Our experiment showed that performance on the IGT depends on the interaction between two separable cognitive components. Although we did not collect individual covariate measures of specific cognitive abilities, our experimental design permits a preliminary characterization of these two components. The first of these includes those higher-level processes involved in tracking and discovering the long-term payoff structure of the task. The second is based on the immediate emotional reactions to wins and losses, and is responsible for participants' sensitivity to loss frequencies and reactions to loss magnitude. These two components can be tracked by aggregating participants' choices according to the decks' payoff or loss frequency, respectively. In fact, our results show a double dissociation between these two processes. The first, but not the second, is harmed by the additional demands of an interfering task; and the second, but not the first, is affected by the recent experience of a monetary loss.
Additionally, our design allowed us to separate the decision stage, where participants evaluate their options and perform a selection, from the learning stage, where participants have the opportunity to learn about each deck's returns. This was achieved by having participants perform the IGT in two phases. Feedback was suspended on the second phase, so that further learning was prevented. Our results show that an interfering task disrupts performance only in the first phase, but not in the second. This shows that the learning stage, during which each decks' wins and losses are estimated, critically requires central cognitive resources. By contrast, the decision phase itself, whether driven by payoff or loss frequency, can be carried out efficiently in spite of an interfering task. As far as we know, this is the first successful attempt to separate these two stages.
Other authors have previously attempted to decompose the IGT into component processes (e.g., [11, 66]). Our framework, however, advances these attempts in two directions. First, it accounts for a richer measurement of performance, based on loss frequency as well as payoff. Second, it includes a more elaborate specification of the type of knowledge required by the two processes, including a description of the conditions under which decision making requires central cognitive resources.
Our conclusions are hinged on the secondary task's disruptive effect on performance, as measured by the P index. In line with previous studies that adopted a dual-task paradigm to the IGT [42–45], this result was interpreted as arising from competition for central cognitive resources, such as working memory. One might argue that the task we adopted is not normally used to saturate these resources. However, as we argued in the introduction, the IGT does not require a constant use of working memory or other cognitive resources. Participants can use these resources intermittently during the task, and perform long streaks of selections building on either their previous estimates or their intuition. The task we designed and used is effective in that it consistently prevents participants from reverting back to the IGT for the amount of time needed to perform satisfactory payoff estimation.
An alternative explanation for our findings could be that the time constraints posed by the secondary task have changed participants' attitude towards risky options. In particular, time constraints might make them less averse to risk and more prone to select from the bad decks. In fact, a recent study has shown that brain responses that correlate with decision risk are slower and can be temporally dissociated  from those that correlate with processing a decision outcome . If risk processing indeed takes more time, it is also more easily affected by the distractions posed by a fast-paced interfering task. Furthermore, this account could explain both our results and those obtained in , where decreased IGT performance was obtained by reducing the available decision time with no interfering task.
One problem with this alternative interpretation arises when we define risk. In the economic literature (e.g.,[67, 68]) as well as in neuroimaging studies of decision making [7, 9], risk is operationalized as the variance of returns within an option. In turn, variance is at maximum when the frequency of losses is higher. It follows that the high loss-frequency decks, B and D, should be perceived as riskier, and that any attitude change towards risk should impact the Q, and not the P, index. On the contrary, our data show that the secondary task affected P, but had no effect on Q.
Another problem with the risk-aversion hypothesis is that effect of risk aversion should be observed during the decision stage, when participants ultimately decided which deck to pick. However, our design allowed us to separate the decision stage from the learning stage within the IGT, and infer that the effects of a secondary task are limited to the learning stage. This conclusion was based on the fact that the secondary task did not alter performance in the second, blind phase – after learning has occurred. Had the secondary task altered participants' attitude towards risk, performance in the second phase should have been affected as well. Therefore, we conclude that our account is preferable to the risk-aversion interpretation of our results.
The findings reported in this paper can also contribute to interpret a number of results in the IGT literature. Importantly, they are relevant to the debate over the existence of unconscious components of decision making in the IGT. In our dual-process framework, implicit and explicit components co-exist, and jointly contribute to the behavioral choices. Identifying one or the other depends on how performance data is examined. For instance, selecting from the advantageous decks requires a demanding evaluation of each decks' payoff that can occur only under conscious control. It is therefore predictable that performance, when measured only by payoff, correlates with participants' conscious knowledge of the task [19, 37]. On the other hand, the low-level process that responds to loss frequency and magnitude does not show up in this sort of measurement. In our results, we detected it in the Q index. It was possibly underlying participants' biases and perseverations in . An interesting possibility is that anticipatory physiological reactions, such as skin conductance responses, can reflect the contribution of this low-level process to some extent. The relative independence of this process from payoff preferences could explain how physiological responses were found in anticipation of disadvantageous choices in the original version of the task [11, 35], but in anticipation of advantageous choices under a different distribution of losses .
Even more importantly, our results might be useful for better understanding the source of impairment in clinical populations performing in the IGT. Our framework suggests that there are at least three potential sources for patients' abnormal behavior.
A first source of impairment is the inability to allocate sufficient cognitive resources to the IGT. This prevents patients from conducting a sufficiently accurate assessment of the decks, and, therefore, leads to detectable performance differences in terms of payoff measures (e.g., the P index) when compared with controls. This is the case, for instance, of patients with working memory impairment due to prefrontal lesions . Although similar impairments are not affective in nature, they can possibly originate from an emotional disorder. In depression, for instance, mental rumination hinders executive functions and problem solving abilities [70, 71], and can thus interfere with patients' efforts to examine the payoff structure of the task.
A second source of impairment is in the emotional reactions to losses. Our results show that experienced monetary losses trigger a temporary shift in decision preferences. In patients with affective disorders, this change might not take place at all (e.g., ), or might occur and have longer behavioral effects. Because of the structure of the IGT, this kind of impairment affects performance measured by the Q index, but does not necessarily impact the traditional payoff-based performance scores. Regrettably, however, most patient studies have assessed IGT performance only in terms of payoff, and might have overlooked important significant differences in this measure.
Our framework allows for a third, potential source of impairment. It states that participants have two types of knowledge, one deriving from careful payoff estimation and one from emotional reaction to losses or wins. Not only these two types of information, but also their relative utilities need to be learned – and ultimately learned from feedback on one's own actions. Therefore, we propose that affective reactions play a double role in decision making, and that they are used not only to evaluate decision options, but also for guiding meta-decisions on how to select among alternative decision policies.
Converging lines of evidence indicate that, for example, VMPFC patients' impairments could originate at this level. For instance, these patients poorly organize their search of information about decision options. In the IGT, they persist in selecting from the disadvantageous decks even when they are aware that the other options have higher returns . Finally, this explanation is consistent with their inability to re-learn deck-outcome associations .
In fact, we suggest that decision-making impairments in patients with affective disorders often originate at this meta-strategic level. As such, they do not betray a malfunction in an isolated emotional decision making system, but problems in exploiting emotions to consolidate decision policies and organize behavior.