Primate anterior insular cortex represents economic decision variables proposed by prospect theory

The study investigates how context-dependent risk attitudes, central to prospect theory, are implemented in neuronal activity. Decisions under uncertainty depend on risk attitude and are modulated by context such as gain versus loss framing and current wealth. Prospect theory posits that outcomes are evaluated relative to a reference point and that losses loom larger than gains. The authors hypothesize that the primate anterior insular cortex encodes the reference point (current wealth) and reference-dependent, asymmetric value signals for gains and losses, influencing risk attitude. To test this, they use a token-based gambling task in macaques and record single-unit activity in AIC during decision making.

Human imaging and lesion studies implicate a network including the anterior insular cortex in risky decision-making, interoception, and risk-averse behavior. AIC damage affects risk attitudes, and monkey recordings show AIC neurons encode reward expectation. Prior work shows probability distortion and context effects on risk preferences across species, and insula involvement in risk signals and prediction errors. These findings motivate testing whether AIC implements prospect theory constructs: reference-dependent valuation and asymmetric sensitivity for gains versus losses.

Two male rhesus macaques performed a token-based gambling task. Tokens served as secondary reinforcers; accumulating six or more tokens yielded a 600 μl water reward. Maximum three tokens could be gained or lost per trial. Choice trials presented a sure option versus a gamble option; loss and gain contexts were separated. Sure options indicated certain token changes; gambles offered two outcomes with indicated probabilities. There were 24 choice combinations (12 gain, 12 loss) and 13 forced-choice options interleaved in blocks. Start token number (0–5) indicated current wealth/reference point. Monkeys made saccades to select options; outcomes updated tokens, and excess over six rolled over to the next trial. Behavioral analyses quantified probability of choosing gamble, response times, and certainty equivalents for matched end-token distributions to distinguish relative versus absolute value frameworks. A prospect theory (PT) model fitted behavior per start-token level with: a utility function with curvature α (risk attitude) and loss aversion λ (steeper in loss if >1), a probability weighting function with curvature γ (inverse S if <1), and a softmax decision policy with inverse temperature s and directional bias. An expected value (EV) model without nonlinearities served as comparison. Model fits were evaluated by cross-validated negative log-likelihood and BIC. Neurophysiology: Single-unit recordings were obtained from anterior insular cortex (AIC) in two monkeys (240 neurons total: 142 and 98). Activity during the choice period (target onset to saccade) was analyzed, primarily in forced-choice trials to isolate encoding from spatial confounds (AIC showed weak spatial selectivity). Multiple linear regression models tested encoding of decision variables: token-related (parametric linear, categorical high/low, numerical Gaussian tuning to specific token count), value-related (general value across contexts, gain/loss category, behavioral salience as absolute EV, gain-only value, loss-only value), risk-related (parametric variance, categorical safe vs gamble), and expected end token (absolute framework). Best models were selected via AIC. Sensitivity of gain vs loss value neurons was assessed via standardized regression coefficients (SRC), and dependence on wealth level was tested by splitting trials into low (0–2) and high (3–5) tokens. Trial-by-trial choice and risk-attitude predictions used ROC/AUC analyses in choice trials; explicit choice AUC contrasted gamble vs sure choices, and implicit risk-attitude AUC contrasted risk-seeking vs risk-avoiding choices in subsets where EV did not favor the chosen option. Correlations between AUCs and contingency of encoding type with predictive signals were assessed (Pearson, chi-square).

Behavior:

• Monkeys preferred gambles overall and more so in gains than losses. P(Gamble): Monkey G 59%, Monkey O 67% (one-sided t-tests both p < 10^-4). Higher in gain than loss context (paired t-tests p < 10^-4 for both).
• Wealth modulation: In gain context, P(Gamble) decreased with start tokens (Monkey G β = -0.044, p < 10^-4; Monkey O β = -0.035, p < 10^-4). In loss context, P(Gamble) increased with start tokens (Monkey G β = 0.028, p < 10^-4; Monkey O β = 0.001, p = 0.8).
• Response times were slower in loss vs gain (G: 205 vs 247 ms, p < 10^-3; O: 175 vs 206 ms, p < 10^-3) and tended to increase with start tokens.
• Relative value framework: Certainty equivalents for matched final token outcomes differed between gain and loss framing (both monkeys p < 10^-4), indicating relative valuation. Prospect theory modeling:
• Utility curvature α decreased with start tokens (G β = -0.16, p < 10^-4; O β = -0.14, p < 10^-4), showing shift from risk seeking at low wealth to neutrality/aversion at higher wealth in gains.
• Loss aversion λ > 1 in Monkey G across token levels (p < 10^-4), indicating greater sensitivity to losses; Monkey O’s λ ~ 1 (no consistent loss aversion). λ did not vary significantly with tokens.
• Probability weighting γ < 1 for both monkeys across tokens (p < 10^-4), inverse S-shaped distortion; slight token effect in G (β = 0.02, p < 10^-4), none in O.
• Decision policy: Negative directional bias toward gambles at low tokens, bias magnitude decreased (less risk-seeking) with tokens (G β = 0.32, p < 10^-4; O β = 0.28, p < 10^-4). Choice stochasticity decreased with tokens (s increased: G β = 0.23, p < 10^-4; O β = 0.22, p < 10^-4).
• PT outperformed EV in cross-validated fits and BIC for many token levels (see Table 1; multiple comparisons significant, e.g., G: BIC better p < 10^-4 at tokens 0–1; O: p < 10^-2 to < 10^-4 at low tokens). Neuronal encoding (choice period, primarily forced-choice trials):
• 62% (149/240) encoded at least one decision-related variable; 34% (50/149) showed mixed selectivity.
• Token/reference point encoding: 105/149 (70%) encoded start tokens—Parametric 13 (12% of token cells), Categorical high/low 11 (11%), Numerical tuning 81 (77%), covering 0–5 tokens.
• Value encoding: 73/149 (49%)—General value 14 (19% of value cells), Gain/Loss categorical 13 (18%), Behavioral salience (|EV|) 13 (18%), Loss-only value 29 (40%), Gain-only value 4 (6%). Loss-value cells were far more prevalent.
• Risk encoding: 19/149 (13%)—Parametric variance 9 (47%), Categorical safe/uncertain 10 (53%).
• Absolute end-token encoding was rare: 9/149 (6%). Prospect-theory-consistent neural asymmetry and wealth effects:
• Loss-value neurons had larger sensitivity (|SRC|) than gain-value neurons (means 2.97 vs 2.05; one-sided permutation p ≈ 0.053–0.054), indicating greater neural sensitivity to losses than gains.
• Sensitivity decreased with higher tokens: loss-value neurons |SRC| low vs high tokens 3.49 vs 2.47 (p = 0.017). Gain-value neurons showed a similar but non-significant trend (2.38 vs 2.06, p = 0.29). Choice and risk-attitude prediction:
• 15% (35/240) showed significant trial-by-trial predictive signals: 15 (6%) predicted explicit choice, 16 (7%) predicted implicit risk-attitude, 4 (2%) predicted both. Across neurons, explicit-choice AUC and risk-attitude AUC were correlated (r = 0.41, p < 10^-4). Encoding category was not significantly associated with being predictive (chi-square χ^2 = 7.61, p = 0.67).

The findings support the hypothesis that the primate anterior insular cortex represents prospect-theory–relevant variables: it encodes the reference point (current wealth) and context-specific value signals for gains and losses, with stronger sensitivity to losses, paralleling behavioral loss aversion and risk-attitude shifts. The predominance of loss-specific value neurons and categorical gain/loss coding suggests partially segregated neural representations for gains and losses, consistent with separate utility functions in prospect theory. The scarcity of absolute end-token encoding and the behavioral certainty-equivalent results indicate a relative valuation framework. Risk signals in AIC, together with trial-by-trial neural prediction of choices and risk attitude, position AIC as part of a circuit that monitors state and context to modulate downstream decision processes under risk. Differences from typical human risk aversion (overall gamble preference here) likely reflect task-specific factors rather than species-wide differences. No clear functional segregation was observed across anterior insula subregions sampled.

This study demonstrates that macaque anterior insular cortex encodes core components of prospect theory during risky decision-making: the wealth-dependent reference point, asymmetric value coding with enhanced sensitivity to losses, and subjective probability distortion reflected behaviorally. Neural activity in a subset of AIC neurons predicts both choices and implicit risk attitude on a trial-by-trial basis, suggesting AIC contributes to context- and state-dependent control of risk taking. Future research should: (1) probe richer parametric ranges of outcomes and probabilities to test neural non-linearities more directly; (2) assess whether similar representations exist in upstream and downstream regions to delineate circuit mechanisms; and (3) use causal manipulations of AIC activity to determine its necessity in shaping risk preferences.

• Limited parametric range of reward magnitudes and probabilities constrained tests of non-linear utility and probability weighting at both behavioral and neural levels.
• Unclear whether observed signals are unique to AIC; recordings from other connected regions are needed to establish specificity and circuit roles.
• No causal interventions were performed; thus, AIC’s necessity in risky decision-making remains to be established.
• The origin of the asset (wealth) effect on risk attitude—strategic optimization versus contextual bias—was not resolved.