The cognitive architecture of anxiety-like behavioral inhibition

Abstract

The combination of reward and potential threat is termed approach/avoidance conflict and elicits specific behaviors, including passive avoidance and behavioral inhibition (BI). Anxiety-relieving drugs reduce these behaviors, and a rich psychological literature has addressed how personality traits dominated by BI predispose for anxiety disorders. Yet, a formal understanding of the cognitive inference and planning processes underlying anxiety-like BI is lacking. Here, we present and empirically test such formalization in the terminology of reinforcement learning. We capitalize on a human computer game in which participants collect sequentially appearing monetary tokens while under threat of virtual "predation." First, we demonstrate that humans modulate BI according to experienced consequences. This suggests an instrumental implementation of BI generation rather than a Pavlovian mechanism that is agnostic about action outcomes. Second, an internal model that would make BI adaptive is expressed in an independent task that involves no threat. The existence of such internal model is a necessary condition to conclude that BI is under model-based control. These findings relate a plethora of human and nonhuman observations on BI to reinforcement learning theory, and crucially constrain the quest for its neural implementation. (PsycINFO Database Record

Figure 1

Experimental setup. (A) In Experiment 1 a human player (green triangle) rests in a safe place on the bottom of grid, while a “predator” is sleeping at the top (gray circle). On each epoch, six successive reward tokens (yellow rhombi) appear. The colored frame indicates the threat level of the sleeping predator with color/threat association balanced across subjects. (B) Tokens are separated from the player by barriers that disappear at a random time point. Once they have disappeared, the time until the token is removed is exponentially distributed. The player can press a key to collect the tokens, and thus accumulate up to six tokens over any given epoch. At any time during the epoch, the predator becomes active with equal probability, but once active it will only reveal itself if the player is currently outside the safe place and outside barriers; the predator can never reach the safe place or cross the barriers. (C) If the player is caught by the predator, it loses all tokens already collected in this epoch, and no more new tokens appear. Magnitude of potential loss therefore corresponds to the number of already collected tokens. Threat level is defined as the wake-up rate, which was different for the three predators. (D) In Group 1, the starting safe is protected from the predator. For Group 2, starting place is outside the safe place. Participants played 270 epochs, thus making up to 1,620 choices. (E) In Experiment 2, participants played the same approach/avoidance Task 1 on a 2 × 2 grid. (F) Following approach/avoidance Task 1, participants in Experiment 2 were instructed to press a key to “expose” the status of the predator in safe predator exposure Task 2. See the online article for the color version of this figure.

Figure 2

Results from Experiment 1. The graphs show responses to the possibility of collecting the nth token after already having collected (n – 1) tokens, which constitutes the potential loss. L = low threat; M = medium threat; H = high threat; Action = Proportion of epochs in which the player chose to collect at least the nth token. Because the players rarely approached after collecting five tokens, approach latency is only shown up to a potential loss of four tokens. As the data are unbalanced, mean approach latencies were estimated in a linear mixed effects model with random intercepts. Approach latency is markedly shorter in Group 2 than Group 1. See the online article for the color version of this figure.

Figure 3

Results from Experiment 2. Top panels: RT distributions for Task 2 in Experiment 2. RT are expressed with respect to the token appearance that preceded the response. Blue lines: RT distribution expected under the null hypothesis. Red line: Fit with the winning model, a combination of an ex-Gauss model with the null distribution. Because the inter-token-interval is a random variable, responses are less likely to be observed at long latencies than at shorter latencies, even under the null hypotheses. However, responses are much more frequent than expected directly after a token has appeared. Middle panels show RT distributions split up between two subsequent blocks. Y-Ticks = Estimated proportion of responses fit by the null distribution in the combined model. Bottom panels: Results from approach/avoidance Task 1. See the online article for the color version of this figure.