The effects of sound in the Balloon Analogue Risk Task

Abstract

In this study, we examined the effects of pairing sounds with positive and negative outcomes in the Balloon Analogue Risk Task (BART). A number of published studies using the BART incorporate sounds into the task, where a slot machine or cash register sound is produced when rewards are collected and a popping sound is produced when balloons pop. However, some studies do not use sound, and other studies do not specify whether sound was used. Given that sensory information contributes to the intensity of experiences, it is possible that outcome-related sounds in the BART influence risk-taking behaviors, and inconsistent use of sounds across the many BART variations may affect how results are interpreted. Therefore, the purpose of this study was to investigate the effects of sounds paired with outcomes in the BART, and whether the presence or valence of a sound would systematically alter participants’ risk-taking. Across two experiments using Bayesian censored regressions, we show that sounds, regardless of the outcomes they were paired with or their valence, did not affect risk-taking in an adult, non-clinical sample. We consider the implications of these results within methodological and theoretical contexts and encourage researchers to continue dissociating the role of auditory stimuli in feedback processing and subsequent responding.

Introduction

Individuals must balance prospective risks and rewards when making choices. Pursuing opportunities to acquire increasingly larger rewards often comes at the risk of decreasing the probability of attaining those rewards or even incurring punishments. For example, the longer a hunter waits for a better shot, the greater the likelihood that the prey will flee. However, if someone is too reluctant to engage in risk, they may be settling for less reward in the long term. Researching risky behavior and modeling deviations from reward-maximizing levels of risk has been a popular topic of examination because it has both clinical and economic implications. For example, risk-seeking tends to be highly correlated with experiencing injuries (Cherpitel, 1999; Ryb et al., 2006), drug and alcohol abuse (Bickel et al., 1999; Bickel et al., 2006; Fernie et al., 2010; Hopko et al., 2006; Lejuez et al., 2002; Reynolds, Richards, & de Wit, 2006b; Richards et al., 1999; Robles et al., 2011), and pathological gambling (Petry & Cassarella, 1999; Reynolds, Ortengren, et al., 2006a).

Although decision-making research has largely entailed presenting participants with independent gambles or events to derive risk propensities (Kahneman & Tversky, 1979; Tversky & Kahneman, 1992), daily life involves repeated opportunities for engaging in risky behavior and experiencing its consequences, thus allowing people to learn from those consequences (e.g., speeding, sleeping late, not backing up a computer, etc.). Therefore, measuring risk-taking as choices over sequential trials is an important methodological tool for understanding how behavior reflects the environment (Bechara et al., 1994; Lejuez et al., 2002; Mishra & Lalumière, 2010; Reilly et al., 2006; Schreiber & Dixon, 2001; Thaler & Johnson, 1990).

A common experimental method used to measure sequential risk-seeking is the Balloon Analogue Risk Task (BART; Lejuez et al., 2002), a decision-making game where participants inflate a series of computerized balloons to obtain reinforcement (e.g., cash or points). The larger the balloon is when the participant stops inflating it (i.e., when they cash in) the more points they earn, but larger balloons also carry a greater risk of popping – if inflated too far – and yielding zero points. Participants who inflate the balloons to larger sizes are judged to be more risk-seeking. The BART can therefore serve as an elegant representation of the dynamic risk-reward tradeoff people experience on a daily basis, and risk propensities as measured by the BART positively correlate with self-reports of impulsivity and sensation seeking (Bornovalova et al., 2009; Holmes et al., 2009; Lejuez et al., 2002; MacLean et al., 2018; Mishra et al., 2010; Mishra & Lalumière, 2010; Vigil-Colet, 2007), risky sexual behavior (Lejuez et al., 2004), alcohol and substance abuse (Aklin et al., 2005; Crowley et al., 2006; Hopko et al., 2006; Lejuez, Aklin, Jones, et al., 2003a; Lejuez et al., 2002) and participating in extreme sports (Keller et al., 2021). Additionally, responding on the BART is positively associated with risk taking in similar laboratory tasks (Mishra & Lalumière, 2010; Reynolds, Ortengren, et al., 2006a; Robles et al., 2014).

However, the BART is not without its limitations; the relationship to real-world risk behaviors is often small (e.g., correlations around .10; see Lauriola et al., 2014 for a review), and other research has revealed mixed or no evidence that clinical behaviors are associated with increased responding on the BART (Ashenhurst et al., 2011; Campbell et al., 2013; Dean et al., 2011; Ryan et al., 2013). Nonetheless, the BART remains an extremely popular and widely used behavioral method (and clinically relevant diagnostic tool) that still has its place in risk-taking research today.

Many variations of the BART have been used to understand the environmental variables that might alter risk-seeking, from manual or automatic inflation of the balloons, balloon points being hypothetical or worth actual monetary rewards, or manipulating the probability that the balloon will pop (e.g., Lejuez et al., 2002; Pleskac et al., 2008; Robles, 2015; Young & McCoy, 2019). However, an unexamined component is the inclusion of sound when experiencing feedback. In Lejuez et al.’s (2002) original study, a popping sound was produced when the balloon popped, and a slot machine sound was produced when the participant cashed in the balloon. While subsequent research often has followed this design (Lejuez, Aklin, Jones, et al., 2003a; Lejuez, Aklin, Zvolensky, & Pedulla, 2003b; Lighthall et al., 2009; Lighthall et al., 2012; Pleskac et al., 2008), other research has not incorporated sound into the game’s consequences (Fein & Chang, 2008; McCoy, 2015; Robles, 2015; Robles et al., 2014; Young & McCoy, 2019). Moreover, many other studies do not specify whether they used sound or not (Acheson et al., 2007; Ashenhurst et al., 2014; Killgore et al., 2011; Rao et al., 2008; Reynolds, Ortengren, et al., 2006a).

Although the discrepancy in sound use across BART designs may seem trivial, it has important theoretical and methodological implications. The world teems with stimulating sensory information that contributes to the intensity of experience and heightens sensitivity to both the anticipated and experienced consequences of actions (Bechara et al., 1994; Bechara & Damasio, 2005; Damasio et al., 1991). Sound therefore may sensitize individuals to the environment’s contingencies, increasing their arousal levels and reactivity to outcomes (Slovic et al., 2007), or even act as a second-order conditioned stimulus that adds to the reinforcement of play (Parke & Griffiths, 2006; Schull, 2005) and arousal (Dixon et al., 2014). Interestingly, although casino sounds can increase emotional and physical arousal, pleasure, and gambling intentions (Dixon et al., 2014; Finlay et al., 2007; Loba et al., 2002), they did not influence actual risk-taking in the Iowa Gambling Task (Brevers et al., 2015), suggesting there may be discrepancies between the effects of sound on feedback processing and subsequent responding.

If, however, outcome-related sounds do significantly impact learning and choices beyond feedback processing, it will be important to understand their function in risky decision-making, and to incorporate this impact into theories and experimental methodologies. A consequence of neglecting this element in the BART is that participants’ behavior may be confounded across sounds and outcomes experienced – is the popping aversive because of the failure to earn points, the aversiveness of the sound, the disappearance of the balloon, or some combination of these factors? This varying use of sound could create inconsistencies across the large body of research using the BART because response patterns observed in some studies may be due, at least in part, to sounds accompanying the outcomes. Moreover, designs with sensitive procedures, such as neuroimaging (Cazzell et al., 2012; Chen & Wallraven, 2017; Crowley et al., 2009; Fein & Chang, 2008; Fukunaga et al., 2012; Kessler et al., 2017; Rao et al., 2008; Schonberg et al., 2012), or studying at-risk populations (Cheng et al., 2012; Crowley et al., 2006; Hopko et al., 2006; Kelley et al., 2010; Killgore et al., 2006) may be especially vulnerable to sound. The presence of sound can add noise to the neural signal collected during the task, possibly startle subjects who abuse substances, suffer from mental disorders or have experienced trauma, or stimulate pathological gamblers to act more riskily.

Therefore, the purpose of the present study is to examine whether sounds in the Balloon Analogue Risk Task influence participants’ response patterns. In Experiment 1, we tested the effects of sound on responding at a more molecular level. The effect of each sound was isolated such that a participant either received a sound when the balloon was cashed in or when it popped – but not both. The particular sound experienced was either consistent with the outcome (e.g., popping sound when the balloon popped) or inconsistent (e.g., popping sound when the balloon was cashed in). Our goals were to determine the degree to which pairing a sound with an outcome might alter risky choice and whether those effects were specific to a consistently valenced sound. Because we found no evidence of these manipulations altering risky behavior, in Experiment 2 we used a stronger manipulation in which a participant either experienced consistent sounds for both outcome types (the common design of the BART) or experienced no sounds in the task.

Experiment 1

Experiment 1 used a 2 × 2 between-subjects design with the individual outcome (cash-ins vs. pops) crossed with the sound’s valence (positive cash register sound vs. negative pop sound) resulting in four groups: 1) a cash register sound played when a balloon was cashed in, 2) a cash register sound played when a balloon popped, 3) a pop sound played when a balloon popped, and 4) a pop sound played when a balloon was cashed in. Note that only one type of outcome was followed by a sound in each condition; the other outcome produced no sound.

We generated a series of hypotheses regarding the potential effects of the sound’s outcome and valence on the participants’ propensity to cash in balloons in Experiment 1. Our expectations were based on subjects’ desires to delay and reduce the probability of aversive events (the popping sound) and to hasten and increase the probability of desirable events (the cash-register sound). The consequences of these hypotheses are illustrated in Table 1. Hypotheses 1 and 2 create some ambiguity regarding the effects of the sound because an increase in pumping may create a desirable outcome of one type (e.g., delaying an aversive sound) while also creating an undesirable outcome of another type (e.g., increasing the likelihood of an aversive sound). The clearest predictions are for Hypotheses 3 and 4 in which the effects of the sound on the timing and probability of an event are in the same direction. If subjects hear a ‘pop’ sound when cashing-in balloons, increasing the number of pumps will delay and decrease the probability of hearing this aversive sound (Hypothesis 3). Conversely, if subjects hear a ‘kaching’ sound when cashing-in balloons, decreasing the number of pumps will hasten and increase the probability of hearing this desirable sound (Hypothesis 4). Finally, Hypothesis 5 predicted that there would be no difference in responding as a function of the outcome and valence conditions, and all groups will cash in balloons at similar pump values. We explicitly included a null effect as a competing hypothesis because subjects may not show any sensitivity to the sound manipulations and Bayesian analyses (see Methods below) allow us to estimate this likelihood, thus providing direct statistical evidence in favor of (or against) Hypothesis 5. Whereas the first four hypotheses are not competing and are specific to each of the four conditions, the final hypothesis is clearly inconsistent with the first four.

Table 1.

Hypothesized consequences for adding sound to outcomes

		Hasten Positive	Increase P(Positive)	Delay Aversive	Reduce P(Aversive)
Hypothesis 1	Pop = “Pop”	-	-	↑	↓
Hypothesis 2	Pop = “Kaching”	?	↑	-	-
Hypothesis 3	Cash-in = “Pop”	-	-	↑	↑
Hypothesis 4	Cash-in = “Kaching”	↓	↓	-	-
Hypothesis 5	All conditions	-	-	-	-

↑ indicates predicted increase in number of pumps; ↓ indicates predicted decrease in number of pumps; ? indicates uncertainty regarding effect; - indicates no effect on pumping

Methods

Participants

A total of 174 adults were recruited from Amazon’s Mechanical Turk (MTurk) Cloud Research (formerly TurkPrime) participant pool and provided a link to Qualtrics where they first read an informed consent form. If participants consented, they were provided task instructions and then redirected to Pavlovia (www.pavlovia.org), an online platform for running PsychoPy-based experiments (www.psychopy.org), to complete the BART. All study methods were approved in advance by the local IRB. Participants were compensated a flat fee of $2 for completing the study; no performance bonus was given. We removed 15 subjects for either failing an attention check (n = 8) or low effort responding (operationalized by averaging fewer than ten pumps before cash-ins and not showing an increase in responding throughout the task; n = 7). A subset of participants (n = 3) were classified as low effort because two did not cash in any balloons and another cashed-in only three balloons. An additional 14 subjects were removed due to a programming error where sounds were not played when balloons were cashed-in or popped. Our final sample size was 145 (62 females, 82 males, one non-binary), aged 20–71 years (M = 40, SD = 12).

Procedure

Participants completed 30 trials of the BART (Fig. 1) where they manually inflated computer-simulated balloons to accumulate hypothetical points by pressing the ‘spacebar’ key. Participants could press the ‘enter’ key to store the balloon’s accumulated points in a permanent bank – this represented cashing-in. Every manual pump increased the physical size of the balloon stimulus and its value by one point, but each pump also increased the probability of the balloon popping and losing its point value:

$$ProbabilityofPopping=\frac{1}{128-n+1},$$ (1)

where is the current number of pumps for the balloon (Lejuez et al., 2002). Using Eq. (1), the maximum number of pumps for each balloon is 128, with 64 pumps optimizing the total points accumulated by the end of the study, although few studies have ever reported subjects reaching this point (Pleskac et al., 2008 observed near-optimal pumping, but they also explicitly informed subjects of this strategy). After each pump, a random number (i.e., uniformly distributed from 0–1) was generated and the balloon popped if this number was less than the pop probability derived from Eq. (1). We seeded the random number generator (see Supplementary Materials for more details) so that every subject started with the same seeding, however what they experienced on each trial was a byproduct of their behaviors and the effects of the seed diverged as the task progressed.

Fig. 1.

Screenshots of the BART used in Experiments 1 and 2. The balloon starts off at a small, non-zero size and increases in both physical size and point value as the participant inflates it, until the balloon is either cashed in (top right) or popped (bottom right)

The points stored in the bank (i.e., points obtained from cashing-in previous balloons) were not lost by popping a balloon. When a balloon was cashed-in, it disappeared from the screen and positive feedback and the point value obtained were visually displayed. When a balloon popped, it disappeared, and negative feedback was visually displayed. Feedback was displayed for 1500 ms, after which a new balloon was presented (see Fig. 1). The main objective of the BART is to accumulate as many points as possible.

We added two sound checks to ensure subjects had their volume on and were attending to the task. Each sound check consisted of a 3–4-s-long sound clip presented immediately following the offset of feedback on the 10th and 20th balloons and then asking the subjects to identify the sound from a list of four options. The sound clip following the 10th balloon was a dog barking, and the sound clip following the 20th balloon was ocean waves on a beach. While failing these attention checks did not terminate the task, any subjects failing them would have their data excluded from the final analysis.

Participants were randomly assigned to one of four groups in the 2 × 2 design: In the first group (n = 34), a cash register sound played when a balloon was cashed in. In the second group (n = 37), a cash register sound played when a balloon popped. In the third group (n = 36), a pop sound played when a balloon popped. Finally, in the fourth group (n = 38), a pop sound played when a balloon was cashed in.

The BART was programmed in PsychoPy (Peirce et al., 2019) and run online using Pavlovia. Because the experiment was hosted online, we could not strictly control the presentation of the sound stimuli, so it is likely that the volume and delivery (e.g., headphones, computer speakers) of the sounds differed across subjects. It is also possible that subjects were distracted by working on other tasks or even listening to other background sounds produced by music players, a TV or the general environment in which they completed the task. However, Cloud Research has been shown to yield high-quality data (Peer et al., 2022) and visual inspection of the raw response data did not generate concerns about validity.

Statistical analyses

The conventional method for assessing risk-taking in the BART is to average participants’ number of responses for all unpopped balloons throughout the entire task, referred to as mean adjusted pumps (MAPs; Lejuez et al., 2002). MAP scores can then be compared across groups using t tests or ANOVAs. While using MAPs to infer risk-taking propensities is statistically convenient, it results in two significant problems. First, MAPs ignore changes in behavior across the task. Given that the BART is a dynamic decision-making task in which participants can learn the contingencies, recording changes in participants’ responding as a function of experiencing reinforcement and punishments is invaluable. Furthermore, prior research has evidenced learning effects in the BART (Jackson et al., 2013; Lejuez et al., 2002; Pleskac & Wershbale, 2012; Robles, 2015; Robles et al., 2014; Wallsten et al., 2005). In the current analysis, multilevel (i.e., repeated measures) regressions were run on the unaggregated number of responses for each balloon to model changes in behavior throughout the task.

The second problem with using MAPs as the dependent variable in the BART is that discarding trials in which the balloon popped reduces the sample size of responses and omits useful information about participants’ risk-propensities that should be accounted for in the analyses (Coon & Lee, 2021; Pleskac et al., 2008; Young & McCoy, 2019). On popped trials we know that participants were willing to respond at least up to the point at which the balloon popped. However, when a balloon pops, the participant’s behavior is interrupted, creating uncertainty as to when exactly they would have cashed in the balloon – the participants could have cashed in the balloon one pump later or perhaps even 20 pumps later. The disrupted cash-in values are considered to be “censored” and still do not correctly model the upper tail of the response distribution, again resulting in downward-biased pump values compared to subjects’ intended inflation rates¹ (Pleskac et al., 2008). Censored regression (Tobin, 1958) can be used to analyze censored responses by factoring the participants’ observed data and the number of censored trials into the model’s predictions, and Young and McCoy (2019) have recently demonstrated how utilizing this analysis within a Bayesian framework can more accurately model participants’ full distribution of responses in the BART.

To test our hypotheses and correct for the biases introduced by measuring MAP scores in the BART, we used a Bayesian analysis to conduct a generalized linear multilevel censored regression using the brms package in R (Bürkner, 2017). We specified a Poisson distribution (with a log-link function) to account for the strong positive skew in the marginal distributions of responses at cash-in (which are count data). Responses were predicted as a full factorial function of the sound’s outcome, the sound’s valence, and balloon (i.e., trial). Balloon was included in the regression model to isolate a partial source of variance, namely whether responding changed over the course of the task. The intercept and slope estimate of balloon were allowed to vary across participants as random effects, which allows us to individually estimate each participant’s average response rate and the degree to which their responding uniquely changes across balloons. Balloon was mean-centered and the outcome and valence groups were effect-coded to minimize multicollinearity and make the model intercept represent the grand mean of responses.

For the Bayesian components of the regression model, trials were classified as censored when the balloon popped (and the pop point of the balloon was then assigned as its response value). We set a burn-in period of 1000 iterations with an additional 15,000 saved iterations to estimate the posterior distribution of each parameter and used a thinning rate of 10 to decrease autocorrelation in the sampled values; four chains were run to derive the parameter estimates. The larger burn-in period, iterations, and thinning rate were implemented to mitigate warnings about model convergence related to subpar sampling of the posterior distributions (see Results). Based on the MAPs observed in previous studies (Table S1, available at https://osf.io/ztuwp/), we specified a moderately informed prior of N(3.5, 0.25) for the intercept. This intercept prior suggests that participants on average will pump balloons about 33 times before cashing in, with approximately 95% of means ranging from 20 pumps to 54 pumps. We allowed a moderate amount of variance on the intercept prior to account for the additional uncertainty due to our use of censored regression. To test for sensitivity to intercept priors, a variety of similar values were also evaluated (e.g., N(3.25, 0.25) and N(3.75, 0.5)) to encompass the range of MAPs observed across BART studies. All tested intercept priors returned comparable estimates and conclusions, confirming that the decided prior is suitable and that data are driving the predicted results. We specified a weaker prior of N(0, 0.1) for all other regression weights, Cauchy(0, 0.5) for the standard deviations (i.e., random effects for individual differences) and LKJ(2) for the random effect correlation. The priors for all regression weights do not initially assume there will be an effect of the sound conditions or their interactions, but these priors still allow for a difference of about 15 pumps in either direction between the four sound groups.

We also computed a Bayes factor (BF₁₀; Morey et al., 2016) using the bayestestR library to estimate the relative evidence in favor of the full factorial model over a simpler model that estimated responses from only the balloon (trial) number. Using Bayes factors allows us to make additional statistical inferences about the sound conditions on responding that avoids issues associated with p values (Jarosz & Wiley, 2014).

Results and discussion

All 145 participants completed all 30 trials of the BART. We removed 29 individual data points when balloons were cashed in without pumping at all representing about 0.6% of the remaining data. Inspections of the zero-pump trials indicated subjects were likely accidentally double-pressing the enter key following cashing in the previous balloon rather than systematically engaging in null responding, so no other data were removed.

The model output issued a series of warnings about convergence regarding high R-hat values, poor mixing of the sampling chains and low effective sample sizes (ESSs). However, visual inspection of the posterior distributions, R-hat values (all below 1.05), sampling chains and ESSs did not reveal any concerns about the reliability of the parameter estimates. Some of the posterior distributions were slightly flattened out or misshaped instead of peaked, which may have posed difficulties for the model to identify the most likely parameter value. Additional models (see https://osf.io/ztuwp/) tested different priors, response distributions and fixed- and random-effects structures to address these warnings, but none of these approaches fully remedied the convergence issues, so we continued with our reporting of the specified model with these caveats in mind.

Figure 2 illustrates the model-predicted responses at cash-in, showing that there does not appear to be any difference in average responding between the outcome and valence groups. Additionally, all groups increased their responding as the session progressed. Table 2 shows the Bayesian posterior parameter estimates of the regression model, and the back-transformed model-predicted averages are discussed next (see Figure S2 for the posterior distributions of the model-predicted responses for each sound group).

Fig. 2.

Model-predicted responses as a function of the sound presented during feedback (left panel = cash register sound and right panel = pop sound), the outcome the sound was paired with (blue solid curves = cash-ins and red dashed curves = pops) and balloon. Shaded regions represent ± 1 standard error

Table 2.

Experiment 1’s parameter estimates and 95% credible intervals. All estimates are in the log-transformed space

	B	95 % CI
Intercept	3.547	[3.478, 3.616]
Balloon	0.010	[0.008, 0.013]
Outcome	− 0.006	[− 0.073, 0.062]
Valence	− 0.024	[− 0.092, 0.044]
Balloon × Outcome	0.001	[− 0.003, 0.003]
Balloon × Valence	0.001	[− 0.002, 0.003]
Outcome × Valence	− 0.001	[− 0.069, 0.067]
Balloon × Outcome × Valence	0.001	[− 0.002, 0.003]

Balloon was centered by subtracting 15.5. Outcome was effect-coded as [Cash-in = +1; Pop = −1] and Valence was effect-coded as [Cash register sound = +1; Pop Sound = −1]

Participants who heard a cash register sound when cashing-in balloons were estimated to cash in balloons after about 33.6 pumps (95% HDI = [29.1, 38.4]), participants who heard a cash register sound when popping balloons after about 34.1 pumps (95% HDI = [29.7, 38.8]), participants who heard a pop sound when popping balloons after about 35.7 pumps (95% HDI = [31.2, 40.9]), and participants who heard a pop sound when cashing in balloons after about 35.3 pumps (95% HDI = [30.8, 40.2]); there was little-to-no evidence that the average number of pumps at cash-in differed between these groups, supporting Hypothesis 5. Participants did increase their average responding as the session progressed (mean slope = 0.010, 95% CI = [0.008, 0.013]), but there was no difference between any of the outcome or valence groups in their slopes. Finally, we found a Bayes factor (BF₁₀) of less than 0.001 (i.e., B F₀₁ > 1000), providing extremely strong evidence in favor of the simpler model (where responses were predicted by balloon only) and affirming that the sound manipulation did not cause differential responding for any of the four experimental conditions (see Jarosz & Wiley, 2014, for interpreting Bayes factors).

These results revealed that responding in our implementation of the BART task was not influenced by the individual outcome with which the sounds were paired or the sounds’ valence. However, we did not directly compare the outcome and valence groups to a no-sound group that did not experience sounds for either the cash-in or popped-balloon outcomes. Moreover, it may be that sound does not alter responding when only paired with one outcome but may have an additive effect when paired with both outcomes. Therefore, in Experiment 2, we assessed responding in the BART as a function of whether sounds were paired with both cash-ins and pops or were absent from feedback.

Experiment 2

In Experiment 2, we created a simpler manipulation wherein participants were randomly assigned to one of two groups: 1) a sound group, where a cash register sound played when a balloon was cashed-in and a pop sound played when a balloon popped (the typical use of sound in the BART), and 2) a no-sound group, where no sound played when a balloon was cashed in or popped. We did not manipulate the sounds’ valence as in Experiment 1.

We generated three hypotheses about participants’ propensity to cash in balloons for Experiment 2. Hypothesis 1 predicts that participants in the sound group will respond less on average than participants in the no-sound group, in order to increase the probability of hearing the cash-register sound and decrease the probability of hearing the pop sound. Note that this hypothesis integrates Hypotheses 1 and 4 from Experiment 1. Hypothesis 2 predicts that participants in the sound group will respond more on average than participants in the no-sound group thus indicating greater sensitivity to the outcomes; as responding approaches 64 pumps at cash-in, the expected value of accumulated points is maximized. Finally, Hypothesis 3 predicts that there will be no difference in responding between the sound and no-sound groups, consistent with Experiment 1’s results.

Participants

A total of 82 adults participated in Experiment 2 and were recruited and instructed in the same manner as Experiment 1. Subjects were again compensated a flat fee of $2 for completing the study and did not receive any performance bonus. Four subjects were removed due to failing an attention check. Two subjects repeated the entire task, so we removed their second set of responses. Our final sample size was 78 (28 females, 49 males, one preferred not to say) aged 21–62 years (M = 36, SD = 9).

Procedure

Experiment 2’s procedure was identical to Experiment 1’s, except that participants were now randomly assigned to one of two groups for the BART: In the sound group (n = 41), a cash-register sound played when a balloon was cashed-in and a pop sound played when a balloon popped. In the no-sound group (n = 37), no sounds played when balloons were cashed-in or popped.

Statistical analyses

We again conducted a Bayesian censored multilevel Poisson regression where responses were predicted as a full factorial function of the sound condition, balloon (i.e., trial) and their interaction. The model intercept and slope estimate of balloon were allowed to vary across participants as random effects. Balloon was mean-centered and the sound condition was effect-coded to minimize multicollinearity and so that the model intercept represented the grand mean of responses. A main effect of the sound condition, where participants in the sound group cashed-in balloons with fewer responses on average than the no-sound group would support Hypothesis 1. If the main effect of the sound condition instead showed that participants in the sound group cashed in balloons with more responses on average than the no-sound group, this would support Hypothesis 2. A lack of a main effect of sound condition, where both groups are responding equivalently, would support Hypothesis 3.

The Bayesian components of the regression model were similar to the values specified in Experiment 1, except we set more informative priors on the model intercept, N(3.5, 0.1), and the balloon slope, N(0.01, 0.005). These updated priors come directly from the posterior estimates in Experiment 1, but with additional variance to account for them being derived from the previous study alone.

Results and discussion

All 78 subjects except one completed all 30 balloons. One subject only completed 25 balloons, but we retained their data because multilevel models can handle incomplete and imbalanced data. A total of 17 data points were removed due to balloons being cashed in with zero pumps representing 0.7% of the data for the retained subjects.

The model output again issued a series of warnings about convergence regarding the posterior estimates given high R-hat values and low ESSs. However, visual inspection of the posterior distributions, R-hat values (all below 1.05), sampling chains and ESSs did not reveal any concerns about the reliability of the estimates, so we continued with our reporting of the specified model.

Figure 3 illustrates the model-predicted responses, showing that there does not appear to be a substantive difference in average responding between the sound and no-sound groups. Additionally, both groups increased their responding as the session progressed as expected if participants optimize their behavior with experience. Table 3 shows the Bayesian posterior parameter estimates of the multilevel Poisson regression, with the back-transformed model-predicted averages presented next (see Figure S2 again for the posterior distributions of the model-predicted responses for each sound group).

Fig. 3.

Model-predicted responses at cash-in as a function of sound condition (blue solid curves = no sound and red dashed curves = sound) and balloon. Shaded regions represent ± 1 standard error

Table 3.

Experiment 2’s parameter estimates and 95% credible intervals. All estimates are in the log-transformed space

	B	95% CI
Intercept	3.508	[3.414, 3.604]
Balloon	0.009	[0.006, 0.012]
Sound	−0.015	[−0.108, 0.082]
Balloon × Sound	0.001	[−0.002, 0.005]

Balloon was centered by subtracting 15.5. Sound was effect-coded [No Sound = +1 and Sound = −1]

Participants in the sound group were estimated to cash in balloons after about 33.9 pumps (95% HDI = [29.7, 38.5]), and participants in the no-sound group were estimated to cash in the balloon after about 32.8 pumps (95% HDI = [28.6, 37.7]); the estimated difference (M = − 1.0, 95% HDI = [− 7.3, 5.3]) provides evidence that sound did not influence responding, supporting Hypothesis 3. Moreover, participants on average increased their responding as the session progressed, but there was little evidence that the sound and no sound groups differed in their slopes. Finally, we found a Bayes factor (BF₁₀) of 0.014 (i.e., BF₀₁ = 71), again providing strong evidence in favor of the simpler model that excludes sound as a source of variance, thus affirming that the sound and no-sound groups did not differ in responding.

General discussion

Overall, the results from these two experiments show that sounds do not materially influence risk-taking in the BART, regardless of which outcome(s) they were paired with or their valence. These findings demonstrate that the variability in the use of sound across BART studies is not likely to pose a methodological concern for adult, non-clinical samples.

While there are always numerous plausible explanations for null results, we believe our findings are due to a lack of effect of sound in the task. First, modeling pumps across balloons using a generalized linear multilevel regression correctly accounts for individual differences and positive skew in the response data that can violate assumptions of the general linear model (Bolker, 2015; Young & McCoy, 2019) thus minimizing bias in the parameter estimates and our conclusions. Next, running our statistical model within a Bayesian framework (Coon & Lee, 2021; Young & McCoy, 2019) provides direct evidence of null results by estimating the likelihood of the model parameters being zero (rather than only providing a p value indicating if the parameter estimates are significantly different than zero) and using Bayes factors to show that there was decisive evidence in favor of simpler models that excluded sound as a source of variance.

We also do not believe that the online nature of our task nor the participant sample biased responding. Even though we could not strictly control subjects’ off-task behavior (e.g., being distracted during the task by other online or offline activities, or even potentially looking up information on the BART) and sound settings in the way that we can in a laboratory setting, subjects passed the sound checks and increased their responding towards optimal across the 30 balloons, which indicates sufficient engagement in the task. Although the current study was not conducted to test the validity of an online sample, our findings show that the Internet-based design and Cloud Research sample effectively reproduced risk-taking behavior observed in BART studies conducted in-person (see the Supplementary Materials for a comparison of our results to prior results from laboratory settings, https://osf.io/ztuwp/).

Limitations and future directions

There were several limitations in these experiments. First, the lack of model convergence in the experiments led to an interesting investigation into the distribution of pop values across balloons. We discovered that a portion of balloon pops were actually occurring earlier in the task than originally anticipated (Figure S1, https://osf.io/ztuwp/), particularly before ten pumps had been reached. Although the number of observed pops for each pump value did not differ from the expected number given the pop algorithm (Lejuez et al., 2002), the distribution of pops may cause problems when estimating the posterior parameters of a censored regression. Future studies may instead prefer to use a superellipsoid function for the pop algorithm (Young & McCoy, 2019) or various alternative algorithms to better reflect naturalistic inflation contingencies (Steiner & Frey, 2021).

In Experiment 1, we did not compare responding in the outcome and valence groups to a no-sound group. In Experiment 2, we did compare a sound group to a no-sound group, but doing so prevented us from determining whether these two groups differed from the groups in Experiment 1. Future studies could therefore directly compare responding in the BART when sounds are paired with both outcomes, only the individual outcomes or when sounds are not paired with any outcomes to test the full range of experimental combinations. However, we consider it unlikely that a significant impact of sound would emerge given the strong evidence in favor of a null effect in Experiments 1 and 2.

A more fruitful follow-up might use sounds with higher amplitudes, but such a design would go beyond the amplitudes historically used in the BART and introduces a risk of hearing damage. A study by Notebaert et al. (2016) documented that participants took action (invested a coin) to avoid the presentation of a loud noise on fewer than 10–15% of trials when the sound was 70 dB (the sound of a vacuum cleaner) or less. Actions to avoid the noise began to noticeably increase only when the noise was 80 dB (the sound of a typical alarm clock) or louder and when these noises occurred with a high probability (80–100%). Note that 85 dB is considered the threshold beyond which a sound can be harmful. The popping of a real balloon has been recorded as 145 dB or louder (Hodgetts & Scott, 2021) which is well beyond the sound levels used in the laboratory. Repeated exposure to such levels is unlikely to merit approval by an Institutional Review Board.

Lastly, our study design varied slightly from how the BART is typically presented (see Table 4). First, we only compensated participants with a flat fee and balloons were worth hypothetical points instead of including performance bonuses based on how much each balloon was inflated. Second, we did not display the participant’s cumulative points or the points obtained on the previous balloon (e.g., Lejuez et al., 2002; Robles, 2015; cf. Young & McCoy, 2019). A study by Ferrey and Mishra (2014) showed that the compensation method for participants can affect risk-taking in the BART, namely that providing a flat fee plus performance bonus resulted in more pumping of balloons on average than only performance bonuses or no payment at all, but we have not found any literature on the effects of displaying participants’ cumulative points and the previous balloon’s points. Although we do not know of any compelling rationale for why these variations would differentially influence risk-taking for the sound groups, it could be that sound becomes important when participants are cued to their current performance and motivated to maximize their potential winnings. Future research could therefore add in further experimental manipulations of the BART’s task attributes and financial compensation types, although it may be more beneficial to couch this investigation within a broader theoretical context studying the interplay between different types of reinforcement contingencies and the environmental stimuli that accompany them.

Table 4.

BART task attributes and our preparation’s similarities/differences with typical implementations

Attribute	Our task	Typical task
Balloon manually inflated	Yes	Yes
Total points displayed	No	Yes
Previous balloon’s points displayed	No	Yes
Balloon outcome feedback	Yes	Yes
Outcome-based soundsa	Yes & No	Mixed
30 balloons total	Yes	Yes (but not always)
128 pump maximum with 64 being optimal	Yes	Yes (but not always)
Financial compensation	Flat fee and hypothetical points	Mixed

^aThe current study’s experimental manipulation

Conclusions

Auditory stimuli provide important environmental information that contribute to the intensity of experience and potentially amplify the hedonic value of outcomes. Many published studies measuring risk taking using the BART have incorporated sound into the design, but doing so may confound whether more or less risk-taking is due to experiencing the task’s outcomes, the sounds, or some combination of both. In the present study, we demonstrated that risk preferences in the BART are not materially affected by outcome-related sounds or the sounds’ valence for an adult, non-clinical sample with mild-to-moderate sound amplitudes. However, given the BART’s popularity for assessing the behavioral risk characteristics in clinical samples (e.g., substance abusers, trauma patients), we cannot say whether such populations will be similarly unaffected. Moreover, the BART is also used to understand the neural signatures of error and feedback processing (Fein & Chang, 2008; Kessler et al., 2017; Rao et al., 2008). With prior evidence that sounds paired with outcomes can enhance feedback-based arousal (Dixon et al., 2014), it is possible that neural signals may be influenced by the sounds through increased physiological arousal. This result has important theoretical implications because neuroimaging studies could help dissociate the impact of sound on feedback-based arousal and subsequent responding. We encourage researchers to further investigate the effects of sound on risk-taking in clinical and neuroscientific domains. However, until there is evidence that sound may have a measurable impact on risk taking under certain conditions, researchers can safely assume that their removal has no discernible impact beyond removing their negative side effects on the processing of neural signals.