Contextual Modulation of Human Associative Learning following Novelty-Facilitated Extinction, Counterconditioning, and Conventional Extinction

Jérémie Jozefowiez; James E Witnauer; Yaroslav Moshchenko; Cameron M McCrea; Kristina A Stenstrom; Ralph R Miller

doi:10.1037/xan0000385

. Author manuscript; available in PMC: 2025 Oct 1.

Published in final edited form as: J Exp Psychol Anim Learn Cogn. 2024 Oct;50(4):267–284. doi: 10.1037/xan0000385

Contextual Modulation of Human Associative Learning following Novelty-Facilitated Extinction, Counterconditioning, and Conventional Extinction

Jérémie Jozefowiez ¹, James E Witnauer ², Yaroslav Moshchenko ³, Cameron M McCrea ³, Kristina A Stenstrom ³, Ralph R Miller ³

PMCID: PMC11911142 NIHMSID: NIHMS2056832 PMID: 39432329

Abstract

The expression of an association between a conditioned stimulus (CS) and an unconditioned stimulus (US) can be attenuated by presenting the CS by itself (i.e., extinction, Ext). Though effective, Ext is susceptible to recovery effects such as renewal, spontaneous recovery, and reinstatement. Dunsmoor et al. (2015; 2019) have proposed that pairing the CS with a neutral outcome (novelty-facilitated extinction, NFE) could offer better protection against recovery effects than Ext. Though NFE has been compared to Ext, it has rarely been compared to counterconditioning (CC), a similar procedure except that the CS is paired with a US having a valence opposite to the US used in initial training. We report two aversive conditioning experiments using the rapid-trial streaming procedure with human participants that compare the efficacies and susceptibilities to ABA renewal of Ext, CC, and NFE. Associative learning was assessed through expectancy learning and evaluative conditioning. CC and NFE equally decreased anticipation of the US in the presence of the CS (i.e., expectancy learning). Depending on how the CS-US association was probed, they were either as or more effective at doing so than Ext. All three interference treatments were equally susceptible to context manipulations. Only CC clearly altered the valence of the CS (i.e., evaluative conditioning). Valence ratings after Ext, CC, and NFE, as well as a no-interference control condition, were all equally susceptible to context effects. Overall, the present study does not support the assertion that NFE is consistently more resistant to recovery effects than Ext.

Keywords: Extinction, counterconditioning, novelty-facilitated extinction, renewal, expectancy learning, evaluative conditioning

In Pavlovian conditioning, pairing an initially neutral conditioned stimulus (CS) with a biologically relevant unconditioned stimulus (US) leads to the formation of a CS-US association (associative learning), the strength of which determines the capacity of the CS to elicit conditioned responses (CR). If the US is an aversive event, the CR typically includes indications of fear and anxiety. For this reason, associative learning is thought to play a central role in anxiety disorders and phobias (e.g., Van Elzakker et al., 2014; Vervliet et al., 2013), and extensive research has been conducted toward reducing the extent to which the CS triggers the CR once the CS-US association has been established.

After conditioning consisting of CS-US pairings, presenting the CS by itself reduces the potential of the CS to elicit a CR (i.e., extinction [Ext]; Pavlov, 1927). Relatedly, counterconditioning (CC; Pavlov, 1927) is a reduction in original conditioned responding that results from the CS being paired with a stimulus having a valence opposite to that of the original US. However, these treatments often wane with procedural changes between extinction and testing (Bouton, 2017). Testing the CS in a context that is distinctly different from the context of Ext or CC increases the target CR strength relative to the CR observed in the Ext or CC context. For example, rats that received CS-US pairings in context A followed by extinction in context B will show renewal of CR strength when tested in the context of the original CS-US pairings (ABA renewal; Bouton & Bolles, 1979). Similar recovery effects are observed when the US is presented in the test context before testing with the CS (reinstatement; Rescorla & Heth, 1975) or when the retention interval between the last Ext or CC session and the test is increased (spontaneous recovery; Pavlov, 1927).

Recently, Dunsmoor and his collaborators (Dunsmoor et al., 2015; Dunsmoor et al., 2019; see also Lucas et al., 2018) have proposed a new procedure, novelty-facilitated extinction (NFE), that reduces expression of the CS-US association. NFE resembles CC in that it involves post-acquisition pairings of the CS with another stimulus, but the paired stimulus is neutral in valence rather than being opposite in valence to the training US. Dunsmoor et al. reported that after initial pairings of a CS with an aversive shock US, NFE is more effective than Ext in reducing the conditioned galvanic skin conductance response (SCR) by humans and the conditioned freezing by rats. Moreover, Dunsmoor et al. (2015, 2019) found NFE to be putatively less susceptible to spontaneous recovery than Ext. Lucas et al. (2018) reached the same conclusion using reinstatement to assess recovery of both SCR and evaluative ratings of CS valence by humans. However, Dunsmoor et al.’s observations might be difficult to replicate; for example, Krypotos and Engelhard (2018) observed that NFE was less effective than Ext in a test immediately after the response-reducing treatment (NFE or Ext. See also Steinman et al., 2022) and equally subject to reinstatement-induced recovery in both affective ratings of the CSs and expectancy of the USs.

An additional limitation of the conclusions reached by Dunsmoor et al. (2015, 2019) and Lucas et al. (2018) is that they did not compare NFE with CC in terms of efficacy or susceptibility to recovery, which is surprising given Dunsmoor et al.’s (2015, 2019) rationale for proposing the use of NFE. Based on the Rescorla and Wagner (1972) model of learning, they posited that change in the CS-US association was a function of the surprisingness of the event following the CS. Although subject to argument, they presumed that the neutral outcome used in NFE was more surprising than the omission of the US in Ext; hence, it should have a greater efficacy than simple extinction which consists of presentations of the CS alone. However, one might suppose that the absence of an outcome as in Ext differs more from target training than does the substitution of the target outcome to an affectively neutral nontarget outcome in NFE. Moreover, extending their theoretical reasoning for expecting NFE to be more effective than Ext, Steinman et al. (2022) suggested that NFE reduces uncertainty more than Ext because detection of a non-event (i.e., no outcome in Ext) is more difficult than detection of a nontarget outcome in NFE. Based on these two factors, Dunsmoor et al. hypothesized that the greater amount of surprise and reduced uncertainty generated by NFE also would result in lower susceptibility to recovery effects, which was consistent with the specific data they presented (see also, Lipp et al., 2020). However, both of their theoretical assumptions are arguable. First, it is not clear why a neutral outcome should be more surprising than the absence of the expected outcome; one might view a cue followed by the target outcome as being more similar to the same cue followed by a different outcome than by no outcome in that the first two have in common the presence of a subsequent event. Second, although the importance of surprise in extinction is central in most models of conditioning including the Rescorla-Wagner (1972) model and is supported by phenomena such as overexpectation (Kremer, 1978), no data to date has linked surprise to recovery effects. Likewise, there is no data supporting a link between recovery effects and uncertainty. Still, suppose we assume that NFE is less susceptible to recovery than Ext because it generates more surprise. If this were true, after pairing the CS with an aversive US in initial training, pairing it with an appetitive US would be arguably more surprising than pairing it with a neutral outcome. Hence, if Dunsmoor et al.’s account is correct, we would expect CC to be even more effective and more resistant to recovery effects than NFE.

Research has shown that CC is frequently more effective than Ext at reducing emotional CRs (for a review, see Jozefowiez et al., 2020). Data on the susceptibility to recovery of reductions in responding due to CC relative to Ext are sparse and, to date, contradictory. Holmes, Leung, and Westbrook (2016) have concluded that CC is more susceptible to renewal than Ext based on experiments that measured freezing in rats. In contrast, Kang et al. (2018), using human participants and measures of both CS valence and US expectancy, have reported that CC is less susceptible not only to renewal (see also Thomas et al., 2012) but also spontaneous recovery and reinstatement. Hence, it seems important to conjointly assess the relative effectiveness of CC and NFE, and their susceptibilities to renewal. Using chocolate as US, Van Gucht et al. (2013) found no difference between Ext and CC regarding their impact on either CS-elicited approach behavior or US expectancy. Otherwise, only CC was able to alter conditioned liking with its effect generalizing between contexts.

Using SCR, Chen et al. (2022) compared NFE with not only Ext but also CC. They found no appreciable difference across the three treatments in either immediate efficacy or susceptibility to spontaneous recovery and reinstatement-induced recovery. However, Chen et al.’s study used electric shock to initially condition the CS, whereas CC consisted of pairing the CS with positive images from the International Affective Picture Scale (IAPS; Lang et al., 2008). Likely, the absolute value of the valence of the electric shock is greater than the valence of positive IAPS pictures. If the valence of a stimulus is coded relative to other stimuli in the same situation, this might have led to the positive IAPS pictures being processed by participants in a way that resembled neutral pictures. This possibility cannot be ruled out because Chen et al. did not ask their participants to rate the valence of the various stimuli they used as USs in their study. A better comparison between CC and NFE could be achieved by using aversive and appetitive stimuli that have at least similar absolute valences and that use the same sensory modality (e.g., visual).

Importantly, the SCR is elicited by a variety of emotionally positive and negative stimuli. In a procedure that includes Phase 1 CS-aversive US pairings and Phase 2 CS-appetitive stimulus CC trials, the participant will learn to expect the appetitive stimulus as a result of CC. However, the expectation of an appetitive stimulus could trigger an SCR that would be indistinguishable from an SCR triggered by an association between the CS and the aversive US. Consequently, other methods for assessing the CS-US association are necessary to compare the efficacies of NFE and CC. Often, participants are asked to rate how likely the CS is to be followed by the US. Another possibility is to have participants rate the valence of the CS: a CS paired with a positive outcome should be rated as more positive than a CS paired with a negative outcome (i.e., evaluative conditioning; for a review, see De Houwer et al., 2001). Such a strategy was followed by Quintero et al. (2024). They paired a visual CS with an aversive sound. The CS was subsequently paired with a positive sound (CC), a neutral one (NFE), or no sound at all (Ext). Assessing associative learning through both expectancy learning (the participants had to rate how likely was the US to appear following the CS) and evaluative conditioning (the participants had to rate the valence of the CS), they did not observe any difference among the three treatments regarding either their efficacies immediately after interference treatment or their susceptibilities to spontaneous recovery. However, Quintero et al. (2024) did not assess the comparative susceptibilities of Ext, CC, and NFE to context effects, more specifically, renewal. This was the goal of the present experiments. We used the rapid-trial streaming procedure developed by Allan and her collaborators (a.k.a. the streamed-trial procedure; Alcala et al., 2023; Crump et al., 2007; Hannah et al., 2009; Jozefowiez, 2021; Jozefowiez & Miller, 2024; Laux et al., 2010; Maia et al., 2018; Murphy et al., 2021; Siegel et al., 2009) that we had previously used in the study comparing CC and Ext (Jozefowiez et al., 2020).

Experiment 1

In Experiment 1, participants were exposed to rapid streams of stimuli divided into two phases. The first phase established an association between the target CS (X) and an aversive US (NEG). Except for the control condition (Ctr), the second phase was intended to reduce expression of the X-NEG association through either Ext (presenting X by itself), CC (pairing X with an appetitive US), or NFE (pairing X with a neutral outcome). The positive and aversive USs, as well as the neutral outcomes, were all IAPS pictures.

Context manipulations were achieved by changing the background over which the stimuli were displayed. Phase 1 of a stream took place in Context A whereas Phase 2 took place in Context B. Testing occurred in either Context A or B depending on experimental condition to assess potential ABA renewal.

During the test, the participants were asked how likely it was that the target CS would be followed by the aversive US in each context (prediction rating). This measured expectancy learning. The participants were also asked to rate how much they ‘liked’ the CS when it was presented in a given context (valence rating). This measured evaluative conditioning. Some authors have suggested that evaluative conditioning is at least highly resistant, if not completely insensitive, to extinction (i.e., Baeyens et al., 1988; Díaz et al., 2005; Gawronski et al., 2015; Vansteenwegen et al., 2006). In a previous study comparing CC and NFE using the streaming procedure (Jozefowiez et al., 2020), we observed that different results are obtained depending on whether associative learning is assessed through expectancy learning or evaluative conditioning.

Methods

Participants

In our previous study (Jozefowiez et al., 2020), the within-subject difference between Ext and CC corresponded to a Cohen’s (1988) d of 0.3, and the correlation between the performance of participants in the Ext and CC conditions varied between 0.3 to 0.7. The power analysis used to select the size of the sample in the present experiment assumed that the effect sizes of Ext, CC, and NFE would be inside of this range. With those assumptions, 80% statistical power for a t-test is achieved with a sample size between 55 and 122 participants, corresponding to correlation strengths of 0.7 and 0.3, respectively. Hence, we sought to obtain data from 100 participants. This and the subsequent experiment were approved by the SUNY-Binghamton Institutional Review Board.

Overall, we ran 180 participants, between 18 and 50 years old without a predisposition for visually-induced seizures or a history of anxiety disorders, recruited from the SUNY-Binghamton subject pool. Eighty participants failed to meet the criterion for inclusion based on ratings during a warmup procedure (see below). The data file from one participant who met the inclusion criterion was corrupted and could not be read. Hence, the analysis below is based on a sample of 99 participants (29 males, 69 females, and one participant who did not indicate gender, with the imbalance reflecting the demographics of the subject pool). The mean age of participants was 18.99 +/− 0.86, ranging from 18 to 21 years old. Even though the target sample size was 100 participants, we recruited 180 participants to compensate for the anticipated loss of participants who failed to meet the inclusion criterion.

Apparatus and Stimuli

The experiment was conducted on Windows PCs in individual cubicles at SUNY-Binghamton. The screen resolution of the computers was 1920 × 1080 pixels, and the monitors were 53.34 cm wide. The experiment used a customized program written in Python using the PsychoPy2 library (Peirce, 2007). Participants used a standard computer mouse to provide their responses.

Six stimuli (cues: X, Y, W; and outcomes: NEG, NEUT, POS) were used in each condition. As there were 8 experimental conditions, 24 neutral cues played the roles of cues X, Y, and W, and 8 aversive, 8 neutral, and 8 appetitive pictures played the roles of NEG, NEUT, and POS, respectively. For each participant, 8 sets of stimuli were created by randomly selecting three images from the pool of 24 cues and randomly assigning them to the roles of Cues X, Y, and W. One stimulus from the pool of 8 aversive pictures, one from the pool of 8 appetitive pictures, and one from the pool of 8 neutral pictures were randomly selected to fill the roles of the NEG, POS, and NEUT outcomes, respectively. All random selection of images was done without replacement. Hence, each condition had a different set of cues and outcomes, with a different random assignment of the cues and outcomes to each condition for each participant. The cues were 450- x 490-pixel neutral symbols. Based on our judgment, they were distinctively different from each other. The aversive, appetitive, and neutral pictures were drawn from the IAPS (Lang et al., 2008; see Table 1). Although the three categories of pictures differed in their valence ratings, matching them regarding arousal was not possible because aversive images are ordinarily more arousing than either appetitive or neutral images, and appetitive images are more arousing than neutral images. Instead, we tried to reduce as much as possible the differences in arousal by picking the most arousing neutral images and the least arousing aversive images. All IAPS pictures were scaled down from their original 1024 × 768 resolution to match the dimensions of the cues.

Table 1.

Identifier, valence, and arousal ratings of the IAPS used as outcomes.

IAPS identifier	Valence	Arousal	Set

#7950 Tissue	4.94	2.28	Neutral
#7059 Keyring	4.93	2.73	Neutral
#7010 Basket	4.94	1.76	Neutral
#7012 Rubber bands	4.98	3.00	Neutral
#7020 Fan	4.97	2.16	Neutral
#7035 Mug	4.98	2.66	Neutral
#7041 Baskets	4.99	2.60	Neutral
#7090 Book	5.19	2.61	Neutral
#3103 Injury	2.07	6.06	Aversive
#3150 Mutilation	2.26	6.55	Aversive
#9571 Cat	1.96	5.64	Aversive
#9301 Toilet	2.26	5.28	Aversive
#9185 Dead Dog	1.97	5.65	Aversive
#9183 Hurt Dog	1.69	6.58	Aversive
#9181 Dead Cows	2.26	5.39	Aversive
#9570 Dog	1.68	6.14	Aversive
#1440 Seal	8.19	4.61	Appetitive
#1441 Polar Bears	7.97	4.61	Appetitive
#1460 Kitten	7.97	3.94	Appetitive
#1710 3 puppies	8.21	4.31	Appetitive
#1750 Bunnies	8.28	4.10	Appetitive
#1920 Dolphins	7.90	4.27	Appetitive
#2045 Baby	7.87	5.47	Appetitive
#2070 Baby	8.17	4.51	Appetitive

Open in a new tab

The stimuli appeared superimposed over a context. The single context for the Warmup conditions and Contexts A and B for each of the 8 experimental conditions (17 contexts in all) consisted of a rectangular image, each framed by a border (see Figure 1 for an example). The background of the image inside the border was 1 of 17 solid colors, each yoked to a distinctly different border pattern. The borders were 1024 × 768 pixels while the internal colored area was 960 × 662 pixels. A white fixation cross (10 × 40 pixels vertically, 46 × 9 pixels horizontally) was displayed at appropriate times at the center of the screen.

Figure 1. — Top: Examples of background and stimuli shown to the participants. Bottom left: Example of a prediction rating screen. Bottom right: Example of a valence rating. The stimuli and background shown in these examples are those used during the warm-up.

A unique set of stimuli (X, Y, W, O1, O2) and a unique context were used during the warmup conditions that preceded the experimental conditions. All stimuli in this set were 450- x 490-pixel, affectively neutral symbols. Those stimuli were used throughout the whole warm-up sequence and were the same for all the participants.

Procedure

All participants were initially required to complete an informed consent form and asked to turn off their cell phones. Upon giving consent, they were seated in individual cubicles. During the experiment, the participants were exposed to various conditions in which they saw streams of trials. After each stream, they were asked to make predictions and provide valence ratings concerning a target cue and the various outcomes to which they had been exposed.

For each condition, a set of three randomly selected cues and three randomly selected outcomes was assigned to a randomly selected pair of contexts (see Tables 2 and 3). Random selections were without replacement across conditions and were made anew for each participant. Each condition, presented in random order, was composed of a stimulus stream followed by the prediction and the valence rating questions.

Table 2.

Composition of the trial streams during warmup in Experiment 1.

Condition	Phase 1 (cycled once)	Phase 2 (cycled twice)

Positive warmup	A: 6 X-O1 / 5 Y- / 6 W-	B: 6 Y- / 5 Y-O2
Negative warmup	A: 6 X- / 5 Y-O2 / 6 W-	B: 6 Y- / 5 Y-O2
Null warmup-1	A: 3 X-O1 / 3 X- / 5 Y-O2 / 6 W-	B: 6 Y- / 5 Y-O2
Null warmup-2	A: 3 X-O1 / 3 X-O2 / 5 Y- / 6 W-	B: 6 Y- / 5 Y-O2

Open in a new tab

Note. ‘n I-J’ means that n trials of cue I and outcome J were presented during each cycle. ‘n I-’ means that n trials of cue I alone were presented during each cycle. The letters A and B indicate in which context the stream was presented. The program cycled once through Phase 1 and twice through Phase 2 before the contingency question was presented. During each cycle, the order of presentation of the trials was determined randomly. X and O1 were the target cues and outcomes. The prediction and valence ratings always occurred in context A.

Table 3:

Composition of the trial streams during experimental conditions in Experiment 1.

Condition	Phase 1 (cycle once)	Phase 2 (cycle twice)

Ctr	A: 5 X-NEG / 5 W-	B: 5 Y-POS / 5 Y-NEUT / 10 Y-
Ext	A: 5 X-NEG / 5 W-	B: 5 Y-POS / 5 Y-NEUT / 5 X- / 5 W-
CC	A: 5 X-NEG / 5 W-	B: 5 X-POS / 5 Y-NEUT / 5 Y- / 5 W-
NFE	A: 5 X-NEG / 5 W-	B: 5 Y-POS / 5 X-NEUT / 5 Y- / 5 W-

Open in a new tab

Note. ‘n I-J’ means that n trials of cue I and outcome J were presented during each cycle.’ n I-’ means that n trials of cue I alone were presented during each cycle. The letters A and B indicate in which context the stream was presented. The program cycled once through Phase 1 and twice through Phase 2 before the contingency question was presented. During each cycle, the order of presentation of the trials was determined randomly. X and O1 were the target cues and outcomes. The prediction and valence ratings always occurred in context A for the ABA streams and context B for the ABB streams. During a block, each of the conditions in Table 3 was presented twice to a participant: testing occurred in Context A on one occasion (ABA condition) and in Context B on the other occasion (ABB condition). Hence, a block was composed of 8 streams: cues, outcomes, and context differed for each of these streams.

Each stream began with a 1000-ms presentation of the context and the fixation cross. The fixation cross remained visible throughout the two training phases of each stream. The Phase 1 trials (see Table 3) were presented in a random order and were followed immediately by the Phase 2 trials which consisted of two cycles of the trials specified in Table 3 with trial order randomized within each cycle. Every trial started with the presentation of a cue (X, Y, or W) for 1000 ms, and potentially 400 ms into the cue, an outcome (POS, NEG, or NEUT) was presented for 600 ms. Trials were separated by a 500-ms intertrial interval (ITI). X was always presented such that its left edge was one stimulus width to the right of the center of the screen; Y was always presented such that its right edge was one stimulus width to the left of center; W was always presented centered in the upper part of the screen, vertically aligned with and between the locations of X and Y. Outcome NEG was always presented in the lower left corner, diagonally opposite the location of X; Outcomes POS and NEUT were always presented centered immediately below the fixation cross (see Figure 1). The intent of Phase 1, which was identical across all conditions, was to establish an association between cue X and outcome NEG. Stimuli were displayed in Context A. Except for the Ctr condition, the aim of Phase 2 was to impair the expression of the X-NEG association established during Phase 1 either through Ext (X presented by itself; we label this event the NULL outcome), CC (X paired with POS), or NFE (X paired with NEUT). During Phase 2, stimuli were displayed in Context B.

Once exposure to a condition was complete, the participant was asked to predict how likely in a specific context (A or B) each of three given cue/outcome configurations was immediately following the presentation of X and to rate the relative valences of the various cues presented during the condition (see Figure 1). As test order might influence responding to the second-asked question, the actual order of testing (i.e., prediction vs. valence) was counterbalanced within-subjects. That is, for half of the streams, the prediction rating occurred before the valence rating, whereas the reverse was true for the remaining streams.

For the prediction ratings (see the lower left panel of Figure 1 for an example), the participant was shown a screenshot of cue X in either Context A (ABA condition) or Context B (ABB condition), as it had appeared during the stimulus stream. Above the cue was the statement: “Imagine you have been shown the following configuration on the screen:” and below the cue was the instruction: “Using the scale below each image, use the mouse to indicate how likely the top configuration is to be followed by each of the below configurations.” Below this instruction, there were four screenshots arranged in a square showing an image of Cue X along with an image of each potential outcome (NEG, POS, NEUT, and the NULL outcome), all presented in either Context A (ABA condition) or Context B (ABB condition). Below the image of each outcome was an 11-point Likert scale ranging from 0 to 100 (incremented by steps of 10) and anchored at 0 (‘Very unlikely’) and 100 (‘Very likely’). Consistently, the mouse pointer initially appeared centered below the “Imagine…” statement at the top of the screen. The location of the four screenshots showing the outcomes was randomized for each participant and each condition. A white circle indicating a specific rating in the Likert scale became black if the participant clicked on it and remained black for 150 ms. Then, that particular outcome image and its associated Likert scale disappeared. 150 ms after the participants had given the last of their four prediction ratings, the prediction rating screen disappeared.

Each valence rating pitted two cues (S1 and S2) against one another (the lower right panel of Figure 1 provides an example). Three valence ratings were presented sequentially in random order: X vs. W, X vs. Y, and W vs. Y. Cue W was always presented exactly as often as Cue X but was never paired with any outcome. Hence, the valence of W should have remained neutral and, therefore, served as a neutral option against which X could be compared by participants. The X vs. W ratings served as the critical dependent variable in data analysis, whereas the other ratings were only fillers that prevented participants from ignoring cue Y in later experienced conditions. During each valence rating, the participant saw screenshots of Cues S1 and S2 in Context A for the ABA condition and Context B for the ABB condition, as they appeared during the condition’s stream. One was displayed on the left of the screen, the other on the right, with cues randomly assigned to a location. Centered above the two cues was the question: “Which stimulus do you prefer?” Below the cue was an 11-point Likert scale, ranging from +5 to 0 to +5, and anchored at +5, 0, and +5. The anchor below 0 always read, “I do not have any preference.” For the series of three valence ratings, the horizontal positions of the two cues were counterbalanced. Under each of the left and right instances of +5, the anchor read, “I much more prefer [W, X, or Y],” where X, Y, or W in the anchors were small images of that cue. The image of the cue in its training context always appeared above the anchor corresponding to the highest preference for that stimulus (so, for instance, the image of X, in either context A or context B depending on the condition, was displayed immediately above the “I much prefer X” anchor). The participants used the mouse to provide their answers. Consistently, the mouse appeared at the center of the screen, equidistant from the leftmost +5 and the rightmost +5 anchors. When the participants clicked on the white circle indicating the various rating levels in the Likert scale, the selected value turned black for 150 ms before the valence rating screen disappeared. Once the prediction and valence ratings were over, another screen appeared, prompting participants to left-click the mouse whenever they were ready to start the next condition.

Warmup.

Before experiencing the experimental conditions, all participants received a series of warmup conditions aimed at familiarizing them with the procedure and providing examples of different relationships between the target cue X and the outcomes (see Table 2). The conditions during the warmup all used the specific context and stimulus sets designated for the warmup, which were distinctly different from those used for the experimental conditions. Notably, the warmup stimulus set contained only two outcomes (O1 and O2), each of them neutral in valence, versus three outcomes (POS, NEG, and NEUT) in the experimental stimulus sets, with only one of them being neutral. Thus, instead of providing four ratings during the prediction test, participants provided three, corresponding to the X-O1, X-O2, and X-NULL contingencies. Otherwise, the structure of the warmup condition was identical to the description above, with O1 appearing at the same location as NEG and O2 appearing at the same location as POS and NEUT. The warmup procedure consisted of two parts. First, participants received warmup training streams that were followed by feedback on the programmed relationships between X and each of the outcomes (O1, O2, and NULL). Second, participants received warmup streams of trials intended to test whether they could reliably report positive, zero, and negative contingencies between X and each of its outcomes. Between warmup streams, participants were tested on both the predictive relationship between X and the various outcomes and on the valence of X. For each valence rating, participants were reminded that there was no right or wrong rating of valence and then prompted to left-click the mouse to advance from one stream to the next.

The first warmup training stream provided an example of a positive contingency between X and O1. The trial composition for the stream presented in this condition (positive warmup) is shown in Table 2. Other than the differences noted in Table 2, all warmup training streams were identical. At the end of each warmup stream, the prediction and valence ratings (in random order) were requested in Context A. Once the participants had provided their prediction and valence ratings, an instruction screen was presented:

In the sequence you have just seen, [X] was always followed by [O1], [X] was never followed by [O2], and a stimulus was always presented following [X]. Hence, you should have judged that it is very likely that [X] would be followed by [O1]. As for which stimulus you prefer, there is no good or bad answer to those questions; just choose according to your preference!

[X], [O1], and [O2] consisted of images of the respective cue and outcomes.

The second warmup training stream provided participants with an example of a negative X-O1 contingency. The trial composition for the negative warmup stream is shown in Table 2. Once participants had provided their responses to the prediction and the valence rating questions, an instruction screen explained to them that their answer should have been 0 for the X-O1 and X-O2 relations but 100 for the X-NULL relation because X had never been paired with an outcome.

The third and fourth warmup training streams provided participants with examples of zero contingencies for X-O1 and X-NULL (third warmup stream) and X-O1 and X-O2 (fourth warmup stream). The trial compositions for these warmup streams are shown in Table 2. Once the participants had provided their responses to the prediction and valence ratings, an instruction screen explained the correct prediction ratings as in the first and second warmup streams. Participants were then prompted to left-click the mouse to start the next condition.

In the final warmup streams, participants were eliminated from the study if they failed to answer all the questions correctly. Before beginning the final warmup streams, participants were told that they would be presented with more examples of each of the four types of warmup conditions, but now there would be no feedback. Upon left-clicking the mouse, participants were exposed to a series of warmup measurement quartets consisting of the positive, negative, and two null warmups in a random order for each quartet. Following each stream, participants provided both prediction ratings and valence ratings. In the positive warmup streams, the participant’s prediction rating was defined as correct if it was between 80 and 100 for the X-O1 contingency and between 0 and 20 for the X-O2 and X-NULL contingency. Following each negative warmup stream, ratings between 0 and 20 for the X-O1 and X-O2 contingencies and between 80 and 100 for the X-NULL contingency were defined as correct. For the first zero contingency warmup stream, a correct response was defined as one between 30 and 70 for the X-O1 and X-NULL contingencies and between 0 and 20 for the X-O2 contingency. For the second zero contingency warmup stream, a correct response was defined as one between 30 and 70 for the X-O1 and X-O2 contingencies and between 0 and 20 for the X-NULL contingency. Participants were repeatedly presented with this quartet of warmup conditions until they were either able to provide a correct response for each condition composing a quartet or until they had been exposed to 10 quartets. In the latter case, they were considered to have failed warmup training and were thanked and then dismissed. As previously mentioned, 80 participants failed the warmup.

Experimental Conditions.

The experimental conditions started immediately after the warmup criterion was reached. Before participants clicked their mouse to advance to the study, an instruction screen informed them that the relation between X and the various outcomes would now be more difficult to identify. They were exposed to three octets of conditions (i.e., each experimental condition was presented three times). The 8 conditions in each octet were composed by factorially crossing the type of interference treatment (Ctr vs. Ext vs. CC vs. NFE) with the context in which the prediction and valence ratings took place (A vs. B). The composition of the streams in the Ctr, Ext, CC, and NFE conditions are shown in Table 3. Phase 1 consisted of trials that established an X-NEG association and a W-NULL association. The last trial of Phase 1 was followed immediately by the first trial of the first cycle of Phase 2 and the last trial of the first cycle of Phase 2 was followed immediately by the first trial of the second cycle of Phase 2. The order of trials within a cycle was randomly determined. Testing occurred at the end of each stream, in either context A or B depending on the condition. Once a participant completed an octet of conditions, the next octet commenced. Condition order was randomized within octets for each participant.

Outcome ratings.

After all the experimental condition streams, the affective valences of the outcomes were assessed as a manipulation check. Participants were asked to evaluate the valence of each outcome on a −5 to +5 scale. To obtain a rating, a stimulus was shown for 400 ms centered on the screen over a grey background. This was followed by the question “How pleasant or unpleasant is this image for you?” along with the presentation, below the question, of an 11-point Likert scale going from −5 to +5 and anchored at −5 (very unpleasant), 0 (neither pleasant nor unpleasant), and +5 (very pleasant). Participants responded by clicking on the Likert scale, which was followed by a 1000-ms ITI screen consisting of a white fixation cross over a grey background that preceded the next question. The ratings of all the outcomes were repeated five times, with the order randomized within each of the five cycles. The goal of these final outcome ratings was to make sure that the valence of the IAPS pictures remained intact despite their smaller scale and their short presentation. Note that, as we were not interested in their arousal properties which, anyway, were roughly equated across outcomes; thus, we did not ask the participants to rate how aroused they were by the pictures. Once these valence ratings were completed, the participants were presented with a debriefing screen informing them of the intent of the experiment.

Data Analysis

Although the Likert scale used to measure participants’ ratings is formally an ordinal scale, parametric statistics were used to analyze ratings as is often done (see Maia et al., 2018, for a critique of this approach). For each participant in each condition (Ctr, Ext, CC, and NFE), a mean prediction rating and a mean valence rating were computed based on the three prediction and valence ratings of that condition. The X vs. W valence ratings were used to compute the valence of X. If the participant indicated that he preferred X over W, the valence of X was, of course, the rating he chose on the Likert scale. If, on the contrary, the participant indicated a preference for W over X, the rating X was assigned was the valence the participant assigned to W with the sign reversed. Hence, every X vs. W valence rating allowed to assign to X a valence ranging from −5 and −1 (when the participant indicated he preferred W over X), equal to 0 (when the participant indicated he was indifferent between the two stimuli), or ranging from 1 to 5 (when the participant indicated he preferred X over W). The other valence ratings were only collected as decoys and were not analyzed. For each stimulus presented during the final outcome rating phase, a mean rating was computed based on the five ratings the participant gave for each outcome.

Inferential analysis of ratings was carried out using repeated-measures analyses of variance (ANOVA). The Huyhn-Feldt correction was applied when the sphericity assumption was not met. Partial eta-squared with a 90% confidence interval (CI) was used as the measure of effect size for each ANOVA, which was computed using resources from Nelson (2016). Based on Steiger’s (2004) analysis, which points out that the one-tailed nature of hypothesis tests based on the F-distribution, supports the use of 90% (two-tailed) CIs, we used in the present study 90% CIs for $η_{p}^{2}$ following ANOVAs with alphas of 0.05. Cohen’s d was used as a measure of effect size whenever two conditions were compared. Following Cummings’ (2012) recommendation, the unbiased estimate of Cohen’s d in the population (sometimes called Hedge’s g) was used as a measure of effect size and computed with the formula:

e s t i m a t e d d = (1 - \frac{3}{4 d f - 1}) \times \frac{M_{d i f f}}{\sqrt{\frac{s_{1}^{2} + s_{2}^{2}}{2}}}

where the right side of the formula divides the mean of the difference scores across participants ( $M_{d i f f}$ ), by an estimate of standard deviation that considers the pooled variances of the two conditions, $\sqrt{\frac{s_{1}^{2} + s_{2}^{2}}{2}}$ , and df is the degrees of freedom on which a paired samples t-test is based (n - 1). The left side of the formula applies a correction for estimating d based on a sample. 95% CIs for Cohen’s d for paired designs can only be approximated. Cousineau and Goulet-Pelletier (2021) reviewed eight different methods to achieve this goal with the conclusion that the best one is the adjusted lambda-prime method, which we implemented using the ESCI module in JAMOVI (www.jamovi.com). Error bars in graphs are always 95% CIs computed using Student’s t distribution.

Transparency and openness statement

All raw data from these experiments and the Python computer code for Experiment 1 are available at https://orb.binghamton.edu/psych_fac/19/ and upon request from Jeremie Jozefowiez or Ralph Miller. All the stimuli are available with the PYTHON program at https://orb.binghamton.edu/psych_fac/19/. We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study.

Results

Outcome Valences

Figure 2 depicts the mean ratings for the valence of the outcomes for each condition. Quite clearly, the IAPS pictures retained their emotional properties despite their reduced size and short presentation time. The POS outcomes were rated as highly positive in all conditions, whereas the NEG outcomes were rated as highly negative. The NEUT outcomes were rated slightly positive.

Prediction ratings

Efficacy and Context Sensitivity of Ext, CC, and NFE.

The top panel of Figure 3 shows the mean prediction ratings as a function of the conditions. Let us first focus on the X-NEG ratings. A repeated-measures ANOVA using the testing context (ABA vs. ABB) and the type of interference treatment (Ctr. vs. Ext vs. CC. vs. NFE) as factors found a main effect of the context manipulation, F(1, 98) = 69. 98, p < 0.001, $η_{p}^{2} = 0.42$ , 90% CI [0.29, 0.51], and of the type of interference, F(1.81, 177.38) =118.05, p < 0.001, $η_{p}^{2} = 0.55$ , 90% CI [0.46, 0.61], but failed to detect an interaction between these two factors, F(2.90, 284.20) = 2.34, p = 0.08, $η_{p}^{2} = 0.02$ , 90% CI [0, 0.05].

Table 4 lists the results of comparisons between conditions. Across different response-reducing treatments (Ext, CC, and NFE), ratings were reliably higher in Context A than in Context B. All three response-reducing treatments (Ext, NFE, and CC) resulted in significantly lower ratings than the control (Ctr) treatment.

Table 4:

Results of Experiment 1.

Measure	Comparison	Mean difference	Cohen’s d	Between-measurement r

Prediction	A vs. B	33.70 [25.70, 41.70]	1.45 [1.09, 1.84]	−0.50
Prediction	Ctr vs. Ext	28.60 [23.50, 33.70]	1.56 [1.26, 1.89]	0.02
Prediction	Ctr vs. CC	32.40 [27.20, 37.50]	1.81 [1.49, 2.17]	−0.06
Prediction	Ctr vs. NFE	33.20 [28.00, 38.30]	1.84 [1.52, 2.20]	−0.03
Prediction	Ext vs. CC	3.77 [1.11, 6.44]	0.23 [0.07, 0.40]	0.66
Prediction	Ext vs. NFE	4.58 [1.66, 7.50]	0.28, [0.10, 0.46]	0.60
Prediction	CC vs. NFE	0.81 [−1.51, 3.12]	0.05 [−0.09, 0.20]	0.73
Valence	Ctr vs. Ext	−0.34 [−0.63, −0.06]	−0.13 [−0.24, −0.02]	0.85
Valence	Ctr vs. CC	−0.78 [−1.17, −0.39]	−0.30 [−0.46, −0.15]	0.71
Valence	Ctr vs. NFE	−0.27 [−0.67, 0.13]	−0.10 [−0.26, 0.05]	0.70
Valence	Ext vs. CC	−0.43 [−0.77, −0.10]	−0.17 [−0.30, −0.04]	0.78
Valence	Ext vs. NFE	0.07 [−0.30, 0.45]	0.03 [−0.12, 0.17]	0.74
Valence	CC vs. NFE	0.51 [0.14, 0.87]	0.20 [0.06, 0.35]	0.73

Open in a new tab

Note. Each mean difference reports the mean of the differences in ratings between the specified conditions. Brackets contain the upper and lower limits of 95% confidence intervals for either the mean difference or estimate of effect size (d, see the Methods section). Between-measurement r reports the correlation between participants’ scores in the comparisons. For valence comparisons, negative values indicate that the first listed condition (e.g., “Ext” in “Ext vs. CC”) produced ratings corresponding to greater unpleasantness. Except for the A vs. B comparison, all comparisons pooled ratings at test in contexts A and B.

Prediction Ratings for Other Outcomes.

Figure 3 also shows the mean prediction ratings for the POS, NEUT, and NULL outcomes as well as the NEG outcome. Besides the X-NEG ratings, only the X-NULL ratings in Ext, the X-POS ratings in CC, and the X-NEUT ratings in NFE differed meaningfully from 0. The NULL outcome in the Ext condition, the POS outcome in the CC condition, and the NEUT outcome in the NFE condition all have a special status in that, just like the NEG outcome, they had been paired with X, albeit in Phase 2. In the Ctr condition, in which no outcome other than NEG was paired with X, no similar pattern could be observed. The mean X-NULL rating was a bit higher than the other ones and was slightly affected by the test context, but this difference was small compared to what happened to the X-NULL ratings in Ext.

The test context affected the X-NULL rating in Ext, the X-POS rating in CC, and the X-NEUT rating in NFE, but the effect was opposite the one observed for the X-NEG rating. In summary, when the X-NEG rating in Ext (or CC or NFE) increased, the X-NULL rating (or X-POS rating or X-NEUT rating, respectively) decreased. Figure 4, which depicts the relationship between the X-NEG ratings and the X-NULL (and X-POS and X-NEUT) ratings in Ext (and CC and NFE, respectively) shows that this relation is not an artifact of averaging across participants.

A striking feature of Figure 4 is that some of the points are aligned linearly along the equation $y = 100 - x,$ with $x$ being the value of the X-NEG prediction rating and $y$ being the value of either the X-NULL, X-POS, or X-NEUT rating depending on whether the condition is Ext, CC, or NFE. This is due to some participants seemingly allocating their prediction ratings across the various outcomes so that the sum was equal to 100%.

Valence ratings

The bottom panel of Figure 3 shows the mean valence ratings of X (relative to W) across conditions as a function of the test context and the type of interference treatment. A repeated-measures ANOVA using the test context (Context A vs. Context B) and the type of interference (Ctr vs. Ext vs. CC vs. NFE) as factors found a main effect of the type of interference, F(2.79, 272.97) = 6.31, p < 0.001, $η_{p}^{2} = 0.06$ , 90% CI [0.02, 0.10], but did not detect an effect of the context, F(1, 98) = 3.40, p = 0.07, $η_{p}^{2} = 0.03$ , 90% CI [0.00, 0.11], nor an interaction between the two factors, F(3, 294) = 0.79, p = 0.50, $η_{p}^{2} = 0.008$ , 95% CI [0.00, 0.02]. The results of comparisons between conditions are summarized in Table 4. Importantly, lower valence ratings (less appetitive) pooled from Contexts A and B were observed after Ext and CC, relative to the control treatment. Moreover, the analysis failed to detect a reliable change in valence resulting from NFE treatment. The valence of X was higher after CC than after either Ext or NFE, and there was no detectable difference between Ext and NFE. Thus, CC, but not NFE, reliably increased valence ratings at test relative to Ext.

Discussion

The goal of this experiment was to directly compare the efficacies of Ext, CC, and NFE in reducing both the valence of a CS previously paired with an aversive outcome (evaluative conditioning) and the predicted likelihood of that aversive outcome given the CS (expectancy learning) at tests in both the context of target association degradation (ABB) and the context of acquisition (ABA).

Examining the prediction ratings, the expression of the target X-NEG association was robustly altered by Ext, CC, and NFE, with both CC and NFE being a little more effective. That is, CC and NFE produced lower prediction ratings than Ext, and no appreciable difference was observed between CC and NFE. All conditions (Ctr, Ext, CC, and NFE) were susceptible to large context effects (estimated d between 1.0 and 1.2).

Examining the valence ratings, the data suggest that CC and Ext increased the valence of X relative to Ctr and that the effect of CC was stronger than either Ext or NFE. Otherwise, there is no detectable context effect although, at the descriptive level, the valence of the cue was slightly higher when the testing occurred in Context B than in Context A, as one would expect. We will postpone further discussion of these conclusions until the data from Experiment 2 have been described. Discussing the results of the two experiments together will remove some of the ambiguities that arise when each experiment is considered by itself.

Experiment 2

The greater efficacy of CC relative to Ext in lowering the prediction ratings observed in Experiment 1 is inconsistent with a result we repeatedly observed in prior experiments that contrasted Ext and CC using the streaming procedure (Jozefowiez et al., 2020). In that series, we found that Ext was more effective than CC, except in one experiment in which CC proved as effective as Ext. We hypothesized that this was because, in that experiment, conditions allowed participants to better discriminate between the initial US paired with the target CS and the interfering US used in CC. Likewise, in Experiment 1, asking the participants to rate the X-POS, X-NEUT, and X-NULL relations in addition to the X-NEG relation should have highlighted the difference in outcomes between Phases 1 and 2, creating exactly the kind of situation hypothesized by Jozefowiez et al. to boost the efficacy of CC relative to Ext, especially as all the outcomes were simultaneously present on screen during the prediction ratings. The same reasoning would also apply to NFE, which could explain why both CC and NFE were more effective at altering the prediction ratings than Ext.

If this hypothesis were correct, the greater efficacies of CC and NFE over Ext observed for the prediction ratings in Experiment 1 should not be observed if participants were probed only about the X-NEG association. Experiment 2 was designed to test this prediction. It replicated Experiment 1, except that during the prediction ratings, the participants only rated how likely it was for X to be followed by NEG. We expected this manipulation to bridge the gap in efficacy between Ext and CC/NFE, perhaps making Ext even more effective than CC and NFE.