Multiple neural mechanisms of decision making and their competition under changing risk pressure

Abstract

Sometimes when a choice is made, the outcome is not guaranteed and there is only a probability of its occurrence. Each individual's attitude to probability, sometimes called risk proneness or aversion, has been assumed to be static. Behavioral ecological studies, however, suggest such attitudes are dynamically modulated by the context an organism finds itself in; in some cases, it may be optimal to pursue actions with a low probability of success but which are associated with potentially large gains. We show that human subjects rapidly adapt their use of probability as a function of current resources, goals, and opportunities for further foraging. We demonstrate that dorsal anterior cingulate cortex (dACC) carries signals indexing the pressure to pursue unlikely choices and signals related to the taking of such choices. We show that dACC exerts this control over behavior when it, rather than ventromedial prefrontal cortex, interacts with posterior cingulate cortex.

Figure 1

Behavioral Task and Results (A) Trial timeline: at the start of trials, subjects were presented with choices on the left and right of the screen. Each option was composed of a reward probability (height of purple bar fill) and magnitude (number next to each bar). This was followed by a choice cue (yellow question mark) that instructed subjects to choose. Subjects chose between a more probable low-magnitude option (safer option) and a less probable high-magnitude option (riskier option) on each trial. After responding, their choice was highlighted with a white frame, and feedback was shown for both options (both the chosen option and the alternative). If the choice was rewarded, then points were added to the blue “current points” bar at the top of the screen (white bar indicated added points), progressing it further toward the target. The number of trials remaining in the block was indicated in the circle at the bottom of the screen. After each trial, the number of remaining trials was reduced by one. The target turned white if it was reached. (B) Two examples of progressions through a miniblock. Points accumulated are shown in green, with target level in red (upper panel) and the resulting risk pressure in magenta (lower panel). In the first example, the target was relatively high, and the risk pressure is highest before a big win after the fourth decision, when the subject selected the less likely but more valuable riskier option. In the second example, the pressure is lower at first but increases after a series of losses until it actually exceeds the risk pressure experienced in the other block. nr., number. (C) Overall proportion of riskier choices as a function of increasing relative value of riskier choice (V_riskier − V_safer). (D) Overall proportion of riskier choices split by optimal risk bonus scaling and binned by increasing relative value of the riskier choice (V_riskier − V_safer). (E) Differences in proportion of riskier choices (left) between low-level and midlevel (green and blue, respectively, in D) and (right) between midlevel and high-level (blue and red, respectively, in D) optimal risk bonus scaling, illustrating how changes in optimal risk bonus scaling are associated with increased frequencies of riskier choices. Additionally, for the first change, choices with large V_riskier − V_safer value differences are affected more, whereas, for the second change, the more difficult decisions involving lower V_riskier − V_safer value differences are more affected. All error bars represent means ± SEM. See also Figures S1 and S2.

Figure 2

Logistic Regression of Riskier Choices against Parameters Defining Each Decision (A) GLM incorporating relative value of riskier choice (V_riskier − V_safer), number of trials already performed in the current block (trial number, nr.), and risk pressure. Increases in both V_riskier − V_safer and risk pressure were associated with significant increased riskier choices. The constant term from the GLM, however, indicates a bias against riskier choices (left bar with negative value). a.u., arbitrary units. (B) An alternative analysis used risk bonus (reflecting the model-based impact of the current risk pressure on V_riskier − V_safer); again, increases in this term were associated with significant increases in riskier choices. All error bars represent means ± SEM. See also Figures S3 and S4.

Figure 3

Decision Making in the Absence of Risk Pressure and Task Progression Effects (A) Decreasing risk bonus was associated with increased vmPFC activity. The impact was present, regardless of subjects’ choices. (B) Activity increases in dACC and elsewhere during the decision phase as number of trials remaining decreased.

Figure 4

dACC at Decision I (A) When just the main effect of choice is considered, then activity in dACC and adjacent dorsomedial frontal cortex increased for riskier as opposed to safer choices. (B) Again, the same dACC region and adjacent dorsomedial frontal cortex exhibited increased activity as a function of the relative value of riskier choices (V_riskier − V_safer). (C) dACC activity during the decision phase increased as a function of risk pressure when subjects did not succumb to it and instead made riskier rather than safer choices. (D) In (i), the dACC group time course of the V_riskier − V_safer effect is shown separately for riskier (continuous line) and safer choices (dotted line). In (ii), subjects who were less biased against riskier choices exhibited a higher dACC V_riskier − V_safer effect at the peak of the group time course. a.u., arbitrary units; RT, reaction time. See also Figure S5.

Figure 5

dACC at Decision II (A) The risk pressure signal in dACC increased when subjects did not act in accordance with it, but it had a negative effect when subjects did act in accordance with risk pressure and took the riskier option. a.u., arbitrary units. (B) Independent of choice (riskier or safer), higher expected value at the end of the block, as estimated using our model, was related to increased dACC activity. See also Figure S6.

Figure 6

FP at Decision (A) The FPl carried a V_riskier − V_safer signal regardless of which choice subjects ultimately made. (B) In (i), the group time course of the V_riskier − V_safer signal (regardless of whether riskier or safer decisions were made) is shown. In (ii), individual subject V_riskier − V_safer signal effect sizes at the peak of the group time course predicted individual behavioral sensitivity to risk bonus. a.u., arbitrary units. (C) The V_riskier − V_safer signal was present in FPl both for riskier choices (continuous line) and safer choices (dotted line). (D) Risk pressure activated PFl similarly for riskier choices (continuous line) and safer choices (dotted line).

Figure 7

The Effect of Outcomes on Activity in the dACC after Riskier or Safer Choices The effect of outcomes is not only dependent on choice (riskier or safer) but also on how much impact they have on future risk pressure (greater on the right than on the left of each panel).

Figure 8

PPI Analyses Demonstrated vmPFC and dACC Interactions with PCC during Different Types of Decisions and the Relationship with IFG Activity (A) Time course illustrating PPI between PCC and vmPFC as a function of decreasing risk bonus on trials when the safer option was taken. a.u., arbitrary units. (B) Time course of PPI between PCC and dACC as a function of V_riskier − V_safer value difference on trials on which the riskier option was chosen (dotted line). (C) Time course of a PPI on PCC, using the IFG signal and the different effects of vmPFC and dACC, the two regressors from (B) and (A), respectively. (D) Illustration of effects in (A)–(C). The PCC couples with dACC and vmPFC during decisions in which the riskier option (red) and safer option (green) were taken, respectively. Left IFG may regulate PCC’s interactions with the vmPFC and dACC by increasing the relative degree of coupling to the former as opposed to the latter during riskier choices. See also Figures S7 and S8.