Defaults and Donations: Evidence from a Field Experiment∗

Steffen Altmann, Armin Falk, Paul Heidhues, Rajshri Jayaraman, and Marrit Teirlinck

July 2018

A BSTRACT
We study how defaults affect charitable donations. In a field experiment that was conducted on a
large online platform for charitable giving, we exogenously vary the default options in the donation
form in two distinct choice dimensions. The first pertains to the primary donation decision, namely,
how much to contribute to the charitable cause. The second relates to a “codonation” decision
of how much to contribute to supporting the online platform itself. We find a strong impact of
defaults on individual behavior: in each of our treatments, the modal positive contributions in both
choice dimensions invariably correspond to the specified default amounts. Defaults, nevertheless,
have no significant effects on average donation levels. This is because defaults in the donation
domain induce some people to donate more and others to donate less. In contrast, higher defaults in
the secondary choice dimension unambiguously induce higher average contributions to the online
platform. We complement our experimental results by setting up and estimating a structural model
that explores whether personalizing defaults based on individuals’ donation histories can help the
online platform to increase donation revenues.

Keywords: Default Options, Online Platforms, Charitable Giving, Field Experiment

JEL codes: D03, D01, D64, C93

∗
Steffen Altmann: University of Copenhagen; Armin Falk: Institute on Behavior and Inequality (briq) and University
of Bonn; Paul Heidhues: Düsseldorf Institute for Competition Economics (DICE); Rajshri Jayaraman and Marrit
Teirlinck: ESMT Berlin. We thank the editor, Brigitte Madrian, and two anonymous reviewers for their thoughtful
comments and suggestions. The paper also benefited from discussions with Thomas Dohmen, Andreas Grunewald,
Joachim Winter, and Florian Zimmermann, and from comments of seminar and conference participants in Aarhus,
Bergen, Berlin, Bristol, Munich, Munster, Oxford, Paris, Ithaca, Montréal, and Vancouver. We are grateful to
Betterplace.org, in particular Joana Breidenbach and Tobias Jordan, for their input, support, and cooperation. Steffen
Altmann gratefully acknowledge financial support from the Volkswagen Foundation and Rajshri Jayaraman gratefully
acknowledges financial support from the DFG project CRC TRR 190.
 1

1. I NTRODUCTION

Online fundraising constitutes a sizable and rapidly growing segment of the market for charitable
giving.1 A pervasive feature on the websites of charities and online fundrasising platforms are
default options that specify the amount to be donated unless a donor actively enters a different
contribution level. The ubiquity of default donation amounts in online fundraising is likely to stem
from a common presumption that “defaults matter”. This presumption—buttressed by famously
documented examples of the importance of defaults for decisions on retirement saving or organ
donation (e.g., Madrian and Shea 2001, Johnson and Goldstein 2003, Thaler and Sunstein 2008)—
has generated a lively discussion in the practitioner community regarding “best practice” for setting
default donation amounts. Yet, this discussion lacks rigorous evidence.2
This paper takes a step towards closing this evidence gap with the help of a natural field experi-
ment on Germany’s largest online platform for making charitable contributions. The experiment is
designed to ask two main questions. First, do defaults affect individual behavior; in other words, do
they influence the distribution of individuals’ contributions? Second, do defaults influence overall
donation revenues?
To address these questions, we exogenously vary default options in two distinct choice dimen-
sions: The main donation decision and an add-on choice, which is a gratuity to support the providers
of the online platform. Regarding the first dimension, website visitors are randomly assigned to
default donation amounts of e10, 20, and 50. These values correspond, respectively, to the 25th,
50th, and 75th percentile of donations on the platform in the six months prior to our experiment.
This allows us to examine whether defaults have stronger or weaker effects on behavior when they
are set relatively high or low compared to what most people would donate otherwise. We also
implement an additional treatment in which the donation field is initially set to zero, such that

1
In 2017, online giving accounted for 7.6% of the total—multi-billion—fundraising volume in the U.S. nonprofit sector
(see Blackbaud 2018). In line with the positive trend from previous years, online giving grew strongly both in absolute
terms as well as compared to the overall increase in charitable giving (the yearly growth rates were 12.1% and 4.1%,
respectively).
2
Perhaps the best existing evidence comes from a disaster relief donation drive conducted by Google.com in 2009
(see http://googlecheckout.blogspot.dk/2009/12/google-checkout-for-non-profits-in-2010.html). While there is no
information on sample sizes and statistical significance, the results indicate that—with the exception of a strong drop
in average donations at a $20 default—overall donation revenues did not differ strongly for different default donation
levels.
2

people who want to make a donation have to make an active decision on their contribution level. In
our second treatment dimension, we randomly assign donation-page visitors to percentage add-ons
of 5%, 10%, or 15% of their main donation. The corresponding contributions, or “codonations”, are
used to support and maintain the online platform, which itself operates as a nonprofit organization.
Over the course of our experiment, we collected data on roughly 680,000 donation-page visits
and almost 23,000 donations, yielding a total of e1.17 million in terms of revenues for charitable
organizations on the platform. Our data show that defaults have a strong impact on individual
donor behavior. In each of our treatments, the modal positive contribution invariably corresponds
to the default specification. This holds for the main donation decision as well as for the add-on
contribution to the online platform, indicating that defaults are important poles of attraction for
donors’ behavior in both decision dimensions.
Despite these systematic effects on the distribution of donations, defaults in our experiment do
not significantly alter overall donation revenues. We neither find systematic differences in average
contributions across the different donation defaults, nor when comparing average donation levels to
the environment where donors have to actively decide on their contribution. The difference between
our individual- and aggregate-level results can be explained by countervailing changes in the
distribution of donations due to defaults. We find that, relative to the active-decision environment,
defaults induce some people to donate more while others donate less or not at all, such that the
two countervailing effects cancel each other out at the aggregate level. For default contributions of
e10 and e20, the changes in the donation distribution operate entirely on the intensive margin. At
the e50 default, we observe an additional extensive-margin effect, with more people opting out of
the donation process altogether. As a result of this higher donor attrition, average donations once
again do not differ significantly from those in the other treatments. By contrast, we do observe
strong average treatment effects in the add-on dimension. Average codonation revenues increase
monotonically in the percentage add-on that is set as the default. This is because the dominant
change in the distribution of codonations at higher default values is an intensive-margin movement
towards the default from lower codonations.
 3

In the final part of our empirical analysis, we examine whether “personalized” defaults could
help to raise donation revenues in our setting. We start by exploring heterogenous treatment effects
based on individual-level characteristics such as gender and the type of donation. Consistent with
a number of earlier findings in the literature (e.g., Madrian and Shea 2001, Levav et al. 2010,
Altmann, Falk, and Grunewald 2013), we find that some groups of donors are more likely to stick
to defaults than others. Our reduced-form estimates, however, also indicate that there is little scope
for making use of this tendency to systematically increase donations. To further explore whether
the online platform could increase donations by differentiating defaults based on individuals’ prior
donation levels, we estimate a simple structural model in which individuals behave as if deviating
from the default were costly (motivated by Carroll et al. 2009 and Bernheim, Fradkin, and Popov
2015). Counterfactual simulations based on our model estimates indicate that there is at best modest
scope for successfully using personalized defaults in a setting like ours, in which information on
(potential) donors is sparse.

Related Literature. Our paper contributes to two main strands of the literature. First, we add
to the body of literature that analyzes the impact of defaults on a variety of economic decisions,
such as retirement saving (e.g., Madrian and Shea 2001, Beshears et al. 2008, and Carroll et al.
2009), organ donor registration (e.g., Johnson and Goldstein 2003 and Abadie and Gay 2006), or
the choice of insurance contracts (Johnson et al. 1993).
An important difference between the mentioned studies and ours is that they consider appli-
cations in which consumers who remain entirely inactive automatically stick to the default. In
contrast, potential donors in our setting must actively confirm the transaction for the default to
affect outcomes. This is similar to how default options are used in web interfaces for configuring
computers, cars, and other customizable products (e.g., Levav et al. 2010, Ebeling 2013). As
we explain in Section 3.4, the difference between the two types of settings is important, as
some common explanations for default effects—such as procrastination in making choices—are
unlikely to play an important role for our results. These differences notwithstanding, we find that
the workhorse model for capturing default effects in the retirement savings literature—a model
involving fixed costs of deviating from the default (see, e.g., Carroll et al. 2009 and Bernheim,
4

Fradkin, and Popov 2015)—does remarkably well in fitting the key features of the donation
distributions in our experiment. At the same time, our estimates indicate that in our setting a
much smaller fraction of potential donors is affected by these as-if costs, which is in line with the
intuition that procrastination is indeed an important factor behind the default effects observed in
the retirement savings context (e.g., Carroll et al. 2009).
Our experiment differs from previous studies on “web defaults” in other economic applications
(e.g., Johnson, Bellman, and Lohse 2002, Löfgren et al. 2012, Ebeling 2013) in that we examine a
setup where consumers do not only face a binary opt-in vs. opt-out decision, but have a continuum
of decision alternatives. This allows us to study a rich set of reactions to defaults along both the
intensive and extensive margin of the donation distribution. Our findings demonstrate that defaults
can have manifold—and, in our case, countervailing—effects, highlighting the importance of a
detailed assessment of distributional effects of defaults for non-binary choices. In particular, our
results indicate that a strategy that attempts to boost donation revenues through higher defaults
based on a simplistic notion that “defaults work” might backfire for charitable organizations.3
The second strand of the literature to which our paper contributes is that on charitable giving
and nonprofit fundraising. While defaults are widely observed on online donation platforms and
many practitioners presume that appropriately specified defaults will help them increase donations,
there has been a lack of rigorous evidence on the causal effects of default options on donation
behavior. A notable exception are two recent papers by Fiala and Noussair (2017) and Goswami
and Urminsky (2016), who study default effects on charitable giving in lab experiments and online
surveys, with mixed results: while Fiala and Noussair (2017) observe no significant differences
in overall donation levels under different defaults, Goswami and Urminsky (2016) report a small
positive effect of higher defaults. One has to bear in mind, however, that these findings are based
on relatively small samples and arguably rather weak incentives.4

3
This is related to a recent result by Haggag and Paci (2014), who analyze tipping behavior in New York City cabs and
find that customers are more likely to leave no tip at all when the payment interface features a high default tip.
4
More distantly related, Smith and Ottoni-Wilhelm (2018) document that default values systematically affect fundrais-
ers’ choices of fundraising goals. Analyzing voluntary cooperation outside a charitable-giving context, Messer et al.
(2007), Altmann and Falk (2011), and Fosgaard and Piovesan (2015) examine the effects of defaults in laboratory
public goods games, and find that that higher default contribution levels tend to enhance cooperation, at least in early
periods of the game.
 5

Beyond defaults, a voluminous literature has examined the impact of other fundraising interven-
tions (see Andreoni 2006 as well as Bekkers and Wiepking 2011 for comprehensive reviews of the
literature). Our paper is related to these studies in that some of the potential mechanisms behind
default effects can also play a role for other fundraising interventions. Specifically, to the extent that
potential donors interpret the default option as a recommended contribution to the charitable cause,
our paper is related to studies that examine how giving is affected by directly requesting (Fraser,
Hite, and Sauer 1988, Edwards and List 2014) or explicitly suggesting (Adena, Huck, and Rasul
2014, Goswami and Urminsky 2016) specific donation levels during solicitation. Similarly, the
literature on “appeal scales” (i.e., providing donors with a vector of multiple suggested contribution
levels; see, e.g., Weyant and Smith 1987, Desmet and Feinberg 2003, Adena and Huck 2016, and
Reiley and Samek 2017) is related in that there is a partial overlap in the channels through which
appeal scales and defaults may affect behavior (e.g., recommendations or anchoring). Lastly,
interventions based on statements like “every penny helps” (e.g., Cialdini and Schroeder 1976,
Fraser, Hite, and Sauer 1988) or the provision of information about other donors’ behavior (e.g.,
Frey and Meier 2004, Shang and Croson 2009) are potentially related to defaults, as they may
also affect behavior by transmitting information or shaping social norms. Since all of these
interventions, however, also introduce aspects that are unrelated to defaults5 and since defaults,
in turn, may work through mechanisms that have little or no relevance for the other interventions,
it is difficult to directly compare the results of these studies to our setting.
The paper proceeds as follows. In the next section, we describe the setup, treatments and
procedures of our experiment. Section 3 presents the main results of our experiment, and Section 4
examines whether the charity could increase aggregate donation revenues by personalizing defaults.
Section 5 concludes.

2. T HE E XPERIMENT

2.1. The Donation Platform. We study the effect of default options on betterplace.org, Ger-
many’s largest platform for making charitable donations over the web. At the time of the
5
For instance, the studies on explicit requests typically provide additional contextual information or employ relatively
strong framing. Similarly, appeal scales open the possibility for “decoy” or “compromise effects” (e.g., Simonson
1989, Ekström 2018).
6

experiment, the platform hosted about 6,000 “project pages” through which charities collect funds
for their activities. The aid projects on the platform cover the whole gamut in terms of geography,
charitable cause, and scale. They range from after-school help for a handful of children in Berlin, to
supporting orphanages in Kenya, to humanitarian aid for victims of natural disasters. Charities that
are present on Betterplace include small local NGOs as well as organizations like UNICEF or the
International Committee of the Red Cross. The platform also hosts pages for “fundraising events”,
which offer individuals, firms, or other organizations the possibility to collect donations for one of
the aid projects by organizing charity runs, benefit concerts, or similar fundraising campaigns.
Visitors to the donation platform can browse individual fundraising or project pages, which
describe the project and overall budget needed to fund it, as well as the amounts of money that
are required for the specific elements of which the overall project consists. Figure 1 provides an
example of a project page. (A full English translation of the original screen shot can be found in
Figure B.1 of the online appendix.) The project title, “Typhoon Haiyan: Emergency Relief in the
Philippines”, is displayed at the top of the page, followed by a picture, a location map, and a project
description. The number of previous donors, the proportion of the overall project budget that has
already been funded, and the amount that is still required for the project are displayed in the upper
right part of the page. Potential donors can contribute directly to the aid project or to one of the
specific project elements, in this example relief packages for the catastrophe zone, displayed at the
bottom right of the figure and further below on the screen (suppressed in Figure 1).6
By clicking on either of two red buttons on the screen—the large button, which reads “Jetzt
spenden” translates to “Donate now” and the smaller one at the bottom right, which reads “Hierfür
spenden” translates to “Donate for this”—the potential donor is redirected to the donation page for
the project (Figure 2; see Figure B.3 in the online appendix for a full English translation). On this
page, the donor specifies the amount that she wishes to contribute to the charitable cause, by filling
in the “Project donation” (“Projektspende”) field on the top left part of the screen. In what follows,
we refer to this amount as the donation or donated amount.

6
The corresponding page for fundraising events has a slightly different layout (see Figure B.2 in the online appendix
for an example). The donation page on which our experimental intervention takes place, however, is exactly the same
for all types of donations (see Figure 2).
 7

F IGURE 1. Screen shot of a project page.

In addition to specifying the donation to the charitable cause, donors can also make a contribution
to support the online platform. In this secondary choice dimension, contributions can be determined
as a percentage add-on or as an absolute Euro amount that is added to the project donation. By
clicking the field below the “Support betterplace.org” (“Fördere betterplace.org”) label on the right
side of the screen, a drop-down menu appears that allows donors to choose between the options
“not this time” (i.e., no contribution), 5%, 10%, 15%, 20%, 25%, or “other amount”. The last of
these options gives the donor the possibility to enter any absolute Euro amount. We refer to the
add-on contributions in support of the platform as codonations. Codonations are used to cover the
costs for developing and sustaining the platform, which itself operates as a nonprofit organization.
The sum of the donation and codonation amount determine the donor’s “Total donation”
(“Gesamtspende”), which is automatically calculated in the second line on the left of the donation
form. In the bottom part of the donation page (suppressed in Figure 2), donors are asked to provide
further information that is required to finalize the transaction, including their name and payment
details. After having completed the donation form, donors confirm the transaction by clicking a
“Donate Now” button at the end of the page.
8

F IGURE 2. Screen shot of the donation page.

2.2. Treatments. Our experimental intervention pertains to the donation page depicted in Fig-
ure 2. For each website visitor who enters the donation page, we exogenously vary the donation
and codonation amounts that are displayed by default in the respective fields of the donation form.
We randomize independently in the two different treatment dimensions. In the donation dimension,
we assign potential donors to one of four different treatments. Specifically, when arriving at the
donation page, the amount displayed in the project donation field is either zero, or corresponds
to a pre-specified donation level of e10, e20, or e50. Note that in each case, donors are free to
contribute any positive amount by simply typing in the desired contribution level into the project
donation field.
The three positive default values correspond, respectively, to the 25th, 50th, and 75th percentile
of all donations on the platform during the six months before our experiment started. This allows
us to examine whether defaults have stronger or weaker effects on behavior when they are set
relatively high or low compared to what most people would donate otherwise.7 In contrast to the
positive donation defaults, the zero treatment implements an active-decision or “forced-choice”
environment: A user who wants to make a donation in this treatment has to actively specify the
amount she wishes to contribute. If a donor tries to finalize the transaction while the donation field
is set to zero, an error message appears and the donor is redirected to the donation form. Active-
decision environments are sometimes argued to have desirable properties, e.g., if preferences in the
7
Coincidentally, rather than by design, the different default amounts also correspond to modes in the historical
distribution of donations (as well as in the active-decision environment). This is the case since—as we will see in
further detail in Section 3—many donors contribute “round” or “prominent” amounts such as e5, e10, e20, etc. It
is unclear whether this may dampen or magnify the impact of defaults on behavior. On the one hand, defaults may
have little traction to increase the mass of donations at these modes, given that relatively many people are already
giving these amounts. On the other hand, it may be easier for defaults to attract potential donors from more “unusual”
donation levels to the common default amounts (e.g., follow a default of e10 or e20 instead of giving e14), or to
make potential donors “jump” from one prominent amount to another.
 9

population are very diverse (see Carroll et al. 2009 or Sunstein 2013). In our empirical analysis
below, this treatment will provide us with a benchmark of actively determined donations, against
which we can compare donors’ behavior in the treatments with positive donation defaults.
There is a second sense in which our setting involves active decision-making: contributions, and
thus potentially also the default donation levels, only become effective after users actively confirm
the transaction. While this is typical for how defaults—or, more specifically, “default options”—
are implemented in a wide variety of online applications, it differs from the use of defaults in other
settings like organ donor registration or 401(k) savings plans. In these environments, defaults—
or what might be coined “default rules”—are typically implemented as a set of rules that are
relevant for the decision maker even if she remains entirely passive. While this difference may
seem subtle, it is potentially important for understanding the channels through which defaults can
affect behavior. In particular, as we will discuss in further detail in Section 3.4 below, the degree
to which present-biased preferences and procrastination of active decisions might affect outcomes
differs between the two different types of default regimes.
In our second treatment dimension, codonations, we independently vary the pre-specified
percentage add-on to support the online platform. Specifically, we randomly assign donation-
page visitors to codonation defaults of 5%, 10%, or 15%. These treatments were chosen based on
historical values and heuristics. The 5% default was the value originally used by the platform.
It was retained in the experiment as a “control” group representing the status quo before the
beginning of the intervention. The remaining two values were chosen based on the intuition that
the codonation is likely to be perceived as a “tip” to Betterplace. The codonation defaults of 10%
and 15% were implemented as they correspond to the tipping conventions in Germany and places
like North America, respectively.

2.3. Implementation of the Experiment. The experiment was conducted over an 11-month
period from June 08, 2012 to April 19, 2013. Overall, we observe roughly 680,000 donation-
page visitors during this period, distributed over the 12 different treatment cells in our 4 × 3 design
(see Table 1 for an overview). Some aspects of our data and procedures are worth noting. First, to
avoid technical errors in the settlement of payments, our experiment is confined to situations where
10

the remaining required budget for the respective project element is at least e50 (i.e., the highest
possible default). Second, to rule out that a small number of extreme contributions may distort our
results, we drop the top 0.2% of donors (n=41) for our empirical analysis.8
Third, we randomize website visitors into treatments at the “website-session” level, such that
they remain in the same treatment throughout their visit to the platform. This minimizes the
possibility that a potential donor who visits more than one donation page—because she wishes to
make multiple donations or browses several project and donation pages before ultimately making a
donation decision—is exposed to different treatments. More precisely, we assign treatments when
a user enters a donation page for the first time. Subsequently, a browser cookie ensures that the
user keeps being exposed to the same treatment. While we cannot perfectly ensure that a donor
never faces another treatment (e.g., when she makes donations from two different computers), this
procedure minimizes donors’ awareness of the experiment and possible treatment spillovers.

Donation Codonation default

default 5% 10% 15% Total
AD 56,894 56,959 56,807 170,660
e10 56,739 57,014 57,017 170,770
e20 56,777 57,083 57,117 170,977
e50 57,183 56,985 57,335 171,503
Total 227,593 228,041 228,276 683,910
TABLE 1. Treatments and number of observations per treatment. Notes: “AD”
denotes the active-decision environment.

Our final sample covers 683,910 observations—roughly 57,000 in each of the 12 treatment cells
(see Table 1). In 99.7% of cases, one observation in the table corresponds to a unique website visitor
or “session”: the 683,910 observations correspond to 681,660 unique sessions. This is the case
since relatively few donors make more than one donation during our observation period. Table A.1
in the appendix indicates that observations are balanced in terms of baseline characteristics for
which we have information, namely whether the potential donation was for a fundraising event, a
8
Each of these donors contributes e2,165 or more. This compares to a median donation level of e20. Some of the
figures reported below (e.g., the exact values of the average donation and codonation amounts) naturally depend on the
specific cutoff used. Unless explicitly noted otherwise, however, our main results and conclusions remain unchanged
when applying different cutoff levels (e.g., excluding the top .1%, .5%, or 1% of donors).
 11

project, or an element within the project, and whether or not potential donations to a given project
are tax deductible (which is typically the case when the charity is registered in Germany).
In total, we observe 22,792 donations coming from 20,542 different participants. Among those
who do make a positive donation, 92.5% (n=19,010) make a single donation in the period of our
intervention, 5.5% (n=1,125) make two donations, and the remaining 2% (n=407) make 3 or more
donations. In what follows, we separately include the different individual donations as our unit
of analysis for participants who make multiple donations within a session. To control for possible
dependencies of observations, all estimation results reported below are clustered at the website-
session level. This clustering has almost no bearing on our empirical findings. Results are also
robust to using alternative approaches to account for donors with multiple contributions (e.g., using
the sum of donations or focusing only on the first decision for each donor).
Table 2 provides an overview of the fraction of participants who make a donation (denoted as
the “donation rate” in what follows), the amount donated to the charitable cause, and the amount
of codonations (for a more detailed overview of summary statistics by treatment, see Table A.1 in
the appendix). Over the course of the experiment, we observe 22,792 donations, corresponding
to an overall donation rate of 3.33%. At first glance, this rate seems relatively low, given that
participants in our experiment are all individuals who visited the online platform, browsed through
the website and, at least at some point, clicked on the “Donate Now” button. According to the
platform providers, however, this figure is in line with historical levels of the donation rate on the
platform. Relatively low donation rates are common in the charitable-giving literature in general
(see, e.g., Karlan and List 2007, Falk 2007, Huck and Rasul 2011), and in online fundraising
more specifically. In a study based on the online fundraising sites of 84 nonprofit organizations,
for instance, M + R Benchmarks (2015) reports a median overall conversion rate on the charities’
websites of 0.76%. Conditional on visiting the organizations’ main donation page, the median
donation rate reported in the study is 13%. We can only speculate on what drives the relatively
lower donation rate in our setting, compared to the last figure. One possible explanation is that we
are studying a “marketplace” for charitable projects where platform visitors may have more diffuse
donation intentions compared to potential donors who visit the website of a specific charitable
12

organization. The latter might also attract a relatively higher fraction of donors who arrive at the
website in response to solicitation emails or other fundraising drives of the charity. Indeed, we also
observe considerably higher donation rates for participants who come to the platform through an
organized “fundraising event” (see Section 2.1); in this case, donation rates are roughly 10-11%.9
As a consequence of the low donation rate, the modal action of participants in our experiment
is not to donate. This holds for all of our treatments (see Table A.1 in the appendix). Conditional
on making a donation, the average (median) donation level in our sample is e51.27 (e20). The
corresponding values for codonations are e2.00 and e0.25, respectively. In sum, these numbers
yield a total of e1.17 million in terms of donations and roughly e45,500 in codonations over the
course of our experiment.

Variable No. Obs. Mean Median SD

Donation rate 683,910 0.033 0.00 0.18
Donation amount 22,792 51.274 20.00 117.17
Codonation amount 22,792 1.998 0.25 7.02
TABLE 2. Summary Statistics.

3. E MPIRICAL R ESULTS

In this section, we present the result of our experiment. We begin, in Section 3.1, by studying
treatment effects in terms of individual donor behavior. In particular, we analyze how the different
defaults affect the distributions of donations and codonations in the experiment. In Sections 3.2
and 3.3, we turn to an aggregate-level perspective and examine the influence of defaults on average
donation and codonation revenues, respectively. We also explore how treatment differences on
the intensive and extensive margin of the donation and codonation distributions can account
for the observed aggregate-level outcomes. We conclude, in Section 3.4, by discussing which
psychological mechanisms may account for donors’ reactions to defaults in our experiment.
9
Reassuringly, our main findings (i.e., strong distributional effects for both donations and codonations, but significant
average treatment effects only in terms of codonations) hold both for participants who respond to fundraising events as
well as those who visit the donation pages of aid projects or project elements. Our empirical analysis in Section 3 thus
concentrates on the pooled data set that includes all participants. In those cases where we find systematic differences
between different donor types for more specific results, we note this explicitly.
 13

3.1. Do Defaults Affect Individual Donor Behavior? In a first step, we examine how defaults
influence individual donation and codonation patterns across treatments. Does the presence of
defaults lead to systematic bunching of donors at the respective default amounts? The answer to
this question is a clear yes. To illustrate this point, we examine the distributions of donations and
codonations, focussing first on the 22,792 cases in which participants in our experiment actually
make a donation. Table 3 summarizes the distribution of donations across treatments. Each column
in the table corresponds to a different treatment cell, denoted by the corresponding default values
for the donation amount and codonation percentage, (de, c%). In the rows of the table, we depict
the fraction of donations in a given treatment that correspond to one of the default donation levels,
e10, e20, and e50, as well as the fraction of donations that differ from these values.
The highlighted cells reveal a strong impact of defaults on individual donations. The likelihood
of making a donation of e10, e20, or e50 is considerably more pronounced when the respective
amount is selected as the default donation level. For instance, 22.9%, 22.8%, and 21.7% of
donors make a contribution of e10 in the three treatment cells where this amount is the default
donation value (see columns 1-3 of Table 3). This compares to only 12-14% of donors making
a e10-contribution when facing a default of e20 or e50 (columns 4-9). Similar effects can be
found for each of the nine treatments that involve a positive default contribution. Comparing the
highlighted fractions of donors who stick to the different defaults to the corresponding numbers in
the treatments where donors have to make an active decision (AD in columns 10-12) shows that
setting the default to a certain value increases the proportion of donors who actually contribute
this amount by roughly 5-10 percentage points. Given that the observed baseline values for the
different donation levels in the active-decision environment lie between 10 and 17%, this implies
that defaults increase donors’ propensity to make the corresponding contribution by 30-90%.
 14

Treatment (de, c%)

(10,5) (10,10) (10,15) (20,5) (20,10) (20,15) (50,5) (50,10) (50,15) (AD,5) (AD,10) (AD,15)
Donated amount (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
e10 .229 .228 .217 .128 .128 .124 .135 .139 .132 .173 .164 .146
e20 .124 .109 .114 .233 .211 .238 .105 .109 .110 .119 .124 .132
e50 .106 .115 .104 .094 .110 .095 .196 .183 .176 .122 .115 .122
other .541 .549 .565 .546 .552 .544 .564 .569 .582 .586 .597 .600

TABLE 3. Donations by treatment. Notes: The table depicts the proportion of donors in a given treatment who contribute e10,
e20, or e50, as well as the fraction of donors who donate a different amount (“other”). Each column in the table corresponds to a
different treatment cell, denoted by the corresponding default values for donations / codonations.

Treatment (de, c%)

(10,5) (10,10) (10,15) (20,5) (20,10) (20,15) (50,5) (50,10) (50,15) (AD,5) (AD,10) (AD,15)
Codonation (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
0 .421 .451 .439 .413 .432 .448 .426 .495 .475 .455 .480 .486
5% .530 .138 .121 .532 .147 .133 .530 .121 .141 .502 .125 .144
10% .033 .385 077 .023 .399 .082 .022 .363 .071 .022 .373 .071
15% 0 .003 .337 .003 .002 .308 .001 .005 .290 .001 .003 .280
other .017 .024 .027 .029 .020 .030 .021 .017 .023 .020 .020 .019

TABLE 4. Codonations by treatment. Notes: The table depicts the proportion of donors in a given treatment who make a
codonation of 0, 5%, 10%, 15%, as well as the fraction of donors who codonate a different amount (“other”). Each column in the
table corresponds to a different treatment cell, denoted by the corresponding default values for donations / codonations.
 15

AD 10
.3

.3
Relative Frequency

Relative Frequency
.2

.2
.1

.1
0

0
0 10 20 50 100 0 10 20 50 100
Donated Amount Donated Amount

20 50
.3

.3
Relative Frequency

Relative Frequency
.2

.2
.1

.1
0

0 10 20 50 100 0 10 20 50 100
Donated Amount Donated Amount

F IGURE 3. Donation distributions by default donation level. Notes: The figure depicts
the relative frequencies of donations for each of the treatments in the donation dimension
(indicated by the panel titles). Default donation levels are highlighted by the dashed lines.
The x-axes of the graphs are censored at e100 (the underlying data are not).

The strong influence of defaults on individual donor behavior is also evident in the overall
distribution of donations. In Figure 3, we present histograms for the active-decision regime and
the three different donation defaults. To facilitate illustration, we right censor the x-axis of the
graphs at e100 and focus our attention on the donation-default dimension. More precisely, we
plot the histograms for subsamples in which we pool observations across the different codonation
treatments, holding the treatment assignment in the donation dimension constant.
The histograms underscore the strong impact of defaults on donation patterns. While the
distributions otherwise look relatively similar—e.g., we observe more or less pronounced spikes
in donations at multiples of e5—there is a marked difference in the proportion of donations at the
default values (indicated by the dashed lines). Indeed, the figure shows that the modal contribution
16

always corresponds to the default donation level. Kolmogorov-Smirnov tests indicate that the
distributions of donations differ significantly across the four different default regimes (p<0.01 for
all pairwise tests).
The systematic influence of donation defaults on the distribution of donations is also evident
when considering the 12 individual treatment cells separately. Figure B.4 in the online appendix de-
picts the full set of histograms for all individual treatment cells. When comparing the distributions
of treatment pairs that differ in terms of donation defaults, but have identical codonation defaults
(i.e., when testing across “columns” within a given “row” of Figure B.4), all but one treatment
comparisons are statistically significant (p=0.138 when comparing (AD,5) vs. (10,5); p<0.05 for all
other pairwise treatment comparisons). At the same time, the distributions of donations generally
do not differ significantly when holding the donation default constant, but varying the default in the
codonation dimension (i.e., comparing the rows within a given column of Figure B.4): only one out
of twelve pairwise treatment comparisons turns out to be significant at the 10% level (Kolmogorov-
Smirnov tests; p=0.084 when comparing the (20,5) and (20,10) treatments).
In a next step, we study how defaults affect behavior in our second treatment dimension—the
add-on contribution to support the online platform. Table 4 depicts the codonation frequencies
across treatments, mirroring the analysis of donations in Table 3.10 The highlighted cells indicate
that defaults also have a pronounced impact on individuals’ behavior in terms of add-on contribu-
tions. For instance, moving from a 5% to a 10% default increases the proportion of donors who
make a 10% contribution from roughly 2-3% to 35-40% (see the third row of Table 4). Another
noteworthy feature of Table 4 is that participants’ choices in the codonation dimension exhibit a
bimodal pattern, with 40-50% of donors in a given treatment making no codonation at all and
another 30-50% of donors sticking exactly to the respective default amount. Comparing differences
in the distributions of codonations using Kolmogorov-Smirnov tests shows that the distributions

10
The corresponding codonation histograms can be found in Figure B.5 in the online appendix. We display individual
treatment cells instead of histograms for subsamples that pool across donation defaults, since Kolmogorov-Smirnov
tests indicate a number of significant differences between individual distributions (e.g., the codonation distribution for
the (10,15) treatment in the third row of Figure B.5 turns out to differ significantly from the (AD,15) as well as the
(50,15) treatment; p<0.01 in both cases).
 17

differ significantly for all pairwise tests of individual treatments that differ in the codonation default,
but have the same default donation (p<0.01 in all cases).
Our data also indicate that donors’ propensity to stick to defaults in the two different choice
dimensions is highly correlated. In particular, the conditional likelihood of donating the default
amount is almost 80% higher for donors who also stick to the default in the codonation dimension
(the respective likelihoods are 28.7% vs. 16.1%; p<0.01). This suggests that some people in
our sample are systematically more affected by defaults than others. It is not to say, however,
that we generally observe no default effects for those participants who actively deviate from the
default in one of the decision dimensions. For instance, among donors who actively opt out of the
codonation default, we still observe bunching at the default for donations: relative to the active-
decision environment, their propensity to donate the stipulated default amount increases by 10-
40%.11 We will return to the discussion of “types” that are generally more likely to stick to defaults
in Section 4.

3.2. Do Defaults Affect Average Donation Levels? In a next step of our analysis, we explore
how defaults affect average donation amounts at the aggregate level. Figure 4 presents average
donation levels across treatments, calculated based on all 683,910 observations in our data set,
i.e., including donors as well as those participants who opted out of the donation process without
making a contribution. Average donations in the different treatments lie in a range between
e1.54 and e1.85 (for more details, see also Table A.1 in the appendix). The confidence intervals
marked at the top of each bar indicate that the observed differences across treatments are generally
insignificant. If we consider all pairwise treatment comparisons that are possible given our 12
different treatment cells, we find that only 1 out of the 66 pairwise t-tests is significant at the 5%
level, and 3 further treatment pairs differ at the 10% level. Specifically, the average donation level
in the (10,5) treatment is significantly lower than in the (50,15) treatment, and weakly lower than
in the (AD,15) and the (10,15) treatment (t-tests accounting for clustering of standard errors at
the session level; p=0.030, p=0.076, and p=0.081, respectively). In addition, contributions in the
11
For further illustration, Figure B.6 in the online appendix depicts separate donation histograms for individuals who
do vs. do not stick to the codonation default. In Section B.1 of the online appendix, we further discuss how people are
affected by the specific default tuples in different treatment cells.
18

(50,15) treatment are marginally higher than in the (AD,5) treatment (p=0.080). The p-values of
all other 62 treatment comparisons, however, are well above conventional levels of significance.

2
1.8
1.6
Average Donation (EUR)
1.4
1.2
1
.8
.6
.4
.2
0

(AD,5)

(AD,10)

(AD,15)

(10,5)

(10,10)

(10,15)

(20,5)

(20,10)

(20,15)

(50,5)

(50,10)

(50,15)
Treatment

F IGURE 4. Average donation by treatment. Notes: The figure depicts average donation
levels across the 12 different treatments, calculated based on all participants in the experi-
ment. 95% confidence intervals, accounting for clustering of standard errors at the session
level, are presented at the top of each bar.

Most importantly, we observe no systematic influence of different donation defaults on average

contribution levels. For instance, average donations under a e10 donation default (bars 4-6 in
Figure 4) are very similar to those in the active-decision environment (cp. the three leftmost bars
in Figure 4). On average, participants in the AD-treatments contribute e1.69. This compares
to e1.70, e1.68, and e1.77 in the treatments with a e10, e20, and e50 donation default,
respectively; see Table 5. As is the case for the comparison of individual treatment cells, these
differences in average contributions for the “pooled” subsamples are not statistically significant
(p>0.3 for all treatment comparisons).12
12
Despite the lack of statistical significance, the relative difference between the observed donation averages might
still seem sizable in economic terms (e.g., contributions increase by roughly 5% between the active-decision envi-
ronment and the e50 default). One has to bear in mind, however, that—our high overall number of observations
notwithstanding—small treatment differences in the number of “top donors” with very high contributions can still
have a non-negligible impact on the precise values of the average donation levels in a given treatment. For instance,
if instead of dropping only the top 0.2% of donors, we would restrict our attention to donors whose contributions
do not lie more than three standard deviations above the overall mean across all donors (adopting an approach by
Edwards and List 2014), the average donation level in the active-decision environment (e1.45) would actually lie
 19

Treatment (Donation Default)

AD e10 e20 e50
(1) Donation rate (%) 3.35 3.39 3.35 3.23
(2) Av. donation (overall) 1.69 1.70 1.68 1.77
(3) Av. donation (donors only) 50.29 50.16 50.17 54.59
(4) Median donation (donors only) 20 20 20 25
(5) No. Obs. 170,660 170,770 170,977 171,503
(6) No. donors 5,725 5,795 5,727 5,545

TABLE 5. Summary statistics by default donation level. Notes: The table gives an
overview of donation behavior for different donation defaults (subsamples pooled across
codonation treatments).

Interestingly, some of the bars in Figure 4 seem to suggest that—for a given donation default—
average donation levels tend to increase in the codonation default. While the effect is relatively
modest and generally not statistically significant, it turns significant for one treatment comparison
if we pool observations across the different default donation levels (in particular, donations under
a 15% codonation default turn out to be significantly higher than under a 5% default (p=0.025) in
pooled data).
One might worry that the overall low donation rate in our sample (i.e., the high number of
“zero contributions” from platform visitors who end up making no donation) could bias our results
towards not finding statistically significant treatment differences at the aggregate level. To address
this potential concern, we repeat our analysis with restricted subsamples in which we drop x% of
observations for each treatment (all of which involving contributions of zero). One way to interpret
this exercise is to assume that x% of participants in our experiment were only browsing the online
platform without an inclination to actually make a donation. Doing so for various cutoff levels,
e.g., x=10, 25, or 50, we generally find no significant differences in average donation levels across
treatments (see Figure B.7 in the online appendix for a more detailed summary of our analysis).
In the most extreme scenario, we only keep 3.39% of participants per treatment. This implies
that we solely retain the 5,795 actual donors in the e10 treatment (where the donation rate is
exactly 3.39%; see Table 5), and no more than 275 non-donors in each of the remaining treatments.
slightly above the value for the e50 treatment (e1.43). Put differently, to conclude that the e50 default systematically
increases donations relative to the active-decision environment, one would need to believe that this effect is driven by
the default’s impact on the fraction (or contribution levels) of participants who donate very high amounts of money.
While we cannot conclusively rule out such an effect, it seems implausible.
20

Nevertheless, we still cannot reject the null hypothesis of no difference in average donation levels
at different donation defaults (the lowest p-value for all pairwise treatment comparisons in this case
is 0.282). In other words, even at an (implicitly) assumed donation rate in the range of 95-100%,
the differences in average donations across treatments are not statistically significant.
Despite the substantial individual-level reactions described in Section 3.1, defaults in our
experiment have no systematic impact on average donation levels. Figure 5 explains how both
of these findings can be reconciled. In the figure, we show how behavior under a given donation
default changes relative to the active-decision environment, and relative to the treatments involving
other default specifications. The three frames in the top row of the figure depict the differences
in the distributions of donations between the active-decision environment and the e10, e20, and
e50 default, respectively. Simply put, we “subtract” the upper-left panel of Figure 3 from the three
other histograms depicted in Figure 3, while additionally taking into account potential differences
in the proportion of non-donors (i.e., a bar at 0). This allows us to examine how defaults affect
the distributions of donations along both the intensive and extensive margin, relative to the active-
decision environment. The frames in the second and third row of Figure 5 depict the corresponding
pairwise differences in the distributions of donations (and non-donors) for different default donation
levels.
If defaults are poles of attraction for people’s behavior but there are no significant differences
in average donation levels, then it must be the case that defaults induce some people to donate
more than they otherwise would have, while others donate less or not at all, such that the two
countervailing effects cancel each other out at the aggregate level. This is exactly what we find.
Figure 5 demonstrates that, relative to the active-decision environment, people move towards the
default from both above and below for each of the different default donation levels.13 For instance,
the spike of additional people donating e20 when this is the default (middle panel in the top row
of Figure 5) comes “at the cost” of fewer people donating e5, e10, e25, e50, and e100.

13
The statistical significance of these movements towards the default is further examined in Section B.2 of the online
appendix.
 10 vs. AD 20 vs. AD 50 vs. AD

.4
.4
.4

.2
.2
.2

0
0
0

-.2
-.2
-.2

Difference (p.p.)
Difference (p.p.)
Difference (p.p.)

-.4
-.4
-.4
0 10 20 50 100 0 10 20 50 100 0 10 20 50 100
Donated Amount Donated Amount Donated Amount

20 vs. 10 50 vs. 10

.4
.4

.2
.2

0
0

-.2
-.2

Difference (p.p.)
Difference (p.p.)

-.4
-.4

0 10 20 50 100 0 10 20 50 100
Donated Amount Donated Amount

50 vs. 20
.4
.2
0
-.2
Difference (p.p.)
-.4

0 10 20 50 100
Donated Amount

F IGURE 5. Change in donation distributions due to defaults. Notes: Each panel of the figure depicts differences (in percentage
points) between the distributions of donations under different donation defaults (as indicated in the panel titles). Default donations
21

are indicated by the dashed lines. Extensive-margin differences in the proportion of non-donors are depicted at zero.
22

Notably, at higher default values, the mass of people who can be “pulled down” by the default
becomes smaller and smaller (recall that the e50 default corresponds to the 75th percentile in
the distribution of historical donations as well as in the AD treatments). As a result, one might
reasonably expect average donation levels to go up. The panels in the right-most column of
Figure 5, however, show that there is a second countervailing effect that works against such an
increase. In particular, under a e50 default, we observe a higher fraction of participants opting
out of the donation process altogether. The donation rate in the treatment with a e50 default is
3.23%. This compares to values of 3.35-3.39% in the remaining treatments (see row 1 of Table 5).
Linear-probability models that analyze the propensity of making a donation across treatments show
that the drop in the donation rate at the e50 default is statistically significant. The corresponding
p-values are p=0.018 when comparing the e50 treatment to all other treatments, and p=0.077 (e50
vs. AD), p=0.022 (e50 vs. e10), and p=0.093 (e50 vs. e20), respectively, for individual treatment
comparisons. While the drop in the donation rate might seem modest in size, it suffices to offset
the increase in donations at the intensive margin that we observe at the e50 default.14 As a result,
we again observe no significant treatment differences in average donation levels.15
In sum, our analysis shows that defaults are important poles of attraction for donors’ contribution
decisions. We observe strong bunching of donations exactly at the respective default in a given
treatment, but no systematic changes in the frequency of contributions in the neighborhood of
the default amount. Defaults tend to push up the contributions of some donors, even if they do
not induce non-donors to become donors. At the same time, they tend to pull others’ donations
down. At relatively low default values, these two effects seem to operate entirely on the intensive
margin. At higher defaults, we find that defaults can also lead to an reduction in donation rates on
the extensive margin. In both cases, the countervailing effects essentially cancel each other out,

14
Focussing only on the subset of participants who do make a donation (row 3 of Table 5), we indeed find a significant
increase in the average donation level at the e50 default (p=0.068, p=0.064, and p=0.061 when comparing the e50
treatment to the e10, e20 and AD treatment, respectively).
15
Interestingly, while this aggregate-level result holds for all of the different donation types, the mechanisms behind
the result are somewhat different for the group of participants who come to the platform in response to an organized
fundraising event (see Section 2.1). In particular, within this group of participants, we observe no significant drop in
the donation rate at the e50 default. Instead, the countervailing effects in this treatment also operate entirely on the
intensive margin—i.e., a decrease in the number of donors who give even higher amounts.
 23

leading to small and statistically insignificant average treatment effects for the different donation
defaults in our experiment.

3.3. Do Defaults Affect Average Codonations? The picture is quite different when considering
overall codonation levels. Figure 6 presents average codonation amounts by treatment for our full
sample (for further information on codonation levels in the subsample of participants who make
positive donations, see Table A.1 in the appendix). The saw-shaped pattern shows that, holding
the donation default constant, codonation revenues increase monotonically for higher codonation
defaults. The 95% confidence intervals presented at the top of each bar indicate that for most of the
relevant pairwise comparisons, these differences are statistically significant. In particular, average
codonation levels are always significantly higher at the 15% relative to the 5% codonation default
(t-tests accounting for clustering at the session level, p<0.01 in all cases). With the exception of the
treatments that involve a e20 donation default, this also holds when comparing the 10% and the
5% codonation treatments (p=0.417 for (20,5) vs. (20,10); p<0.01 in the remaining cases). When
comparing the 15% and the 10% codonation treatments, we find that codonations do not differ
significantly in the active-decision environment (p=0.458), whereas the differences are significant
for the treatments with positive donation defaults (p=0.062, p=0.012, and p=0.001 for the e10,
e20, and e50 donation default, respectively).16
The magnitude of the observed differences in codonation levels is substantial. For the 15%
codonation default, overall codonation levels are roughly 80% higher than under the 5% default and
still lie about 30% above the values for the 10% treatments. Comparing the codonation revenues to
the overall donation levels in the corresponding treatments underscores this effect. When facing a
5% codonation default, participants on average make an add-on contribution to the online platform
that amounts to 2.94% of their donation. This value increases to 3.88% and 4.78%, respectively,
under the 10% and 15% codonation defaults. Given that the donation levels themselves are not
lowered by higher codonation defaults (see Figure 4), our findings indicate that higher defaults

16
When comparing average codonation levels under different donation defaults, we find no systematic evidence for a
spillover from donation defaults to codonation behavior. In only one case, codonations are significantly higher than in
another treatment that features the same codonation default, but a different donation default (specifically, codonations
in the (20,5) treatment are higher than in (AD,5) [p=0.048] and weakly higher than in (50,5) [p=0.064]).
24

.1
Average Codonation (EUR)
.08
.06
.04
.02
0

(AD,5)

(AD,10)

(AD,15)

(10,5)

(10,10)

(10,15)

(20,5)

(20,10)

(20,15)

(50,5)

(50,10)

(50,15)
Treatment

F IGURE 6. Average codonation by treatment. Notes: This figure describes average

codonation levels across the 12 different treatments. 95% confidence intervals, accounting
for clustering of standard errors at the session level, are presented at the top of each bar.

in the codonation dimension increase overall revenues for the online platform without hampering
donations to the charitable cause.
Table 4 in Section 3.1 as well as Figure B.5 in the online appendix illustrate how participants’
reactions to codonation defaults bring about this positive overall effect. Notably, we find that donors
essentially never deviate from a codonation default in order to make a higher contribution to the
platform. Across all treatments, the fraction of donors doing so is at most 6%. Furthermore, we
only observe a modest increase in the proportion of donors who opt out of making a codonation
altogether when facing higher default values. The corresponding fraction changes from 42.9%
in case of a 5% default to 46.4% and 46.2% for the 10% and 15% default, respectively. While
this increase of about 3 percentage points is statistically significant (p<0.01 in both cases),17 it
is far from being able to offset the boost in codonations that is caused by the roughly 30-35%
of additional donors who make a 10% or 15% codonation when facing these values as a default
contribution (see the bold figures in Table 4). This implies that most of the behavioral reactions

17
These tests are based on linear-probability estimations that compare the propensity of making an add-on contribution
for the different codonation defaults, controlling for potential differences across the donation-default treatments.
Standard errors are clustered at the session level.
 25

to defaults in the codonation dimension happen on the intensive margin, with movements to the
default from below. As a result, we not only observe strong individual-level effects of defaults, but
also substantial increases in overall revenues in the codonation dimension.

3.4. Why Do Defaults Affect Behavior? A natural question to ask in view of our empirical
findings is why defaults matter in our setting. Although our experiment is not designed to pin
down the precise mechanisms through which defaults affect behavior, our data do permit some
informed speculation regarding the psychological mechanisms at work. The literature on default
effects has identified numerous mechanisms that may cause people to stick to defaults, such as
status-quo biases, attentional limitations, or a tendency to procrastinate (see Dinner et al. 2011
and Sunstein 2013 for comprehensive reviews of the literature). In what follows, we briefly assess
the relevance of some frequently discussed candidates in light of our empirical findings. A more
detailed discussion of various mechanisms and their predictions for our setting can be found in
Online Appendix C.
First, while our data looks as if deviating from the default were costly for some agents, it
seems highly unlikely that the treatment differences in our experiment can be explained by direct
(neoclassical) transaction costs of opting out of the default. For one thing, these costs are essentially
zero in online applications, since consumers are in an environment where alternative choices are
just “one click away”. For another, the direct costs of altering the donation amount seem negligible
in comparison to the other costs that donors incur in order to finalize the transaction, such as filling
out the payment details in the donation form.
Second, since we are dealing with an environment where defaults only become relevant in
the final stage of a sequence of active choices, explanations based on present-biased preferences
and procrastination seem of limited relevance in our setting. Specifically, while a tendency to
procrastinate active decision making may contribute to the low overall donation rate that we
observe, it seems unlikely that consumers bear the short-run costs of actively going to the platform,
selecting a project, etc., but then procrastinate on determining the actual donation amount.
Third, the finding that defaults have no effect on overall donation revenues is inconsistent with a
class of psychological mechanisms that predict a monotone increase in average donation levels at
26

higher default values. As we explain in Section C in the online appendix, these mechanisms include
explanations based on anchoring as well as simple models of reference-dependent preferences,
consumer inattention, or information transmission and recommendations.
More involved variants of these models—e.g., featuring non-linear gain-loss utility or allowing
for more general information structures—may be able to rationalize our data. The same holds for
some formalizations of the idea that defaults may signal or directly shape prevailing social norms.
All of these more involved formulations, however, can rationalize a very wide range of behavioral
responses. In this sense, they lack meaningful predictive power.
In sum, none of the predictive mechanism mentioned above is able to account for all of our
main empirical findings. Yet, our data suggest that defaults systematically affect people’s choices,
and that some individuals are systematically more prone to stick to defaults than others. These
individuals thus behave as if deviating from the default were costly.18 In the next section, we will
examine this more closely by analyzing whether the platform could make use of the heterogeneity
in individuals’ reactions to defaults, in order to increase donation revenues.

4. P ERSONALIZED DONATION DEFAULTS

Our results above show that defaults can be used to increase codonations, but not donations.
Defaults in the donation dimension serve as strong attractors, but they make some people donate
more than they would otherwise have, and others donate less or not at all. On aggregate, it is a
wash. This problem would obviously not arise if the online platform could personalize defaults so
that some donations are pushed up but none are pulled down.
In this section we ask whether Betterplace could make use of such a personalization strategy
to increase donations revenues. A natural starting point is a reduced-form approach, investigating
whether some types of donors are more likely to stick to defaults than others; and whether there
are heterogeneous treatment effects in donation levels under different defaults, which might in
principle be exploited to target defaults based on trackable individual characteristics. A drawback
to this approach is that many personal characteristics of interest are only observable for individuals
18
A fixed as-if cost of deviating from the default could also rationalize why we observe stronger aggregate-level effects
in the codonation dimension in which stakes are smaller and people have a relatively low baseline inclination to
contribute (cp. Section 3.3).
 27

who end up making a donation, but not for potential donors who visit the website. More generally,
the scope for personalizing defaults is limited on (publicly accessible) online fundraising sites
since charities—often motivated by privacy and transparency concerns—typically cannot observe,
or do not track, potentially relevant individual characteristics. This is in contrast to some “offline”
settings (such as alumni fundraising), and it is certainly a limiting factor given the data architecture
of Betterplace. In focusing on donor characteristics, our reduced-form approach is therefore bound
to ignore responses along the extrinsic margin. That being said, for all realized donations, we can
track the calendar time of the donation; whether the donation was towards a fundraising event, a
project, or an element within the project; whether the donor was a registered user; and, among
registered users, whether a donor has donated on multiple occasions during the observation period,
i.e., is a repeat donor. In addition, in the donation form, donors provide their first names, from
which we use a name recognition algorithm to deduce their gender.
We explore donors’ propensities to stick to defaults by estimating linear probability models in
which the dependent variable is a dummy variable equal to 1 if a donor sticks to a positive donation
default or—in an alternative specification—if she sticks to both the donation and the codonation
defaults. The estimates (presented in Table B.1 of the online appendix) indicate that donors are
not significantly more likely to stick to positive donation defaults in the year-end holiday season
(December); donors who contribute as part of a group fundraising event are 2-3 percentage points
less likely to stick to the default relative to those who contribute to a particular element or project
(p<0.05); female donors are 2-3 percentage points more likely to stick to a donation default than
males (p<0.01); and registered users as well as repeat donors are 1-4 percentage points less likely to
stick to donation defaults than their unregistered or non-repeat counterparts, respectively (p<0.05
and p<0.01). This pattern is qualitatively identical when it comes to sticking to both the donation
and codonation defaults, indicating again that some types of donors are generally more likely to
stick to defaults than others.
These findings suggest that there may be some scope for personalizing defaults on the basis
of gender, donation type and frequency, and user registration. But it doesn’t say much about
what that default should be. Heterogeneous responses to different defaults in terms of donation
28

levels are potentially more informative, but our reduced-form results are not very promising in
this regard. More specifically, although we observe statistically significant level effects—women,
notably, donate e9 less on average than men, and donations are e26 higher in December than
at other times of year—none of the interactions between these characteristics and the donation
defaults are statistically significant at the 5% level (see Table B.2 in the online appendix).19
Our reduced-form results indicate that it may, in principle, be possible to increase donations
by targeting defaults based on personal characteristics, but that this strategy requires substantially
richer data on donation histories and personal characteristics, not to mention large sample sizes.
More generally, personalizing defaults in this manner is unlikely to be a successful strategy given
the data constraints on Betterplace and other charitable-giving websites. In the remainder of this
section, therefore, we adopt a structural approach that has more modest and arguably more realistic
data requirements for the personalization of donation defaults. The model we build requires that
the charity can store historical data on individual donations. The thought experiment we wish
to conduct is the following. Suppose that the platform can first track individuals’ donations in
a default-free environment akin to the active-decision treatment (which was the status quo for
Betterplace prior to our experiment).20 This information can be used to recover the donors’
underlying “generosity”—how much they are inclined to donate in the absence of a default. The
platform can then use this information to personalize defaults, ensuring that they never set a default
that is below a donor’s baseline generosity level.
To personalize defaults in this manner, one needs to predict how individuals would respond to
different default donation levels. The structural exercise below accomplishes this by setting up a
model, in Section 4.1, in which donors differ in their generosity levels. They also differ in terms of
the “as-if” costs they face when either deviating from the default to a different donation amount,
or opting out of donating altogether. Structural estimates of the distributions of donors’ underlying

19
In line with our aggregate-level results for the subset of participants who make a donation (see Table A.1 and Foot-
note 14), contributions under the e50 default are significantly higher than in the AD treatment in some specifications
of Table B.2. This result, however, neglects the extensive-margin reduction in donation rates under the e50 default,
illustrating again the limitations of focussing only on the intensive margin of donations.
20
This kind of tracking is technically feasible in many online settings, by requiring one-step logins upon website
entry—e.g., through linked social media accounts—or, as a second-best alternative, using cookies to track IP-addresses.
 29

generosity and as-if costs, described and derived in Sections 4.2 and 4.3, are then used to make out-
of-sample predictions of donations under different defaults, which are personalized as a function
of a potential donor’s underlying generosity, as captured by his or her past contribution in the
AD environment. Based on these predictions, we are able to examine, in Section 4.4, whether
Betterplace could increase aggregate donation revenues by personalizing defaults.

4.1. A Simple Model of As-If Costs. In this section, we study a stylized model to derive a
potential donor’s optimal contribution in the presence and absence of a default. In so doing, we
remain agnostic about individuals’ behavioral motivations for adhering to defaults (see Section 3.4
for a discussion of these). Instead, we set up a simple model of “costly opt-out”—in the tradition of
Carroll et al. (2009) and Bernheim, Fradkin, and Popov (2015)—in which individuals who deviate
from the default incur as-if costs that can stem from a variety of possible underlying psychological
or economic mechanisms.
In order for this model to be useful, its predictions must match three key empirical features
of our data regarding distributional differences under different defaults, summarized at the end
of Section 3.2. First, defaults generate bunching at the default but not in its neighborhood, and
movements to the default come from both sides of the default. Second, donation rates in the AD
environment are not lower than those in the different default treatments: Defaults do not induce non-
donors to become donors. Third, high but not low defaults lead to movements along the extensive
margin, reducing the donation rate by inducing some potential donors to opt out of the donation
process altogether.
Formally, let x ≥ 0 be the donation made by an individual to the charitable cause. We suppose
that there is a stable underlying trait—donor generosity ρ—that determines how much an individual
donates in the absence of a default.21 In the AD environment, without defaults, an individual of
type ρ ≥ 0 maximizes her donation utility V(x, ρ). We suppose that

x2
(1) V(x, ρ) = ρx − .
2

21
In as much as this trait is not stable, we are bound to overestimate the benefits from personalizing defaults based on
donors’ past donations under an active-decision policy.
30

This structure enables us to uncover the generosity type ρ from observing the chosen donation x in
the AD environment, as the utility-maximizing donation in this case is simply x = ρ.
Now consider an agent who faces a default d > 0. We suppose that, independent of the default, an
ungenerous type (ρ = 0) obtains a utility of 0 when opting out of donating altogether. Conversely,
a generous type (ρ > 0) receives a fixed “opt-out utility” of −α when making no donation (with
α ≥ 0). Intuitively, a generous type may feel bad when donating nothing. An individual who
deviates from the default (x , d) but still donates a positive amount (x > 0) incurs a deviation cost
δ ≥ 0, so that her overall utility is V(x, ρ) − δ.
Note that the above structure allows us to capture the three key features of the empirical donation
distributions in the different treatments of our experiment. First, a person donating a positive
amount will either stick to the default—thus avoiding the deviation cost δ—or donate an amount
equal to her generosity level ρ. This implies that defaults increase the frequency of donations
exactly at the default amount from above or below the default, but that donors are not drawn to
other positive donation amounts. Second, an ungenerous agent (ρ = 0) cannot be induced to give
a positive amount. This is in line with our observation that positive defaults do not increase the
observed number of donors. Third, defaults may induce donors to opt out of the donation process
altogether. This is the case when the fixed cost of doing so (α) is low relative to both the utility
V(ρ, ρ) − δ of giving one’s preferred amount, and the utility V(d, ρ) of sticking to the default. This
captures the extensive-margin reduction in donation rates at higher defaults.
Simple algebra establishes that the optimal donation xo ≥ 0 for a generous agent with ρ > 0 in
the presence of a default option d > 0 is given by


d if V(d, ρ) > V(ρ, ρ) − δ and V(d, ρ) ≥ −α







xo = 

(2) ρ if V(ρ, ρ) − δ ≥ V(d, ρ) and V(ρ, ρ) − δ ≥ −α






0 if max{V(ρ, ρ) − δ, V(d, ρ)} < −α




To simplify notation, let

ρ2 + d2
∆(ρ, d) ≡ V(ρ, ρ) − V(d, ρ) = − ρd.
2
 31

We assume that there is a share λ1 of agents who act as if deviating from the default is costless.
These agents experience no deviation costs δ and therefore always donate their preferred amount
ρ.22 The remaining 1 − λ1 share of agents has positive deviation costs. The optimal donation
decision of agents who face deviation costs will depend on their generosity level. For relatively
generous agents, it is never optimal to opt out of making a donation altogether. Specifically, for
agents with a generosity level ρ ≥ d/2, sticking to the default yields utility ρd − d2 /2 ≥ 0, whereas
the utility from opting out and making no donation is −α < 0. These relatively generous agents
will therefore either stick to the default or donate their preferred amount ρ, depending on whether
their deviation costs δ are larger or smaller than ∆(ρ, d). Agents with relatively low generosity,
0 < ρ < d/2, also determine their choice of sticking to the default or deviating to their preferred
amount depending on which side of the cutoff, ∆(ρ, d), their δ lies. For low-generosity agents,
however, the utility of donating d and the utility of deviating to ρ may both be smaller than the opt-
out utility −α. When this is the case, it is optimal for low-generosity agents to opt out of donating
altogether. Finally, ungenerous agents with ρ = 0 will never make a positive donation.

4.2. Estimation. There are three unknown parameters in this model: generosity types ρ, deviation
costs δ, and opt-out costs α. A key advantage of our data is that we can identify the distribution
function f (ρ) non-parametrically from the observed donation distribution in the AD treatment,
which features no default and therefore entails no deviation costs.
Modeling costs requires more structure, since costs are not directly observed. In keeping with
Bernheim, Fradkin, and Popov (2015), we allow for heterogeneous deviation costs that follow an
exponential distribution. Specifically, we assume that conditional on belonging to the share 1 − λ1
of agents that have positive deviation costs (whom we index by z = 1), the cost δ of deviating from
the default to a positive donation amount is distributed according to the cumulative distribution
function Φ, where 
−λ2 δ
for δ ≥ 0,

(1 − e ),


Φ(δ|z = 1) = 


for δ < 0.

0,


22
Note that these agents might still be subject to opt-out costs (α), but the latter are irrelevant for agents’ choices. This
is because agents can costlessly deviate from the default to their preferred donation amount ρ, which guarantees strictly
positive utility.
32

Once again following Bernheim, Fradkin, and Popov, we further assume that ρ and δ are indepen-
dently distributed. As for the costs of opting out of the donation process altogether, we assume
that making no donation entails no costs for ungenerous types (α = 0 for agents with ρ = 0). For
generous types (ρ > 0), the opt-out cost α is distributed according to the cumulative distribution
function Ω, where 
(1 − e−λ3 α ), for α ≥ 0,



Ω(α|ρ > 0) = 


for α < 0,

0,


and is distributed independently of δ and ρ.

Given that f (·) is already non-parametrically identified from the AD treatment, the estima-
tion problem boils down to identifying the three parameters of the model that define the cost
distributions—a proportion λ1 which is inure to deviation costs; and the parameters of the
exponential distributions λ2 and λ3 , which define the deviation costs and opt-out costs, respectively.
We estimate these parameters by maximizing a log-likelihood function of the following form:

N
X
L(λ) = log(Pr(xi |d, λ, f (·)))
i=1

where i = 1, ..., N are the individual observations in the treatments with positive donation defaults.
Section D.1 in the online appendix provides a detailed derivation of the log-likelihood function. In
essence, the likelihood function in our setting consists of the different cases involved in the optimal
donation decision described in Equation (2), weighted by the corresponding probabilities with
which they occur, given the (estimated) parameters of the model. For example, the probability of
observing an individual donating e15 in a treatment with a e50 default depends on the prevalence
of individuals with a generosity type ρ = 15, the fraction of individuals that are subject to deviation
costs (1 − λ1 ), as well as the distributions of the deviation and opt-out costs (determined by λ2 and
λ3 ).
Note that the log-likelihood function is only defined if f (ρ) takes on positive values for all ρ. As
the empirically observed donations in the AD treatment take discrete values and the data becomes
sparse at donations above e300, we restrict the sample in our estimation to ρ ∈ [0, 300] and
 33

“smooth” our data by assigning donations to integer bins, such that each bin has positive mass.
The observations used in the estimation amount to 99.9% of the total sample. Parameters λ =
(λ1 , λ2 , λ3 ) are identified through changes in the donation distribution under the different donation
defaults in our experiment, relative to the AD treatment (cp. Figure 5).

4.3. Estimation Results. Maximum likelihood estimates for λ are presented in Table A.2 of the
appendix. The estimate for λ1 indicates that 89% of potential donors are inure to, or simply
ignore the default. This proportion, though seemingly high, is entirely in line with the empirically
observed responses to defaults in our experiment. In particular, recall that the proportion of donors
who contribute the default amount under different donation defaults increased by roughly 5-10
percentage points relative to the AD environment (cp. Table 3). Default options thus substantially
affect the behavior of some individuals, but they leave a majority of potential donors untouched.23
Among the 11% of potential donors who do incur costs, the estimates for λ2 and λ3 indicate that
deviation costs and opt-out costs are rather high.24 High deviation costs imply that agents who are
subject to these costs are inclined to stick to the default, generating modes in the distributions of
donations at the corresponding default values. At the same time, the estimate for λ3 is larger than
the estimate for λ2 , indicating that opt-out costs tend to be smaller than deviation costs. The upshot
of this is that opting out of donating altogether may be preferable to donating a non-default amount,
especially for potential donors who are less generous. At high enough defaults, this generates
movements along the extensive margin.
The model with these parameter estimates performs remarkably well. As Figure 7 shows, the
fitted model successfully reproduces all of the key features of the data from our experiment: it
generates the empirically observed modes at the e10, e20, and e50 defaults, reproduces the

23
This finding might seem surprising in light of the evidence that defaults affect a large share of the population in
applications like retirement savings or organ-donor registration. Note, however, that some of the potential mechanisms
behind default effects in these settings are, by design, less relevant in ours (such as hassle costs of opting out or present-
biased procrastination; cp. Section 3.4). Hence, our relatively high estimate for λ1 may indicate that procrastination is
indeed a major driver of default effects in those other settings.
24
These costs are measured in utils and are therefore not directly interpretable in monetary terms. Figure D.1 in the
online appendix, however, gives a sense of what “high” means in this context by plotting the ∆(ρ, d) functions under
the three different defaults, as well as the mean and median of δ implied by our estimates. The figure shows that for
agents of type ρ ≤ e158 (168) [198], the median deviation costs are high enough to make the agents stick to a default
of e10 (e20) [e50], rather than donating their preferred amount ρ.
34

.3
10 20

.3
Relative Frequency

Relative Frequency
.2

.2
.1

.1
0

0
0 10 20 50 100 0 10 20 50 100
Donated Amount Donated Amount

50
.3
Relative Frequency
.1 .2

Data
0

0 10 20 50 100 Simulation
Donated Amount

F IGURE 7. Fitted versus Actual Donation Distribution. Notes: The figure compares
the empirically observed distributions of donations in the experiment to (simulated) fitted
distributions, using the non-parametric distribution f (·) from the AD treatment for integer
values of ρ and the maximum likelihood estimates of λ from Table A.2 in the appendix.
For the simulated ρ’s, a random sample of the distribution function f (ρ) is drawn with
N=500,000. For illustrative reasons (i.e., to avoid extreme spikes at 0 due to the high share
of non-donors), the analysis focuses on a subsample containing 3.5% of the overall sample
in each treatment (all actual donations and the corresponding number of zero contributions
per treatment), excluding donations above 300.

decrease in the donation rate (i.e., the spike at zero) under the e50 default, and closely matches the
empirical distributions elsewhere.

4.4. Personalized Defaults. Using the parameter estimates for λ, we can now make out-of-
sample predictions regarding how donation revenues would change under a system of personalized
defaults. Specifically, we examine how the platform could optimally condition defaults on an
individual’s (past) donation level in the absence of defaults, captured in our model by the parameter
ρ. As is immediately clear from our derivations above—as well as from the empirical observation
 35

that defaults pull some donors’ contributions down relative to the AD treatment—it never makes
sense to set the donation default for a given individual below his or her generosity type ρ. We
therefore simulate two types of personalized defaults. The first is additive: an individual of type ρ
is assigned a default of ρ + a, where a ≥ 0. The second is proportional: an individual of type ρ is
assigned a default of ρ · b, where b ≥ 1.

2 2

1.9 1.9
Mean donation in Euros

Mean donation in Euros

1.8 1.8

1.7 1.7

1.6 1.6

1.5 1.5

1.4 1.4
10 20 30 40 50 60 70 80 1 1.5 2 2.5 3 3.5 4
a b

( A ) Additive: d(ρ) = ρ + a ( B ) Proportional: d(ρ) = ρ · b

F IGURE 8. Donations under Personalized Defaults Notes: This figure plots means and
confidence intervals of donations under different personalized default-setting policies, based
on simulated data. Personalized defaults are determined by adding a constant a to the agent’s
baseline generosity level ρ (left panel) or multiplying it with a a factor b (right panel).
Agents’ response to the personalized default is simulated for a range of different values
of a and b. The confidence interval accounts for uncertainty in the estimated parameters.
We re-estimate the λ parameters with bootstrap samples from the e10,20, and 50 default
treatments and replicate this a 1000 times. For each of these estimated λ parameters, we
simulate the agents’ response to the personalized defaults and construct a confidence interval
around these means. The simulations are carried out with a large sample of ρ’s drawn from
the AD treatment to approximate well the distributions of the as-if costs.
.

Figure 8 furnishes our model’s predictions of mean donation levels under personalized defaults
for the additive (Panel A) and proportional case (Panel B). Under the additive default option,
donations are maximized at a∗ = 31.3. The optimal scaling factor for the proportional personalized
default is b∗ = 2.0.25 Under these defaults, our model predicts that overall donation revenues
25
The optimal add-on and scaling factor may seem high, but they follow naturally from our parameter estimates in the
previous section, our model specification, and the empirical results from Section 3: since a majority of potential donors
is generally inure to defaults, it doesn’t matter for them when this default if high. For the rest, a high personalized
default may lead some to completely opt out of the donation process, but those that don’t are induced to contribute a
36

would increase by at most 6.2% relative to the AD environment (4.7% in the additive and 6.2%
in the proportional case). While this increase in donation revenues seems economically relevant,
we think of it as an upper bound for the potential gains to be made from personalization in our
setting. Specifically, if the donors’ generosity level fluctuates over time or varies by project, the
gains from personalized defaults are bound to be lower. Moreover, our simple functional form
of individuals’ donation utility V(·) implies that increasing an potential donor’s default to up to
twice her preferred amount does not trigger an extensive-margin reaction. The fact that we see a
sharp drop in predicted revenues for proportional scaling factors above 2 suggests that individuals
may drop out somewhat earlier if we allowed for a more flexible specification of their preferences.
Hence, despite our model’s ability to replicate key data patterns of our experiment, the potential
benefits of personalizing defaults in our setting seem rather limited.

5. C ONCLUSION

We conclude by discussing practical implications of our findings for charitable organizations

and providers of online donation platforms. Most importantly, our results highlight the possibility
that defaults may have both desired and undesired effects on the distribution of donations and
overall donation revenues. They also demonstrate that defaults may have an influence on people’s
decisions, even if this influence might not be directly apparent in aggregate-level data.
Both observations caution against a simplistic use of defaults based on the notion that “defaults
work”. This, of course, does not to imply that positive defaults may never increase donation
revenues. The use of personalized or adaptive defaults seems promising in this respect, but our
results from Section 4 suggest that it is challenging to successfully increase donations through
personalized defaults in a setting like ours. While our reduced-form results indicate heterogeneity
across groups in terms of potential donors’ proclivity to stick to defaults as well as in the overall
propensity to contribute, we find no compelling evidence of heterogeneous donation responses to
defaults. The results from our structural estimates are not much more encouraging. They indicate
higher amount. Optimal personalized defaults are those that best manage the tradeoff between driving some people
out and others up. This can be readily seen in the case of proportional personalization. The scaling factor of 2 implies
that nobody is induced to opt out entirely under their personalized default, since individuals with ρ ≥ d/2 always stay
in the donation process (see Section 4.1). At the same time, individuals with high deviation costs δ will stick to the
default, thereby generating higher revenues for the charity.
 37

that personalized defaults have the potential to avert downward movements in donations by setting
defaults neither too high nor too low for a given donor, but they also suggest that successful
personalization requires much richer data. While the data constraints that limit our analysis in
this respect are, at present, shared by many other online charitable giving platforms, better tracking
of data from donors and their reactions to different features of the platforms—as well as linked data
from other sources—might eventually make such an approach feasible.

R EFERENCES

Abadie, Alberto and Sebastien Gay. 2006. “The Impact of Presumed Consent Legislation on
Cadaveric Organ Donation: a Cross-country Study.” Journal of Health Economics 25 (4):599–
620.
Adena, Maja and Steffen Huck. 2016. “Online Fundraising, Self-deception, and the Long-term
Impact of Ask Avoidance.” WZB Discussion Paper No. SP II 2016-306 .
Adena, Maja, Steffen Huck, and Imran Rasul. 2014. “Charitable Giving and Nonbinding
Contribution-level Suggestions: Evidence from a Field Experiment.” Review of Behavioral
Economics 1 (3):275–293.
Altmann, Steffen and Armin Falk. 2011. “The Impact of Cooperation Defaults on Voluntary
Contributions to Public Goods.” Working Paper, University of Bonn .
Altmann, Steffen, Armin Falk, and Andreas Grunewald. 2013. “Incentives and Information as
Driving Forces of Default Effects.” IZA Discussion Paper No. 7610 .
Andreoni, James. 2006. “Philanthropy.” Elsevier, 1201 – 1269.
Bekkers, René and Pamala Wiepking. 2011. “A Literature Review of Empirical Studies of
Philanthropy: Eight Mechanisms That Drive Charitable Giving.” Nonprofit and Voluntary Sector
Quarterly 40:924–973.
Berndt, Ernst R, Bronwyn H Hall, Robert E Hall, and Jerry A Hausman. 1974. “Estimation and
Inference in Nonlinear Structural Models.” In Annals of Economic and Social Measurement,
Volume 3, number 4. NBER, 653–665.
38

Bernheim, B Douglas, Andrey Fradkin, and Igor Popov. 2015. “The Welfare Economics of Default
Options in 401 (k) Plans.” American Economic Review 105 (9):2798–2837.
Beshears, John, James J. Choi, David Laibson, and Brigitte C. Madrian. 2008. “The Importance of
Default Options for Retirement Saving Outcomes: Evidence from the United States.” In Lessons
from Pension Reform in the Americas, edited by Stephen J. Kay and Tapen Sinha. Oxford: Oxford
University Press, 59–87.
Blackbaud. 2018. “Charitable Giving Report: How Fundraising Performed in 2017.” Tech. rep.
Carroll, Gabriel D., James J. Choi, David Laibson, Brigitte C. Madrian, and Andrew Metrick. 2009.
“Optimal Defaults and Active Decisions.” Quarterly Journal of Economics 124 (4):1639–1674.
Cialdini, Robert B and David A Schroeder. 1976. “Increasing Compliance by Legitimizing Paltry
Contributions: When even a Penny Helps.” Journal of Personality and Social Psychology
34 (4):599.
Desmet, Pierre and Fred M Feinberg. 2003. “Ask and Ye Shall Receive: The Effect of the Appeals
Scale on Consumers’ Donation Behavior.” Journal of Economic Psychology 24 (3):349–376.
Dinner, Isaac, Eric J Johnson, Daniel G Goldstein, and Kaiya Liu. 2011. “Partitioning Default
Effects: Why People Choose not to Choose.” Journal of Experimental Psychology: Applied
17 (4):332.
Ebeling, Felix. 2013. “Non-binding Defaults and Voluntary Contributions to a Public Good—Clean
Evidence from a Natural Field Experiment.” Working Paper, University of Cologne (66).
Edwards, James T. and John A. List. 2014. “Toward an Understanding of Why Suggestions Work
in Charitable Fundraising: Theory and Evidence from a Natural Field Experiment.” Journal of
Public Economics 114:1–13.
Ekström, Matthias. 2018. “The (Un)compromise Effect.” NHH Discussion Paper .
Falk, Armin. 2007. “Gift Exchange in the Field.” Econometrica 75 (15):1501–1511.
Fiala, Lenka and Charles Noussair. 2017. “Charitable Giving, Emotions, and the Default Effect.”
Economic Inquiry 55 (4):1792–1812.
Fosgaard, Toke R. and Marco Piovesan. 2015. “Nudge for the (Public) Good: How Defaults Can
Affect Cooperation.” PLoS One, December 30, 2015 .
 39

Fraser, Cynthia, Robert E Hite, and Paul L Sauer. 1988. “Increasing Contributions in Solicitation
Campaigns: The Use of Large and Small Anchorpoints.” Journal of Consumer Research
15 (2):284–287.
Frey, Bruno S. and Stephan Meier. 2004. “Social Comparisons and Pro-Social Behavior: Testing
“Conditional Cooperation” in a Field Experiment.” American Economic Review 94 (5):1717–
1722.
Goswami, Indranil and Oleg Urminsky. 2016. “When Should the Ask Be a Nudge? The Effect of
Default Amounts on Charitable Donations.” Journal of Marketing Research 53 (5):829–846.
Haggag, Kareem and Giovanni Paci. 2014. “Default Tips.” American Eonomic Journal: Applied
Economics 6 (3):1–19.
Huck, Steffen and Imran Rasul. 2011. “Matched Fundraising: Evidence From a Natural Field
Experiment.” Journal of Public Economics 95 (5):351–362.
Johnson, Eric J., Steven Bellman, and Gerald L. Lohse. 2002. “Defaults, Framing and Privacy:
Why Opting In , Opting Out.” Marketing Letters 13 (1):5–15.
Johnson, Eric J. and Daniel Goldstein. 2003. “Do Defaults Save Lives?” Science 302:1338–1339.
Johnson, Eric J., John Hershey, Jacqueline Meszaros, and Howard Kunreuther. 1993. “Framing,
Probability Distortions, and Insurance Decisions.” Journal of Risk and Uncertainty 7:35–51.
Karlan, Dean and John A. List. 2007. “Does Price Matter in Charitable Giving? Evidence from a
Large-Scale Natural Field Experiment.” American Economic Review 97 (5):1774–1793.
Levav, Jonathan, Mark Heitmann, Andreas Herrmann, and Sheena S. Iyengar. 2010. “Order in
Product Customization Decisions: Evidence from Field Experiments.” Journal of Political
Economy 118 (2):274–299.
Löfgren, Ä, Peter Martinsson, Magnus Hennlock, and Thomas Sterner. 2012. “Are Experienced
People Affected by a Pre-set Default Option—Results from a Field Experiment.” Journal of
Environmental Economics and Management 63 (1):66–72.
M + R Benchmarks. 2015. “The 2015 M + R Benchmarks Study.” Tech. rep.
Madrian, Brigitte C. and Dennis F. Shea. 2001. “The Power of Suggestion: Inertia in 401(k)
Participation and Savings Behavior.” Quarterly Journal of Economics 116:1149–1187.
40

Messer, Kent D., Homa Zarghamee, Harry M. Kaiser, and William D. Schulze. 2007. “New
Hope for the Voluntary Contributions Mechanism: The Effect of Context.” Journal of Public
Economics 91 (9):1783–1799.
Reiley, David and Anya Samek. 2017. “Round Giving: A Field Experiment on Suggested
Charitable Donation Amounts in Public Television.” Working Paper, University of Southern
California .
Shang, Jen and Rachel Croson. 2009. “A Field Experiment in Charitable Contribution: The
Impact of Social Information on the Voluntary Provision of Public Goods.” Economic Journal
119 (540):1422–1439.
Simonson, Itamar. 1989. “Choice Based on Reasons: The Case of Attraction and Compromise
Effects.” Journal of Consumer Research 16 (2):158–174.
Smith, Sarah and Mark Ottoni-Wilhelm. 2018. “Gender Differences in Goal Setting: Evidence
from Fundraising.” Working Paper, University of Bristol .
Sunstein, Cass R. 2013. “Deciding by Default.” University of Pennsylvania Law Review 162 (1):1–
57.
Thaler, Richard H and Cass R Sunstein. 2008. Nudge: Improving decisions about health, wealth,
and happiness. New Haven: Yale University Press.
Weyant, James M and Stephen L Smith. 1987. “Getting More by Asking for Less: The Effects of
Request Size on Donations of Charity.” Journal of Applied Social Psychology 17 (4):392–400.
 A PPENDIX

Treatment (De, C%)

(AD,5) (AD,10) (AD,15) (10,5) (10,10) (10,15) (20,5) (20,10) (20,15) (50,5) (50,10) (50,15) F-Stats
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)
Donations
(1) Donation rate (%) 3.34 3.27 3.45 3.35 3.45 3.38 3.31 3.39 3.35 3.14 3.35 3.21
(2) Av. donation (overall) 1.59 1.68 1.79 1.54 1.76 1.81 1.63 1.71 1.70 1.71 1.74 1.85
(3) Av. donation (donors only) 47.65 51.44 51.76 45.90 51.10 53.44 49.31 50.33 50.85 54.53 51.82 57.54
(4) Median donation (donors only) 20 20 20 20 20 20 20 20 20 23 25 25

Codonations
(5) Conation rate (%, all) 1.82 1.70 1.77 1.94 1.89 1.90 1.94 1.93 1.85 1.80 1.69 1.69
(6) Conation rate (%, donors only) 54.50 52.04 51.43 57.91 54.89 56.12 58.68 56.77 55.25 57.39 50.55 52.52
(7) Av. codonation (e, all) 0.042 0.070 0.076 0.045 0.069 0.086 0.060 0.067 0.090 0.043 0.061 0.090
(8) Av. codonation (e, donors only) 1.26 2.13 2.20 1.35 2.00 2.53 1.80 1.98 2.69 1.38 1.83 2.80

Project Background Characteristics

(9) Fundraising Event 0.118 0.116 0.117 0.118 0.116 0.116 0.115 0.116 0.117 0.114 0.117 0.118 0.658
(10) Project 0.339 0.339 0.338 0.338 0.338 0.338 0.340 0.340 0.336 0.337 0.339 0.336 0.827
(11) Element 0.542 0.544 0.545 0.544 0.545 0.546 0.545 0.543 0.548 0.549 0.544 0.546 0.620
(12) Tax Deductible 0.661 0.657 0.655 0.659 0.660 0.658 0.660 0.655 0.660 0.658 0.657 0.656 0.317

(13) N. Obs. 56,894 56,959 56,807 56,739 57,014 57,017 56,777 57,083 57,117 57,138 56,985 57,335
(14) N. donors 1,901 1,864 1,960 1,903 1,964 1,928 1,878 1,936 1,913 1,793 1,909 1,843

TABLE A.1. Summary Statistics. Notes: The top and middle panel provide an overview of key outcome variables in the different
treatments. The bottom panel provides an overview of observed project background characteristics. Rows (9)–(11) report the
proportion of observations in a given treatment which pertain to a fundraising event (e.g., a birthday or charity run), an aid project, or
a specific element within a project (e.g., a first aid kit within a Red Cross Project), respectively. Row (12) pertains to the proportion
of observations per treatment group which were eligible for a tax deduction (typically the case when the charity is registered in
Germany). The final column reports p-values from F-tests for treatment differences in background characteristics, based on separate
regressions of each of the characteristics on dummies for the different treatments.
41
42

Parameter Estimate Standard error

λ1 89.0×10−2 43.1×10−4
λ2 62.8×10−6 34.4×10−6
λ3 99.1×10−5 30.2×10−5
TABLE A.2. Parameter Estimates. Notes: Results obtained from maximum likelihood
estimation. Standard errors are computed numerically by using a variant of the Berndt et al.
(1974) algorithm to approximate the Hessian. Estimation are based on all observations with
ρ ∈ [0, 300], which accounts for 99.92% of the overall sample.