Risk behind the veil of ambiguity: Decision‐making under social and nonsocial sources of uncertainty

Abstract

Research has long documented how decision‐making in risky environments differs between environments where the probabilities of uncertain outcomes are known and where the probabilities are unknown, the latter often referred to as “ambiguous” environments. Yet, there is a dearth of research examining how decisions may be affected by the source responsible for the distribution of uncertain outcomes. The source responsible for generating distributions of uncertain outcomes may be generated by another person (i.e., is social in nature) or by a nonsocial probabilistic mechanism. While a few studies examine how the source responsible for uncertain outcomes affects decisions when probabilities are known, the present study extends prior research to the realm of ambiguity by testing how the source of uncertainty affects both decisions when probabilities are fully known and when probabilities are partially unknown using a within‐subjects experimental design. We calculate a general measurement of Social Risk Sensitivity to capture how individuals differ in their sensitivity across three uncertainty environments: risk with no ambiguity, risk with low ambiguity, and risk with high ambiguity. We find evidence showing strong correlations between Social Risk Sensitivity across all three levels of ambiguity. Our results corroborate the previous literature regarding ambiguity effects on decision‐making behavior and extend prior work for the first time in this area by demonstrating that individual decisions are shaped by their individual sensitivity to the source from which uncertainty is derived.

1. INTRODUCTION

The advance of artificial intelligence provides firms and individuals with increasing opportunities to invest in nonhuman technologies to automate at least some tasks that might otherwise be performed by a human. This type of tradeoff is not new. For instance, in the past, farmers had to choose between hiring labor to milk cows or investing in automatic milking technology, which involved different risks (Sauer & Zilberman, 2012). Today, individuals face many such choices, from using content generation services like ChatGPT to choosing between hiring a financial planner or using a robo‐advisor (Rühr et al., 2019). Thus, these sorts of decisions will continue to be commonplace both in the workplace and at home. For example, managers in charge of purchasing company vehicles may need to weigh the costs and benefits of standard vehicles, artificial intelligence‐driver‐assisted features, the potential for future fully automated driverless vehicles, or the option to forego the purchase of new vehicles altogether and the provision for eligible employees to use a subscription service instead (e.g., Abraham et al., 2017; Schoettle & Sivak, 2015). In all these scenarios, individuals must choose between investing in an uncertain environment where the outcome probabilities are sourced from the actions of other people (such as a financial planner) or from nonhuman actors (such as robo‐advisors). Investigating such scenarios requires asking whether individual preferences over different sources (social and nonsocial) differ, and, if so, how these preferences differ across different types of uncertainty (such as between uncertainty with known vs. unknown probabilities of outcomes). These latter questions, until now, have been only sparsely explored in the literature. Our study aims to fill this gap.

Understanding human preferences for the uncertainty source is essential for risk and uncertainty management. For example, there is evidence that a multitude of factors, including personal experience with extreme weather (a nonsocial source of uncertainty), inform individuals’ climate change risk perceptions (Steynor et al., 2021). Given the differing narratives that surround climate change, it is plausible that investment in preferred interventions will depend in part on individuals’ preferences for social sources of uncertainty. For instance, individuals who prefer socially sourced uncertainty over nonsocially sourced uncertainty may prefer to invest in climate change mitigation techniques that depend on collective action and messaging, while those who prefer nonsocial uncertainty may prefer to invest in monitoring technologies and technological interventions. There is also evidence that humans who have high general trust perceive fewer risks from COVID‐19 than those with low general trust; however, this finding is flipped when comparing individuals with high social trust to individuals with low social trust (Siegrist et al., 2021). This implies that individuals’ preferences for uncertainty source may also play a role in their investments in pandemic prevention measures. It may be that individuals who prefer nonsocially sourced uncertainty may prefer to invest in mitigation techniques that do not rely upon the choices of others (such as vaccines to protect oneself), while individuals who prefer socially sourced uncertainty may prefer to invest in techniques that involve others generating the uncertainty (such as the widespread provision of masks). To understand whether preferences between social and nonsocial sources of uncertainty affect such types of decision‐making, we must first establish whether these preferences exist in both risky environments without ambiguity (e.g., where the probabilities of outcomes are fully known) and risky environments with ambiguity (e.g., environments where the probabilities of outcomes are at least partially unknown) as we do in this study.

Decisions under uncertainty are influenced by the consequences of potential choices. A large body of research (see Krain et al., 2006 for a meta‐analysis and our deeper background discussion in the next section) has focused on how types of uncertainty influence choices (namely, the role of ambiguity in decision‐making). However, there is a dearth of research examining how the source that determines the probabilistic outcomes of uncertainty affects decision‐making. By “determines,” we do not mean that a particular source (social or nonsocial agent) estimates or calculates what the probabilities are, but instead mean that these are the actual source of the presented probabilities. For instance, in social decisions to trust, the source of uncertainty is another person who may or may not reciprocate investments. A customer service representative who provides high‐quality service 70% of the time and low‐quality service 30% of the time has determined the probability distribution by the distribution of their decisions. This situation can be distinguished from other decisions in which the source of uncertainty is a nonsocial mechanism, such as the role of a six‐sided die. A “fair” die would be the source of a nonsocial probability distribution.

Lauharatanahirun et al. (2012) (henceforth referred to as LCK) show that individuals do respond differently based upon the source of risk and that individuals differ in their Social Risk Sensitivity (SRS) in risk environments with no ambiguity. There are several critical questions that remain unanswered: (1) does SRS carry over into risk environments that contain varying degrees of ambiguity?; (2) do individuals have a stable SRS or does an individuals’ SRS vary as the degree of ambiguity changes?; and (3) if SRS differs with various degrees of ambiguity, are such sensitivities correlated? Our study explores these important questions by investigating how individual preferences (as defined by their investment decisions) vary as a function of both the type of uncertainty environment (using three different types of uncertainty domains: risk with no ambiguity, risk with low ambiguity, and risk with high ambiguity) and source of uncertainty (where uncertainty stems from either a social or nonsocial source), while avoiding possible confounds from participants’ other social preferences. We discuss the framework of and the related literature surrounding both the type and source of uncertainty in the next section.

While in the past, these two lines of literature have been examined relatively independently, our experimental design uses a within‐participants 3 × 2 (degree of ambiguity condition × source of uncertainty condition) design to investigate differences in investment decisions under risk with three levels of ambiguity with both social and nonsocial sources of uncertainty. Our analyzed sample comprises data from 96 participants made across 132 trials. One might expect decision‐making behavior on average to be similar across social and nonsocial sources of uncertainty in risk (as it was in LCK) and expect this to extend to situations where ambiguity is present as well. We construct a measure of SRS (as done in LCK) within each uncertainty environment, calculated as the difference in the percentage of times a subject invests under a social uncertainty source, less the percentage of times the subject invests under a nonsocial uncertainty source. This yields three measures of SRS: SRS‐No Ambiguity; SRS‐Low Ambiguity; and SRS‐High Ambiguity.¹

Our work in this article is distinguished from the literature on trust and risk in that while our design incorporates a social source of uncertainty, it does not constitute a trust environment as the probabilities are pre‐existing and not made as reciprocity to the decision‐maker. These probabilities are ex‐ante fully known (in the case of the risk‐only environment) or partially known (in the case of the ambiguity including environments). Further, participants know that, even though they are matched with human and nonhuman sources of uncertainty, their decisions affect only themselves and thus they do not have social partners, thus eliminating the potential introduction of other social preferences like inequality aversion or betrayal aversion.

We find that SRS extends into uncertainty domains with ambiguity and that knowing an individual's SRS in one uncertainty domain does not perfectly predict what their sensitivity will be in another uncertainty domain with a different level of ambiguity. However, there are strong correlations observed between SRS across levels of ambiguity. In other words, taken together, SRS appears relevant across uncertainty domain types but is dependent upon the level of ambiguity. The remainder of the article is organized as follows. Section 2.1 introduces and describes the framework of uncertainty we will be using in our study followed by background literature of the two elements of uncertainty most closely related to our study, differences between risk and ambiguity (Section 2.2), and differences between social and nonsocial sources of uncertainty (Section 2.3). Section 3 describes our experimental design, procedures, hypotheses, and methodology. Section 4 gives the results of our statistical analyses, and Section 5 discusses the article's findings, highlights our contributions, and concludes the article.

2. BACKGROUND

2.1. General framework of risk and ambiguity in uncertainty

Uncertainty enters decision‐making environments in multiple ways and can take many different forms. The term “risk” is often used interchangeably with “uncertainty” to describe the uncertainty present in a variety of environments, ranging from the risk of severe weather to wars to financial markets to gambling. The extensive variety of these situations of uncertainty means that these situations vary along many different dimensions such as the valence/size of the outcomes and associated probabilities, whether the probabilities and outcomes are known or unknown, the source of uncertainty (or where the probabilities come from) itself, whether the outcomes affect one or more people, whether an individual or a group makes decisions, and much more. As we discuss in this article, and as past studies have also shown, decision‐making under uncertainty is highly dependent upon these different dimensions. In the current study, we focus on two dimensions. We explore how decision‐makers are differentially sensitive to the source of uncertainty (whether social or nonsocial) and how that sensitivity differs when underlying probabilities of outcomes are known or unknown. Since our questions are focused upon these two dimensions, we leave the considerations of other dimensions of uncertainty (such as temporal differences, divisions of outcomes, etc.) to future work, and form a conceptualization of uncertainty and an experimental design that centers around the two dimensions relevant to the question at hand.

Figure 1 illustrates the conceptualization of uncertainty that we will use in this article. This framework separates uncertainty situations into either situations containing only risk or situations containing risk with some degree of ambiguity, as frequently done within the field of economics. Frank Knight argued that risk constitutes “a measurable uncertainty,” which is distinct from other types of uncertainty (Knight, 1921, p. 21). Ellsberg (1961) later summarized Knight's definition of risk as uncertainty, where the underlying probability distribution of potential outcomes is known ex‐ante. Ellsberg (1961) referred to Knight's “uncertainty with unknown probabilities” as ambiguity. In this framework, situations where a decision‐maker has little experience or little knowledge of the probability of outcomes occurring would find themself in the type of uncertainty we call “risk with ambiguity” and would be on the left side of Figure 1. Situations where the decision‐maker has formed a well‐defined belief of the probabilities of outcomes (objectively known, like a fair six‐sided die or due to an experience of a sufficient number of past observations of outcomes from the same uncertainty environment) would find themselves in the type of uncertainty environment we call “risk only.” While we use the word “risk” in this particular way in this article, we note that we are not limiting the scope of what is traditionally studied in the risk analysis research community which would encompass environments of both risk and ambiguity and delve into other rich dimensions of uncertainty that we do not tackle in this article, such as the size of the stakes and whether those stakes are in the gain or loss domain (e.g., like Kahneman & Tversky, 1979).

FIGURE 1.

Differing sources of uncertainty across uncertainty environments.

Note: This figure was generated using Microsoft Office.

The gap we seek to fill with this article is predominantly associated with the bottom level of Figure 1. We particularly care about understanding how differences in the source of uncertainty affect decisions. In other words, we address how differences pertaining to where the probabilities of potential outcomes arise from affect ultimate decisions. We further characterize the source of uncertainty into two distinct types: social sources (where a human or humans determine the probabilities) and nonsocial sources (where a nonhuman determines the probabilities). Socially sourced uncertainty environments would include decision‐making situations dependent upon human actors, like facing the possibility of terrorist attacks, diplomacy, or even a game of poker. Nonsocially sourced uncertainty environments would include decision‐making environments where the relevant uncertainty comes from factors like weather or a roulette wheel. These two different sources of uncertainty could underlie environments where probabilities are well‐known to the decision‐maker (risk without ambiguity) and environments where probabilities are not well‐known (risk with ambiguity). There is no guarantee that there is symmetry in how individuals perceive and are affected by the source of uncertainty between risk environments containing or not containing ambiguity. Therein lies the heart of our exploration in this study.

2.2. Background on differences between risk and ambiguity in uncertainty

A healthy past literature has explored uncertainty in both environments where the uncertainty probabilities are known (often called “risk” in studies like Eckel & Grossman, 2006; Holt & Laury, 2002), and ambiguity environments, where the probabilities are not wholly known (e.g., Binmore et al., 2012; Trautmann et al., 2008; Fox & Weber, 2002). Studies in this literature have shown evidence that individuals differ when making decisions in risk versus ambiguity environments (Blankenstein et al., 2017; Borghans et al., 2009; Chakravarty & Roy, 2009; Fujino et al., 2017; Halevy, 2007; Huettel et al., 2006; Krain et al., 2006; Moreno & Rosokha, 2016; Smith et al., 2002). People tend to exhibit both risk‐averse and ambiguity‐averse preferences (e.g., Chakravarty & Roy, 2009). It is intuitive to assume that individuals who exhibit risk‐averse preferences are also likely to exhibit ambiguity‐averse preferences. Further, there is evidence that neural processing differences exist between risky and ambiguous environments, though there is some overlap (Blankenstein et al., 2017). A large body of evidence also demonstrates that people make different decisions in risk and ambiguity environments especially when the outcomes of such choices affect one's self only (nonsocial environments) or others instead of or alongside oneself (social environments—for examples, see: Aimone & Pan, 2020; Evans & Krueger, 2017; Li et al., 2020; Andersson et al., 2016; Beisswanger et al., 2003; Chakravarty et al., 2011; FeldmanHall et al., 2015; König‐Kersting & Trautmann, 2016; Pahlke et al., 2012, 2015; Stone & Allgaier, 2008; Trump et al., 2015; Vieider et al., 2016; Zhang et al., 2017).² However, few studies have provided empirical support showing that individuals are internally consistent across uncertainty environments. This is important to understand and can have potential implications for designing public health communications and policies.

2.3. Background on social and nonsocial sources of uncertainty

Prior work has focused on whether outcomes of choices made under risk and ambiguity affect oneself and/or others (e.g., Füllbrunn and Luhan, 2020; Polman & Wu, 2020). Social preferences (inequity aversion, altruism, etc.) that deal with how payments and outcomes are divided between participants are thus factors that can influence decisions in such social environments. However, research on whether people consider where the underlying probabilities come from has been largely understudied. In everyday life, some uncertainty is generated by random chance (not connected to humans), which we call “nonsocial uncertainty,” and some by the actions of humans, which we call “social uncertainty.” Whether the source of uncertainty is social or nonsocial does not implicitly determine whether the outcomes of the uncertainty are social or nonsocial (and thus whether common social preferences like inequity aversion might also influence decisions). Consider a few examples. In an iterative trust game (e.g., Berg et al., 1995), an investor engages in a repeated social exchange with an anonymous partner. Within this game, the investor does not have a well‐established prior of their partner's likelihood of repayment, which is akin to an ambiguous uncertainty environment, and the decisions that either the investor or the trustee make affect the payoffs for both parties. Moreover, the human trustee's unknown propensity to reciprocate or betray the source is an example of a social source of uncertainty. On the other hand, consider a parent's decision to drive their child to school with the roof of their convertible down when the sky is threatening rain. This is also an ambiguous uncertainty environment (if the parent does not know the exact probabilities of rain), and the parent's decision to drive with the roof down has potential consequences for both the parent and child (i.e., exposure to rain for both the parent and child). However, in this environment, the source of uncertainty is nonsocial, as the outcome (whether it rains or not) is not coming from a social (e.g., human) source. In both examples, the uncertainty environment was ambiguous, and the decision‐maker's choices affected another person. The key feature we are interested in that differed between the two examples is the source of uncertainty, which was either social (trust game) or nonsocial (driving with the roof down and potential for rain).

Some existing research has shown that people are sensitive to the source of uncertainty within risk (known probability) environments that lack ambiguity. For instance, LCK used a within‐subjects design to investigate differences in decision‐making behavior in socially risky environments (a modified trust game) and nonsocially risky environments (lotteries) while undergoing functional magnetic resonance imaging. In this study, they estimated an individual's risk attitude using a constant relative risk aversion utility function when outcomes were determined by another person (social source) and when outcomes were determined by a lottery (nonsocial source). They quantified an individual's SRS as the difference between estimated risk aversion in response to social versus nonsocial sources. This study showed that individuals differed in the degree to which they were sensitive to social and nonsocial sources of risk and that activation of decision‐making‐related brain networks was congruent with these individual sensitivities. Similarly, another study conducted by Fetchenhauer and Dunning (2012) found that in a comparison between a lottery and trust‐game with matched probabilities of gain, participants were highly sensitive to the probability of reward in the lottery condition, but not the trust‐game. Taken together, these findings suggest that the source of uncertainty indeed impacts decisions made within risk environments. In addition to decision‐making behavior, there is also research showing that the source of uncertainty affects information search behavior prior to deciding. For instance, Fleischhut and colleagues (2022) showed that individuals spend less time searching for information in social risk environments relative to nonsocial risk environments.

While an increasing number of studies examine how the source of uncertainty affects decisions in risk environments, limited attention has focused on environments including ambiguity. A recent study by Fairley and colleagues (2022) showed that within ambiguity environments, people tend to show greater ambiguity aversion when they believe that the probability of reciprocity is higher. This study suggests and very nicely provides evidence that people may be sensitive to the source of uncertainty in ambiguity environments. Their design used four environments with social outcomes (multiple human participants were affected by each uncertainty decision), which nicely controls for outcome‐based social preference considerations, like inequity aversion. Fairley et al. mention³ betrayal aversion as a social factor that may intensify ambiguity aversion in social interactive contexts (socially sourced uncertainty) compared to lottery contexts (nonsocially sourced uncertainty). It thus remains an open question whether sensitivity to the source of ambiguity remains in the absence of these additional social factors when decisions do not affect others. Our study continues this important line of work by uniquely exploring how investment decisions may change as a function of risk and ambiguity, uncertainty environments, and sources of uncertainty within situations where the decision‐maker's choices do not impact other people. This is particularly important because social preferences (such as betrayal aversion) have been shown to influence decision‐making behavior when choices affect not only the decision‐maker but others as well. Since our design involves individuals making choices where consequences are isolated to the individual, we remove the potential for social preferences like betrayal aversion to influence investment behavior.

Another related line of prior work is the literature on the relationship between trust and risk. The evidence for a link between trust and risk is mixed. Some articles, mainly those arising out of Eckel and Wilson (2004), generally find that risk and trust preferences are separate preferences (Blais & Weber, 2006; Fairley et al., 2016; Farjam, 2019; Houser et al., 2010). Others argue that risk and trust preferences are linked (Chetty et al., 2021).⁴ The literature on betrayal aversion has attempted to explain some of these differences. Betrayal aversion focuses upon the importance of the source of uncertainty within trust situations, and how the outcome is specifically directed at the recipient, for example, betrayals or reciprocations (Aimone & Houser, 2012; Bohnet & Zeckhauser, 2004; Bohnet et al., 2008, 2010). In addition, it is also possible that ambiguity attitudes may explain some findings from the trust game: Li et al. (2020) find no evidence of betrayal aversion in trust games but do find that social ambiguity explains motivational behavior in trust games, and Li et al. (2019) find that both people's beliefs and ambiguity preferences affect behavior in trust games. While this work has not specifically focused in on the difference between social and nonsocially sourced uncertainty, this work motivates the importance of considering the source of uncertainty as well.

3. METHODOLOGY

3.1. Participants

A total of 103 participants (Female = 31.25%), ages 18−66 (M = 33.86; SD = 9.15), were recruited to complete the present study (see Table 1) via the Amazon Mechanical Turk platform. Amazon Mechanical Turk is an online crowdsourcing system where anonymous human workers (also known as MTurk Workers) can complete a wide array of tasks for money. Prior research has shown that human responses on cognitive research tasks collected from the Amazon Mechanical Turk platform are similar to human responses collected in the laboratory, and thus is a valid methodology for scientific research (e.g., Buchheit et al., 2019; Hauser and Schwartz, 2016; Crump et al., 2013). To ensure higher data quality, we included eight attention checks throughout the online experiment (see Procedure for attention check details). Participants were excluded if they failed more than 25% of the eight attention checks (N = 0), experienced technical difficulties during the experiment (N = 4), or made the same decision in all trials of the experiment (N = 3). We also excluded any trials in which participants failed to decide. After these exclusions, our final sample consisted of a total of 96 participants. All methods and procedures in the present study were approved by the Institutional Review Board at Carnegie Mellon University and carried out in accordance with this approved research protocol. All participants in this study provided implied informed consent to participate in this approved research protocol.

TABLE 1.

Sample characteristics. ^a

Personality	Mean
Average score IUS	39.53
Average score ITS	78.36
Demographics	Percent
Female	31.25%
Asian	13.54%
Black or African American	16.67%
Other race or ethnicity	9.38%
White	60.42%
Resides in the United States	89.58%
Education	Percent
High school diploma or GED	6.25%
Some college	7.29%
Associates	4.17%
Bachelors	61.46%
Graduate degree (master's, professional, doctorate)	20.83%
Income before taxes (yearly)	Percent
Below $35k	44.79%
35k−50k	16.67%
50k−70k	20.83%
Above 70k	12.50%
Not reported	5.21%
Age group	Percent
18–24	5.21%
25–34	48.96%
35–44	36.46%
Age group	Percent
45–54	5.21%
55 or older	3.13%
Unknown	1.03%

^a The Intolerance of Uncertainty Scale Score (IUS score) is scored out of 60; the Interpersonal Trust Scale Score (ITS Score) is scored out of 100. Age was recorded as a continuous variable but is reported as a discrete variable here due to the existence of outliers.

3.2. Procedure

Participants completed this online study remotely using their own personal electronic devices. This online study was listed as a Human Intelligence Task (HIT) on the Amazon Mechanical Turk platform. To be eligible to complete this HIT, MTurk workers were at least 18 years of age, had not previously completed this HIT, and were residents of the United States. After selecting to complete the HIT, MTurk workers were then provided with an online consent form. If the MTurk worker provided consent, they were officially enrolled in the study. Participants were then given a link to complete all the study components, which included the economic lottery choice task (described below), demographics survey, and personality questionnaires. The experiment was coded in JavaScript and presented via a web interface. Prior to the experiment, participants completed a demographic questionnaire that included questions about gender, race and ethnicity, education level, age, gross income, and whether the subject was a U.S. resident. At the beginning of the experiment, participants read a set of instructions (see Supplementary Appendix B). Participants were instructed that they could receive an additional bonus payment of $1 based on the winnings earned from one randomly selected trial. Next, participants completed the Social Uncertainty Lottery Choice Task. If a subject failed to make a decision in each trial, they were given another opportunity to decide in an identical trial later in the experiment, ensuring that each subject completed trials in all experimental conditions. Throughout the online experiment, participants completed a total of eight attention checks. Participants were asked to copy and paste a series of letters and numbers into a blank field. The attention checks appeared randomly, and there were at least 15 trials between any two attention checks. All participants scored at least 75% or higher on average across the eight attention checks. Following the experiment, participants completed the Intolerance of Uncertainty Scale (hereafter, “IUS”: Carleton et al., 2007) and Interpersonal Trust Scale (hereafter, “ITS”: Rotter, 1967). Higher scores on the IUS denote higher uncertainty intolerance, and higher scores on the ITS denote lower trust. We include these two personality measures to assess their relationship with investment decisions in our study. After completing all study components, participants were provided with a code to input into MTurk to receive compensation, which was electronically disbursed by Amazon. One randomly selected trial was chosen, and participants were compensated based on the outcome of their decision made on that trial. Points were translated to monetary outcomes (e.g., 1.50/22*1), where participants could earn a maximum of $1 in bonus points.

3.3. Social uncertainty lottery choice task

In this task, participants were instructed that their goal was to earn as many points as possible. The Social Uncertainty Lottery Choice Task was adapted from a single‐shot unidirectional trust style game (related to Berg et al., 1995; Lauharatanahirun et al., 2012). All participants were assigned to the role of the investor. As the investor, participants were asked to make decisions to either keep an initial endowment that ranged between 5 points and 15 points or invest their endowment in another player who had made a “return” decision in a prior experiment (social source condition) or a gamble (nonsocial source condition), in which the goal was to earn as many points as possible. For each gamble, there was a high and low monetary outcome, each associated with a specific probability. The associations between outcomes and probabilities are represented with corresponding colors (blue and yellow, see Figure 2). Probabilities associated with each potential payoff are represented visually with pie charts. There were 10 slices in each pie, each corresponding to a probability of 10%. Participants were presented with slices representing the probability of receiving a high (yellow) or low (blue) outcome (see Figure 2).

FIGURE 2.

Social uncertainty lottery choice task.

Note: This task was conducted within‐subjects. Each subject completed an equal number of trials within each dimension of the experiment. The Risk with 30% Ambiguity environment and Risk with 60% Ambiguity environment are denoted in the remainder of the article as the “Risk with Low” and “Risk with High” Ambiguity uncertainty environments. This figure is a schematic that illustrates the different choice environments that participants were exposed to within the task. Face stimuli representing prior participants were from the Chicago Face Database (Ma et al., 2015). This figure was generated using Adobe Illustrator software.

Participants could either invest their endowment in another person (social source of uncertainty) or a nonhuman randomization device (such as a roulette wheel; a nonsocial source of uncertainty; see Figure 2). The face images used in the social source of uncertainty condition were standardized male and female faces from a diverse array of ethnicities from the Chicago Face Database (Ma et al., 2015) that were rated at the midpoint of a trustworthiness scale to attenuate potential trustworthiness priming from facial stimuli. While these face images were not the actual photos of participants from the previous study, we did use the actual payoff distributions of these participants from LCK. Since we wanted to be very cautious about not deceiving our participants, we were careful to use precise language and did not tell them that the picture was the picture of the participants. Participants were told in the instructions (see the Supplementary Appendix) that “Each picture and pie chart represent responses of a real person that participated in this study previously.”⁵

For the social source of uncertainty, participants were matched with the distribution of reciprocation decisions from a real participant from LCK. Participants were matched with a lottery task (with identical probabilities and payoffs to the modified trust game) for the nonsocial source of uncertainty. In the “Risk with No Ambiguity” uncertainty environment condition, the exact probabilities of a given outcome were shown to the participant, and there was no occlusion of the pie. Extending on LCK, an additional two conditions were added, representing low and high ambiguity (see Figure 2). In the ambiguity conditions, the probabilities of each outcome were partially known to the participant. Ambiguity was represented in the task by the percentage of occlusion or “white slices” of the pie. The “Risk with Low Ambiguity” and “Risk with High Ambiguity” conditions were represented by a 30% (three white slices) and 60% (six white slices) occlusion of the pie.

The participants did not know how much of the occluded probabilities were split between the high and low values. Functionally, the occluded probabilities were always split equally between the high and low value pieces. The probabilities and outcomes for each trial were identical to those used in LCK. Each participant completed a total of 132 trials, yielding 66 trials in each of the social and nonsocial source conditions. In each of these 66 trials, there were 22 Risk with No Ambiguity trials, 22 Risk with Low Ambiguity trials, and 22 Risk with High Ambiguity trials. Values and associated probabilities were matched across the six conditions (social risk with no ambiguity, nonsocial risk with no ambiguity, social risk with low ambiguity, nonsocial risk with low ambiguity, social risk with high ambiguity, nonsocial risk with high ambiguity), and their order was randomly varied throughout the task. Specifically, there were 11 unique gambles that were paired with each endowment level, ranging from 5 to 15 points. Each condition had the same 11 gambles that were each shown twice, with their order pseudo‐randomized within each condition, so the same gamble would not appear successively.

3.4. Experimental design and hypotheses

This study was a within‐subjects 3 (uncertainty environment: risk with no ambiguity, risk with low ambiguity, risk with high ambiguity) X 2 (source of uncertainty: social, nonsocial) experimental design (see Figure 2). The dependent measure was the type of investment decision made by the participant (i.e., whether they chose to keep their initial endowment or invest in a gamble). This experimental design allowed us to address three primary hypotheses that were focused on better understanding the relationship between participants’ preferences between social and nonsocial sources of uncertainty within and between different types of uncertainty environments. We know from prior work that participants have preferences over the source of uncertainty in the risk domain when there is no ambiguity (Lauharatanahirun et al., 2012); we would expect this type of preference structure to extend into the ambiguity domains and see SRS in risk domains that include ambiguity as well. This leads to Hypothesis 1.

Hypothesis 1

Participants will have preferences over the source of uncertainty in the ambiguity uncertainty environments.

Knowing whether SRS is present in both risk domains lacking or including ambiguity is important as in any situation of uncertainty, be it due to a pandemic, political unrest, financial upheaval, business troubles, or some other cause. Individuals subject to that uncertainty will likely shift from perceiving the environment as one with a lot of ambiguity to relatively little ambiguity as their exposure to the environment provides them with better defined perceptions of the probabilities of different outcomes. Thus, people will likely be moving from an uncertainty environment of high ambiguity to an environment of little to no ambiguity. Whether SRS is only important in a risk environment lacking ambiguity or is also important in ambiguity environments is important to know. If there is indeed SRS in both risk domains lacking or containing ambiguity, it is not necessarily the case that one's preferences in one domain are predictive of their preferences in another domain. We expect individuals to be similarly sensitive to the source of uncertainty across domains, which leads to Hypothesis 2.

Hypothesis 2

Participants’ preferences for uncertainty source will be correlated across ambiguity domains.

Knowing whether SRS is correlated across various ambiguity domains (as explored in Hypothesis 2) may be critically important for policy makers, aid providers, and support agencies to understand. In particular, if there is no correlation, the way one communicates about the uncertainty may need to be modified or changed as individuals become more educated or experienced with the environment. Understanding such correlations (or lack thereof) can help ease the transition and improve outcomes.

Simply knowing whether SRS is correlated across ambiguity domains is not sufficient, however. The degree of correlation and directionality of correlation are also critically important for understanding how preferences differ across ambiguity domains. If the correlation is near perfect, then measuring SRS initially at the beginning of a situation of uncertainty, when there is a high degree of ambiguity, is perhaps sufficient, as it can be assumed that the level of SRS will persist as the degree of ambiguity decreases over time. If the correlation is of low magnitude (e.g., not perfectly correlated), then there are remaining important differences in SRS across ambiguity domains that would need to be considered. If such a situation exists, it may continue to be important for decision‐makers, aid groups, or policy makers to elicit or measure SRS in an uncertainty situation as time progresses on, if they want to have a good idea of what is leading to differential decision‐making differences between agents. With this important distinction, between perfect and nonperfect correlation, in mind, we make our third testable hypothesis:

Hypothesis 3

Correlation in social risk sensitivity between ambiguity domains will be imperfect, in other words, there will be heterogeneity within a participant's preferences for socially sourced uncertainty as opposed to nonsocially sourced uncertainty across all three ambiguity domains.

4. RESULTS

As our main goal of this article is to explore differential sensitivities to social and nonsocial sources of uncertainty, it is important to establish whether our experiment demonstrates standard behavioral differences in tolerance of uncertainty based upon the type of uncertainty environment (i.e., risk with no ambiguity, risk with low ambiguity, or risk with high ambiguity). We begin by examining investment behavior on average across the three ambiguity domains and then systematically explore our main questions of individual heterogeneity to social and nonsocial sources of uncertainty. Finally, we explore how demographic and personality measures are related to investment behavior.

Figure 3 gives the mean investment rate in each ambiguity domain (see the Supplementary Appendix A for additional descriptive statistics). As expected, participants were significantly less likely to invest in the Risk with Low Ambiguity or Risk with High Ambiguity uncertainty environment conditions than in the Risk with No Ambiguity uncertainty environment conditions. Participants’ investment decisions also decreased as the level of ambiguity increased (i.e., 30% or 60%). Wilcoxon Sign‐Rank two‐tailed tests provide evidence that participants invested at significantly different rates between the Risk with No Ambiguity uncertainty environment (62.4% of the time) and both ambiguity uncertainty environments (50.7% of the time in the Risk with Low Ambiguity uncertainty environment and 43.1% of the time in the Risk with High Ambiguity uncertainty environment; p ≤ 0.001 in both cases). There was also a significant difference in investment between the Low and High Ambiguity uncertainty environments (p = 0.002). However, we found no difference in investment behavior between the social and nonsocial uncertainty source conditions overall (p = 0.255), nor between the social and nonsocial sources within any of the uncertainty environment conditions (Risk with No Ambiguity: p = 0.155; Risk with Low Ambiguity: p = 0.478; Risk with High Ambiguity: p = 0.376).

FIGURE 3.

Investment conditional on uncertainty environment and source of uncertainty. Note: Each column gives the percent of times a subject chooses to invest within a given treatment and source of uncertainty. Error bars are standard errors. This figure was generated in Stata version 18.

These results are echoed in a random‐effects logit model of investment behavior across conditions; we report the marginal effects of this model in Table 2. Models (1) and (2) report coefficients from the logit models. Model (1) reports the coefficients of a model with “Invest” as the primary dependent variable (where “Invest” is an indicator variable that takes a value of 1 if a participant invested in a given trial and 0 otherwise), and indicator variables for the uncertainty environment condition and the source condition as explanatory variables, along with interactions between the uncertainty environment and the source of the uncertainty. The nonsocial Risk with No Ambiguity uncertainty environment serves as our reference condition (as we include interaction terms for both ambiguity uncertainty environments with the Social Source condition, we can interpret the standalone Social Source condition indicator as capturing the effect of the social Risk with No Ambiguity uncertainty environment). Model (2) reports the coefficients of a similar model that also includes controls for race, gender, and the participant's Intolerance of Uncertainty measure (“IUS Score”), as well as her Interpersonal Trust Score (“ITS Score”).

TABLE 2.

The effect of source and condition on investment decisions. ^a

Model	(1)	(2)	(3)	(4)
Variables	Invest	Invest	Invest	Invest
Risk with Low Ambiguity Condition	−0.700 ***	−0.700 ***	−0.126 ***	−0.131 ***
	(0.138)	(0.138)	(0.025)	(0.026)
Risk with High Ambiguity Condition	−1.142 ***	−1.142 ***	−0.206 ***	−0.213 ***
	(0.186)	(0.186)	(0.034)	(0.036)
Social Source	−0.003	−0.003	−0.000	−0.000
	(0.075)	(0.075)	(0.013)	(0.014)
Risk with Low Ambiguity X Social Source	−0.012	−0.012	−0.002	−0.002
	(0.086)	(0.086)	(0.016)	(0.016)
Risk with High Ambiguity X Social Source	0.050	0.050	0.009	0.009
	(0.091)	(0.091)	(0.016)	(0.017)
Female		0.484		0.090
		(0.349)		(0.065)
Black or African American		0.441		0.082
		(0.408)		(0.076)
Other Nonwhite		−0.213		−0.040
		(0.266)		(0.049)
IUS Store, Standardized		0.394 ***		0.074 ***
		(0.147)		(0.027)
ITS Score, Standardized		−0.133		−0.025
		(0.132)		(0.025)
Constant	0.806 ***	0.601 ***
	(0.176)	(0.221)
Observations	12,672	12,672	12,672	12,672
Number of Participants	96	96

^a Robust standard errors, clustered by subject, are reported below each coefficient. Model (1) reports the coefficients of a random‐effects logit regression model where the independent variable is an indicator for the decision to invest. Model (2) gives a similar random‐effects logit to Model (1), but with controls for race, gender, IUS, and ITS included. Model (3) reports the marginal effects from Model (1), and Model (4) reports the marginal effects from Model (2). All marginal effects denote the effect of a one standard deviation increase from the mean.

^*** denotes p < 0.01, ^** p < 0.05 and ^* p < 0.1

However, we focus our discussion on Models (3) and (4) of Table 2, which report the marginal effects from Model (1) and Model (2), respectively. Relative to the nonsocial Risk with No Ambiguity uncertainty environment, participants are less likely to invest in both the nonsocial Risk with Low Ambiguity and Risk with High Ambiguity conditions; however, we find no effect of the social source in any condition. This echoes what we find in Figure 3, which indicates that there is no statistically significant difference between the social and nonsocial source conditions at any uncertainty level.

We know from LCK that some individuals have a greater sensitivity to nonsocially sourced risk and others have a greater sensitivity to socially sourced risk. Aggregation would be expected to wash out these differences.⁶ This means we next need to turn to calculating and exploring individual differences in SRS within risk without ambiguity and with ambiguity uncertainty environments. To explore individual differences, we calculate a measure of SRS in each of our three uncertainty environments (Risk with No Ambiguity, Risk with Low Ambiguity, and Risk with High Ambiguity) for each participant. We base this calculation on the “Social Risk Sensitivity” measure in LCK. We calculate SRS as the difference between the proportion of investments made when presented with a social source and a nonsocial source of uncertainty:

$$SRS=PercentInveste{d}_{Social}-PercentInveste{d}_{Nonsocial}$$ (1)

A positive SRS value indicates a relatively greater preference for investment in a social over a nonsocial source (and vice versa for a negative SRS). Given that we have three uncertainty environment conditions, we calculate three different SRS variables. We refer to SRS in the Risk with No Ambiguity uncertainty environment condition as “Social Risk Sensitivity‐No Ambiguity” or “SRS‐NA”; SRS in the Risk with Low Ambiguity uncertainty environment condition as “Social Risk Sensitivity—Low Ambiguity” or “SRS‐LA,” and SRS in the Risk with High Ambiguity uncertainty environment condition as “Social Risk Sensitivity—High Ambiguity” or “SRS‐HA.”

We see the extent of heterogeneity across subjects in Figure 4’s histograms of SRS. Only a minority of participants are strictly indifferent to the source of uncertainty in each (SRS = 0; 17.7% in Risk with No Ambiguity, 21.9% in Risk with Low Ambiguity, and 19.8% in Risk with High Ambiguity). Whether in risk with or without ambiguity environments, most individuals either have an SRS below zero (42.70% in Risk with No Ambiguity, 35.40% in Risk with Low Ambiguity, and 39.60% in Risk with High Ambiguity) or above zero (39.60% in Risk with No Ambiguity, 42.70% in Risk with Low Ambiguity, and 40.60% % in Risk with High Ambiguity). This is primary evidence toward Hypothesis 1, in demonstrating that SRS extends out of risk without ambiguity environments and into risk with ambiguity environments as well.

FIGURE 4.

Distribution of social risk sensitivities.

Result 1

Social Risk Sensitivity exists in risk environments with no ambiguity, low ambiguity, and high ambiguity.

Figure 4 does not, however, provide insight into Hypothesis 2, whether individuals are similarly sensitive to uncertainty source as the degree of ambiguity changes. Table 3 illustrates the intercorrelational relationships between SRS‐NA, SRS‐LA, and SRS‐HA. There are statistically significant (p < 0.001), positive correlations between SRS within each uncertainty environment, indicating that individuals are likely to carry their SRS across different ambiguity domains.

TABLE 3.

Pearson−Watson correlations between social risk sensitivity across ambiguity environments.^a

	Social Risk Sensitivity, No Ambiguity	Social Risk Sensitivity, Low Ambiguity	Social Risk Sensitivity, High Ambiguity
Social Risk Sensitivity, No Ambiguity	1.000 ***
Social Risk Sensitivity, Low Ambiguity	0.376 ***	1.000 ***
Social Risk Sensitivity, High Ambiguity	0.284 **	0.354 ***	1.000 ***

^a This table reports the correlation coefficients of a Pearson−Watson correlation between the Social Risk Sensitivity measures within each level of ambiguity.

^***denotes p < 0.01, **p < 0.05, and *p < 0.1

Result 2

Individuals’ Social Risk Sensitivities are significantly and positively correlated across ambiguity domains.

Note that while there is a highly significant correlation of SRS across ambiguity domains, these correlations are not perfect. The low magnitude of the correlation coefficients suggests that these sensitivities are far from identical, leading to Result 3, connected directly to Hypothesis 3. The variation of SRS across ambiguity domains indicates that SRS is not a general social preference but is ambiguity domain variant.

Result 3

Social Risk Sensitivity in one ambiguity domain is not perfectly correlated with (or reflective of) Social Risk Sensitivity in another ambiguity domain.

The correlations in Table 3 provide evidence that SRS is imperfectly correlated across ambiguity domains but does not incorporate the richness of our within‐subject design. To further investigate these correlations across ambiguity domains, we estimate an OLS regression (Table 4) that takes a given participant's SRS–LA as the dependent variable and includes the other two SRS measures as independent variables. As they are calculated from the participants’ own decisions, it is clear that these two are endogenous and not independent variables, but these regressions are useful as a robustness test. We then explore whether our previously identified correlations appear and whether they are robust to the inclusion of control variables. Table 4 shows our regression results that, in line with the earlier correlation statistics, provide robust support for our results.

TABLE 4.

Correlations between SUS types and controls.^a

	(1)	(2)	(3)
Model	Social Risk Sensitivity,	Social Risk Sensitivity,	Social Risk Sensitivity,
Variables	Low Ambiguity	Low Ambiguity	Low Ambiguity
Social Risk Sensitivity, No Ambiguity	0.338 **	0.344 **
	(0.143)	(0.143)
Social Risk Sensitivity, High Ambiguity	0.285	0.279 *
	(0.175)	(0.160)
Female		−0.028	−0.020
		(0.035)	(0.039)
Black or African American		−0.009	−0.004
		(0.036)	(0.036)
Other Race		−0.106*	−0.094
		(0.063)	(0.080)
IUS Score, Standardized		0.014	0.016
		(0.012)	(0.012)
ITS Score, Standardized		−0.010	−0.017
		(0.014)	(0.018)
Constant	−0.005	0.015	0.013
	(0.014)	(0.020)	(0.022)
Observations	96	96	96
R‐squared	0.208	0.274	0.069

^a Robust standard errors are reported below each coefficient. Column (1) gives the coefficients of an OLS regression of the individual participants’ SRS‐NA and SRS‐HA on the participants’ SRS‐LA. Column (2) estimates the same regression with controls added, while Column (3) estimates only the effects of our control variables.

^***denotes p < 0.01, **p < 0.05, and *p < 0.1

While one's SRS in one ambiguity domain is significantly correlated across ambiguity domains, it is also generally not the same as their SRS in another domain (note the R‐square in Model 1 is only about 0.2). The models in columns 1 and 2 both show that one's SRS is significantly positively correlated with SRS‐LA (p < 0.05 in both models without and with external controls). Similarly, there is a marginally significant positive correlation between

SRS‐HA and SRS‐LA in Model 2, when controls are included.⁷ We highlight that these significant relationships are robust to the addition of control variables, which include variables indicating gender, race, and standardized measures of the participant's IUS (Intolerance of Uncertainty Scale Score) and ITS (Interpersonal Trust Scale) scores. In Model 3, we estimate the pure effects of the control variables, demonstrating that they do not appear to affect an individual's SRS‐LA, giving additional confidence to our results from Model 1.

As discussed in the previous section, the distinction between what is learned in Result 2 and Result 3 is very important, albeit perhaps a bit subtle. Result 2, involving a correlation between with and without ambiguity SRS scores, suggests that sensitivity to social (relative to nonsocial) sources of uncertainty in one ambiguity domain partly explains sensitivity in another ambiguity domain. Table 2 shows that if we see an individual (or a population) having a positive SRS in one domain, say a risk without ambiguity domain, they are likely to also have a positive SRS in a different domain as well, say a low ambiguity domain. However, this result alone does not tell us whether knowing one's SRS in one ambiguity domain is sufficient to predict preferences for social or nonsocial uncertainty in a different ambiguity domain. Result 3, which demonstrates that the degree of correlation is not perfect, tells us something different. Result 3 indicates that SRS is ambiguity domain‐dependent and varies based upon the degree of ambiguity present within the environment.

The importance of what is learned from Results 1, 2, and 3 can be illustrated by a simple thought experiment. Consider an individual who has received news about a newly diagnosed medical condition. This person may receive advice from their doctor (social source of uncertainty), or they may receive the same advice from an AI system (nonsocial source) that has been developed to help people with that condition. Hypothesis 1 (Result 1) indicates that individuals are likely to have sensitivity to the source of the uncertainty (doctor or AI) when the probabilities are objectively known (uncertainty is best described as a “Risk with No Ambiguity” environment) and when the probabilities are little known (when uncertainty is best described as “Risk with High Ambiguity”). However, evidence toward Hypothesis 1 (Result 1) does not tell us whether individuals will be similarly attracted to or away from socially/nonsocially uncertainty sourced environments as the environment shifts from being relatively more or less ambiguous. Evidence for Hypothesis 2 (Result 2) indicates that, if one obtains information about SRS in a domain with one level of ambiguity, this will provide some information about what SRS will be under a different level of ambiguity. For instance, if SRS was elicited under a low ambiguity environment and the individual had a positive SRS‐LA, our Result 2 suggests that SRS‐NA, when probabilities are all objectively known, will be positively correlated with the initial positive SRS‐LA. Result 3, however, goes further in demonstrating that the SRS‐NA in the no ambiguity environment would likely be different, even though it would be correlated with the SRS‐LA in the low ambiguity environment. Result 3 suggests that if one wanted to utilize SRS information for some form of policy or intervention, it would not be sufficient to measure SRS in one domain alone. SRS would likely need to be measured under each type of ambiguity domain.

5. DISCUSSION AND CONCLUSIONS

In this experimental study, we highlight three new main findings about decision‐making in risk environments with and without ambiguity. First, the results illustrate that sensitivity to the source of uncertainty (whether from a social or nonsocial source) influences decision‐making both in risk domains with and without ambiguity. Second, our results provide strong evidence that SRS is correlated between domains with varying degrees of ambiguity. Third, while such sensitivity is correlated across domains, we observe that the degree of sensitivity varies between domains. Thus, one cannot rely upon a measured or observed sensitivity in one domain to make accurate predictions about those individuals’ sensitivities in other domains. These results suggest that the source of uncertainty is critically important for an individual's decision‐making and should be considered when exploring risk environments with any degree or absence of ambiguity. Overall, this work provides an important step toward disentangling the effects of risky decision‐making under social and nonsocial sources of uncertainty and indicates that more research is needed. It is particularly important considering that, intuitively, “real‐world” social situations of risk are likely to be described as containing some degree of ambiguity (probabilities of outcomes are unknown). Past studies, like Lauharatanahirun et al. (2012), provided evidence that social uncertainty sensitivity was important in the risk domain when there was no ambiguity. Our study provides needed evidence and further suggests that we cannot rely upon data in a risk domain without ambiguity to understand behavior in risk domains that vary in their levels of ambiguity.

As is the case with every study, our study is not without its limitations. For example, like all laboratory, field, and online experiments, we cannot completely control participants’ intrinsic beliefs about some factors pertinent to the experiment, such as default probability reference points, personal attitudes toward gambling, or attitudes toward gambling devices. Nonetheless, we took steps to mitigate inattention, facilitate similar priors across subjects, and ensure that participants understood the task by providing detailed instructions with clear examples. Future studies would benefit from including additional measures that assess participants’ beliefs about the experiment as well as replicating the current study within the laboratory and within the field. Additionally, more research is needed to explore how other subjective factors like prior beliefs, gender biases, AI biases, the complication of the risk environment, and so on might affect individuals’ SRS; we highlight some particularly promising potential avenues below.

An important area of future research may involve understanding what the predictors of SRS are and what explains the between‐participant heterogeneity, because it may be important to understand the makeup of such sensitivities in different populations. “Nudge”‐based (Thaler & Sunstein, 2009) policy interventions are likely to intentionally place decision‐makers into either an environment of socially sourced uncertainty or nonsocially sourced uncertainty to make these decisions feel more or less social (consider computer algorithm‐mediated healthcare). Our present study indicates that such efforts may change the fundamental decision‐making processes underlying such decisions. For instance, a nudge involving the usage of social accountability to enhance addiction recovery (social sourced uncertainty) may have no effect (or backfire) for individuals whose SRS is negative (which implies they prefer nonsocially sourced risk). However, the same nudge may have a strong effect in the intended direction for individuals who prefer socially sourced risk (positive SRS). Other policy areas where our findings could be built upon include areas such as regulation of drivers. The act of operating a motor vehicle on a public roadway is an exercise in both socially sourced risk with potentially high levels of ambiguity (from the actions of other drivers) and nonsocially sourced risk with potentially high levels of ambiguity from weather, animals in the road, or automobile malfunctions).⁸ We leave to future research to explore how new policies about automobile design or road engineering can take these differences into account based upon the characteristics of drivers and local conditions. Our study shows the importance of understanding the individual SRSs under different levels of ambiguity within a given population.

Another promising avenue of future research would be to investigate the possible links between SRS and expressive trust (as outlined in Dunning et al., 2014) or betrayal aversion. It may be that betrayal aversion is one channel through which SRS influences decision‐making in ambiguous social environments, particularly given that betrayal aversion has both negative and positive valences (Aimone et al., 2015). Another potentially promising avenue of research could examine whether SRS is related to differential tolerance to occupations and investments. Different jobs (and different investment opportunities) expose workers (and investors) to different types of uncertainties. Consider an agricultural company, such as a farm. The day‐in‐day‐out uncertainties faced by workers and investors center largely around nonsocially sourced risks (with varying degrees of ambiguity) like drought, machinery hazards, and so on, and perhaps relatively few social uncertainties compared to other jobs (businesses) like education. Though we cannot answer those questions here, our work provides a foundation for future researchers to investigate these questions.

CONFLICT OF INTEREST STATEMENT

The authors declare no competing financial interests.

Supporting information

Supporting Information

Lauharatanahirun, N. , Aimone, J. A. , & Gately, J. B. (2025). Risk behind the veil of ambiguity: Decision‐making under social and nonsocial sources of uncertainty. Risk Analysis, 45, 3144–3159. 10.1111/risa.70081