Abstract
Avoidance of potential danger is crucial to an individual’s survival, but in the absence of threat, avoidance contributes to the development and maintenance of anxiety-related psychopathology. Researchers typically study avoidance of neutral stimuli by using conditioning procedures, where participants learn to emit an experimenter-defined response (e.g., pressing a button) to avoid a threatening cue (e.g., presentation of loud noise). Although such tasks can help to discover the relevant learning mechanisms, it is unclear whether they can also reveal individual differences in learning. Here, we recruited a large sample of individuals (>400) and tested the relationships between anxiety-related characteristics (i.e., trait anxiety, intolerance of uncertainty, and neuroticism) and performance in an avoidance conditioning task, where participants could avoid unpleasant cues by means of a button press. Using network and cluster analyses, we were unable to identify a strong relationship between avoidance conditioning and the individual characteristics. We discuss our findings and suggest that a paradigm shift in this experimental task is urgently needed for studying individual variations in avoidance learning.
Excessive avoidance of essentially innocuous stimuli is a diagnostic feature of anxiety-related disorders (American Psychiatric Association, 2013). Therefore, laboratory avoidance learning procedures are used to study how such avoidance is acquired (Krypotos et al., 2018; LeDoux et al., 2017). In avoidance learning procedures, participants typically undergo a combination of Fear Conditioning and Instrumental phases (Krypotos, 2015; LeDoux et al., 2017). During the Fear Conditioning Phase, participants encounter neutral stimuli (Conditioned Stimuli or CS+; e.g., a picture of a lamp). Some of these are followed by the presentation of an unpleasant stimulus (Unconditioned Stimulus or US; e.g., the sound of a scream) and others (CS-) are not. During the Instrumental Phase, participants can cancel the presentation of the US by emitting an experimenter-defined response (e.g., button press) during the CS + presentation. These procedures typically result in more fear of the CS + than the CS- and more relief after avoiding the US presentation.
Given the growing popularity of avoidance learning procedures (LeDoux et al., 2017), we recently evaluated their validity and found that, while avoidance learning procedures demonstrate good face validity, they may be insufficient for revealing individual differences in avoidance (Krypotos et al., 2018). Although some studies have provided strong evidence of a relationship between individual differences and avoidance learning, San Martín et al. (2020), Morriss et al. (2021), and Wong et al. (2023), in their recent systematic review, found only mixed or even null evidence that a plethora of individual differences (e.g., trait anxiety and anxiety sensitivity) influence avoidance learning. As argued by Wong et al. (2023), these findings could be due to the simplicity of avoidance procedures, which result in the rapid acquisition of “avoidance-no US” contingencies and therefore offer only limited heterogeneity for investigating individual fluctuations in learning (see Beckers et al., 2013; Lissek et al., 2006 for similar arguments in fear conditioning procedures). Although all individuals will undergo multiple threatening events in their lives, not everyone will go on to develop an anxiety-related disorder. Therefore, to ensure its construct validity, an avoidance learning task should account for this heterogeneity in responses.
To overcome these problems, it has been suggested that ambiguous learning procedures should be used (Beckers et al., 2013; Krypotos et al., 2018; Lissek et al., 2006), such as avoidance generalization. In avoidance generalization tasks (Glogan et al., 2020; Meulders & Vlaeyen, 2013), participants need to decide whether to avoid stimuli that resemble the CS + or the CS- (Generalization Stimuli or GSs) but have never been followed by a US. This generalization is critical for understanding how stimuli associated with an unpleasant event (e.g., being bitten by a Labrador) could contribute to the spread of fear and avoidance of similar stimuli (e.g., any dog). Indeed, there is some emerging evidence (e.g., San Martín et al., 2020) that several anxiety traits (e.g., distress tolerance and intolerance of uncertainty) predict avoidance generalization (see Wong et al., 2023). However, this has not been examined yet in a large sample size, using multiple trait and state characteristics, and employing a multiverse analytic approach (Lonsdorf et al., 2022), that allows to see whether the results depend on the analytic strategy.
The goal of this paper is to extend previous findings (e.g., Wong et al., 2023) by systematically testing whether performance in avoidance learning procedures relates to individual differences relevant for anxiety-related disorders. For doing that, we tested a large number of participants online (>400) with an avoidance acquisition and generalization procedure. All participants filled in anxiety-relevant questionnaires that have been previously associated with avoidance learning and/or anxiety disorders. These were Intolerance of Uncertainty (IU) (Flores et al., 2018), Distress Tolerance (DT) (San Martín et al., 2020), Anxiety Sensitivity Index (ASI) (Wilson & Hayward, 2006), State-Trait anxiety inventory (STAI) (Klein et al., 2020), Neuroticism (Lommen et al., 2010), and Positive and negative affect (I-PANAS-SF) (Schlauch et al., 2013). It was hypothesized that higher, compared to lower, levels of anxiety traits would predict stronger avoidance learning, operationalized as higher avoidance responses to the CS + compared to the CS-, and avoidance generalization, operationalized as higher avoidance responses to the stimulus closer to the CS + than the CS-. In view of the absence of consensus on how to detect individual differences in conditioning tasks (Lonsdorf et al., 2022), we employed different analytic procedures for investigating this hypothesis, including network analysis, that have been previously used in avoidance and fear learning literature (e.g., Krypotos, Baas, & Engelhard, 2020; Krypotos & Engelhard, 2018; Lonsdorf et al., 2022).
Methods
Participants
Participants fluent in English between the ages of 18 and 35 were approached through social media advertising, personal contacts, and the participation exchange website “surveyswap.io,” as well as through the Utrecht University SONA system. Exclusion criteria were self-reported pregnancy, cardiovascular or other medical conditions, past or present psychiatric disorder, being instructed to avoid stressful situations by medical professionals, color blindness, or any hearing problem. As in our previous studies (Krypotos, Baas, & Engelhard, 2020; Krypotos & Engelhard, 2018), we excluded individuals with past or present mental psychiatric disorders due to ethical reasons related to the content of the stimuli. Participants were reimbursed with course credit and they also had the option of participating in a lottery for a 50€ amazon gift card. The study was approved by the Faculty Ethical Review Board Utrecht University (Reference number 21-2190). The study design was preregistered at OSF: https://osf.io/yn8vh. Please note that this was an exploratory study and we had not preregistered all the analyses, and performed more analyses in addition to the preregistered ones.
A total of 402 participants were tested within the preregistered time period. We excluded four participants whose reported age was outside the set age range for the study. In addition, we excluded 40 participants who reported that they had removed their headphones during the experiment. The final sample size was thus 358 individuals (age: M = 24.32, SD = 3.19; sex: 212 females, 145 males, and 1 other). Data collection took place in the whole academic year of 2021–2022.
Materials
Questionnaires
Anxiety Sensitivity Index
The Anxiety Sensitivity Index (ASI-3) was used to measure sensitivity to anxiety-related sensations, (Taylor et al., 2007). Individuals can rate 16 items on a 5-point Likert scale ranging from 0 (very little) to 4 (very much). In the current study, the Cronbach alpha coefficient was α = .875.
Distress Tolerance Scale
The Distress Tolerance Scale (DTS) is a 15-item self-report measuring four factors: tolerance, absorption, appraisal, and regulation (Simons & Gaher, 2005). The results indicate how capable a person perceives themselves to withstand stress. Each item can be rated from 1 (strongly agree) to 5 (strongly disagree). In the current study, DTS showed good internal consistency with Cronbach’s alpha of 0.88.
Intolerance of Uncertainty
Intolerance of uncertainty scale (IUS) (Buhr & Dugas, 2002; Freeston et al., 1994) of 12 items was used to measure a person’s negative beliefs about uncertainty and its consequences. The participant is instructed to indicate how strongly they identify with statements, like “Unforeseen events upset me greatly,” on a 5-point Likert scale (1 = “not at all characteristic of me” to 5 = “entirely characteristic of me”). The computed Cronbach alpha coefficient for the present study was α = .904.
Positive and Negative Affect Schedule
The Positive and Negative Affect Schedule (PANAS) is a 20-item scale that measures 10 positive (i.e., active, determined, attentive, inspired, and alert) and 10 negative (i.e., afraid, nervous, upset, hostile, and ashamed) affect states (Karim et al., 2011). These states are indicated on a five-point Likert scale (1 = very slightly or not at all, 5 = extremely). In the current study, Cronbach alpha was α = .740 for the positive affective portion of the scale and α = 0.728 for the negative affective subscale.
Neuroticism
The neuroticism subscale of the Big Five Inventory (BFI) was used to measure the personality trait neuroticism (John & Srivastava, 1999). It consists of 8 questions rated on a 5-point scale ranging from 1 (“disagree strongly”) to 5 (“agree strongly”). In our study, the subscale had excellent reliability (a = .872).
Spielberger State-Trait Anxiety Inventory
The trait anxiety portion of the Spielberger State-Trait Anxiety Inventory (Spielberger, 1983) consists of 20 statements. Participants are asked to rate to what extent each statement applies to themselves (e.g., I feel pleasant) in general, using a 5-point scale ranging from “almost never” (0) to “almost always” (4). Cronbach alpha of the questionnaire was α = 0.714.
Stimuli
CSs and GSs
We used the same stimuli as in San Martín et al. (2020) (see Figure 1). CSs were three pictures depicting an office room with a desktop lamp in three different colors: yellow (580 nm), green (502 nm), or red (642 nm). GSs consisted of morphed colors: two colors were between yellow and red (600 nm and 620 nm) and two colors were between yellow and green (560 nm and 540 nm). (A) Schematic Representation of the Main Experimental Procedure. During Fear Conditioning, the Yellow or Red Lamps Were Followed by the US (CS +), whereas the Green Lamp was Not (CS-). During Avoidance Conditioning and Avoidance Generalization, Whenever Participants Would See the Red Button (Avoidance Cue) They Could Press the Spacebar and This Would Cancel the US During Presentation of the Yellow Lamp (CS + av) but Not During the Red Lamp (CS + un). During Avoidance Generalization, Two Morphed Color Lamps Between CS- and CS + av and Two Morphed Color Lamps Between CS + av and CS + un Were Also Presented. None of These GSs Were Followed by a US, but Participants Could Still Use the Spacebar. The Numbers Within Brackets Denote the Number of Trials for Each stimulus. (B) Trial Flow During the Fear Conditioning Phase. (C) Trial Flow During the Avoidance Conditioning and Avoidance Generalization Phases
US
The US was a fear-inducing female scream from the International Affective Digitized Sound (IADS) (Stevenson & James, 2008). Using a sound as aversive stimulus has been shown to be effective as a US (Morriss et al., 2016).
Avoidance cue
The avoidance cue was a picture of a red button.
Ratings
Fear ratings
Fear of CSs and the GSs was measured with the question “How afraid were you of the color of this lamp?”, with anchors “not at all” (0) to “very much” (10).
US expectancy ratings
Expectancy of a US-occurrence was measured after each block with the question “How much do you expect to hear a sound after this image?”, rated on a scale ranging from “not at all” (0) to “very much” (10). During the instrumental and generalization phases, participants were also asked to rate their expectancy of a sound occurrence after they have pressed the button with the following question: “How much did you expect a scream during the presentation of this lamp AFTER you have pressed the spacebar?” on a scale ranging from “not at all” (0) to “very much” (10).
Relief ratings
Relief about avoiding the US was assessed using a slider ranging from “neutral” (0) to “very pleasant” (10) and the question “How pleasant was the relief that you felt?”.
Sound ratings
Unpleasantness of the sound was assessed with the question “On a scale from 0 to 10, how unpleasant did you find the sound?”. Participants could use a slider that ranged from “Not at all unpleasant” (0) to “Extremely unpleasant” (10).
Procedure
The online experiment started with the informed consent procedure. By accepting the consent form, participants also agreed that none of the exclusion criteria applied. Then, they were asked to fill in the PANAS, which was followed by instructions to put on their headphones, and turn up the computer volume to 70%. Then, they were introduced to the aversive sound and were asked to rate its unpleasantness.
The main experimental paradigm consisted of three phases: Fear Conditioning, Avoidance Conditioning, and Avoidance Generalization. During the Fear Conditioning Phase, participants were instructed that they would see colored office lamps and that some of them potentially would be followed by the aversive sound, whereas others would not. Then, they saw the baseline stimulus (lamp switched off) followed by CS + un, CS + av, or CS- for 4s. In this phase, both CS + un and CS + av were followed by the US, whereas CS- was not. The phase included 16 trials (4x CS + un, 4x CS + av, 8x CS-) divided into two blocks, with an equal presentation of CSs per block. The US was presented for 4s. The inter-trial interval across phases randomly ranged between 2 and 4s. In the avoidance conditioning/ instrumental phase, participants were told that in the following phase, if they would see the red button on the screen, they could use the spacebar to potentially omit the presentation of the aversive sound. This phase consisted of two blocks of 12 trials (4x CS + un, 4x CS + av, 4x CS-). Stimulus duration was 2s for the baseline picture, 1 s for each CS, 2s for the red button, and 1s for the picture without the button. Again, the CS + un and CS + av were followed by the US, unless the spacebar was pressed during the CS + av. In that case, US presentation was omitted, and a blank screen was presented for 1s instead. We used 2 CSs, in line with previous studies (e.g.,San Martin et al., 2020), which allows testing the action→ omission association to CS + av and an action→no_omission association to CS + un. After each no-US trial, a relief rating question was presented.
Then, in the generalization phase, to measure fear and avoidance responses towards stimuli similar to the CSs, participants were given the same instructions as in the Instrumental Phase. Then, three blocks of 21 trials (3xCS, 4xGS, per block) were presented. GS1 through GS4 were presented like the CS + av with an option to avoid, although generalization stimuli were never followed by the US. Apart from that, the phase was identical as the avoidance conditioning phase.
After each block, participants, across phases, were presented with US expectancy ratings and CS fear ratings. After task completion, they were asked to fill in the rest of the questionnaires: ASI, DTS, IOU, BFI-N, and STAI. To ensure that participants followed the instructions, they were asked whether they removed their headphones or altered the volume during the experiment. Participants were also asked about their age, sex, and educational status. Lastly, participants were provided with a debriefing form, which included the option to fill in an email address to participate in the lottery.
Data Analysis
As in our previous studies, US-expectancy, fear, and relief were analyzed separately for each phase using separate CS/GS x Block repeated measures Analyses of Variance (ANOVAs), with 2 within subject factors (CS and Block). The levels CS/GS and the Block factors were adjusted according to the number of blocks of each phase (see Procedure section). In case the assumption of sphericity was violated, the Greenhouse–Geisser correction was used. In case of significant interactions, we followed the results with post-hoc tests with Bonferroni’s correction. To reassure the readability of the text, we do not report all post-hoc tests here. However, all results are available online: https://osf.io/txcp3/.
We investigated the relationships among the anxiety trait characteristics in a series of confirmatory and exploratory analyses. For the former, we conducted a series of cluster analyses, where we grouped participants based on differential scores of CS + av and CS- during the avoidance conditioning phase using the k-means clustering. Then, we followed the winning number of clusters and performed a non-parametric ANOVA, given that the data were not normally distributed, where we compared the number of clusters to the scores of each one of the questionnaires separately. Similar approaches have been used before (Gazendam et al., 2020).
Exploratory analyses included a series of ANCOVAs as well as running network models. For the ANCOVAs, we performed the same statistical analyses as in the main task, but now after adding trait characteristics as covariates (for a similar approach, see San Martín et al., 2020). Here, we focus only on the CS effects, and not the interactions with block so as to reduce the amount of tests performed and make more meaningful comparisons between tests. For the network analyses, we ran network models to explore the relationship between performance in the main task (i.e., avoidance, relief, and fear ratings) and the sum scores of all questionnaires. For the performance variables, we computed mean scores for each variable (i.e., avoidance, relief, and fear ratings) across each CSs for each block separately. We used EBICglasso and Correlation Network methods in JASP (Love et al., 2019). Centrality measures, including Betweenness, Closeness, Strength, and Expected Influence, were applied to evaluate the relative importance of nodes in the network. Clustering coefficient measures, including Barrat, Onnela, WS, and Zhang, were also used to assess the degree of grouping among nodes.
Results
Fear Ratings
During the Fear Conditioning Phase, there was a difference in fear ratings between the different CSs, F (1.407, 502.442) = 816.922, p < .001, η
2
p
= .0696. The effect of Block × CS interaction on fear ratings was also significant, F (1.829, 652.797) = 5.696, p = .005, η
2
p
= .016, with participants giving higher scores for the CS + un and CS + av compared to the CS- (see Figure 2A). The post-hoc tests of the Block × CS showed that all comparisons were significant, except for the comparisons between CS + av at block 1 and block 2, and CS- at blocks 1 and 2. The higher responses for both types of CS + compared to the CS- indicate that the CS fear manipulation in the Fear Conditioning Phase was successful. Performance Results Across the Different Experimental Phases. (A) Mean Fear Ratings, (B) Mean Expectancy Ratings When the Avoidance Response was Performed, (C) Mean Expectancy Ratings When the Avoidance Response was Not Performed, (D) Mean Avoidance Proportions, and (E) Mean Relief Ratings. Note: FC = Fear Conditioning phase, I = Instrumental Phase, G = Generalization Phase, and 1–3 = block number. Error bars represent standard errors
In the Instrumental Phase, the means of fear ratings also differed significantly between the 3 CSs, F (1.851, 660.894) = 946.149, p < .001, η 2 p = .726. Importantly the CS × Block interaction was significant, F (1.869, 667.284) = 39.789, p < .001, η 2 p = .100. The post-hoc tests for the CS × Block interaction showed that all comparisons were significant except for CS + un for block 1 and block 2, and CS- for block 1 and block 2. The mean difference between block 1 and block 2 for the CS + av was more pronounced than the other 2 CS levels, meaning that the CS + av now caused less fear because participants could in principle cancel the US presentation. These results indicate that manipulation in the Instrumental Phase for fear was successful: participants now showed less fear for the CS + un from the first to the second block, in combination with fear for the CS + av remaining high and fear for the CS- remaining low.
Lastly, for the generalization phase, we observed again a Block × CS interaction, F (8.056, 2876.156) = 38.557, p < .001, η 2 p = .010.
US-Expectancy
The means of US-expectancy in the Fear Conditioning Phase differed significantly between the 3 CSs, F (1.624, 579.607) = 3,214.094, p < .001, η 2 p = .846. The Block × CS interaction was also significant, F (1.853, 661.561) = 11.300, p < .001, η 2 p = .031. The mean difference between block 1 and block 2 for the CS + av was more pronounced than the other 2 CS levels (see Figure 2B). The post-hoc comparisons were all significant except from CS + un for block 1 and block 2, CS + un for block 1 for CS + av for block 2, and CS- for blocks 1 and 2.
In the Instrumental Phase, the means of US-expectancy also differed significantly. Importantly, there was a significant CS × Block × avoidance interaction, F (1.978, 706.296) = 17.088, p < .001, η 2 p = .046. All meaningful post-hoc interactions between the different stimuli, blocks, and avoidance responses were significant except for the comparisons between CS- in block 1 for when participants could and could not avoid (p = .443), CS + un for block 2 when they could and could not avoid (p = .396), CS + un for block 2 when they would and would not avoid (p = .294), and CS- for block 2 when they would and would not avoid (p = .793). Given that participants gave higher US-expectancy responses for the CS + s, compared to the CS-, when they did not press the button, as well that the US expectancy ratings dropped for the CS + av, compared to the CS + un, when they rated their expectancies after they had pressed the button, showed successful manipulation.
For the generalization phase, again there was a CS × Block × avoidance interaction, F (8.516, 3,040.039) = 22.112, p < .001, η 2 p = .058. All meaningful post-hoc tests were significant except for CS + av between blocks 1 and 2 (p = 0.182), GS2 between blocks 1 and 2 (p = .947), GS3 between blocks 1 and 2 (p = .130), and GS4 between blocks 1 and 2 (p = .172). Figure 2D also shows the expected generalization pattern, with more expectancy of a US after stimuli that resembled the CS + un more than the CS-.
Avoidance Proportions
The means of avoidance proportions in the Instrumental Phase differed significantly between the 3 CSs, F (1.515, 714.000) = 1,381.868, p < .001, η 2 p = .795. Post-hoc comparisons showed that there were significant differences across all stimuli, all ps < .001. The Block × CS interaction was not significant, F (1.971, 703.650) = 1.909, p = .150, η 2 p = .005. The post-hoc comparisons showed that all comparisons were significant except for CS + un blocks 1 and block 2, as well as CS + un block 1 and CS + av for block 2. Also, non-significant were the comparisons between CS + av for block 1 and CS + un for block 2, CS- for block 1 and CS- for block 2, and lastly, CS + un block 2 and CS + av for block 2. The higher avoidance proportions for both types of CS + compared to CS- indicate that the manipulation in the Instrumental Phase for avoidance proportions was successful.
During the generalization phase, the CS main effect was significant, F (3.812, 1,361.053) = 369.647, p < .001, η 2 p = .509, with all post-hoc comparisons showing significant differences between stimuli except between CS + av and GS4, CS- and GS1, and GS3 and GS4.
Relief Ratings
For relief ratings in the Instrumental Phase, the CS main effect was significant, F (1, 357) = 125.932, p < .001, η 2 p = .261The mean of relief ratings was significantly higher for CS + av than for CS-, suggesting that the manipulation in the Instrumental Phase for relief was successful. No significant CS × Block interaction was found, F (1, 357) = 1.922, p = .166, η 2 p = .0005. All post-hoc comparisons were significant except from CS + av for block 1 and block 2, and CS- for block 1 and block 2.
During the generalization phase, the CS main effect was statistically significant, F (3.812, 1,361.053) = 369.647, p < .001, η 2 p = .509. All post-hoc differences were significant except from the comparison between CS + av and GS4 (p = .062). The gradient of responses showed that participants increasingly reported more relief for the generalization stimuli that resembled the CS + un more, with most relief being reported for the CS + un and the least for CS-. The post-hoc comparison showed non-significant interactions only for CS + av and GS4, CS- and GS1, and GS3 and GS4.
Individual Differences
Descriptive Statistics of all Used Questionnaires
p-Values per Questionnaire the Cluster Analysis and ANCOVA Analyses. We Denote Significant Results with a Star
Clustering
A cluster analysis was performed based on the avoidance proportions in the Instrumental and Generalization Phase. A K-Means Clustering, which was run in R, showed a result of 8 clusters when BIC was the lowest, BIC = 1,563.000, R 2 = .602. The Hierarchical Clustering was also conducted and showed a result of 8 clusters when BIC was the lowest, BIC = 1,916.210, R 2 = .491. Since the R 2 of K-Means Clustering was larger than that of Hierarchical Clustering, indicating that the grouping of K-Means was able to explain more variance, the grouping of K-Means was used as the grouping for the following analyses. Although the sample sizes of all clusters are larger than the number of clusters, the size of cluster 8 is only slightly larger than 8, which requires caution.
In the main analyses, a one-way ANOVA was performed for each questionnaire (i.e., ASI, DTS, STAI, IUS, ASI, I-PANAS-SF, and Neuroticism). The independent variable was the resulting cluster. The main effect of grouping was not significant for ASI scores, (F (7, 189) = 1.844, p = .081, η 2 p = .064, BF10 = .762), DTS scores (F (7, 189) = .540, p = .803, η 2 p = .020, BF10 = .023), IUS scores (F (7, 189) = 1.681, p = .116, η 2 p = .059, BF10 = .251), PA scores (F (7, 189) = 1.527, p = .160, η 2 p = .054, BF10 = .171), NA scores (F (7, 189) = 1.580, p = .143, η 2 p = .055, BF10 = .186), and Neuroticism scores (F (7, 189) = 1.517, p = .163, η 2 p = .053, BF10 = .136). The grouping did have a significant main effect for STAI scores (F (7, 189) = 2.729, p = .010, η 2 p = .092, BF10 = 1.305). Group 8 scored higher on STAI compared to other groups. Collectively, we found only limited evidence for the prediction of performance during the avoidance phase by individual differences scores.
Analyses of Covariates
We explored whether we would be able to find individual differences based on ANCOVA analyses. We repeated our main analyses including all questionnaire scores in our analysis and only report significant results here (see Supplementary results for the full results tables in JASP).
For fear ratings, we found a significant interaction between CS and STAI (F(1.422, 497.763) = 5.381, p = .011, η 2 p = .015) in the Fear Conditioning phase.
For avoidance data, we found a significant interaction between CS and PA, (F (1.530, 535.541) = 3.526, p = .042, η 2 p = .010) in the Instrumental Phase.
Regarding relief ratings, there was a significant interaction of the CS with neuroticism (F (1, 350) = 39.318, p = 0.038, η 2 p = .012) in the Instrumental Phase.
Lastly, regarding expectancy ratings in the Fear Conditioning phase, there was a significant interaction between CS and Intolerance of Uncertainty (F (1.628, 571.357) = 4.423, p = .018, η 2 p = .012). For the instrumental Phase, there was a significant interaction in CS × PA (F (1.869, 654.282) = 4.489, p = .013, η 2 p = .013). For the generalization phase, we found a significant interaction in CS x PA, (F (6, 3.458) = 4.022, p = .005, η 2 p = .011).
All in all, and despite some significant results, we did not find systematic individual differences between the anxiety traits and any individual differences in the avoidance task. The only exception was the influence of Positive and Negative effect, neuroticism, and (in the Fear Conditioning phase) intolerance of uncertainty.
Network results
The full results of the network analyses can be found in the supplementary material. In summary, we did not find any strong relations between performance in the task and any of the questionnaires. The results suggest that the variables of interest (i.e., performance in the avoidance task and the questionnaires) are not well-connected or easily accessible within the network, and are unlikely to have a significant influence on the network according to all aforementioned measures. Furthermore, the weighted matrix representing the connections or relationships between nodes in the network, which contains information about the strength or importance of each connection or relationship, confirms our conclusions. Overall, the network analysis does not indicate a substantial relationship between these three variables and the others within the network.
Discussion
We investigated whether individual differences in avoidance learning relate to differences in anxiety-related traits, using a large sample size (N >400), and confirmatory and exploratory analytic techniques. Although our confirmatory analyses did not show any such differences, exploratory analyses showed some evidence for the influence of positive and negative affect, neuroticism, and intolerance of uncertainty. We expand on those findings below.
Previous research has focused on finding individual differences in avoidance with various outcomes (San Martín et al., 2020). For example, Morriss et al. (Morriss et al., 2021) found that subscales of Intolerance of Uncertainty could be relevant, among others, for the extinction and generalization of avoidance. However, please note that the previous studies and our study are methodologically different (e.g., lab studies with electrocutaneous stimuli), with different research questions and often employing different statistical approaches (e.g., ANCOVAs with different covariates than us). Our findings do not contrast those previous findings but corroborate also the extensive systematic review of Wong et al. (Wong et al., 2023) who did not find consistent evidence between individual differences and avoidance learning. Taken together, those findings call for further investigation as to why individual differences in conditioned avoidance remain elusive.
We have previously argued about the validity of avoidance learning paradigms and whether they are able to elucidate individual differences (Krypotos, 2015; Krypotos et al., 2018). In light of our findings, we suggest that future studies could use different paradigms to increase the variability in performance so that individual differences may arise. Conditioning procedures typically use a single CS, although in everyday situations individuals may avoid a sequence of stimuli in various degrees (e.g., social anxiety, where individuals avoid social situations because of embarrassment) or rely on more subtle forms of avoidance (i.e., safety behaviors). Examples that use sequence of stimuli (Mobbs et al., 2007) or dimensional avoidance measures (Wong & Pittig, 2022) could be more helpful here. We have also argued that one limitation of the present paradigms is that they do not test the dynamic nature of avoidance as this is expressed in real-life settings, where avoidance is sometimes helpful and in some other cases not. This could be done by different paradigms such as bandit tasks (Krypotos et al., 2022). This could give more room for variations in performance, potentially also making room for studying individual differences.
We note that the deviation between the confirmatory and the exploratory analysis is hard to interpret given that in exploratory analyses someone samples from a much larger analytic space (Krypotos et al., 2019). As such, we urge for future confirmatory analyses that can confirm or disconfirm the present findings.
Our data put into question as to whether “weak” situations, such as that of avoidance generalization, are suitable for investigating individual differences. This of course remains an open question but calls for inquiries beyond generalization procedures and invites further thoughts on alternative procedures, as well maybe as to whether conditioning procedures are equipped in the first place to shade light on individual differences in the first place.
There are some limitations in our study. First, we used an online design, rather than a lab setting, and a scream, instead of an electrocutaneous stimulus that is often used. This means that we did not have full experimental control of the experimental conditions. However, the replication of learning patterns typically found in the lab as well as the high Cronbach’s alphas for all questionnaires showed that our methodology was adequate to draw reliable conclusions. Second, due to the online nature of the study, and despite the manipulation checks, we cannot be sure that participants indeed set the computer volume to 70%. Until there is way to measure to gather such a measure, future studies could include more manipulation checks. One of those is a hearing procedure, where individuals need to score whether they heard, not, tones with different intensities, with different tone intensities being able to be heard only whenever the sound level is above 70%. Third, in our confirmatory and exploratory analyses, we used statistical models commonly found in the literature (e.g., ANOVAs). Given the recent rise of more complicated models for decomposing avoidance performance (Krypotos, Crombez, et al., 2020), future studies could preregister and employ computational models for further exploring individual differences patterns.
It could be mentioned that due to our strict inclusion criteria, our sample size consists of individuals scoring low in anxiety traits, something that could explain the absence of evidence regarding the relationship between the tested traits and performance in the experimental task. Future studies could take advantage of our freely available data and perform further analyses in our data (e.g., Gazendam et al., 2020) or use the available materials for collecting data from individuals with even higher levels in anxiety traits.
All in all, we investigated potential individual differences in conditioned avoidance task. Although confirmatory analyses did not show such differences, exploratory analyses showed some potentially encouraging outcomes. These sporadic findings may call for a potential update in our experimental paradigms or may underline the limits of conditioned learning procedures as capable of shedding light on individual differences in avoidance learning. We hope our work inspires a reconsideration of conditioned avoidance procedures as valid paradigms for modeling excessive avoidance in anxiety-related disorders.
Footnotes
Acknowledgments
We would like to thank San Martin for sharing their experimental materials with us. The authors want to thank the following students for data collection and help during the setup of the project: Bella, Andrade Lima Thomas, Minea Rutar, Katja Schneider, Insa Stangier, Athanasia Kotzasavva, Sarah Saborowski, Georgia Lepenioti, Emma Koek, Margreet Bloemen, Tirza Mulder, Stella Marceta, Jochem Goosensen, Julia Rietra, Delaram Abbasi, Tara Sieling, and Yexin Peng.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Nederlandse Organisatie voor Wetenschappelijk Onderzoek: # 401.18.056, 453-15- 005.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.