Two roads leading to the same evaluative conditioning effect? Stimulus-response binding versus operant conditioning

ABSTRACT

Evaluative Conditioning (EC) refers to changes in our liking or disliking of a stimulus due to its pairing with other positive or negative stimuli. In addition to stimulus-based mechanisms, recent research has shown that action-based mechanisms can also lead to EC effects. Research, based on action control theories, has shown that pairing a positive or negative action with a neutral stimulus results in EC effects (Stimulus-Response binding). Similarly, research studies using Operant Conditioning (OC) approaches have also observed EC effects. The aim of the present study is to directly compare EC effects elicited by two different response-based approaches – S-R bindings and OC. To this end, participants were randomly assigned to an S-R binding procedure and an OC procedure. EC effects were measured in conditions and compared. Implications for EC theory are discussed.

KEYWORDS:

• Evaluative conditioning
• action control
• stimulus-response binding
• operant conditioning

Our attitudes towards objects, people or concepts are relevant to our actions. However, we are generally not born with those attitudes, rather we learn them over time (Rozin & Millman, 1987). Attitudes can be learned via associations with other objects towards which we might already have formed a liking or disliking. Attitude formation, i.e. the like or dislike of stimuli due to their pairing with other positive or negative stimuli, is called Evaluative Conditioning (EC, De Houwer, 2007; Walther et al., 2019). In a typical EC study, a neutral stimulus (or conditioned stimulus, CS) is paired with a stimulus which already has strong positive or negative meaning (unconditioned stimulus, US). After many pairings, CS paired with positive US is more positively evaluated than CS paired with negative US (e.g. Baeyens et al., 1992; Blask et al., 2016, 2017; Hughes et al., 2016; Levey & Martin, 1975; Pleyers et al., 2007).

Most theories explaining EC are based on stimulus–stimulus associations, (e.g. Baeyens et al., 1992; Jones et al., 2009; Levey & Martin, 1975; Mitchell et al., 2009; however, see Hughes et al., 2016). Human beings, however, are not just passive receivers of information, rather they interact with their environment. Thus, positive or negative actions may transfer valence to stimuli too. Thus, it is not only the case that we conclude the stimulus from the observation of a positive response but also that the stimulus itself is associated with a valent response. This hypothesis is based on action control theories postulating that perceptual and action representations rely on the same pool of feature codes (Prinz, 1997), and assuming that features of stimuli and responses are integrated into short-term episodic traces that can influence actions (e.g. Frings et al., 2020; Hommel et al., 2001). Based on this idea, valence, as a feature of a stimulus, would be integrated with the response and could lead to a valence transfer between stimulus and response. Conversely, if an action or response (R) is positive or negative, this valence might transfer to a stimulus (S) to which the R is executed. This is exactly what Blask et al. (2016) observed (see Walther et al., 2019 for a detailed theoretical discussion). They had participants respond to positive and negative US with a left or a right button press. Later, these responses were used to respond to CS. They observed that CS paired with the R previously paired with positive US was more positively evaluated than CS paired with paired with the R previously paired with negative US. Using an impression formation paradigm, Gast and Rothermund (2011) observed EC effects via S-R associations, lending further support to the idea that valence transfer can occur not only between stimuli but also between stimuli and responses. This idea, adopted from action theory, considerably enlarges the scope of EC and is, therefore, of great theoretical interest.

Nevertheless, S-R pairings may be seen as atypical since they include a response as the US (or indeed a response outcome in the operant conditioning condition below). However, according to action control frameworks (e.g. Frings et al., 2020) stimuli and responses share the same representation codes; therefore, it does not matter whether two stimuli are paired, a stimulus and a response are paired or indeed two responses. Additionally, the so-called functional accounts (Walther et al., 2019), for example, the intersecting regularities accounts (e.g. Hughes et al., 2016), postulate that EC occurs due to the overlap of contingencies in the environment. For instance, if the same response elicits the same outcome when executed with two different stimuli, one positive and one neutral, the neural stimulus might be evaluated more positively due to the overlap of responses and outcomes (Hughes et al., 2016). Including responses in an EC procedure not only extends it to a very relevant component of human actions, namely responses, but also extends it to more real-life like situations, where responding is often an inevitable part of human interactions with the environment. In the present study, we refer to conditioning effects via the pairing of some US (stimulus, response or outcome) with a CS as an EC effect.

The action control idea of EC shares some similarities with other types of conditioning processes, e.g. Operant Conditioning (OC). Since OC is based on learning consequences, the response is crucial. A response leading to a positive outcome will be performed more often than one leading to a negative outcome. In the case of EC, outcome (O) valence may be transferred to a CS (Eder et al., 2019). Eder et al. (2019) first used OC to condition two responses. Participants could freely execute one of two responses in a free-choice task. One response was always followed by a positive O and the other with a negative O. In the following phase, participants classified fictitious groups of people using the same responses from the previous phase. The group classified using the response, which was previously associated with a positive O, was more positively evaluated than the group that was classified with the response which was previously associated with a negative O. They explained their findings with the intersecting regularities account (Hughes et al., 2016). The intersecting regularities account postulates that relations between objects in the environment may sometimes intersect or overlap. This overlap or intersection can lead to a transfer of valence if the intersecting regularities consist of a valence or evaluative component (Hughes et al., 2016). Thus, in Eder et al. (2019), a valence transfer occurred in the first phase via operant conditioning, and later this valence component was transferred to neutral stimuli due to the pairing of the conditioned response and neutral stimulus – they termed this operant evaluative conditioning. Thus, while OC has been used to condition a response, a transfer of O valence to a stimulus has not been tested. Within operant conditioning literature itself, associations between a discriminative stimulus and an outcome have been observed (e.g. Colwill & Rescorla, 1988), additionally, there is also evidence suggesting the existence of hierarchical associations between S, R and O (Rescorla, 1990). On the basis of outcome devaluation studies, it was suggested that the S is associated with a specific R-O contingency (Rescorla, 1990). That is, the organism didn’t just learn a new specific behaviour, rather, they showed a specific behaviour in the presence of a specific CS, indicating more complex associative structures. Within the context of EC effects, this would mean that either the imperative stimulus (CS) is associated with the outcome valence directly or the CS could be associated with a compound response-outcome association, thereby resulting in valence transfer. That is, valence may be transferred via a simple S-O association or a more complex structure of S-(R-O) associations.

Present study

The aim of the present study was to replicate previous findings of S-R binding EC effects and to directly compare them with EC effects elicited via OC. While in both cases, response valence plays a central role – either as a feature of the response or as its outcome– there are significant differences between both processes. Firstly, the structure of the associations is different. Action control theories generally assume binary bindings between S and R features, whereas, in OC, the relation might be hierarchical, with the S being associated with a particular R-O contingency or the S may be directly associated with the O. Even though the latter case is a binary association it still bears examination since evaluative judgements of the stimulus are not the typical dependent variable in OC studies. Furthermore, such S-O associations have not been observed in studies examining short-term S-R associations (e.g. Moeller & Frings, 2019). Secondly, in OC the causal relationship between the R and O is of central importance, the R must be seen to cause the O (Colwill & Rescorla, 1990; Staddon & Cerutti, 2003), whereas in SR binding, a causal relationship is not necessary, the temporal coexistence or contiguity is enough to form bindings (Frings et al., 2020; Hommel et al., 2001). Similarly, unlike OC, SR binding does not assume any kind of reinforcement (Frings et al., 2020). Additionally, since response-outcome contingencies are central to OC, the temporal order of the R, S and O is very clearly specified. In SR binding, however, the temporal order should not be of central importance as SR bindings as well as response-effect bindings have been observed (Dutzi & Hommel, 2009) and bindings do not store temporal order information (Moeller & Frings, 2019). Thus, while responses are relevant to SR binding and OC, there are some differences between the two processes.

To directly test EC effects via SR binding and OC, the present study included both an SR binding procedure based on Blask et al. (2016) and an OC procedure based on classical OC learning procedures (e.g. Eder et al., 2019). The S-R condition is very similar to the procedure used in Blask et al. (2016). For the OC condition, the procedure deviated somewhat from the procedure used in Eder et al. (2019). Instead of first conditioning a response and then executing the conditioned response to an imperative stimulus (Eder et al., 2019), in the present study correct responses were followed either by a positive outcome (positive images) or negative outcome (negative images). This procedure more closely mirrors classical OC procedures wherein an organism is directly presented with an outcome upon executing a response to an imperative stimulus. In the S-R and OC conditions, EC effects can be measured; however, their emergence is due to different kinds of associations. While in both tasks, the CS is always a neutral image, the US differs. In the S-R task, the US is a (conditioned) response to the CS, whereas in the OC condition, the US is a response outcome. Thus, in the S-R condition, the EC effect occurs due to automatic S-R association whereas in the OC condition, it occurs due to automatic S-(R-O) or S-O associations. A comparison of the EC effects in both conditions would shed more light on the role of the response in EC effects, either as the US (S-R condition) or as a being instrumental in producing the US (OC condition), the kind of associations leading to EC effects S-R associations (S-R condition) or S-(R-O)/ S-O (OC condition) and whether they are equally effective in producing EC effects. The present study might also help understand the relation between S-R bindings and longer-term learning processes like OC. S-R bindings have already been discussed as a possible first step in long-term learning (Frings et al., 2023 for a discussion) and have been discussed together with Pavlovian Conditioning (Giesen & Rothermund, 2014; Singh et al., 2016) as well as OC (Singh et al., 2018).

Methods

Participants

In considering the sample size calculation, two theoretical considerations were taken into account; firstly, in the OC condition, the dependent variable being measured was not the typical dependent variable usually tested in OC studies, i.e. behaviour or responses, rather evaluative judgements of the CS were tested. Secondly, action control studies that have been discussed together with Pavlovian/Operant Conditioning, do not find any evidence for S-O associations (Moeller & Frings, 2019). Based on these considerations, we expected smaller EC effects in the OC condition. The sample size was pre-registered [https://aspredicted.org/1QY_T46] and calculated based on an effect size of f = .2, α = .05 and power of 1-β = .90, leading to a sample size of 68 participants. To have an equal number of participants in all conditions a total sample of 72 participants was recruited. A total of seventy-two students from the University of Trier (53 female and 19 male) participated in the study. Psychology students were granted partial course credit. The median age of the participants was 24 years (range 18–36 years). All participants gave informed consent before starting the experiment. As per the pre-registration, participants with accuracy lower than 80% in the response formation phase or the evaluative conditioning phase were excluded, this led to the exclusion of two participants (Appendix A). Results of the analysis including the entire sample of 72 participants are reported in Appendix B.

Design

The design consisted of two factors with two levels each 2 (Condition: Stimulus-Response binding vs. Operant Conditioning) x 2 (US valence: positive vs. negative). Additionally, the mapping of valence onto the response side (left vs. right) and the combination of US and CS (first half of the CS mapped to positive US and second half to negative CS vs. first half of the CS mapped to negative US and second half to positive CS) were counterbalanced across participants. For all the analyses, the data were collapsed across the last two factors.

Apparatus and materials

The experiment was run with E-Prime 3 Software (Psychology Software Tools) on tower PCs with an Intel Core i5-8400 processor and 16 GB RAM attached to a 23.8-inch monitor with a 1920 × 1080 pixel resolution and a white background and a standard German language QWERTZ keyboard. Stimuli were taken from Blask et al. (2016). The US consisted of 20 images (10 positive and 10 negative) depicting persons. The US subtended a visual angle of 16° by 11.99° from a distance of 60 cm. The CS consisted of 16 images of fictitious water brands. The CS subtended a visual angle of 12.08° by 9.48° from a distance of 60 cm. The fixation marker subtended a visual angle of 1.5° by 1.3° and the arrow cue subtended a visual angle of 4.1° by 2.1° from a distance of 60 cm.

Procedure

Participants were tested individually in sound-proof chambers. Participants gave their informed consent before starting the experiment. Experimental instructions were presented on the screen. The experiment consisted of three phases in the SR binding condition and two phases in the OC condition. The procedure for each phase for both conditions is depicted in Figure 1.

Figure 1. Trial procedure in SR (upper panel) and OC (lower panel) conditions.

SR binding condition

Response formation phase: In the SR binding condition, the first phase consisted of a response formation phase. Participants were presented with the US and required to categorise the stimuli as positive or negative using the “w” and “p” keys. One-half of the participants responded to the positive US with the “w” key and negative US with the “p” key, while the other half of the participants received the opposite mapping. Each trial began with a fixation marker for 500 ms, which also served as the inter-trial interval. Following the fixation marker, the US was presented for 500 ms, after which the participants were presented with a blank display for a further 1000 ms during which time they could still execute the response, resulting in a total response window of 1500 ms. If participants failed to respond within the time, they were presented with feedback encouraging them to respond faster. Participants first worked through a practice block in which each US was presented once, thus resulting in 20 trials during which they were presented with feedback after every response. Following the practice phase, participants worked through the test phase consisting of 10 repetitions of each US resulting in 200 trials. In the test phase, participants were only provided feedback after erroneous responses.

Evaluative Conditioning Phase: The second phase consisted of a conditioning phase. 8 of the 16 CS were paired with a positive response side, i.e. the response side with which the participant responded to positive images in the first phase, and the other 8 CS were paired with the negative response side, i.e. the response side with which the participant responded to negative images in the first phase. Participants were not informed of this pairing. CS-valence pairings were counterbalanced across participants such that for one-half of the participants, the first 8 CS were paired with positive US and the second with negative US, and the other half of the participants received the opposite mapping. An arrow cue indicated the response to be executed. Participants were instructed to prepare the response indicated by the arrow but to carry it out only when the CS was presented. A single trial consisted of a fixation marker for 500 ms which also served as the inter-trial interval, followed by an arrow pointing either to the left or right for 500 ms. After the arrow the CS was presented for 500 ms, followed by an additional 1000 ms blank screen during time participants could still respond. This phase consisted of 10 repetitions of each of the 16 CS resulting in 160 trials in total. Participants were provided feedback after erroneous responses as well as if they responded to them early.

Test phase: The third phase of the experiment was the test phase, in which participants rated each of the CS and US. Participants were presented first with the CS and then with the US and instructed to rate how positive or negative they found the stimuli. Participants were instructed to respond as spontaneously as possible. A fixation marker was presented for 500 ms, after which the stimuli were each presented for 1000 ms; however the rating scale remained on the screen until participants made a rating. The rating scale had a minimum of −10 and a maximum of 10 and was only labelled, the two ends to indicate which end was positive and which was negative. The indicator was only visible once participants clicked somewhere on the scale. Once participants made a rating they clicked on the “next” button (weiter in German) to proceed to the next stimulus.

Operant conditioning condition

The OC condition consisted of two phases

Evaluative Conditioning Phase: The first phase in the OC phase consisted of a conditioning phase similar to the conditioning phase in the SR binding condition. Participants were presented with a response cue (arrow) pointing to the left or the right for 500 ms, followed by the CS for 500 ms with a 1000 ms blank during which time participants could still respond. Following the response, either a positive or negative image was presented when participants responded according to the cue. CS-outcome valence pairings were counterbalanced across participants such that for one-half of the participants, the first 8 CS were paired with positive outcomes, the second with negative outcomes and the other half of the participants received the opposite mapping. If they did not respond according to the cue, they were presented with an error feedback. One-half of the CS was paired with positive images and the other half with negative images. Each CS was presented a total of 10 times.

Test phase: The test phase was the same as in the SR binding condition.

Results

US ratings

All Data were analysed using R (R Core Team, 2023). US ratings were analysed in a 2 (condition: SR binding vs. OC) x 2 (US Valence: positive vs. negative) using the ez Package (Lawrence, 2016). The ANOVA revealed a non-significant main effect of condition, F(1,68) = 1.11, p = .296, η_p² = .02, indicating that the mean ratings did not differ between both conditions (M_SR = −0.3, SD_SR = 8.14; M_OC = −0.14, SD_OC = 8.05). The main effect of valence was significant F(1,68) = 1712.63, p < .001, η_p² = .96, indicating that the mean ratings for negative US were significantly different from the mean ratings of positive US, (M_neg = −8.07, SD_neg = 1.50; M_pos = 7.64, SD_pos = 1.88). The interaction of condition by US valence was not significant, F(1,68) = 0.04, p = .833, η_p² = .00, indicating that there was no difference in the mean ratings for positive and negative US between the two conditions.

CS ratings

CS ratings were analysed in a 2 (condition: SR binding vs. OC) x 2 (US Valence: positive vs. negative) ANOVA using the ez Package (Lawrence, 2016) . The ANOVA revealed a non-significant main effect of condition, F(1,68) = 0.20, p = .654, η_p² = .00, indicating that the mean ratings did not differ between both conditions (M_SR = 0.00, SD_SR = 1.90; M_OC = −0.16, SD_OC = 1.89). The main effect of valence was significant F(1,68) = 4.74, p = .033, η_p² = .07, indicating that the mean ratings for CS paired with negative US were significantly different from the mean ratings of CS paired with positive US, i.e. an overall EC effect, (M_neg = −0.36, SD_neg = 1.76; M_pos = 0.19, SD_pos = 1.98). The interaction of condition by valence was not significant, F(1,68) = 0.08, p = .784, η_p² = .00. The mean EC effects for each condition, computed as the difference between the mean positive and negative ratings, are depicted in Figure 2, the mean EC effect for the SR binding condition was 0.62 (SD = 2.56) and the mean EC effect for the OC condition was 0.48 (SD = 1.59). EC effects in each condition were tested against null, the EC effect in the SR binding condition was not significant from zero, t(33) = 1.45, p = .166, Cohen’s d = .24, the EC effect in the OC condition was not significant from zero, t(33) = 1.82, p = .078, Cohen’s d = .30.

Figure 2. Mean Evaluative Conditioning (EC) effects, computed as the difference between mean positive and negative ratings, in OC and SR binding conditions.

Discussion

The aim of the present study was to compare EC effects elicited via SR associations and OC. To this end, participants were assigned either to an SR binding or OC procedure. Evaluative ratings for CS paired with positive US were compared to those for CS paired with negative US. CS paired with a positive US was rated more positively than CS paired with a negative US. A non-significant interaction indicated the absence of a difference of the magnitude of d = 0.4 or larger in the present sample. The present results indicate that EC can occur by SR binding and OC; however, the data do not allow any conclusions on differences between the EC effects of the magnitude of d > =  0.4.

Nevertheless, the present results shed further light on the mechanisms behind EC effects by, firstly, providing evidence that other kinds of associations can elicit EC effects – namely S-O/S-(R-O) associations. In traditional accounts of EC (Baeyens et al., 1992; Levey & Martin, 1975; Pleyers et al., 2007) there is a direct association of the US and CS on a stimulus-stimulus level. Similarly, in the action control account, direct S-R associations are postulated (Blask et al., 2016, 2017). In OC, however, the association of an S to a specific R-O is not necessarily direct. While direct S-O associations have been observed (e.g. Colwill & Rescorla, 1988), indirect associations wherein the S is associated with a specific R-O contingency, but not directly with the O have also been observed (Colwill & Rescorla, 1990; Rescorla, 1990). Both of these kinds of associations are interesting for different reasons. While direct S-O associations have been observed in the OC studies (e.g. Colwill & Rescorla, 1988), they have to date not been observed in action control literature (e.g. Moeller & Frings, 2019). This indicates that while there are some parallels between conditioning processes and short-term S-R associations (for detailed discussions, see e.g. Giesen & Rothermund, 2014; Singh et al., 2016; Singh et al., 2018) there would appear to be some significant differences between them. The case of hierarchical associations is also interesting since to date, the proposed EC mechanisms and action control literature have only relied on direct associations, with no evidence of higher-order associations being observed. Given that S-R binding has been discussed as potentially being early stages of longer-term learning (Frings et al., 2023), it might be an indication that hierarchical associations might exist under some circumstances in S-R bindings as well. Secondly, the present study provided no evidence that EC effects elicited via different kinds of associations S-R and S-O/S-(R-O) differ in magnitude, indicating that they are equally efficient in producing condition effects. However, the present data only allow the conclusion that no effect larger than/equal to Cohen’s d 0.4 was found in the present sample. Further testing for an absence of an effect was non-conclusive. A Bayesian ANOVA provided anecdotal evidence for the null hypothesis, while a two-one-sided test was non-significant (see Appendix C). Taken together, while the present study does not provide any evidence for an interaction of Cohens’d > =  0.4 with the present sample, future research might test for the presence of a smaller effect. Finally, the present results also provide further evidence for a response-based mechanism of EC. Even though the exact process differs, in both the SR binding and the OC conditions, executing a response was central to the elicited EC effect. One caveat with the present study, however, is that there is no direct test of the effect of the response outcomes on the responses themselves. Therefore, we cannot unequivocally claim that the response itself was conditioned in the OC condition since we did not test this. However, this does not preclude the kinds of associations mentioned above (i.e. S-O and S-(R-O)).

The present results are also interesting as they indicate that a causal relation between S-R is irrelevant for EC effects. In the SR condition, there is no causal relationship between the R and CS. The R is simply executed in the presence of the CS. In the OC condition, however, the R is instrumental in causing the O, and the S is associated with this R-O contingency. This is in line with S-S accounts like the referential account (Baeyens et al., 1992), in which an association between the US and CS is formed without assuming contingent relations between them (Walther et al., 2019). The present results also indicate that the temporal order of events is not relevant to EC effects. That is, it is not relevant whether the US was presented together with the CS (SR condition), or whether it was presented after the CS (OC condition). While temporal order is highly relevant to OC since the response must be seen to cause the outcome (Staddon & Cerutti, 2003), it is not so in action control theories (e.g. Moeller & Frings, 2019).

While the present results do not provide evidence for a difference in EC effects between SR Binding and Operant Conditioning larger than Cohen’s d> = 0.4 in the present sample, there are significant implications of the type of learning mechanism for the resulting EC effects which must be further examined. For instance, it is well known that reinforcement schedules play an important role in OC (Staddon & Cerutti, 2003), whereas action control theories might predict that since temporal contiguity is enough (e.g. Frings et al., 2020), EC effects might not be modulated by reinforcement schedules. Furthermore, the question of the durability of EC effects is also relevant. While SR binding effects are usually tested at relatively short durations, EC effects via OC mechanisms are possibly more durable than SR binding EC effects.

In conclusion, the present results replicate previous findings of EC effects elicited via S-R binding and OC. Furthermore, no evidence was observed that these effects differed in magnitude, at least not for differences larger than or equal to Cohen’s d 0.4 in the present sample. Thus, EC effects can be elicited via response-related processes, even if these are indirect (S-O) or hierarchical (S-(R-O)) rather than direct associations (S-R). Future research is required to further elucidate these effects.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.

Appendices

Appendix A

Analysis of accuracy data.

Accuracies in the response formation phase (SR binding condition) and the evaluative conditioning phase (SR binding condition and OC condition) were analysed to exclude participants with an accuracy below 80%. The mean accuracy in the response formation phase was 90.77%. Two participants were identified with accuracies below 80% (76.86% and 77.32%) and thus excluded from the analysis of CS ratings. No participants were excluded based on accuracy in the evaluative conditioning phase (mean accuracy in the SR binding condition = 99.19%, SD = 1.07%, mean accuracy in the OC condition = 98.74%, SD = 2%).

Appendix B

Analysis of CS ratings for the entire sample.

A 2 (condition: SR binding vs. OC) x 2 (US Valence: positive vs. negative). The ANOVA revealed a non-significant main effect of condition, F(1,70) = 0.13, p = .715, η_p² = .00, indicating that the mean ratings did not differ between both conditions (M_SR = −0.03, SD_SR = 1.89; M_OC = −0.16, SD_OC = 1.89). The main effect of valence was significant F(1,70) = 4.79, p = .032, η_p² = .06, indicating that the mean ratings for CS paired with negative US were significantly different from the mean ratings of CS paired with positive US, i.e. an overall EC effect, (M_neg = −0.37, SD_neg = 1.75; M_pos = 0.17, SD_pos = 1.98). The interaction of condition by valence was not significant, F(1,70) = 0.05, p = .818, η_p² = .00, indicating that there was no difference in the EC effects for SR binding and OC. The mean EC effect for the SR binding condition was 0.60 (SD = 2.49) and the mean EC effect for the OC condition was 0.48 (SD = 1.59).

Appendix C

To test for evidence of the absence of the valence by condition interaction, two separate tests were run. Firstly, a Bayesian analysis was run, and secondly, the Two One-Sided Test (TOST) was run.

Bayesian ANOVA: A Bayesian analysis with the factors Valence (positive vs. negative) and Condition (SR bindings vs. Operant Conditioning) with the Subject as the random factor was run with the BayesFactor Package (Morey & Rouder, 2024). The Bayes factor for the model with the interaction and the main effects and the random effect; Condition + Valence + Condition: Valence + Subject was 0.11. The Bayes factor for the model with only the two main effects and the random effect was 0.42. Thus, the Bayes factor for the interaction alone (Condition + Valence + Condition: Valence + Subject / Condition + Valence + Subject) is 0.26. Since the Bayes factor for the null is the inverse of the Bayes factor for the alternative hypothesis, the Bayes factor for the null hypothesis for the interaction is 3.85, indicating evidence for the null hypothesis (Kass & Raftery, 1995).

Two-one-sided-test: To run the TOST, the first EC effects were computed for each participant by subtracting the mean negative rating from the mean positive rating. The TOST was run with the TOSTER package (Caldwell, 2022; Lakens, 2017). A TOST with summary statistics was run with a lower bound of Cohen’s d = −0.4 and an upper bound of Cohen’s d = 0.4, the correlation between the groups of 0.5 and alpha level of .05. The TOST was non-significant, TOST lower, t(59.47) = 1.47, p = .074 and TOST upper, t(59.47) = −1.93, p = .029.