Social Recognition: Experimental Evidence from Blood Donors
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Social Recognition:
Experimental Evidence from Blood Donors*
Lorenz Goette Egon Tripodi
July 14, 2021
Abstract
Does social recognition motivate repeat contributors? We conduct large-scale experi-
ments among members of Italy’s main blood donors association, testing social recogni-
tion both through social media and in peer groups. We experimentally disentangle visi-
bility concerns, peer comparisons and simple ask effects, and we study how exposure to
different norms of behavior affects giving. Overall, asking increases giving by 20.4 per-
cent. However, a simple ask to donate is at least as effective as asks that emphasize social
recognition. Using peer groups, we show that visibility concerns do not generate addi-
tional donations, and we find that donations respond weakly and non-monotonically to
social norms of giving. In an original study and two subsequent replications, we find
no evidence that donors give blood for the prospect of their donations being observed
by others on social media. Overall, our findings caution against over-reliance on social
recognition to promote good citizenship.
Keywords: Prosocial behavior, blood donations, social recognition, field experiments,
social media, WhatsApp.
* Goette: University of Bonn and National University of Singapore; lorenz.goette@uni-bonn.de. Tripodi:
University of Essex and JILAEE; egontrpd@gmail.com. We thank Roland Bénabou for comments and sugges-
tions that helped improve the manuscript. We are also grateful to participants of the Belief Based Utility work-
shop, the ECONtribute workshop on Social Image and Moral Behavior, the WZB Economics of Philantropy
workshop, and a seminar at Milano Bicocca for helpful feedback. We acknowledge research funding from
the Deutsche Forschungsgemeinschaft through grants 411047502, CRC TR 224 (Project B07), and Germany’s
Excellence Strategy (EXC 2126/1 390838866). Giacomo Manferdini provided outstanding research assistance.
This study could have not been possible without the cooperation of Avis Toscana; we are especially grateful
to Adelmo Agnolucci, Filomena Autieri, Luciano Franchi and Donata Marangio, as well as the staff and di-
rectorship of each of the 67 local chapters of Avis Toscana that took part in the study. This paper combines
two studies, both pre-registered at the AEA Social Science Registry (#AEARCTR-0004890 and #AEARCTR-
0007266). Ethics approval was obtained for the first study from the German Association for Experimental
Economic Research (#jN7oi2yd), and for the second study from the Ethics Sub Committee 1 at the University
of Essex (ETH2021-0488). All studies were also reviewed and approved by the data protection office at Avis
Toscana. All errors are our own.
Well functioning societies rely on the contributions of good citizens, who routinely help
their community even if they are not monetarily compensated for the full social value of
their actions. These include mission oriented workers, community leaders and volunteers.
Complementary forms of compensation that render such transactions viable include fu-
ture career benefits, psychological utility from helping others and social rewards. An open
agenda investigates the motivational effects of non-monetary forms of compensation.
This paper is concerned with how social rewards shape good citizenship. We study how
one commonly used social reward – recognition – motivates repeat blood donors. Theories
of identity and social recognition emphasize conformity with a socially relevant group and
costly signaling of a socially desirable identity as reasons why people respond to social
recognition (Bernheim, 1994; Akerlof and Kranton, 2000; Bénabou and Tirole, 2006, 2011).
However, different forces generate a tension that makes unclear whether repeat contributors
are more sensitive to social recognition than occasional contributors. On the one hand,
stronger identification with a group can raise the stakes of norm adherence; on the other
hand, once a strong reputation is established, the marginal signal may carry little weight.1
Through partnership with the largest Italian association of blood donors—Avis—we lever-
age the infrastructure of one regional branch to conduct an ideal experiment that provides
clean causal evidence on how social recognition affects giving among repeat blood donors.2
We select eligible donors from Tuscany that can be contacted through WhatsApp, asking
them to donate blood in the following month. We offer them social recognition using either
of two main approaches: First, we follow a classical intervention that informs donors of
their peers’ recent engagement in a civic activity along with a promise that an update will
follow to make visible future activity (Gerber et al., 2008). We put this in practice by creating
random groups of twenty donors with whom participants in the experiment can relate to.
Second, we introduce a social media campaign that rewards participants who donate in the
study period by acknowledging their name as a recent donor on highly subscribed Face-
1 Seealso Levine (2021) for an illustration of how different types compositions in the population can make
good reputations more or less easy to lose.
2 Several features make the setting ideal: (i) we can eliminates concerns that results could be driven by
awareness of being observed by researchers and concerns of participants acting out of desire to please the
experimenter, (ii) we can ensure high levels of participants’ engagement with our intervention thanks to
the availability of official trusted communication channels, and (iii) we have direct access to administrative
records of blood donations from the regional health authorities that we can precisely link to experimental data.
2book public pages of Avis Toscana. We test social recognition against two natural bench-
marks, not being solicited and being solicited with a simple ask. We find that participants
give more in the social recognition treatments than when they are not solicited.3 However,
the simple ask is at least as effective at encouraging giving as any of the social recognition
interventions.
The social media recognition intervention replicates a study conducted two years prior,
where we also observed no effect of social recognition on giving. In that study, the prospect
of social recognition lead participants to give more than they would have without solicita-
tion, but a simple ask was at least as effective. The earlier study had some limitations that we
discuss. In particular, poor engagement of donors with the organization’s treatment emails,
which led us to estimate an intention-to-treat effect off of the 23 percent of compliers. In
the new study we replicate the original experiment and we vary whether eligible partici-
pants are contacted via email or WhatsApp. We obtain similar results in the replication of
the email experiment: engagement via email is low (17 percent email opening rate) and the
prospect of recognition on social media does not inspire giving any more than a simple ask.
The replication through WhatsApp instead delivers much greater engagement (91 percent
of participants read our messages) and finds that donations decrease with social recognition
relative to a simple ask (p = 0.023). These results were unexpected, but can be interpreted
as an overjustification effect of social image concerns. If agents care differently about being
seen as altruistic by others, actions signal both the agent’s prosocial type and their desire
to impress others. Increasing publicity of good actions backfires when it becomes clear that
good actions constitute image-seeking behavior (Bénabou and Tirole, 2006, p. 1665).4
The part of our intervention that leverages social recognition in groups of twenty donors
also sheds new light on mechanisms. First, we design the content of social recognition
treatments in two modules to disentangle the effects of peer comparisons from visibility.
Second, we construct groups of twenty donors also for treatment arms in which we do not
provide peer comparisons, which allow us to provide a clean test of status seeking theories
(Immorlica et al., 2017) after isolating mechanical mean reversion. Third, we leverage rich
3 To rule out that these effects could be explained as a pure reminder effect, we asked the organization to
contact all study participants through the same channel for a communication unrelated to any of the treat-
ments a couple of days prior to the intervention roll-out.
4 We thank Roland Bénabou for this helpful suggestion.
3experimental variation in the social norms of giving that subjects are exposed to to map the
giving schedule along the full support of possible descriptive norms. We find that varying
visibility in groups, in addition to peer comparisons, does not increase giving (p = 0.218).
Participants give more by 7 percent of a standard deviation (p = 0.011) if peer comparisons
place them below the group norm: 54 percent of this effect can be causally attributed to
exposure to peer information, while the remaining 46 percent of the effect is due to mechan-
ical mean reversion. Finally, we do not observe a strong linear response to different social
norms. Estimating the treatment effects of different quintiles of social norms indicate a non-
monotonic response, with subjects being most motivated to donate in response to social
norms that are closer to the median level of giving.
We causally investigate social proximity as a moderating factor of social recognition. Half
the participants are randomly assigned to twenty-donors groups that include people from
all over Tuscany (distant), while the other half is assigned to groups that only include people
who donate at the same blood collection center and potentially know each other (close).
We show that greater social proximity to a group increases the chance that fellow group
members know each other, but does not make people more eager to impress other group
members with their donations and does not lead to a stronger response to the social norm
of giving within the group.
While social recognition did not generate the intended effects, our treatment messages
still increased overall giving by 20.4 percent during the study period and had positive
spillovers in the following months. We employ a Differences-in-Differences strategy that
leverages excluded local branches to rule out congestion effects, confirming that the inter-
vention generated excess donation. Moving to inter-temporal effects, we show that a simple
ask leads to harvesting from the most engaged donors for whom donations are crowded out
in the coming months, whereas social recognition appeals to all donors and leads to positive
spillovers on future donations. We discuss how social image concerns can lead to negative
immediate and positive delayed effects to reconcile the evidence.
This paper contributes to a broad literature on social recognition (see Bursztyn and Jensen,
2017, for a review).5 Few of these papers focus on settings where people already have a
5 Relevant to our work is the extensive literature on how prosocial behavior responds to image concerns
(Ariely et al., 2009), social norms (Frey and Meier, 2004; Fellner et al., 2013), and social pressures (DellaVigna
4strong identity and reputation as good citizens. Exceptions include Ager et al. (2017) who
study status seeking on the intensive margin of performance among World War II pilots and
Soetevent (2005) who study church offerings in a repeated experiment. Among repeat blood
donors, we provide surprising evidence that social recognition does not motivate people to
give. By compiling a meta-analyses of field experiments on the effects of social recognition
on acts of ’good citizenship’, we highlight that this study is the only evidence of significant
net crowding-out.
Our work complements a large literature on peer and social influence (e.g. Frey and
Meier, 2004; Shang and Croson, 2009; Krupka and Weber, 2009; Allcott, 2011; Ferraro and
Price, 2013; Kessler, 2017; Cantoni et al., 2019; Drago et al., 2020), by leveraging rich variation
in the behavior of groups of peers we are able to provide a test of status seeking preferences
(Immorlica et al., 2017) and we contribute with granular evidence on how donors respond
to different levels of descriptive social norms.6 We also provide the first causal test from
the field of social proximity as a moderator of social influence, which we find to be weak in
comparison to existing correlational results (e.g. Bond et al., 2012).7
In most countries, blood donors cannot be monetarily compensated and giving blood is
purely a civic-minded act (World Health Organization, 2009). Blood products are essential
in medicine and cannot yet be generated artificially (Shaffer, 2020). This landscape has led
to growing interest from social scientists in investigating blood donations as a measure of
social capital of a community (Guiso et al., 2004), and in addressing a key policy objective
by developing non-monetary instruments that can be applied to address blood shortages
(Heger et al., 2020). This study offers the first experimental evidence on the effects of so-
cial recognition on blood donations, and offers guidance on how natural features of local
et al., 2012). Support for these social influence mechanisms comes also from studies on other forms of good
citizenship, such as child immunization (Karing, 2018), energy conservation (Allcott and Rogers, 2014) and
voting (Gerber et al., 2008).
6 In comparison, existing research identifies adherence to social norms primarily from variation in informa-
tion about one level of average group behavior (e.g. Frey and Meier, 2004), or information about the behavior
of one peer in dyadic comparisons (e.g. Schultz et al., 2007). Papers that leverage rich group comparisons
include Allcott (2011); Ferraro and Price (2013); Beshears et al. (2015), but to the best of our knowledge no
existing paper maps the causal effects of different levels of group norms.
7 Goette and Tripodi (Forthcoming) and Bicchieri et al. (2021) also report evidence from more controlled
settings that is consistent with social proximity being a moderator of social influence.
5collection systems can be leveraged to harness social recognition.8
The rest of the paper is organized as follows. Section 1 describes the setting and research
design. Section 2 provides results. Section 3 contextualizes the results. Section 4 concludes.
1. Research design
1.1. Institutional setting
The study was conducted in partnership with the Tuscany branch of Avis (Avis.it) – the
largest Italian association of blood donors. Blood collection in Italy relies on blood donor
associations. In 2018, 92 percent of active donors in Italy are affiliated with a blood donor
association (Catalano et al., 2019). Blood donor associations are in charge of donor recruit-
ment and they support local hospitals in the scheduling of appointments for donations. In
some regions these associations also handle the collection of blood directly and re-sell inter-
mediate blood products to the health care system (this is not the case in Tuscany).
Working with Avis Toscana allows us to reach the vast majority of blood donors in the
region, and it gives us access to several official communication channels to contact donors.
Avis Toscana is also one of the few regional branches to have access to a rich data infras-
tructure that links administrative individual-level donation data from the universe of hos-
pitals and other blood collection points that donors have access to. In turn, this setting is
particularly suitable for investigations that combine experimental interventions to accurate
administrative data on actual donation behavior.
We conduct the main part of the experiment using Avis Toscana’s official WhatsApp ac-
count. With the support of Twilio (Twilio.com), we were able to leverage a new approach
for conducting experiments, using the WhatsApp Business API through which we can si-
multaneously contact very large numbers of registered donors with personalized messages.
This tool was introduced by WhatsApp in August 2018, and is primarily used by large firms
and organizations for personalized service communications with their customers and ben-
8 Seminal work from Lacetera and Macis (2010) uses non-linear social incentives to show with retrospective
data that social recognition affects blood giving. Meyer and Tripodi (2018) conduct a field experiment that
manipulates the visibility of pledges to give blood (but not giving itself).
6eficiaries.9 We started searching for alternative tools after experiencing the limitations of
email campaigns (see Section 1.3). This tool presents at least four substantive advantages
over sms, mail and email experiments: availability of reliable information on subject en-
gagement with the experiment, ease of conducting longitudinal studies, ease in establishing
trust with recipient through official verification of the organization’s account (green check
mark), and relatively low cost (4.70 USD every 100 messages).10
1.2. Mechanisms of social recognition
This experiment takes several step to understand how to harness social recognition in a
relevant setting. First, we take a classical social recognition intervention that bundles peer
comparisons and visibility in small social groups (Gerber et al., 2008), and we decompose
it to separately identify the two motivational mechanisms. Second, we introduce experi-
mental variation in the social proximity of the group of peers to whom donations are made
visible and with whom subjects compare their donation behavior. Third, we benchmark the
effects of social incentives to a simple ask. Fourth, we include in the experiment the repli-
cation of a study we conducted in 2019, which used social media to motivate donors with a
form of broader social recognition.
The experimental design comprises of five main treatments and is summarized in Fig-
ure 1. A few days before the experiment Avis conducted a short survey, unrelated to this
9 Facebook reviews all templates of messages that organizations want to send to their contact lists and does
not allow advertisement or mass campaigns.
10 The WhatsApp Business API allows organizations to reach out to their beneficiaries or customers only
when consent for being contacted is provided. An ideal feature of our setting is that all Avis donors provide
consent to be contacted by Avis for calls and initiatives that encourage them to donate blood.
We systematically review prior studies that have used WhatsApp as a channel to conduct an experiment.
From a Google Scholar search that we ran on April 20, 2021 with keywords “whatsapp” and “experiment”,
we scraped the 996 most relevant search results. We read the abstract of these articles and find that very few
of these studies are actually experiments. Moura and Michelson (2017) conduct a series of get-out-the-vote
experiments and manually send videos via WhatsApp to about 1,000 eligible voters. Hoffmann and Thommes
(2020) combine WhatsApp, and standard text messages to contact repeatedly 41 truck drivers over a two
months period. The closest complentary approach is introduced in Bowles et al. (2020), who use the ’broadcast
list’ feature of private WhatsApp accounts to contact 27,000 newsletter subscribers with non-personalized
messages. This method has the clear advantage that it does not require paying text message fees to WhatsApp.
However, it has the key disadvantages that messages cannot be tailored at the individual level, longitudinal
studies are impractical, and the reputational benefits of a verified business accounts are not available.
7particular study, which allows us to identify donors who use WhatsApp. Donors selected
to participate in the study are randomly partitioned in groups of twenty and treatment is
assigned at the group level. All experimental messages sent to donors encourage them to
donate in the month of March 2021. The No ask treatment serves as a pure control.11 The
Peer + Visibility treatment mirrors the social recognition intervention in Gerber et al. (2008),
providing a social comparison with fellow group members over donations made in the past
11 months and promises an image reward at the end of the month – by making March 2021
donations publicly observable within the group. The Peer treatment features a message
with the same contents except for the paragraph on the visibility incentive. In the Facebook
treatment, donors are informed that their donations of March 2021 will be acknowledged
broadly through the highly visited Facebook page of the organization. The Simple ask treat-
ment includes only the last paragraph of the message that is common to all treatments.
In addition, the groups of 20 donors that constitute the unit of randomization are con-
structed using one of two possible protocols: there are close groups composed of people
who typically donate at the same collection center, and distant groups composed of people
from all over Tuscany – a region of 3.7 million inhabitants spread over an area of 23,000
square kilometers (9,000 square miles). As an illustration, Figure A.1 plots the mailing ad-
dress of members of the two median average distance close and distant groups.
For two of the treatment messages, Simple ask and Facebook, we also randomized the com-
munication channel – between WhatsApp and email. In each treatment, half of the groups
are close and half are distant. Treatment assignment is stratified by gender, age groups, and
past donation behavior.
Procedures. A first screening for including Avis donors in the study was done before the
pre-intervention survey. We include only donors registered at one of the 65 largest local
Avis in the region, if their last donation was done in the last 5 years at a blood collection
center from which we can select at least 500 donors. We exclude from the study donors that
have not provided a phone number to Avis. This leaves us with 43,731 donors, a sample
covering 51.64 percent of active Avis Toscana donors (members who donated in the 5 years
11 We do not need an active control with a completely unrelated message, because donors are contacted by
the organization very infrequently and the survey sent a few days before the experiment already makes the
No ask condition play as if we had an active control.
8No ask
Simple ask
WhatsApp Facebook Facebook post
Peer
5/7
Peer + Visibility Group disclosure
Survey Experiment Donation data
2/7
Simple ask
Email
Facebook Facebook post
Spring ’21
Feb 26 – Mar 1 Mar 2 – Mar 4 Apr 1 Apr 8
Figure 1: Design overview
prior to the experiment).
Between February 26 and March 1 of 2021, our partner invites through the official What-
sApp account of Avis Toscana this pool of donors to answer a short survey. With this invite
donors are informed of an upcoming research study, and offered a simple procedure to
opt-out by simply replying with keyword “NORICERCA” into the WhatsApp chat. After
excluding donors that are not eligible to donate and did not receive the message (4,315) and
those who opted out prior to treatment (357) our final study sample includes 38,991 donors:
25,479 of which were assigned to being contacted via WhatsApp, 9,049 via email and 4,463
received no further message (No ask).
Over three days, from March 2 to 4, our partner sent all treatment messages out evenly
spread over business hours, randomizing the order at which messages were sent to different
donors.12 We programmed a bot that handled responses from donors and automatically
provided receipt of confirmation when subjects indicated preference to opt-out from the
research.
On March 31 our partner solicited feedback on the degree of satisfaction for the initiative
in a random sample of 10,000 subjects from treatments Simple ask, Facebook, Peer and Peer
12 Per WhatsApp usage limits, we could reach up to 10,000 distinct contacts per day.
9+ Visibility.13 On April 1, our partner shared with us individual-level donation data and
sent a message to all donors in the Facebook and Peer + Visibility treatments to remind them
that they could opt-out of their name being shared publicly next to their donation behavior
in March. We take opt-out requests until April 7. On April 8, participants in the Peer +
Visibility treatments receive tables reporting which members of their group have and have
not donated in March 2021 (see e.g. Figure A.3, panel a), and our partner posted on the
Facebook page of Avis Toscana the donations of donors in the Facebook treatments (see e.g.
Figure A.3, panel b). On April 8, our partner also sent a short follow-up question to 1,384
subjects in the Peer + Visibility treatment to measure their beliefs about how likely they think
it is that they know at least one member of their group, to confirm the differential social
relevance of the group composition in close vs. distant groups (see panel c of Figure A.5).
Sample. Table A.1 shows that strong predictors of donation behavior, such as past dona-
tion behavior and gender (women are allowed to donate less frequently), are well balanced
across treatments. In the treatments administered via WhatsApp the share of participants
that have received the message is very close to 100 percent, and about 90 percent of sub-
jects included in the study read the message. In comparison, the emailing opening rate is
much lower, around 17 percent. Interest in graphical contents that provide visual illustra-
tion of the social rewards is relatively low and varying across treatments. Finally, we see
virtually no opt-out from the research study after treatment (68 out of 38,991 participants),
though dropout is somewhat larger (between 0.2 percent and 0.4 percent) in social recog-
nition treatments administered through WhatsApp. All intention-to-treat results that we
present are unchanged if we include dropouts.
1.3. Recognition on social media
This experiment (conducted in 2019) investigates whether the prospect of social recognition
on social media encourages repeat blood donors to give blood when asked. The two main
treatments are comparable to the Simple ask and Facebook treatments in the experiment pre-
sented in Section 1.2, which are compared to a control group of donors who do not receive
13 Participants were asked for a feedback on a quantitative scale (from 0 to 10) for half of the sample and on
a qualitative scale for the other half. Exact wording of the question and results are presented in Figure A.2.
10any message. Randomization for this experiment is conducted at the Avis center level. The
objectives of this experiment were slightly broader, but beyond the scope of this particu-
lar paper. The full experimental design and set of findings for this study are presented in
Appendix B.
Procedures. Avis Toscana donors were included in the study if they satisfied the follow-
ing criteria: they had provided an email address and they were eligible to donate in Novem-
ber both blood and plasma. This leaves us with 15,326 donors, a sample covering 15.62
percent of active Avis Toscana donors (who donated in the 5 years prior to the experiment).
Between October 29 and October 31 of 2019, our partner sent out treatment messages via
email to this pool of donors. Donors were encouraged to donate in the month of November.
On December 1 we obtain donation data for the month of November, and email participants
to collect their opt-in consent for sharing their name on social media posts. On December
16 Avis makes the Facebook posts to recognize the donations of donors that have given
consent. The timeline of this experiment is described in Figure B.1. Online Appendix B
describes this experiment in greater detail.
Table 1 summarizes the main design differences between this initial experiment and the
more recent replications.
Table 1: Design differences
Original Email Replication Email Replication WhatsApp
Randomization Avis center level. 67 clusters. Twenty-donors group level. 681 Twenty-donors group level. 677
level clusters. clusters.
Privacy policy Opt-in at the end of study pe- Generic research opt-out offered at the beginning of the study period
riod. Donors choose between and specific privacy opt-out offered at the end of the study period.
revealing full name, first name Donors choose between revealing full name with last name initial and
only, neither. not revealing.
Active email re- 86.3% 77.0% 76.0%
cipients
Sample. Table B.1 shows that the final sample of 15,326 donors is well balanced across
treatments on age and gender. However, there are small imbalances in past donation be-
havior. We account for these by controlling for individual characteristics in econometric
specifications. Post treatment balance tests show that the opening rate of treatment emails
11is very similar across treatments. This is important in a design with treatment random-
ization at the level of Avis centers, where contact lists are maintained, because it rules out
differential quality of contact lists across centers as a confound.
1.4. Identifying (the mechanisms of) social recognition
To fix ideas, we present two simple conceptual frameworks that highlight our approach for
identifying social recognition and distinguishing its mechanisms. Our premise is that so-
cial recognition entails both elements of peer comparisons and social image concerns. Peer
comparison models typically resemble the following (see e.g. Bernheim, 1994; Immorlica et
al., 2017; Goette and Tripodi, Forthcoming):
U ( Di ) = ai Di − c( Di ) − sµi p( Di , D−i )
where agents experience consumption utility due to potentially heterogeneous joy of giving
and a private cost, and social utility that depends on the salience of social comparisons s.
The extent to which people care about social comparisons can also be heterogenous and the
social comparison function p(.) has been modeled in different ways. Perhaps most promi-
nent is the distinction between models of norm adherence where people experience benefits
to behave like others to fit-in p( Di , D−i ) = | Di − D−i |, and models of status seeking where
people care about being better than others p( Di , D−i ) = max ( D−i − Di , 0).
Models of social signaling (e.g. Bénabou and Tirole, 2006) feature a similar consumption
utility term. A key parameter of the social utility term is how visible the agent’s good
deeds are, and agents choose how prosocial to act in order to signal that they care about the
positive externalities of their actions (high a).
U ( Di ) = ai Di − c( Di ) − vµi r ( a| Di )
It has been hard in this literature to distinguish between social image concerns and peer
comparisons, primarily because equilibrium outcomes are sensitive to unobservable param-
eters.14 Our approach for this study is to test a social recognition intervention that features
14 The only paper we are aware of that can test social signaling separately from peer comparisons is Birke
12both visibility and peer comparisons, and experimentally manipulate visibility v (as in e.g.
Andreoni and Petrie, 2004; Ariely et al., 2009) and the salience of peer comparisons s (as
proposed by Cialdini et al., 1990) to distinguish mechanisms.
2. Results
We begin this section with an overview of the behavioral response to different forms of
social recognition. We then move to investigating status seeking and response to social norm
information, and to investigating heterogeneity in response depending on social proximity.
Finally, we shift our focus to a series experiments on leveraging social recognition through
social media.
2.1. Peer effects, visibility, and when a simple ask is enough
In the control treatment where donors receive no experimental message from the organi-
zation the average number of donations is 0.103.15 Offering donors the prospect of being
recognized on social media (Facebook) increased donations by 9.7 percent. Donations also
increase by when we offer donors peer comparisons (in Peer and Peer + Visibility), by in-
forming them of how much they donated recently relative to a random set of peers, and
we find that making prospective donations visible among group members makes little dif-
ference. However, none of these social rewards are more effective than a Simple ask, which
increases donations by 26 percent. Mean comparisons are summarized in Figure 2.
These effects are then estimated using the following linear regression framework:
Donationsi = α + ∑ β g 1{ γ = g } + ε i
g ∈T
where Donationsi is a count variable for individual donations during the study period
(March 2021) and T = { Facebook, Peer, Peer + Visibility, Simple ask}. The model we es-
(2020), who shows that anti-bunching (going the extra mile) in the presence of monetary bonuses for reaching
a donation target is a unique prediction of social signaling models.
15 This roughly coincides with the share of participants who donate in the 30-days study period, and slight
differences are due to a small number of participants who gave plasma twice.
13.15 Simple ask
+23.3% +26.3%
+19.6%
No ask +10.4%
.1
Donations
.05 0
No ask Facebook Peer Peer + Visibility Simple ask
N = 4463 N = 4244 N = 8511 N = 8497 N = 4227
Figure 2: Main Effects on Donations in WhatsApp Experiment
Notes: Average number of donations in the one-month study period and 95 percent confidence intervals.
timate also includes a vector of controls for individual characteristics that were balanced
through stratification, to account for the randomization method and improve precision
(Bruhn and McKenzie, 2009). Standard errors ε i are allowed to be correlated within ran-
domization units – the randomly formed twenty-donors groups.
This model is estimated in Table A.2, both using an intention-to-treat design (ITT) and
by conditioning on whether subjects were actually exposed to the treatment (LATE). In line
with the visual evidence, these estimates confirm the evidence that none of the social re-
wards motivate giving beyond what is achieved by a simple ask. Both ITT and LATE esti-
mates confirm that the Facebook treatment actually crowds out donations relative to a Simple
ask. The effects of other social reward treatments are slightly smaller but not statistically
different from the Simple ask.
Visibility holding constant peer comparisons. Our study disentangles the two main
mechanisms that are characteristic of social recognition: peer comparisons and visibility
incentives. We achieve this experimentally by comparing donations in Peer and Peer + Visi-
bility. In our preferred specification from column 3 of Table A.2, we estimate that visibility in
addition to peer comparisons leads donors to make 0.006 (p = 0.218) extra donations during
the study period. An effect that is positive but of no economic or statistical significance.
142.2. Do donors react to social comparisons?
We study two types of responses to social comparisons with peers: First, whether donors
exhibit status seeking preferences by reacting more strongly to social norm information in-
dicating that they are less charitable than their peers. Second, whether exposure to more
positive social norms systematically inspires more donations.
Table 2: Peer comparisons, effects on compliers
Panel A (1) (2) Panel B (3) (4)
Peer All Peer Peer
info info info
Below buddies’ average 0.024∗∗ 0.011 Buddies’ average -0.005
(0.009) (0.010) (0.008)
Peer info -0.003 Buddies’ average 1st quintile -0.013
(0.008) (0.008)
Below buddies’ average × Peer info 0.013 Buddies’ average 2nd quintile -0.005
(0.009) (0.008)
Buddies’ average 4th quintile -0.020∗∗
(0.009)
Buddies’ average 5th quintile -0.018∗∗
(0.009)
Age, gender, and past donations Yes Yes Age, gender, and past donations Yes Yes
Local branches FE Yes Yes Local branches FE Yes Yes
Observations 15497 23171 Observations 15497 15497
Clusters 911 1363 Clusters 911 911
R2 0.056 0.059 R2 0.056 0.056
F-test Buddies’ average quintiles 0.100
Notes: “Below buddies’ average” is a binary indicator taking the value of 1 if the donors has donated less
than the average of the group in the last 11 months, 0 otherwise. “Peer info” is a binary indicator taking the
value of 1 if the donor is in the Peer or Peer + Visibility treatment, 0 otherwise. “Buddies’ average” is the average
number of donations made only by the other members of the groups in the last 11 months. In the last row of
Panel B, “F-test Buddies’ average quintiles” is an F-test for the joint significance of the coefficients related to the
quintiles indicator variables. Standard errors in parentheses are clustered at the group level.
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Status seeking. Identifying status seeking preferences requires variation in the observed
behavior of peers, particularly in whether peers are (on average) more or less generous.
The main empirical challenge is that being more generous than the average to individual
giving behavior is endogenous. Going from column 1 to 2 of Table A.3, we can account
for endogeneity by estimating the effect of recent donations being below the group average
conditional on individual recent donation history. We leverage the fact that conditional on
individual recent history, being above or below average is almost as good as random by
random group matching. We estimate that donors who recently donated below the average
15of their group make 0.023 (p = 0.011) donations more in the study period.
This effect could be inflated by mechanical mean reversion if donors who donated less re-
cently were already planning to donate in the study period even without our intervention.
In fact, column 3 shows that also participants who do not receive peer information (No ask,
Simple ask and Facebook treatments) are more likely to donate if they randomly happen to be
below their group average (without knowing).16 To disentangle mechanical mean reversion
from status seeking, we use a model that interacts having donated less than the social norm
with being one of the donors that is informed of that. Column 4 of Table A.3 shows that 83
percent of the naive effect estimated in column 2 can be attributed to mechanical mean re-
version. This is an upper bound of mechanical mean reversion because the model implicitly
assumes that everyone who is assigned to receive peer information actually engages with
the contents of the intervention. We estimate local average treatment effects among partici-
pants who read the experimental messages that are sent to them. Table 2 column 2 provides
estimates for the share of the treatment effect that can be attributed to status seeking that
are less prone to downward bias: 54 percent of the naive effect is due to status comparisons.
Social norms. We can also study more systematically how random exposure to different
social norms can shape behavior (see distribution of social norms that donors in our sample
are exposed to in Figure A.4). In Table A.4 we estimate the effect of social norms on giving
using two complementary approaches. First, we use a model where social norms are as-
sumed to linearly affect behavior. Because the distribution of social norms is not uniform,
this specification has the disadvantage of being excessively sensitive to extreme values of
the social norms. Second, we use a model where social norms are discretized into quintiles.
We find that donors do not linearly respond to the social norms that they are exposed
to, both if we look at ITT and LATE effects (Table A.4, columns 1 and 3). The analysis with
discretized social norms reveals that behavior responds to norms non-monotonically, driven
by a strong response to social norms in the third quintile and a discouragement when norms
in the forth and fifth quintiles are observed.
16 Our experimental design forms groups of 20 donors irrespective of whether donors in the experiment
receive peer information.
162.3. Social proximity
Using experimental variation in group composition, whether group members are fellow
donors of the same donations center (Close) or they are scattered across the region (Distant),
we study how social proximity moderates status seeking, response to social norms, and
visibility concerns.
In column 4 of Table A.5, we cannot reject the interaction effect between the binary indica-
tor for social proximity Close and the social norm to be different from 0 (p = 0.603). Panels
b and c of Figure A.8 estimate the effect of being exposed to social norms from different
quintiles of the distribution without point at substantial heterogeneity.
We show evidence in support of the idea that the experimental variation in social proxim-
ity is meaningful to subjects. Panel a and b of Figure A.5 show that donors are more likely to
provide unsolicited responses to the treatment message if they are in Close treatments, and
conditional on responding they write more. At the end of the experiment, in a sub-sample
of 895 subjects from the Peer + Visibility treatment, we also measure beliefs about knowing
at least one group member (out of 20). Panel c of Figure A.5 shows that donors are almost
twice as likely to believe that they know someone from their group (p < 0.001). Taken
together, these results paint social proximity as a weak moderator of social influence.
2.4. Recognition on social media
In this section we report the results of three experiments. The later two experiments were
conducted simultaneously and should be seen as conceptual replications of the first, to-
wards understanding whether low engagement and low statistical power to identify av-
erage treatment effects conditional on engagement (LATE) explain the findings from the
original study. There are slight differences in implementations, as explained in Appendix B.
However, all three experiments offer a clean comparison between Simple ask and Facebook to
identify the effect of social media recognition on giving.
In the first experiment, we find that a Simple ask causes donations to increase. How this
effect compares to recognition on social media (Facebook) unfortunately depends strongly
on whether the specification includes local branch fixed effects. A feature of the first ex-
periment was that social recognition was cluster-randomized at the local branch level. If
17fellow donors interact during the study period, this feature has the important advantage of
strongly attenuating treatment contamination. The cost of this design with a small number
of clusters is worse quality of randomization, which does not allow us to reliably estimate
the model with local branch fixed effects. The informativeness of this experiment is also
plagued by poor deliver rates of the experimental communications that were sent: these
effects are estimated off of a relatively small number of 2577 compliers.
In the new experiments we take several steps to improve delivery of experimental com-
munications.17 The change that had the most meaningful impact was to use WhatsApp as
the communication channel through which experimental interventions were delivered. This
in itself improved delivery of our communications from 17 percent to 91 percent. Further-
more, as the lack of intra-class correlation in the first experiment suggested little scope for
treatment contamination (in Facebook for November 2019 donations, ICC=0.016), we were
able to implement a higher quality randomization for the new study, by clustering the ran-
domization at the twenty-donors group level.
Both in the second and in the third experiment we estimate a positive and significant ITT
effect of a Simple ask on donations. The effect is larger, though not significantly so, in the
WhatsApp experiment (t-test, p = 0.135), and in both experiments we find that making
donations more visible tends to backfire. While we have relatively low power to detect
meaningful effects of social recognition on compliers in the first two experiments, in the
third experiment we have 80 percent power to detect a 0.02 unit change. The local average
treatment effect that we estimate shows that social recognition substantially crowds out
donations (β = −0.015, p = 0.038).
Finally, we follow the approach proposed by Fowlie et al. (forth.) to establish whether the
crowding out effect is due to marginal compliers of the WhatsApp experiment responding
(a) more negatively to the social recognition treatment or (b) more positively to the Simple
Ask. We want to estimate what is the effect of knowing that donations are rewarded with
social recognition among participants who read our WhatsApp messages but would have
not read our emails.18 An obvious challenge in answering this question is that we do not
17 We extensively test HTML contents of treatment emails to avoid that our messages end up in spam, we
restrict the sample to ’active’ donors (that have donated at least once in the past 5 years), and we vary the
communication channel (email vs. WhatsApp).
18 Table A.8 shows differences in the composition of compliers across communication channels.
18Table 3: Facebook experiments
Panel A: Treatment effects of social recognition
ITT LATE
(1) (2) (3) (4) (5) (6)
2019 2021 2021 2019 2021 2021
Email Email WhatsApp Email Email WhatsApp
Simple ask 0.015∗ 0.014∗∗ 0.026∗∗∗
(0.007) (0.006) (0.007)
Facebook 0.017∗∗ 0.010 0.011∗ -0.005 0.017 -0.015∗∗
(0.007) (0.006) (0.007) (0.016) (0.019) (0.007)
Age, gender, and past donations Yes Yes Yes Yes Yes Yes
Local branches FE No Yes Yes No Yes Yes
Observations 14993 13512 12934 2577 1550 7674
Clusters 67 681 677 65 445 452
R2 0.059 0.058 0.065 0.058 0.082 0.072
Opening rate 22.62% 17.13% 90.59% 100% 100% 100%
MDE Facebook 0.020 0.018 0.018 0.045 0.052 0.020
Simple ask - Facebook -0.002 0.005 0.015 0.005 -0.017 0.015
p-value(Simple ask - Facebook) 0.829 0.492 0.028 0.754 0.353 0.045
Panel B: Effects of social recognition and single ask awareness, by channel and on marginal WhatsApp compliers
(7) (8) (9) (10) (11) (12)
2021 2021 Only 2021 2021 Only
Email WhatsApp Facebook Email WhatsAppSimple ask
Social recognition aware -0.027 -0.016∗∗ 0.001
(0.039) (0.007) (0.009)
Simple ask aware 0.026 0.016∗∗ 0.019∗∗
(0.038) (0.007) (0.009)
No ask -0.014∗∗ -0.026∗∗∗ -0.010 -0.011∗
(0.007) (0.007) (0.006) (0.006)
Age, gender, and past donations Yes Yes Yes Yes Yes Yes
Local branches FE Yes Yes Yes Yes Yes Yes
Observations 13512 12934 8809 13512 12934 8711
Clusters 681 677 457 681 677 451
R2 0.057 0.065 0.061 0.058 0.065 0.075
Notes: Panel A reports the impact of treatments on donations in the study period. “Simple ask” and “Face-
book” are binary indicators taking the value of 1 if the donor is in the particular treatment, 0 otherwise.
In columns related to the 2019 Email experiment, standard errors in parentheses are clustered at the local
branch level. In columns related to the 2021 experiments, standard errors in parentheses are clustered at
the group level. Panel B shows results from instrumental variable regressions. “Social recognition aware”
is a binary indicator taking the value of 1 if the donor is in the Facebook treatment and read the treatment
communication, 0 otherwise. “No ask” is a binary indicator taking the value of 1 if the donor is in the No
ask treatment, 0 otherwise. “Simple ask aware” is a binary indicator taking the value of 1 if the donor is in
the Simple ask treatment and read the treatment communication, 0 otherwise. In columns 7 and 10 (8 and
11) the estimation sample includes all treatments in the email (WhatsApp) experiment, including the No
ask. In columns 7 and 8 (10 and 11) awareness variable of interest is instrumented for whether the donor
is assigned to the Facebook (Simple ask) treatment. In columns 9 (12) the estimation sample includes both
Facebook (Simple ask) treatments, administered via email and WhatsApp; the awareness variable of interest
is instrumented for whether the donor is assigned to an email treatment.
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
19observe both counter-factuals to identify who are the marginal compliers on WhatsApp.
As Fowlie et al. (forth.) show, one solution is to exploit the feature of our intervention that
communication channel was randomized to estimate the effect of awareness about the treat-
ment by instrumenting awareness with treatment messages sent.19 The exclusion restriction
corresponds to assuming that receiving information about the social recognition interven-
tion through WhatsApp (instead of email) only affects donations by increasing awareness
about the social recognition campaign. Instrumenting awareness of social recognition in the
email treatments with the randomized communication channel identifies the effect of social
recognition on the marginal compliers.
Table 3 (9) shows that social recognition has no effect on marginal compliers. The crowding-
out effect is due to the response of marginal compliers to the simple ask (column 12).
3. Discussion
3.1. Privacy concerns, social signaling and inter-temporal substitution
Privacy concerns and sophisticated social signaling considerations can weaken the effective-
ness of social recognition. In buddy donors, where both the donation and lack of donation are
revealed, participants may choose to not donate to avoid appearing narcissistic – by being
associated with people that donate because of the visibility incentive. Bénabou and Tirole
(2006) formalize this intuition and show that when the degree of image concerns is hetero-
geneous in the population, high visibility can lead to an over-justification effect. This can be
particularly pronounced in a setting with low variance in the distribution of altruistic types
and high variance in the distribution of narcissism: good public actions end up signaling
narcissism. The Facebook intervention only recognizes people that do donate, potentially
also leading privacy concerned donors to avoid donating in the study period simply to
prevent their name to appear on social media. If donors do engage in either of these consid-
erations, we should expect social recognition interventions to have positive delayed effects.
We re-estimate our main regression model for donations in the first month after the study
period, the second month, and the two months combined with the study period (Table A.6).
19 Straightforwardly, we define awareness of social recognition as a binary indicator that takes value 1 for
participants that are (a) assigned to a Facebook treatment and (b) read the treatment message.
20We now observe that donations one month later are higher in Peer + Visibility than Simple
ask (β = 0.013, p = 0.032) and No ask (β = 0.010, p = 0.048). Interestingly, these delayed
effects also reveal that the Simple ask generates harvesting while social recognition does
not: the increase in March donations due to the Simple ask is largely offset by a decrease in
donations in the following two months. Heterogeneity analyses reveal that the Simple ask is
very effective, but only among the more active donors (Table A.6, column 4): Active donors
come early to the donation center if they receive a simple ask, but do so at the expense
of future donations, whereas social recognition may be more effective at increasing overall
donations because also less frequent donors respond to it.
3.2. Social recognition and good citizenship: A meta-analysis
To put our evidence into context, we conduct a selective meta-analysis of the field experi-
mental literature on social recognition and good citizenship. Studies included in this meta-
analysis are summarized in Table A.7, where we remark whether the design includes ad-
ditional control treatments–like in our study–to disentangle simple reminder or salience
effects from social recognition.
In Figure 3 we plot standardized effect sizes across studies, with respect to sample size,
and make two observations. First, in line with other meta-analyses (e.g. DellaVigna and
Linos, 2020) documenting a ’voltage effect’ (Al-Ubaydli et al., 2019), we observe a nega-
tive correlation between standardized effect size and sample size. In spite of that, larger
studies in this literature still find the effects of social recognition on good citizenship to be
positive–indicating that smaller studies may overstate actual effect size but are unlikely to
be false-positives. Second, ours is the only study to provide evidence that social recogni-
tion can have negative effects on targeted good citizenship outcomes.20 We view this as an
important and surprising result that should spur interest in understanding how to harness
social recognition and in mapping the environmental factors that can make social recogni-
tion welfare enhancing.
20 In an internal replication on the effect of social recognition on voting in presidential elections, DellaVigna
et al. (2016) also report negative effects. However, these are rather small and imprecisely estimated. Lambarraa
and Riener (2015) finds that charitable donations decrease with visibility. However, this result attains in a
cultural setting with a strong religious prescription of anonymous giving. We flip the sign of this effect to
capture that a ’good citizen’ in Islam is someone who gives money to charity in private.
21Alpizar et al. (2008). T: Non-anonymous voluntary contribution (N = 502). C: Anonymous
p ≤ 0.05 p > 0.05
voluntary contribution (N = 495).
Ashraf et al. (2014)
ρsample size = -0.185, p-value = 0.422
Andreoni et al. (2017). T: Impossibility to avoid ask of charitable giving (N = 8896). C: Possibility
to avoid the ask of charitable giving (N = 8766).
.4
ρ log(sample size) = -0.516, p-value = 0.017 Ashraf et al. (2014). T: Social recognition for sale of condoms (N = 185). C: No social recognition
for sale of condoms (N = 212).
Chen et al. (2020). T: Social recognition for Wikipedia contribution (N = 1, 329). C: No social
recognition for Wikipedia contribution (N = 1, 315).
Yoeli et al. (2013)
Chetty et al. (2014). T: Review times publicly posted (N = 347). Standard review request (N =
432).
DellaVigna et al. (2016) A. T: Informed about subsequent survey regarding turnout for 2010 elec-
.3
tions (N = 9, 039). C: Simple reminder of upcoming 2010 elections (N = 10, 805).
DellaVigna et al. (2016) B. T: Informed about subsequent survey regarding turnout for 2012 elec-
tions (N = 23, 436). C: Simple reminder of upcoming 2012 elections (N = 23, 501).
Gallus (2017). T: Public award for being a new Wikipedia contributor (N = 1, 617). C: No award
for being a new Wikipedia contributor (N = 2, 390).
Gerber et al. (2008). T: Peer information and image pressure to vote at the 2006 elections (N =
38, 201). C: Simple reminder about the 2006 elections (N = 191, 243).
Lambarraa and Rienner (2015)
.2
Standardized effect
Goette and Tripodi (2021) A. T: Blood donors contacted via WhatsApp are prospected social
recognition on Facebook (N = 4, 234). C: Blood donors contacted via WhatsApp receive a simple
Panagopoulos (2010) B ask to donate (N = 4, 219).
Karing (2018) Goette and Tripodi (2021) B. T: Blood donors contacted via WhatsApp are prospected social
Soetevent (2005) B Gallus (2016) recognition vis-a-vis other donors (N = 8, 476). C: Blood donors contacted via WhatsApp re-
Gerber et al. (2008) ceive a simple ask to donate (N = 4, 219).
Chetty et al. (2014)
Andreoni et al. (2017) Goette and Tripodi (2021) C. T: Blood donors contacted via email are prospected social recog-
nition on Facebook (N = 4, 544). C: Blood donors contacted via email receive a simple ask to
22
.1
donate (N = 4, 473).
Karing (2018). T: Bracelet informative of child’s vaccination status (N = 319). C: Bracelet unin-
Soetevent (2005) A
formative of child’s vaccination status (N = 318).
Panagopoulos (2010) A Lambarraa and Riener (2015). T: Non-anonymous charitable donations (N = 269). C: Anony-
Alpizar et al. (2008) mous charitable donations (N = 265).
DellaVigna et al. (2016) A Panagopoulos (2010) A. T: Social recognition on newspaper for voters in 2007 general elections
Perez-Truglia and Troiano (2018) (N = 2, 951). C: Simple reminder about 2007 general elections (N = 15, 090).
Rogers et al. (2016)
Panagopoulos (2010) B. T: Social recognition on newspaper for non-voters in 2007 general elec-
0
Goette and Tripodi DellaVigna
(2021) B et al. (2016) B
tions (N = 674). C: Simple reminder about 2007 general elections (N = 685).
Goette and Tripodi (2021) C
Chen et al. (2020) Perez-Truglia and Troiano (2018). T: Tax delinquents exposed to social pressure (N = 171, 79).
Tax delinquents not exposed to social pressure (N = 17, 155).
Goette and Tripodi (2021) A
Rogers et al. (2016). T: Informed about a possible subsequent survey regarding turnout for 2010
elections (N = 347, 054). C: Simple reminder of upcoming 2010 elections (N = 346, 929).
Soetevent (2005) A. T: Open basket to gather donations for the parish during service (N = 406).
C: Closed bag to gather donations for the parish during service (N = 428).
-.1
Soetevent (2005) B. T: Open basket to gather donations for an external cause during service (N =
400 3,000 22,000 160,000 1,200,000 380). C: Closed bag to gather donations for an external cause during service (N = 401).
Sample size Yoeli et al. (2013). T: Participation in an energy preservation program is observable. C: Participa-
tion in an energy preservation program is unobservable.
Figure 3: Selected meta-analysis of field experimental evidence for social recognition on good citizenship
Notes: Standardized effects are calculated as the ratio between the difference in means and the standard deviation of the control group. P-values are obtained from t-tests for the equality of
means. For Ashraf et al. (2014), where the standard deviation of the outcome in treatment and control is not reported, we use the standard deviation of the overall sample. In Yoeli et al. (2013), a
breakdown for the number of observations by treatment is not reported; we assume treatment groups to be equally sized. For Chen et al. (2020) and Perez-Truglia and Troiano (2018) values are
currently being extrapolated from charts or regression tables and are subject to revision.You can also read
