Effects of a fearful emotional state on financial decisions in the presence of prior outcome information

Abstract

Negative emotions have been shown to influence financial risk-taking. However, how receiving salient information about prior outcomes interacts with a decision-maker’s emotional state is not well known. In a laboratory experiment, we induced a fearful emotional state to investigate its effects on financial investment when outcome probabilities are unknown but decision-makers observe prior outcomes. The effects of fear on investment depended on whether the sequence of previous outcomes was favorable or unfavorable and contained weak or strong information. Our findings suggest that fear affected investment, at least in part, through changes in expectations of success.

1. Introduction

Individuals constantly choose among alternatives with uncertain consequences. Sometimes, before making a choice, they obtain information about prior outcomes of the choice options. For example, a financial investor usually observes previous returns of several financial assets before choosing a portfolio. Similarly, a physician learns how different medical treatments have affected a given patient in the past before deciding on a new treatment course for that patient. The way in which individuals process prior outcome information and use it to reach a decision might be influenced by their emotional state, even if current emotions are incidental, that is, driven by circumstances that are unrelated to the decision at hand. How do emotional states influence decision-making under uncertainty when individuals observe prior outcomes? We address this question in this paper, focusing on the effects of a negative emotion—fear—on financial decisions.

The influence of emotional states on decision-making in the absence of information about prior outcomes has been widely studied in psychology (see Lerner et al. (2015) for an overview). More recently, economics studies have investigated temporary variations in risk preferences and risk-taking resulting from shifts in emotional states (see Schildberg-Hörisch (2018) for an overview and Meier (2022) for a recent study using survey data). By contrast, less is known about how emotional states affect decision-making when individuals observe prior outcomes. While psychologists and neuroscientists have proposed that high-arousal/negative-valence emotional states affect the attention paid to specific pieces of information (Forgas, 1989, 1991; Wichary et al., 2015), there is scant systematic evidence on how emotions affect processing of prior outcome information and subsequent financial decisions.

To help fill this gap in the existing research, we conducted a laboratory experiment on financial risk-taking under negative emotion—predominantly fear—in the presence of information about prior outcomes. Our design allowed us to study whether the influence of induced fear on investment varies with the type of outcome information (i.e., good/bad news) and its strength (i.e., weak/strong news). Our focus on fear was motivated by several previous studies establishing a causal relationship between fear and financial risk-taking (Lee and Andrade, 2011; Cohn et al., 2015; Guiso et al., 2018).

Our study builds on previous experimental research analyzing systematic effects of changes in the emotional state on financial decision-making. Such research focused on risk-taking in the absence of prior outcome information (see Marini (2022) for a review and meta-analysis). Some studies investigated the influence of emotions arising from the decision at hand (i.e., integral emotions), for example, during the decision-making process and in anticipation of the resolution of uncertainty (van Winden et al., 2011; Nguyen and Noussair, 2014; Lucarelli et al., 2015; Breaban and Noussair, 2018). These studies, however, were unable to identify the causal effects of the emotional state on decision-making because integral emotional experiences could be correlated with other determinants of financial decisions, such as baseline risk preferences or expectations. To overcome this limitation, other studies directly manipulated subjects’ emotional state appealing to stimuli unrelated to the decision at hand, thereby focusing on the role of incidental emotions in risk-taking (see, e.g., Lee and Andrade (2011), Campos-Vazquez and Cuilty (2014), Cohn et al. (2015), Conte et al. (2018), Guiso et al. (2018), and Kassas et al. (2022)). We took this approach too. Importantly, we complemented these studies by independently manipulating information about previous outcomes, with the aim of assessing how a fearful emotional state affects financial risk-taking in the presence of such information.

In our experiment, we hypothesized that the interaction between the induced fear and prior outcome information would influence investment behavior. In fact, we found that the effect of fear on investment depended on whether the sequence of previous outcomes was favorable or unfavorable and contained weak or strong information. Specifically, compared to control subjects, those who were treated to experience fear invested less when observing weakly bad outcomes but more when the outcomes were strongly bad. By contrast, good prior outcomes seemed to counterbalance the effect of fear on risk-taking, resulting in no difference in investment between treated and control subjects. Interestingly, we observed the emergence of contrarian behavior among subjects who witnessed strongly bad outcomes following the induction of fear. This small group of subjects increased their investment compared to their baseline investment (which they had made prior to the induction of fear and when prior outcome information was unavailable). Overall, the observed effects of fear on investment operated, at least in part, through changes in expectations of success.

The rest of the paper is organized as follows. Section 2 presents the conceptual framework, which lays the groundwork for our experiment. Section 3 describes the design and procedures of the experiment. Section 4 presents the results. Section 5 discusses the findings and some limitations of our study and outlines suggestions for future research.

2. Conceptual framework

In this section, we lay out a simple framework to analyze the link between fear and financial risk-taking.¹ We first discuss the potential effects of fear on financial risk-taking when salient information about prior outcomes is absent, which has been the focus of previous research. We then discuss how these effects might change when individuals have information about prior outcomes.

2.1. Setup

Consider an individual (henceforth, the decision-maker) who is endowed with $X$ monetary units and decides how much to invest in a lottery $L$ with an unknown probability of success. We denote the amount invested by $x$. If the outcome of the lottery is successful, the decision-maker obtains $2.5\hspace{0.17em}x$. Otherwise, she loses $x$. The remaining amount that is not invested in the lottery (i.e., $X-x$ ) is kept by the decision-maker as a safe payment.

At the moment of choosing her investment, the decision-maker experiences incidental fear (i.e., fear triggered by a source that has no relation to the lottery) with intensity $f\ge 0$, where $f=0$ represents a neutral emotional state. Previous research has found that incidental emotions pervasively carry over from one situation to the next, typically without awareness, affecting decisions that should be unrelated to that emotion (Lerner et al., 2015). One mechanism through which the carryover effect occurs in decisions under uncertainty is a change in risk judgment. In our setup, this means that fear may affect the decision-maker’s expectation of a successful lottery outcome. Let $p\text{(}f\text{)}\in \text{(}0\text{,}\hspace{0.33em}1\text{)}$ denote the subjective probability of success as a function of the intensity of fear.²

Given an investment $x$, the value of $L$ as perceived by the decision-maker is

$$U\text{(}L\text{)}=p\text{(}f\text{)}\hspace{0.33em}u\text{(}2.5x+X-x\text{)}+\text{(}1-p\text{(}f\text{))}\hspace{0.33em}u\text{(}X-x\text{),}$$ (1)

where $u$ is an increasing and concave utility function over outcomes. The decision-maker chooses her investment to maximize the value of $L$ as represented in (1). Intuitively, optimal investment increases with the expectation of success.³ We next discuss how the experience of fear (i.e., $f>0$ ) could affect the expectation of success and hence investment compared with a neutral emotional state (i.e., $f=0$ ).

2.2. Decision-making without information about prior outcomes

First, consider a situation in which the decision-maker lacks information about previous outcomes of $L$ and hence forms a prior expectation of success, $p_0\left(f\right)$. Two prominent psychological theories, namely the affect infusion model (Forgas, 1995) and the appraisal-tendency framework (Lerner and Keltner, 2000), concern the role of the emotional state in risk judgments.

The affect infusion model argues that the emotional state may affect decisions through its influence on selective attention to information. In particular, this model predicts that subjects in a negative emotional state will be more pessimistic than subjects in a positive emotional state because they are primed to pay more attention to downside risks. The affective generalization hypothesis (Johnson and Tversky, 1983) similarly postulates that negative emotions trigger more pessimistic risk judgments than positive emotions, although it does not elaborate on the mechanism.

The appraisal-tendency framework posits that incidental emotional states predispose individuals to assess future events in line with the central appraisal dimensions that triggered the emotion. Key appraisal dimensions triggered by fear are certainty (i.e., the degree to which future events seem predictable and comprehensible) and control (i.e., the degree to which events seem to be brought about by individual agency vs situational agency). Specifically, incidental fear leads individuals to perceive possible negative events in the future as unpredictable (i.e., featuring low certainty) and under situational control (i.e., featuring low control). Consequently, incidental fear results in pessimistic risk judgments.

In summary, in the current setup, the aforementioned psychological theories imply that fear triggers a more pessimistic prior judgment of the probability of success compared with a neutral emotional state: $p_0\text{(}f\text{)}<p_0\text{(}0\text{)}$. Such pessimistic judgment, in turn, leads to lower investment: $x\left(p_0\text{(}f\text{)}\right)<x\left(p_0\text{(}0\text{)}\right)$.

With a few exceptions, empirical evidence on shifts in decision-making under uncertainty due to changes in the intensity of fear tends to support these predictions. First, Lerner and Keltner (2001) found that induced fear creates a tendency to hold pessimistic risk perceptions in relation to several life events. Second, several experimental studies on financial decisions found that fear reduces individual risk-taking in the absence of information about prior outcomes (Lee and Andrade, 2011; Cohn et al., 2015; Guiso et al., 2018). Additionally, studies that manipulated chronic stress (e.g., Kandasamy et al. (2014)), acute psychosocial stress (e.g., Cahlíková and Cingl (2017)), or financial worries (e.g., Dalton et al. (2020)) typically found that these factors, like fear, tend to reduce risk-taking in the absence of prior outcome information.

2.3. Decision-making with information about prior outcomes

Now consider a situation in which the decision-maker observes a sequence of prior outcomes of $L$, after which she forms a posterior expectation of success, $p_1\left(f\right)$. While new information may affect expectations on its own (i.e., $p_1\left(f\right)$ may differ from $p_0\left(f\right)$ ), this is not our main focus here. Rather, we are interested in analyzing how incidental fear affects $p_1$, and hence final investment, once the new information has been accounted for. Specifically, we pose the following questions: (1) Compared with a neutral emotional state, does fear still affect risk judgment and investment when the decision-maker observes a sequence of prior outcomes? (2) If so, does the effect depend on whether the sequence of previous outcomes was favorable or unfavorable and contained weak or strong information? Because previous research abstracted from the interaction between emotional state and prior outcome information, there is no systematic evidence regarding these particular questions. Nevertheless, we next advance a few exploratory hypotheses.

One hypothesis is that observation of prior outcomes does not significantly interfere with the influence of incidental fear. In this case, we would expect similar qualitative effects of fear on decision-making regardless of the outcome information possessed by the decision-maker. Thus, to make a prediction, we can extrapolate the insights from previous research (discussed in Section 2.2), which concern situations without information about prior outcomes. This hypothesis then implies that fear results in more pessimistic expectations and lower investment even in the presence of prior outcome information: that is, $p_1\left(f\right)<p_1\left(0\right)$ and $x\text{(}{p}_1\text{(}f\text{))}<x\text{(}{p}_1\text{(}0\text{))}$ for both favorable and unfavorable outcome sequences.

However, because transparent information about prior outcomes reduces outcome uncertainty (regardless of whether the news is good or bad), it could debias the decision-maker from the “mental contamination” produced by incidental fear (Wilson and Brekke, 1994). According to this hypothesis, the new information conveyed by prior outcomes offsets the influence of fear on risk judgment and investment: that is, $p_1\left(f\right)=p_1\left(0\right)$ and $x\text{(}{p}_1\text{(}f\text{))}=x\text{(}{p}_1\text{(}0\text{))}$ for both favorable and unfavorable outcome sequences. A more nuanced version of this hypothesis is that new information about prior outcomes has a debiasing effect only if it significantly reduces the original outcome uncertainty. When new information is weak or ambiguous, the tendency to reduce expectations and investment as a result of fear remains.

Finally, it is also possible that, compared with a neutral emotional state, fear has an asymmetric effect on decision-making between situations with good news and those with bad news. In particular, while fear may still induce more pessimism and lower investment conditional on receiving favorable news, it may increase the expectation of success and hence investment conditional on receiving unfavorable news: that is, $p_1\left(f\right)<p_1\left(0\right)$ and $x\text{(}{p}_1\text{(}f\text{))}<x\text{(}{p}_1\text{(}0\text{))}$ when the news is good but $p_1\left(f\right)>p_1\left(0\right)$ and $x\text{(}{p}_1\text{(}f\text{))}>x\text{(}{p}_1\text{(}0\text{))}$ when the news is bad.

This hypothesis draws on the mood maintenance hypothesis (Isen and Patrick, 1983), which posits that individuals who are in a negative affective state take risks to improve how they feel. We speculate that the combination of incidental fear and bad news could exarcebate this pattern. Intuitively, people who are already experiencing fear and then confront bad news may feel quite distressed, which may lead them to become contrarians, holding overly optimistic expectations and taking excessive risk in an attempt to alleviate the distress. Although this hypothesis has not been directly tested by previous research, laboratory studies did find contrarian behavior among a small subset of traders in experimental financial markets (Drehmann et al., 2004; Cipriani and Guarino, 2009).

While we have drawn on previous research to advance hypotheses on the role of fear when individuals observe prior outcomes, we emphasize that a unifying theoretical framework is still missing in the literature and that proposing one is beyond the scope of this paper. Therefore, we adopted an exploratory approach to investigate the hypotheses. In the next section, we describe the laboratory experiment that we employed to conduct such exploration.

3. Experiment

3.1. General aspects

Using the Online Recruitment System for Economic Experiments (ORSEE) (Greiner, 2004), we recruited a total of 401 undergraduate students from several universities in the city of Bogotá to participate in one of several study sessions, which lasted for approximately 75–90 minutes. Sessions were conducted in three waves: February 2020, September/October 2020, and October/November 2022. While most sessions (encompassing 284 subjects) were conducted remotely through the Zoom platform, some (encompassing 117 subjects) were conducted in person at the Rosario Experimental and Behavioral Economics Laboratory (REBEL) from Universidad del Rosario. We ensured that the treatment and control groups were balanced in the proportion of subjects who attended each type of session.

The experiment was computerized and the interface was programmed in oTree (Chen et al., 2016). Payments were made through an online bank transfer within 48 hours of the end of the session.

For the in-person sessions, on arrival at the laboratory, each subject was seated at a separate carrel in front of a computer terminal and received a set of headphones. Computer carrels were separated by panels that ensured that subjects could not see each other.

In the remote sessions, the subjects were instructed to stay in a quiet room, keep their cameras off, and use their own set of headphones. They received the link to the experimental interface through the Zoom chat box. To minimize procedural differences with in-person sessions, the instructors kept their cameras on to provide verbal instructions and display physical procedures.

The Online Appendix contains the experimental procedures and instructions translated into English. Additionally, a package of replication materials for this study is available at https://osf.io/7swfk. Figure 1 summarizes the timeline of an experimental session, which we further describe below.

Figure 1:

Timeline of an experimental session

Notes: The experiment began with the random assignment of a red or blue colored ticket to each participant as their winning color for the lottery. By design, in each session half of the participants were assigned to either color. After this, participants received their endowment and were told they could invest any amount in the lottery and that if the color of the ball drawn from the urn (lottery) matched their ticket color, they would receive 2.5 times the amount invested. After submitting their baseline investment amount and their prior expectation of success (i.e., the probability that the lottery color would match their ticket), participants watched either the control video or the treatment video. After watching the video, participants were shown three practice draws from the urn that did not count toward their payment. The purple bubble indicates the type and strength of the new information for each possible sequence of the draws for a participant with a red colored ticket. After the practice draws, participants were allowed to revise their investment and expectation of success. Before resolving the lottery, subjects filled out their emotional rating self-report. The final draw, which was performed in front of all the subjects in the session, determined each subject’s payment.

3.2. Timeline of an experimental session

Task: Baseline investment decision

Subjects participated in an investment task, adapted from Gneezy and Potters (1997) and Cohn et al. (2015). They were endowed with 20,000 Colombian pesos (COP) (equivalent to approximately 5.50 USD at the time of the experiment) and decided how much to invest in a lottery.

To set up the lottery, using a computer, half of the subjects were randomly assigned a red ticket and the other half received a blue ticket. This procedure was common knowledge but the color of the ticket received was private information. Subjects were told that they would use the ticket to participate in a lottery and were then shown an empty black urn, which one of the instructors later filled in private. Subjects knew only that the urn contained a total of 10 balls, each of which was either red or blue, and they did not receive any information about its composition.

The outcome of the lottery was determined by randomly drawing a ball from the urn. If the color of the ball drawn matched the color of the subject’s ticket, the subject won 2.5 times the amount invested. Otherwise, he or she lost the amount invested. The remaining amount that was not invested in the lottery was kept by subjects as a safe payment. Because the composition of the urn was completely unknown to subjects, the lottery was an ambiguous gamble (Ellsberg, 1961). The subjects were allowed to use a calculator as an aid in making their decision. Before indicating their investment, they answered a few comprehension questions and were given time to resolve any doubts.

Announcement of practice trials

After the subjects made their baseline investment decision, they were truthfully told that, before resolving the lottery, one of the instructors would perform three unpaid practice trials with replacement.

Task: Elicitation of the prior expectation of success

Subjects then answered the following question: “What do you think is the probability that the FIRST practice draw will be successful? You must express your answer as “ $x$ percent probability,” where $x$ is a number between zero (inclusive) and 100 (inclusive).” The subjects provided their answers by placing a cursor on a slider ranging between 0% and 100% (with one percentage-point increments). This task was not monetarily incentivized. We used this question to discover what the participants’ prior expectations of success might have been when they made their baseline investment choice.

This belief elicitation task, which draws on the one used by Sautua (2020), has two noteworthy aspects. First, the reference to the first practice trial, rather than the draw that would resolve the lottery, was intended to disguise the true purpose of the question to prevent post hoc rationalization of investment choices through stated expectations. Second, we did not incentivize subjects’ answers because paying subjects for both choices and reported expectations could have induced them to hedge across tasks, thus contaminating the reports.

Once the subjects had reported their prior expectations of success, they were informed that they would have the opportunity to revise their initial investment decisions after observing the outcomes of the three unpaid practice trials.

Treatment: Induction of fear

Next, subjects underwent a video-based manipulation of emotional state. Passive viewing of short videos with emotional content is generally considered to be effective in inducing both negative- and positive-valence incidental emotions (Uhrig et al., 2016; Gilman et al., 2017; Fernandez-Aguilar et al., 2019). Moreover, video clips were used in several previous studies examining the role of emotional states in risk-taking (see, e.g., Lee and Andrade (2011), Conte et al. (2018), and Kassas et al. (2022)). Given that most of the sessions were remote, we note that induction of emotional states using videos in online experiments has been shown to be a robust method, including for studies aimed at eliciting fear (Ferrer et al., 2015). To implement our intervention, we randomly allocated subjects to either a treatment group or one of two control groups. We included two control groups to verify the robustness of our key findings.

Subjects in the Treatment group watched a 3.5-minute excerpt from the short horror film “I Heard It Too.” The horror film excerpt was intended to induce a significantly more negative emotional state, dominated by fear, in the Treatment group compared with the two control groups. The video was chosen after other candidate videos were piloted (e.g., YouTube video compilations of mid-flight intense turbulence). We chose this specific video based on length, simplicity in the narrative, low anticipated familiarity with the video in our targeted study sample, and effectiveness at inducing the intended emotion of fear.

Subjects in the Control-1 group watched a 2.5-minute excerpt from the video “1 Full Hour of Birds in the Birdbath.” Subjects in the Control-2 group watched the 2-minute BBC documentary “World’s Smallest Cat.” Both control videos had nature-related content and were chosen for their soothing nature, which was intended to induce positive-valence emotions of low to mild intensity. As we show in Section 4.2 below, the two control groups were indeed comparable.

The Online Appendix provides additional details on the video selection process and the content of the video clips.

Implementation of practice trials

One of the instructors then performed the three trials in front of all the subjects. The practice draws provided two independent sources of random variation in the information about prior outcomes observed by the subjects.

First, because the tickets had been randomly assigned in a given session, the same sequence of trials yielded a majority of good draws for half of the subjects and a majority of bad draws for the other half. This feature of the design provided exogenous between-subjects variation in the type of information conveyed by the sequence of practice draws.

Second, note that all the subjects in a given session randomly observed either two or three draws of the same color. This feature of the design provided exogenous between-subjects variation in the strength of the information conveyed by the sequence of practice trials.

For present purposes, a “weakly good (bad) sequence” consists of two good (bad) draws and one bad (good) draw, whereas a “strongly good (bad) sequence” consists of three good (bad) draws. When we do not want to distinguish between weak and strong sequences, we simply refer to sequences with a majority of good (bad) draws as “good (bad) sequences.”

Task: Revision of investment decision

Immediately after observing the outcomes of the three unpaid practice trials, the subjects were given the opportunity to revise their initial investment decision if they chose to. Subjects made their final investment choices knowing that the draw that was to resolve the lottery would follow the observed sequence of practice draws.

Task: Elicitation of the posterior expectation of success

After revising their investment decision, and before the lottery was resolved, the subjects were asked to imagine a hypothetical fourth practice trial. Subjects then answered the following question: “What do you think is the probability that the HYPOTHETICAL fourth practice draw will be successful? You must express your answer as “ $x$ percent probability,” where $x$ is a number between zero (inclusive) and 100 (inclusive).” The subjects indicated their answers using the slider on the computer screen. We used this question to discover what the participants’ posterior expectations of success might have been when they made their final investment choice. As discussed in our conceptual framework (Section 2), posterior expectations of success are a key channel through which fear could affect investment when subjects observe previous outcomes. The elicitation of such expectations allowed us to evaluate this mechanism.

Similar to the elicitation of prior expectations, the reference to a hypothetical draw was intended to prevent post hoc rationalization of investment choices through stated expectations, and reported beliefs were not monetarily incentivized. In Section 4.4.2, we show that despite the lack of monetary incentives for truthful revelation, there was a strong link between reported expectations of success and investments, which suggests that subjects’ reports were truthful.

Emotional state self-report

Before resolving the lottery, we asked subjects to rate the intensity with which they had experienced six emotions while watching the video clip. Of the six emotions, four were of negative valence—fear, rage, sadness, and disgust—and two were of positive valence—amusement and happiness. For each emotion, the subjects reported the intensity on a 10-point scale ranging from one (no emotional experience) to ten (the most intense experience of their lives). We used the subjects’ self-reported emotional ratings to check the effectiveness of our manipulation of emotional states.

We asked the subjects to provide the emotional ratings after they completed all tasks, rather than earlier, to avoid contaminating the intrinsic effect of the induced emotional state on decision-making. Prompting subjects to label their emotions earlier would have raised their awareness, perhaps leading to an increased cognitive monitoring and faster regulation of emotions (Keltner et al., 1993). Alternatively, the mere passing of time associated with the act of completing the questionnaire may have diluted the emotions that had been induced just minutes earlier (Kassas et al., 2022). In any case, the attenuation of the intensity of emotions could have reduced the impact of the emotional state on expectations of success and investment decisions.

Lottery resolution

At the end of the session, one of the instructors made the final draw from the urn to resolve the lottery and determine payments.

4. Results

Motivated by the conceptual framework discussed in Section 2, our analysis of the experimental data focused on answering the following questions: (1) Did fear affect investment in the presence of prior outcome information? (2) Did the effect of fear depend on whether the sequence of previous outcomes was favorable or unfavorable and contained weak or strong information? (3) Did fear influence final investment by affecting posterior expectations of success? We address these questions after briefly describing our study sample and verifying that the treatment was effective at inducing the intended emotion of fear.

4.1. Sample description

Data collection ended in November 2022 when we reached a target of 400 participants, which we deemed reasonable given the exploratory nature of our study. The sample consisted of 142 subjects in the Treatment group, 129 subjects in the Control-1 group, and 130 subjects in the Control-2 group. Table 1 displays the distribution of subjects by group and observed sequence. (Online Appendix Table A1 further breaks down the distribution by type of session.) Overall, 60% of subjects were women, only 29% were economics or finance majors, and the average subject was 21.3 years old.

Table 1:

Distribution of subjects by group and observed sequence

Group	Good sequences			Bad sequences			All sequences
	Weak	Strong	Total	Weak	Strong	Total	Weak	Strong	Total
Treatment	38	33	71	35	36	71	73	69	142
Control-1	25	40	65	23	41	64	48	81	129
Control-2	35	32	67	34	29	63	69	61	130
Total	98	105	203	92	106	198	190	211	401

4.2. Manipulation check: Emotions

We first checked whether the treatment was effective at inducing a more negative emotional state, dominated by fear, in the Treatment group relative to the control groups. Figure 2 displays the mean reported score for each emotion for each group. On average, the treatment had the intended effect on subjects’ emotional states.

Figure 2:

Average score of each emotion by group

Note: Bars indicate standard errors of the means.

We highlight three key features of the emotional state self-reports. First, fear was the emotion with the largest mean score in the Treatment group (all p-values < 0.001, using Wilcoxon signed-rank tests). Second, the mean score of fear was significantly larger in the Treatment group than in the Control-1 and Control-2 groups (p < 0.001 in both comparisons, using Wilcoxon rank-sum tests). These two features can be summarized by noting that fear was the dominant emotion for 53.5% of subjects in the Treatment group but only for 12.4% and 17.7% of subjects in the Control-1 and Control-2 groups. Third, compared with subjects in the two control groups, those in the Treatment group reported larger scores on disgust, sadness, and rage and smaller scores on amusement and happiness (all p-values ≤ 0.002, using Wilcoxon rank-sum tests).

We summarized the treatment effects on emotional states by constructing three indexes for each subject: (1) a negative-valence emotion index, which averages the scores on fear, disgust, sadness, and rage; (2) a positive-valence emotion index, which averages the scores on amusement and happiness; and (3) an index of excess negative emotion, defined as the difference between the negative-valence and positive-valence emotion indexes. We standardized all three indexes by subtracting their means and dividing by their standard deviations (z-scores). In Online Appendix Table A2, we used these indexes to show that the aforementioned patterns in emotional state self-reports are robust to controlling for the type of session attended and subjects’ gender in a regression framework.

Overall, the treatment increased the intensity of negative emotion by more than one standard deviation and decreased positive emotion by approximately one standard deviation, thus increasing excess negative emotion by approximately 1.3 standard deviations. All effects were statistically significant at p < 0.001 (using two-tailed t-tests) and were very similar using either control group as the benchmark.⁴ As additional evidence of the large treatment effect on excess negative emotion, we note that while 79.6% of subjects in the treatment group displayed a positive value of the excess negative emotion index, only 13.2% and 19.2% did so in the Control-1 and Control-2 groups.

Result 1. Compared with control subjects, those who were treated with the horror film excerpt experienced a significantly more negative emotional state. Fear was the dominant emotion among treated subjects.

4.3. Treatment effects on investment

This section presents the main results, which concern the effects of exposure to the horror film excerpt on investment in the presence of prior outcome information. Result 1 suggests that treatment effects on investment, if they exist, can be attributed to a more negative emotional state dominated by fear.

We define a subject’s investment rate as the share of the endowment invested by the subject in the lottery. The average investment rate at baseline (that is, before the treatment’s implementation and without outcome information) was 40%. Figure 3 shows, for each observed sequence and group, the average within-subjects change in investment rates relative to the baseline (in percentage points). We highlight three key patterns.

Figure 3:

Average change in investment rate by observed sequence and group

Notes: The change in the investment rate is defined as the difference between a subject’s final and baseline investment rates. Bars indicate standard errors of the means.

First, as expected, average investment rates increased following good sequences and decreased following bad sequences (except in the Control-1 group after a weakly bad sequence, where there was a slight increase).

Second, among control subjects, average investment rates were more responsive to strong information than to weak information following both good and bad sequences. By contrast, while the average investment rate in the Treatment group responded more to strongly good sequences than to weakly good sequences, it was equally responsive to strongly bad and weakly bad sequences.

Third, while there were no systematic treatment effects following good sequences, the treatment led to a larger reduction in the average investment rate following weakly bad sequences and a smaller reduction in the average investment rate following strongly bad sequences.

Next, we assess the statistical significance of the average treatment effects on investment decisions. For ease of exposition, we focus on group differences in final investment rates. Using ordinary least squares (OLS), we separately estimated the following regression model on the subsample of subjects who observed a good sequence and on those who observed a bad sequence:

$$I_i={\alpha }_0+{\alpha }_1{T}_i+{\alpha }_2\hspace{0.33em}Strong{}_i+{\alpha }_3\left(T_i\times Strong{}_i\right)+X_i^{\prime }{\alpha }_4+{ϵ}_i\text{.}$$ (2)

The dependent variable $I_i$ is the final investment rate of subject $i$. $T_i$ is a dummy variable for the Treatment group. $Stron{g}_i$ is a dummy variable for strong sequences. $X_i$ is a vector of control variables that includes subject $i\text{’s}$ baseline investment rate, a dummy variable that equals 1 if subject $i$ attended an in-person session and 0 otherwise, and a dummy variable that equals 1 if subject $i$ is female and 0 otherwise. Finally, ${ϵ}_i$ is the idiosyncratic error term. We included the interaction between $T_i$ and $Stron{g}_i$ to test whether treatment effects depended on the strength of the observed sequence. The parameter ${\alpha }_1\text{(}{\alpha }_1+{\alpha }_3\text{)}$ captures the average treatment effect on final investment rates when subjects received weak (strong) outcome information. A non-zero value of ${\alpha }_3$ indicates that the average treatment effect was different when subjects received strong rather than weak information.

The results are shown in Table 2. Columns 1 and 4 (2 and 5) display the estimates for good and bad sequences, respectively, using the Control-1 (Control-2) group as the benchmark. Recall from Section 4.2 that the treatment effects on subjects’ emotional states were very similar using either control group for comparison. Hence, we expected similar treatment effects on investment using either control group as the benchmark, and, in fact, this is what we found. On the basis of these results, hereafter we combine observations from both control groups to boost statistical power. Columns 3 and 6 display the estimates for good and bad sequences, respectively, using the Pooled Control group as the benchmark. Unless otherwise noted, p-values from tests regarding regression coefficients are from two-tailed t-tests.

Table 2:

Average treatment effects on final investment rates

	Good sequences			Bad sequences
	(1)	(2)	(3)	(4)	(5)	(6)
T (=1)	0.029	−0.026	−0.006	−0.080*	−0.044	−0.061*
	(0.047)	(0.034)	(0.033)	(0.034)	(0.033)	(0.028)
Strong (=1)	0.116	0.075	0.083*	−0.124**	−0.111*	−0.113***
	(0.066)	(0.040)	(0.034)	(0.045)	(0.045)	(0.032)
T×Strong	−0.041	0.003	−0.011	0.138*	0.135*	0.133*
	(0.075)	(0.062)	(0.056)	(0.062)	(0.062)	(0.053)
Baseline investment	0.709***	0.800***	0.730***	0.784***	0.811***	0.776***
	(0.087)	(0.057)	(0.065)	(0.084)	(0.072)	(0.066)
In-person session (=1)	0.041	0.041	0.026	−0.013	−0.031	−0.020
	(0.047)	(0.037)	(0.029)	(0.027)	(0.024)	(0.023)
Female (=1)	−0.058	−0.049	−0.048	−0.029	0.032	0.003
	(0.044)	(0.040)	(0.035)	(0.033)	(0.031)	(0.027)
Constant	0.161*	0.178***	0.189***	0.131*	0.049	0.096*
	(0.064)	(0.047)	(0.046)	(0.052)	(0.042)	(0.039)
Control group	1	2	Pooled	1	2	Pooled
Number of subjects	136	138	203	135	134	198

Notes: This table displays OLS estimates of equation (2). Columns 1–3 (4–6) show estimates for the subsample of subjects who observed a good (bad) sequence. Columns 1 and 4 (2 and 5) [3 and 6] show estimates using the Control-1 (Control-2) [Pooled Control] group for comparison. Robust standard errors are shown in parentheses.

^{*, **, ***}The symbols denote that p < 0.05, p < 0.01, and p < 0.001, respectively, using a two-tailed t-test.

While there were no statistically significant group differences for good sequences, there were sharp treatment effects for bad sequences. Moreover, consistent with Figure 3, the sign of the treatment effect varied with the strength of the bad news. For weakly bad sequences, the average final investment rate was approximately 6 percentage points smaller in the Treatment group compared with the Pooled Control group (p = 0.031). The sign of the treatment effect was reversed for strongly bad news, as indicated by the large positive estimate of ${\alpha }_3$ (p = 0.012). Indeed, for strongly bad news, the average final investment rate was approximately 7 percentage points larger in the Treatment group compared with the Pooled Control group (p = 0.049 using a one-tailed $t$ -test).

The varying treatment effects for bad sequences are related to sharp group differences in the response of investment to strong vs weak information. This is seen by comparing the estimates of ${\alpha }_2$ and ${\alpha }_2+{\alpha }_3$. The former (latter) captures the difference in the average final investment rate between strong and weak sequences among subjects in the Pooled Control (Treatment) group. Subjects in the Pooled Control group invested approximately 11 percentage points less when the bad news was strong than when it was weak (p < 0.001). By contrast, subjects in the Treatment group invested virtually the same amount regardless of the strength of the bad sequence, as shown by the small and statistically insignificant estimate of ${\alpha }_2+{\alpha }_3$ (p = 0.604).⁵

Result 2. Exposure to the horror film excerpt had no significant effects on final investment rates for good sequences. By contrast, it had clear effects for bad sequences, and these varied with the strength of the outcome information. Compared to control subjects, those who were treated with the horror film excerpt invested a smaller amount when the bad news was weak and a larger amount when the bad news was strong.

4.4. Further investigation of the main findings

In this section, we report additional analyses to gain more insight into our main findings. We begin with a more nuanced characterization of the treatment effects on investment, further examining investment changes relative to the baseline. We then explore the role of expectations of success as an underlying mechanism.

4.4.1. Investment changes relative to the baseline

Figure 4 shows histograms of within-subjects investment changes relative to the baseline for each observed sequence and group. It uncovers two patterns of investment behavior that were masked by the preceding analysis of average treatment effects on final investment rates.

Figure 4:

Histogram of changes in investment rates by observed sequence and group

Note: The change in the investment rate is defined as the difference between a subject’s final and baseline investment rates.

First, for all sequences and groups, the modal response to the practice draws was to make no change in investments. Overall, 44% of subjects decided not to revise their investments. Inertia, however, significantly dropped from 51% after weak sequences to 38% after strong sequences (p = 0.008, ${\chi }^2$ -test pooling all groups).⁶

Second, some subjects made contrarian investment revisions, that is, they decreased (increased) their investment after good (bad) sequences. Although only 9.2% of all subjects made a contrarian investment revision, 75.7% of contrarian choices were made after bad news. Most notably, for bad sequences, the treatment appears to have reduced the prevalence of contrarian behavior when outcome information was weak and increased the prevalence of such behavior when information was strong. This helps to explain our finding that the average final investment rate of treated subjects was smaller than that of control subjects for weakly bad sequences and larger for strongly bad sequences (see Result 2).

To assess the statistical significance of these patterns of contrarian investment behavior, we estimated the following regression model using OLS:

$$CI{}_i={\beta }_0+{\beta }_1{T}_i+{\beta }_2\hspace{0.33em}Strong{}_i+{\beta }_3\left(T_i\times Strong{}_i\right)+X_i^{\prime }{\beta }_4+{ϵ}_i\text{.}$$ (3)

The dependent variable $C{I}_i$ is a dummy variable that equals 1 if subject $i$ made a contrarian investment revision and 0 otherwise. $T_i$, $Stron{g}_i$, and $X_i$ are the same as for equation (2). Again, we separately estimated equation (3) on the subsample of subjects who observed a good sequence and on those who observed a bad sequence. The parameter ${\beta }_1\text{(}{\beta }_1+{\beta }_3\text{)}$ captures the average treatment effect on the probability of being a contrarian investor when subjects received weak (strong) outcome information.

The results are shown in Table 3. Column 1 (2) displays the estimates for good (bad) sequences. There were no statistically significant treatment effects on the prevalence of contrarian investment revisions following good news. Turning to bad sequences, treated subjects were significantly more likely to be contrarians when outcome information was strong, as shown by the large positive estimates of ${\beta }_3$ (p = 0.023) and ${\beta }_1+{\beta }_3$ (p = 0.026 using a one-tailed t-test).

Table 3:

Treatment effects on contrarian investment behavior

	Good sequences	Bad sequences
	(1)	(2)
T (=1)	−0.036	−0.093
	(0.039)	(0.074)
Strong (=1)	−0.017	−0.136*
	(0.036)	(0.062)
T×Strong	0.026	0.241*
	(0.052)	(0.105)
Baseline investment	0.046	−0.222*
	(0.059)	(0.090)
In-person session (=1)	0.003	−0.049
	(0.028)	(0.057)
Female (=1)	−0.071	−0.046
	(0.038)	(0.051)
Constant	0.087	0.338***
	(0.053)	(0.086)
Control group	Pooled	Pooled
Number of subjects	203	198

Notes: This table displays OLS estimates of equation (3). Column 1 (2) shows estimates for the subsample of subjects who observed a good (bad) sequence. Robust standard errors are shown in parentheses.

^{*, **, ***}The symbols denote that p < 0.05, p < 0.01, and p < 0.001, respectively, using a two-tailed t-test.

Similar to our main findings, the varying treatment effects for bad sequences are related to sharp group differences in the propensity for contrarian behavior after strong vs weak information. Subjects in the Pooled Control group were significantly less likely to be contrarians when the bad news was strong than when it was weak, as shown by the large negative estimate of ${\beta }_2$. However, subjects in the Treatment group displayed the opposite pattern.

Result 3. There were no treatment effects on the prevalence of contrarian investment revisions after good sequences. By contrast, compared to control subjects, those who were treated with the horror film excerpt were more likely to be contrarians following strongly bad sequences. This behavioral pattern helps to explain the larger average final investment rate of treated subjects for strongly bad sequences.

4.4.2. Mechanisms: The role of expectations of success

Last, we tested whether the observed treatment effects on investment were mediated, at least in part, by changes in posterior expectations of success. We began by estimating the following regression model using OLS:

$$E_i={\gamma }_0+{\gamma }_1{T}_i+{\gamma }_2\hspace{0.33em}Strong{}_i+{\gamma }_3\left(T_i\times Strong{}_i\right)+Z_i^{\prime }{\gamma }_4+{ϵ}_i\text{.}$$ (4)

The dependent variable $E_i$ is the posterior probability of a successful lottery outcome expected by subject $i$. $E_i$ is expressed as a number between 0 and 1. $T_i$ and $Stron{g}_i$ are the same as for equations (2) and (3). $Z_i$ is a vector of control variables that includes subject $i\text{’s}$ prior expectation of success and the dummy variables for in-person sessions and female subjects. Again, we separately estimated equation (4) on the subsample of subjects who observed a good sequence and on those who observed a bad sequence.

The results are shown in Table 4. Column 1 (2) displays the estimates for good (bad) sequences. Consistent with the null treatment effects on investment for good sequences (see Result 2), there were no significant group differences in posterior expectations in this case. After weakly bad sequences, treated subjects held lower expectations of success than did control subjects, although ${\gamma }_1=0$ cannot be rejected at p ≤ 0.10. Following strongly bad sequences, treated subjects maintained higher expectations of success than did control subjects (p = 0.049 for ${\gamma }_3=0$ and p = 0.067 for ${\gamma }_1+{\gamma }_3\le 0$ ), which helps to explain the larger average final investment rate in the Treatment group (see Result 2). Note that, for bad sequences, control subjects held significantly lower expectations when information was strong than when it was weak (p = 0.012 for ${\gamma }_2=0$ ), whereas treated subjects held similar expectations in both cases (p = 0.686 for ${\gamma }_2+{\gamma }_3\le 0$ ).

Table 4:

Treatment effects on posterior expectations of success

	Good sequences	Bad sequences
	(1)	(2)
T (=1)	0.024	−0.040
	(0.034)	(0.030)
Strong (=1)	0.041	−0.090*
	(0.040)	(0.035)
T×Strong	−0.027	0.106*
	(0.054)	(0.053)
Prior expectation	0.393***	0.585***
	(0.095)	(0.096)
In-person session (=1)	−0.037	−0.049
	(0.031)	(0.025)
Female (=1)	−0.048	−0.038
	(0.029)	(0.027)
Constant	0.380***	0.216***
	(0.064)	(0.057)
Control group	Pooled	Pooled
Number of subjects	203	198

Notes: This table displays OLS estimates of equation (4). Column 1 (2) shows estimates for the subsample of subjects who observed a good (bad) sequence. Robust standard errors are shown in parentheses.

^{*, **, ***}The symbols denote that p < 0.05, p < 0.01, and p < 0.001, respectively, using a two-tailed t-test.

Recall that one determinant of the larger average investment rate of treated subjects for strongly bad sequences was the higher prevalence of contrarian investment behavior (see Result 3). We tested whether such contrarian behavior was backed with a contrarian expectation of success, that is, a posterior expectation that exceeded the prior expectation despite the bad news. To this end, we estimated the following regression model using OLS:

$$C{E}_i={\delta }_0+{\delta }_1{T}_i+{\delta }_2\hspace{0.33em}Strong{}_i+{\delta }_3\left(T_i\times \hspace{0.33em}Strong{}_i\right)+Z_i^{\prime }{\delta }_4+{ϵ}_i\text{.}$$ (5)

The dependent variable $C{E}_i$ is a dummy variable that equals 1 if subject $i$ held a contrarian posterior expectation of success and 0 otherwise. (Online Appendix Figure A1 shows histograms of within-subjects expectation changes relative to the baseline for each observed sequence and group.) $T_i$ and $Stron{g}_i$ are the same as for equations (2), (3), and (4), whereas $Z_i$ is the same as for equation (4). Although we now focus on bad sequences, for completeness we also report the results for good sequences. (In the latter case, a posterior expectation was classified as contrarian if it was smaller than the prior expectation despite the good news.)

The results are shown in Table 5. Column 1 (2) displays the estimates for good (bad) sequences. Focusing on strongly bad sequences, the results indicate that, compared to control subjects, those in the Treatment group were more likely to hold contrarian expectations (p = 0.014 for δ₃ = 0 and p = 0.046 for δ₁ + δ₃ ≤ 0). This provides further evidence that the treatment effects on investment operated, at least in part, through changes in expectations of success.

Table 5:

Contrarian expectations of success

	Good sequences	Bad sequences
	(1)	(2)
T (=1)	−0.121	−0.159
	(0.078)	(0.085)
Strong (=1)	−0.031	−0.123
	(0.078)	(0.082)
T×Strong	0.034	0.320*
	(0.110)	(0.130)
Prior expectation	0.460***	−0.481**
	(0.141)	(0.158)
In-person session (=1)	0.066	−0.201**
	(0.065)	(0.064)
Female (=1)	0.124*	−0.026
	(0.055)	(0.060)
Constant	−0.041	0.611***
	(0.093)	(0.113)
Control group	Pooled	Pooled
Number of subjects	203	198

Notes: This table displays OLS estimates of equation (5). Column 1 (2) shows estimates for the subsample of subjects who observed a good (bad) sequence. Robust standard errors are shown in parentheses.

^{*, **, ***}The symbols denote that p < 0.05, p < 0.01, and p < 0.001, respectively, using a two-tailed t-test.

Result 4. The treatment effects on investment were mediated, at least in part, by changes in expectations of success.

5. Discussion

In this study, we experimentally investigated the effects of a fearful emotional state on financial investment in the presence of prior outcome information. Our results revealed a complex pattern of effects. We found that fear had differential effects on investment depending on whether the observed sequence of outcomes was favorable or unfavorable and contained weak or strong information. Specifically, compared to control subjects, those who were treated to experience a fearful emotional state invested the same amount after good news, a smaller amount after weakly bad news, and a larger amount after strongly bad news.

We highlight the following general conclusions. First, the common finding from previous research that fear tends to reduce risk-taking in the absence of prior outcome information (see Section 2.2) did not generally extend to situations in which individuals observed prior outcomes. In fact, we replicated this finding only among subjects who received weakly bad news about previous outcomes. Second, fear had an asymmetric effect on financial investment between situations with favorable news and those with unfavorable news: while good news counterbalanced the influence of fear, bad news did not. Third, when the news was unfavorable, the effect of fear on risk-taking was crucially affected by the strength of information. In particular, while weakly bad news was qualitatively similar to no news (the benchmark from previous studies), a reversal occurred after strongly bad news, when fear increased risk-taking among some subjects instead of reducing it.

We now elaborate on the effects of fear on investment when subjects observed strongly bad outcome sequences. The finding that fear raised the prevalence of contrarian decision-makers in this situation is consistent with prior experimental evidence that emotional distress sometimes increases risk-taking (Leith and Baumeister, 1996; Raghunathan and Pham, 1999; Lerner and Keltner, 2001; Fessler et al., 2004; Bruyneel et al., 2009). One possible explanation, which draws on the mood maintenance hypothesis (Isen and Patrick, 1983), is that people who are already experiencing fear and then confront strongly bad news sometimes find the news intolerable, and hence prioritize the immediate goal of feeling better over other goals. Thus, contrarian decision-makers in the Treatment group seem to have followed a strategy of “affective compensation,” seeking out a high-risk financial gain that would alleviate the distress resulting from the combination of negative affect and strongly bad news.⁷ Future studies may seek to better understand the mechanisms by which incidental fear compounded with negative outcome information can bias some individuals toward risk-seeking behavior.

A related question is what individual-specific factors are predictive of contrarian financial decisions. A minority in our sample exhibited contrarian behavior, but we speculate that this phenomenon could be more prevalent among individuals with psychopathology who engage in risky behavior in moments of distress (e.g., in substance use disorders, bipolar disorder, or oppositional defiant disorder). Another interesting question is whether financial contrarian behavior is related to specific personality traits or personality disorders such as borderline or antisocial personality disorder. The nature of our study precluded us from capturing information about individual differences. Further research could investigate the connection between decision-making sensitivity to emotional states and individual characteristics, including psychopathological dimensions.

Another finding of our study was that the effects of fear on investment following bad news were mediated, at least in part, by changes in expectations of success. Indeed, the rise in contrarian investment behavior under the influence of fear and strongly bad news was backed with a rise in contrarian expectations, as if distressed contrarians were unwilling to acknowledge the bad news. We note, however, that a direct influence of fear on investment—that is, an effect that did not operate through the expectations channel—cannot be disregarded. A common view in psychology states that emotions may influence risk-taking by directly shaping risk preferences or affecting the salience of potential outcomes regardless of individuals’ beliefs about their likelihood (see, for example, Loewenstein et al. (2001)).⁸ A formal mediation analysis would attempt to separate the total effect of our treatment on investment into the direct effect and the indirect effect mediated by expectations (Baron and Kenny, 1986). However, such separation is not possible if unobserved individual characteristics, such as general optimism, affected both expectations of success and investment. Therefore, we refrained from conducting a statistical mediation analysis.⁹

While our intervention appeared to be successful at inducing fear, we are limited by the fact that we relied entirely on participants’ self-report of their discrete emotional ratings. It is worth noting that measuring emotional states via a single item (word) for each emotion, as we did here, has been criticized for reduced reliability and validity across studies due to variability in the interpretation of words, even as straighforward as “happiness” (Weidman et al., 2017). We found that even though the most highly-rated emotion in the Treatment group was fear, the horror video clip also induced other negative emotions. One explanation is that the lack of specificity in the six items we used led to the mapping of a single emotion onto multiple items.¹⁰ We note that it is not possible to assert that this is the case here, as studies such as Weidman et al. (2017) have not been performed in our study’s language (Spanish). Furthermore, their thorough review reports no significant overlap between fear and any of the other three negative-valence items we employed (rage, disgust, and sadness), suggesting that our targeted emotion—fear—was likely to be independently elicited by the fearful components of the video. It is conceivable that the use of a componential scale or a single-item scale with the addition of a definition might have resulted in an even cleaner result.

Another limitation of our assessment of emotional state is that the accuracy of self-reported ratings is often hindered by heterogeneity in the participants’ insight and by expectations on their part about the nature of the experiment, which are hard to control for. Furthermore, there is evidence that individuals engage, sometimes involuntarily, both cognitive control and emotional regulation strategies when reflecting on their emotional state (Mauss and Robinson, 2009). With this in mind, we chose to ask participants to provide emotional ratings after concluding their investment decisions and reporting their expectations of success. While this reordering strategy delayed any instructed awareness of emotions, we could not control for the effects that the investment decisions and questions about expectations themselves may have had on participants’ emotional ratings. Recent studies have addressed this limitation by incorporating objective measures of emotional arousal and facial expression (Ben-Shakhar et al., 2007; Nguyen and Noussair, 2014; Lucarelli et al., 2015; Breaban and Noussair, 2018; Kassas et al., 2022). These measures complement self-report and yield a continuous monitoring of emotional correlates that could also be useful for controlling for the timecourse of the experimental emotional induction. Future investigations of the interplay between emotional states and information in decision-making under uncertainty may combine subjective and objective measures of emotion.

In closing, our study is a first empirical step towards an understanding of the interplay between incidental emotional states and prior outcome information in financial decisions. Our investigation was largely exploratory because a theoretical framework that explicitly incorporates both incidental emotions and previous outcomes has not been developed yet. While the nuanced behavioral patterns we reported here suggest that developing a unified framework is a challenging task, they also underscore that progress on the theoretical side is much needed. Hopefully, such progress will guide future empirical studies of decision-making under uncertainty in situations where decision-makers experience incidental emotions but also have information about prior outcomes.