb.org/articles?id=10.
1257/aer.90.1.166.
Bun, Maurice J. G. and Teresa D. Harrison (2019), \Ols and iv estimation of regression models
including endogenous interaction terms." Econometric Reviews, 38, 814{827, URL https://
EconPapers.repec.org/RePEc:taf:emetrv:v:38:y:2019:i:7:p:814-827.
Buraschi, Andrea and Francesca Cornelli (2013), \The economics of donations and enlightened
self-interest." SSRN, URL https://ssrn.com/abstract=317688.
B enabou, Roland and Jean Tirole (2006), \Belief in a Just World and Redistributive Politics*."
The Quarterly Journal of Economics, 121, 699{746, URL https://doi.org/10.1162/qjec.
2006.121.2.699.
B enabou, Roland and Jean Tirole (2010), \Individual and corporate social responsibility." Eco-
nomica, 77, 1{19, URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0335.
2009.00843.x.
Cameron, Lisa, Jennifer Seager, and Manisha Shah (2020), \Crimes Against Morality: Unintended
Consequences of Criminalizing Sex Work*." The Quarterly Journal of Economics, 136, 427{469,
URL https://doi.org/10.1093/qje/qjaa032.
Cappelen, Alexander W., Astri Drange Hole, Erik Srensen, and Bertil Tungodden (2007), \The
pluralism of fairness ideals: An experimental approach." American Economic Review, 97, 818{
827, URL http://www.aeaweb.org/articles?id=10.1257/aer.97.3.818.
Ciacci, Riccardo (2018), \Banning the purchase of prostitution increases rape: evidence from Swe-
den." Working Paper., URL https://mpra.ub.uni-muenchen.de/100393/.
Ciacci, Riccardo and Maria Micaela Sviatschi (2018), \The Eect of Adult Entertainment Es-
tablishments on Sex Crime: Evidence from New York City." Working Paper., URL https:
//repositorio.comillas.edu/xmlui/handle/11531/41905.
131
Cruces, Guillermo, Ricardo Perez-Truglia, and Martin Tetaz (2013), \Biased perceptions of income
distribution and preferences for redistribution: Evidence from a survey experiment." Journal of
Public Economics, 98, 100{112, URL https://EconPapers.repec.org/RePEc:eee:pubeco:v:
98:y:2013:i:c:p:100-112.
Cunningham, Scott and Greg DeAngelo (2019), \Craigslist reduced vio-
lence against women ." URL https://www.semanticscholar.org/paper/
Craigslist-Reduced-Violence-Against-Women-*-Cunningham-DeAngelo/
36229421fc4f63c7aa0e298e8f710510d8710b02.
Cunningham, Scott and Manisha Shah (2017), \Decriminalizing Indoor Prostitution: Implications
for Sexual Violence and Public Health." The Review of Economic Studies, 85, 1683{1715, URL
https://doi.org/10.1093/restud/rdx065.
Engelmann, Dirk and Martin Strobel (2004), \Inequality aversion, eciency, and maximin pref-
erences in simple distribution experiments." American Economic Review, 94, 857{869, URL
http://www.aeaweb.org/articles?id=10.1257/0002828042002741.
Federal Student Aid (2021), \Wondering how the amount of your federal student aid is determined?"
URL https://studentaid.gov/complete-aid-process/how-calculated. (accessed January
5, 2021).
Fehr, Ernst and Klaus M. Schmidt (1999), \A theory of fairness, competition, and cooperation." The
Quarterly Journal of Economics, 114, 817{868, URL http://www.jstor.org/stable/2586885.
Fischbacher, Urs and Simon Gachter (2010), \Social preferences, beliefs, and the dynamics of
free riding in public goods experiments." American Economic Review, 100, 541{56, URL http:
//www.aeaweb.org/articles?id=10.1257/aer.100.1.541.
Forsythe, Robert, Joel L. Horowitz, N.E. Savin, and Martin Sefton (1994), \Fairness in simple
bargaining experiments." Games and Economic Behavior, 6, 347 { 369, URL http://www.
sciencedirect.com/science/article/pii/S0899825684710219.
Frey, Bruno S. and Stephan Meier (2004), \Pro-social behavior in a natural setting." Journal
132
of Economic Behavior & Organization, 54, 65 { 88, URL http://www.sciencedirect.com/
science/article/pii/S0167268103002257.
Glazer, Amihai and Kai Konrad (1996), \A signaling explanation for charity." American Economic
Review, 86, 1019{28, URLhttps://EconPapers.repec.org/RePEc:aea:aecrev:v:86:y:1996:
i:4:p:1019-28.
Hart, Claire M., Timothy D. Ritchie, Erica G. Hepper, and Jochen E. Gebauer (2015), \The bal-
anced inventory of desirable responding short form (bidr-16)." SAGE Open, 5, 2158244015621113,
URL https://doi.org/10.1177/2158244015621113.
Karadja, Mounir, Johanna Mollerstrom, and David Seim (2017), \Richer (and holier) than thou?
the eect of relative income improvements on demand for redistribution." The Review of Eco-
nomics and Statistics, 99, 201{212, URL http://www.mitpressjournals.org/doi/10.1162/
REST_a_00623.
Klor, Esteban and Moses Shayo (2010), \Social identity and preferences over redistribution."
Journal of Public Economics, 94, 269{278, URL https://EconPapers.repec.org/RePEc:eee:
pubeco:v:94:y:2010:i:3-4:p:269-278.
Lacetera, Nicola and Mario Macis (2010), \Social image concerns and prosocial behavior: Field
evidence from a nonlinear incentive scheme." Journal of Economic Behavior & Organization,
76, 225{237, URL https://EconPapers.repec.org/RePEc:eee:jeborg:v:76:y:2010:i:2:p:
225-237.
Nair, Gautam (2018), \Misperceptions of relative auence and support for international redistri-
bution." The Journal of Politics, 80, 815{830, URL https://doi.org/10.1086/696991.
Nicola and Mario Macis (2008), \Motivating altruism: A eld study." IZA Dis-
cussion Paper No. 3770. Available at SSRN: https://ssrn.com/abstract=1290039 or
http://dx.doi.org/10.1111/j.0042-7092.2007.00700.x.
Nizalova, Olena and Irina Murtazashvili (2016), \Exogenous treatment and endogenous factors:
Vanishing of omitted variable bias on the interaction term." Journal of Econometric Methods, 5,
133
71{77, URL https://EconPapers.repec.org/RePEc:bpj:jecome:v:5:y:2016:i:1:p:71-77:
n:2.
Perez-Truglia, Ricardo (2020), \The eects of income transparency on well-being: Evidence from
a natural experiment." American Economic Review, 110, 1019{54, URL https://www.aeaweb.
org/articles?id=10.1257/aer.20160256.
Piketty, Thomas (1995), \Social mobility and redistributive politics." The Quarterly Journal of
Economics, 110, 551{584, URL http://www.jstor.org/stable/2946692.
Tonin, Mirco and Michael Vlassopoulos (2013), \Experimental evidence of self-image concerns
as motivation for giving." Journal of Economic Behavior & Organization, 90, 19 { 27, URL
http://www.sciencedirect.com/science/article/pii/S0167268113000565.
Urminsky, Oleg, Christian Hansen, and Victor Chernozhukov (2016), \Using double-lasso regres-
sion for principled variable selection." Available at SSRN, URL https://ssrn.com/abstract=
2733374.
134

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays in the economics of education and conflict

PDF

Essays on development economics and adolescent behavior

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Student mobility in policy and poverty context: two essays from Washington

Asset Metadata

Creator Schneider, Jacob Joseph (author)

Core Title Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 03/15/2021

Defense Date 02/26/2021

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag biased perceptions,comparative economic need,craigslist.com,FOSTA,Me Too movement,OAI-PMH Harvest,pro-social behavior,Public Policy,rape,scholarship requests,subjective perception

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Oliva, Paulina (committee chair), Melguizo, Tatiana (committee member), Ridder, Geert (committee member)

Creator Email jacob.j.schneider@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-427768

Unique identifier UC11668229

Identifier etd-SchneiderJ-9317.pdf (filename),usctheses-c89-427768 (legacy record id)

Legacy Identifier etd-SchneiderJ-9317.pdf

Dmrecord 427768

Document Type Dissertation

Rights Schneider, Jacob Joseph

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA