Discussion
The present study on decision-making under risk and ambiguity investigated the decisions of participants under different levels of uncertainty. We aimed to contrast risk and ambiguity on the behavioral and the neural level by experimentally varying these two types of uncertainty. Subjects were exposed to three different levels of risk: i.e., low, high, and mixed-risk and two levels of ambiguity: i.e. high and low ambiguity. Probabilities and outcomes were composed such that the expected value for all decisions was the same. Further, we defined the different levels of risk based on a suggestions of Rothschild & Stiglitz (1970) as the spread of potential outcomes. On the behavioral level, data confirmed classical effects of risk and ambiguity (see Ellsberg, 1961; Kahneman et al., 1997; Kahneman & Tversky, 1979; Knight, 1921). Subjects behaved significantly more cautiously when they had to decide between one high-risk option and one low-risk option. In addition, we found an effect of ambiguity on decision-making. High ambiguity – i.e., during a condition where any probability information was absent – led to the highest percentage of cautious decisions.
On the neuronal level, we found increased activity in the parahippocampal region during mixed-risk trials compared to high- and low-risk trials in the risk block indicating an effect specific for the risk difference between options. This region is closely linked to the amygdala (e.g. Roy et al., 2009; Stein et al., 2007) and part of the network that became activated during loss aversion in previous research (e.g. Canessa et al., 2013) and we are thus able to replicate earlier findings.
In the ambiguity block we found activation of the supplementary motor area (SMA), which we showed for mixed-risk trials with low ambiguity in comparison to high- and low-risk trials with low ambiguity. SMA activation has been repeatedly related to reinforcement signals that may be used to guide decision-making and action planning under risk and uncertainty. Recently, Canessa and colleagues (2013) found greater activation in SMA for individuals with greater loss aversion. Our data also corroborated earlier findings that activity of the insular plays an important role in the emotional evaluation of ambiguous decision in conditions for which information about probabilities of options are missing.
Higher ambiguity activated the dorsolateral prefrontal cortex (DLPFC). Activation of this region was previously found to be related to ambiguity (Krain et al., 2006). This activation is likely related to rule-based action selection and likely indicates the conscious or cognitive evaluation and manipulation of the relevant information in working memory and the integration of negative evaluative emotional information from the amygdala under ambiguous choice conditions. In addition, the inferior parietal cortex (IPC) was during mixed trials with high ambiguity in comparison to mixed trials with low ambiguity as well. Accordingly, DLPFC and IPC accordingly seem to respond specifically to higher ambiguity independent of risk. This region may be directly linked to the DLPFC activity and was described as neural index of working memory functions (e.g. Friedman & Goldman Rakic, 1994) and mathematical operations (e.g. Arsalidou & Taylor, 2011). Further, we found significant inferior frontal gyrus activation under high ambiguity when participants where exposed to mixed risk blocks as compared to low- and high-risk blocks, which indicates a role in the interaction of risk and ambiguity. This also replicates previous research (e.g. Bach et al., 2009; Hsu et al., 2005).
We could also identify greater activity in the dACC for gain and in the insula and lateral OFC for loss in the mixed-risk condition as compared to the average activation during low- and high-risk levels. This ensured that our activations were independent from the overall level of risk and the overall level of outcomes or from EV. Activation of dACC was also observed during action selection under uncertain reward (Hampton & O’Doherty,(2007) and by Christopoulos et al. (2009) who characterized dACC activity as an objective metric of risk. Furthermore, Canessa et al. (2013) showed increased activation in clusters of the dACC for the parameter representing the magnitude of gains and deactivation for the magnitude of losses. Our results replicate these findings and are particularly well in line with those of Christopoulos et al. (2009) who observed stronger activity in dACC and more cautious decisions during conditions of pure risk. It has further been suggested that dACC activity reflects conflict (Botvinick et al., 2001) or a decrease in reward expectation (Holroyd & Coles, 2002) when the choice condition is risky.
Further, activation of the OFC was found during decisions under risk (Bach et al., 2009; Bhatt & Camerer, 2005; Krain et al., 2006). Bhatt et al. (2005) associated the OFC activity to the probability of gains, which was positively correlated with the level of ambiguity, whereas Krain et al. (2006) described a bilateral OFC activation for decisions under risk. We found activations in the OFC in mixed-risk trials when the possible loss was integrated into the analysis. Our results are in compliance to Kringelbach & Rolls (2004) suggesting that activity in the lateral OFC “is related to the evaluation of punishers which may lead to a change in ongoing behavior”. Therefore, we suggest that the present orbitofrontal activation relates to the coding of punishment contingencies and may be related to the emotional evaluation of losses in the decision alternatives. This is underlined by additional activation of the insular cortex, which is an important structure for the integration of emotions, subjective awareness, and experience, and the representation of internal states of affect and arousal (e.g. Craig, 2011; Duerden et al., 2013). In summary, the effects of losses seem to be primarily of emotional nature reflecting the representation of the value of losses and at the same time the effect of losses on subjective emotional awareness in conditions of mixed risk.
The fMRI analysis of integrated gain and loss parameters that compared the two ambiguous situations, identified significant activations for gains. Here, the amygdala was activated. This is in compliance with studies both on risk and ambiguity (Breiter et al., 2001 ; Hsu et al., 2005; Krain et al., 2006; Kuhnen & Knutson, 2005; Platt & Huettel, 2008). Since our experimental design controls for the overall level of risk, outcomes and EV, our result may suggest that increasingly high wins under ambiguity may act as a conditioned cue for excessive risk. This may bias subsequent processing towards a more negative evaluation of the risky gamble in total. This seems to be in line with the amygdala’s prominent role in affective conditioning (e.g. LeDoux, 2003).
Further, we asked the participants why they chose an option. Correlational analyses indicated that focusing on the amount and the probability of the gain in pure-risk situations led to incautious decisions. This provides further evidence that risk-taking becomes modulated by the expected value of options. Thus, we observed higher risk-taking under decision-conditions where a gain is more likely (for related evidence see Fochmann et al., 2017; Vorhold et al., 2007). In contrast, when participants realize a greater likelihood of loss in ambiguous situations, the behavior is more cautious. One limitation of the present study is certainly sample size in particular, in relation to the correlational findings. Future research should aim to corroborate these findings in larger samples.
In summary, for a risky context dorsal anterior cingulate cortex (dACC), insula, and OFC were activated in mixed trials that led to the highest percentage of cautious decisions in risk trials. Accordingly, these regions might reflect risk aversion in decision-making under risk and support previous research (e.g. Krain et al., 2006; Platt & Huettel, 2008; Tobler et al., 2007). Because we integrated potential losses, our findings do not provide evidence for a clear differentiation of risk aversion and loss aversion. Recent evidence suggests that these two kinds of effects have the same neural basis. For example, Canessa et al. (2013) found a significant correlation of loss aversion with risk aversion and Rabin (2000) postulated that loss aversion explains risk aversion in gambles with wins and losses. In particular, our experimental design focuses on the evaluation of humans in decision situations with different amounts of uncertainty that may all drive cautious decisions The behavioral phenomenon of risk aversion may be motivationally driven by loss aversion. In line with this argument, we found activation in regions that were demonstrated being relevant for loss aversion by Canessa et al. (2013) and/or linked to these areas, i.e., activation of the amygdala and the posterior insula; the SMA, the parahippocampal region. In summary, our results corroborate and extend the findings of Canessa et al. (2013) and supports a link between loss aversion and risk aversion and highlights a neural network for it. For a more extensive analysis and interpretation of individual differences in our data, our sample was too small.
Concerning ambiguity, we found that activity of the amygdala and activation of the IFG was present in mixed trials with high ambiguity when participants made the most cautious decisions. Further, an increase in ambiguity from low to high was indicated by increased activity in DLPFC and parietal cortex. Again our findings corroborate previous results showing activity for dorsolateral prefrontal cortex (DLPFC), amygdala, inferior frontal cortex, and parietal cortex (e.g. Bach et al., 2009; Hsu et al., 2005; Krain et al., 2006) in the ambiguous context and not in the risky context for trials inducing cautious decisions.
Taken together, we replicated that both kinds of uncertainty – risk and ambiguity – induce more cautious decision-making. As argued by Rothschild & Stiglitz (1970), the risk of options can be described by the spread of outcomes and risk avoidance behavior is primarily driven by the difference of risk between options. Ambiguity amplifies behavioral risk avoidance in such situations entailing risk differences between options in a non-additive, interactive way. Although some of our evidence suggests that risk avoidance under both ambiguity and risk may be driven by neural systems related to loss aversion, the systematic differences in patterns of neural responses between conditions of risk and ambiguity suggest basic differences between risk and ambiguity (Ellsberg, 1961; Knight, 1921) and their underlying neural sources (e.g. Bach et al., 2009; Hsu et al., 2005; Krain et al., 2006). Our own findings particularly support a special role of working memory related structures in DLPFC and parietal cortex that showed significantly different activity between high and low ambiguity conditions. The latter structures might be involved in integrating information from affective structures like posterior insula and amygdala in order to compensate for the missing cognitive numerical information to assess uncertainty and potential losses.
Acknowledgements: This study was supported by a grant from VolkswagenStiftung to JH.