The cognitive architecture of anxiety-like behavioral inhibition.

Experiment 1

This experiment addressed the question whether anxiety-like BI is reduced when it leads to negative consequences, and therefore, whether it is under Pavlovian or instrumental control. To this end, we adapted a previously developed operant approach/avoidance conflict paradigm, embedded in a “scoop-and-run” computer game ( Bach, 2015 ). This task affords analysis of individual actions, and thus differs from our previous “stay-and-play” approach/avoidance conflict task ( Bach et al., 2014 ). One group played the game in a version in which BI had no influence on threat probability, as in our previous work. In this version, participants started each trial in a safe place where the predator could not reach them. In the second group, participants would start outside the safe place. The predator could thus catch them if they hesitated to make a response; thus, BI would increase threat probability. We hypothesized that participants in Group 2 would show reduced BI after experiencing the unfavorable consequences. We expected to find this both in an overall group difference, and also in a Group × Time interaction, that is, participants in Group 2 would reduce approach latency over time. A between-subjects design was used because behavior in approach/avoidance conflict tasks has been described to become stereotypical with extended practice ( Gray & McNaughton, 2000 ).

We recruited 38 participants from the student and general population (19 female, mean age ± standard deviation: 23.7 ± 3.7 years), and assigned them to two gender-balanced groups. The sample was independent from Experiment 2 and from a previous study using the same setup ( Bach, 2015 ). State anxiety values for all but one participant, and trait anxiety for all but three participants, were within 2 standard deviations around the reference sample mean (586 individuals between 15- and 29-years-old, both sexes; Laux, Glanzmann, Schaffner, & Spielberger, 1981 ). State anxiety values were comparable to the reference sample (35.07 vs. 36.8, p = .13, Welch’s t test) and trait anxiety values slightly higher than the reference sample (39.25 vs. 35.1, p = .002). All participants gave written informed consent after being fully informed about the purpose of the study. The study protocol, participant information, and form of consent, were in full accordance with the Declaration of Helsinki and approved by the competent research ethics committee (Kantonale Ethikkommission Zurich).

Design and procedure

The experiment followed a 2 × 3 × 6 factorial design with the between-subjects factor starting place (inside/outside safe place) and the within-subjects factors threat level (low/medium/high) and possible loss (0–5 tokens). Threat level corresponds to wake-up rate of the predator, and thus, to loss probability. Potential loss corresponds to number of already collected tokens which would be lost if the player got caught. A modified version of a previously developed approach/avoidance computer game ( Bach, 2015 ) was presented on a standard LCD monitor (see Figure 1 ) in six blocks of 45 epochs. In each epoch, a sequence of six reward tokens appeared at random time points; the player could decide each time whether or not to approach and collect the token. The primary dependent measure was approach latency on those trials on which participants chose to approach. Participants received a fixed payment and an additional reward for the number of retained tokens of one randomly drawn epoch at the end of the experiment. A “sleeping predator” was waiting above the token and could become active with a probability that was constant over time. Once active, if the human player was inside the safe place, it deactivated itself. If the human player was outside the safe place (regardless whether or not a reward token was present), it revealed itself and moved to the human player’s grid block. The human player was “eaten” and all previously collected reward tokens from this epoch were removed. Once the predator was active, the human player had no possibility to escape if he or she was in the neighboring grid block. In order to remove any time benefit (opportunity cost) from getting caught by the predator, the active predator staid visible on the screen for the remaining time of the epoch while the human player had to wait.

Stimuli and task statistics

The game was presented on a 1 × 4 grid in vertical orientation (∼1.5° horizontal visual angle). The player was placed on the starting position, confined by one (Group 1) or two (Group 2) “barriers” to prevent the player from moving. Starting position was in the safe place for Group 1, and in the grid block above the safe place for Group 2. A token was visible in the grid block above the player.

After a variable interval drawn from an exponential distribution with a mean of 1.25 s, the barriers were removed, and the player was free to move and collect the token, under risk of getting caught. Note that the presence of barriers, and of the token in the delay period, is different from Experiment 2 and from a previous report ( Bach, 2015 ). The reason for making the token visible in the delay period in this version of the game was to simplify the graphical setup such that only one visual event (barrier removal) occurred to signal the possibility for token collection. The interval during which the reward token was present and could be collected was also drawn from an exponential distribution with a mean of 1.25 s. If not collected, the token disappeared at the end of this interval. Whether or not it was collected, the player was moved to the starting position 250 ms after the token disappeared, and the barrier above the player put in place. The next trial within the epoch started 250 ms later. The wake-events of the predator followed a homogenous Poisson process, independently determined in successive time bins of 20 ms duration. The wake-up rate was set such that the probability of getting caught was p 1 = 0.1, p 2 = 0.2, and p 3 = 0.3, respectively for the three threat levels, if the player stayed outside the safe place for 100 ms (Group 1) or 600 ms (Group 2). These latency values approximated values found in previous experiments. Thus, the event rate for the three threat levels was, respectively, λ 1 = 1.0536, λ 2 = 2.2314, and λ 3 = 3.5667 for Group 1, and was divided by 6 for Group 2. Actual catch rates depend on participants’ response latencies and turned out to be 0.08/0.18/0.25 for Group 1, 0.05/0.10/0.17 for Group 2 when making an approach response, and 0.02/0.03/0.04 for Group 2 when making an escape response. The human player was controlled with the up/down cursor keys on a standard computer keyboard. The player could move between grid blocks at all times unless restricted by barriers or caught by the predator, but it could never reach the top grid block occupied by the sleeping predator.

Data analysis

All data are necessarily unbalanced because the number of data points for each cell in the design depends on behavioral choices and on chance.

When the participant approached the token, we extracted the approach latency as main dependent variable. We also analyzed return latencies, that is, time passed between approaching the token and moving back. In Group 2 the player had to make two movements to go to the safe place; hence there are two return latencies. For escape choices in Group 2 we extracted escape latency. To avoid response latencies being biased by extreme values, they were only analyzed if they fell into response windows of 150 ms < approach/escape latency < 2,000 ms and 0 ms < return latency < 2,000 ms, as in a previous study ( Bach, 2015 ). This excluded, in Group 1, 1.7% of approach latencies and 3.5% of return latencies. In Group 2, this excluded 4.5% of approach latencies, 5.7% of escape latencies, and 1.5% or 7.6% of the first or second return latencies, respectively. Choices were reconstructed by creating six data points for each epoch, corresponding to the possibility of collecting six tokens. For each of these six tokens, we recorded 1 if the individual chose to collect up to, or more than, this number of tokens on this epoch, and 0 if the individual chose to collect less than this number of tokens on this epoch. Choices in epochs on which the player was caught cannot be reconstructed and were therefore not analyzed. The resulting data are serially correlated by design. Most players rarely collected the sixth token such that some design cells were empty and the parameters could not be estimated reliably. Therefore, the sixth token was excluded for all reaction time (RT) analysis. The resulting model followed a 2 (Group) × 3 (Threat Level) × 5 (Potential Loss) factorial design, and for choice data a 2 × 3 × 6 design. To analyze changes in approach latency over time, we split the data into the 6 blocks and added the main effect of block and the Block × Group interaction to the model. Finally, in addition to the full factorial model, we also analyzed data from both groups separately in 3 × 5 or 3 × 6 factorial models, in order to facilitate comparison with previous publications.

The lme4 package in the software R ( www.r-project.org ) was used for all inference statistics as it provides meaningful parameter estimators for unbalanced data sets. Choice data were analyzed using a generalized linear mixed effects model (glmer) for binomial data, and RTs in linear mixed effects models (lmer). We did not transform RTs, as we had no a priori reason to do so and a previous report demonstrated that analysis of transformed RTs replicates analysis of raw RTs ( Bach, 2015 ). All models included a random subject intercept. Fixed-effects F -statistics were extracted using unpartitioned error variance and the R function ANOVA; p values were calculated by using a (conservative) lower bound on the effective denominator degrees of freedom as df = N − K, where N is the number of observations, and K is the number of all modeled fixed and random effect parameters. No p values were computed for the choice data as they are autocorrelated across the “potential loss” factor by construction, and therefore have reduced effective numerator degrees of freedom. Mean RTs were reconstructed from the linear mixed effects model using the function lsmeans. In a nutshell, this function averages the data for each subject and experimental condition separately, and then averages over subjects, while correcting for missing values in individual subjects.

Results

Approach latency in Group 2 was around 150 ms shorter than in Group 1, a highly significant difference ( Figure 2 , Table 1 ). While threat level increased approach latencies in both groups, this influence was smaller in Group 2, as indicated by a significant Group × Threat Level interaction. Additionally, the effect of potential loss on approach latencies was different between the groups, and there was a significant three-way interaction. Next, we analyzed how BI developed over time, in both groups, by splitting the data into six blocks of 45 epochs. We found a significant Group × Block interaction (see Results in supplemental materials). Approach latency was reduced between Blocks 1–2 from 515 ms to 507 ms in Group 1, and from 404 ms to 358 ms in Group 2. For comparison with a previous report we analyzed Group 1 separately. Threat level, potential loss and their interaction, influenced approach latencies with a similar pattern as in previous reports, but there was no linear effect of potential loss ( Table 1 , Figure 2 , see Results in supplemental materials).

Discussion

We asked whether BI is reduced when it increases threat probability. We found that in this case, approach latencies were about 150 ms shorter than in a control group in which BI had no impact on threat probability. This suggests that BI adapts to unfavorable consequences. Alternatively, this overall group difference could be explained if participants used a Pavlovian, but model-based strategy ( Dayan & Berridge, 2014 ) to precompute their behavior even before the experiment started, rather than instrumentally learn from their actions. In other words, according to this explanation they would not take into account consequences of BI, but respond to the Pavlovian cues of being inside or outside the safe place when they started. To exclude such possibility, we showed that participants in Group 2 adapted their behavior over time to a greater degree than in control Group 1. In particular, we observed a pronounced reduction in approach latencies from Block 1 to Block 2 in Group 2 (46 ms) but not Group 1 (8 ms). Furthermore, we note that in spatial approach/avoidance conflict tasks, anxiety-like BI is elicited also outside safe compartments in rodents ( Fonio et al., 2009 ) and humans ( Bach et al., 2014 ). Thus it appears unlikely that in the current task, BI should depend on the qualitative aspect of being in a protected starting position. Finally, the starting place in Group 2 was quantitatively no less safe than in Group 1: In case of an escape response in Group 2, participants were rarely caught, just as when making no response in Group 1. In case of an approach response, participants in Group 2 adapted their approach latency to an extent that overall, they were caught less often than in Group 1. All in all, it appears that instrumental consequences of BI lead to its reduction, rather than Pavlovian cues.

Behavior in the control group was comparable to a previous report ( Bach, 2015 ), underlining the validity of the modified experimental setup. Different from the previous report and from Experiment 2, however, we observed that the influence of possible loss on approach latency was not linear. A possible explanation is that the token was already visible before the participant could make a movement. According to our previous model, BI arises from subjective assumptions on threat/reward correlations, corresponding to a situation in which a predator is alerted by the occurrence of his prey’s reward. The influence of possible loss on approach latency in this model depends on the curvature (second derivative) of the temporal evolution of threat probability. It appears possible that the temporal evolution of subjective threat probability after a reward occurs is different from the evolution after a barrier is removed, and this would lead to a different impact of possible loss.

Experiment 2

After having shown that anxiety-like BI is likely under instrumental control, we asked whether it is based on an explicit (although not necessarily conscious) model of the environment. Experiment 2 therefore addressed whether possible assumptions about threat/reward correlations in the approach/avoidance task are explicitly expressed in a different task, not involving any threat and thus not involving BI. Such threat/reward correlations exist in natural environments ( Prevedello et al., 2013 ; Sofaer et al., 2013 ) but they are objectively absent from our task. However, we have previously shown that anxiety-like BI would be adaptive from the perspective of an agent if the agent subjectively assumed such correlations. Experiment 2 had two tasks. In approach/avoidance Task 1, participants were familiarized with the computer paradigm and collected tokens. Next, they engaged in safe predator exposure Task 2. Here, they were could expose the status of the predator by key press, without threat of getting caught. They were rewarded if they exposed the predator just at the moment when it was awake. Under a null hypothesis that participants had no assumptions on threat/reward correlations, the timing of their exposure attempts should be independent from the occurrence of incidental and unobtainable tokens. We hypothesized that such assumptions exist, and that exposure attempts would be more frequent immediately after tokens.

We recruited 20 participants from the student and general population (10 female, mean age ± standard deviation: 23.6 ± 3.7 years). The sample did not overlap with Experiment 1 or a previous report ( Bach, 2015 ). State anxiety values for all and trait anxiety for all but two participants, were within 2 standard deviations around the reference sample mean ( Laux et al., 1981 ). State anxiety values were slightly lower (33.4 vs. 36.8, p = .04), and trait anxiety values slightly higher than the reference sample (38.6 vs. 35.1, p = .05). All participants gave written informed consent after being fully informed about the purpose of the study. The study protocol, participant information, and form of consent, were in full accordance with the Declaration of Helsinki and approved by the competent research ethics committee (Kantonale Ethikkommission Zurich).

Design and procedure: Approach/avoidance Task 1

This part realized a 3 × 6 factorial design with the within-subjects factors threat level (low/medium/high) and possible loss (0–5 tokens). Participants played four blocks (Blocks 1–2, 5–6) of 45 successive epochs of the previously reported computer game ( Bach, 2015 ). The game was the same as in Group 1 of Experiment 1, with the only difference that the playing field was a 2 × 2 grid in diamond orientation (∼4.0° horizontal angle), there were no barriers, and the timing was therefore slightly different. Specifically, at the start of each epoch, the player was in a safe place in the bottom grid block. A token could appear either to the left or to the right. The sleeping predator was waiting in the top grid block. As there were no barriers, the player was free to move during the entire epoch unless caught be the predator. The interval during which the reward token was present and could be collected was drawn from an exponential distribution with a mean of 1.25 s. If not collected, the token disappeared at the end of this interval. After this, whether or not the token was collected, a waiting interval started that lasted 500 ms plus a random sample from an exponential distribution with mean of 1.25 s, before the next token came on the screen or the epoch ended. The predator wake-up rates were the same as for Group 1 in Experiment 1. The human player was controlled with the left/right cursor keys on a standard computer keyboard.