I. INTRODUCTION

Charitable giving, in the form of money, time, or in kind, constitutes a nonnegligible fraction of the economy in many countries. Nevertheless, there does appear to be scope to increase donations through the application of behavioral economics. Experimental methods have been used to test a number of techniques for increasing donations, including lotteries (Landry et al. 2006), tontines (Lange, List, and Price 2007), challenge gifts (Andreoni 2006), and matching gift campaigns (Eckel and Grossman 2003; Karlan and List 2012). The study reported here considers (1) the potential influence of default donation levels on charitable giving, and (2) whether informing potential donors about the presumed effect of defaults mitigates their impact.

By a default, we refer to an action which will be taken unless the decision maker overrides it by actively choosing another option. If individuals are presented with a default donation level that becomes binding unless they change it, it may serve as a reference point, or anchor (Tversky and Kahneman 1974), from which individuals insufficiently adjust. Moreover, if the cost of changing a choice outweighs the benefit of lowering a donation, one might choose to maintain giving at the default level. If these effects are substantial, it indicates that a charity can intentionally choose default donation levels to increase donations. Indeed, defaults have been shown to be powerful attractors in a number of other domains. (1)

A number of previous experimental studies have explored the effect of specifying defaults on cooperation or donations. The results are mixed. Some studies find that defaults affect cooperation. For example, Altmann and Falk (2009) report that a relatively high default, in a linear public good game, increases contributions relative to both a baseline and to a low, noncooperative default. (2) Ebeling (2013) finds a large effect of a default on the number of people signing up for green energy. However, in the previous study most closely related to ours, Grossman and Eckel (2015) find that assigning the initial endowment to a charity and allowing an individual to draw from it, generates the same final allocation as awarding the initial endowment to a potential donor and allowing her to donate from it.

Other studies find that a default can influence decisions under some conditions but not others. In a field experiment on the default effect and public good giving, Carlsson, Johansson-Stenman, and Pham Khanh (2011) find that people in a low default condition contribute substantially less to a public good, a new bridge, than under a baseline treatment. Their results from a high default treatment are inconclusive, however. Altmann and Falk (2009), in a field experiment, find that introducing multiple defaults for donations to cover different components of a charity's expenses fails to increase the total amount of charitable giving, though it can affect the amount assigned to different expenses. Messer et al. (2007) find that defaults do not affect contributions in a public good game when implemented alone, but do have an impact in conjunction with a voting process or communication. Haggag and Paci (2014) study default tips for cab drivers in New York, and find that defaults can increase tipping unless they are set too high, in which case they reduce tip levels. Caplin and Martin (2012) find that defaults affect decisions, but reduce the quality of decision making, because some individuals reduce the attention that they pay to the task and follow the default instead.

There is also related work that considers the effect of providing either suggested donations or information about other individuals' donations. These can be viewed as similar to defaults in the sense that they may serve to create a reference point that influences decisions. Fosgaard and Piovesan (2015) find that a suggested contribution level influences cooperation in a public good game. Cappelletti, Gueth, and Ploner (2011) also study the effect of advice in a social dilemma and find that it has an effect, but that it is short term in nature. Edwards and List (2014) find that when charitable donations for a specific donation level are requested, donated amounts are more likely to be at that level. Croson and Shang (2008) study the effect of providing social information, regarding other individuals' prior donations relative to one's own, on renewals of subscriptions to public radio, and find that subscribers adjust their decisions in the direction of the social information. Shang and Croson (2009) report mixed results on the effect of social information on donation levels, finding a larger effect on new subscribers, and on those making an intermediate level of donation.

However, even if a default is effective in increasing donations in the short run, potential donors may become more sophisticated with experience. It stands to reason that they might eventually become aware of default effects and thus may correct for them when they make their decisions. To examine the implications of such awareness, we include treatments in which we explain the default effect to participants.(3) One way to view this procedure is as a form of de-biasing. A large literature exists on de-biasing methods (see, e.g., Babcock, Loewen-stein, and Issacharoff 1997; Cason and Plott 2014). Some popular methods for de-biasing are to ask participants to generate alternative scenarios, or to make decisions in steps (see, e.g., Kivetz and Simonson 2000; Koriat, Lichtenstein, and Fischhoff 1980). Another method to help people make better choices is to provide more information about the decision situation. However, the evidence on the effectiveness of this is mixed (Elbel etal. 2009; Schwarz etal. 2007; Zikmund-Fisher et al. 2008). The presumed effect of explaining default effects would be to reduce their magnitude. It is possible, however, that the information about defaults may make them even stronger attractors, and that disclosure may well strengthen the default effect. The subjects might perceive the information as an implicit recommendation and perhaps be more likely to select the default option.

In our experiment, we monitor the emotional correlates of donation decisions. It is clear that there may be emotional underpinnings to the decision to donate. As Small and Verrochi (2009) show, emotionally charged images can have a profound influence on donations. They observe that subjects viewing sad pictures donate more on average to a cause. Emotions are also presumably one of the mechanisms generating the identifiable victim effect that is known to increase donations (Genevsky et al. 2013; Jenni and Loewenstein 1997). Furthermore, the act of donation or failure to donate may generate an emotional response on the part of the donor Rubaltelli and Agnoli (2012).

A large body of work has shown that emotions do affect decisions in related contexts. Positive mood can significantly increase people's willingness to help others (see, e.g., Isen and Levin 1972). In turn, giving can make people happier and these two effects can reinforce each other (Anik etal. 2010). A few authors have considered the relationship between emotional state and the power of defaults. Lemer, Small, and Loewenstein (2004) demonstrate that sadness and disgust can eliminate the endowment effect, which is otherwise a robust type of default effect (Kahneman, Knetsch, and Thaler 1991). Lin et al. (2006) obtain similar results. Martinez, Zeelenberg, and Rijsman (2011) find that while regret eliminates the endowment effect, disappointment reverses it.

To study the relationship between emotions and donation, we gather objective physiological measures of participants' emotions, registered with Facereading software, before and after they make their donation decisions. We test whether a positive overall emotional state predicts a larger donation. We also consider whether one or more of the six basic universal emotions (Ekman 2007) of happiness, sadness, fear, disgust, anger, or surprise, is predictive of donations. We test whether behaving prosocially correlates with people feeling more positively afterward.

In our experiment, each subject is given an endowment and is asked to split it between herself and a charity. (4) Our design is in part based on Carpenter, Connolly, and Myers (2008), who study a dictator game, in which the receiver is a charitable organization. Subjects are given the option to choose their preferred charity to contribute to. They can either pick one from a menu of prespecified options or indicate a different charity as a destination for a donation. As indicated earlier, the most closely related work to ours is by Grossman and Eckel (2015). In their study, there are 20 dollars to be divided between a participating subject and a charity. Subjects are put in one of two situations: (1) they are allocated the full endowment of 20 dollars and can donate any part of it to charity, or (2) the charity is allocated the full endowment of 20 dollars and the subject can take away any portion of the money for herself. They observe no difference in the final allocations under the two conditions.

In our study, we find no evidence that defaults affect donation levels. Similarly, informing potential donors of the hypothesized effect of defaults also has no effect. We observe that a more positive emotional state correlates with making greater donations. Higher donations lead to a more positive self-reported emotional state, but to a less positive state as registered by our physiological measure. While our experiment was not designed to detect the reason for such a discrepancy, two possible explanations for the pattern are (1) that there is reporting bias in the self-reports, perhaps because individuals feel pressured to indicate that donating makes them feel more positive, or (2) the physiological measure exclusively captures an immediate affective state, while the self-reports also integrate other sources of subjective wellbeing. We also find that women donate more than men.

In the next section, we describe our experiment. In Section III, we report our results, and in Section IV we present our conclusions.

II. EXPERIMENTAL DESIGN AND IMPLEMENTATION

A. Procedures Common to All Treatments

The experiment was conducted in the CentER laboratory at Tilburg University in the Netherlands, in groups of sizes of between 1 and 16, in the years 2014 and 2016. In total, 363 subjects (185 male) participated. They were recruited using the Tilburg University online recruitment system. The subjects were bachelor's and master's students majoring in economics, business, and law. In the majority of cases, the sessions of this experiment were run after other, longer experiments. (5) In those cases, participants were not informed about the second session until the first one was completed. A number of sessions (encompassing 117 subjects) were run with subjects who had not participated in any experiment beforehand. When possible, all sessions were run in a separate room from the first, in order to emphasize the independence of the two experiments, and to thereby enhance isolation effects. All sessions were run by the same (female) experimenter in order to eliminate possible experimenter effects which, especially in the context of charitable giving, can affect behavior (Landry et al. 2006). We employed the z-Tree software (Fischbacher 2007). Subjects earned 3.20 [euro] (1 [euro] equals approximately US $1.20) on average, not including the participation fee, in this experiment.

At the beginning of the session, the experimenter turned on video cameras on each computer with the consent of the subjects. Subsequently, all subjects received written instructions, which were read aloud by the experimenter. While subjects could ask the experimenter about something that was unclear, no between-subject communication was allowed throughout the experiment. Subjects were paid by bank transfer after the session.

In all conditions, participants were presented with the screen illustrated in Figure 1, and asked to split 5 [euro] (in 25-cent increments) between themselves and a charitable organization. They could either contribute to KNGF Geleidehonden, a Dutch charity which trains dogs for visually impaired people, chosen by the experimenter, or could indicate any other charity of their choice to contribute to. All donations were sent to the respective charities after all experimental sessions were finished. Participants could request a receipt proving that the money was indeed sent to charity, though no subject did so.

Our experimental design allows, but does not require, individuals to specify a charity of their choice as the destination for their donation. We consider whether it correlates with donation levels, though we cannot isolate the effect of the choice of charity on donations from that of self-selection. That is, people who choose their preferred charity may be better informed about charities and be more frequent or larger donors. (6)

B. Treatments and Treatment-Specific Procedures

There were seven conditions (see Table 1). There were three default levels--low, medium, and high. A low default corresponded to a default donation equal to 0, medium default to a donation of 2.50, and high default to a donation of 5. A default level of donation was implemented by preselecting a button in the display on subjects' computer screens, as in Figure 1. Subjects were free to change the selection, but if they took no action to change it, the default level of donation was implemented. Subjects were required to click on a button labeled Confirm to submit their decision. Each default level was studied under two information levels, informed (info), and not informed, about the default effect. In the informed (info) conditions, participants received the following information:

   Please note that one of the options may be randomly
   preselected for you. A number of published
   studies have shown that many people will be influenced
   by this selection: i.e., they will donate more
   if the preselected choice offers a high donation, and
   they will donate less if the preselected choice offers a
   low donation.

A baseline treatment, in which there was no default and no information about defaults, was also included. We also ran a number of sessions with an additional survey asking subjects to self-report their emotions immediately after the donation decision. In all other respects, these sessions were identical to the corresponding treatments. Subjects were not informed that they would be asked about their emotions until after they made the donation decision. On a scale from 0 to 10, subjects had to state how happy, sad, angry, scared, and disgusted they felt. In total, 146 subjects in 15 sessions participated in these treatments.

Table 1 lists the treatments, the number of subjects who participated in each treatment, the number of sessions conducted, and the number of subjects for whom we have self-reported data on emotional state after their decision. Low default corresponded to no donation, medium to a donation of 2.50 [euro], and high to a donation of 5 [euro].

Our design has sufficient power to detect substantial treatment effects. The results of power calculations are given in Table A4 in Appendix B. We assume an [alpha] = .05. Then the power to detect a difference of one standard deviation (Cohen's d = 1) in average donation between the Low and High treatments, is 97%. The probability of detecting an effect size of d = 1 between Low-info and High-info is greater than 99%. The probability that we detect an effect size of d = .5, one-half of a standard deviation, is .47 between Low and High, .77 between Low-info and High-info, and .91 between the pooled data from Low and Low-info and the pooled data from High and High-info.

C. Facereader Software

All participants were videotaped, with their consent, during the entire session, using the webcam on their computer screen. The videotapes were analyzed later using Noldus Facereader 5. Facereader classifies a facial expression by how much it reflects the six basic emotions of happiness, sadness, anger, fear, disgust, and surprise. It also measures conformity to a neutral expression. (7) The facial expressions that correspond to the six basic emotions appear to be universal and innate, in that they are common across all cultures and different primates (Ekman 2007; Ekman and Friesen 1986), as well as between blind and sighted humans (Ekman 2007; Matsumoto and Willingham 2009).

Facereader has been shown to reliably identify the emotion an individual intends to show (Bijlstra and Dotsch 2011). It also correlates highly with self-reported emotions (Den Uyl and Van Kuilenburg 2005). Furthermore, it classifies human expressions as well as do human observers (Kudema-Iulian, Marcel, and Valeriu 2009; Lewinski, Fransen, and Tan 2014; Terzis, Moridis, and Economides 2010). Most relevant for this study, it correctly identifies happiness as the emotion an actor intends to express, with an accuracy rate of 97% (Den Uyl and Van Kuilenburg 2005). Facereader has been used in experimental economics to study risk aversion (Nguyen and Noussair 2014), asset markets (Breaban and Noussair 2017), and rejection decisions in ultimatum games (Van Leeuwen et al. forthcoming).

Facereading has a number of attractive features as a tool to measure emotions. It provides objective physiological correlates of subjects' emotional states (in contrast to self-reports where subjects may have incentives to misreport their true feelings). It yields a quantitative measure for the intensity of emotions. It operates unobtrusively, meaning that the data collection would likely not be noticed by the subjects had they not been informed about it. Finally, it operates in nearly real time, with current emotional states registered 30 times per second.

Facereader reports values for the six emotions and neutrality on a scale between 0 and 1. From these, we derive a measure of valence at time t that equals:

(1) [valence.sub.t] = [happiness.sub.t] - 0.25 x ([anger.sub.t] + [sadness.sub.t] + [fear.sub.t] + [disgust.sub.t]).

This is a measure of the net positivity of an emotional state, in that it equals the difference between the only positive basic emotion, happiness, and the average of the four negative emotions: anger, sadness, fear, and disgust. Valence ranges from -1 to +1. This measure differs somewhat from the standard measure that Facereader computes, which is [valence.sub.t] = [happiness.sub.t] - max{[anger.sub.t], [sadness.sub.t], [fear.sub.t], [disgust.sub.t]}. Surprise and neutrality are not considered as positive or negative emotions, and therefore do not enter into the valence calculation. (8)

In our experiment, we measure participants' emotional state both before and after their donation decision. To register subjects' emotions before the decision, we measure emotional state from the beginning of the reading of the experimental instructions until they make their decisions. The recorded value of each emotion is then averaged over this time period. To compute emotions after the decision, we average the recorded value of each emotion over the 5 seconds immediately following the donation choice.

The timing of the sessions, and in particular of the emotion measurement, is described in Figure 2. At the beginning of the session, subjects are seated with the video cameras already turned on. The instructions are then read by the experimenter, and all video from this point onward is included in the Emotions Before measure of emotional state. (9) Subjects then submit their donation decision. The Facereader data for the 5 seconds immediately after the donation decision constituted the Emotions After data. Afterward, in some sessions, the survey regarding subjects' emotional state was administered. For some more detail on the operation of Facereader, see Appendix C.

III. HYPOTHESES

The three hypotheses that we evaluate originate in previous research. The first hypothesis is that the default pull on donations would appear in our setting. As we described earlier, the results from prior work are mixed regarding whether defaults influence decisions. Nevertheless, because a majority of the previous studies conducted in related settings have reported an effect of defaults under at least some conditions, we maintain, a priori, the hypothesis that default donation levels shift decisions closer to the default level.

HYPOTHESIS 1: People facing a lower (higher) default contribute less (more) to charity.

As indicated earlier, explaining to an individual that a default may affect her decisions could alter her behavior. It may cause her to try compensate for the effect. That is, she may adjust her donation consciously to attempt to offset the impact of the default. If a substantial fraction of individuals respond in this manner, the effect of defaults on donations will be reduced. This is in line with previous research (Zikmund-Fisher et al. 2008). However, it is also possible that the information will render the default level an even stronger anchor or reference point. Thus, our hypothesis about the difference between conditions with and without information is that there is no effect and thus the tests we conduct are two-sided.

HYPOTHESIS 2: Informing participants about the default effect does not affect donations.

Our third hypothesis has two parts and concerns the relationship between emotional state and donation level. The first part is that prior emotional state correlates with subsequent donations, with more positive emotions accompanying higher levels of giving. A number of previous studies have reported that positive emotional state, as indicated in self-reports, correlates positively with subsequent charitable giving (Anik et al. 2010; Isen and Levin 1972). This leads us to hypothesize that the same relationship would exist when emotional state is measured with Facereading software. The second pattern is a correlation between the amount donated and subsequent emotional state, with greater donations leading to more positive emotions. Anik et al. (2010) find that individuals report greater satisfaction after making a donation. However, DellaVigna, List, and Malmendier (2012) argue that if giving takes place out of social pressure, which may lye present in our experiment, it can lower the utility of the donor. Because of these potentially offsetting effects, we hypothesize that making a donation has no overall relationship with subsequent emotional state.

HYPOTHESIS 3a: Individuals with more positive emotional valence, and greater happiness, donate more.

HYPOTHESIS 3b: Emotional valence and happiness after the donation decision is uncorrelated with the amount donated.

IV. RESULTS

This section is organized as follows. We first briefly describe the overall patterns in the data, and then evaluate our two hypotheses about the relationship between defaults, information, and donation levels. We then turn our attention to the relationship between emotions and giving. Finally, in Section IV.B, we report a number of observations from an exploratory analysis of the data.

A. Description of the Data and Tests of Hypotheses

The average donation in each treatment is shown in Figure 3, and histograms of donations in the pooled data from all treatments, as well as from each treatment separately, are given in Figures 4 and 5. At least six prominent patterns appear in the overall donation data. First, out of a total of 363 subjects, approximately a fifth (78) donated nothing; the share of people donating zero ranged between 19% (Medium, High, Medium info, and High info) and 25% (Baseline) in different treatments. Second, about a tenth (41) of the subjects donated everything; their share per treatment ranged between 6% (Low info) and 16% (Medium). Third, 13% (48) of subjects offered a 50:50 split; the lowest share of these subjects was in Low (6%), and the highest in the High condition (19%). Fourth, 15% (56) of our subjects donated 1 [euro], corresponding to 20% of their endowment; 20% (74) of subjects donated between 1 and 1.5, a typical average range of giving in dictator games (Engel 2011). Their share ranged between 10% (Medium) and 26% (Low) of subjects. Fifth, the treatment averages lie within a fairly small range, from 1.6 in Low to 2.0 in High. Sixth, the distributions in each of the seven treatments give the impression of being quite similar.

Our first hypothesis is that the default level of donation would affect the decisions of participants. In contrast, however, the first result that we report is that the presence or the level of a default donation has no significant effect on giving.

RESULT 1: Defaults do not affect donations.

Support for Result 1: Figure 3 shows the average donation by treatment. The impression conveyed by the figure is that differences among treatments are relatively small while the standard deviations within each treatment are large. Table 2 reports the results of Wilcoxon rank-sum tests of treatment differences. None of the treatments is significantly different from any other at any conventional significance level. Additionally, the Kruskal-Wallis test also cannot reject the null hypothesis of there being no difference between treatments (p value of .918 with ties). Furthermore, Kolmogorov-Smimov tests of the equality of distributions fail to reject the hypothesis that the distributions of donations are equal in any pairwise comparison of two treatments at the 5% level. This confirms the visual impression in Figures 4 and 5, which contain detailed histograms of donations by treatment, and for the pooled data from all treatments.

The lack of a default effect reported in Result 1 suggests, but does not necessarily imply, that informing individuals that defaults may exert an effect on their decisions would have no impact. Indeed, as reported in Result 2, we find that the information does not affect donation levels.

RESULT 2: Informing individuals about the anticipated effect of defaults has no effect on donations.

Support for Result 2: We make pairwise comparisons between the Low and Low-Info, Medium and Medium-Info, and High and High-Info treatments, using Wilcoxon rank-sum tests. The results are reported in the bottom portion of Table 2. None of the differences are significant at conventional levels.

To check that the lack of treatment effects is not a consequence of differing session size or gender composition among treatments, we run ordinary least squares (OLS) regressions with contribution as the dependent variable, and treatment dummies as independent variables. The estimates are reported in Table 3. The table shows that, regardless of the control variables used, the treatment dummies are never significant.

Results 1 and 2 indicate an absence of support for Hypothesis 1, but do support Hypothesis 2. However, though defaults do not affect decisions, other factors may well do so. Results 3 and 4 consider the emotional correlates of donations. Result 3 reports that positive valence, as well as the specific emotion of happiness, as registered with Facereading software, are correlated with greater subsequent donations.

RESULT 3: Emotional valence, as well as happiness, is marginally significantly correlated with subsequent donations for the overall sample.

Support for Result 3: Figure 6 illustrates the donation level for each quartile of happiness, sadness, and overall valence. Individuals in the least happy quartile donate an average of 1.6 [euro] while those in the happiest quartile donate 2.1 [euro]. Valence exhibits a similar pattern, with those in the highest quartile donating roughly 24% more than the lowest. Sadness does not exhibit a consistent relationship with donation level. The same is the case for fear, anger, and disgust, which are not illustrated.

The Spearman rank correlation coefficients, between both happiness and valence prior to the decision and donation level are positive, at [rho] = .101 and .104, respectively. Those correlations are marginally significant, yielding p values of .060 and .052. Additional evidence for these relationships can be seen in Table 3. In Equations (5) and (6), the coefficient of the variable Happy Before is significant in both regressions, albeit only at the p <. 1 level in one of the estimated equations. If Happy Before is replaced by Valence Before, in otherwise identical specifications for Equations (5) and (6), Valence Before is significant at p < .05 in both regressions.

There are two special supgroups of particular interest for whom the relationship between emotional state and giving is stronger than for the full sample. The first is those who donated very high amounts, at least three-fourth of the 5 [euro] endowment. They are significantly happier and have more positive valence than the rest of the sample (p = .012 and .025). This is shown in Table 4. The second consists of those individuals who chose their charity (5.5% of all subjects). They exhibit a positive correlation of 0.539 (p = .017) between prior happiness and donation level. While relatively small in number, and the decision to specify a charity is endogenous, this is an important subgroup, because they are presumably more likely than other participants to be active donors outside the experiment. For this subgroup, prior sadness is also negatively correlated with donation ([rho] = -.406, p = .084), and for them, valence exhibits a large and significant correlation as well ([rho] = .581, p = .009).

We now consider whether donations affect the emotional state of the donor, and the pattern we observe is described in Result 4.

RESULT 4: Higher donations are positively correlated with subsequent self-reported emotional valence, as well as with happiness. However, higher donations are also correlated with subsequent decreases in our physiological measure of valence and of happiness, so that after the giving decision, donors and nondonors are in a similar emotional state.

Support for Result 4: Figures 7 and 8 show the relationship between donations and subsequent self-reported emotional state. They contain histograms of donation levels for the lowest, second-lowest, second-highest, and highest quartile of individuals in terms of how strongly they report happiness and sadness, as well as in terms of the positivity of their emotional valence, which is derived from the emotions they report. The figures show that happiness and positive valence are associated with higher donations. Those who donated 0 are considerably more likely to report low happiness. Similarly, those who made the maximum possible donation of 5 [euro] are more likely to describe themselves to be in a happy emotional state than other participants.

Donations and happiness have a correlation of 0.214 (p = .01). Donation correlates with valence at 0.288 (p < .000), and with sadness at -0.188 (p = .023). When comparing the individuals in the lowest and highest quartiles of self-reported emotions, the average donation is significantly different between the groups for happiness (p = .047) and valence (p = .003), though not significantly different for sadness (p = .135).

The Facereader data are illustrated in Figure 9. The figure suggests an absence of a relationship between the amount of donation and emotional state. Table 4 shows the magnitude of valence, as well as happiness and sadness, both before and after the decision to donate, for different ranges of donations. Columns 1 and 2, which reflect data from the most observations, show that valence and happiness decline after the decision.

A Wilcoxon signed-ranks test is used to consider whether valence differs before and after making the donation decision. The results are reported in Table 5. Valence is lower after than prior to the decision (p <.001). However, this result only holds for the pooled data from donors and nondonors, and for instances in which the donation is positive. When the subject's donation equals zero, the values for the valences before and after the donation decision are statistically indistinguishable (p = .257). The tests in the last three columns suggest that higher donors experience particularly large decreases in valence after their decision, bringing them to similar levels of valence and happiness as others who donated less. While this last relationship is based on relatively few observations and the analysis is exploratory in nature, it does suggest that at least some donors experience less positive emotions after making a decision to donate.

B. Other Correlates of Giving

Although the treatments do not correlate with decisions, we do find that a number of other factors do exhibit significant relationships with donation decisions. The first of these is that there is a small but significant correlation between the number of subjects in a session and average donations.

OBSERVATION 1: The number of subjects in a session correlates negatively with average donation.

Support for Observation 1: The correlation between donation and session size, equal to -0.164 (p=.002), is negative and significant. As illustrated in Figure 10, this effect is driven particularly by the smallest session sizes (four or fewer subjects in a session), in which donations are higher.

There are at least two plausible explanations for this effect. First, it may be the case that subjects perceive greater anonymity in a larger group and interpret this as a license to behave selfishly. Second, is that if the subjects think that the experimenter is aiming to raise a fixed amount of money, such as 20 [euro], then subjects in larger groups feel less responsible individually for reaching this target, reasoning that there are sufficiently many other people who can donate instead.

However, as can be seen in Table 3, this correlation does not account for our Results 1 and 2 that show a lack of differences in donations between treatments. Even when controlling for session size, which has a significantly negative coefficient, the results that defaults or information about them have no effect on donations are unaltered.

Another strong correlate of donations is gender. We find that females donate significantly more than males. This pattern is summarized as Observation 2.

OBSERVATION 2: Male subjects donate significantly less than female subjects.

Support for Observation 2: The average donation of males is 1.5 [euro], while the average for females is 2.1 [euro]. A Wilcoxon rank-sum test yields a p value of <.001, indicating that the difference is significant. (10)

These results contrast, for example, with those of Bolton and Katok (1995), who find no evidence for gender differences in generosity in the dictator game, while corroborating the results of List and Price (2009) and DellaVigna et al. (2013), and the overall conclusion of Engel (2011), that women give more than men in dictator games.

In addition, we observe some evidence that the genders exhibit some differences in how their donations change in response to receiving information about the presumed effect of defaults.

OBSERVATION 3: There is some evidence of gender differences in the response to information about defaults. When women in the low default condition learn that a low default is thought to lower donations, they respond by increasing their donations. When men in the high default condition learn that a high default is thought to increase donations, they respond by lowering their donation.

Support for Observation 3: All p values for the comparisons reported in Table 2, when conducted for the subsamples consisting of each gender separately, remain above .05 with two notable exceptions. First, when facing a low default and being informed about it, women increase their donations relative to when they have no information (1.7 [euro] in the Low treatment compared to 2.3 in Lowinfo, p value of .034). Second, when facing a high default and being informed about it, men decrease their donations from 1.7 [euro] in High to 1.2 in High-info (p value of .042).

The next observation concerns the subset of the sample (5.5%), which chooses their own charity. We find that they donate more than others. The relationship is correlational rather than causal, because the decision to name a recipient's charity is endogenous. The presence of such a correlation is quite reasonable, as these individuals are presumably more likely to be active donors outside of the experiment than the rest of the sample.

OBSERVATION 4: People who choose their charity to contribute to donate more than those who do not.

Support for Observation 4: While the average donation of those who do not choose their charity is 1.8, those who specify a destination for their donation give an average of 2.5. A Wilcoxon rank-sum test rejects the hypothesis that the two levels are equal (p = .019). (11)

V. CONCLUSION

In our study, we do not detect any effect of a default on donation decisions. Those who were given relatively low, medium, and high default donation levels behave in a similar manner. We also observe that informing individuals about defaults has no impact on their decisions. This does not suggest that de-biasing techniques are not effective; there was no behavioral bias to remove.

We have also studied the relationship between emotions and giving. We corroborate, using Facereading software to measure emotional state, the finding that greater happiness, prior to the donation decision, is correlated with higher donations. A similar relationship holds for overall valence. Those in a more positive overall emotional state prior to the donation decision tend to donate more.

Self-reports indicate that those who donate are in a more positive emotional state afterward. However, our facereading data show those who donate experience a subsequent drop in emotional valence, which returns them to the average level of the whole sample. The disagreement between the self-reports and Facereader may be due to a number of factors. One interpretation of this pattern is that Facereader data are unreliable measures of emotional state. However, this claim is belied by a number of studies that have documented intuitive relationships between Facereader data and economic decisions (Breaban and Noussair 2017; Nguyen and Noussair 2014; Van Leeuwen et al. forthcoming), as well as the studies cited in Section II.C that have validated Facereader as a tool to measure emotional state.

Another possibility is that the self-report data include responses from individuals who feel compelled to make donations against their wishes. Della Vigna, List, and Malmendier (2012) note that some individuals try to avoid donating when they can, suggesting that confronting a decision about giving is an unpleasant experience. Indeed, we do observe higher donations in sessions where few subjects are present, perhaps because individuals felt more pressure to give in small sessions.

Alternatively, the discrepancy may reflect the fact that facereading captures an immediate snapshot of current emotional state that puts much weight on the money that one has just spent or the social pressure present at the moment, while the self-reports reflect a more comprehensive sense of well-being. This may reflect future expectation of positive emotions and other dimensions of satisfaction resulting from the donation that are not necessarily manifested in immediate positivity of emotional state.

Prior results are mixed with regard to the effectiveness of defaults. While defaults seem to have an effect in some settings, these appear to have a tendency to be situations in which the context is relatively complex and unfamiliar to participants (Choi et al. 2004; Schweitzer 1995). Our reading of the literature and our own work leads us to believe that these two factors, (1) complexity and (2) unfamiliarity, are conducive to defaults having an effect. Our task is very simple and rather familiar to many participants, and this may make it more likely that people would form decisions with confidence and override the defaults.

We make no claims that providing information about the role that defaults are thought to have would never have an effect on decisions in other contexts. A research strategy for future work to demonstrate this might be to use a protocol of a previous study that is known to produce large default effects, and to add a treatment in which subjects are informed of the effect that a default is known to have. One would then measure the extent to which the default effect is mitigated by this information.

APPENDIX A: EXPERIMENTAL INSTRUCTIONS

This section contains the instructions used at the experiment. The paragraph written in italics was added for the three informed conditions.

This is an experiment in the economics of decisionmaking. You will be asked to split 5[euro] (in 25 cent increments) between yourself and a charitable organization. Your earnings depend only on the decision you yourself make and will be kept private from other participants.

The charitable organization that will receive any money you and other students decide to donate is KNGF Geleidehonden (The Royal Dutch Guide Dog Foundation), which trains guide dogs for visually-impaired people. If you prefer, we can donate it to another charitable organization. Please indicate which one here:...

You will now be presented with a screen similar to the following one. You can only select one option.

The amount I wish to donate to the charity is...

* 5.00
* 4.50
* 4.00
* 3.50
* 3.00
* 2.50
* 2 00
* 1.50
* 1.00
* 0.50
*0 00

Please note that one of the options may be randomly
preselected for you. A number of published studies have
shown that many people will be influenced by this selection:
i.e., they will donate more if the preselected choice offers a
high donation, and they will donate less if the preselected
choice offers a low donation.

You will be paid within two working days via a
bank transfer.

If you have any questions, please raise your hand and
ask now.

Note: You will be video-taped throughout the experiment.

APPENDIX B: ADDITIONAL TABLES

This appendix contains two types of supplementary tables. Tables A1, A2, and A3 provide more information on how subjects were randomized into sessions and treatments, and the various experiments that they participated in prior to ours. Table A4 provides power calculations for three different effect sizes.

The studies referred to in Tables A1 and A2 are the following:

Y11 = Noussair, C. N., and Y. L. Xu. "Information Mirages and Financial Contagion in an Asset Market Experiment." Journal of Economic Studies, 42(6), 2015, 1029-55.

Y12 = Noussair, C. N., S. Tucker, and Y. L. Xu. "Futures Markets, Cognitive Ability and Mispricing in Experimental Asset Markets." Journal of Economic Behavior and Organization, 130,2016, 166-79.

Ays = Terzi, A., C. Koedijk, C. N. Noussair, and R. Pownall. "Reference Point Heterogeneity." Frontiers in Psychology, 7(September), 2016,1-10; article 1347.

Mrkt1 = Unpublished market experiment.

Mrkt2 = Unpublished market experiment.

Eli = Unpublished, experiment about environmental economics.

Gon1 = Unpublished, macroeconomic experiment.

Gon2 = Unpublished, macroeconomic experiment.

Car = Unpublished, art appreciation experiment.

Gyu = Unpublished, auction experiment.

Vik = Unpublished, experiment about goal setting.

VanK = Unpublished, experiment about individual decision-making under uncertainty.

Sig = Unpublished, experiment about social preferences.

APPENDIX C: THE FACEREADER SOFTWARE

In this appendix, we provide some more information on using Noldus Facereader[TM] to measure emotional state. Facereader can be used to analyze images, previously recorded video, or videos taken in real time. It can analyze video at a rate or up to 30 frames a second. For more information, please see http://www.noldus.com/human-behaviorresearch/products/facereader.

TIME COURSE OF EMOTIONS

We use a 5-second window after an event to capture the emotional reaction to the stimulus in our experiment, the donation decision. According to Ekman (2007), most emotional expressions are approximately 2 seconds long; with very rarely the expressions being shorter than 0.5 a second, or longer than 4 seconds.

Ekman (2007) also notes that short but very intense expressions can be a sign that a subject is trying to conceal the emotion. In contrast, low intensity expressions that last for a long period of time can be a sign that a subject is deliberately controlling the emotion. Note that typically in Facereader experiments, such as ours, subjects are aware that they are videotaped but not that their emotions are being analyzed, and therefore should have few reasons, if any, to conceal emotions.

HOW FACEREADER OPERATES

Noldus Facereader[TM] uses the Active Template Method (ATM) to locate the position of a face in an image. Should this method fail to locate a face, a second algorithm (Viola Jones cascaded classifier algorithm) takes over. Then, the Active Appearance Model (AAM) locates 530 key points on the face (see Figure A1), using a database of several thousand annotated images. It then, using a proprietary algorithm, correlates the degree to which the distances between these points conforms to canonical profiles for the six basic universal emotions. It then derives a measure of overall valence of emotional state based on the values of the individual emotions. For video, it can calculate these emotional states at a rate of up to 30 times per second.

It is possible to calibrate Facereader for each individual subject, by taking video of her at the outset of the experiment and measuring deviations from this initial profile. Alternatively, it is also possible to calibrate based on a library of several thousand images that accompanies the software.

RUNNING EXPERIMENTS WITH FACEREADER

Our experience has been that Facereader functions more effectively when a camera with HD resolution is used. The camera is best positioned on top of the computer screen that the subject will be using, and adjusted such that it gets an unobstructed full frontal view of the subject's face (though Facereader is able to follow the face to some extent). Note that new versions of Facereader (versions 5 onward) are able to read facial expressions for subjects with glasses, facial hair, or scars/tattoos.

The experimenter should not stand behind or next to subjects during the experiment, as the camera might capture his/her face as well, and the emotional reading might shift from the subject to the experimenter. This problem could be exacerbated if the laboratory has poor quality lighting (which includes the light being positioned behind the subject, thereby putting him/her in the shadow), or if the subject does not look at the camera. For this reason, it is important not to analyze the first few seconds after turning the camera on and the last few seconds before the camera is turned off if the experimenter is in front of it during those moments.

Recently, a new software program, [mu] Cap was developed. It can merge z-tree and video footage, which makes analyzing the data much easier, because it coordinates the timing of the ztree data and the video. This software can be downloaded for academic purposes free of charge at http://mucap.davidschindler.de/. See the working paper of Doyle and Schindler (2015) for more details.

Caption: FIGURE A1 AAM Facereader Analysis

TABLE A1

Cross-Tabulation: Subjects, Sessions, and Prior Experiments I

Condition       None     Yll   Y12      Mrktl

Baseline          1      5,7         2(2x), l(8x)
Low              2,1      8     3         1
Medium          l(3x)                  2,l(4x)
High             2,3      8     8         2
Low info                  9            3,l(3x)
Medium info      2,1
High info      5,l(2x)               2(2x),l(2x)

Condition        Mrkt2        Eli      Car    Gonl

Baseline                                       6
Low                                     6
Medium             1           7              6 *
High                                           8
Low info        2,l(5x)                        9
Medium info                                  5,9,10
High info        l(5x)                        6,7

Condition      Gon2   Ays    Gyu    Vik   VanK     Sig

Baseline                    6(3x)    9    10 **   1,1,8
Low             2
Medium          3      2
High
Low info
Medium info            4
High info

Notes: The table lists, for each treatment, the size of the
sessions conducted, and which experiment preceded them, if any. In
case multiple sessions of the same size in the same cell were run,
this is indicated in the parentheses. * and ** are cases where one
subject's data was excluded from analysis, because it was
discovered after the session that he/she had participated in the
experiment previously. For these two sessions, although the session
sizes were 6 and 10, we only have 5 and 9 data points,
respectively.

TABLE A2

Cross-Tabulation: Subjects, Sessions, and Prior Experiments
II

Condition      None   Ays   Gon1
Baseline
Low             B            B
Medium          A      A
High
Low info
Medium info
High info

Notes: The table shows the two sessions where some subjects
participated in one experiment and others participated in
a different experiment, prior to ours. In session A, three
subjects previously participated in the Ays experiment, and one
in no experiment. In session B, seven subjects participated in
the Gon 1 experiment, while one subject in no experiment.

TABLE A3

Cross-Tabulation: Subjects in Mixed Treatment Sessions, No
Prior Experiment

               Ses-     Ses-     Ses-     Ses-
Condition     sion 1   sion 2   sion 3   sion 4

Low info        6        4        4        4
Medium info     4        4        4        6
High info       4        4        8        4

               Ses-     Ses-     Ses-
Condition     sion 5   sion 6   sion 7

Low info        2        6        5
Medium info     5        4        4
High info       4        3        4

Notes: The table lists how many subjects participated in each
of the sessions, conducted in 2016, in which different subjects
were randomized to different treatments within the sessions.

TABLE A4

Power Calculations

Assumed     Low     Low-info    Low and Low-info
Effect     versus    versus     versus High and
Size        High    High-info      High-info

1           0.97        1              1
0.5         0.47      0.77            0.91

Notes: The table gives the calculated power for each of our
three key comparisons. We assume a two-tailed comparison,
parent distribution normal, and alpha equal to .05.

ABBREVIATIONS

AAM: Active Appearance Model

ATM: Active Template Method

OLS: Ordinary Least Squares

LENKA FIALA and CHARLES N. NOUSSAIR

Fiala: Ph.D. Candidate, Faculty of Economics and Business, Tilburg University, Warandelaan 5000LE, Netherlands. Phone +31 13 466 2416, Fax +31 13 466 3042, E-mail [email protected]

Noussair: Professor, Department of Economics, University of Arizona, Tucson, AZ 85721. Phone 1-520-621-6629, Fax 1 520621 8450, E-mail [email protected]

10.1111/ecin.12459

REFERENCES

Altmann, S., and A. Falk. "The Impact of Cooperation Defaults on Voluntary Contributions to Public Goods." University of Bonn Working Paper, 2009.

Andreoni, J. "Leadership Giving in Charitable Fund-raising." Journal of Public Economic Theory, 8(1), 2006, 1-22.

Anik, L., L. B. Aknin, M. I. Norton, and E. W. Dunn. "Feeling Good about Giving: The Benefits (and Costs) of Self-Interested Charitable Behavior," in The Science of Giving: Experimental Approaches to the Study of Charity, edited by D. M. Oppenheimer and C. Y. Olivola. London: Psychology Press, 2010.

Babcock, L., G. Loewenstein, and S. Issacharoff. "Creating Convergence: Debiasing Biased Litigants." Law & Social Inquiry, 22(4), 1997, 913-25.

Basil, D. Z., N. M. Ridgway, and M. D. Basil. "Guilt and Giving: A Process Model of Empathy and Efficacy." Psychology & Marketing, 25(1), 2008. 1-23.

Bennett, R. "Factors Underlying the Inclination to Donate to Particular Types of Charity." International Journal of Nonprofit and Voluntary Sector Marketing, 8(1), 2003, 12-29.

Bijlstra, G., and R. Dotsch. "FaceReader 4 Emotion Classification Performance on Images from the Radboud Faces Database." Radboud University Nijmegen Working Paper, 2011.

Bolton, G. E., and E. Katok. "An Experimental Test for Gender Differences in Beneficent Behavior." Economics Letters, 48(3), 1995, 287-92.

Breaban, A., and C. N. Noussair. Forthcoming. "Emotional State and Market Behavior." Review of Finance, 2017.

Caplin, A., and D. Martin. "Defaults and Attention: The Drop Out Effect." NBER Working Paper No. 17988, 2012.

Cappelletti, D., W. Gueth, and M. Ploner. "Being of Two Minds: An Ultimatum Experiment Investigating Affective Processes." Journal of Economic Psychology, 32, 2011,940-50.

Carlsson, F., O. Johansson-Stenman, and N. Pham Khanh. "Funding a New Bridge in Rural Vietnam: A Field Experiment on Conditional Cooperation and Default Contributions." University of Gothenburg Working Paper No. 503, 2011.

Carpenter, J., C. Connolly, and C. K. Myers. "Altruistic Behavior in a Representative Dictator Experiment." Experimental Economics, 11(3), 2008, 282-98.

Cason, T. N., and C. R. Plott. "Misconceptions and Game Form Recognition: Challenges to Theories of Revealed Preference and Framing." Journal of Political Economy, 122(6), 2014, 1235-70.

Choi, J. J., D. Laibson, B. C. Madrian, and A. Metrick. "Optimal Defaults." American Economic Review, 93(2), 2003, 180-5.

--. "For Better or for Worse: Default Effects and 401 (k) Savings Behavior," in Perspectives on the Economics of Aging, edited by D. A. Wise. Chicago: University of Chicago Press, 2004, 81-126.

Croson, R., and J. Shang. "The Impact of Downward Social Information on Contribution Decisions." Experimental Economics, 11(3), 2008, 221-33.

DellaVigna, S., J. A. List, and U. Malmendier. "Testing for Altruism and Social Pressure in Charitable Giving*." Quarterly Journal of Economics, 127(1), 2012, 1-56.

DellaVigna, S., J. A. List, U. Malmendier, and G. Rao. "The Importance of Being Marginal: Gender Differences in Generosity." American Economic Review, 103(3), 2013, 586-90.

Den Uyl, M. J., and H. Van Kuilenburg. "The FaceReader: Online Facial Expression Recognition," in Proceedings of Measuring Behavior, Vol. 30, Wageningen, The Netherlands, 2005, 589-90.

Dhingra, N., Z. Gom, A. Kener, and J. Dana. "The Default Pull: An Experimental Demonstration of Subtle Default Effects on Preferences." Judgment & Decision Making, 7(1), 2012, 69-76.

Dinner, I., E. J. Johnson, D. G. Goldstein, and K. Liu. "Partitioning Default Effects: Why People Choose Not to Choose." Journal of Experimental Psychology: Applied, 17(4), 2011, 432.

Doyle, L., and D. Schindler. "Mucap: Connecting facereader[TM] to z-tree." Munich Discussion Paper No. 2015-4, 2015.

Ebeling, F. "Non-binding Defaults and Voluntary Contributions to a Public Good--Clean Evidence from a Natural Field Experiment." Working Paper Series in Economics No. 66, University of Cologne, Department of Economics, 2013.

Eckel, C. C., and P. J. Grossman. "Rebate versus Matching: Does How We Subsidize Charitable Contributions Matter?" Journal of Public Economics, 87(3), 2003, 681-701.

Edwards, J., and J. List. "Toward an Understanding of Why Suggestions Work in Charitable Fundraising: Theory and Evidence from a Natural Field Experiment." Journal of Public Economics, 114, 2014, 1-13.

Ekman, P. Emotions Revealed: Recognizing Faces and Feelings to Improve Communication and Emotional Life. New York: Macmillan, 2007.

Ekman, P., and W. V. Friesen. "A New Pan-Cultural Facial Expression of Emotion." Motivation and Emotion, 10(2), 1986, 159-68.

Elbel, B., R. Kersh, V. L. Brescoll, and L. B. Dixon. "Calorie Labeling and Food Choices: A First Look at the Effects on Low-Income People in New York City." Health Affairs, 28(6), 2009, w1110-21.

Engel, C. "Dictator Games: A Meta Study." Experimental Economics, 14(4), 2011, 583-610.

Fischbacher, U. "z-Tree: Zurich Toolbox for Ready-made Economic Experiments." Experimental Economics, 10(2), 2007, 171-78.

Fosgaard, T. R., and M. Piovesan. "Nudge for (the Public) Good: How Defaults Can Affect Cooperation." PLoS ONE, 10(12), 2015.

Genevsky, A., D. Vastfjall, P. Slovic, and B. Knutson. "Neural Underpinnings of the Identifiable Victim Effect: Affect Shifts Preferences for Giving." Journal of Neuroscience, 33(43), 2013, 17188-96.

Grossman, P., and C. Eckel. "Giving Versus Taking for a Cause." Economics Letters, 132, 2015, 28-30.

Haggag, K., and G. Pad. "Default Tips." American Economic Journal: Applied Economics, 6(3), 2014, 1-19.

Isen, A. M., and P. F. Levin. "Effect of Feeling Good on Helping: Cookies and Kindness." Journal of Personality and Social Psychology, 21(3), 1972, 384-88.

Jenni, K., and G. Loewenstein. "Explaining the Identifiable Victim Effect." Journal of Risk and Uncertainty, 14(3), 1997,235-57.

Johnson, E. J., and D. Goldstein. "Do Defaults Save Lives?" Science, 302(5649), 2003, 1338-39.

Kahneman, D., J. L. Knetsch, and R. H. Thaler. "Anomalies: The Endowment Effect, Loss Aversion, and Status Quo Bias." The Journal of Economic Perspectives, 5(1), 1991, 193-206.

Karlan, D., and J. A. List. "How Can Bill and Melinda Gates Increase Other People's Donations to Fund Public Goods?" National Bureau of Economic Research Working Paper No. 17594, 2012.

Kivetz, R., and I. Simonson. "The Effects of Incomplete Information on Consumer Choice." Journal of Marketing Research, 37(4), 2000, 427-48.

Koriat, A., S. Lichtenstein, and B. Fischhoff. "Reasons for Confidence." Journal of Experimental Psychology: Human Learning and Memory, 6(2), 1980, 107-18.

Korobkin, R. "Status Quo Bias and Contract Default Rules." Cornell Law Review, 83, 1997, 609-87.

Kudema-Iulian, B., C. Marcel, and T. Valeriu. "Towards an Affective Aware Home," in Ambient Assistive Health and Wellness Management in the Heart of the City. Berlin Heidelberg: Springer, 2009, 74-81.

Landry, C. E., A. Lange, J. A. List, M. K. Price, and N. G. Rupp. "Toward an Understanding of the Economics of Charity: Evidence from a Field Experiment." Quarterly Journal of Economics, 121(2), 2006, 747-82.

Lange, A., J. A. List, and M. K. Price. "A Fundraising Mechanism Inspired by Historical Tontines: Theory and Experimental Evidence." Journal of Public Economics, 91(9), 2007, 1750-82.

Lemer, J. S., D. A. Small, and G. Loewenstein. "Heart Strings and Purse Strings Carryover Effects of Emotions on Economic Decisions." Psychological Science, 15(5), 2004,337-41.

Lewinski, R, M. L. Fransen, andE. S. Tan. "Predicting Advertising Effectiveness by Facial Expressions in Response to Amusing Persuasive Stimuli." Journal of Neuroscience, Psychology, and Economics, 7(1), 2014, 1-31.

Lin, C., S. Chuang, D. TT. Kao, and C. Kung. "The Role of Emotions in the Endowment Effect." Journal of Economic Psychology, 27(4), 2006, 589-97.

List, J. A., and M. K. Price. "The Role of Social Connections in Charitable Fundraising: Evidence from a Natural Field Experiment." Journal of Economic Behavior & Organization, 69(2), 2009, 160-69.

Loewenstein, G., C. Bryce, D. Hagmann, and S. Rapal. "Warning: You Are About to Be Nudged." Behavioral Science & Policy, 1(1), 2015, 35-42.

Madrian, B. C., and D. F. Shea. "The Power of Suggestion: Inertia in 401 (k) Participation and Savings Behavior." Quarterly Journal of Economics, 116(4), 2001, 1149-87.

Martinez, L. F., M. Zeelenberg, and J. B. Rijsman. "Regret, Disappointment and the Endowment Effect." Journal of Economic Psychology, 32(6), 2011, 962-8.

Matsumoto, D., and B. Willingham. "Spontaneous Facial Expressions of Emotion of Congenitally and Noncongenitally Blind Individuals." Journal of Personality and Social Psychology, 96(1), 2009, 1-10.

McKenzie, C. R. M., M. J. Liersch, and S. R. Finkelstein. "Recommendations Implicit in Policy Defaults." Psychological Science, 17(5), 2006, 414-20.

Messer, K. D" H. Zarghamee, H. M. Kaiser, and W. D. Schulze, "New Hope for the Voluntary Contributions Mechanism: The Effect of Context." Journal of Public Economics 91(9), 2007, 1783--99.

Nguyen, Y., and C. N. Noussair. "Risk Aversion and Emotions." Pacific Economic Review, 19(3), 2014,296-312.

Rubaltelli, E., and S. Agnoli. "The Emotional Cost of Charitable Donations." Cognition & Emotion, 26(5), 2012, 769-85.

Sargeant, A., D. C. West, and J. B. Ford. "Does Perception Matter? An Empirical Analysis of Donor Behaviour." The Service Industries Journal, 24(6), 2004, 19-36.

Schwarz, N., L. J. Sanna, I. Skumik, and C. Yoon. "Metacognitive Experiences and the Intricacies of Setting People Straight: Implications for Debiasing and Public Information Campaigns." Advances in Experimental Social Psychology, 39, 2007, 127-61.

Schweitzer, M. "Multiple Reference Points, Framing, and the Status Quo Bias in Health Care Financing Decisions." Organizational Behavior and Human Decision Processes, 63(1), 1995,69-72.

Shang, J., and R. Croson. "A Field Experiment in Charitable Contribution: The Impact of Social Information on the Voluntary Provision of Public Goods." Economic Journal, 119(540), 2009, 1422-29.

Small, D. A., and U. Simonsohn. "Friends of Victims: Personal Experience and Prosocial Behavior." Journal of Consumer Research, 35(3), 2008, 532-42.

Small, D. A., and N. M. Verrochi. "The Face of Need: Facial Emotion Expression on Charity Advertisements." Journal of Marketing Research, 46(6), 2009, 777-87.

Terzis, V., C. N. Moridis, and A. A. Economides. "Measuring Instant Emotions during a Self-assessment Test: The Use of FaceReader," in Proceedings of the 7th International Conference on Methods and Techniques in Behavioral Research. Eindhoven, The Netherlands: ACM, 2010, 18-22.

Tversky, A., and D. Kahneman. "Judgment under Uncertainty: Heuristics and Biases." Science, 185(4157), 1974, 1124-31.

Van Leeuwen, B., C. N. Noussair, T. J. S. Offerman, S. Suetens, M. Van Veelen, and J. Van de Ven. Forthcoming. "Predictably Angry: Facial Cues Provide a Credible Signal of Destructive Behavior." Management Science.

Zikmund-Fisher, B. J., P. A. Ubel, D. M. Smith, H. A. Derry, J. B. McClure, A. Stark, R. K. Pitsch, and A. Fagerlin. "Communicating Side Effect Risks in a Tamoxifen Prophylaxis Decision Aid: The Debiasing Influence of Pictographs." Patient Education and Counseling, 73(2), 2008,209-14.

(1.) Defaults have been shown to exert effects on the choice of whether or not to donate organs (Johnson and Goldstein 2003), as well as on savings decisions (Madrian and Shea 2001). As many researchers have argued, defaults can be seen as implicit recommendations or an experimenter's expectation (McKenzie, Liersch, and Finkelstein 2006), or as social norms (Carlsson, Johansson-Stenman, and Pham Khanh 2011), which may be costly to violate. Defaults may anchor participants' decisions (Dhingra etal. 2012; Dinner et al. 2011), and switching away from them might involve cognitive effort (Johnson and Goldstein 2003). People can also suffer from status quo bias (Kahneman, Knetsch, and Thaler 1991; Korobkin 1997) and hence be unwilling to override the default. People can simply procrastinate and keep on postponing a decision, effectively making the default option apply (as argued by Choi et al. 2003).

(2.) Andreoni (2006) shows that framing a social dilemma as an opportunity to contribute to provide a positive externality generates higher contributions than a payoff-equivalent framing that allows players to withdraw funds from a group account.

(3.) Loewenstein etal. (2015) study whether informing individuals about default effects negates these effects in the choice of advance directives about medical care. These directives specify instructions for physicians in the event that a patient loses the capacity to make his own decisions. For such decisions, default effects were observed, and these persisted when it was disclosed beforehand that defaults affect decisions.

(4.) Engel (2011), in a metastudy of dictator game experiments, reports that dictators give away on average around 28% of their endowment. If the recipient is needy or deserving (such as a charity), donations are higher.

(5.) See Appendix B for a cross-tabulation of the treatment to which subjects were assigned and the previous experiment in which they participated.

(6.) The characteristics of the charity in question matter, and must be taken into account when considering how to increase donations. A charity seen as effective and providing quality services is likely to attract greater donations (Sargeant, West, and Ford 2004). Personal involvement or experience can induce people to choose one charity over another (Bennett 2003). Empathy with a particular cause leads to greater donations (Basil, Ridgway, and Basil 2008; Small and Simonsohn 2008).

(7.) The newest version of Facereader, version 7, can also measure contempt and arousal.

(8.) While there is consensus in the literature that happiness is positive in valence, and that anger, fear, sadness, and disgust are negative in valence, there is no consensus about how an aggregate measure of valence should be constructed from these specific emotions. We chose our particular measure because, unlike the measure that Facereader computes, it always places some weight on each of the negative emotions.

(9.) Facereader requires a nearly full frontal view of the subject to compute an emotional state. Before the reading of instructions, most subjects are looking away from their computer screen in other directions. During the reading of the instructions, facial expression data can typically be registered part of the time. We were able to obtain data for most individuals in the few seconds before submitting their donation and in the few seconds afterward.

(10.) We conducted several regressions to verify that the effects of gender and of session size are robust when we control for the other effect. Specifically, we estimated the following equations: [Donation.sub.i] = [[beta].sub.1]Male + [[beta].sub.2]Sessionsize + [[epsilon].sub.i]; [Donation.sub.i] = [[beta].sub.1]Male + [[beta].sub.2]Sessionsize + [[beta].sub.3]Owncharity + [[SIGMA].sub.k] [[beta].sub.k][D.sub.k] + [[epsilon].sub.i]; and [Donation.sub.i] = [[beta].sub.1]Male + [[beta].sub.2]Sessionsize + [[beta].sub.3]Owncharity + [[SIGMA].sub.k] [[beta].sub.k][D.sub.k] + [[SIGMA].sub.m] [[beta].sub.m][E.sub.m] + [[epsilon].sub.i]. In the second and third regressions, there are six dummy variables [D.sub.k], one for each treatment other than the Baseline. In the third regression, there are also six variables [E.sub.m], where m[member of] (Happy, Sad, Scared, Angry, Disgusted, Surprised), corresponding to the level of each of the emotions experienced by participant i before the donation decision. The estimated coefficients, associated t statistics, and p values in the first equation are [[beta].sub.1] = -.6013, t = -3.69, p< = .001, and [[beta].sub.2] = -.0673, t = -3.46, p = .001. For the second equation, the estimated coefficients are [[beta].sub.1] = -.657895, t=-3.96, p< .001; [[beta].sub.2] = -.800, t = -3.63, p <.001; and [[beta].sub.3] = -.9412, t = -2.57, p < .01; For the third equation, the estimated coefficients are [[beta].sub.1] =-.7267, t = -4.11, p < .001; [[beta].sub.2] = -.0816, t = -3.37, p = .001; and [[beta].sub.3] = -.9513, t = .3749, p = .012. See also estimated Equation (6) in Table 3 for another specification. Thus, there is strong evidence that both gender and session size have independent and significant effects on donation.

(11.) Three percent of women and 8% of men choose their charity. Among the individuals who specify a charity, there is no gender difference in donation level. Males donate 2.4, and females 2.8, on average, and a rank-sum test fails to reject equality (p = .357).

Caption: FIGURE 1 Decision Screen

Caption: FIGURE 2 Timing of Emotion Measurements

Caption: FIGURE 3 Average Donation by Treatment

Caption: FIGURE 4 Donation Distribution

Caption: FIGURE 5 Donation Distribution by Treatment

Caption: FIGURE 6 Emotions Prior to Decision (Facereader)

Caption: FIGURE 7 Donations and Subsequent Self-Reported Emotions

Caption: FIGURE 8 Self-Reported Emotions after Decision

Caption: FIGURE 9 Facereader Emotions after Decision

Caption: FIGURE 10 Average Donation by Session Size

TABLE 1

The Treatments

               Number of   Number of    Number of
Condition      Subjects    Sessions    Self-Reports

Baseline          84          22            53
Low               31           8            0
Medium            31          14            0
High              31           6            0
Low info          62          19            31
Medium info       62          13            31
High info         62          21            31

Notes: The table reports summary statistics of the number
of subjects in each treatment, on how many sessions were
run for each treatment (there were seven sessions where
multiple treatments were run at once), and how many subjects
completed a questionnaire about their emotional state after the
donation decision was completed.

TABLE 2

Wilcoxon Rank-Sum Test of the Effect of
Defaults and Information on Donations

Null Hypothesis                      p Value

Baseline = Low                        .651
Baseline = Medium                     .506
Baseline = High                       .372
Baseline = Low info                   1.000
Baseline = Medium info                .920
Baseline = High info                  .936
Low = Medium                          .338
Medium = High                         .910
Low = High                            .206
Low info = Medium info                .986
Medium info = High info               .958
Low info = High info                  .994
Low = Low info                        .543
Medium = Medium info                  .559
High = High info                      .388
Info = No info                        .757
Low & Low info = High & High info     .480

Notes: The table gives the p values of rank-sum tests conducted
between different pairs of treatments. The baseline
treatment is compared to all conditions. Each default level
with no information is compared its counterpart with information.
Within each information condition, each of the three
default levels is compared to the two other default levels. The
results show that no treatment is significantly different from
any other. Hence, we find no effect of defaults, and no effect
of information, on donations.

TABLE 3

Robustness Check: Is There a Default Effect after All?

                      (1)         (2)          (3)

Low                 -0.172       -0.236      -0.218
                    (0.343)     (0.338)      (0.342)

Medium               0.247       0.241        0.253
                    (0.343)     (0.338)      (0.341)

High                 0.255       0.230        0.286
                    (0.343)     (0.338)      (0.342)

Low*info             0.129       0.158        0.206
                    (0.359)     (0.354)      (0.359)

Medium*info         -0.242       -0.261      -0.204
                    (0.359)     (0.354)      (0.358)

High*info           -0.246       -0.284      -0.272
                    (0.359)     (0.354)      (0.357)

Male                           -0.590 ***
                                (0.169)

Own charity                                 0.792 **
                                             (0.377)

Session size

Session size sq.

Happy before

Constant           1.777 ***   2.107 ***    1.720 ***
                    (0.178)     (0.199)      (0.179)

Observations          363         363          363

R-squared            0.005       0.038        0.017

                      (4)          (5)         (6)

Low                  -0.218      -0.166       -0.322
                    (0.335)      (0.355)     (0.341)

Medium               -0.086       0.347       0.006
                    (0.343)      (0.343)     (0.336)

High                 0.317        0.336       0.419
                    (0.335)      (0.360)     (0.344)

Low*info             0.196        0.228       0.341
                    (0.369)      (0.369)     (0.372)

Medium*info          0.222       -0.189       0.240
                    (0.379)      (0.357)     (0.371)

High*info            -0.383      -0.217       -0.507
                    (0.369)      (0.376)     (0.375)

Male                                        -0.673 ***
                                             (0.168)

Own charity                                  0.936 **
                                             (0.371)

Session size       -0.277 ***               -0.272 ***
                    (0.075)                  (0.075)
Session size sq.   0.013 ***                0.013 ***
                    (0.005)                  (0.005)
Happy before                    1.243 **     0.997 *
                                 (0.523)     (0.522)
Constant           2.910 ***    1.515 ***   2.972 ***
                    (0.324)      (0.203)     (0.353)

Observations          363          349         349

R-squared            0.056        0.023       0.119

Notes: The table contains estimates of OLS regressions with
donations as the dependent variable, ranging from 0 [euro] to
5 [euro]. The unit of observation in the individual decision
maker. Low, Medium, and High are dummy variables for the default
level, and Low*info, Medium*info, and High*info are interaction
terms for the informed treatments. Own charity is a dummy variable
that equals 1 if an individual chooses her own charity. Session
size is the number of individuals participating in the experimental
session. Standard errors in parentheses. Happy Before denotes the
happiness level recorded by Facereader prior to the donation
decision.

* Significant at the 1% level; ** significant at the 5% level;
*** significant at the 10% level.

TABLE 4

Donations and Emotions as Registered with Facereader Both Prior
and Subsequent to the Donation Decision

                               0      >0    0.25-1.75

No. of observations before     74    275       116
No. of observations after      67    215       92
Valence before                0.11   0.12     0.11
Valence after                 0.10   0.11     0.11
Happiness before              0.14   0.15     0.14
Happiness after               0.13   0.14     0.14
Sadness before                0.04   0.05     0.06
Sadness after                 0.06   0.06     0.05

                              2.00-3.50   3.75-5.00   Overall

No. of observations before       103         56         349
No. of observations after        81          42         282
Valence before                  0.10        0.19       0.12
Valence after                   0.10        0.10       0.11
Happiness before                0.12        0.22       0.15
Happiness after                 0.13        0.14       0.14
Sadness before                  0.04        0.05       0.05
Sadness after                   0.06        0.07       0.06

Notes: The table lists the average level of valence, happiness, and
sadness, before and after donation for those who donate varying
amounts. The number of observations may differ between before and
after the session depending on whether we were able to obtain any
readings during the relevant period. Missing Facereader data are
due to a failure of the subject to look straight at the screen
during the measurement period, or to the occasional webcam
malfunction.

TABLE 5

Signed-Rank Test: Change in Emotions Conditional
on Donation Amounts

                         0       >0     0.25-1.75

No. of observations     66      215        92
Valence                0.257   <0.001     0.152
Happiness              0.177   <0.001     0.123
Sadness                0.831   0.163      0.385

                      2.00-3.50   3.75-5.00   Overall

No. of observations      81          42         281
Valence                 0.098      <0.001     <0.001
Happiness               0.276      <0.001     <0.001
Sadness                 0.193       0.925      0.187

Note: The table lists the p values from signed-rank tests
comparing whether the level of emotion prior to donation decision
equals the level of the emotion after the decision.