The Subjective Value of Cognitive Effort is Encoded by a Domain-General Valuation Network

Abstract

Cognitive control is necessary for goal-directed behavior, yet people treat cognitive control demand as a cost, which discounts the value of rewards in a similar manner as other costs, such as delay or risk. It is unclear, however, whether the subjective value (SV) of cognitive effort is encoded in the same putatively domain-general brain valuation network implicated in other cost domains, or instead engages a distinct frontoparietal network, as implied by recent studies. Here, we provide rigorous evidence that the valuation network, with core foci in the ventromedial prefrontal cortex and ventral striatum, also encodes SV during cognitive effort-based decision-making in healthy, male and female adult humans. We doubly dissociate this network from frontoparietal regions that are instead recruited as a function of decision difficulty. We show that the domain-general valuation network jointly and independently encodes both reward benefits and cognitive effort costs. We also demonstrate that cognitive effort SV signals predict choice and are influenced by state and trait motivation, including sensitivity to reward and anticipated task performance. These findings unify cognitive effort with other cost domains, and suggest candidate neural mechanisms underlying state and trait variation in willingness to expend cognitive effort.SIGNIFICANCE STATEMENT Subjective effort costs are increasingly understood to diminish cognitive control over task performance and can thus undermine functioning across health and disease. Yet, we are only beginning to understand how decisions about cognitive effort are made. A key question is how subjective values are computed. Recent work suggests that the value of cognitive effort might be computed by networks that are distinct from those involved in other domains like intertemporal and risky decision-making, implying distinct mechanisms. Here we demonstrate that the domain-general network also encodes effort-discounted value, linking cognitive effort closely with other domains. Our results thus elucidate key mechanisms supporting decisions about cognitive effort, and point to candidate neural targets for intervention in disorders involving impaired cognitive motivation.

Keywords: cognitive control; cognitive effort; decision-making; motivation; subjective value; vmPFC.

Figure 1.

Decreasing indifference points (here normalized by, and averaged over base amounts) reflect rising subjective costs as N-back level increases. Gray bars and black lines reflect group means ± SEM. Gray dashed lines show individual participants' discounting curves, illustrating clear individual differences. AUC values provided for two example participants.

Figure 2.

A, Proportion of high-demand options selected as a function of offer biasing, and N-back level. B, Median reaction times as a function of proximity and whether choices went with (pro-bias) or against (anti-bias) offer biasing (High-load: γ < 0; Low-load: γ > 0).

Figure 3.

Imaging trials begin with a high-amount, high-load offer (load is indicated by a color label, e.g., red for 2-back). After 6 s the 1-back (black task) offer is presented. Participants have 5.25 s to respond. After response is indicated briefly, a fixation cross is presented until the end of the trial.

Figure 4.

A, Six-millimeter-radius spheres centered at all a priori ROIs. Colors indicate origination from either of two prior meta-analyses of domain-general SV encoding (yellow, Levy and Glimcher, 2012; Bartra et al., 2013) or two recent studies on SV encoding during cognitive effort decision-making (blue, Massar et al., 2015; Chong et al., 2017). Note that the amygdala is projected to the surface for display purposes only. B, Percentage of ROIs from each set tracking SV 6–8 s after first offer onset by p value. Map shows ROIs reliably tracking SV at p < 0.05. C, Percentage of ROIs from each set with reliably more activity on difficult, regular versus easy catch trials by p value. Map shows ROIs with a reliable difficulty contrast at p < 0.05. T Pole, Temporal pole; Amyg, amygdala; Supr Gyrus, supramarginal gyrus; ITG, inferior temporal gyrus; Brstm, brainstem.

Figure 5.

Residual time courses in a priori ROIs, averaged by first offer amount or load. Error bands reflect ± SEM across participants. Reliability of amount and load effects in separate hierarchical multiple regression (trials and ROIs, nested within participants) at each time point indicated by *p < 0.05. Gray region highlights 6–8 s after first offer onset; vertical dashed line indicates second offer onset.

Figure 6.

Trait willingness to select high cognitive load for reward varies with reward sensitivity. A, Residual time courses in a priori ROIs, averaged by first offer amount and divided by above (High) or below median (Low) AUC. Error bars reflect SEM across participants. Reliability of the interaction between AUC and first offer amount in separate hierarchical multiple regression (trials nested within participants) at each time point indicated by *p < 0.05 and · p < 0.075. Gray region highlights 6–8 s after first offer onset; vertical dashed line indicates second offer onset. B, Cross-session AUC predicts the average amount effect on mean activity at 6–8 s following first offer onset in both the bilateral ventral striatum (B = 2.49 × 10⁻², p = 0.027) and bilateral amygdala (B = 2.43 × 10⁻², p = 0.023). Shaded regions show 95% CI.

Figure 7.

Higher mean N-back performance (measured by sensitivity index d′) predicts shallower load effects in bilateral AI across participants (B = 2.11 × 10⁻²; p = 0.085). The shaded region shows 95% CI.

Figure 8.

Mean residual time courses at vmPFC loci, averaged by whether participants chose the high or low-demand offer and whether the low-demand offer amount biased them toward the high- or low-demand offer. Error bands reflect SEM across participants. Reliability of choice-bias effects in separate, fully random hierarchical multiple regressions at each time point indicated by *p < 0.05, and · p < 0.075. Based on a Choose Low, Pro-Bias < Choose Low, Anti-Bias < Choose High, Pro-Bias < Choose High, Anti-Bias coding scheme.