The Role of Trait Inferences in Evaluative Conditioning

A well-studied pathway for establishing and modifying evaluations is known as Evaluative Conditioning (EC). In a typical EC task, a (conditioned) stimulus (CS) is repeatedly paired with a positive or negative (unconditioned) stimulus (US), and as a result, the former typically acquires a similar valence as the latter (i.e., an EC effect; De Houwer, 2006; Hofmann et al., 2010). People encounter stimulus pairings throughout the course of their everyday lives, from exposure to advertisements and media, to social interactions. As such, EC may play a role in many behaviors such as consumer behavior (e.g., Gibson, 2008) and phobias (e.g., Merckelbach et al., 1993).

In the attitude literature, EC is often contrasted with a second evaluative learning pathway: persuasion (e.g., Petty & Cacioppo, 1984; Skowronski & Carlston, 1989). EC is typically viewed as fundamentally different from persuasion and these two pathways have long been studied in isolation from one another (for a discussion, see De Houwer & Hughes, 2016). Whereas EC research examines changes in evaluative responding that are due to stimulus pairings, persuasion research mainly studies changes in evaluative responding that are due to verbal arguments.

EC and persuasion are also assumed to strongly differ in the mental processes that underlie their effects. In the past, EC effects were often said to be mediated by the automatic formation and activation of associations in memory, a process that is assumed to take little cognitive effort (e.g., Baeyens et al., 1992, 1995; Levey & Martin, 1975; Martin & Levey, 1994; Olson & Fazio, 2001). In contrast, many persuasion effects were thought to be mediated by inferences that require effortful processing of words and sentences (e.g., Briñol et al., 2009). One well-studied type of inferences in persuasion research is known as trait inferences. Trait inferences involve the assignment of traits or attributes (psychological or non-psychological qualities or features that regarded as a characteristic or inherent part of someone or something) to a target person or object based on some (source) information (e.g., Kelley, 1973; Skowronski & Carlston, 1989). Persuasion research on trait inference traditionally focused on a specific type of trait inference: inferring the traits of a person or an object based on information other than merely telling participants about the person possessing this trait. For example, this research finds that people often draw inferences about another person’s personality traits or other characteristics based on their physical appearance (e.g., Olivola & Todorov, 2010), their behavior (e.g., Uleman et al., 2008), or their facial expressions (e.g., Knutson, 1996).

Research on trait inferences and their role in persuasion has flourished over the years (e.g., Nisbett & Ross, 1980; Uleman et al., 2008). It is not our goal here to provide a comprehensive literature review of trait inferences research, however, to give a few examples, we now know that people can spontaneously derive trait knowledge (e.g., cruel) about others from behavioral descriptions (e.g., kicking puppies) of those individuals (spontaneous trait inference effect; Carlston & Skowronski, 2005; Uleman et al., 2008). Likewise, behaviors which are considered highly diagnostic (i.e., have high informational value) exert far more impact on trait inferences than those that are less diagnostic (e.g., Menon et al., 1995; Skowronski & Carlston, 1989). Some research examined how people infer specific personality traits (e.g., introvert; Trope & Bassok, 1983) while other research focused on how people infer more general valence traits (e.g., positive; Gregg et al., 2006), based on verbal behavioral descriptions.

As mentioned above, EC effects were viewed in the past as being mediated by the automatic formation and activation of mental associations between CS and US representations in memory (e.g., Baeyens et al., 1992; Levey & Martin, 1975). Yet, recent theoretical developments have called this idea into question (De Houwer, 2009, 2018). According to a propositional perspective, EC effects are not due to associative processes, but instead reflect the formation and activation of propositions in memory. These propositional representations differ from associative representations in that they (a) encode relational information (i.e., they specify how stimuli are related) and (b) have a truth value (i.e., they can be evaluated as being true or false).

Propositional accounts make several assumptions about EC effects that are not shared by associative models. One such assumption is that inferential processing plays a key role in EC effects (e.g., De Houwer et al., 2020). Because propositions have a truth value they allow for inferential reasoning to occur (i.e., the activation or construction of new propositional information based on its compatibility with other activated information). Thus, the propositional accounts make the unique prediction that inferences can underlie EC effects (Van Dessel et al., 2019). For example, when stimuli are paired with one another, a proposition is formed about this pairing (e.g., “A co-occurs with B”). This proposition can be combined with pre-existing propositions about the meaning of pairings (e.g., “similar things tend to co-occur”), which can lead to the formation of an inference about how those stimuli are related (e.g., “A and B are similar”). This then sets the stage for an evaluative inference (e.g., “A has the same valence as B”; see De Houwer, 2018) that transfers into evaluative responding.

From this perspective, the spatio-temporal pairing of stimuli, in combination with one’s prior knowledge, functions as a sort of argument. For humans, these “regularity-based” arguments can function in the same way as verbal arguments that consist of words and sentences (De Houwer & Hughes, 2016). If this is the case, then EC and persuasion effects may be more alike than previously assumed, insofar as both are driven by inferences. This idea could apply also to trait inferences given that they are a subclass of inferences that may be crucial when determining valence. Specifically, the proposition that is formed about pairings (e.g., “an unknown object [A] co-occurs with a positive stimulus [B]”) can be combined with pre-existing propositions about the meaning of pairings (“co-occurrence indicates similarity”) to form trait inferences (e.g., “A possesses similar traits as B”) which lead to changes in evaluative inferences (e.g., “A possesses similar valence as B”).¹

Our goal in the present research is to explore the potential role of trait inferences in EC effects. It is important to note that we use the term “trait inference” slightly differently than how it is traditionally used in persuasion research. As mentioned above, persuasion research usually uses the term “trait inference” to refer to inferring the traits of a person or an object based on information that does not involve telling participants about the person possessing the trait itself. In fact, there is typically no mention of the target trait itself. In the present research, in the context of EC, we also use the term “trait inference” to refer to the conclusion that a person or an object holds a trait. Importantly, we argue that this inference can be based on the spatio-temporal pairing of this person or object (CS) with a US. While we also do not tell participants about the person possessing the trait, for our purposes, the US can be a trait word as this can promote inferences of trait possession based on (pairing) information. Indeed, our goal is not to test if people can infer traits based on information that does not mention the trait but rather to test if people infer traits based on spatio-temporal pairing and if these inferences contribute to EC effects.

The idea that trait inferences can influence EC effects seems plausible given that many EC studies provide participants with instructions and cover stories prior to the pairing phase (e.g., Balas & Gawronski, 2012; Gast & Rothermund, 2011; Heycke & Gawronski, 2020; Hu et al., 2017; Hughes et al., 2019; Mitchell et al., 2003; Zanon et al., 2014). These instructions may serve as a (persuasive) context that promotes trait inferences. For example, some researchers (Balas & Gawronski, 2012; Mitchell et al., 2003; Zanon et al., 2014) told participants that the study concerned language acquisition and that they would be presented with words (CSs) together with pictures (USs) illustrating their meaning. Others informed participants that they would see images of pharmaceutical products (CSs) and visual information about their effects (USs) (Heycke & Gawronski, 2020; Hu et al., 2017). Such instructions or cover stories may direct participants to infer that the CSs have traits related to the USs (e.g., the pharmaceutical product causes positive health effects).

It is plausible that, even in those EC studies that do not use a cover story or provide pre-pairing instructions, the types of stimuli used (i.e., the specific CSs and USs) create conditions that promote trait inferences (e.g., Gibson, 2008; Kim et al., 1996; Sweldens et al., 2010). For example, it is possible that the presentation of Coca-Cola brand (CS) with the word “awesome” (US; Gibson, 2008) leads participants to infer that Coca-Cola is awesome (and therefore positive). Similarly, it is possible that the presentation of Belgium beers (CSs) with pictures of adults having fun in various ways (USs; Sweldens et al., 2010) leads participants to infer that the Belgium beers are fun and positive. Although there are also EC procedures that do not seem to promote trait inferences via instructions or biased selection of stimuli (e.g., Moran et al., 2021; Olson & Fazio, 2001), the frequent use of EC procedures that do promote trait inferences in this way suggests that trait inferences might have been crucial in many prior EC studies. Interestingly, however, no prior study has attempted to manipulate trait inferences to systematically investigate their potential role in EC effects.

It is worth noting that traits have been used as USs in prior EC studies (e.g., Gibson, 2008; Hughes et al., 2019; Milner et al., 2017). For instance, Wagner et al. (2020) set out to test the effectiveness of an EC procedure that was previously found to change child-related attitudes and expectations in adults (see Milner et al., 2017). To do so, the authors replicated the same EC procedure as Milner et al. but replaced the trait words that were originally used as USs with emoji. EC effects were found to be weaker (relative to the original research) and the authors concluded that trait inferences mediated the original effects. However, because Wagner et al.’s study did not directly compare the two types of USs, it is difficult to make strong conclusions regarding the role of trait inferences in their EC effects. The present research is the first to systematically test the potential role of trait inferences in EC by directly comparing the effect of different conditions that promote versus undermine trait inferences on EC effects.

Examining the role of trait inferences in EC could shed new light on the moderators of, and the processes underlying, EC as well as the relation between EC and persuasion. Moreover, it can inform us about a potentially important and widespread source of trait inferences that so far received little attention: the spatio-temporal pairing of stimuli.

Across four studies, we tested the idea that trait inferences contribute to EC effects. We did so in two ways. In Experiments 1-2, we manipulated trait inferences by pairing neutral stimuli (male faces) with valenced traits compared to valenced nouns. In Experiment 3, we provided information about the validity of pairings as an information source that can be used for trait inferences. Finally, in Experiment 4, we combined the two manipulations to test their interactive effect.

Experiments 1-2 used a job-hiring cover story and unfamiliar men (potential job candidates) as CSs. We targeted trait inferences by manipulating the content of the USs by either using as USs (a) positive (e.g., Puppies) and negative nouns (e.g., Cemetery) that are unlikely to give rise to trait inferences about the unfamiliar men (which was also verified in a pretest, see more details below) or (b) positive (e.g., Agreeable) and negative trait adjectives (e.g., Reckless) that are more likely to give rise to trait inferences (also see the pretest described below). The logic behind this manipulation was that that pairing of person with a personality trait word will lead to an inferential leap that involves seeing the personality trait word as being a property of the person (e.g., James is reckless). Such an inferential leap would be less likely to occur when persons are paired with nouns (e.g., James is cemetery).

EC effects were indexed using both self-report and automatic evaluation measures.² Discussions regarding the mental processes that might (differentially) impact performance in self-report and automatic evaluation measures are still ongoing (e.g., Bar-Anan & Nosek, 2012; Bar-Anan & Vianello, 2018; Payne et al., 2013; Schimmack, 2021). Therefore, rather than using these different measures to make claims about different mental processes, we used both types of measures primarily to provide convergent evidence for our claims, that is, to generalize the findings beyond one specific measure. However, it is also worth noting that different theoretical perspectives diverge in their predictions regarding the sensitivity of automatic versus self-reported evaluation to inference processes (e.g., De Houwer, 2014; Gawronski & Bodenhausen, 2018). We will return to this in the General Discussion. Experiments 1-2 were identical except for the automatic evaluation measure used. Experiment 1 used a modified version of the Affect Misattribution Procedure (AMP; Mann et al., 2019) and Experiment 2 used the Implicit Association Test (IAT; Greenwald et al., 1998).

Research on impression formation indicates that the perceived validity of encountered information influences subsequent evaluations (e.g., Golding et al., 1990; Peters & Gawronski, 2011). Specifically, when instructions suggest that a certain piece of information is not true, people usually discount or ignore this information, both when making trait inferences and when forming impressions. Elsewhere, others have found that when the source of information is considered to be low in credibility, the information has less of an impact on trait inferences and impression formation compared to when the source is considered to be high in credibility (e.g., Smith et al., 2013).

If trait inferences can also influence EC effects then similar findings should emerge when neutral male faces (CS) are paired with traits (USs) and the validity of the information source (i.e., pairings) is manipulated. Specifically, inferences that the CS has the US trait should be reduced when the validity of pairings is undermined compared to when it is promoted. With this in mind, Experiment 3 manipulated the validity of pairings as an information source. Participants in the low validity condition read that a hacker had been hired by a competing company to create chaos in their company (e.g., by deleting all the original [genuine] information about the two candidates and inserting fake information in its place). Participants in the high validity condition were additionally informed that the security department had identified this hacking problem and fixed it. As such, they would see the original (genuine) information about the candidates.

To further test the role of trait inference in EC effects, Experiment 4 combined the trait versus nouns as USs manipulation of Experiments 1-2 and the validity manipulation of Experiment 3. Combining these manipulations allowed us to test the boundaries of the trait inference idea. First, we tested whether the effect of using traits rather than nouns as USs is moderated by the perceived validity of pairings as an information source. To illustrate, imagine that people are told prior to an EC procedure that the pairings they will soon encounter are a non-valid piece of information. In such a case, does the content of the USs still matter (i.e., would pairing the men with trait adjectives rather than nouns still moderate EC effects)? If trait inferences contribute to EC effects, then the effect of CS-US pairings on evaluative responding should depend on the assumed relationship between the paired stimuli. If people know in advance that pairings do not indicate possession, then the content of the US (whether it is an adjective or noun) should not matter. Critically, no prior work has tested the mutual effects of validity and US content on EC effects. Finally, we examined if EC effects would still emerge under conditions designed to strongly reduce trait inferences (i.e., when the pairings are said to be an invalid source of information and the USs are unlike to elicit strong trait inferences [nouns]).

Materials, sampling plans, exclusion rules, and analysis plans for all experiments were pre-registered.³ We report all data exclusions, manipulations, measures, and how we determined our sample sizes. Materials and data of the whole project can be found at the Open Science Framework (https://osf.io/dphac/).

In Experiments 1-2, we targeted trait inferences by manipulating the content of the USs. We tested if the manipulation of US content would moderate EC effects. The two experiments were identical with the exception of the automatic evaluation measure used: Experiment 1 employed a modified version of the AMP (Mann et al., 2019). Experiment 2 replicated Experiment 1 while substituting the modified AMP for an IAT (Greenwald et al., 1998).⁴

Participants and design. Participants in Experiment 1-4 were recruited online via the Prolific Academic website (https://prolific.ac). Each experiment took about 14 minutes to complete and all participants were paid £1.40. Pre-registered materials are available at https://osf.io/p8cvf (Experiment 1), and https://osf.io/nuhw6 (Experiment 2).⁵ In Experiments 1-3, we used a Sequential Bayes Factor (SBF) design with a maximal N to determine sample size (Schönbrodt & Wagenmakers, 2018). We used a threshold of BF > 6 or BF < 0.16, a minimum of 60 participants, an addition of 60 participants for each test, and a maximum of 240 participants. In total, 241, and 239 participants completed Experiment 1 and 2, respectively. In line with the preregistrations, in Experiment 1, we excluded participants who had more than 95%, or less than 5%, pleasant responses on the AMP (i.e., 14 participants) (Moran et al., 2017). In Experiment 2, we excluded participants who (1) had more than 10% fast trials in the IAT (RT< 300ms; Greenwald et al., 2003) or (2) did not complete all the measures (total 7 participants). The final samples consisted of 227 participants for Experiment 1 (51% Women, M_age= 33.63, SD_age= 11.60) and 232 participants for Experiment 2 (56% Women, M_age= 34.10, SD_age= 10.86).

Experiment 1 involved a three-factor between-subjects design: 2 (US Content: nouns vs. adjectives) x 2 (CS-US assignment: James paired with positive, Chris paired with positive) x 2 (Measures order: self-report measure first, automatic evaluation measure first), with self-reported and automatic evaluations as the dependent variables. The design of Experiment 2 was the same, except for adding a fourth between-subjects factor (IAT block order: compatible block first, incompatible block first). The preregistered analytic strategy, in all the experiments, was to use an ANOVA model that includes the main manipulation and all the counterbalancing procedural factors. As specified in the preregistration, the counterbalancing procedural factors were included in the models to control for their influence, but their effects were not analyzed.

Materials. CSs were pictures of two males selected from an open database of facial stimuli (Minear & Park, 2004; pre-tested by Bar-Anan & Amzaleg-David, 2014). We named them Chris and James. The USs were positive and negative words (half of the words were adjectives and half were nouns; see Table 1), selected based on a pretest (N = 92, see the online-supplement and https://osf.io/v6kbj/ for details), wherein participants were asked to rate different adjectives and nouns on valence and the ability to infer evaluative traits from their pairing with a person (question: ‘how diagnostic is this word for inferring the true valence of the person? When we ask how ’diagnostic’ you consider the word to be, we are asking you to carefully think about how much presenting a person with this word tells you something about him [i.e., how positive or negative he is likely to be]’). The two positive sets did not differ significantly in valence ratings, t(91) = -0.45, p = .65, d = -0.04, BF₁₀ = 0.12, but they did differ significantly in trait-inference ratings, t(91) = 11.82, p < .001, d = 1.23, BF₁₀ > 1000. The positive adjectives were rated as more diagnostic (M = 3.10, SD = 0.75) than the positive nouns (M = 1.67, SD = 0.87). Likewise, the two negative subsets did not differ in valence ratings, t(91) = 0.26, p = .80, d = 0.02, BF₁₀ = 0.11, but did differ in trait-inference ratings, t(91) = 11.99, p < .001, d = 1.25, BF₁₀ > 1000. The negative adjectives were rated as more diagnostic (M = 3.20, SD = 0.75) than the negative nouns (M = 1.69, SD = 0.86).

As targets in the AMP (Experiment 1), we used 120 abstract paintings adapted from Mann et al. (2019). As primes in the AMP (Experiment 1) and men target stimuli in the IAT (Experiment 2), we used four different versions of the picture of Chris and of James that varied in their color (colored vs. grayscale) and the presence of a surrounding frame (no frame vs. yellow frame). The IAT also used four negative and four positive words (see Table 1).

Procedure.

Instructions. The full procedure, instructions, and materials of all the experiments are explained in detail in the online-supplement. At the beginning of the experiment, participants read a cover story asking them to imagine that they are working for a large company and that it is their job to recruit new employees for the company. We informed them that they would receive information about two potential candidates for the job: Chris and James and that their task was to judge which of the two men they would like to hire.

EC. The EC procedure consisted of two blocks of 24 trials (48 trials in total). Each trial simultaneously presented a male face and his name in compound (CS) on the left side of the screen and a valenced word (US) on the right side of the screen for 2500ms. The intertrial interval was 1000ms. During each block, one of the men (CS_pos) was presented three times with each of four different positive words, whereas the second man (CS_neg) was presented three times with each of four negative words. Assignment of the men to positive or negative USs was counterbalanced across participants. We manipulated the type of US words (nouns vs. adjectives) between participants.

Self-reported ratings. Participants rated their liking of each of the two men by answering three questions: ‘How good or bad do you consider the above individual?’, ‘How much do you like or dislike the above individual?’ and ’How positive or negative do you consider the above individual? Response scale was a 9-point Likert scale (-4 = Very bad/ I dislike him a lot/ Very negative; +4 = Very good/ I like him a lot/ Very positive). Rating scores for each target person were calculated by averaging the three questions (Cronbach’s Alpha > .92 in Experiments 1-4). We calculated EC scores on self-report ratings by subtracting the mean score rating for CS_pos from the mean score rating for CS_neg.

AMP. In Experiment 1, to measure automatic evaluation, we used a modified version of the AMP (Mann et al., 2019) that was designed to reduce intentional responding in the task. Each trial of the AMP displayed stimuli in the following sequence: (1) A photograph of one of the two targets for 100ms, (2) a blank screen for 100ms, (3) an abstract painting for 100ms, and (4) a pattern mask of black-and-white noise until participants responded. Upon presentation of the mask, participants were prompted to indicate if the painting was more pleasant or more unpleasant than average using two response keys on the keyboard signifying pleasant and unpleasant. Participants completed three blocks of 40 trials, each with 20 primes for each man. EC scores were computed by subtracting the proportion of pleasant responses after primes of the CS_pos from the proportion of pleasant responses after CS_neg primes. The internal consistency, as computed on the basis of two parcels (split-half) of the task, was α = .84.

IAT. In Experiment 2, we used the IAT to measure automatic evaluation. In the IAT, participants categorized stimuli using two computer keys. In the critical blocks, participants responded with the left key to stimuli of two categories (e.g., “James” and “Good”), and with the right key to stimuli of two other categories (e.g., “Chris” and “Bad”). In two of these blocks, James and Good shared the same response key, and in the other two critical blocks, Chris and Good shared the same response key. The seven-block IAT followed the procedure described in Nosek et al. (2005). The D4 algorithm was used to create IAT scores (Greenwald et al., 2003). We computed these scores such that they indicate a more positive evaluation for CS_pos than for CS_neg. The IAT internal consistency, as computed on the basis of two parcels (split-half) of the task, was α = .79.

Exploratory questions. At the end of the experiment, the participants answered exploratory questions related to their experiences in the study (for a full description of these questions and associated findings see the online-supplement).

Self-report ratings. The mean EC scores on self-report ratings (and automatic evaluations) as a function of US content condition in Experiments 1-2 are detailed in Table 2. The ANOVA model (see Table 3 for details on the ANOVA models in all the experiments) revealed a main effect of US content in Experiment 1: F(1, 219) = 73.33, p < .001, η_p² = .25, 90%CI [.16, .31], BF₁₀ >1000, and in Experiment 2: F(1, 216) = 53.18, p < .001, η_p² = .20, 90%CI [.11, .25], BF₁₀ >1000. In all cases, EC effects were stronger when USs were trait-adjectives (Ms = 4.56, 3.76, SDs = 2.22, 2.50, in Experiments 1-2, respectively) than nouns (Ms = 1.61, 1.51, SDs = 2.91, 2.39). In both experiments, EC effects significantly differed from zero both in the adjectives and the nouns conditions (see Table 2).

Automatic evaluation.

AMP scores. The ANOVA (see Table 3) on AMP scores in Experiment 1 did not show an effect of US content, F(1, 219) = 1.21, p = .27, η_p² = .01, 90%CI [0, .03], BF₁₀ = 0.26, indicating no significant difference between the size of EC effects when USs were adjectives (M = 0.04, SD = 0.19) or nouns (M = 0.07, SD = 0.22) (see Figure 1). EC effects significantly differed from zero in both conditions (Table 2)

IAT scores. Unlike the results of Experiment 1, the ANOVA (Table 3) on automatic evaluations in Experiment 2 revealed main effects of US content, F(1, 216) = 7.48, p = .007, η_p² = .03, 90%CI [.004, .07], BF₁₀ = 7.39, indicating a stronger IAT effect when USs were adjectives (M = 0.30, SD = 0.46) than nouns (M = 0.14, SD = 0.41). IAT effects significantly differed from zero in the adjectives and nouns conditions (Table 2).

Direct comparison between self-reported and automatic evaluation effects.⁶ To directly compare the effect of US content on self-reported and automatic evaluations, we first computed a preference for James over Chris for both measures and then standardized these scores. Next, we recoded the standardized scores to reflect a preference for CS_pos over CS_neg. The mean standardized EC scores as a function of US content condition and measure type, in Experiments 1-2 are illustrated in Figure 1 (panels A-B, respectively). We submitted these standardized scores to mixed ANOVA models (see Table 3). Table 4 summarizes the full results of the ANOVAs. Importantly, the interaction between US content and measure type was significant in Experiment 1, F(1, 219) = 37.86, p < .001, η_p² = .15, 90%CI [.08, .21], BF₁₀ >1000⁷, and Experiment 2, F(1, 216) = 5.60, p = .019, η_p² = .03, 90%CI [.002, .06], BF₁₀ = 1.05. In both experiments, the interaction indicated that the effect of US content on self-report ratings, (Experiment 1: F(1, 219) = 73.33, p < .001, η_p² = .25, BF₁₀ >1000; Experiment 2: F(1, 216) = 53.18, p < .001, η_p² = .20, BF₁₀ > 1000) was stronger than the effect on automatic evaluations, (Experiment 1: F(1, 219) = 1.21, p = .272, η_p² = .01, BF₁₀ = 0.27; Experiment 2: F(1, 216) = 7.48, p = .007, η_p² = .03, BF₁₀ = 7.84).

Results indicate that the nature of the US (nouns vs. trait-adjectives) strongly moderated EC effects as indexed by self-report ratings. Effects were stronger when USs were trait-adjectives compared to when they were nouns. This was also the case for automatic evaluations when measured by an IAT, but not when measured using a modified AMP (Mann et al., 2019). A direct comparison of this moderating effect across self-reported and automatic evaluation measures revealed a stronger effect of US content on the former than on the latter types of measures.

These findings support the idea that trait inferences moderate self-reported and automatic evaluations (at least when measured with the IAT), and the impact of such a manipulation is stronger for self-reported than automatic evaluations. It is also worth noting that, for all measures and in all experiments, EC effects always significantly differed from zero when USs were nouns (Table 2). This suggests that people preferred the CS paired with positive USs over the CS paired with negative USs even under the conditions that did not encourage trait inferences. In Experiment 3, we tested if trait inferences contribute to EC effects by using a different manipulation of trait inferences. Specifically, we manipulated the perceived validity of pairings as an information source.⁸

Participants and design. Preregistered materials for Experiment 3 are available at https://osf.io/jsg2h. 268 participants completed Experiment 3. We excluded participants who had more than 10% of trials that were too quick in the IAT or those who did not complete all measures (i.e., 28 participants). The final sample included 240 participants (55% Women, M_age= 34.20, SD_age= 11.40). The experiment involved a four factor between-subjects design: 2 (Validity: high vs. low) x 2 (CS-US assignment) x 2 (Measures order) x 2 (IAT block order), with self-reported and automatic evaluations as the main dependent variables.

Materials and procedure.

Instructions. To manipulate the validity of the pairing as an information source, we presented different instructions before the EC phase. In the low validity condition, we modified the cover story to suggest that a hacker, hired by a competing company, deleted the original information about Chris and James and inserted new fake information about them. In the high validity condition, the instructions also informed the participants about the hacker but also informed them that the security department identified the problem and fixed it and that the database now included the original and correct information about Chris and James. Prior to the study, the two cover stories were pretested (N = 351, see https://osf.io/mpyqv/ and the online-supplement for more details) for their believability and the extent to which they promote trait inference (i.e., to what extent was the pairing of James and Chris with positive or negative information reflective of something about their true character).

Immediately after reading the validity instructions, participants answered two open-ended questions: “Please describe - in a few words - what you were just told”, and “What implications does this information have for what you are soon going to learn about Chris and James?” The purpose of including these two questions was to increase the chance that participants would elaborate on the implications of the cover story for the pairings they were about to see. We did not analyze the question responses nor exclude participants based on their responding.

Other experiment phases. All other procedural details (EC procedure, self-report and IAT measures) were the same as in Experiment 1-2 except that in the EC procedure, for all participants, USs were always adjectives. Participants also answered the same exploratory questions as in Experiments 1-2 with the addition of a believability question about the cover stories (see online supplement). The internal consistency of the IAT was α = .80.

Self-report ratings. The ANOVA (see Table 3) revealed a main effect of validity, F(1, 224) = 53.90, p < .001, η_p² = .19, 90%CI [.11, .25], BF₁₀ >1000, indicating a stronger effect when participants read the high validity instructions (M = 4.10, SD = 2.40) than the low validity instructions (M = 1.27, SD = 3.45). Effects significantly differed from zero in the high and low validity conditions (see Table 2).

IAT scores. The ANOVA on the IAT scores (Table 3) revealed a significant effect of validity, F(1, 224) = 7.41, p = .007, η_p² = .03, 90%CI [.004, .07], BF₁₀ = 5.14, indicating a stronger effect when participants received the high validity instructions (M = 0.32, SD = 0.44) than low validity instructions (M = 0.18, SD = 0.44). Again, effects significantly differed from zero in both the high and low validity conditions (Table 2).

Direct comparison between self-reported and automatic evaluation effects. Standardized scores were computed and analyzed as in Experiment 1-2 with replacement of the US content factor with the Validity factor. The mean standardized scores as a function of validity condition and measure type, are illustrated in Figure 2, the ANOVA model is specified in Table 3, and the ANOVA results are summarized in Table 4. Importantly, the interaction between validity and measure type was significant, F(1, 224) = 8.78, p = .003, η_p² = .04, 90%CI [.007, .08], BF₁₀ = 10.38. The interaction revealed a stronger effect of validity on the self-report ratings, F(1, 224) = 53.90, p < .001, η_p² = .19, BF₁₀ > 1000, than on IAT effects, F(1, 224) = 7.41, p = .007, η_p² = .03, BF₁₀ = 4.22.

Manipulating how valid pairings are as an information source strongly moderated EC effects as indexed by self-report ratings and IAT scores. A direct comparison between self-reported and automatic evaluative measures revealed a stronger impact of validity on the former relative to the latter. In addition, both automatic and self-reported EC effects were significantly different from zero in the low validity condition. Taken together, the results of Experiment 3 indicate that creating conditions that undermine trait inferences leads to weaker (but still significant) EC effects. In Experiment 4, we sought to test the boundary conditions of trait inferences on EC effects by combining the trait inference (US content) manipulation of Experiments 1-2 with the trait inference (validity) manipulation of Experiment 3.

Participants and design. Preregistered materials are available at https://osf.io/ys8t4. Based on power analyses, we estimated that 517 participants would provide 90% power to detect a small interaction effect of η_p²= 0.02. To account for possible exclusions, we decided to collect data from at least 600 participants. As planned, we excluded participants who had more than 10% of trials in the IAT that were too quick, and those who did not complete all measures (i.e., 15 participants). The final sample consisted of 585 participants (56% women, M_age = 35.43, SD_age = 12.80). The experiment involved a five-factor between-subjects design: 2 (US Content: adjectives vs. nouns) x 2 (Validity: high vs. low) x 2 (CS-US assignment) x 2 (Measures order) x 2 (IAT block order).

Materials and procedure. To manipulate US content, we used the same stimuli and procedures as in Experiments 1-2. To manipulate validity, we used the same cover stories as in Experiment 3. IAT and self-report tasks were similar to Experiments 2-3. The internal consistency of the IAT was α = .78. Exploratory questions were the same as in Experiment 3.

Analytic strategy. To test the interactive effect of US content and validity on EC effects, we submitted self-reported ratings and IAT scores to a between-participants ANOVA (see details on Table 3). We also preregistered specific hypotheses for a number of comparisons. Details on these comparisons and their results are summarized in Table 5 and in the online-supplement.

Self-report ratings. The mean EC scores on self-report ratings (and automatic evaluations) as a function of validity and US content conditions in Experiment 4, are detailed in Table 2. Effects were positive and significantly different from zero in all four conditions (Table 2). The ANOVA (Table 3) revealed a significant interaction between US content and validity, F(1, 552) = 20.77, p < .001, η_p² = .04, 90% CI [.01, .06], BF₁₀ >1000. The interaction revealed an effect of US content in the high validity condition, F(1, 262) = 56.90, p < .001, η_p² = .18, BF₁₀ >1000, but not in the low validity condition, F(1, 290) = 1.03, p =.31, η_p² < .01, BF₁₀ = 0.21. The ANOVA also found main effect of US content, F(1, 552) = 36.00, p < .001, η_p² = .06, 90%CI [.03, .09], BF₁₀ >1000, indicating stronger effects when USs were adjectives (M = 2.62, SD = 3.26) than nouns (M = 1.29, SD = 2.51). A main effect also emerged for validity, F(1, 552) = 57.93, p < .001, η_p² = .09, 90%CI [.05, .12], BF₁₀ >1000, indicating a stronger effect when participants received high validity instructions (M = 2.83, SD = 2.86) compared to low validity instructions (M = 1.14, SD = 2.86).

IAT scores. IAT effects were positive and significantly different from zero in all four conditions (see Table 2). The ANOVA (Table 3) revealed a significant interaction between US content and validity, F(1, 552) = 8.61, p = .003, η_p² = .02, 90%CI [.002, .03], BF₁₀ = 5.49. The interaction revealed a significant impact of US content on IAT effects under high validity conditions, F(1, 262) = 23.64, p < .001, η_p² = .08, BF₁₀ >1000, but not under low validity conditions, F(1, 290) = 0.19, p = .663, η_p² < .01, BF₁₀ = 0.14. The ANOVA also revealed a main effect of US content, F(1, 552) = 12.78, p < .001, η_p² = .02, 90%CI [.006, .04], BF₁₀ = 14.02, indicating a stronger effect when USs were adjectives (M = 0.23, SD = 0.43) than nouns (M = 0.12, SD = 0.40). A main effect also emerged for validity, F(1, 552) = 9.49, p = .002, η_p² = .02, 90%CI [.003, .03], BF₁₀ = 8.81, indicating a stronger effect when participants received high validity (M = 0.23, SD = 0.40) compared to low validity instructions (M = 0.13, SD = 0.43).

Direct comparison between self-reported and automatic evaluation effects. Standardized scores were computed as in previous experiments. The mean standardized scores as a function of validity, US content, and measure type, are illustrated in Figure 3. We submitted these scores to a mixed ANOVA (see Table 3 for details). Table 4 summarize the ANOVA results. Importantly, the interaction between validity and measure type was significant, F(1, 552) = 8.91, p = .003, η_p² = .02, 90%CI [.003, .03], BF₁₀ = 5.07, such that validity more strongly impacted self-reported ratings, F(1, 552) = 57.93, p < .001, η_p² = .09, than IAT effects, F(1, 552) = 9.49, p = .002, η_p² = .02. The interactions between US content and measure type, F(1, 552) = 1.85, p = .18, η_p² < .01, 90%CI [0, .01], BF₁₀ = 0.29, and the three-way interaction between US content, validity, and measure type, F(1, 552) = 0.71, p = .40, η_p² < .01, 90%CI [0, .01], BF₁₀ = 0.21, were all non-significant.

Our key findings from Experiments 1-3 replicated. Two types of trait inference (US content and validity) manipulations moderated EC effects as indexed by self-report ratings and IAT scores. Validity exerted a stronger impact on EC effects on self-reported compared to IAT scores, whereas the effect of US content was consistent across measures. US content and validity also interacted, such that the former exerted a stronger influence on evaluative responses whenever pairings were considered to be a valid piece of information. When pairings are considered to be an invalid source of information, US content did not moderate evaluative responding. EC effects were significant in all the conditions, including when the pairings were low in validity and US were nouns.

There was variability in how US content and validity manipulations moderated EC effects across our six experiments (the four experiments reported in the main text and the two experiments reported in the online-supplement). We therefore decided to meta-analyze (fixed effects) those experiments that manipulated US content (Experiments 1, 2, S1, and 4) and validity (Experiments S2, 3 and 4). The meta-analytic effect of US content on self-reported ratings, Hedges’ g = 0.70, SE = 0.06, 95%CI [0.58, 0.82], Z = 11.51, p < .001, and automatic evaluations, Hedges’ g = 0.23, SE = 0.06, 95%CI [0.11, 0.35], Z = 3.86, p < .001, were both significant. The meta-analytic effect of validity on self-reported ratings, Hedges’ g = 0.65, SE = 0.06, 95%CI [0.53, 0.78], Z = 10.32, p < .001, and automatic evaluation, Hedges’ g = 0.27, SE = 0.06, 95%CI [0.15, 0.39], Z = 4.36, p < .001, were also significant. In short, US content and validity were genuine and general moderators of EC effects.

Across four studies, we examined the role of inferential processes – and specifically trait inferences – in EC effects. Results consistently supported the idea that manipulations designed to promote (Experiments 1-2) or undermine (Experiment 3) trait inferences influenced the magnitude of self-reported and automatic evaluations, consistently influencing the former more so than the latter. Interestingly, we found evidence for (self-reported and automatic) EC effects under conditions that were designed to undermine trait inferences: when the USs were nouns (Experiments 1-2), when the validity of pairings as an information source was low (Experiment 3), and when these two conditions were combined (Experiment 4).

As is the case for many mental processes, it is difficult if not impossible to demonstrate beyond any doubt that participants made specific trait inferences (e.g., De Houwer, 2011). That said, there are several reasons to substantiate the conclusion that the impact of US type (noun vs. adjective) on EC was mediated by trait inferences. First, the pretest we conducted for choosing the adjective and noun words suggests that the adjectives are more likely to lead to evaluative trait inferences than the nouns. Second, the fact that the validity manipulation influenced the effect of US type (Experiment 4: adjective USs led to stronger EC effects than noun USs only when the pairings were described as valid but not when the pairings were described as invalid) also fits very well with the idea that inferences were crucial to this effect. Although one could always argue that our results reflect some other difference between adjectives and nouns (e.g., word concreteness), we can find no arguments to support such alternative accounts.

As we mentioned in the introduction, our work was inspired by a propositional perspective on EC according to which stimulus pairings function as an argument in much the same way as verbal arguments do and are combined with prior knowledge to inform evaluative responding (see De Houwer, 2018; De Houwer & Hughes, 2016). Building on this idea, we argued that EC and persuasion effects may be more alike than previously assumed and may therefore also be moderated by manipulations that target trait inferences.

The current findings support this idea on several fronts. First, moderation of EC effects by US content is similar to the moderation of persuasion effects by information diagnosticity. For example, Skowronski and Carlston (1987) argued that the perceived informativeness of behavior is related to its trait diagnosticity - the extent to which a behavior promotes the conclusion that an actor possesses one trait and prevents the conclusion that the actor possesses the alternative trait. Being told that “Lisa engages in lots of volunteer work” increases the probability that Lisa is considered generous and lowers the probability that Lisa is considered selfish and therefore is said to be diagnostic for inferences about these traits (Skowronski & Carlston, 1987). Other research that focused on stereotypes formation, treated diagnosticity as the extent to which attributes provide the most valuable information for differentiating between groups (Ford & Stangor, 1992). Cone and Ferguson (2015), that tested the role of diagnosticity in the revisions of automatic evaluation, defined diagnosticity as the extent to which information is revealing of a person’s true nature, traits, or character. The common ground for all of this research is that it treats diagnosticity as the extent to which information is considered relevant for identifying the nature and cause of something.

Work in this area indicates that information which is considered to be highly diagnostic exert far more impact on judgments than information considered to be less diagnostic (e.g., Cone & Ferguson, 2015; Menon et al., 1995; Skowronski & Carlston, 1992). The manipulation of US content used in the present research can also be viewed as a type of manipulation of diagnosticity of trait inferences. Specifically, both the pretest of Experiments 1-2 and the exploratory questions indicated that participants perceived the adjectives as more diagnostic (relevant and informative) for inferring the men’s true character than the nouns. As such, the present research demonstrates that the same factor (i.e., diagnosticity) has similar moderation effects in the contexts of EC as in persuasion.

Second, moderation of EC effects by the validity of pairings as an information source is also similar to the moderation of persuasion effects by validity (and source credibility) information. Persuasion research indicates that validity information exerts a strong impact on evaluations (e.g., Cone et al., 2019; Gregg et al., 2006; Peters & Gawronski, 2011). For example, Cone and colleagues (2019) found that negative information suggesting that a man committed a crime leads to stronger negative evaluations when people were told that this information was derived from police reports than from a coworker who had reason to spread negative rumors. The current work suggests that EC effects, like persuasion effects, are moderated by information about the validity of the evaluative information (regardless if this information constitutes verbal arguments or stimulus pairings). In short, there is marked similarity between the factors that influence EC and those that influence persuasion effects.

In the current research, we used both self-reported and automatic evaluation measures mainly to generalize our findings beyond one specific measure. Moreover, discussions regarding which mental processes underlie self-report versus automatic evaluation measures are still ongoing (e.g., Bar-Anan & Nosek, 2012; Schimmack, 2021). Nevertheless, it is still useful to look at the difference between self-report versus automatic evaluation measures in terms of operating conditions: automatic evaluation measures provide less optimal processing conditions than self-report measures (e.g., Moors et al., 2010; Van Dessel et al., 2020). From this perspective, the results from self-reported and automatic evaluation measures in the current research can impose empirical constraints on evaluation theories (e.g., Gawronski et al., 2020).

The main findings that (1) EC effects on self-reported and automatic evaluation measures are moderated by manipulations that may promote or undermine trait inferences, (2) the moderation effects are stronger on self-report measures, and (3) EC effects on both types of measures are significant also in conditions that are designed to undermine trait inferences, can be used to constrain associative, dual-process, and propositional accounts of EC. First, if EC is mediated purely by the formation (and activation) of mental associations between CS and US (valence), then the specific content of the USs or the credibility of the source who provides the pairings should not influence EC effects. This is because associative mental representations only encode the CS and US information (valence), and not how relevant are those pairings for trait inferences. As such, we find it difficult to reconcile the present findings with such a perspective.

Second, our results constrain dual-process accounts of EC, such as the associative-propositional evaluation (APE) model (Gawronski & Bodenhausen, 2018), which claim that EC effects can result from two distinct mental processes: associative and propositional. Associative processes refer to the formation and activation of a mental association between CS and US, based on the spatio-temporal relationship between the two. Propositional processes, on the other hand, refer to the formation and activation of mental representations based on inferential processes and the perceived validity of the encoded CS-US relation. A central premise of the APE model is that mere pairings influence mental representations via associative processes even when pairings in the environment are considered invalid during encoding (Gawronski & Bodenhausen, 2018). Moreover, according to the APE model, automatic evaluations mainly reflect the outcome of these associative processes whereas self-reported evaluations mainly reflect the outcome of the propositional processes. From such a perspective, trait inferences (manipulated by US content or validity) should influence self-reported evaluations directly but should not have a direct effect on automatic evaluation. The findings that EC effects were significant also in conditions that are designed to prevent trait inferences, and that the trait inference manipulations had a stronger effect on self-reported evaluations, can be seen as a support for associative processes that influence EC in addition to propositional processes. However, the finding of the present research, that manipulations that promote or undermine trait inferences influenced both self-reported and automatic EC effects do not align with this perspective. The APE model does allow for an indirect influence of trait inferences on automatic evaluation if, during the encoding of information, these inferences lead to the formation of mental associations that reflect diagnosticity or validity information. Critically, however, the APE model, at the moment, does not provide clear predictions regarding the specific conditions under which trait inferences would impact EC effects on automatic evaluation.

Finally, several of our results fit well with the propositional perspective of EC which provides a priori predictions regarding a moderating impact of trait inferences on both self-reported and automatic evaluations. The propositional perspective argues that CS evaluation depends on the extent to which participants can infer CS valence or related traits based on CS-US pairings. It therefore assumes that EC effects are mediated by the formation (and activation) of propositions that can encode information about whether stimulus content or pairings are relevant (or credible) for inferring CS valence. From this perspective, both self-reported and automatic evaluations reflect the formation and activation of such propositions (e.g., De Houwer, 2014). According to this perspective, the content of the US and the validity of CS-US pairings should moderate EC effects on both self-reported and automatic evaluations, as was found in the present research.

The finding that EC effects were significant also in conditions that were designed to undermine trait inferences fits less well with the propositional perspective. One could account for such findings by assuming that even in the conditions that were designed to undermine trait inferences some sort of trait inference occurred. Specifically, some of the nouns used in Experiments 1-2 and 4 might have been used to indirectly infer traits. For example, the pairing of Chris with the word “puppies” could have led to the inference that “Chris is like puppies”, or the inference that “Chris hangs around puppies” which both suggest that he is a nice person. It is also possible that in the condition in which the pairing was described as invalid (Experiments 3-4), some participants did not pay attention to the instructions and still used trait inferences.

Alternatively, it is possible that trait inferences are not the only inferences that drive EC effects. Indeed, the propositional perspective also highlights other regularity based inferences (e.g., inferences that co-occurrence of a valenced stimulus and another stimulus indicates similarity in valence; e.g., Hughes et al., 2016) as central to EC effects. Further, the finding that trait inference manipulations had a stronger influence on self-reported evaluations than on automatic evaluation can explained by assuming that when evaluation conditions are suboptimal (i.e., when evaluation) there is less control of which of the inferences will influence performance. However, this alternative account needs to be validated in future research before it can be given much weight.

EC procedures have been used in marketing contexts (e.g., Gibson, 2008; Sweldens et al., 2010). Some have even argued that EC procedures can be used as an intervention to change unwanted attitudes and behavior in areas such as alcohol and snacks consumption (e.g., Houben et al., 2010; Walsh & Kiviniemi, 2014). Until now, inferential processes have rarely – if ever – been taken into account when designing applied EC-based interventions. In many cases, USs are not chosen based on their relevance to the CSs. Consider, for instance, a study testing whether EC can increase actual fruit (CS) consumption (Walsh & Kiviniemi, 2014). In this case, the USs were generic IAPS pictures (Lang et al., 2008) of a waterfall, a chipmunk, and an astronaut. Other applied studies have sought to minimize awareness of (and thus inferences about) the CS-US pairings, by embedding those pairings in a rapid stream presentation of unrelated stimuli (e.g., Gibson, 2008).

The current findings suggest that EC-based interventions can be ‘supercharged’ by actively shaping the types of inferences that people will make in the context of an EC procedure, and directing those inferences in a more precise manner. For instance, the effectiveness of an EC task designed to increase fruit (CS) consumption could be improved by choosing USs that are directly relevant for fruit consumption (e.g., words and images that refer to healthiness rather than general positive but irrelevant words or images). Similarly, the effectiveness of EC procedures may increase when those tasks create optimal conditions for inferring that the CS-US pairing is a relevant and credible piece of information instead of directing attention away from that’s information source.

From the perspective of persuasion research, our studies highlight a new way of inducing trait inferences and thereby changing evaluations and behavior. Whereas persuasion typically involves verbal message that blatantly communicate information about behaviors or traits, merely pairing stimuli can provide a more subtle way to induce trait inferences. Especially in cases where persuasive messages are likely to evoke reactance, stimulus pairings may provide an alternative pathway for inducing trait inferences that is less likely to evoke reactance (De Houwer & Hughes, 2016).

The present research was inspired by a propositional perspective on EC which assumes that EC effects are driven by inferences. Importantly, as we mentioned before, trait inferences are probably not the only type of inferences that drive EC effects. For example, conditions that promote inferences about positive or negative appearance (e.g., using “beautiful” and “ugly” as USs) or inferences about positive or negative states (e.g., using “energetic” and “tired” as USs) may increase EC effects relatively to conditions that do not promote inferences. Future research can test if other types of inferences contribute to EC effects.

The current research is subject to a number of limitations. First, our investigation used a restricted set of stimuli (i.e., CSs were men, USs were words), one specific context (i.e., hiring decision) and a specific sample (Prolific participants). Future research should extend this line of research by examining the role of trait inferences in EC with different stimuli, contexts, and samples. Indeed, the observation that trait inferences can play a role in EC does not constitute that they play a role in all types of EC (e.g., in EC procedures which does not mention traits or cover stories: e.g., Moran et al., 2021; Olson & Fazio, 2001). Second, US content and validity effects were observed on only one self-report measure (liking ratings) and one automatic evaluation measure (the IAT). Given that US content did not impact evaluations assessed using another measure (a modified version of the AMP), it is possible that the observed effects on automatic evaluation are due to specific properties of the IAT and will not transfer to other automatic evaluation measures. Future research might examine this question by using different automatic evaluation measures such as the Evaluative Priming task (Fazio et al., 1995) or the original version of the AMP (Payne et al., 2005). Third, the current research design does not allow to test if EC effects and their moderation by trait inference were due to changes in the evaluation of the CS that was paired with positive stimuli, the CS that was paired with negative stimuli, or both. Future research can shed more light on the bi-directionality of the explored effects by including baseline ratings of the CSs before the conditioning procedure.

Inspired by a propositional perspective of EC, the present research examined the role of inferential processes, and specifically trait inferences, in EC. We found that conditions which promote trait inferences led to stronger EC effects than conditions which undermine such inferences. We also found that conditions that undermine trait inferences do not eliminate EC effects. The results of the present research suggest that inferential processes play an important role in EC effects but also that trait inferences may not be the only pathway to EC effects.

TM contributed to the experimental design, creation of the materials, data collection, analysis, and writing of the paper. SH and PVD contributed to the experimental design and writing of the paper. JDH contributed to the writing of the paper. All authors contributed to the conceptualization of the research idea.

The authors declare that they have no competing interests.

This manuscript is supported by Ghent University grant BOF16/MET_V/002 to JDH and by grant FWO19/PDS/041 of the Scientific Research Foundation, Flanders to PVD.

All the stimuli, presentation materials, participant data, and analysis scripts can be found on the Open Science Framework (https://osf.io/dphac/).