Integration of conditioned threat with pre-existing memories

Abstract

How does negative affect spread through existing memories? Whereas many studies have investigated generalization of learned threat responses across perceptual and semantic dimensions, little attention has been given to the possibility that Pavlovian threat responses may spread beyond what is directly learned to previously encoded memories that overlap in content. Here, we increased the demand on associative memory in a modified sensory preconditioning task to investigate this. First, participants encoded 40 unique episodes, each consisting of two neutral stimuli. On the following day, one of each pair was newly associated with either an aversive or a neutral stimulus. Another day later, both stimuli of the original memories were found to trigger enhanced pupil dilation if one was indirectly linked to an aversive stimulus. This effect was independent of whether the associations encoded on day 1 were accurately retained on the day of testing, and confined to trials on which the indirectly associated stimulus was consciously brought to mind, suggesting the formation of a link that directly connects preconditioned stimuli to subsequently learned aversive outcomes. The present study demonstrates that the human defensive system is remarkably adept at quickly anticipating threat based on information acquired over separate events, and gives a first glimpse into the associative structures that enable this ability.

Since the earliest writings on human memory, a central idea has been that different past experiences can become linked together in the mind through overlapping content, meaning an element of one memory can reinstate another (James 1890). Such an ability may be particularly adaptive when the presence of threat can only be inferred by combining information across multiple memories. For example, if one day you meet a couple in a park, and the next day you see one of them walking a dog that attacks you, future encounters with the other partner (who was never directly seen with the dog) could bring back the memory of the attack and trigger psychophysiological defense mechanisms. Aside from being important to understanding how humans learn to predict aversive outcomes in complex and dynamic environments, the question of when and how associations that are predictive of threat reach beyond direct experience also has a clear clinical relevance. So far, a vast literature has investigated anticipatory changes in physiology in response to stimuli that bear some resemblance (e.g., in terms of shape, color, or semantic category) to a learned predictor of threat, and whether this is amplified in patients with anxiety disorders (Lissek et al. 2014; Duits et al. 2015; Cooper et al. 2022). Much less attention has been given to scenarios such as the one described above, where a defensive response has to come about through an indirect association between elements of two distinct memories. Indeed, as Dunsmoor et al. pointed out more than 10 years ago (2011), human studies into this phenomenon were nearly nonexistent then, and can still be counted on one hand today (White and Davey 1989; Vansteenwegen et al. 2000; Dunsmoor et al. 2011; Wong and Pittig 2022).

The spread of defensive physiological threat responses to indirectly associated elements in memory has previously been studied primarily in animals with the sensory preconditioning (SPC) paradigm (Brogden 1947). In an SPC study, two neutral stimuli are repeatedly paired together, after which one of them takes on the role of a conditioned stimulus (CS) that predicts a motivationally significant outcome (unconditioned stimulus, US). Following this procedure, the other neutral stimulus (preconditioned stimulus, PS) is presented again and typically evokes a conditioned response similar to the CS. Since the initial association between two neutral stimuli is formed in the absence of reward or punishment, it does not elicit a conditioned response, and its presence and/or strength can therefore not be gauged by conventional threat conditioning methods. However, there may be a solution in applying methods from other fields. The current lack of research into how and when predictive threat responses spread to existing associations in memory may, in part, be explained by the historical and methodological separation between threat conditioning and declarative memory research (Dunsmoor and Kroes 2019). Here, we aimed to investigate the premise conditions for learned threat to spread through pre-existing associations. To do so, we developed a novel SPC-inspired paradigm capable of measuring both conditioned responses and declarative associative memories (de Vries et al. 2022a).

Two mechanisms by which an initially neutral association between two elements can enable a preconditioned response are frequently discussed in the SPC literature. The first is by “chaining.” This entails the PS cueing retrieval of the CS representation at the time of testing, which then in turn activates the associated US representation and triggers a physiological response (Rizley and Rescorla 1972). A second option is “online integration,” which states that the retrieval of the PS is cued by the CS during conditioning, after which both the PS and CS are able to trigger a physiological response (Wong et al. 2019). The first systematic investigation into the relational structure of preconditioned stimuli in animals showed that extinction of the CS also extinguished conditioned responding to the PS (Rizley and Rescorla 1972). This was taken as evidence against the theory that SPC procedures create a direct link between PS and US, and in favor of the idea that PS responses are mediated by retrieval of the CS at the time of testing (chaining). Yet, recent neuroscientific studies have cast doubt on this conclusion (Wong et al. 2019; Robinson et al. 2022). For example, a series of SPC experiments in rats demonstrated that inhibiting the perirhinal cortex during the conditioning phase abolished freezing responses to the PS, but not the direct CS (Wong et al. 2019). The fact that preconditioned responding was impaired by a manipulation at the time of conditioning, rather than at testing, was taken as evidence for online integration. Similar to work in animals, human brain imaging studies using the SPC paradigm with monetary reward as the reinforcing stimulus have found evidence for both online integration (Wimmer and Shohamy 2012) and chaining (Barron et al. 2020; Wang et al. 2020). Overall, these findings highlight that online integration and chaining are not mutually exclusive strategies, and which one underlies a particular PS response may depend on a variety of factors during all three phases of an SPC experiment (Holmes et al. 2022).

In cases when the mechanism is online integration, an important question to ask is “What exactly has been integrated?” Here, too, there are two options. The first possibility is that the affective value of the US has been directly integrated with the PS representation. This means that, following the conditioning phase, a previously neutral PS has become an affectively charged stimulus that may trigger physiological responses without the need for retrieval/prediction of the US at the time of testing. Alternatively, the PS forms an association with the stimulus representation corresponding to the US. In the declarative memory literature, the integration of stimuli that are indirectly connected by association with a shared element (like in SPC) is typically studied with triads of all neutral stimuli, and variably referred to as “integrative encoding” (Shohamy and Wagner 2008) or “retrieval-mediated learning” (Zeithamova et al. 2012). In this form of integration, a PS remains itself affectively neutral, but may lead to retrieval/prediction of an indirectly associated US, which in turn triggers a physiological response. Consistent with contemporary terminology of reinforcement learning, we will refer to these different forms of integration as “model-free integration” and “model-based integration.”

The three hypothesized mechanisms, chaining, model-free integration, and model-based integration, are summarized below in Figure 1A–C. Chaining differs from the other two mechanisms in that it critically depends on the preconditioned association at the time of testing, meaning it is the only mechanism that cannot generate preconditioned responses when the original PS → CS association is forgotten. We used a large number of PS → CS → US stimulus triads that were learned and tested across three consecutive days. The idea behind this paradigm is that by increasing the number of associations to be learned, we can expect a substantial proportion to be forgotten at the time of testing, such that preconditioned responses can be contrasted between trials for which the preconditioned association is still intact versus not intact. Additionally, we aimed to test whether preconditioned responses come about through model-free or model-based predictions at the time of testing. Model-free predictions are defined by the fact that they do not require taking mental “steps ahead” (Bach and Dayan 2017). Rather, they are intrinsically tied to a currently perceived state. In such a case, it can be said that a CS (or PS) has itself acquired a certain affective value somewhere in the learning process, which is independent of the US representation. To enable a test of model-free versus model-based preconditioned responding, participants indicated whether or not each PS triggered conscious awareness of its indirectly associated US at the time of testing. If preconditioned responses can occur in the absence of US prediction, this constitutes evidence for model-free learning. We used pupil dilation as an index of conditioned and preconditioned responding. Finally, as a first exploration of the clinical relevance of observed effects, participants completed the State-Trait Anxiety Inventory (STAI-T) to test whether individual differences in self-reported anxiety are associated with enhanced preconditioned threat responses.

Figure 1.

Proposed mechanisms by which associative memory may enable preconditioned threat responses. Neutral events are shown in blue, affective events in orange. Dark colored events are directly perceived, and light colored events are mental. (A) Chaining: at the time of testing, PS presentation results in retrieval of the CS, which in turn activates the US representation, leading to pupil dilation. (B) Model-free integration: during conditioning, presentation of the CS results in retrieval of its associated PS, which integrates with the affective value of the US. At testing, the affectively charged PS triggers pupil dilation without the requirement of retrieving any associated stimuli. (C) Model-based integration: during conditioning, the PS is retrieved and forms a direct association with the US. At the time of testing, the PS leads to retrieval/prediction of the US, which in turn triggers pupil dilation.

Results

Participants were excluded from analyses when over 50% of their pupil responses during the test phase, in either PS condition, were based on over 50% interpolated data (n = 4). The final sample thus included 34 participants.

Manipulation checks

Negative stimuli (US⁺) trigger increased pupil dilation

An ANOVA with Condition (US⁺, US⁻) and Block (1, 2, and 3) as factors revealed that, during conditioning on day 2, pupil dilation in response to US⁺ stimuli was larger than to US⁻ (F_(1,2593) = 172.52, P < 0.001), and decreased over learning blocks (F_(2,2531) = 31.54, P < 0.001). There was no significant interaction (F_(2,2593) = 1.32, P = 0.268), indicating that the effect of Condition remained stable across repeated stimulus presentations of the aversive unconditioned stimuli, as compared to their neutral counterparts (Fig. 2A).

Figure 2.

Directly conditioned and unconditioned responses on day 2 of the experiment. (A) Pupil responses to CS⁺ and CS⁻ did not differ in block 1, but diverged after the predictive associations were learned over the course of blocks 2 and 3, indicating successful threat acquisition. (B) Pupil responses to the US⁺ did not habituate—they remained consistently higher than to US⁻ throughout the experiment, even though pupil responsiveness gradually decreased over learning blocks. Error bars represent 95% confidence intervals. Asterisks indicate statistically significant differences between conditions (*P < 0.05, **P < 0.01, and ***P < 0.001).

Conditioned responses are acquired on day 2 and retained on day 3

To assess whether conditioned stimuli came to elicit anticipatory physiological responses after repeated CS → US pairings, we again conducted an ANOVA with Condition (CS⁺, CS⁻) and Block (1, 2, 3) as factors (Fig. 2B). There were significant main effects of both Condition (F_(1,2654) = 19.97, P < 0.001) and Block (F_(2,2653) = 22.28, P < 0.001). Additionally, the interaction between block and condition was significant (F_(2,2653) = 4.83, P = 0.008), indicating that the effect of condition varied across learning blocks. Tukey-adjusted comparisons of marginal means showed a pattern consistent with Pavlovian threat conditioning: Before any stimulus pairing had taken place, there was no significant difference in pupil dilation triggered by CS⁺ or CS⁻ (difference = 0.007 mm, CI₉₅ = [−0.024, 0.038], P = 0.661). Then, after each CS → US association had been presented in the first learning block, pupil dilation to CS⁺ was higher than to CS⁻ in block 2 (difference = 0.040 mm, CI₉₅ = [0.009, 0.072], P = 0.013) and block 3 (difference = 0.078 mm, CI₉₅ = [0.045, 0.110], P < 0.001). This observation replicates our earlier work using the episodic conditioning paradigm (de Vries et al. 2022a,b). Crucially, pupil dilation to CS⁺ remained higher than to CS⁻ the following day (difference = 0.053 mm, CI₉₅ = [0.026, 0.079], P < 0.001), indicating that direct conditioned responses were retained at the time SPC was tested. Average pupil dilation during the third block to CS⁺ numerically exceeded that to CS⁻ for 76.5% (26 out of 34) of the participants.

Associative memory for preconditioned associations

There was no significant difference in memory performance for PS → CS pairs that were later linked to US⁻ or US⁺ stimuli (difference = 1.05%, CI₉₅ = [−6.58, 8.68], P = 0.784).

Main analyses

Negative affect transfers to preconditioned stimuli through episodic associations

We first tested whether directly conditioned responses transferred to previously associated items, whether this effect was moderated by the intact episodic PS–CS awareness, and/or the magnitude of the directly conditioned response, by running a multilevel model with three interacting predictors: Condition (PS⁺, PS⁻), whether the PS → CS association was remembered on day 3 (Premise Remembered: true, false), and magnitude of the retained CR on day 3 (continuous). There was a main effect of Condition (β = 0.073 mm, CI₉₅ = [0.000, 0.145], P = 0.050), but not of Premise Remembered (β = 0.032 mm, CI₉₅ = [−0.026, 0.089], P = 0.285), nor of CR magnitude (β = −0.012 mm, CI₉₅ = [−0.178, 0.155], P = 0.889). There were no statistically significant interactions (all P > 0.082). These results are visualized in Figure 3A–C. This analysis suggests that stimuli which are indirectly predictive of aversive events trigger anticipatory physiological responses through indirect associations. Given that this effect was independent of whether the PS → CS association was still remembered at the time of testing, the present analysis provides no evidence for the “chaining hypothesis,” under which this is a requirement.

Figure 3.

Mean pupil dilation to PS⁻ and PS⁺ during test at day 3 as estimated by the multilevel regression. Panel A shows the main effect of PS condition on pupil dilation. PS⁺, which are indirectly predictive of aversive stimuli, evoke significantly more pupil dilation than PS⁻. This effect was independent of B, whether the declarative association between PS and CS was remembered on the day of testing, and C, the magnitude of the directly conditioned response evoked by the intermediary CS. Error bars and shaded areas represent 95% confidence intervals. Asterisks indicate statistically significant differences between conditions (*P < 0.05, **P < 0.01, and ***P < 0.001).

Preconditioned responding requires correct threat prediction

We next assessed the role of US awareness in indirect threat prediction. After each trial, participants indicated whether or not they correctly imagined the incoming US when the PS was on screen. On average, participants reported to have predicted the correct US on 14.7 trials (SD = 9.3), a wrong US on 12.8 trials (SD = 8.9), and no US at all on 13.5 trials (SD = 8.1). We tested whether this binary factor Imagined (correct, incorrect) interacted with Condition (US⁺, US⁻). The only significant effect to emerge was the interaction between Imagined and Condition (β = 0.069 mm, CI₉₅ = [0.00, 0.149], P = 0.032), indicating that pupil dilation in the PS⁺ condition was only augmented when the outcome was correctly and consciously predicted. As such, the present analysis does not support a role for model-free integration, which is the only hypothesized mechanism that may enable preconditioned responses without bringing the corresponding threat to mind.

STAI-T does not moderate preconditioned responses

To test the hypothesis that high-anxiety individuals are more likely to transfer negative value to harmless elements of overlapping memories, we tested whether participants’ STAI-T scores moderated the effect of Condition. We found no effect of STAI-T (β = −0.001 mm, CI₉₅ = [−0.008, 0.006], P = 0.706), nor an interaction between STAI-T and Condition (β = 0.002 mm, CI₉₅ = [−0.002, 0.005], P = 0.317). Additionally, an analysis of the effect of STAI-T on the likelihood of imagining a negative outcome during PS presentation showed no main effect of STAI-T (β = 0.013 mm, CI₉₅ = [−0.070, 0.092], P = 0.902), nor an interaction with Condition (β = −0.050 mm, CI₉₅ = [−0.150, 0.101], P = 0.538). These results suggest that trait anxiety does not affect either physiological or subjective measures of threat prediction.

Discussion

The present study demonstrates that stimuli which indirectly predict threat through overlapping associations can quickly trigger defensive physiological responses. This effect only occurred on trials for which participants indicated that they correctly predicted the negative outcome, and was independent of declarative memory for premise associations. This pattern of findings is mostly consistent with model-based integration, which states that elements of overlapping memories are proactively associated together before such a link is actually required. Contrary to chaining, model-based integration allows one memory to reinstate the other without depending on mediation by a shared element, and, contrary to the model-free integration, requires correct US prediction. Against our expectations, we found no moderating effect of trait anxiety, thus providing no evidence for a connection between individuals’ self-reported level of anxiety and their tendency to transfer learned threat within a network of pre-existing memories. Together, these findings shed the first tentative light on the associative structures that enable generalization of conditioned physiological responses to previously encoded elements that are indirectly associated with threat.

Previous studies using SPC paradigms in humans have shown that conditioned threat responses transfer to arbitrary neutral stimuli that were previously associated with the CS (White and Davey 1989; Vansteenwegen et al. 2000; Dunsmoor et al. 2011; Wong and Pittig 2022). Our results corroborate and extend these findings by showing that indirect predictors of threat can trigger defensive responses within the same brief time window that is used to detect automatic threat responses in Pavlovian conditioning experiments (Lonsdorf et al. 2017), despite using a cognitively demanding hippocampus-dependent task (Zeithamova et al. 2012) and relatively long time intervals between encoding and testing. This approach enabled testing of the role of pre-existing memory associations in the transfer of defensive responses, and thereby highlights the efficacy of human relational memory, which must encode overlapping events in some way that preemptively facilitates indirect threat prediction.

One way by which model-based integration may be implemented is by means of a cognitive map that integrates the PS, CS, and US in an abstract stimulus space. Subsequently, presentation of the PS may then bring an aversive US to mind through a mental shortcut, quickly leading to enhanced pupil dilation. Once this shortcut has evolved, retention of the original premise memories is not essential anymore, like we also show here. Another computational implementation is provided by the temporal context model (TCM; Howard and Kahana 2002). Originally designed to account for free recall data, the TCM has been successfully extended to capture findings from a wide variety of episodic memory tasks, including inference across overlapping associative memories (Howard et al. 2005, 2009). It posits an encoding process in which items are incorporated within a drifting contextual representation. Then, items can reinstate their encoding context and vice versa through Hebbian association matrices. Consistent with our data, the TCM is a model of relational memory in which the associations between indirectly related elements of memories are created at the time of encoding by means of an overlapping temporal context (Kumaran and McClelland 2012), after which it can be utilized without having to retrieve a mediating stimulus.

Although model-based integration is the mechanism that is most consistent with our findings, we also hypothesized that such a proactive encoding mechanism could be disproportionally active in high-anxiety individuals, but did not find any effect of trait anxiety on preconditioned responses. A meta-analysis on the relation between anxiety and biased retrieval of threat information has shown that the effect is only consistently present when memory is tested via free recall (Herrera et al. 2017), whereas we made use of cued associative recall. Additionally, studies have demonstrated that anxiety may only enhance the retrieval of episodic threat information when there is a clear causal connection between the predictor and the aversive outcome (Toffalini et al. 2015), or when it can be controlled (Large et al. 2016). Neither of these factors were present in the current design. But although not supported by our data, further investigations into the question whether threat generalization through overlapping memories plays a role in anxiety disorders remain warranted by its clinical relevance, and could nicely complement the relatively vast literature aiming to explain individual differences in anxiety by the tendency to overgeneralize threat responses across perceptual and semantic dimensions (Lissek et al. 2014; Duits et al. 2015; Cooper et al. 2022).

There are five methodological factors that potentially limit the generalizability of our results: First, we rule out chaining as an underlying mechanism on the basis that pupil dilation in response to preconditioned stimuli was independent of declarative memory for the preconditioned association. However, chaining may be significantly slower than integration mechanisms as it requires two successive steps of associative recall. It is therefore possible that by presenting each PS for just 4 sec, our paradigm is restricted by-design to detect only those preconditioned responses that arise from faster mechanisms. It could therefore still be the case that premise-memory-dependent chaining is a common strategy by which defensive physiological responses arise when more time is given. Second, some studies on whether anticipatory responding in more conventional threat conditioning studies is facilitated by model-free or model-based learning suggest that it may vary depending on what measure is used (Sevenster et al. 2012). Whereas fear-potentiated startle seems sensitive to tests of model-free learning, skin conductance and pupil dilation may reflect the workings of a model-based learning system, as we also find in the present study. Research on dissociations between conditioning measures and learning systems is in its early days, but do imply that our conclusions might have been different had we chosen another physiological outcome. If such dissociations indeed prove to be robust, future studies into the integration of threat with pre-existing memories are recommended to include both a measure that reflects model-free, and one that reflects model-based learning. Third, similar to Pavlovian threat conditioning studies, we instructed participants to actively predict the outcome that would follow each PS (Lonsdorf et al. 2017). There has been considerable debate in the emotional episodic memory literature, particularly by clinically oriented researchers, on whether voluntary and involuntary retrieval rely on the same (Berntsen 2010; Cohen and Kahana 2022), or distinct neural substrates (Brewin 2014; Visser et al. 2018). If the latter holds, our findings and interpretations are constrained to settings where an individual has already been alerted about the presence of threat and is actively attempting to retrieve the correct details in anticipation. Note however that this argument equally applies to the findings of any Pavlovian threat conditioning study that used US expectancy ratings, which similarly probe participants to actively predict US onset on the basis of previous experience. Fourth, the fact that just before the testing session participants were instructed on how the 40 triads of PS, CS, and US stimuli were related makes the present study considerably different from conventional SPC experiments. However, we do not consider it plausible that our conclusions regarding associative mechanisms were influenced by this decision. While in principle it is possible that in the few seconds between these instructions and task onset, participants were able to retrieve some of the stimuli and realized an integrated PS and US that otherwise would not have been integrated, it is unlikely they were able to do this for an amount of trials that substantially impacted our analyses. Finally, it is possible that the retention of conditioning and declarative memory test on day 3 were influenced by the tests that preceded them. Future studies using a similar paradigm could consider counterbalancing the order of these tests across participants.

In conclusion, the present study demonstrates that the human defensive system is remarkably adept at quickly anticipating threat based on information acquired over separate occasions, and gives a first glimpse into how this ability is facilitated by associative memory. The weight of the evidence leans more toward model-based integration than chaining and model-free integration. However, it remains possible that even within the same experimental context a single individual employs a mix of these strategies that is difficult to detect by means of behavioral read-outs. Extending the current paradigm with neuroimaging data and representational similarity analysis of data acquired during days 2 and 3 could be a promising method to gain deeper insight into the mechanisms that underly indirect threat prediction, and how these are triggered by factors that vary across encoding and testing.

Materials and Methods

Participants

Previous studies demonstrating preconditioned threat responding have included at least 20 participants per experimental group (White and Davey 1989; Vansteenwegen et al. 2000; Dunsmoor et al. 2011; Wong and Pittig 2022). As sample sizes of 30 or higher have been demonstrated to yield reliable parameter estimates in multilevel models (Maas and Hox 2005), we aimed to include at least 30 participants in our final sample. Anticipating some drop-out, 49 participants were recruited through the University of Amsterdam's online system and passed a screener survey that tested for the following exclusion criteria: physical or neurological illness, having received treatment for a mental disorder recognized by the DSM-5 in the last year, ever having experienced a traumatic event, average consumption of over 21 units of alcohol per week, and recreational drug use more than once per week. All participants signed informed consent forms and completed the first session of the experiment. During the second session, four quit the experiment due to the aversiveness of the stimuli. Additionally, one participant could not start the second session because of a technical error with the eye tracker, and another was excluded for not complying with task instructions. Three participants did not show up to their appointment for the final session of the experiment. Finally, two participants that completed all sessions had to be excluded for having participated in other studies that made use of the same emotional stimuli. This resulted in a final sample consisting of 38 participants. However, pupil data of the second session was lost for seven participants because of a technical error. These are thus not included in analyses of unconditioned responses and acquisition of conditioned responses on the second day, which served as manipulation checks. This study was approved by the Ethics Review Board, Faculty of Social and Behavioral Sciences, University of Amsterdam.

Stimuli

Eighty images of neutral objects were selected from the Bank of Standardized Stimuli (BOSS; Brodeur et al. 2010) to function as preconditioned (PS) and conditioned stimuli (CS). As unconditioned and control stimuli, we used the negative (US⁺) and neutral (US⁻) image/sound combinations from the episodic conditioning paradigm (de Vries et al. 2022a), which reliably elicit anticipatory responses after only few CS → US pairings, despite the relatively large number of CS–US associations to be learned. To reduce variance in luminance of the images, which is a potential source of noise in pupil data, the mean luminance of all images (PS, CS, and US) was set equal using the SHINE toolbox for MATLAB (Mathworks 2004; Willenbockel et al. 2010). Sound settings in our laboratory were such that the maximum amplitude of each US was capped at 72 dB to avoid confounding effects of loud noises (Liao et al. 2016).

Tasks and procedures

Day 1—Preconditioning

On the first day, participants were provided an information brochure concerning the experiment and signed an informed consent form, after which they took place in front of the computer screen. Four questionnaires were then presented in the following order: the trait component of the STAI-T (Spielberger et al. 1983), Beck's Depression Inventory (BDI; Beck et al. 1996), the Spontaneous Use of Imagery Questionnaire (SUIS; Reisberg et al. 2003), and the Plymouth Sensory Imagery Questionnaire (Psi-Q; Andrade et al. 2014). These questionnaires were included for exploratory purposes and, with exception of the STAI-T, not further analyzed. The experimenter then explained that in the first part of the session, various pairs of stimuli would be sequentially presented on the screen, and the participant's task was to use each pair to imagine a vivid story in which they play a central role. Actively imagining relations between stimuli within a coherent narrative recruits elaborative processes that are beneficial to associative memory (Craik and Lockhart 1972), and thereby increases the odds of forming a link between PS and CS through which a conditioned response may transfer. At the beginning of each trial a fixation cross was presented for 500 msec. Then, the PS of a given pair was presented for 4 sec, immediately followed by the to-be CS for another 4 sec, and an intertrial interval that varied randomly between 8 and 12 sec. After all 40 pairs had been presented, they were repeated in a second and third learning block. Participants were given 1-min breaks between learning blocks. During the third learning block, participants were asked to rate the vividness of their imagined story for each PS → CS pair on a visual analog scale ranging from 0 (Not at all vivid) to 100 (Very vivid). Following the encoding task, participants’ declarative memory for all PS → CS pairs was assessed by means of an associative recognition test. Each PS was presented for 4 sec, after which six CSs appeared on the screen from which the paired image had to be selected using the buttons 1–6 on the numpad (self-paced). The session concluded with a brief exit questionnaire on participants’ motivation to follow the instructions for each task.

Day 2—Conditioning

Participants returned to the laboratory for the conditioning session the day after preconditioning. First, the eye tracker was calibrated, after which participants were instructed that today's task would again be to use pairs of stimuli to imagine stories in which they themselves play a central role, with the notable difference that the second stimulus of each pair can be either emotionally neutral (US⁻) or negative (US⁺), and would always be accompanied by a matching sound played through a pair of headphones. Each trial started with a fixation cross for 500 msec, after which a CS was presented for 4 sec, followed by a US for 4 sec, and an intertrial interval ranging between 8 and 12 sec. After each trial, participants used VASs to rate both the arousal (calm to tense) and valence (negative to positive) they experienced on a scale from 0 to 100 while imagining their story. All CS → US pairs were repeated in a second and third learning block, and during the third learning block participants were again asked to indicate the vividness of each imagined story on a visual analog scale ranging from 0 (Not at all vivid) to 100 (Very vivid). Participants were given 1-min breaks between learning blocks. The task was programmed such that no more than three pairs from the same condition were presented consecutively. The conditioning session was also followed by an associative recognition test in which each CS was presented for 4 sec, after which the paired US was to be selected from six options presented on the screen: the target US, two USs from the same affective condition as the target US, and three USs from the other condition. Finally, participants completed an exit questionnaire inquiring after their motivation to comply with the instructions for each task, as well as the degree to which they experienced the stimuli as emotionally distressing.

Day 3—Testing

The final session took place the day after conditioning, and began with calibrating the eye tracker. Participants then completed three tests. The first was to measure pupil dilation to indirectly conditioned stimuli (SPC test: PS → US) indicating the degree of affective transfer, the second to measure the retention of directly conditioned responses (CS → US), and the third to assess retention of declarative memory for preconditioned associations (PS → CS). There were 1-min breaks between tests. The session concluded with an exit questionnaire regarding participants’ motivation to comply with instructions for each task, after which they were debriefed on the study's true aims.

SPC test

Participants were reminded that on the first day of the experiment they used pairs of pictures to imagine stories, and that on the second day, the latter picture of each pair was presented again to be used in another imagined story involving a third picture. The experimenter then explained that the pairs of stimuli encoded on days 1 and 2 were connected by an overlapping element, and asked the participant to repeat this back in their own words to make sure they had fully understood the connection between PS, CS, and US (or stimulus A, B, and C in the language of the cover-up “imagination task”). The reason why participants were told about the task structure is that, even though they were not instructed this way, after some number of trials they most likely would still realize how the stimuli they associated on each day were connected. We intended to keep participants’ understanding of the task consistent throughout the testing phase, and therefore opted for these instructions.

During each trial of the preconditioning test, a PS was presented for 4 sec, and immediately followed by 4 sec of its indirectly paired US (Fig. 4). Participants were instructed to, while the first picture of a stimulus pair was on screen, imagine as vividly as possible which picture/sound combination would be presented next. Then, after the US had disappeared off the screen, they answered three questions about what they imagined during the 4 sec of PS presentation: They were first asked to indicate the vividness of whatever it was they imagined would be presented next on a visual analog scale ranging from 0 to 100, then whether what they imagined was indeed the stimulus that followed and third, and finally, whether what they imagined was emotionally neutral or negative. Questions two and three were answered by pressing buttons on the keyboard corresponding to (Yes or No) and (Neutral or Negative), respectively. An additional button (None) could be used to answer both questions when they did not imagine anything in particular.

Figure 4.

Overview of task structure on both learning days and the associative inference test phase on day 3. In this example, the pineapple (PS) is presented first, followed by the guitar (CS). Then, on day 2, each CS item is presented and followed by a picture/sound combination (US) that can either be neutral or aversive. All pairs on both days 1 and 2 are presented a total of three times. During the last block of pair presentations on each day, participants are asked to indicate the vividness of their imagined story for each pair. Finally, during the associative inference test, PS items are presented followed by the corresponding US. Participants are asked to predict and imagine the incoming US when the PS is on the screen, after which they are asked whether they indeed imagined the correct item, whether they imagined something neutral or negative, and how vivid their imagination was.

Retention of conditioned responses

The test for retention of conditioned responses was identical to the SPC test described above, except that the directly conditioned associations (CS → US) were presented instead of indirect associations. Participants were again instructed to vividly imagine the upcoming stimulus while the first stimulus of a CS → US pair was presented, after which they answered the same three questions regarding the vividness, accuracy, and valence of what they imagined.

Retention of preconditioned associations

To test whether PS → CS were still remembered on the final day of the experiment, participants completed a final associative recognition test that was identical to the memory test of the first day. Each PS was presented for 4 sec, followed by six CSs from which the correct stimulus had to be selected using the buttons 1–6 on the numpad.

Data acquisition and processing

Task presentation and recording of behavioral data were done using Presentation software (Neurobehavioral Systems Inc.), and pupil data were measured with a Tobii Pro Nano eye tracker sampling at 60 Hz. We used the Python programming language to preprocess the pupil data in the following steps: First, samples within 100 msec from missing values were removed, after which the data was linearly interpolated. Then, a band-pass filter was applied (0.01–6 Hz, third-order Butterworth) to remove high-frequency noise and slow drift from the signal (Knapen et al. 2016). Pupil responses were computed by averaging the values in the time series within the frame of interest (3–4 sec from onset in the case of PS and CS, 2–4 sec from onset for US [Visser et al. 2013; de Vries et al. 2022a]) and subtracting a trial-specific baseline (average of 0.5 sec prior to PS or CS until onset). Trials were excluded if either the mean response value or its corresponding baseline were computed on the basis of more than 50% interpolated data (Leuchs et al. 2019). Participants of which over 50% trials were excluded in either condition were excluded from analysis altogether (de Vries et al. 2022a,b).

Data analysis

First, as manipulation checks, we tested whether the emotionally negative US⁺ stimuli triggered more pupil dilation than their neutral counterparts (US⁻) during conditioning on day 2, and whether this effect habituated over time by running a multilevel model with Condition (US⁺, US⁻) and Learning Block (1, 2, 3) as interacting factors, nested in participants. Similarly, to assess whether the Episodic Conditioning procedure had resulted in threat acquisition, we analyzed changes in pupil dilation to CS⁺ relative to CS⁻ across repeated CS presentations, again by running a multilevel model with Condition (CS⁺, CS⁻) and Learning Block (1, 2, 3) as interacting factors. Since the factor “block” consists of three levels, ANOVA was used to test for significance. We additionally tested for a significant difference in pupil dilation triggered by CS⁺ and CS⁻ during the retention test on the third day, as without any retention it would be unlikely that conditioned responses would transfer to the indirectly associated CS.

To test the main hypotheses, we used multilevel modeling to estimate effects at the level of specific memories (i.e., trials) while simultaneously accounting for the nested structure of trials within participants. Intercepts were thus allowed to vary across participants. All analyses were conducted in R (R Core Team 2021) and made use of the package “lme4” (Bates et al. 2015) for multilevel modeling. The variables Condition (PS⁺, PS⁻), Premise Remembered (true, false), and the CR measured in the retention test on day 3 (continuous, centered within participants) were included in the main model as interacting variables. Including the factor Premise Remembered tests whether memory for the preconditioned association (PS → CS) moderates the effect of Condition. If memory for the initial PS → CS association is required for preconditioned threat responding, as is predicted by the chaining hypothesis, there will be a significant interaction between Condition and Premise Remembered, but no main effect of Condition, whereas a main effect of Condition is possible under both model-free and model-based integration. We included the CR corresponding to each trial in the model as well for two reasons: (1) it is expected to increase statistical power by zooming in on those trials where direct conditioning was successful, and (2) to test whether preconditioned responding is proportional to the magnitude of the corresponding CR. To also differentiate model-free integration from model-based integration, we tested whether Condition interacts with participants' own indication of whether they correctly predicted the incoming US. Only model-free integration allows for a preconditioned response to occur in the absence of correct prediction of the US, such that a main effect of Condition would be expected. Finally, we assessed if preconditioned responses are enhanced in high-anxiety individuals by running a regression with STAI and Condition as interacting predictors.