Can I Rely on You?

Abstract

This paper introduces a strategic element into the dictator game by allowing recipients to select their dictator. Recipients are presented with the photographs of two dictators and the envelopes containing their allocations, and are then asked to select which dictator’s gift they would like to receive. The recipient is paid the contents of the envelope they select. The photographs carry information about the gender and race/ethnicity of the dictators, and we ask an independent sample of raters to evaluate the photographs for other characteristics. While gender and ethnicity do not affect the recipient’s choice, one characteristic inferred from the photos makes them significantly more likely to be selected: Their perceived reliability.

1. Introduction

Many interactions take place online. For example, people create profiles on LinkedIn to network professionally, on ziprecruiter.com to apply for jobs, on eBay to sell items, and on Tinder to find their soulmates. With the level of convenience and popularity of these online platforms, our faces become our brands and provide the very first impression in online interactions. Most people, if they are asked, believe that they are able to read other’s faces, and can judge trustworthiness, generosity and a willingness to cooperate in others. These assessments may be based on the person’s gender, ethnicity, attractiveness or other facial features. It is reasonable to ask just how good people are at making economically meaningful choices based on quick judgments from facial images. In this study we investigate the perceived generosity of dictators using the behavior of recipients who are able to choose their dictators.

In this study our primary focus is on subjects who are recipients in a modified dictator game. Each recipient is presented a pair of envelopes. The envelopes contain the allocation choices of dictators from a prior experiment, and have a photograph of the dictator affixed to the corner. Recipients decide which envelope, from the pair, that they wish to keep. Given that the subjects have no other basis by which to judge the contents of the envelopes, we assume that the photos are the basis of the choice. A rational recipient would select the envelope that she expects to contain the largest amount of money.

The design blocks the pairings of dictator photos based on gender and ethnicity. We have three findings. First, there is no obvious pattern in the choice of a dictator by gender or ethnicity. We expected that ethnicity would be a powerful cue, but that is not the case. Second, and related to this point, is that subjects do not choose dictators that are like themselves. For example, a white woman does not automatically choose a dictator who is also a woman, or a white person. Finally, subjects are reading something systematic from the photos. They are more likely to select photos that others have rated as being “reliable”, where the reliability measure is the outcome of a factor analysis of a set of survey questions used to evaluate the photos. However, this judgement is not accurate, in that the “reading” of the dictator does not lead to a larger payoff for the recipient.

2. Motivation

There is an extensive literature on reading faces in dyadic games (particularly the trust game). For example, Scharlemann et al. show that first-movers in a trust game are more likely to extend their trust to a smiling face [1] (also see Centorrino et al. [2] and the references therein). People exhibit a greater trust in a hypothetical trust game where kinship is induced by morphing a trustor’s face with the counterpart’s face [3]. People are more likely to trust an attractive counterpart, and expect higher levels of trust from attractive first movers in the game [4] (see also Hancock and DeBruine [5], and Smith et al. [6]). Cultural cues are also important, ranging from skin color to the gender or ethnicity of one’s counterpart [7,8,9,10,11,12]. All affect beliefs about trustworthiness.

Separate studies have focused on whether there are particular elements of human faces that signal trustworthiness. By systematically searching for facial markers of trustworthiness, Todorov and Duchaine [13] generate a set of computerized faces to match these markers. Subjects were able to pick out such faces as trustworthy, even when they were exposed to the image for as little as 100 milliseconds [14,15,16]. Several papers confirm that subjects are more likely to send money to faces rated as more trustworthy [17,18,19]. This finding is replicated in part by Tingley [20] who allows subjects to pick an avatar image to play the trust game. The images are taken from the stock of images generated by Todorov and Duchaine [13]. Subjects are more likely to choose an image that is rated as trustworthy and to act on the basis of the perceived trustworthiness of the avatar selected by their counterpart (see also Desteno et al. [21] and the use of a different type of avatar). In short when playing the trust game subjects are attentive to the facial features signaling the trustworthiness of their counterpart. Indeed, Eckel and Petrie [22] measure the willingness of subjects to pay to see the photographs of their counterparts in the trust game, indicating that subjects do expect to be able to infer trustworthiness from their reading of a person’s face.

There are a number of studies using faces in the dictator game, typically (but not always) using photos of the recipient to convey information to the dictator. Additionally, Habyarimana et al. [23] include a dictator game in a multi-stage experiment for the purpose of measuring ethnic bias. In this experiment dictators view the images of two recipients and decide how to allocate a fixed pie between themselves and the recipients. The authors vary the ethnicity of the recipients and find little evidence for ethnic bias (in part because subjects had difficulty identifying differences among ethnic groups in the study site of Uganda). They find that about 40 percent of the total pie is allocated to the photographed recipients. Attractiveness of the recipient may also play a role in dictator’s decision. For example, Rosenblat [24] uses a photo condition in a dictator experiment in which the dictator sees a headshot of the recipient. She focuses on the attractiveness of the recipient and finds a slight beauty penalty in that condition—more attractive people receive less. But when dictators also hear the voice of the recipient, the beauty premium reappears, with attractive people receiving more. However, Bhogal et al. [25] find little effect in allocations accounted for by beauty in an experiment in which the dictator can see the recipient, but not vice-versa (using a lab with one-way mirrors). They find similar results in a second study in which both dictators and recipients can see one another. The closest study to our own is Burnham [26]. They take photos of the dictators and use them to examine the effect on giving of revealing the dictator’s identity to the recipient.1 (Also see Buchanan et al. [27] for a videotaped version of this experiment). Furthermore, research finds that just looking at abstract eyes, which may create the impression that dictator’s actions are being watched, increases dictator giving [28]. Almost all of the interest by researchers focuses on dictator behavior, and how dictator giving varies depending on cues embedded in the photo or video images of the recipient. In this design we instead allow the recipients to choose between dictators based on their photos. We focus on recipient behavior, and examine whether recipients infer something about the dictator based on their photographic image. We are interested in whether recipients, when forced to choose between two dictators, make systematic choices that reveal a preference for gender, ethnicity or some other observable characteristic of the photograph.

3. Materials and Methods

In this experiment, subjects played the role of the recipient in the dictator game. We provided them a pair of envelopes. Each envelope contained the amount left by a dictator from a previous study as explained in Appendix A, as well as a small color picture of the dictator that was affixed in the upper right corner, outside the envelope. Using neutral language, we explained how the dictators made decisions in a previous experiment, and asked recipients to choose one of the two envelopes in the pair to keep (instructions are provided in Appendix B). In other words, each recipient was provided two envelopes and chose their dictators, keeping whatever was in the envelope. Each recipient was provided four unique pairs of envelopes, one at a time, and chose one from each pair. Thus, they made a total of four choices. Next, they answered a brief survey. At the end, two of the four envelopes the recipients picked were randomly selected. Subjects opened these envelopes and the experimenter recorded their selections and their earnings.

Each envelope contained 15 objects (a combination of one-dollar bills and blank paper slips of the same size and texture as one-dollar bills). The number of one-dollar bills in each envelope was determined by the dictator whose photo was attached to the envelope. For example, if the dictator left $7, then the envelopes with this dictator’s photo contained seven one-dollar bills and eight paper slips. This was to ensure that the envelopes themselves did not signal to the recipient anything about the contents.

The dictator decisions presented to the subjects were selected from the decisions of two groups of subjects who participated as dictators prior to this experiment. One group of subjects at the University of Houston (UH) (August 2001) and another group at Virginia Tech (VT) (February 2002) participated in a dictator game experiment as dictators (details are provided in Appendix A). We collected the dictator decision data from these two different locations to ensure diversity in our dictator sample. After we collected the decisions from the initial set of 51 dictators, we selected 22 dictators to be used in this study. We used the dictators’ ethnicity, gender and the giving behavior in the dictator game as the basis for generating our envelope pairs.

Table 1. Experimental design.

(a) Envelope Pairings
Ethnicity			Gender			Fairness
Black/MSA			male/male			self-interested/self-interested
White/MSA			female/female			generous/generous
MSA/MSA			mixed			generous/self-interested
White/White						self-interested/generous
Black/Black
(b) Number of pairings in each cell
		Black			White			MSA
		Female		Male	Female		Male	Male	Female
Black	Female	3
	Male	4		2 *
White	Female	-		-	3
	Male	-		-	4		3
MSA	Female	4		3 *	4		4	3
	Male	4		4	4		4	4	3

* These two cells are missing an envelope pairing each. The missing pairings are indicated in Appendix C. We have only three fairness pairings in the pairs with the same gender and the same ethnicity cells. This is because generous/self-interested combination is the same as self-interested/generous combination in these cells. MSA, Muslim Student Association.

presents the details of the pairings presented to subjects in our design. These consist of 5 (ethnicity) × 3 (gender) × 4 (fairness) dictator conditions. In the aftermath of 11 September 2001, we focused on photos of dictators that were recruited from the Muslim Student Association at Virginia Tech (we will refer to this group as MSA); and we also included African-American (black) and Caucasian-American (white) subjects. From our reduced set of dictators, we created five ethnicity pairings, as seen in the first column of Table 1 (a). We also controlled for the gender of the dictators, which gives us three gender pairings, as seen in the second column. Finally, we classified the photos based on what they sent to their counterpart. By design, this final categorization was not revealed to the recipients. However, we were interested to study whether people would be able to detect the more generous decisions by just looking at the dictators’ photos. Dictators were classified as either self-interested (sending five dollars or less) or generous (sending seven dollars or more). We had a total of 60 unique envelope pairings. The breakdown of these pairings can be seen in Table 1 (b) and the full list is provided in Appendix C. Due to the fact that we had a limited number of dictators, we were careful to ensure that no subject saw the same picture twice. The choice of the photo-tagged envelope (when presented with its paired envelope) constitutes the main experimental outcome and the dependent variable in our analysis below.

Evaluation of Dictator Photos

An independent group of students evaluated the photos of the dictators in Fall 2001 and Spring 2002. On average each image was evaluated by 21.8 subjects. VT students evaluated the photos of the dictators from the UH; and Rice University students evaluated the photos of the dictators from VT. This was designed to ensure that subjects doing the evaluations did not know the person being evaluated. The details of the photo evaluation are provided in Appendix D.

Using 15 opposite word/pair items, students were asked to indicate how well they thought each specific word pair fit the person in the photo. For example, given the word pair Honest/Dishonest, students were asked to rate the photo on a six-point scale. In this instance if the student thought the person in the photo was honest, then she would choose values from 1 through 3, with 1 being the highest category of honesty. However, if the student thought the person in the image was dishonest, then she would pick values from 4 through 6, with the latter representing the most dishonest category. Photos are rated on these dimensions: Motivated, trusting, competitive, trustworthy, hardworking, intelligent, selfish, honest, respectful, attractive, excitable, friendly, happy, accepting, and dependable (see Appendix D for the set of items and a sample decision sheet).

We use the results of these evaluations as control variables in the analysis. Because of the focus on attractiveness in prior studies, the attractiveness rating enters directly. We also used factor analysis to identify a second measure. First, we take the remaining 14 opposite word-pair items and scale them in the same direction. Exploratory factor analysis indicate that these items loaded on a single dimension (only one factor has an eigenvalue greater than 1). Using Cronbach’s alpha (α = 0.86) as a guide we removed four items, leaving the factor containing the following measures: Motivated, trusting, trustworthy, hardworking, intelligent, honest, respectful, friendly, happy, accepting, and dependable. Considering the word-pair items that load on this factor, we believe it is related to one’s perceived level of reliability. Thus, we term this factor “reliable”. From this we generate factor scores for a single factor and calculate the average score for each dictator. We use these values as a covariate.

4. Results

A total of 98 subjects (59% male) played the dictator game as recipients and made a total of 278 decisions.2 At Rice University (41 subjects), subjects were recruited from the Behavioral Research Laboratory subject pool database and from commons areas on the Rice campus. At Virginia Tech (VT) (57 subjects), the experiments were conducted in the Lab for the Study of Human Thought and Action (LSHTA), and recruited from the LSTA subject pool by email. The experiments were conducted in March 2002 and in November 2002.

We first study whether gender or ethnicity of the dictator affected their probability of being chosen by the recipients. The probability of each dictator being chosen given the characteristics of the other envelope in the pair is provided in

Table 2. The probability of a dictator being chosen given the paired dictator’s characteristics.

			Not Chosen
			Black		White		MSA
			Female	Male	Female	Male	Female	Male
CHOSEN	Black	Female	not reported	0.50	-	-	0.31	0.76
		Male	0.50	not reported	-	-	0.61	0.45
	White	Female	-	-	not reported	0.50	0.57	0.44
		Male	-	-	0.50	not reported	0.45	0.35
	MSA	Female	0.69	0.39	0.43	0.55	not reported	0.68
		Male	0.24	0.55	0.56	0.65	0.32	not reported
		Overall	0.47		0.55

The number of times each dictator was chosen given the paired dictator’s characteristics can be found in Table A2 (Appendix E). MSA, Muslim Student Association.

. On average, when MSA dictators were matched with black dictators, MSA dictators were chosen 47% of the time. When they were matched with White dictators, MSA dictators were chosen 55% of the time. Additionally, when females and males were matched, females were picked 56% of the time. Using a binomial probability test, we find that the probabilities presented in Table 2 are not statistically different than a random draw of 50% except one. When male MSA dictators were paired with female black dictators, they were significantly less likely to be chosen (p-value: 0.049).

We further explore the data using a fixed effects conditional logit model.3 This model lets us study subjects’ choice behavior based on the characteristics of the alternative envelope (dictators), as well as the decision makers (recipients).

Table 3. Conditional fixed effects logit regression results.

DV: Chosen	(1)	(2)	(3)	(4)
Dictator Characteristics
Female Dictator	1.293	1.288	1.206	1.196
	(0.336)	(0.345)	(0.340)	(0.349)
White Dictator	0.830	0.921	1.174	1.329
	(0.244)	(0.370)	(0.384)	(0.555)
White Female Dictator	1.092	1.112	0.781	0.793
	(0.424)	(0.447)	(0.328)	(0.345)
Black Dictator	1.394	1.395	1.252	1.254
	(0.445)	(0.451)	(0.407)	(0.411)
Black Female Dictator	0.749	0.752	0.692	0.696
	(0.330)	(0.335)	(0.312)	(0.318)
Self-interested	1.135	1.130	1.021	1.014
	(0.215)	(0.213)	(0.203)	(0.201)
Dictator and Recipient Characteristics
Same Gender		0.986		0.985
		(0.204)		(0.206)
Same Ethnicity		0.823		0.823
		(0.408)		(0.410)
Same Ethnicity and Gender		1.078		1.028
		(0.445)		(0.423)
Dictator Face Evaluation Variables
Attractive			0.886	0.888
			(0.205)	(0.205)
Reliable			1.874 ***	1.874 ***
			(0.443)	(0.443)
Number of Subjects	98	98	98	98
Observations	556	556	556	556
Log pseudolikelihood	–191.3	–191.3	–186.8	–186.8

*** p < 0.001. Odds ratios are presented. Standard errors clustered at individual level are in parentheses.

presents estimated odds ratio for each variable; significance tests are for differences from 1.

The dependent variable in all models is “Chosen”, which takes the value of one if the envelope was chosen, otherwise zero. The variables Female Dictator, White Dictator, White Female Dictator, Black Dictator, and Black Female Dictator are dummy variables that take the value of one if the picture on the envelope was female, white, white female, black, or black female respectively. Self-Interested is a dummy variable that takes the value of one if the dictator sent $5 of less, and zero otherwise. This was not observed by the recipients. We include this to check whether the recipients were good at guessing which envelopes contained more money. Same Gender, Same Ethnicity and Same Ethnicity and Gender are dummy variables that take the value of one if the dictator and the decision maker shared the same gender, ethnicity and both respectively. Finally, we include two variables that are the evaluations of the dictator’s face. The variable Attractive is the average attractiveness evaluation of the dictator. The higher the value the more attractive the photo was rated. We also created a factor (Reliable) using a factor analysis of the facial evaluations of each dictator. Reliable is constructed using an exploratory factor analysis, as described above. The higher the value, the more the dictator was perceived as reliable.

Looking at the first column in Table 3, we see that neither the ethnicity nor the gender of the dictator has an impact on the odds of a dictator being chosen. Additionally, we include Self-interested to check whether the recipients are good at guessing which envelopes contain more money. We conclude that subjects are not good at picking the envelope with the larger amount left, since the coefficient is not significant. In fact, they do not do better than randomly selecting one of the envelopes.4 Column 2 shows that whether the decision maker (the recipient) and the dictator share the same Ethnicity, gender, or both have no significant impact on the odds of an envelope being chosen. Finally, in columns 3 and 4, we include variables related to the facial evaluations of the dictators. Like Bhogal et al. [25], we find that attractiveness of the dictator does not play a significant role on recipient’s decision. However, we do find that dictators with higher reliability ratings have higher odds of being chosen.

We also analyze the data using a mixed logit model. Mixed logit model has some advantages over the conditional logit model, since it relaxes some underlying assumptions (see Hensher and Greene [29] for a detailed discussion of this model). Here, we are interested in the preference heterogeneity in our sample. Appendix E

Table A3. Random Parameter (Mixed) Logit Estimation Results.

Variables	(1)		(2)		(3)		(4)
DV: Chosen	mean	s.d.	mean	s.d.	mean	s.d.	mean	s.d.
Female Dictator	1.361	1.230	1.372	0.835	1.234	1.007	1.241	1.007
	(0.559)	(2.552)	(0.822)	(1.957)	(0.443)	(0.027)	(0.502)	(0.027)
White Dictator	0.807	1.730	0.892	1.675	1.165	2.182	1.298	2.166
	(0.270)	(4.887)	(0.469)	(6.536)	(0.466)	(3.123)	(0.654)	(3.565)
White Female Dictator	1.085	2.467	1.105	2.711	0.692	0.962	0.698	0.962
	(0.553)	(6.341)	(0.597)	(10.174)	(0.396)	(0.113)	(0.432)	(0.131)
Black Dictator	1.417	1.067	1.418	1.071	1.263	1.042	1.265	1.042
	(0.533)	(0.068)	(0.548)	(0.086)	(0.443)	(0.078)	(0.444)	(0.082)
Black Female Dictator	0.716	0.995	0.716	1.005	0.651	1.044	0.651	1.043
	(0.399)	(0.068)	(0.457)	(0.096)	(0.342)	(0.091)	(0.355)	(0.085)
Self-interested	1.125		1.121		0.984		0.979
	(0.249)		(0.255)		(0.233)		(0.231)
Same Gender			1.022				1.019
			(0.416)				(0.259)
Same Ethnicity			0.826				0.846
			(0.501)				(0.505)
Same Ethnicity and Gender			1.074				1.032
			(0.503)				(0.481)
Attractive					0.892	1.393	0.893	1.422
					(0.225)	(0.810)	(0.226)	(0.860)
Reliable					2.203 **	2.021	2.196 **	1.978
					(0.714)	(1.230)	(0.737)	(1.229)
Number of Subjects	98		98		98		98
Observations	556		556		556		556
Log likelihood	–190.9		–190.9		–185.9		–185.8

** p < 0.01. Standard errors clustered at the individual level are presented in parentheses.

. presents the constant parameter estimates, as well as the standard deviation of random parameters. We do not see any heterogeneity across individuals in the extent to which their decisions are affected by the gender or ethnicity of the photos. We confirm the positive impact of Reliable. However, there is no heterogeneity across individuals in their response to Reliable.

From these findings, we conclude that the recipients do not react to ethnicity or gender as a signal of generosity. Recipients, however, seem to be reading something else into the photos: Whether the dictator are deemed reliable or not. The reliability rating of the dictators significantly affects the recipients’ choice. However, dictators who have higher reliability ratings are not more generous. Estimating a simple linear regression of the amount left in the envelope (Amount Left) on Reliable (see

Table 4. Linear regression results.

DV: Amount Left
Reliable	0.550 (1.220)
Constant	4.504 *** (0.690)
Number of Subjects	22

*** p < 0.001.

), we do not find any significant effect. Thus, although recipients pick the envelopes with dictators who seem more reliable, these dictators are not more generous.

5. Conclusions

The literature finds that dictators are attentive to their recipient’s photograph when deciding an allocation, but we find little evidence that recipients use gender and ethnicity information (embodied in a photograph) when choosing among dictators. We expected that recipients would be more likely to pick people similar to themselves. This is not the case for either gender or race/ethnicity. It is also the case that people do not use these obvious signals in making their choice—that is, we find no overall preference for individuals of a specific gender or ethnicity. However, recipients do rely on something they read into the photographs. As distilled from raters of the photos, we find that a combination of items relating to an assessment of the reliability of the image in the photo matters. This combination of factors (survey questions evaluating the extent to which the photos appear motivated, trusting, trustworthy, hardworking, intelligent, honest, respectful, friendly, happy, accepting, and dependable) may constitute a kind of heuristic for selecting partners who are more likely to be cooperative. Along a similar line, Eisenbruch et al. [30] find that, in the ultimatum game, subjects use heuristics, such as the appearance of strength, kindness, productivity, health and social status, inferred from photographs of their counterparts, in deciding how much to offer in the game.

We do not find that such a signal translates into economic benefits. Recipients make no more in their choice of an image that is viewed as “reliable” than if they had picked the less reliable in the pair. This result may indicate that subjects are unable to infer true reliability from the photos, or it may be merely an artifact of the experimental design in which only one side of the transaction views the photos. In our study subjects who were dictators did not know who their potential recipients might be. A hint at what might be going on can be seen in the results reported in Eckel and Petrie [22]. In their study, subjects can pay to view a photo of their counterpart. Many subjects purchase the photos, but doing so does not increase their earnings on average; subjects are not able to infer trustworthiness of a partner just by looking. Earnings of the pair (trustor and trustee) are only higher when both parties have decided to view the other’s photo—when they are face-to-face. This hints at the possibility that accurate inferences may depend, not on the photo per se, but on the pairing. Future research may discover whether “reliability” is an honest signal of cooperativeness when both players see each other’s faces.