Cognitive Load Impairs Evaluative Conditioning, Even When Individual CS and US Stimuli are Successfully Encoded

Evaluative Conditioning (EC) consists of a change in the valence of conditioned stimuli (CS) following their pairing with valent stimuli (unconditioned stimuli, or USs). The EC procedure has been regularly used for testing the automaticity of attitude formation (for a recent review, see Corneille & Stahl, 2019). Automaticity refers to four operating conditions: (1) unawareness of the CS-US pairings, (2) efficiency, (3) goal-independence, and (4) uncontrollability (Bargh, 1994). In this research, we focus on the second one: efficiency (i.e., the possibility for the process to operate with minimal attentional resources) (Bargh, 1994). If EC relies on an efficient learning process, then it should be observed even when cognitive resources are depleted at encoding. Against the efficiency view, previous research indicates the EC effect as well as memory of CS-US pairings are drastically reduced, often to non-significance, under cognitive load conditions (Dedonder et al., 2010; Mierop, Hütter & Corneille, 2017; Pleyers et al., 2009). As cognitive load was implemented during participants’ exposure to the CS-US pairs in these experiments, the availability of cognitive resources appears necessary for the successful encoding of CS-US pairings and, presumably as a result, for the establishment of an EC effect. CS-US relations encoding would thus be a critical, yet non-efficient, process in the establishment of EC effects. Importantly, this account assumes load hampers the encoding of the CS-US relation while it preserves the encoding of the individual CS and US stimuli that compose CS-US pairs. This assumption has not been tested yet, despite its importance. That is, should load impact the very encoding of individual stimuli, then the effect of load on EC would not conclusively speak to the efficiency of attitude acquisition. Rather, it would speak to the efficiency of the encoding of individual CS and US stimuli. Blask, Walther, Halbeisen, and Weil (2012) correctly pointed out that a secondary task may prevent the encoding of the individual stimuli. Consistent with this, Halbeisen and Walther (2015) observed EC but not CS-US pairing memory resists a load manipulation when the load materials rely on a different modality than CS-US materials. These authors concluded EC can be established through an efficient encoding of the individual stimuli entering CS-US pairs, despite the unsuccessful encoding of the CS-US relation. Results from that experiment, however, are at odds with many other experiments that failed to find efficient EC effect without CS-US memory when also relying on dissimilarity conditions (e.g., Dedonder, 2010; Mierop et al., 2017). Also consistent with the importance of encoding CS-US relations in EC effects is an experiment by Kattner (2012). This author observed no EC when participants’ attention was directed to individual CS and US stimuli but was diverted from their pairings. Importantly, however, this research manipulated the focus of attention by using a processing goals modification. As was also the case in other experiments manipulating processing goals (e.g., Gast & Rothermund, 2011), participants were instructed to pay attention to irrelevant pairings (i.e., links of each CS and US with a random digit). Hence, this research speaks to goal-dependency (see also Corneille et al., 2009; Mierop, A., et al. (2020). Cognitive Load Impairs Evaluative Conditioning, Even When Individual CS and US Stimuli are Successfully Encoded. International Review of Social Psychology, 33(1): 3, 1–7. DOI: https://doi.org/10.5334/irsp.339

encoding of the individual CS and US stimuli that compose CS-US pairs. This assumption has not been tested yet, despite its importance. That is, should load impact the very encoding of individual stimuli, then the effect of load on EC would not conclusively speak to the efficiency of attitude acquisition. Rather, it would speak to the efficiency of the encoding of individual CS and US stimuli. Blask, Walther, Halbeisen, and Weil (2012) correctly pointed out that a secondary task may prevent the encoding of the individual stimuli. Consistent with this, Halbeisen and Walther (2015) observed EC but not CS-US pairing memory resists a load manipulation when the load materials rely on a different modality than CS-US materials. These authors concluded EC can be established through an efficient encoding of the individual stimuli entering CS-US pairs, despite the unsuccessful encoding of the CS-US relation. Results from that experiment, however, are at odds with many other experiments that failed to find efficient EC effect without CS-US memory when also relying on dissimilarity conditions (e.g., Dedonder, 2010;Mierop et al., 2017).
Also consistent with the importance of encoding CS-US relations in EC effects is an experiment by Kattner (2012). This author observed no EC when participants' attention was directed to individual CS and US stimuli but was diverted from their pairings. Importantly, however, this research manipulated the focus of attention by using a processing goals modification. As was also the case in other experiments manipulating processing goals (e.g., Gast & Rothermund, 2011), participants were instructed to pay attention to irrelevant pairings (i.e., links of each CS and US with a random digit). Hence, this research speaks to goal-dependency (see also Corneille et al., 2009;Gast & Rothermund, 2011;Stahl, Haaf & Corneille, 2016;Verwijmeren et al., 2012), not to efficiency. Because automaticity features do not perfectly overlap with each other (Melnikoff & Bargh, 2018;Moors & De Houwer, 2006), the goal-dependency of EC does not imply EC is non-efficient.
To advance our understanding of the efficiency of attitude learning, we examined whether attitude formation is observed in an evaluative conditioning procedure when the encoding of individual CS and US stimuli is successfully achieved but the encoding of their pairing is not. This was investigated by manipulating the presence or absence of an auditory two-back task at encoding and by measuring the memory of individual CS and US stimuli, as well as of their pairing. Based on growing evidence questioning automatic attitude learning (Corneille & Stahl, 2019), we predicted we would replicate the findings that cognitive load (1) reduces EC and (2) weakens CS-US pairings memory. In addition, we (3) tested whether cognitive load preserves the memory of individual CS and US stimuli entering CS-US pairings and (4) examined whether EC effects are reduced in the load condition when memory for individual CS and US memory is preserved. Mediational analyses were additionally conducted to determine if the load effect on EC was mediated by participants' memory for individual CSs and USs and CS-US relations.

Method
We report how we determined our sample size, all data exclusions, all manipulations, and all measures. The pre-registration, program script, raw data, and analytic script are publicly available on Open Science Framework (osf.io/g9wds).

Participants
Eighty-one undergraduate students participated in the experiment (58 women, 23 men, M age = 20.49, SD age = 1.91). They were randomly assigned to a control (N = 41) or load (N = 40) condition. The 'pwr.f2.test' function of the 'pwr' package (Champely et al., 2013) in R (R Development Core Team, 2017) determined this sample size allows for detecting with a statistical power of (1-β) = 0.8 and a Type I error probability of α = 0.05 effect sizes as low as η 2 p = 0.09, which is an intermediate effect size according to Cohen's (1988) norm. In an Integrative Data Analysis conducted on data from three experiments, Mierop and colleagues (2017) observed the overall effect of load on EC is of intermediate size (η 2 p = 0.08).

Procedure and Material
The experiment was programmed with EPrime 2.0. Participants went through three phases. In the first-conditioning-phase, eight CS-US pairs were presented seven times, each for 1000 ms, in a random order. The CSs represented neutral consumption products (e.g., chewing-gums); whereas, the USs were pictures taken from the International Affective Picture System (IAPS, Lang, Bradley & Cuthbert, 1997). Four CSs were paired with negative USs (IAPS references 2715(IAPS references , 2750(IAPS references , 6360, 6561, 2550 and four CSs were paired with positive USs (IAPS references 4608, 4700, 8200, 8460, 2141, 2900.1, 6315, 6510; CS-US pairings were counterbalanced across participants). At the trial level, a CS was presented at the bottom center of the screen; whereas, the US was simultaneously presented in the background. During this conditioning phase, half of the participants had to perform an auditory two-back task involving numeric information communicated via headphones and responses produced on a keyboard (for a description of the procedure, see Pleyers et al., 2009). In a second-evaluation-step, participants were invited to rate the CSs. In addition to the eight CSs that were presented during the conditioning phase, eight additional CSs were included in this evaluation phase (previously and newly presented CSs were counterbalanced across participants). In this evaluation phase, participants had to rate each of the 16 CSs on a scale ranging from 1 (very negative) to 9 (very positive).
In a third-memory probe-step, we added 8 filler USs in addition to the 8 filler CSs that were not presented in the conditioning phase (old vs new USs were also counterbalanced across participants), resulting in 16 CSs and 16 USs in the 2 subsequent memory tasks. Participants were first asked for each of the 16 CSs and 16 USs to report if (1) it was presented during the conditioning phase, (2) it was not presented during the conditioning phase, or (3) they do not remember.
Finally, for each of the 16 CSs, participants were asked to identify which of the 16 USs was paired with the CS. On each trial, participants had the opportunity to answer that they did not remember. Memory for individual CSs and USs was probed before CS-US pairing memory to avoid confusions (e.g., participants reporting they saw a given CS that was not presented during the conditioning phase because it was presented in the pairing memory phase).

Analytical strategy
Data were analyzed using the ezANOVA, anovaBF, and lmBF functions (from the ' ez' and the 'BayesFactor' packages, (Lawrence, 2016;Morey, Rouder & Jamil, 2015) in R (R Development Core Team, 2017). We report the Bayes factors associated with the model comparison made in the frequentist analyses. 1 An augmented model containing the tested factor was compared to a constrained model not containing this factor. The Bayes factors in favor of the alternative hypothesis (or BF10) are presented when the conventional p-value of 0.05 is encountered. The Bayes factors in favor of the null hypothesis (or BF01) are reported when the p-value is above this threshold.

Load effect on CS-US pairing memory
We computed an accuracy score for the memory of the CS-US pairings. Correct responses were coded '1'; whereas, incorrect responses and ' don't know' responses were coded '0' at the item-level. We computed the proportion of correct responses for each participant. These proportion scores were submitted to a t-test as a function of the Depletion condition (see Figure 1B). Contingency memory was higher in the control (M = 0.75, SD = 0.27) than in the depletion (M = 0.1, SD = 0.16, t(79) = 13.14, p < 0.001, η 2 g = 0.69, BF10 > 1000) condition. Hence, again consistent with previous research, cognitive load strongly reduced CS-US pairing memory.

Load effect on the memory of individual CSs and USs
To probe the memory of the individual stimuli, we computed an accuracy score for the memory of the individual CSs and USs after coding them into one of four possible categories: true positive (an old stimulus reported as old), false positive (a new stimulus reported as old), true negative (a new stimulus reported as new), and false negative (an old stimulus reported as new). We computed an accuracy score for each participant with the following formula: Memory accuracy was submitted to a t-test as a function of the load condition (see Figure 1D). It was higher for participants in the control condition (M = 0.95, SD = 0.08) than in the depletion condition (M = 0.58, SD = 0.17, t(79) = 12.29, p < 0.001, η 2 g = 0.66, BF10 > 1000). Cognitive load thus substantially reduced memory accuracy for individual CS and US stimuli composing a pair.

Evaluative effects of load on correctly retrieved individual CSs and USs
We examined the role of US valence and Depletion condition on the evaluation of the subset of correctly retrieved 2 individual CS and US stimuli. We observed a US valence × Depletion interaction, F(1,73.22) = 5.96, p = 0.02, η 2 g = 0.06, BF10 = 6.01 ± 7.45%. EC was observed in the control condition, F(1,38.27) = 31.81, p < 0.001, η 2 g = 0.44, BF10 > 1000, but not in the depletion condition, F(1,30.57) = 0.63, p = 0.43, η 2 g = 0.00, BF01 = 3.29 ± 0.02% (see Figure 1C). Hence, and critical to the present research endeavor, cognitive load impaired EC even when the encoding of individual CS and US stimuli composing a CS-US pair was preserved. This reduction of EC to non-significance in the depletion condition could be due to a weak statistical power. Indeed, only 30 CS-US pairs in the depletion condition for which both the CS and US stimuli were correctly retrieved were available. The number of available data points, however, allows the detection of an effect size as small as η 2 g = 0.05 with a statistical power of (1-β) = 0.80 given the one-sided test (i.e., positive CSs are preferred to negative CSs) and a type I error probability of α = 0.05. As a matter of fact, the present study had 99% power for detecting an EC effect size of a magnitude similar to the one observed in the control condition (i.e., η 2 g = 0.44). Furthermore, the Bayesian analysis suggests 'substantial evidence in favor of the null-hypothesis', which adds to the proposal that the non-significant EC under the depletion condition is not due to data insensitivity (Wagenmakers et al., 2011).

Memory and EC
In a further-exploratory and non-preregistered-analysis, we examined CS evaluative ratings as a function of US valence and Pair types. Thus, we were able to test what type of memory (i.e., memory of individual CS, of individual US, and of their pairing) is necessary for EC. We first categorized each CS-US pair along five categories: zero-memory (neither the CS nor the US was correctly retrieved), CS-only memory pairs (the CS but not the US was retrieved), US-only memory (the US but not the CS was retrieved), CS-and-US memory (both CS and US were retrieved, but their pairing was not), and CS-US pairing memory (the CS, the US, and their pairing were correctly retrieved).

Discussion
Previous studies suggest load disrupts EC by hampering the encoding of CS-US pairing, making it a non-efficient attitude formation effect (Dedonder et al., 2010;Mierop et al., 2017;Pleyers et al., 2009). It was correctly pointed out, however, the inability to encode individual components of CS-US pairings under load, rather than the inability to encode their pairing, may have been responsible for the effect of load on the EC effect and CS-US pairing memory. Here, we observed the auditory two-back task reduced memory for individual CS and US stimuli, as well as for their pairing. However, we also found load reduces EC to non-significance even when individual stimuli entering a pairing are successfully retrieved. We more generally found EC is observed only when there is a successful retrieval of CS-US pairings. This finding, along with the mediation results obtained here, is fully consistent with previous research supporting the critical role of attention paid to CS-US pairings at encoding for establishing EC effects (e.g., Kattner, 2012;Stahl et al., 2016). In contrast, the present findings go against previous research conclud-ing the correct encoding of individual stimuli may be sufficient for EC to occur (Halbeisen & Walther, 2015). By relating EC effects to a Load manipulation and to memory measures for both CS-US pairings and individual CSs and USs, the present research constitutes the first empirical indication EC is non-efficient because it relies on a non-efficient encoding of CS-US relations. This, in turn, is consistent with a broader range of findings that question the automaticity of attitude formation (see Corneille & Stahl, 2019). Because dual-learning of attitude theories explicitly state the efficiency of associative attitude learning (e.g., Gawronski & Bodenhausen, 2014), the present findings challenge this theorization. Recently, however, attitude learning researchers have noted operating conditions may be agnostic to whether learning effects are driven by propositional or associative processes (De Houwer, 2018;Gawronski & Bodenhausen, 2018).
An important assumption we made is memory measures are valid indicators of encoding effects. Memory measures capture encoding, but also consolidation and retrieval effects. This may be a problem when looking at simple correlations between memory and EC measures (Gawronski & Walther, 2012;Sweldens, Corneille & Yzerbyt, 2014). However, the memory measures were not collected here in an experimental vacuum. They were examined as a function of a Load manipulation implemented during exposure to CS-US pairs. Because past research has demonstrated the immediate effect of a two-back task on encoding capacities (e.g., Jonides et al., 1997;Kane & Engle, 2002), we can confidently assume the memory measures we used here reliably captured a Note: A significant b-weight indicates the beta-weight and semi-partial correlation are also significant. b represents unstandardized regression weights. sr 2 represents the semi-partial correlation squared. LL and UL indicate the lower and upper limits of a confidence interval, respectively. * indicates p < 0.05. ** indicates p < 0.01. Load effect on encoding, one that was evidenced later at retrieval. As a matter of fact, the combined use of an experimental manipulation and memory measure has been specifically recommended for a validation of these measures as indicators of encoding effects (see Sweldens, Corneille & Yzerbyt, 2014). More generally, however, the present findings are fully consistent with a retrieval-based approach to evaluations, which states memory contents retrieved at the evaluation stage, which vary as a function of what information is encoded, underlie evaluations (Gast, 2018;Stahl & Aust, 2018).
One could argue the memory measures we used could be sensitive to affect-as-information heuristic, which may lead to memory estimates inflation. That is, when the valence of a US paired with a given CS cannot be retrieved, participants may rely on the valence conditionally acquired by the CS to infer US valence. In recent research, Hütter, Sweldens, Stahl, Unkelbach, and Klauer, (2012) argued valence memory performance may be contaminated by an affect-as-information heuristic. By relying on US identity measures, however, we strongly reduced an artefactually inflated pairing memory. It should also be noted, even assuming such problem may have to some extent arose, there is no reason to postulate it would have differentially influenced low and high load conditions. Finally, because identity measures are more conservative than valence identity measures (Stahl, Unkelbach & Corneille, 2009), our current choice was facilitating EC in the (presumed) absence of successful CS-US encoding. That is, CS-US pairs we considered unsuccessfully encoded (based on US identity measures) could actually have been successfully encoded (based on US valence measures). The current findings are consistent with previous research stressing the importance of processing the CS-US relation in EC effects Kattner, 2012;Mierop et al., 2019;Stahl et al. 2016) and suggest this processing is non-efficient. Future research may identify conditions under which attitude formation is efficient. It cannot be excluded efficient EC might be found when using other evaluative measures, sensory modalities, pairing procedures, or types of stimuli. To date, however, the detrimental impact of load on EC has been observed using indirect evaluative measures (Davies et al., 2012), conditioning paradigms presumably conducive to implicit misattribution (Mierop et al., 2017), unfamiliar CSs (Dedonder et al., 2010), and CS from both visual and gustatory modalities (Davies et al., 2012).

Notes
1 Priors and methods of computation are the defaults provided in Rouder, Morey, Speckman, and Province (2012). For the models reported here, the r scale values were set to ½, which corresponds to "medium" priors. 2 Due to the uneven proportions of observations across the different type of pairs, we report the F tests and degrees of freedom based on Kenward-Roger approximation. This analytical strategy was also used for the next and last statistical model.

Additional File
The additional file for this article can be found as follows: • Data Accessibility Statement. All materials (databases, scripts, measures, manipulations, and online appendix) are publicly available at. DOI: https://doi. org/10.5334/irsp.339.s1